elifeeLifeeLife2050-084XeLife Sciences Publications, Ltd9096410.7554/eLife.9096410.7554/eLife.90964.3Research ArticleComputational and Systems BiologyNeuroscienceMyelin dystrophy impairs signal transmission and working memory in a multiscale model of the aging prefrontal cortexIbañezSarahttps://orcid.org/0000-0001-7563-6968123†SenguptaNilapratimhttps://orcid.org/0000-0003-1024-559X14†LuebkeJennifer Ihttps://orcid.org/0000-0003-1399-60731WimmerKlaushttps://orcid.org/0000-0003-2973-346223‡WeaverChristina Mhttps://orcid.org/0000-0002-7744-4289christina.weaver@fandm.edu4‡https://ror.org/05qwgg493Department of Anatomy & Neurobiology, Boston University Chobanian & Avedisian School of MedicineBostonUnited Stateshttps://ror.org/020s51w82Centre de Recerca Matemàtica, Edifici C, Campus BellaterraBellaterraSpainhttps://ror.org/052g8jq94Departament de Matemàtiques, Universitat Autònoma de Barcelona, Edifici CBellaterraSpainhttps://ror.org/04fp4ps48Department of Mathematics, Franklin and Marshall CollegeLancasterUnited StatesOstojicSrdjanReviewing Editorhttps://ror.org/05a0dhs15École Normale Supérieure - PSLFranceGoldJoshua ISenior Editorhttps://ror.org/00b30xv10University of PennsylvaniaUnited States
Co-first authors: All authors consent that for their CV/Resume and other purposes, co-first authors may list themselves in either order
Normal aging leads to myelin alterations in the rhesus monkey dorsolateral prefrontal cortex (dlPFC), which are positively correlated with degree of cognitive impairment. It is hypothesized that remyelination with shorter and thinner myelin sheaths partially compensates for myelin degradation, but computational modeling has not yet explored these two phenomena together systematically. Here, we used a two-pronged modeling approach to determine how age-related myelin changes affect a core cognitive function: spatial working memory. First, we built a multicompartment pyramidal neuron model fit to monkey dlPFC empirical data, with an axon including myelinated segments having paranodes, juxtaparanodes, internodes, and tight junctions. This model was used to quantify conduction velocity (CV) changes and action potential (AP) failures after demyelination and subsequent remyelination. Next, we incorporated the single neuron results into a spiking neural network model of working memory. While complete remyelination nearly recovered axonal transmission and network function to unperturbed levels, our models predict that biologically plausible levels of myelin dystrophy, if uncompensated by other factors, can account for substantial working memory impairment with aging. The present computational study unites empirical data from ultrastructure up to behavior during normal aging, and has broader implications for many demyelinating conditions, such as multiple sclerosis or schizophrenia.
multicompartment modelbump attractormyelin dystrophyworking memoryagingrhesus monkeyResearch organismNonehttp://dx.doi.org/10.13039/100000049National Institute on AgingR01 AG059028LuebkeJennifer IMCIN/AEI/10.13039/501100011033 and the European Union "NextGenerationEU"/PRTRPCI2020-112035IbañezSaraWimmerKlaushttp://dx.doi.org/10.13039/501100011033Spanish State Research Agency, through the Severo Ochoa and María de Maeztu Program for Centers and Units of Excellence in R&DCEX2020-001084-MIbañezSaraWimmerKlaushttp://dx.doi.org/10.13039/501100008982National Science Foundation1925192WeaverChristina Mhttp://dx.doi.org/10.13039/501100007601Horizon 2020 - Research and Innovation Framework ProgrammeICEI project grant agreement No. 800858IbañezSaraWimmerKlaushttp://dx.doi.org/10.13039/100000049National Institute on AgingR01 AG071230LuebkeJennifer IThe funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.Author impact statementBiologically plausible levels of myelin dystrophy induce substantial working memory impairment in a computational model of brain aging across two spatial and temporal scales.publishing-routeprcIntroduction
Normal aging often leads to impairment in some cognitive domains, as evidenced by reduced performance on learning and memory tasks in both humans (Albert, 1993; Salthouse et al., 2003; Fisk and Sharp, 2004; Rhodes, 2004; Sorel and Pennequin, 2008) and non-human primates (Moore et al., 2006; Shamy et al., 2011; Moore et al., 2017; Comrie et al., 2018; Chang et al., 2022; Moore et al., 2023). In the rhesus monkey, age-related working memory decline is accompanied by sublethal structural and functional changes in vascular elements, individual pyramidal neurons, glial cells, and white matter pathways (reviews: Hof and Morrison, 2004; Luebke et al., 2010; Peters and Kemper, 2012; Morrison and Baxter, 2012). It is well documented that cortical neurons do not die during normal aging but rather undergo a number of morphological and physiological alterations, particularly in the monkey dorsolateral prefrontal cortex (dlPFC), the critical cortical circuit for working memory. For example, during normal aging Layer 3 pyramidal neurons in rhesus monkey dlPFC exhibit a significant loss of dendritic spines and synapses (Chang et al., 2005; Peters et al., 2008; Chang et al., 2022), and electrophysiological changes observed both in vitro (Chang et al., 2005; Ibañez et al., 2019; Chang et al., 2022) and in vivo during a spatial working memory task (Wang et al., 2011). Perhaps most strikingly, extensive myelin dystrophy during normal aging has been observed in both gray and white matter (Peters et al., 2001; Bowley et al., 2010; review: Peters, 2007), including in monkey dlPFC (Peters and Sethares, 2002; Peters and Sethares, 2003; review: Peters, 2009). Ultrastructural studies reveal that 3–6% of myelin sheaths in dlPFC exhibit age-related alterations including splitting of the major dense line of the myelin sheath, balloons, and redundant myelin (Figure 1; Peters and Sethares, 2002). Remyelination has also been observed across the adult lifespan in monkey dlPFC and there is a 90% increase in the number of paranodal profiles in aged versus young monkeys, indicating higher numbers of internodal myelin sheaths with aging (Peters and Sethares, 2003). Aged subjects also had a significant proportion of abnormally short and thin myelin sheaths (Peters and Sethares, 2003). The hypothesized mechanism to explain these findings (review:Peters, 2009) is that myelin degradation begins as oligodendrocytes degenerate due to oxidative stress, and that axons accumulate dense inclusions in spaces between the lamellae of their associated myelin sheaths. As oligodendrocytes die, the associated sheaths detach from the axolemma, leaving bare axonal sections (complete demyelination). Subsequently, surviving mature oligodendrocytes remyelinate the bare segments, but with shorter and thinner sheaths. It is highly plausible that the altered sheaths lead to a slowdown of signal propagation that contributes to cognitive slowing/impairment with aging. Indeed, several of the changes that pyramidal neurons undergo with aging correlate with the degree of observed cognitive impairment (review: Luebke et al., 2010; Peters and Kemper, 2012; Shobin et al., 2017; Moore et al., 2023), including myelin dystrophies and remyelination (Peters and Sethares, 2002; Peters and Sethares, 2003; Dimovasili et al., 2023). However, which changes are the key determinants of age-related cognitive decline has not yet been firmly established (Konar et al., 2016; Motley et al., 2018; Cleeland et al., 2019). This is in part due to the difficulty of isolating some individual neuronal features (e.g. firing rate, synapses, myelin) empirically, while controlling for concomitant changes in others. Thus, computational models become essential to predict how age-related changes in individual neurons affect cognitive impairment.
Electron photomicrographs (transverse sections) depicting age-related alterations in myelinated nerve fibers of area 46 of the rhesus monkey dorsolateral prefrontal cortex (dlPFC).
(A) Neuropil from a 10-year-old monkey. Healthy and compact myelin is visible as thick outlines surrounding nerve fibers which have been sectioned at their internodes. (B) Neuropil from a 27-year-old monkey. Arrows indicate dystrophic myelin surrounding nerve fibers, presenting a splitting of the major dense line of the myelin sheaths (left and right arrows) and balloons (left and middle arrows). Scale bar = 5 μm. Images are from the archives of Alan Peters and prepared as in Peters and Sethares, 2002.
Numerous modeling studies have explored the potential relationships between axon parameters and action potential (AP) conduction velocity (CV) (Rushton, 1951; Goldman and Albus, 1968; Brill et al., 1977; Moore et al., 1978; Waxman, 1980; Chomiak and Hu, 2009). Others have sought the most appropriate way to model the axon (e.g. Blight, 1985; Richardson et al., 2000; McIntyre et al., 2002; Gow and Devaux, 2008; Dekker et al., 2014). Demyelination has been modeled frequently in the context of disease (review: Coggan et al., 2015), showing that loss of myelin leads to slower AP CV and sometimes to AP failure. Remyelination of axons with shorter, thinner internodes decreases CV in models of both large- and small-diameter axons (Lasiene et al., 2008; Powers et al., 2012; Scurfield and Latimer, 2018). However, to our knowledge, the effects of widespread demyelination and remyelination on AP propagation in a broad population of axons have not been explored systematically. There have also been studies of how CV and myelin alterations impact network synchrony in phase oscillator models (Karimian et al., 2019; Pajevic et al., 2014; Noori et al., 2020; Pajevic et al., 2023), but not in spiking neural networks that can predict the mechanisms underlying working memory. Our recent network model revealed that the empirically observed age-related increase in AP firing rates in prefrontal pyramidal neurons (modeled through an increased slope of the f-I curve) and loss of up to 30% of both excitatory and inhibitory synapses (modeled as a decrease in connectivity strength) can lead to working memory impairment (Ibañez et al., 2019), but this model did not incorporate the known changes to myelin structure that occur during normal aging.
This study unites models at two different scales – multicompartment models of dlPFC pyramidal neurons and a spiking network model simulating spatial working memory – to investigate how age-related myelin degradation (represented by demyelination) and remyelination affect signal transmission and working memory precision. Higher degrees of demyelination led to slower propagation and eventual failure of APs along the axons of the multicompartment models. In the network models, an increase in AP failure rate resulted in progressive working memory impairment, whereas slower conduction velocities, in the range observed in the multicompartment models, had a negligible effect. Sufficient remyelination of all previously demyelinated segments led to a recovery of signal transmission and working memory performance. However, our study indicates that empirically observed levels of myelin changes, if uncompensated by other factors, would lead to substantial working memory impairment with aging.
ResultsProgressive demyelination causes CV slowing and AP failures in model neurons
To simulate myelin alterations in individual neurons, we adapted our multicompartment model tuned to data from rhesus monkey dlPFC (Rumbell et al., 2016) by attaching an axon model that captured nodes and detailed myelinated segments (Gow and Devaux, 2008; Scurfield and Latimer, 2018). Myelinated segments included an internode with adjacent juxtaparanodes and paranodes, and tight junctions between the innermost myelin lamella and axolemma (see Methods; Figure 2A and B). We applied demyelination and remyelination perturbations to a cohort of 50 young (control) neuron models, with axonal parameters varying within biologically plausible ranges (Table 1; Figure 2—figure supplement 1A and B). To simulate demyelination, we removed lamellae from selected myelinated segments; for remyelination we replaced a fraction of myelinated segments by two shorter and thinner segments with a node in between. As such, a ‘fully remyelinated axon’ had all the demyelinated segments subsequently remyelinated, but with fewer lamellae and additional nodes compared to the unperturbed control case, consistent with empirical observations (Peters, 2009). The CV in control models varied across the cohort and in response to myelin alterations (Figure 2C; Figure 2—figure supplement 1). Myelin alterations could also cause AP failures. CV changes and AP failures were more sensitive to variations along some dimensions of the parameter space than to others (e.g. myelinated segment length versus axon diameter), explored further below.
Action potential (AP) transmission in the single neuron model.
(A) Cartoon of the model with a close-up view of unperturbed, demyelinated, and remyelinated segments (not to scale). The paranodes, juxtaparanodes, and internodes (shown in different shades of red) were insulated by myelin lamellae, adjacent to unmyelinated nodes (dark gray). During demyelination, lamellae were removed from a subset of segments; middle cartoon shows two lamellae remaining, indicating 50% lamellae removed relative to an unperturbed myelinated segment. During remyelination, select myelinated segments were replaced with two shorter myelinated segments separated by a new node; bottom cartoon shows remyelination with 50% of lamellae restored relative to unperturbed segments. At right are shown membrane potential traces simulated at the initial segment (top, dashed line) and near the distal end of one axon (here, 1.9 cm long) in the unperturbed, demyelinated, and remyelinated cases. Traces correspond to signals in a distal node and subsequent paranode, juxtaparanode, and internode respectively (colors indicating the axonal sections as in left panels). Demyelinating 75% of segments by removing 50% of their lamellae resulted in a 70% reduction in conduction velocity (CV), and failure of one AP. Remyelination of all affected segments with the same 50% of lamellae recovered the failed AP, and 98% of the CV delay relative to the demyelinated case (in 1 of the 30 simulated trials). (B) Close-up view of an AP simulated in the distal end of the unperturbed axon: suprathreshold in the node and subthreshold along the myelinated segment, indicating saltatory conduction. (C) Distribution of the 50 models of the cohort across two dimensions of parameter space: myelinated segment length and axon diameter. Grayscale shade of each model represents the mean CV change across three demyelination conditions: 25%, 50%, 75% of segments losing lamellae, averaged over 30 randomized trials and lamellae removal conditions.
Distribution of parameters and conduction velocities in the single neuron model cohort.
(A) Histograms of axon morphology parameters of models selected for the single neuron cohort (n=50). Top: axon diameter; middle: length of unperturbed myelinated segments; bottom: total myelin thickness in unperturbed segments, computed as the product of lamella thickness and number of lamellae. (B) Histograms of the conduction velocity (CV) for the 50 axons of the unperturbed model cohort (top), and representative demyelination and remyelination perturbations: mild demyelination (removing 25% of lamellae from 25% of the myelinated segments, second row); severe demyelination (removing all lamellae from 75% of the myelinated segments, third row); and complete (100%) remyelination (where the demyelinated segments from the third row were remyelinated by two shorter segments with 75% of lamellae). CVs averaged over 30 trials in each case. (C) Changes in CV (measured in %) in response to demyelination and remyelination versus the magnitude of current clamp step (+180, +280, or +380 pA). Shown are mean ± SEM (n=50) for demyelinating 50% of myelinated segments (removing all lamellae), and subsequent remyelination of those segments by shorter segments with 75% of lamellae.
Axon parameter ranges for Latin hypercube sampling (LHS) construction.
Parameter
Values
Minimum
Maximum
Axon diameter (μm), measured at nodes
0.5
1.02
Node length (μm)
0.25
2.02
Myelinated segment length (μm)
50
200
Number of myelin lamellae
5
20
Lamella thickness (μm)
0.013
0.019
Scale factor for leak conductance
0.1
1
Scale factor for NaF maximal conductance
0.1
1
Scale factor for KDR maximal conductance
0.1
1
AP propagation was progressively impaired as demyelination increased (Figure 3): CV became slower, eventually leading to AP failure. Removing 25% of lamellae had a negligible effect on CV, regardless of how many segments were affected. However, when all lamellae were removed, CV slowed drastically – by 38±10% even when just 25% of the segments were demyelinated in this way, and 35±13% of APs failed. When 75% of segments lost all their lamellae, CV slowed by 72±8% and 45±13% of APs failed. Responses to demyelination sometimes varied widely across the cohort. We employed Lasso regression to identify key parameters that contributed to CV changes, since those changes preceded AP failures (Figure 3—figure supplement 1A and C). Five of the 12 parameters analyzed contributed to CV changes during demyelination. Among them, myelinated segment length had the largest magnitude with a negative weight indicating that models with longer myelinated segments show more CV slowing in response to a given demyelination perturbation. Scale factors for leak and sodium conductance, axoplasm resistance, and tight junction resistance also controlled CV changes during demyelination.
Effects of demyelination on conduction velocity (CV) and action potential (AP) failures in the single neuron model.
(A) Heat maps showing CV change (reduction relative to the CV of the corresponding unperturbed models, measured in %) in response to select demyelination conditions across the 50 cohort axons (see Methods). Axons arranged vertically in increasing order of myelinated segment length (longest at the bottom). The three blocks from left to right show increasing numbers of demyelinated segments in each axon (25%, 50%, and 75% of segments respectively), illustrated by cartoons on top. Within each block, individual columns correspond to the percentage of myelin lamellae removed from each demyelinated segment (shown in cartoons below). Color of each box indicates the mean CV change across 30 trials of each condition, ranging from 0% (no effect) to –100% (AP failure). Overall, AP propagation was increasingly impaired with increasing levels of demyelination. Mean CV change (B) and percentage of AP failures (C) versus the percentage of lamellae removed for all demyelination conditions simulated. Colors represent the percentages of segments demyelinated, from 10% (light red) to 75% (black). Error bars represent mean ± SEM, averaged across all cohort axons (n=50) and 30 trials each.
Statistical analysis of parameters contributing to conduction velocity (CV) changes after demyelination and remyelination.
Coefficients of Lasso regression models with 10-fold cross-validation for demyelination (A) and remyelination (B). Parameters with non-zero coefficients are important factors underlying the response, critical in ascertaining the susceptibility of axons to respective perturbations. (C–D) The Lasso models effectively predicted how demyelination and remyelination affect CV. The models from (A–B) were applied to a novel test set (n=50 axons). Shown are predicted versus observed CV changes (z-scored; slowdown due to demyelination in (C), recovery due to remyelination in (D)) for the 50 novel axons. Adjusted R2=0.61 (C) and 0.87 (D) respectively.
Remyelination leads to partial recovery from CV slowing and AP failures
We next examined the extent to which remyelination with shorter and thinner segments, occurring after demyelination, restored axonal AP propagation (Figure 4). We first assumed that affected segments had been previously completely demyelinated, i.e., losing all their lamellae (see Methods; Figure 4A–C). Most remyelinated models showed CV recovery from 0% to 100%, except for a few cases in which CV increased relative to the unperturbed models (Figure 4—figure supplement 1A; see Discussion). The CV recovered more as both the remyelination and the lamellae restoration percentages increased. When all demyelinated segments were subsequently remyelinated with sufficient lamellae – and none of the perturbed segments were bare – the CV recovered substantially and almost no AP failed (Figure 4B and C). The initial fraction of demyelination also affected CV recovery, but in a more subtle way. When all demyelinated segments were remyelinated, there was a positive relationship between the initial demyelination rate and the CV recovery: CV recovered more when 75% of the segments were demyelinated (Figure 4B, black lines) than when only 25% were affected. This finding is consistent with observations of Scurfield and Latimer, 2018, in which axons with more transitions between long (unperturbed) and short (remyelinated) segments had slower CV (Figure 4—figure supplement 2). When incomplete remyelination left some segments bare (Figure 4B, colored lines), there was a negative relationship between the initial amount of demyelination and CV recovery: axons with more bare segments had reduced electrical insulation, and therefore recovered less.
Conduction velocity (CV) recovery in response to remyelination.
(A) Cartoons illustrating representative remyelination conditions after select segments were completely demyelinated. Top row shows an unperturbed axon with eight myelinated segments. Second row: 50% of segments are completely demyelinated. Third row: 25% of the demyelinated segments in second row (one in total) are remyelinated with two shorter segments, each with 25% of lamellae restored. Fourth row: 75% of the demyelinated segments in the second row (three in total) are remyelinated with two shorter segments, each with 50% of lamellae restored. Mean CV recovery (B) and percentage of action potential (AP) failures (C) versus the percentage of lamellae restored for all simulated remyelination conditions after complete demyelination. (D) Cartoons illustrating representative remyelination conditions after partial demyelination. Top row shows an unperturbed axon with eight myelinated segments. Second row: 50% of segments are partially demyelinated (with 50% of lamellae removed). Third row: 50% of the demyelinated segments in second row (two in total) are remyelinated with two shorter segments, each with 50% of lamellae restored. Mean CV recovery (E) and percentage of AP failures (F) versus the percentage of lamellae restored for all simulated remyelination conditions after partial demyelination. CV recovery in both cases (B and E) was calculated with respect to the CV change for the complete demyelination (see Methods). In panels (B, C, E, and F), the x-axis refers to the percentage of myelin lamellae restored relative to unperturbed segments, starting at 0% (no remyelination). Line styles represent the percentage of segments initially demyelinated, from 25% (dashed) to 75% (thick solid). Colors represent the extent of remyelination, from 25% (light red) to 100% (black). Shown are mean values, averaged across all cohort axons (n=50) and 30 trials each. For readability, error bars (representing ± SEM) are shown only for the condition of 50% demyelination of segments.
Conduction velocity (CV) recovery in response to remyelination across the model cohort.
Axons arranged vertically in increasing order of myelinated segment length (longest at the bottom). The three groups of heat maps from left to right represent how many segments were demyelinated initially (25%, 50%, and 75%, respectively). Within each group, the position in each of the four blocks indicates what proportion of demyelinated segments were remyelinated. Individual columns correspond to the percentage of lamellae restored to each remyelinated segment. Color of each box indicates the mean CV recovery from the corresponding demyelinated case across 30 trials (see Methods). (A) Recovery after complete demyelination, when all lamellae had been removed from affected segments. Columns highlighted in blue and green respectively correspond to the two remyelination cases shown in Figure 4A. CV recovery was positive, representing improvement, in all remyelination cases except one. For axon 22, the CV was worse for the mildest remyelination situation: when 25% of segments were initially demyelinated, then 25% of those segments were remyelinated by restoring just 25% of the myelin lamellae. A few axons showed recovery above 100%, suggesting faster conduction than in the unperturbed condition, when either 75% of the lamellae were restored. (B) Recovery after partial demyelination, when half of lamellae had been removed from affected segments. Column highlighted in green corresponds to the remyelination case shown in Figure 4D. CV recovery was positive, representing improvement, in all remyelination cases except one (axon 46, far left simulation condition).
Transitions between myelinated segments of dissimilar lengths affect response to perturbations.
(A) Mean conduction velocity (CV) change (measured in %) versus the number of transitions from unperturbed to demyelinated segments in 30 randomized trials for a given demyelination condition (50% of the segments affected, 50% of lamellae removed). (B) Mean CV change (measured in %) versus the number of transitions from unperturbed (long) to remyelinated (short) segments in 30 randomized trials for complete remyelination of 50% of the segments with 50% of lamellae restored. CV change was more severe as the number of transitions between segments of unequal lengths increased.
We also simulated remyelination after a milder partial demyelination, where affected segments initially lost only half their lamellae (Figure 4D). Overall trends were similar to, but less severe than, those for the complete demyelination case (Figure 4E and F; Figure 4—figure supplement 1B). The variability in CV recovery across different remyelination conditions and across the model cohort was similar. There were also fewer AP failures under partial demyelination conditions, relative to the corresponding complete demyelination cases (Figure 4C vs. Figure 4F).
Results for the percentage of AP failures (Figure 4C and F) were consistent with those for CV recovery. Remyelinating all previously demyelinated segments, even adding just 10% of lamellae, brought AP failure rates down to 14.6±5.1%. Remyelinating all affected segments with 75% of lamellae (the maximal amount of remyelination) nearly eliminated AP failures (1.8±1.1%). Incomplete remyelination, where some segments were still demyelinated, still had relatively high AP failure rates. For example, when one eighth of segments were remyelinated with the maximal amount of lamellae and one eighth were left bare, 25.7±11.5% of APs failed across the cohort (Figure 4C, red dashed line and arrow). AP failure rates were slightly lower when starting with partial demyelination: 10.6±7.6% of APs failed in the analogous paradigm (Figure 4F, red dashed line and arrow). In short, combinations of demyelinated and remyelinated segments often led to sizable CV delays and AP failures. Applying Lasso regression to CV recovery after remyelination (Figure 3—figure supplement 1C and D) found 10 of the 12 parameters contributed significantly; only myelin length and axoplasm resistance were omitted. Comparing these parameters with empirical data, when available, may help estimate the severity of CV delays and AP failures in the cortex of aging rhesus monkeys.
AP failures impair performance in a neural network model of working memory
Equipped with the quantification of impaired AP transmission due to myelin alterations in the single neuron model, we next elucidated how these impairments affected neural circuit function in PFC. We focused on spatial working memory because the underlying neural network mechanisms have been studied and modeled in detail (e.g. Compte et al., 2000; Hansel and Mato, 2013; Ibañez et al., 2019). We built on a previous spiking neural network model that accounts for many experimental findings (Hansel and Mato, 2013). It consists of 16,000 excitatory and 4000 inhibitory integrate-and-fire neurons (see Methods). Neurons are coupled through excitatory synapses with AMPA and NMDA receptors, and inhibitory synapses with GABA receptors. Recurrent excitatory synapses are facilitating, as has been empirically observed in PFC (Hempel et al., 2000; Wang et al., 2006), which promotes robust and reliable persistent activity despite spatial heterogeneities in the connectivity or in the intrinsic properties of the neurons.
We first simulated the classical oculomotor delayed response task (DRT, Figure 5A) in a cohort of 10 young (control) networks with different random connectivities and intact AP transmission (see Methods). For appropriate levels of excitation and inhibition, a localized activity bump forms during the cue period, and this bump persists through the delay period. The center of the bump encodes the remembered spatial location (Figure 5B, i; Compte et al., 2000; Hansel and Mato, 2013). Successful trials require a sufficiently strong activity bump throughout the delay period, quantified by the memory strength (Figure 5C, blue line). If the memory strength decreases over time (e.g. caused by the demyelination/remyelination conditions discussed below) the memory duration – the period during which the network can retain the stimulus– becomes limited (Figure 5C). Moreover, due to random fluctuations, the activity bump diffuses along the network during the delay period. This leads to trial-to-trial variability in the cue position read out from the network activity, modeling the variability of recalled spatial locations observed empirically (Figure 5B, right panels; Wimmer et al., 2014). This memory diffusion increases with the delay duration, consistent with decreasing working memory precision observed experimentally (Figure 5D; Funahashi et al., 1989). The bump movement during the delay also has a directed component, i.e., a systematic bias clockwise or counterclockwise away from the cue location, caused by heterogeneities in the network connectivity (Figure 5B, right panels and Figure 5E; Hansel and Mato, 2013). This memory drift is a possible explanation for delay-dependent systematic biases in working memory observed experimentally (see Discussion). However, because of their established relationship with working memory performance, in the following we focus on memory duration, corresponding to complete forgetting, and memory diffusion, corresponding to working memory precision.
Action potential (AP) failures impair working memory performance in a spiking neural network model.
(A) Schematic of the delayed response task. Subjects fixate at the center of a computer screen and need to remember a cue stimulus, presented at one out of eight locations throughout the delay period, before indicating the remembered location with an eye movement. (B) Excitatory neuron activity for a cue stimulus presented at 135° of an (i) unperturbed control network, (ii) a network with demyelination, and (iii) a network with remyelination. Left: Single-trial raster plot showing the activity for each neuron (labeled by its preferred direction) during the precue (fixation), cue and delay periods of the task. The cue period is indicated by the gray shading. Middle: Average spike counts of the excitatory neurons during the delay period. The points show average spike rates of individual neurons and the solid line the average over 500 nearby neurons. Right: Trajectory of the bump center (i.e. the remembered cue location) read out from the neural activity across the cue and delay periods using a population vector (see Methods). Thin lines correspond to individual trials and the solid line to the trial average. (ii) Shows the effect of AP failure probabilities corresponding to demyelination of 25% of the myelinated segments by removing 75% of the myelin lamellae. (iii) Corresponds to AP failure probabilities for remyelination of 50% of the demyelinated segments by adding 75% of the myelin lamellae back, after previous partial demyelination of 25% of the segments. (C) Memory strength as a function of time and corresponding memory duration (horizontal bars; memory strength ≥0.4; see Methods). (D) Working memory diffusion (trial-to-trial variability of bump center) during the cue and delay periods. The inset shows a close-up of the diffusion for control networks. A similar increase of working memory diffusion with demyelination is also observed in networks with overall higher diffusion (Figure 5—figure supplement 1). When demyelination is restricted to a part of the network, diffusion only increases in the perturbed zone (Figure 5—figure supplement 3). (E) Working memory drift (systematic memory errors). Note that the remyelination curve (purple dotted line) in (E) superimposes the young curve (blue solid line). The red dashed line represents the demyelination case. The performance measures in (C–E) were obtained by averaging across 280 trials and 10 networks, either control (B, i) or perturbed (B, ii–iii).
Increased working memory diffusion in spiking networks with spatially correlated background inputs.
Bump position during the cue and delay periods for 280 trials and the eight possible cue directions for (A) a young, control network and (B) a perturbed network when remyelinating 50% of the segments after partial demyelination of 25% of the segments along the neuronal axons, by adding 75% of the myelin lamellae back. Simulations were done with spatially uncorrelated background inputs as in all other simulations in Figures 5—8 (left panels), and with spatially correlated background inputs (see Supplementary Methods; right panels). These simulations show that, as expected from theory, bump diffusion increases for spatially modulated noise correlations (Rosenbaum et al., 2017; Stein et al., 2021). Importantly, the effect of myelin alterations is an increase in diffusion in both cases, showing the robustness of our results (diffusion constant = 0.018 deg2/s in (A) left panel; 0.957 deg2/s in (B) left panel; 3.115 deg2/s in (A) right panel; and 9.605 deg2/s in (B) right panel).
Effect of propagation delays on control and perturbed networks.
(A) Memory strength (left panels) and diffusion (right panels) for the young, control networks with zero propagation delays (blue solid line), as in Figure 5, and with propagation delays from a uniform distribution with a range between 0 and 100 ms (yellow dashed line). (B) Memory strength and diffusion for perturbed networks when demyelinating 50% of the segments along the axons of model neurons, by removing 60% of the myelin lamellae without delays (red solid line), and with delays from a uniform distribution with a range between 0 and 40 ms (gray dashed line) and between 0 and 85 ms (black dash-dotted line). The measures of working memory performance were calculated by averaging across 20 networks and 280 trials for each network. Shaded areas indicate SEM for each case. For the young, control networks, there was no difference with and without propagation delays, even though the delays used in the network simulations were much larger than the delays quantified in the single neuron model (the longest delays found for the most extreme perturbation condition – demyelination of 75% of the segments by removing 100% of the myelin lamellae – were of 49.9 ms on average; A). Working memory performance was also unaffected in the perturbed network with action potential (AP) failures for delays ranging between 0 and 40 ms, also larger than the ones quantified in the single neuron model (for the case of 50% of the segments demyelinated by removing 60% of the myelin lamellae, the average delay in the cohort was 4.6 ms and the maximum delay was 15.7 ms; B). However, including extremely long delays of up to 85 ms did further impair memory compared to the impairment level introduced by AP failures alone (B).
Effect of spatially heterogeneous demyelination of the model neurons according to their preferred angle.
We also tested working memory performance in the network when demyelination affects only parts of the network. The figure shows the decoded bump center position during the cue and delay period for the eight possible cue directions when a fraction of neurons was perturbed and the rest of the neurons in the circuit were unaltered (Figure 5B). We perturbed 10% of the neurons around the neuron with preferred direction 90° (left panel), 25% of the neurons around –90° (middle panel), and 50% of the neurons around 180° (right panel). Bump traces for cues that lie inside the perturbed portion of the circuit are shown in blue. Network perturbation in the three cases consisted in demyelinating 25% of the segments along the axons of model neurons, by removing 70% of the myelin lamellae. In each case, 280 trials were simulated for one network. These simulations show an increased drift and diffusion inside the perturbed zone, consistent with the increased drift and diffusion when perturbing the entire network (Figure 6B and Figure 6—figure supplement 2). In particular, spatially heterogeneous demyelination in our network leads to a bias away from the affected zone and to increased trial-to-trial variability. Note that this is a model prediction, but we are not aware of empirical data showing heterogeneous demyelination with aging. Further, note that while our network model has a topological ring structure, neurons in PFC are not anatomically arranged depending on their preferred features. Thus, spatially heterogeneous demyelination would likely affect neurons with different feature preferences (i.e. neurons throughout our ring model).
Action potential (AP) failures impair working memory performance in a network model with activity-silent memory traces.
(A) Spiking and synaptic activity in an unperturbed, activity-silent working memory model. Top: Raster plot showing the activity for each excitatory neuron (labeled by its preferred direction) in a single trial with a cue stimulus presented at 180°. We modified our spiking neural network model such that it does not show elevated persistent firing throughout the delay period (see Figure 5B for comparison). In particular, we reduced the external background input to excitatory neurons IEext by a factor of 3.61% and we increased the cue stimulus amplitude by 12.5%. Even though spiking activity decays to baseline (close to 0 Hz), a memory trace is imprinted in enhanced synaptic strength due to short-term synaptic facilitation (Mongillo et al., 2008). Selective spiking activity is recovered by a non-selective constant input applied during 300 ms to all excitatory neurons during the two reactivation periods (marked by yellow and green rectangles in the raster plot). The amplitude of the input was 11 mV during the first and 13 mV during the second reactivation period. Reactivation periods are marked in light gray shading in the remaining panels below and the cue period is indicated by dark gray shading. Firing rates (second row), synaptic facilitation variable u (third row), and synaptic depression variable x (bottom row) for the same trial, averaged for 500 neurons around the neuron with 180° as preferred direction (solid lines) and around the neuron with 0° as preferred direction (dashed lines). Note that reactivation recovers the activity bump (C) but also causes elevated firing and subsequent enhancement of synapses at all positions in the networks. (B) Activity in a network with demyelination of 50% of the myelinated segments by removing 60% of the myelin lamellae. AP failures lead to reduced firing rates in the cue and early delay periods and consequently to weaker synaptic enhancement. (C) Average spike counts of the excitatory neurons during the cue period (black lines), and the two reactivation periods indicated in the raster plots in (A) and (B) (yellow and green lines). Solid lines correspond to the control network and dashed lines to the perturbed network. (D) Memory strength as a function of time for the control and perturbed networks. (E–F) Trajectories of the bump center (i.e. remembered cue location) read-out from the neural activity across the cue and delay periods using a population vector (see Methods). Cue position was 180° in all trials. The perturbed network (F) shows larger working memory errors toward the end of the delay period compared to the control network (E).
Effect of propagation delays on control and perturbed activity-silent network models.
(A) Memory strength during the whole simulation time for the young, control networks relying on activity-silent working memory (Figure 5—figure supplement 4) with zero propagation delays (blue line), and with propagation delays from a uniform distribution with a range between 0 and 40 ms (yellow line) and between 0 and 100 ms (orange line). (B) Memory strength for perturbed networks when demyelinating 25% of the myelinated segments by removing 50% of the myelin lamellae, without delays (red line), and with uniformly distributed delays between 0 and 40 ms (light gray line) and between 0 and 100 ms (black line). The cue period is indicated by dark gray shading and reactivation periods are marked in light gray. Memory strength was calculated by averaging across 280 trials for one network. Shaded areas indicate SEM for each case. For the young, control networks (A), working memory was not affected by including delays of up to 40 ms. Unrealistically long delays ranging up to 100 ms did cause an impairment (the longest delays found for the most extreme perturbation condition – demyelination of 75% of the segments by removing 100% of the myelin lamellae – were of 49.9 ms on average). When also incorporating AP failures to the networks (B), we observed a similar trend. For this perturbation condition, delays of up to 40 ms were already much larger than the delays quantified in the single neuron model (for the case of 25% of the segments demyelinated by removing 50% of the myelin lamellae, the average delay in the cohort was 3.75 ms).
To investigate how myelin alterations affect working memory maintenance, we explored in the network model the same demyelination and remyelination conditions as we did in the single neuron model. Because our network model consists of point neurons (i.e. without detailed axons), we incorporated CV slowing as an effective increase in synaptic transmission delays (see Methods). To simulate AP failures, we adjusted the AP failure rate to the values given by the single neuron model, by creating a probabilistic model of spike transmission from the excitatory presynaptic neurons to both the excitatory and inhibitory postsynaptic neurons (see Methods). We found that propagation delays, even larger than those quantified with the single neuron model, had no effect on the network performance. Only unrealistically long delays led to a slight decrease in performance (Figure 5—figure supplement 2). This was as expected because, by design, the network operates in an asynchronous state with irregular neural activity in which the timing of individual spikes does not affect network function. AP failures, on the other hand, did have a large impact. For AP failure probabilities matched to the distribution of AP failure probabilities across the cohort of 50 single neurons above, we observed a decay of the activity bump (i.e. reduced memory strength over time) in the network model. This can ultimately lead to extinction of the activity bump during the delay period, representing complete forgetting of the remembered stimulus (reduced memory duration; Figure 5B, ii and iii, Figure 5C, and Figure 6—figure supplement 1A). For a mild reduction of the bump amplitude, memory duration was not affected (Figure 5C, purple line) but memory diffusion increased compared to the control network (Figure 5B, iii and Figure 5D). This increase in memory diffusion is consistent with mathematical analysis of firing rate models showing that the bump diffusion depends inversely on the squared bump amplitude (Kilpatrick and Ermentrout, 2013; Esnaola-Acebes et al., 2022).
Impact of demyelination and remyelination on working memory
We then systematically characterized changes in memory duration and memory diffusion by comparing working memory performance in the cohort of 10 control networks with the performance of those networks perturbed corresponding to varying degrees of demyelination and remyelination. In each of the 10 networks, we set the AP failure rate of the excitatory neurons according to the distribution of failure probabilities of the neurons in the single neuron cohort for the given demyelination or remyelination condition. Thus, we took into account the heterogeneity of demyelination and remyelination effects from our single neuron cohort (Figure 3A; Figure 4—figure supplement 1). Note that this heterogeneity originates from differences in axon properties, but probabilities of failure for all neurons in the network correspond to the same degree of demyelination (Figure 6). We will also consider networks that contain different combinations of axons with either intact or perturbed myelin (Figure 7 and Figure 8).
Working memory function in the network model is impaired by demyelination and recovered by sufficient remyelination.
(A) Memory duration and (B) diffusion constant for simulations of the delayed response task, as in Figure 5, for a systematic exploration of the effect of action potential (AP) failure probabilities corresponding to the different demyelination and remyelination conditions explored with the single neuron model. Left panel: Demyelination, realized by removing a fraction of myelin lamellae from a fraction of myelinated segments. Middle panel: Remyelination with two shorter and thinner myelin sheaths, with a node in between, of the previously completely demyelinated segments. Right panel: Same as the middle panel but for partial demyelination (removal of 50% of the myelin lamellae) rather than complete demyelination. In all cases, the performance measures were obtained by averaging across the 10 perturbed cohort networks and the 280 trials simulated for each network. The average memory duration for the 10 unperturbed, control networks in the cohort (averaged across 280 trials) was 4 s, and the average diffusion constant was 0.064 (both values corresponding to the case of 0% of myelin lamellae removed in the left panels of (A) and (B), respectively; not shown). Error bars represent mean ± SEM, averaged across all networks and trials.
Memory strength decreases for different degrees of demyelination and remyelination.
(A) Progressive memory strength reduction, due to higher degrees of demyelination (left panel) and lower degrees of remyelination (right panel), leads to a progressive shortening of the memory duration. Left panel: Demyelination of 25% of the myelinated segments by systematically removing myelin lamellae. Removing up to 50% of the lamellae causes no effect on the memory strength compared to the control networks (see Figure 5C; memory duration = 4 s). However, removing over 75% of the myelin lamellae progressively shortens memory duration (<2 s). Right panel: Systematic remyelination of the previously partially demyelinated 25% of the segments. Specially adding back more myelin lamellae recovers the memory strength and thus, the memory duration. Complete remyelination leads to control-like values (black lines; Figure 5C). (B) Memory strength at the end of the delay period for simulations of the delayed response task (DRT), for a systematic exploration of the effect of AP failure probabilities corresponding to the demyelination/remyelination conditions explored with the single neuron model (see Figure 6). Left panel: Demyelination. Middle panel: Remyelination of the previously completely (removal of 100% of the myelin lamellae) demyelinated segments. Right panel: Remyelination of the previously partially (removal of 50% of the myelin lamellae) demyelinated segments. In (A) and (B) memory strength was obtained by averaging across the 10 perturbed cohort networks and the 280 trials simulated for each network. The average memory strength for the 10 control networks in the cohort (averaged across 280 trials) was 0.765, corresponding to the case of 0% of lamellae removed in the left panel of (B). Error bars in (B) represent mean ± SEM, averaged across all networks and trials.
Increase of memory drift for different degrees of demyelination and remyelination.
Drift rate for simulations of the delayed response task (DRT), for a systematic exploration of the effect of action potential (AP) failure probabilities corresponding to the demyelination/remyelination conditions explored with the single neuron model (see Figure 6). Left panel: Demyelination. Middle panel: Remyelination of the previously completely (removal of 100% of the myelin lamellae) demyelinated segments. Right panel: Remyelination of the previously partially (removal of 50% of the myelin lamellae) demyelinated segments. Drift rate was obtained by averaging across the 10 perturbed cohort networks and the 280 trials simulated for each network. The average drift rate for the 10 control networks in the cohort (averaged across 280 trials) was 2.075 deg/s, corresponding to the case of 0% of lamellae removed in the left panel. Error bars represent mean ± SEM, averaged across all networks and trials.
Reduced normal myelin is associated with decreased working memory performance in the network model.
(A) Schematic of the quantification of unperturbed, normal myelin sheaths in groups of neurons containing intact and demyelinated axons with different proportions of demyelinated segments (see Methods). Vertical red lines indicate cross-sectional planes that mimic electron microscopy images capturing cross sections of different axonal parts. (B) Memory duration and (C) diffusion constant vs. the percentage of normal myelin sheaths. Linear regressions show significant positive correlations in both cases (memory duration: r=0.703, p=3.86 × 10–10; diffusion constant: r=–0.802, p=1.26 × 10–14). Circles: All the demyelinated segments in the perturbed axons in the groups were bare segments (all myelin lamellae removed). Squares: All the demyelinated segments in the perturbed axons had 75% of the myelin lamellae removed. Black horizontal bars indicate the percentage of normal sheaths observed in electron microscopy images from young and aged rhesus monkeys dorsolateral prefrontal cortex (dlPFC) (Peters and Sethares, 2002).
A higher proportion of new myelin sheaths impairs working memory in the network model.
(A) Schematic of the quantification of new myelin sheaths in groups of neurons containing intact and partly remyelinated axons. Vertical purple lines indicate cross-sectional planes that model electron microscopy images capturing cross sections of different axonal parts. (B) Memory duration and (C) diffusion constant vs the percentage of new myelin sheaths. Linear regressions show significant negative correlations in both cases (memory duration: r=–0.852, p=4.92 × 10–7; diffusion constant: r=0.607, p=0.003). The remyelinated axons in the groups have different proportions of segments remyelinated after partial demyelination, by adding 25% of the myelin lamellae back. Black horizontal bars indicate the percentage of paranodal profiles observed in electron microscopy images from young and aged rhesus monkeys dorsolateral prefrontal cortex (dlPFC) (Peters and Sethares, 2003).
Demyelination impairs working memory performance compared to control networks
We found that the memory duration was not affected when removing up to 55% of myelin lamellae per myelinated segment, regardless of the percentage of axonal segments that were altered along an axon (Figure 6A, left panel). However, when between 55% and 75% of the myelin lamellae were removed, the memory duration began to decrease, depending on the percentage of axonal segments that were demyelinated. In this case, increasing the percentage of demyelinated segments from 10% to 50% led to a progressive impairment, whereas an increase to 75% of the segments did not impair memory duration further. Finally, in cases where 100% of the myelin lamellae were removed, the memory duration dropped to ≤1 s, regardless of the percentage of segments that were demyelinated. In a similar trend, memory diffusion increased, i.e., working memory became less precise starting when removing between 25% and 50% of the myelin lamellae (Figure 6B, left panel). Again, we found a progressive impairment, depending on the percentage of myelin lamellae removed and the percentage of myelinated segments affected, with a ceiling effect when more than 50% of segments were demyelinated.
Complete remyelination recovers network function
We observed that remyelination of all previously demyelinated segments (100%), independently of the degree of demyelination, recovered the memory duration to the control networks-like performance (Figure 6A, middle and right panels; black lines). However, working memory precision is not fully recovered in all these cases, indicated by an increase in the diffusion constant (Figure 6B). The performance of control networks was completely recovered only when 75% of the myelin lamellae were added back to the remyelinated segments. Thus, despite the new shorter and thinner myelin sheaths compared to the original intact ones, complete remyelination is able to recover control, unperturbed network function.
Incomplete remyelination leads to partial recovery
We studied the effect of incomplete remyelination (remyelination of 25–75% of previously demyelinated segments) after both complete and partial demyelination (Figure 6A and B, middle and right panels, respectively). When we remyelinated between 25% and 75% of the previously completely demyelinated segments, we did not observe a significant recovery of the memory duration (Figure 6A, middle panel; memory duration ≲ 1 s in all cases). Memory diffusion was partly recovered compared to the complete demyelination case (compare left and middle panels in Figure 6B) when 50–75% of the demyelinated segments were remyelinated, but it remained far from the performance of the control network. In sum, incomplete remyelination was unable to restore network function when bare axon (completely demyelinated) segments were present. However, when we remyelinated between 25% and 75% of the previously only partially demyelinated segments, both memory duration and memory diffusion were restored closer to the values of the control networks (Figure 6A and B, right panel).
Alternative working memory mechanisms
Working memory in our neural network is maintained in an attractor state with persistent neural activity (Compte et al., 2000; Hansel and Mato, 2013). Other mechanisms have been proposed, including that working memory maintenance may rely on activity-silent memory traces (Mongillo et al., 2008; Stokes, 2015; Barbosa et al., 2020). In activity-silent models, a slowly decaying transient of synaptic efficacy preserves information without the need for persistent ongoing activity. We implemented an activity-silent model, to our knowledge the first one for continuous spatial locations, and tested how working memory performance is affected by AP failures and propagation delays. We found that AP failures corresponding to demyelination caused working memory errors qualitatively similar to the delay-active network (Figure 5—figure supplement 4). On the other hand, increasing propagation delays did not lead to additional working memory errors, unless we include unrealistically high values (uniform distribution in the range of 0–100 ms; Figure 5—figure supplement 5). These results are qualitatively similar to the delay-active network model. Thus, our main findings do not critically depend on the exact working memory mechanism (active vs. activity-silent).
Simulated heterogenous myelin alterations match empirical data
Up to this point we have studied network models with AP failure probabilities corresponding to a single degree of myelin alterations (i.e. with all excitatory neurons in the network having AP failure rates matched to those of the single neuron cohort for one particular demyelination or remyelination condition). Next, we sought to reveal the effect on working memory performance of more biologically realistic network models, where excitatory neurons in the networks were perturbed according to a combination of different demyelination or remyelination conditions. That is, we simulated networks with excitatory neurons having AP failure probabilities matched to both neuronal axons with intact and with altered myelin sheaths in different degrees, as likely occurs in the aging brain (Figure 1).
Fewer normal myelin sheaths lead to decreased network performance
We ran network model simulations combining AP failure probabilities corresponding to groups of neurons containing either intact axons or axons presenting different degrees of demyelination (Figure 7A; Methods). Quantifying the average degree of demyelination in each simulated network allowed us to predict working memory deficits for a degree of demyelination that is within the empirically observed range of 90–100% normal myelin (Peters and Sethares, 2002). We observed that the performance was impaired – memory duration significantly decreased (Figure 7B), and memory diffusion significantly increased (Figure 7C) – when the percentage of normal sheaths decreased. These results are consistent with an experimentally observed increased cognitive impairment in various learning and working memory tasks (including the delayed recognition span task, a spatial working memory task) with an age-related decrease of the percentage of normal sheaths in dlPFC of rhesus monkeys (Peters and Sethares, 2002). Importantly, our results indicate that myelin alterations alone can account for significant working memory impairment, pointing to demyelination as a key factor in age-related working memory decline.
Shorter and thinner myelinated segments impair working memory
To predict the effects of remyelination on working memory for the empirically observed range of axon remyelination, we simulated network models that contained a combination of model neurons with intact axons and with axons containing different proportions of remyelination (Figure 8A; Methods). Empirical studies have found an age-dependent increase in the percentage of paranodal profiles, indicative of more and shorter myelin sheaths, from 5% in young monkeys to 17% in aged monkeys (Peters and Sethares, 2003). Here, we used different proportions of incomplete remyelination as studied in Figure 6 (axons with either 25%, 50%, or 75% of their segments remyelinated), and we quantified the overall percentage of new myelin sheaths in the networks. We simulated networks with up to a 45% of new myelin sheaths and we observed that performance was significantly impaired compared to the young, control networks with intact myelin: memory duration decreased to almost 1 s (Figure 8B) and diffusion constant increased (Figure 8C). This is consistent with empirical findings showing that cognitive impairment for a cohort of 18 young and aged rhesus monkeys increased with an increase of the percentage of paranodal profiles in the dlPFC (Peters and Sethares, 2003). Therefore, both demyelination and incomplete remyelination lead to impaired performance in our networks, compared to networks with intact myelin sheaths.
Discussion
This multi-level computational study explored how age-related myelin degradation (demyelination) and remyelination affect AP propagation in individual axons and working memory precision in spiking neural networks. We found that these myelin changes lead to AP failures which, if uncompensated by other factors, predict working memory decline with aging.
Myelin changes affect AP propagation in a cohort of model neurons
The novelty of our neuron model lies in its systematic exploration of a combination of different myelin perturbation types known to occur in myelin dystrophies, across a wide range of biologically feasible models. Our single neuron model assumed that age-related myelin dystrophies (e.g. Figure 1) alter the insulative properties of lamellae analogously to demyelination, and examined interactions between demyelination and remyelination. Past studies of myelin dystrophy examined how either demyelination or remyelination of all segments affected AP propagation for a few representative axon morphologies. For example, Scurfield and Latimer, 2018, explored how remyelination affected CV delays, finding that axons with more transitions between long and short myelinated segments had slower CV (Figure 4—figure supplement 2), and was first to explore how remyelination interacts with tight junctions. However, their study did not couple remyelination and demyelination together or examine AP failures. Other basic findings from our single neuron cohort are consistent with past modeling studies, including that demyelination caused CV slowing and eventual AP failures (Stephanova et al., 2005; Stephanova and Daskalova, 2008; Naud and Longtin, 2019), and, separately, that remyelination with shorter and thinner myelinated segments led to CV slowing (Lasiene et al., 2008; Powers et al., 2012; Scurfield and Latimer, 2018). However, by assuming that some previously demyelinated segments were remyelinated while others were not, we found that models could have much higher AP failure rates than previously reported. Such a scenario, in which individual axons have some segments that are normal, some demyelinated, and some remyelinated, is likely to occur. We also found a few neurons in our cohort showing a CV increase after remyelination, which has not generally been reported before and is likely due to an interplay between ion channels in the new nodes and altered electrotonic lengths in the perturbed myelinated segments (e.g. Moore et al., 1978; Naud and Longtin, 2019).
Since our single neuron cohort sampled a wide range of parameter space, we used Lasso regression to identify which of the complex, interacting parameters contributed most to CV delays (which preceded AP failures). Parameters including axon diameter, node length, length of myelinated segments, and nodal ion channel densities predicted how our models responded to demyelination and remyelination; these findings are consistent with past modeling studies over more limited parameter ranges (e.g. Goldman and Albus, 1968; Moore et al., 1978; Babbs and Shi, 2013; Young et al., 2013; Schmidt and Knösche, 2019). Better empirical measurements of these parameters in monkey dlPFC, e.g., from three-dimensional electron microscopy studies or single neuron axon studies combined with markers for myelin, would help predict the extent to which myelin dystrophy and remyelination along individual axons with aging affect AP propagation.
Another important feature of our multicompartment model is that it was constrained by morphological and physiological data in rhesus monkey dlPFC –an extremely valuable dataset from an animal model with many similarities to humans (Upright and Baxter, 2021; Tarantal et al., 2022). While beyond the scope of the current study, this computational infrastructure – with a detailed axon, initial segment, soma, and apical and basal dendrites – enables simultaneous investigations of signal propagation through the dendritic arbor and axon. Our model can also be extended to explore interactions between spatially localized myelin perturbations (such as those seen in multiple sclerosis) and axon collateralization (Sengupta et al., 2023), which would affect the distance dependence of AP failures. Integrating such results from single neuron models into network models of working memory, as we have done here, is a powerful way to connect empirical data across multiple scales.
Myelin changes impair working memory function
With the spiking neural network model, we found that increasing probabilities of AP failure, corresponding to higher degrees of demyelination, gradually impaired working memory precision and the time during which stimulus information can be held in memory. Complete remyelination of all previously demyelinated segments restored performance to the control network level. In contrast to the strong impact of AP failures on network function, introducing propagation delays to mimic AP slowing, in the time range of the delays quantified with the single neuron model, did not have an effect (Figure 5—figure supplement 2 and Figure 5—figure supplement 5). This is because the network operates in an asynchronous state in which the dynamics are primarily governed by the statistics of neuronal activity (e.g. firing rates), rather than precise spike timings (van Vreeswijk and Sompolinsky, 1996; Hansel and Mato, 2013). While highly irregular persistent activity is indeed observed in PFC during working memory tasks (Compte et al., 2003), at the mesoscopic level oscillations of the local field potentials and synchronization also play an important role (Gregoriou et al., 2009; Liebe et al., 2012; Buschman et al., 2012; Salazar et al., 2012). AP slowing may alter these neural oscillations and synchrony which could lead to further working memory impairment not captured by our network model. In addition to age-related changes in memory duration and precision, our network model predicts an age-related increase in systematic errors (bias) due to an increased drift of the activity bump (Figure 6—figure supplement 2). Moreover, if demyelination is spatially localized in a part of the network, the model predicts a repulsive bias away from the memories encoded in the affected zone (Figure 5—figure supplement 3). Delay-dependent systematic working memory errors have been observed in behavioral experiments (Panichello et al., 2019; Bae, 2021; Stein et al., 2021) and it would be interesting to test whether those biases also change with aging. In addition to the prefrontal cortex, working memory is likely sustained by interactions between several fronto-parietal brain areas (Leavitt et al., 2017; Christophel et al., 2017). Our future work will include development of large-scale models of working memory (Mejías and Wang, 2022), incorporating myelin alterations in local circuits and in inter-area connections.
For biologically realistic combinations of neurons with intact and demyelinated axons, our network model predicts that myelin dystrophies alone would lead to spatial working memory impairment with aging (Figure 7). This result supports the observation that cognitive impairment increases as the percentage of normal myelin sheaths decreases (Peters and Sethares, 2002). In addition, combining neurons containing either intact or incompletely remyelinated axons, we found that spatial working memory is still impaired in the context of incomplete remyelination (Figure 8). Our result explains two separate empirical findings. One study showed a positive correlation between cognitive impairment and a higher percentage of new, shorter and thinner, myelin sheaths (Peters and Sethares, 2003). In our single neuron model, insufficiently remyelinated axons with an array of long unperturbed, long thin, and new shorter and thinner segments had slower CV and more AP failures compared to unperturbed axons. This caused WM errors in the network model. A second study found that, with aging, oligodendroglia has a decreased capacity for effective maturation, remaining in a progenitor state without being able to produce new myelin. This low capacity for remyelination was correlated with spatial working memory impairment in the monkeys (Dimovasili et al., 2023). Our models showed that axons with demyelinated segments, or otherwise poorly insulated myelin sheaths, experience delayed and even failed AP propagation, which in turn led to working memory impairment. Therefore, it is reasonable to assume that ineffective remyelination may lead to working memory impairment. In fact, complete remyelination of all previously demyelinated segments with sufficient myelin, with fewer transitions between long and short segments, led to full recovery of working memory function. Our findings also suggest that differences in the degree of demyelination or remyelination (Figures 7 and 8) may account for the cognitive variability observed across individuals with aging, which encompass individuals with both good (successful agers) and impaired (unsuccessful agers) cognitive function (Lacreuse et al., 2005; Moore et al., 2006; Moss et al., 2007; Moore et al., 2017).
Conclusions
The multiscale modeling approach we employed here extends our prior framework for studying how changes in the aging monkey dlPFC might affect working memory (review: Luebke et al., 2010). Our previous work (Ibañez et al., 2019) modeled increased AP firing rates observed in vitro together with the loss of both excitatory and inhibitory synapses, and quantified how these alterations affected working memory performance. Aged networks which compensated the loss of excitatory and inhibitory synapses with higher firing rates in individual pyramidal neurons successfully retained memory of the DRT stimulus. In addition, networks in which we also decreased the overall excitatory drive to pyramidal neurons reproduced the lower firing rates during performance of DRT in aged monkeys, reported in Wang et al., 2011. Such decreased excitation could arise from AP failures induced by myelin dystrophy, as shown here. It has also been hypothesized that the hyperexcitability of dlPFC pyramidal neurons observed in vitro with aging could be a homeostatic mechanism to compensate for AP disruptions due to dystrophic myelin (Luebke et al., 2010). The multicompartment model used here is particularly well suited to simultaneous modeling of alterations in the soma, dendrites, synapses, and axon of individual pyramidal neurons. This can include pathological changes to the nodes of Ranvier (reviewed in Arancibia-Carcamo and Attwell, 2014Arancibia-Carcamo and Attwell, 2014) or to the metabolism of axons after demyelination (Gerevich et al., 2023), which were not modeled here. The effects of such changes can then be incorporated into our DRT model, or models of other working memory tasks.
Myelin dystrophy occurs in many neurological conditions, including multiple sclerosis, schizophrenia, bipolar disorder, autism spectrum disorders, and after traumatic brain injury (Franklin and Ffrench-Constant, 2008; Takahashi et al., 2011; Armstrong et al., 2016; Gouvêa-Junqueira et al., 2020; Galvez-Contreras et al., 2020; Simkins et al., 2021; Valdés-Tovar et al., 2022). As in normal aging working memory is often one of the most vulnerable cognitive functions in these conditions, especially schizophrenia or autism spectrum disorders (Wang et al., 2017; Hahn et al., 2018; Gold and Luck, 2023). Since we found that myelin changes alone can account for working memory impairment, our study points to myelin degradation as a key factor in working memory decline with normal aging and, perhaps, in neuropathological conditions.
Methods
The single neuron and network models for this study are available on ModelDB (https://modeldb.science), accession number 2014821.
Single neuron model
To simulate age-related myelin dystrophies in individual neurons, we used a biophysically detailed multicompartment model of a rhesus monkey dlPFC Layer 3 pyramidal neuron (Rumbell et al., 2016), executed in the NEURON simulation environment (Carnevale and Hines, 2006). The soma, apical, and basal dendritic arbors were constructed schematically, scaled to the overall surface area of a three-dimensional morphological reconstruction from empirical data (68 compartments total; details in Rumbell et al., 2016). Ion channel conductances and kinetics were fit to electrophysiological data obtained in vitro from the same neuron. In addition to passive membrane dynamics, the model included two sodium channels (fast inactivating, NaF; and non-inactivating persistent, NaP), three potassium channels (delayed rectifier, KDR; muscarinic receptor-suppressed, KM; transient inactivating A-type, KA), a high-threshold non-inactivating calcium channel (CaL), a calcium-dependent slow potassium channel (KAHP), and the hyperpolarization-activated anomalous rectifier channel (AR). Because our prime focus was quantifying alterations in AP propagation along the axon under dystrophic myelin conditions, we augmented the axon hillock and initial segment of the Rumbell et al., 2016, model (5 compartments each) by attaching nodes and myelinated segments from a detailed axon model (Gow and Devaux, 2008; Scurfield and Latimer, 2018). To the initial segment we attached 101 nodes (13 compartments each) alternating with 100 myelinated segments. Each myelinated segment was bound by a group of four paranodes (5 compartments each) on either side, with an internode flanked by two juxtaparanodes (9 and 5 compartments each, respectively) sandwiched between the paranode groups (Figure 2A). On both ends of a myelinated segment the extreme outward paranode interfaced with the adjoining node. Segments between successive nodes were endowed with myelin lamellae (wraps), and with tight junctions between the innermost lamella and the axolemma necessary for improving insulation and accurate modeling of AP propagation in nerve fibers with diameters less than 0.9 μm as often observed in PFC. Axon compartments included passive membrane dynamics as well as the NaF and KDR channels in the nodes. Simulations used a fixed 0.025 ms time step.
As done during in vitro electrophysiological experiments (Chang et al., 2005; Ibañez et al., 2019) and past modeling studies (Coskren et al., 2015; Rumbell et al., 2016), we first applied a holding current to stabilize the somatic membrane potential at –70 mV, then injected a current step into the somatic compartment for 2 s. We recorded the somatic firing rate, as well as the CV of APs propagating from the first node after the initial segment (proximal to the soma) and the penultimate node (at the distal end), and the percentage of APs that failed. The CV changes in response to myelin alterations were relatively insensitive to variations in the magnitude of suprathreshold somatic current steps (Figure 2—figure supplement 1C), and whether the current was constant or included Gaussian noise. Therefore, here we quantified CV changes and AP failures from responses to constant +380 pA current steps only.
Building a cohort of control model neurons
Recognizing that the effects of myelin dystrophy may depend on the physical dimensions of an axon, we constructed a cohort of neuron models that were consistent with empirical measurements of axonal morphology in young adult rhesus monkeys. To form this ‘control’ model cohort, we identified five parameters of axonal morphology which have been measured empirically in rhesus monkey cortex: axon diameter (Peters et al., 2001; Bowley et al., 2010; DL Rosene et al., unpublished observations), node length (Peters and Sethares, 2003), length of myelinated segments (Waxman, 1980), number of myelin lamellae (Peters and Sethares, 2003), and thickness of each lamella (Moore et al., 1978; Peters and Sethares, 2003). We also defined three scale factors defining the ratio of leak, NaF, and KDR conductances in our nodes relative to those of the Scurfield and Latimer, 2018, model, assuming an upper bound of 1 for each. This gave a total of eight parameters with associated feasible parameter ranges that we varied (Table 1); other parameters including tight junction resistivity, myelin resistivity, and myelin capacitance were held fixed at values from Scurfield and Latimer, 2018. We then used the space-filling Latin hypercube sampling design (Morris and Mitchell, 1995; Rumbell et al., 2016; Ibañez et al., 2019) to identify a set of 1600 points in parameter space that maximized the minimum distance between all pairs of points. We simulated each model specified by these points, and had several selection criteria for identifying biologically plausible models. First, we constrained somatic firing rates within ranges observed empirically in Layer 3 pyramidal neurons of rhesus monkey dlPFC: firing 13–16 Hz in response to the +380 pA current step and silent when no current was injected (Chang et al., 2005; Ibañez et al., 2019). We also required CVs of 0.3–0.8 m/s (rhesus monkey corpus callosum, DL Rosene et al., unpublished observations), and ensured that simulated APs were suprathreshold in the nodes and subthreshold in the juxtaparanode and internode regions, indicating saltatory conduction. Of the 1600 simulated models, 138 met these criteria; for the present study, we randomly selected 50 models to comprise the young, control model cohort. Along most dimensions, the chosen cohort was approximately normally distributed (Figure 2—figure supplement 1). The g-ratio (ratio of axon to fiber diameter) among models in the cohort was 0.71±0.02, with total axon lengths of 1.2±0.1 cm.
Simulating myelin alterations
We assumed that the effect of all myelin dystrophies observed empirically (e.g. Figure 1) could be modeled by removing lamellae from myelinated segments (demyelination, which reduces electrical insulation), and that remyelination could be modeled by replacing myelinated segments with two shorter and thinner segments with a node in between. Evidence suggests that aging affects oligodendrocytes in several ways, including the ability for oligodendrocyte precursor cells to mature (Dimovasili et al., 2023). Knowing that individual oligodendrocytes myelinate axons of many different neurons, but without data quantifying how oligodendrocyte dystrophy affects myelination in individual axons, we assumed that myelin alterations were randomly distributed. To simulate demyelination, we varied two independent factors: the percentage of myelinated segments selected for demyelination along an axon (demyelination percentage), and the percentage of myelin lamellae removed from those segments (lamellae removal percentage). For each demyelination percentage (10%, 25%, 50%, and 75%; Figure 3), we generated 30 randomized lists of segments to demyelinate. Then for each of the 30 trials, each lamellae removal percentage was applied (25%, 50%, 55%, 60%, 65%, 70%, 75%, and 100%) to the chosen segments, for all 50 models in the control cohort. We then simulated the +380 pA current step, calculating the CV and the number of APs that propagated to the distal end of the axon. For each perturbation, we defined the CV change as the percentage change in CV induced relative to the CV of the corresponding unperturbed model. We also computed, for each perturbation, the percentage of AP failures at the distal axon end, relative to the number of APs observed at the first node.
To simulate remyelination three factors were varied: the percentage of myelinated segments initially demyelinated (demyelination percentage: 25%, 50%, 75%); the percentage of those affected segments which were then remyelinated (remyelination percentage: 25%, 50%, 75%, 100%); and the percentage of lamellae restored with remyelination (lamellae restoration percentage: 10%, 25%, 50%, 75%). We performed these remyelination simulations under two demyelination conditions: where the initially demyelinated segments had lost all their lamellae (‘complete demyelination’), or had lost half their lamellae (‘partial demyelination’). Sample remyelination perturbations shown in Figure 4. Remyelination was performed by replacing an affected (previously demyelinated) segment with two shorter segments, each including paranodes, juxtaparanodes, and an internode, and a new node between them that was identical to existing nodes. The number of lamellae on the shorter segments was determined by the lamellae restoration percentage. As before, for each demyelination percentage (25%, 50%, and 75%), we generated 30 randomized lists of segments to demyelinate. Segments to remyelinate were decided based on remyelination percentage. For example, in the case of 50% remyelination, every alternate (previously) demyelinated segment was remyelinated. Then for each trial, we applied each combination of remyelination percentage and lamellae restoration percentage to the chosen segments, for all models in the cohort, and simulated the current clamp protocol. For each remyelination condition, in addition to computing the percentage of AP failures, we defined CV recovery as the percentage improvement in CV relative to the CV change for the completely demyelinated case. For example, if a control model had a CV of 1 m/s, and a CV of 0.6 m/s after complete demyelination, then the CV change was –0.4/1.0 = –40%. If a subsequent remyelination condition led to a CV of 0.8 m/s (an increase of 0.2 m/s over the demyelinated case), the CV recovery was 0.2/0.4=50%.
Statistical assessment of parameter importance
To identify which parameters of the multicompartment model had the greatest influence on axonal responses to demyelination and remyelination perturbations, we used least absolute shrinkage and selection operator (Lasso) regression (Tibshirani, 1996; James et al., 2021) implemented in MATLAB R2022a (Mathworks, Natick, MA, USA). Lasso fits a regression model to response variable, given predictor variables (observations of predictors) by selecting coefficients that minimize the quantity∑i=1n(yi−β0−∑j=1pβjxij)2+λ∑j=1p|βj|,
for a given value of James et al., 2021. The former quantity is the residual sum of squares of the model. The latter quantity is the absolute sum of coefficients; including it in the minimization shrinks some coefficients to zero when λ is sufficiently large, enabling Lasso to perform feature selection. The tuning parameter was selected to minimize the 10-fold cross-validation error; then for the selected value of λ, the regression was repeated using all available data. A total of 12 parameters served as predictor variables for the Lasso regression. We started with 6 of the 8 parameters used to construct the cohort (axon diameter, node length, myelin length, and scale factors for leak, NaF, and KDR conductances), and combined two other parameters (number of myelin lamellae and lamella thickness) into one quantity: myelin thickness (their product). Past studies (Goldman and Albus, 1968; Koles and Rasminsky, 1972; Moore et al., 1978; Gow and Devaux, 2008) identified several other parameters that affect axonal propagation, including the g-ratio, axoplasmic resistance, axon capacitance, myelin resistance, myelin capacitance, and tight junction resistance. We added five of these parameters to the seven cohort parameters as predictor variables, omitting only g-ratio due to its high correlation with axon diameter and myelin thickness. We created response variables summarizing the effects of demyelination and remyelination on CV in each of the 50 members of the model cohort. AP failures were recorded as CV change of –100%. We did not include AP failure rates as a response variable, since CV changes precede AP failures. For demyelination, we averaged the CV change for all randomized trials in which 100% of lamellae were removed from 25%, 50%, and 75% of segments. For remyelination, we averaged the CV recovery in all randomized trials in which 25%, 50%, or 75% of segments were completely demyelinated, and then all affected segments were remyelinated as two shorter segments with 75% of lamellae added back. To facilitate comparison, we z-scored all predictor and response variables before performing Lasso separately on each response. We tested the predictive ability of the Lasso models by randomly selecting another 50 of the 138 models from the original hypercube that met the inclusion criteria, then simulating the specific myelin alteration protocols that comprised the demyelination and remyelination response variables. We computed Pearson’s correlation coefficient between the z-scored predicted vs. observed responses.
Spiking neural network model
We adapted a neural network model (Hansel and Mato, 2013) to simulate the neural circuit in the dlPFC underlying spatial working memory during the oculomotor DRT (Figure 5A). The network model is composed of N=20,000 leaky integrate-and-fire neurons, NE=16,000 excitatory (E) neurons (80%) and NI=4000 inhibitory (I) neurons (20%) with sparse probabilistic connections among all neuronal populations. A full description of this probabilistic version of the ring model can be found in Hansel and Mato, 2013. Here, we describe the essential features of the network and summarize all the parameter values in Table 2. The range of the interactions is represented by the parameters σEE , σEI , σIE , and σII . Each neuron receives K=500 total inputs, where KE=0.8K are excitatory inputs and KI=0.2K are inhibitory inputs. The subthreshold membrane potential of each excitatory and inhibitory neuron (i) in the network is described byτEdVE,idt=−VE,i+IE,irec(t)+IEext+IE,icue(t),τIdVI,idt=−VI,i+II,irec(t)+IIext,
where τE and τI are the membrane time constants for the E and I neurons, respectively. IE/I,irec(t) is the total recurrent synaptic current that each neuron receives from all the other neurons in the network connected to it. IE/Iext is a constant background input, representing internal brain currents that come from outside the network. IE,icue(t) represents the transient sensory input to each E neuron, associated to a given direction of the stimulus, and only active during the cue period of the task. An AP is fired each time that the membrane potential of a neuron reaches the threshold value, VT . The voltage of the membrane is reset to the baseline value, VR , immediately after.
Network model parameters.
Parameter
Value
gEea
533.3/KEmV⋅ms
gEen
490.64/KEmV⋅ms
gEia
67.2/KEmV⋅ms
gEin
7.4/KEmV⋅ms
gIE
−138.6/KImV⋅ms
gII
−90.6/KImV⋅ms
τE
20ms
τI
10ms
τa
3ms
τn
50ms
τg
4ms
τd
200ms
τf
450ms
U
0.03
VT
20mV
VR
−3.33mV
σEE
30°
σEI
35°
σIE
30°
σII
30°
IEext
1.66KEmV
IIext
1.5355KEmV
Imax,E
0.24mV
εE
61.2°
The total recurrent synaptic current for each neuron, IE/I,irec(t), is given byIE/I,irec(t)=IE/I,ia(t)+IE/I,in(t)+IE/I,ig(t),
where subindices a and n represent the excitatory AMPA and NMDA glutamatergic receptors, and g an inhibitory GABA receptor. Each synaptic current is given as in Hansel and Mato, 2013, with synaptic decay time constants for each receptor, τa,τn , and τg , respectively. Short-term plasticity was incorporated to the excitatory-to-excitatory synaptic connections through the variables u and x, given by Markram et al., 1998τfduidt=U−uIτddxidt=(1−xI).
x and u are updated as x→x(1−u) and u→u+U∗(1−u) , each time that there is a presynaptic spike. The variable x represents the amount of available neurotransmitter resources in the presynaptic terminal and u is the utilization parameter, indicating the residual calcium level (Bertram et al., 1996; Zucker and Regehr, 2002; Mongillo et al., 2008). With each spike, amount ux of the available resources is used to produce the postsynaptic current. Thus, x is reduced, representing neurotransmitter depletion, and u is increased, representing the calcium influx into the presynaptic terminal and its effect on release probability. Between spikes, x and u recover to their baseline levels (x=1 and u=U; 0<x<1) with time constants τd and τf . We set τf>τd so that they facilitate signal transmission (Tsodyks et al., 1998).
The sensory input current to excitatory neuron i during the cue period, and for a specific location of the stimulus, IE,icue(t), is given byIE,icue(t)=Imax,Ee−12(θi−θcueεE)2,
where θi is the preferred direction of the neuron i, given by its position on the ring, θcue is the cue direction, and Imax,E and εE are the amplitude and the width of the sensory input current, respectively.
Network simulations with spatially modulated correlations
To introduce spatially modulated correlations in the model (Figure 5—figure supplement 1), we reduced the strength of the constant background inputs to excitatory and inhibitory neurons, IEext and IIext , by a factor of 0.5 and provide additional external input from a population of Poisson neurons. The parameters were chosen such that the time-averaged total input (the sum of the reduced constant input and the inputs from the Poisson population) is the same as in our default network without Poisson inputs. The external population is composed of Next = 16,000 Poisson neurons that fire with a constant firing rate of rext = 9.72 Hz. These neurons are connected through AMPA synapses to excitatory neurons in the network with a strength of gEaext=0.5gEia and to inhibitory neurons with a strength of gIaext=0.4625gEia . The average total number of synaptic inputs from Poisson neurons that a neuron receives is Kext = 1000. Crucially, the connections from the Poisson population to neurons in the network are spatially structured, with a Gaussian connection profile similar to the recurrent connections in the network (ring structure) and with the interaction range determined by σext = 20°. Thus, neurons in the network receive shared inputs with a spatial structure and this leads to spatial correlations in the network neurons (Rosenbaum et al., 2017).
Simulating a cohort of young (control) networks
Using the Brian2 simulator (based on Python), we simulated 10 different networks, all with the same parameters, but each with a different connectivity profile. For each network, we ran 280 trials, with different initial conditions and cue positions. All networks performed the DRT – with 2 s fixation period, 1 s cue period, and 4 s delay period – and maintained the memory of the stimulus during the whole delay period of the task (Figure 5B, i and Figure 5—figure supplement 1A, left panel).
Modeling the effects of demyelination and remyelination in the network modelModeling AP propagation failures in the network
The network model is composed of point neurons without an explicit model of the axon. To effectively model the AP failures at the distal end of the axons quantified with the single neuron model under the different demyelination and remyelination conditions, the AP failure rate was adjusted to the values produced by the single neuron model. To do this, we perturbed the 10 control networks by designing a probabilistic model of spike transmission from the excitatory presynaptic neurons to both the excitatory and inhibitory postsynaptic neurons. From the single neuron model, for each demyelination/remyelination condition, we quantified the probability of AP failure for each of the neurons in the control cohort, as well as the percentage of those neurons that shared the same probabilities of failure. That is, the percentage of neurons that had probability of failure = 0, probability of failure = 1 or any other probability. Then, we computed the probability of transmission, ptransmission=1−pfailure , and we specified ptransmission for the corresponding percentages of excitatory neurons in the networks. Thus, in the network model, we took into account the heterogeneity observed in the single neuron model under each demyelination/remyelination condition.
Modeling conduction velocity slowing in the network
To explore the effect of CV slowing along the axons of model neurons, we simulated 20 young, control networks and 20 perturbed networks with AP failure rates adjusted for the case of single model neurons with 50% of the segments demyelinated along the axons by removing 60% of the myelin lamellae (we ran 280 trials for each network). Then, we added random delays uniformly distributed with a minimum value of 0 ms in both cases, a maximum value of 100 ms in the control networks, and a maximum values of 40 and 85 ms in the perturbed networks, in both the AMPA and NMDA excitatory connections to both E and I neurons (Figure 5—figure supplement 2). These large values were chosen because we wanted to illustrate the potential effect of CV slowing in our network and smaller, more realistic, values did not have any effect.
Quantification of normal myelin sheaths in groups of model neurons containing both intact and demyelinated axons
We created different groups of 50 total model neurons containing, among those 50, different random amounts of neurons with intact axons and with demyelinated axons, with either 10%, 25%, 50%, or 75% of segments demyelinated. In each group, the distribution of the demyelinated segments along the altered axons was also randomly chosen among the 30 possible distributions simulated with the single neuron model. We sorted the axons in each group by locating their origin aligned in the same position and, up to the maximum length of the shortest axon, we divided the axons longitudinally in sections of 0.5 μm, and calculated the percentage of normal, unperturbed myelin sheaths in each section (Figure 7A). Then, we averaged across all sections and we picked 60 groups that had over 80% of normal myelin sheaths. According to the proportion of intact and demyelinated axons (with either 10%, 25%, 50%, or 75% of segments affected) in each group, we adjusted the distribution of AP transmission probability (ptransmission ;ptransmission=1 for intact axons) in 1 of the 10 control networks. To do this we considered that the perturbed axons in 40 of the groups were completely demyelinated (all lamellae removed) and that in the remaining 20 groups, they had 75% of lamellae removed.
Quantification of new myelin sheaths in groups of model neurons containing both intact and remyelinated axons
We again created different groups of 50 total model neurons containing different random amounts of neurons with intact and remyelinated axons. The remyelinated axons had either 25%, 50%, or 75% of segments remyelinated (with two shorter and thinner myelin sheaths), following previous demyelination of either 25%, 50%, or 75% of the segments. In each group, the distribution of the remyelinated segments along the altered axons was randomly chosen among the 30 possible distributions. As before, we sorted the axons in each group by locating their origin aligned in the same position, we divided the axons longitudinally in sections of 0.5 μm (up to the maximum length of the shortest axon), and calculated the percentage of remyelinated segments in each section (Figure 8A). Then, we averaged across all sections, we multiplied by two to calculate the number of new myelin sheaths, and we picked 22 groups that had below 45% of new myelin sheaths. According to the proportion of intact and remyelinated axons (with different percentages of remyelinated segments) in each group, we adjusted the distribution of ptransmission in the same control network as before for the case of remyelination following partial demyelination by adding 25% of the lamellae back.
Measures of working memory performance
The remembered location was obtained using a population vector decoder and network performance was quantified by the following four measures that describe the quality of the activity bump maintaining a memory of the stimulus during the delay period: the memory strength, the memory duration, the drift rate, and the diffusion constant.
Memory strength
The memory strength was defined as the modulus M(t) of the population vector, Z(t), which characterizes the spatial modulation of the excitatory neuronal activity at time t, given byZ(t)=∑jrj(t)eiθj∑jrj(t)=M(t)eiΨ(t).
rj(t) is the firing rate of neuron j with preferred direction θj . Firing rates rj(t) were estimated as spike counts in a 250 ms window. A memory strength M(t) close to 0 indicates homogeneous activity of the network and M(t) close to 1 indicates a sharply modulated activity (Figure 5C).
Decoded cue location
Ψ(t) is the argument of Z(t) and indicates the remembered stimulus location. That is, the center of the activity bump (Figure 5B).
Memory duration
The memory duration was defined as the time from the delay onset until the time where the memory strength decayed below a fixed threshold that we set at 0.4. That is, the memory duration is the interval during which the bump of neural activity is reliably maintained (Figure 5C).
Memory drift
The drift was defined as the bias of the estimator, which is given bybest(t)=⟨Φ(t)⟩−θcue.
That is, the difference of the trial average of the estimates Φ(t) and the true value θcue (Esnaola-Acebes et al., 2022; Figure 5E). The drift rate is the slope of a linear fit of the drift during the memory duration time (Figure 6—figure supplement 2).
Memory diffusion
The diffusion was defined as the variance of the estimator, given byσest2(t)=⟨(Φ(t)−⟨Φ(t)⟩)2⟩,
the variation of the estimates about their mean value (Esnaola-Acebes et al., 2022; Figure 5D). The diffusion rate is the slope of a linear fit of the diffusion during the memory duration time (Figure 6B).
Additional informationCompeting interests
No competing interests declared
Author contributions
Conceptualization, Resources, Data curation, Software, Formal analysis, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing
Conceptualization, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing
Conceptualization, Resources, Data curation, Supervision, Funding acquisition, Investigation, Writing – original draft, Project administration, Writing – review and editing
Conceptualization, Resources, Data curation, Software, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing
Conceptualization, Resources, Data curation, Software, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing
Additional filesData availability
The single neuron and network models for this study are available on ModelDB (https://modeldb.science), accession number 2014821.
The following dataset was generated:
IbañezSSenguptaNLuebkeJIWimmerKWeaverCM2024Myelin dystrophy impairs signal transmission and working memory decline in a multiscale model of the aging prefrontal cortexModelDB2014821
Acknowledgements
We acknowledge the use of Fenix Infrastructure resources, which are partially funded from the European Union’s Horizon 2020 research and innovation program through the ICEI project under the grant agreement No. 800858, and the research cluster at Franklin and Marshall College, funded through NSF grant 1925192. We thank David Latimer and Albert Compte for sharing computer code, as well as Jason Brooks and Tony Weaver for technical assistance. We thank the CERCA Programme/Generalitat de Catalunya for institutional support. This work was supported by NIH/NIA grant R01 AG059028, NIH 1R01AG071230-01, and grant PCI2020-112035 from MCIN/AEI/10.13039/501100011033 and the European Union 'NextGenerationEU'/PRTR. This work was supported by the Spanish State Research Agency, through the Severo Ochoa and María de Maeztu Program for Centers and Units of Excellence in R&D (CEX2020-001084-M). This work utilized the research cluster at Franklin and Marshall College which was funded through NSF grant 1925192.
ReferencesAlbertM1993Neuropsychological and neurophysiological changes in healthy adult humans across the age rangeNeurobiology of Aging1462362510.1016/0197-4580(93)90049-h8295666Arancibia-CarcamoILAttwellD2014The node of Ranvier in CNS pathologyActa Neuropathologica12816117510.1007/s00401-014-1305-z24913350ArmstrongRCMierzwaAJMarionCMSullivanGM2016White matter involvement after TBI: Clues to axon and myelin repair capacityExperimental Neurology275 Pt 332833310.1016/j.expneurol.2015.02.01125697845BabbsCFShiR2013Subtle paranodal injury slows impulse conduction in a mathematical model of myelinated axonsPLOS ONE8e6776710.1371/journal.pone.006776723844090BaeGY2021Neural evidence for categorical biases in working memory for location and orientationbioRxiv10.1101/2020.12.15.422978BarbosaJSteinHMartinezRLGalan-GadeaALiSDalmauJAdamKCSValls-SoléJConstantinidisCCompteA2020Interplay between persistent activity and activity-silent dynamics in the prefrontal cortex underlies serial biases in working memoryNature Neuroscience231016102410.1038/s41593-020-0644-432572236BertramRShermanAStanleyEF1996Single-domain/bound calcium hypothesis of transmitter release and facilitationJournal of Neurophysiology751919193110.1152/jn.1996.75.5.19198734591BlightAR1985Computer simulation of action potentials and afterpotentials in mammalian myelinated axons: the case for a lower resistance myelin sheathNeuroscience15133110.1016/0306-4522(85)90119-82409473BowleyMPCabralHRoseneDLPetersA2010Age changes in myelinated nerve fibers of the cingulate bundle and corpus callosum in the rhesus monkeyThe Journal of Comparative Neurology5183046306410.1002/cne.2237920533359BrillMHWaxmanSGMooreJWJoynerRW1977Conduction velocity and spike configuration in myelinated fibres: computed dependence on internode distanceJournal of Neurology, Neurosurgery, and Psychiatry4076977410.1136/jnnp.40.8.769925697BuschmanTJDenovellisELDiogoCBullockDMillerEK2012Synchronous oscillatory neural ensembles for rules in the prefrontal cortexNeuron7683884610.1016/j.neuron.2012.09.02923177967CarnevaleNTHinesML2006The NEURON BookCambridge University Press10.1017/CBO9780511541612ChangYMRoseneDLKillianyRJMangiameleLALuebkeJI2005Increased action potential firing rates of layer 2/3 pyramidal cells in the prefrontal cortex are significantly related to cognitive performance in aged monkeysCerebral Cortex1540941810.1093/cercor/bhh14415749985ChangWWeaverCMMedallaMMooreTLLuebkeJI2022Age-related alterations to working memory and to pyramidal neurons in the prefrontal cortex of rhesus monkeys begin in early middle-age and are partially ameliorated by dietary curcuminNeurobiology of Aging10911312410.1016/j.neurobiolaging.2021.09.01234715442ChomiakTHuB2009What is the optimal value of the g-ratio for myelinated fibers in the rat CNS? A theoretical approachPLOS ONE4e775410.1371/journal.pone.000775419915661ChristophelTBKlinkPCSpitzerBRoelfsemaPRHaynesJD2017The distributed nature of working memoryTrends in Cognitive Sciences2111112410.1016/j.tics.2016.12.00728063661CleelandCPipingasAScholeyAWhiteD2019Neurochemical changes in the aging brain: A systematic reviewNeuroscience and Biobehavioral Reviews9830631910.1016/j.neubiorev.2019.01.00330625337CogganJSBittnerSStiefelKMMeuthSGPrescottSA2015Physiological dynamics in demyelinating diseases: unraveling complex relationships through computer modelingInternational Journal of Molecular Sciences16212152123610.3390/ijms16092121526370960CompteABrunelNGoldman-RakicPSWangXJ2000Synaptic mechanisms and network dynamics underlying spatial working memory in a cortical network modelCerebral Cortex1091092310.1093/cercor/10.9.91010982751CompteAConstantinidisCTegnerJRaghavachariSChafeeMVGoldman-RakicPSWangXJ2003Temporally irregular mnemonic persistent activity in prefrontal neurons of monkeys during a delayed response taskJournal of Neurophysiology903441345410.1152/jn.00949.200212773500ComrieAEGrayDTSmithACBarnesCA2018Different macaque models of cognitive aging exhibit task-dependent behavioral disparitiesBehavioural Brain Research34411011910.1016/j.bbr.2018.02.00829432794CoskrenPJLuebkeJIKabasoDWearneSLYadavARumbellTHofPRWeaverCM2015Functional consequences of age-related morphologic changes to pyramidal neurons of the rhesus monkey prefrontal cortexJournal of Computational Neuroscience3826328310.1007/s10827-014-0541-525527184DekkerDMTBriaireJJFrijnsJHM2014The impact of internodal segmentation in biophysical nerve fiber modelsJournal of Computational Neuroscience3730731510.1007/s10827-014-0503-y24827400DimovasiliCFairAEGarzaIRBattermanKVMortazaviFMooreTLRoseneDL2023Aging compromises oligodendrocyte precursor cell maturation and efficient remyelination in the monkey brainGeroScience4524926410.1007/s11357-022-00621-435930094Esnaola-AcebesJMRoxinAWimmerK2022Flexible integration of continuous sensory evidence in perceptual estimation tasksPNAS119e221444111910.1073/pnas.221444111936322720FiskJESharpCA2004Age-related impairment in executive functioning: updating, inhibition, shifting, and accessJournal of Clinical and Experimental Neuropsychology2687489010.1080/1380339049051068015742539FranklinRJMFfrench-ConstantC2008Remyelination in the CNS: from biology to therapyNature Reviews. Neuroscience983985510.1038/nrn248018931697FunahashiSBruceCJGoldman-RakicPS1989Mnemonic coding of visual space in the monkey’s dorsolateral prefrontal cortexJournal of Neurophysiology6133134910.1152/jn.1989.61.2.3312918358Galvez-ContrerasAYZarate-LopezDTorres-ChavezALGonzalez-PerezO2020Role of oligodendrocytes and myelin in the pathophysiology of autism spectrum disorderBrain Sciences1095110.3390/brainsci1012095133302549GerevichZKovácsRLiottaAHasam-HendersonLAWehLWallachIBerndtN2023Metabolic implications of axonal demyelination and its consequences for synchronized network activity: An in silico and in vitro studyJournal of Cerebral Blood Flow and Metabolism431571158710.1177/0271678X23117074637125487GoldJMLuckSJ2023Working memory in people with schizophreniaCurrent Topics in Behavioral Neurosciences6313715210.1007/7854_2022_38135915383GoldmanLAlbusJS1968Computation of impulse conduction in myelinated fibers; theoretical basis of the velocity-diameter relationBiophysical Journal859660710.1016/S0006-3495(68)86510-55699798Gouvêa-JunqueiraDFalvellaACBAntunesASeabraGBrandão-TelesCMartins-de-SouzaDCrunfliF2020Novel treatment strategies targeting myelin and oligodendrocyte dysfunction in schizophreniaFrontiers in Psychiatry1137910.3389/fpsyt.2020.0037932425837GowADevauxJ2008A model of tight junction function in cns myelinated axonsNeuron Glia Biology430731710.1017/S1740925X09990391GregoriouGGGottsSJZhouHDesimoneR2009High-frequency, long-range coupling between prefrontal and visual cortex during attentionScience3241207121010.1126/science.117140219478185HahnBRobinsonBMLeonardCJLuckSJGoldJM2018Posterior parietal cortex dysfunction is central to working memory storage and broad cognitive deficits in schizophreniaThe Journal of Neuroscience388378838710.1523/JNEUROSCI.0913-18.201830104335HanselDMatoG2013Short-term plasticity explains irregular persistent activity in working memory tasksThe Journal of Neuroscience3313314910.1523/JNEUROSCI.3455-12.201323283328HempelCMHartmanKHWangXJTurrigianoGGNelsonSB2000Multiple forms of short-term plasticity at excitatory synapses in rat medial prefrontal cortexJournal of Neurophysiology833031304110.1152/jn.2000.83.5.303110805698HofPRMorrisonJH2004The aging brain: morphomolecular senescence of cortical circuitsTrends in Neurosciences2760761310.1016/j.tins.2004.07.01315374672IbañezSLuebkeJIChangWDraguljićDWeaverCM2019Network models predict that pyramidal neuron hyperexcitability and synapse loss in the dlpfc lead to age-related spatial working memory impairment in rhesus monkeysFrontiers in Computational Neuroscience138910.3389/fncom.2019.0008932009920JamesGWittenDHastieTTibshiraniR2021An Introduction to Statistical Learning: With Applications in RSpringer10.1007/978-1-0716-1418-1KarimianMDibenedettoDMoerelMBurwickTWestraRLDe WeerdPSendenM2019Effects of synaptic and myelin plasticity on learning in a network of Kuramoto phase oscillatorsChaos2908312210.1063/1.509278631472483KilpatrickZPErmentroutB2013Wandering bumps in stochastic neural fieldsSIAM Journal on Applied Dynamical Systems12619410.1137/120877106KolesZJRasminskyM1972A computer simulation of conduction in demyelinated nerve fibresThe Journal of Physiology22735136410.1113/jphysiol.1972.sp0100364675037KonarASinghPThakurMK2016Age-associated cognitive decline: insights into molecular switches and recovery avenuesAging and Disease712112910.14336/AD.2015.100427114845LacreuseAKimCBRoseneDLKillianyRJMossMBMooreTLChennareddiLHerndonJG2005Sex, age, and training modulate spatial memory in the rhesus monkey (Macaca mulatta)Behavioral Neuroscience11911812610.1037/0735-7044.119.1.11815727518LasieneJShupeLPerlmutterSHornerP2008No evidence for chronic demyelination in spared axons after spinal cord injury in a mouseThe Journal of Neuroscience283887389610.1523/JNEUROSCI.4756-07.200818400887LeavittMLMendoza-HallidayDMartinez-TrujilloJC2017Sustained activity encoding working memories: not fully distributedTrends in Neurosciences4032834610.1016/j.tins.2017.04.00428515011LiebeSHoerzerGMLogothetisNKRainerG2012Theta coupling between V4 and prefrontal cortex predicts visual short-term memory performanceNature Neuroscience1545646210.1038/nn.303822286175LuebkeJBarbasHPetersA2010Effects of normal aging on prefrontal area 46 in the rhesus monkeyBrain Research Reviews6221223210.1016/j.brainresrev.2009.12.00220005254MarkramHWangYTsodyksM1998Differential signaling via the same axon of neocortical pyramidal neuronsPNAS955323532810.1073/pnas.95.9.53239560274McIntyreCCRichardsonAGGrillWM2002Modeling the excitability of mammalian nerve fibers: influence of afterpotentials on the recovery cycleJournal of Neurophysiology87995100610.1152/jn.00353.200111826063MejíasJFWangXJ2022Mechanisms of distributed working memory in a large-scale network of macaque neocortexeLife11e7213610.7554/eLife.7213635200137MongilloGBarakOTsodyksM2008Synaptic theory of working memoryScience3191543154610.1126/science.115076918339943MooreJWJoynerRWBrillMHWaxmanSDNajar-JoaM1978Simulations of conduction in uniform myelinated fibers: relative sensitivity to changes in nodal and internodal parametersBiophysical Journal2114716010.1016/S0006-3495(78)85515-5623863MooreTLKillianyRJHerndonJGRoseneDLMossMB2006Executive system dysfunction occurs as early as middle-age in the rhesus monkeyNeurobiology of Aging271484149310.1016/j.neurobiolaging.2005.08.00416183172MooreTLBowleyBShultzPCalderazzoSShobinEKillianyRJRoseneDLMossMB2017Chronic curcumin treatment improves spatial working memory but not recognition memory in middle-aged rhesus monkeysGeroScience3957158410.1007/s11357-017-9998-229047012MooreTLMedallaMIbañezSWimmerKMojicaCAKillianyRJMossMBLuebkeJIRoseneDL2023Neuronal properties of pyramidal cells in lateral prefrontal cortex of the aging rhesus monkey brain are associated with performance deficits on spatial working memory but not executive functionGeroScience451317134210.1007/s11357-023-00798-237106282MorrisMDMitchellTJ1995Exploratory designs for computational experimentsJournal of Statistical Planning and Inference4338140210.1016/0378-3758(94)00035-TMorrisonJHBaxterMG2012The ageing cortical synapse: hallmarks and implications for cognitive declineNature Reviews. Neuroscience1324025010.1038/nrn320022395804MossMBMooreTLSchettlerSPKillianyRRoseneD2007Successful vs. unsuccessful aging in the rhesus monkeyRiddleDRBrain AgingCRC Press10.1201/9781420005523.ch2MotleySEGrossmanYSJanssenWGMBaxterMGRappPRDumitriuDMorrisonJH2018Selective loss of thin spines in area 7a of the primate intraparietal sulcus predicts age-related working memory impairmentThe Journal of Neuroscience38104671047810.1523/JNEUROSCI.1234-18.201830355632NaudRLongtinA2019Linking demyelination to compound action potential dispersion with a spike-diffuse-spike approachJournal of Mathematical Neuroscience9310.1186/s13408-019-0071-631147800NooriRParkDGriffithsJDBellsSFranklandPWMabbottDLefebvreJ2020Activity-dependent myelination: A glial mechanism of oscillatory self-organization in large-scale brain networksPNAS117132271323710.1073/pnas.191664611732482855PajevicSBasserPJFieldsRD2014Role of myelin plasticity in oscillations and synchrony of neuronal activityNeuroscience27613514710.1016/j.neuroscience.2013.11.00724291730PajevicSPlenzDBasserPJFieldsRD2023Oligodendrocyte-mediated myelin plasticity and its role in neural synchronizationeLife12e8198210.7554/eLife.8198236976252PanichelloMFDePasqualeBPillowJWBuschmanTJ2019Error-correcting dynamics in visual working memoryNature Communications10336610.1038/s41467-019-11298-331358740PetersASetharesCKillianyRJ2001Effects of age on the thickness of myelin sheaths in monkey primary visual cortexThe Journal of Comparative Neurology43524124810.1002/cne.120511391644PetersASetharesC2002Aging and the myelinated fibers in prefrontal cortex and corpus callosum of the monkeyThe Journal of Comparative Neurology44227729110.1002/cne.1009911774342PetersASetharesC2003Is there remyelination during aging of the primate central nervous system?The Journal of Comparative Neurology46023825410.1002/cne.1063912687688PetersA2007Golgi, Cajal, and the fine structure of the nervous systemBrain Research Reviews5525626310.1016/j.brainresrev.2006.12.00217270274PetersASetharesCLuebkeJI2008Synapses are lost during aging in the primate prefrontal cortexNeuroscience15297098110.1016/j.neuroscience.2007.07.01418329176PetersA2009The effects of normal aging on myelinated nerve fibers in monkey central nervous systemFrontiers in Neuroanatomy31110.3389/neuro.05.011.200919636385PetersAKemperT2012A review of the structural alterations in the cerebral hemispheres of the aging rhesus monkeyNeurobiology of Aging332357237210.1016/j.neurobiolaging.2011.11.01522192242PowersBELasieneJPlemelJRShupeLPerlmutterSITetzlaffWHornerPJ2012Axonal thinning and extensive remyelination without chronic demyelination in spinal injured ratsThe Journal of Neuroscience325120512510.1523/JNEUROSCI.0002-12.201222496557RhodesMG2004Age-related differences in performance on the Wisconsin card sorting test: a meta-analytic reviewPsychology and Aging1948249410.1037/0882-7974.19.3.48215382998RichardsonAGMcIntyreCCGrillWM2000Modelling the effects of electric fields on nerve fibres: influence of the myelin sheathMedical & Biological Engineering & Computing3843844610.1007/BF0234501410984943RosenbaumRSmithMAKohnARubinJEDoironB2017The spatial structure of correlated neuronal variabilityNature Neuroscience2010711410.1038/nn.443327798630RumbellTHDraguljićDYadavAHofPRLuebkeJIWeaverCM2016Automated evolutionary optimization of ion channel conductances and kinetics in models of young and aged rhesus monkey pyramidal neuronsJournal of Computational Neuroscience41659010.1007/s10827-016-0605-927106692RushtonWAH1951A theory of the effects of fibre size in medullated nerveThe Journal of Physiology11510112210.1113/jphysiol.1951.sp00465514889433SalazarRFDotsonNMBresslerSLGrayCM2012Content-specific fronto-parietal synchronization during visual working memoryScience3381097110010.1126/science.1224000SalthouseTAAtkinsonTMBerishDE2003Executive functioning as a potential mediator of age-related cognitive decline in normal adultsJournal of Experimental Psychology13256659410.1037/0096-3445.132.4.566SchmidtHKnöscheTR2019Action potential propagation and synchronisation in myelinated axonsPLOS Computational Biology15e100700410.1371/journal.pcbi.100700431622338ScurfieldALatimerDC2018A computational study of the impact of inhomogeneous internodal lengths on conduction velocity in myelinated neuronsPLOS ONE13e019110610.1371/journal.pone.019110629329312SenguptaNMedallaMLuebkeJIWeaverCM2023Age-Related Myelin Dystrophies Impair Action Potential Propagation in Collateralized dlPFC Pyramidal Cell AxonsProgram No. PSTR368.04.2023 Neuroscience Meeting PlannerShamyJLHabeckCHofPRAmaralDGFongSGBuonocoreMHSternYBarnesCARappPR2011Volumetric correlates of spatiotemporal working and recognition memory impairment in aged rhesus monkeysCerebral Cortex211559157310.1093/cercor/bhq21021127015ShobinEBowleyMPEstradaLIHeyworthNCOrczykowskiMEEldridgeSACalderazzoSMMortazaviFMooreTLRoseneDL2017Microglia activation and phagocytosis: relationship with aging and cognitive impairment in the rhesus monkeyGeroScience3919922010.1007/s11357-017-9965-y28238188SimkinsTJDuncanGJBourdetteD2021Chronic demyelination and axonal degeneration in multiple sclerosis: pathogenesis and therapeutic implicationsCurrent Neurology and Neuroscience Reports212610.1007/s11910-021-01110-533835275SorelOPennequinV2008Aging of the planning process: the role of executive functioningBrain and Cognition6619620110.1016/j.bandc.2007.07.00617884265SteinHBarbosaJCompteA2021Towards biologically constrained attractor models of schizophreniaCurrent Opinion in Neurobiology7017118110.1016/j.conb.2021.10.01334839146StephanovaDIDaskalovaMAlexandrovAS2005Differences in potentials and excitability properties in simulated cases of demyelinating neuropathies: part IClinical Neurophysiology1161153115810.1016/j.clinph.2004.12.01115826856StephanovaDDaskalovaM2008Membrane property abnormalities in simulated cases of mild systematic and severe focal demyelinating neuropathiesEuropean Biophysics Journal3718319510.1007/s00249-007-0215-617786424StokesMG2015“Activity-silent” working memory in prefrontal cortex: a dynamic coding frameworkTrends in Cognitive Sciences1939440510.1016/j.tics.2015.05.00426051384TakahashiNSakuraiTDavisKLBuxbaumJD2011Linking oligodendrocyte and myelin dysfunction to neurocircuitry abnormalities in schizophreniaProgress in Neurobiology93132410.1016/j.pneurobio.2010.09.00420950668TarantalAFHartigan-O’ConnorDJNoctorSC2022Translational utility of the nonhuman primate modelBiological Psychiatry. Cognitive Neuroscience and Neuroimaging749149710.1016/j.bpsc.2022.03.00135283343TibshiraniR1996Regression shrinkage and selection via the lassoJournal of the Royal Statistical Society Series B5826728810.1111/j.2517-6161.1996.tb02080.xTsodyksMPawelzikKMarkramH1998Neural networks with dynamic synapsesNeural Computation1082183510.1162/0899766983000175029573407UprightNABaxterMG2021Prefrontal cortex and cognitive aging in macaque monkeysAmerican Journal of Primatology83e2325010.1002/ajp.2325033687098Valdés-TovarMRodríguez-RamírezAMRodríguez-CárdenasLSotelo-RamírezCECamarenaBSanabrais-JiménezMASolís-ChagoyánHArguetaJLópez-RiquelmeGO2022Insights into myelin dysfunction in schizophrenia and bipolar disorderWorld Journal of Psychiatry1226428510.5498/wjp.v12.i2.26435317338van VreeswijkCSompolinskyH1996Chaos in neuronal networks with balanced excitatory and inhibitory activityScience2741724172610.1126/science.274.5293.17248939866WangYunMarkramHGoodmanPHBergerTKMaJGoldman-RakicPS2006Heterogeneity in the pyramidal network of the medial prefrontal cortexNature Neuroscience953454210.1038/nn167016547512WangMGamoNJYangYJinLEWangXJLaubachMMazerJALeeDArnstenAFT2011Neuronal basis of age-related working memory declineNature47621021310.1038/nature1024321796118WangYZhangYBLiuLLCuiJFWangJShumDHKvan AmelsvoortTChanRCK2017A meta-analysis of working memory impairments in autism spectrum disordersNeuropsychology Review27466110.1007/s11065-016-9336-y28102493WaxmanSG1980Determinants of conduction velocity in myelinated nerve fibersMuscle & Nerve314115010.1002/mus.8800302076245357WimmerKNykampDQConstantinidisCCompteA2014Bump attractor dynamics in prefrontal cortex explains behavioral precision in spatial working memoryNature Neuroscience1743143910.1038/nn.364524487232YoungRGCastelfrancoAMHartlineDK2013The “Lillie transition”: models of the onset of saltatory conduction in myelinating axonsJournal of Computational Neuroscience3453354610.1007/s10827-012-0435-323306554ZuckerRSRegehrWG2002Short-term synaptic plasticityAnnual Review of Physiology6435540510.1146/annurev.physiol.64.092501.1145471182627310.7554/eLife.90964.3.sa0eLife assessmentOstojicSrdjanReviewing EditorÉcole Normale Supérieure - PSLFranceSolidValuable
This manuscript reports a valuable computational study of the effects of axon de-myelination and re-myelination on action potential speed and propagation failure. The manuscript presents solid evidence for the effects of de- and re-myelination in different models of working memory, with potential implications in disorders such as multiple sclerosis. The exposition of the manuscript is targeted for researchers interested in biophysical models of cognitive deficits.
The authors study the effects of myelin alterations in working memory via the complementary use of two computational approaches: one based on the de- and re-myelination in multicompartmental models of pyramidal neurons, and one based on synaptic changes in a spiking bump attractor model for spatial working memory. The first model provides the most precise angle (biophysically speaking) of the different effects (loss of myelin lamella or segments, remyelination with thinner and shorter nodes, etc), while the second model allows to infer the consequences of myelin alterations in working memory performance, including memory stability, duration, and bump diffusion, while also exploring the case of myeling alterations in a novel silent working memory model. The results indicate (i) a slowing down and failure of propagation of spikes with demyelination and partial recovery with remyelination, with detailed predictions on the role of nodes and myelina lamella, and (ii) a decrease in memory duration and an increase in memory drift as a function of the demyelination, in agreement with multiple experimental studies.
Strengths:
Overall, the work offers a very interesting approach of a topic which is hard to accomplish experimentally --therefore the computational take is entirely justified and extremely useful. The authors carefully designed the computational experiments to shed light into the demyelination effects on working memory from multiple levels of description, increasing the reliability of their conclusions. I think this work provides now convincing evidence and has the potential to be influential in future studies of myelin alterations (and related disorders such as multiple sclerosis).
Weaknesses:
In its current form, the authors have improved the clarity of the results and the model details, and have provided a new set of simulations to complement and reinforce the original ones (including the development of a new spatial working memory model based on silent working memory principles). I do not appreciate any significant weaknesses at this point.
This paper analyzes the effect of axon de-myelination and re-myelination on action potential speed, and propagation failure. Next, the findings are then incorporated in a standard spiking ring attractor model of working memory.
I think the results are not very surprising or solid and there are issues with method and presentation.
The authors did many simulations with random parameters, then averaged the result, and found for instance that the Conduction Velocity drops in demyelination. It gives the reader little insight into what is really going on. My personal preference is for a well understood simple model rather than a poorly understood complex model. The link between the model outcome of WM and data remains qualitative and is further weakened by the existence of known other age-related effects in PFC circuits.
Comments on revised version:
The paper has improved in the revision, although I still think a reduced model would have been nice.
10.7554/eLife.90964.3.sa3Author responseIbañezSaraAuthorBoston University Chobanian & Avedisian School of MedicineBostonUnited StatesSenguptaNilapratimAuthorBoston University Chobanian & Avedisian School of MedicineBostonUnited StatesLuebkeJennifer IAuthorBoston UniversityBostonUnited StatesWimmerKlausAuthorCentre de Recerca MatemàticaBarcelonaSpainWeaverChristina MarieAuthorFranklin & Marshall CollegeLancasterUnited States
The following is the authors’ response to the current reviews.
We thank the reviewers for their overall careful evaluation of our work, the constructive criticism, and their many helpful suggestions. We feel that our revision built on the strengths identified by the reviewers, and addressed all the concerns they have raised. Both reviewers recognize that our revisions have improved the paper. Since the first submission we have:
Rewritten large parts of the papers to improve clarity and make it more concise where possible
Simulated an alternative working memory model, as recommended by Reviewer 1
Included 4 new/revised supplementary figures, following the reviewer’s suggestions for additional analysis.
Below we provide a brief response to the Reviewers’ comments on our manuscript revision.
Reviewer #1: Public Review:
Strengths:
Overall, the work offers a very interesting approach of a topic which is hard to accomplish experimentally --therefore the computational take is entirely justified and extremely useful. The authors carefully designed the computational experiments to shed light into the demyelination effects on working memory from multiple levels of description, increasing the reliability of their conclusions. I think this work provides now convincing evidence and has the potential to be influential in future studies of myelin alterations (and related disorders such as multiple sclerosis).
Weaknesses:
In its current form, the authors have improved the clarity of the results and the model details, and have provided a new set of simulations to complement and reinforce the original ones (including the development of a new spatial working memory model based on silent working memory principles). I do not appreciate any significant weaknesses at this point.
We thank the reviewer for these positive comments on our revision and for the suggestion of adding the silent memory model, as we feel this has strengthened our findings.
Reviewer #2: Public Review:
This paper analyzes the effect of axon de-myelination and re-myelination on action potential speed, and propagation failure. Next, the findings are then incorporated in a standard spiking ring attractor model of working memory.
I think the results are not very surprising or solid and there are issues with method and presentation.
The authors did many simulations with random parameters, then averaged the result, and found for instance that the Conduction Velocity drops in demyelination. It gives the reader little insight into what is really going on. My personal preference is for a well understood simple model rather than a poorly understood complex model. The link between the model outcome of WM and data remains qualitative and is further weakened by the existence of known other age-related effects in PFC circuits.
Comments on revised version:
The paper has improved in the revision, although I still think a reduced model would have been nice.
As noted above, in addition to our spiking bump attractor model, our revision includes a second network-level model: an activity-silent working memory model for continuous features. We found qualitatively similar effects as in our bump attractor network model, showing that our main conclusions do not critically depend on the exact working memory mechanism (active vs. activity-silent). This new model was described in two new supplementary figures and a new paragraph in the Results section.
We did not add a reduced model in our revision to this paper, since neither reviewer explicitly recommended that we add one. As we noted in our private response to reviewers that accompanied our revision: we share the view that understanding simple models can provide critical insights into brain function (and we believe that many of our papers related to attractor dynamics in working memory and decision-making fall into this category, e.g. Wimmer et al. 2014, Esnaola-Acebes et al. 2022, Ibañez et al 2020). We disagree with the reviewer on an important point: we feel that the model complexity that we have chosen is appropriate and necessary to study the phenomenon at hand. Our modeling efforts are principled, with complexity added as necessary. We started with a biophysical single neuron model with firing dynamics fit to empirical data in pyramidal neurons of rhesus monkey dlPFC (Rumbell et al. 2016) – the same type of neurons and cortical region analyzed in the Peters et al. work on structural changes to myelin seen during aging (e.g., Figure 1). Because simple models do not accurately capture the CV along thin axons like those in the PFC, we attached a multicompartment axon with detailed myelinated segments, and constructed a cohort of feasible models. We then used this cohort to get quantitative estimates of the effects of variable degrees of demyelination and remyelination. This would not be possible with a simpler model. We then study the consequences of de- and re-myelination in a spiking neural network model. Again, we could not use a simpler model (e.g. a firing rate attractor model) without making gross assumptions about how demyelination affects circuit function. In sum, we believe that our models are relatively simple but comprehensive given the phenomenon that we are studying.
The reviewer is correct in that there exist “known other age-related effects in PFC circuits”. These are reviewed in the introduction and we discuss future extensions of our model that would incorporate those effects as well. It is important to note that this is the first comprehensive study of demyelination effects in aging PFC, demonstrating that myelin changes alone predict working memory changes associated with aging.
While we agree that averaging results about different parameter sets provide a limited understanding of the system, we persist in our belief that such analyses provide an important baseline. We acknowledge that results vary across our model cohort; this is why we included the heatmaps of our single cell model perturbation results (Figure 3 and Supplementary Figure 3), and simulated network models representing a heterogeneity of neuronal axons with healthy and altered myelin sheaths in different degrees, as likely occurs in the aging brain (Figures 7 and 8). The model framework we present here is well-suited for more targeted analyses and better insights, including those which we are pursuing currently.
The following is the authors’ response to the original reviews.
We thank the reviewers for their careful evaluation of our work, the constructive criticism, and their many helpful suggestions. We feel that our revision builds on the strengths identified by the reviewers, and addresses all the concerns they have raised. We have:
Rewritten large parts of the papers to improve clarity and make it more concise where possible
Simulated an alternative working memory model
Included 4 new/revised supplementary figures, following the reviewer’s suggestions for additional analysis
Reviewer #1 (Public Review):
Summary:
The authors study the effects of myelin alterations in working memory via the complementary use of two computational approaches: one based on the de- and re-myelination in multicompartmental models of pyramidal neurons, and one based on synaptic changes in a spiking bump attractor model for spatial working memory. The first model provides the most precise angle (biophysically speaking) of the different effects (loss of myelin lamella or segments, remyelination with thinner and shorter nodes, etc), while the second model allows to infer the consequences of myelin alterations in working memory performance, including memory stability, duration, and bump diffusion. The results indicate (i) a slowing down and failure of propagation of spikes with demyelination and partial recovery with remyelination, with detailed predictions on the role of nodes and myelina lamella, and (ii) a decrease in memory duration and an increase in memory drift as a function of the demyelination, in agreement with multiple experimental studies.
Strengths:
Overall, the work offers a very interesting approach of a topic which is hard to accomplish experimentally --therefore the computational take is entirely justified and extremely useful. The authors carefully designed the computational experiments to shed light into the demyelination effects on working memory from multiple levels of description, increasing the reliability of their conclusions. I think this work is solid and has the potential to be influential in future studies of myelin alterations (and related disorders such as multiple sclerosis).
We thank the reviewer for these positive comments on our manuscript.
Weaknesses:
In its current form, the study still presents several issues which prevent it from achieving a higher potential impact. These can be summarized in two main items. First, the manuscript is missing some important details about how demyelination and remyelination are incorporated in both models (and what is the connection between both implementations). For example, it is unclear whether an unperturbed axon and a fully remyelinated axon would be mathematically equivalent in the multicompartment model, or how the changes in the number of nodes, myelin lamella, etc, are implemented in the spiking neural network model.
We thank the reviewer for these suggestions to improve the clarity of our manuscript. A ‘fully remyelinated’ axon is not mathematically equivalent to the unperturbed axon: it has shorter and thinner myelinated segments, and additional nodes in between. This is consistent with empirical observations in rhesus monkey dlPFC, as reviewed in Peters et al. (2009): a 90% increase in paranode profiles, and myelin sheaths that were thinner than expected for the size of the enclosed axon. With no empirical observations of fewer numbers of nodes (but rather, the opposite) or bare sections of axon, we assumed that the remyelination process also creates new nodes (which are identical to existing nodes), as also modeled in Scurfield & Latimer (2018). We have added two new sentences to the results to clarify this fact, before presenting the first set of results for the single cell model: (starting at line 137):
“To simulate demyelination, we removed lamellae from selected myelinated segments; for remyelination we replaced a fraction of myelinated segments by two shorter and thinner segments with a node in between. As such, a ‘fully remyelinated axon’ had all the demyelinated segments subsequently remyelinated, but with fewer lamellae and additional nodes compared to the unperturbed control case, consistent with empirical observations (Peters,2009).”
We also state the maximal amount of remyelination more explicitly in the Results, starting on lines 164-165: "We next examined the extent to which remyelination with shorter and thinner segments, occurring after demyelination, restored axonal AP propagation (Figure 4).”
Also on line 192-193: “Remyelinating all affected segments with 75% of lamellae (the maximal amount of remyelination) nearly eliminated AP failures (1.8 ± 1.1%).”
Finally, in Methods we also clarified the structure of the added node (starting at line 634): “Remyelination was performed by replacing an affected (previously demyelinated) segment with two shorter segments, each including paranodes, juxtaparanodes, and an internode, and a new node between them that was identical to existing nodes.”
We have also provided further details describing how myelin dystrophy was simulated in the network model in Results (lines 243 - 249) and in Methods (lines 722 - 747). How myelin alterations have been implemented in the network model is one of the questions of the reviewer (Question 5 in Reviewer #1: Recommendations for the Authors_)._ We have addressed this question by describing in detail how we adjusted CV and AP failure rate to the values produced by the multicompartment neuron model. Please see our answer to Question 5 for the details.
Second, it is unclear whether some of the conclusions are strong computational predictions or just a consequence of the model chosen. For example, the lack of effect of decreasing the conduction velocity on working memory performance could be due to the choice of considering a certain type of working memory model (continuous attractor), and therefore be absent under other valid assumptions (i.e. a silent working memory model, which has a higher dependence on temporal synaptic dynamics).
Whether some conclusions are strong predictions or just a consequence of the model chosen is an important concern and indeed a general problem of computational modeling of working memory. For example, Stein et al. (Stein et al. Towards biologically constrained attractor models of schizophrenia, Curr. Opin. Neurobiol. 2021) showed that opposed manipulations of E/I ratio can produce the same behavioral pattern in different alternative, plausible biological network models. As long as we do not fully understand the neural mechanisms underlying working memory, modeling studies of how alterations (e.g. in E/I ratio or in the reliability and timing of axonal transmission, as we did here) affect circuit function need to be interpreted critically and tested against new experimental data.
One way to strengthen model predictions is by showing that different computational models make similar predictions. To do this, we implemented an activity-silent working memory model for continuous features, as suggested by the reviewer, and we found qualitatively similar effects as in our bump attractor network model. Thus, our main conclusions do not critically depend on the exact working memory mechanism (active vs. activity-silent).
In the revised manuscript, we have added two new supplementary figures (Supplementary Figure 8 and 9, see the next page) and a new paragraph in the Results section about activity silent working memory (starting at line 319):
“Alternative working memory mechanisms. Working memory in our neural network is maintained in an attractor state with persistent neural activity (Compte et al., 2000; Hansel and Mato, 2013). Other mechanisms have been proposed, including that working memory maintenance may rely on activity-silent memory traces (Mongillo et al., 2008; Stokes, 2015; Barbosa et al., 2020). In activity-silent models, a slowly decaying transient of synaptic efficacy preserves information without the need for persistent ongoing activity. We implemented an activity-silent model, to our knowledge the first one for continuous spatial locations, and tested how working memory performance is affected by AP failures and propagation delays. We found that AP failures corresponding to demyelination caused working memory errors qualitatively similar to the delay-active network (Supplementary Figure 8). On the other hand, increasing propagation delays did not lead to additional working memory errors, unless we include unrealistically high values (uniform distribution in the range of 0 to 100 ms; Supplementary Figure 9). These results are qualitatively similar to the delay active network model. Thus, our main findings do not critically depend on the exact working memory mechanism (active vs. activity-silent).”
Action potential failures impair working memory performance in a network model with activity-silent memory traces.
(A) Spiking and synaptic activity in an unperturbed, activity-silent working memory model. Top: Raster plot showing the activity for each excitatory neuron (labeled by its preferred direction) in a single trial with a cue stimulus presented at 180°. We modified our spiking neural network model such that it does not show elevated persistent firing throughout the delay period (see Figure 5B for comparison). In particular, we reduced the external background input to excitatory neurons by a factor of 3.61% and we increased the cue stimulus amplitude by 12.5%. Even though spiking activity decays to baseline (close to 0 Hz), a memory trace is imprinted in enhanced synaptic strength due to short-term synaptic facilitation (Mongillo et al., 2008). Selective spiking activity is recovered by a non-selective constant input applied during 300 ms to all excitatory neurons during the two reactivation periods (marked by yellow and green rectangles in the raster plot). The amplitude of the input was 11 mV during the first and 13 mV during the second reactivation period. Reactivation periods are marked in light gray shading in the remaining panels below and the cue period is indicated by dark gray shading. Firing rates (second row), synaptic facilitation variable u (third row), and synaptic depression variable x (bottom row) for the same trial, averaged for 500 neurons around the neuron with 180° as preferred direction (solid lines) and around the neuron with 0° as preferred direction (dashed lines). Note that reactivation recovers the activity bump (C) but also causes elevated firing and subsequent enhancement of synapses at all positions in the networks. (B) Activity in a network with demyelination of 50% of the myelinated segments by removing 60% of the myelin lamellae. AP failures lead to reduced firing rates in the cue and early delay periods and consequently to weaker synaptic enhancement. (C) Average spike counts of the excitatory neurons during the cue period (black lines), and the two reactivation periods indicated in the raster plots in A and B (yellow and green lines). Solid lines correspond to the control network and dashed lines to the perturbed network. (D) Memory strength as a function of time for the control and perturbed networks. (E-F) Trajectories of the bump center (i.e., remembered cue location) read out from the neural activity across the cue and delay periods using a population vector (see Methods). Cue position was 180° in all trials. The perturbed network (F) shows larger working memory errors towards the end of the delay period compared to the control network (E).
Effect of propagation delays on control and perturbed activity-silent network models.
(A) Memory strength during the whole simulation time for the young, control networks relying on activity-silent working memory (Supplementary Figure 8) with zero propagation delays (blue line), and with propagation delays from a uniform distribution with a range between 0 and 40 ms (yellow line) and between 0 and 100 ms (orange line). (B) Memory strength for perturbed networks when demyelinating 25% of the myelinated segments by removing 50% of the myelin lamellae, without delays (red line), and with uniformly distributed delays between 0 and 40 ms (light gray line) and between 0 and 100 ms (black line). The cue period is indicated by dark gray shading and reactivation periods are marked in light gray. Memory strength was calculated by averaging across 280 trials for one network. Shaded areas indicate SEM for each case. For the young, control networks (A), working memory was not affected by including delays of up to 40 ms. Unrealistically long delays ranging up to 100 ms did cause an impairment (the longest delays found for the most extreme perturbation condition – demyelination of 75% of the segments by removing 100% of the myelin lamellae – were of 49.9 ms on average). When also incorporating AP failures to the networks (B), we observed a similar trend. For this perturbation condition, delays of up to 40 ms were already much larger than the delays quantified in the single neuron model (for the case of 25% of the segments demyelinated by removing 50% of the myelin lamellae, the average delay in the cohort was 3.75 ms).
With additional simulations to address these issues, I consider that the present study would become a convincing milestone in the computational modeling of myelin-related models, and an important study in the field of working memory.
Again, we would like to thank the reviewer for the positive comments. We have addressed all the main issues raised (see below our response to the “recommendations for the authors”).
Reviewer #2 (Public Review):
This paper analyzes the effect of axon de-myelination and re-myelination on action potential speed, and propagation failure. Next, the findings are then incorporated in a standard spiking ring attractor model of working memory.
I think the results are not very surprising or solid and there are issues with method and presentation.
The authors did many simulations with random parameters, then averaged the result, and found for instance that the Conduction Velocity drops in demyelination. It gives the reader little insight into what is really going on. My personal preference is for a well understood simple model rather than a poorly understood complex model. The link between the model outcome of WM and data remains qualitative, and is further weakened by the existence of known other age-related effects in PFC circuits.
We thank the reviewer for the critical assessment of our work. We share the view that understanding simple models can provide critical insights into brain function (and we believe that many of our papers related to attractor dynamics in working memory and decision making fall into this category, e.g. Wimmer et al. 2014, Esnaola-Acebes et al. 2022, Ibañez et al 2020). However, we respectfully disagree with the reviewer on an important point: the model complexity that we have chosen is appropriate and necessary to study the phenomenon at hand. Our modeling efforts are principled, with complexity added as necessary. We started with a biophysical single neuron model with firing dynamics fit to empirical data in pyramidal neurons of rhesus monkey dlPFC (Rumbell et al. 2016) – the same type of neurons and cortical region analyzed in the Peters et al. work on structural changes to myelin seen during aging (e.g., Figure 1). Because simple models do not accurately capture the CV along thin axons like those in the PFC, we attached a multicompartment axon with detailed myelinated segments, and constructed a cohort of feasible models. We then used this cohort to get quantitative estimates of the effects of variable degrees of demyelination and remyelination. This would not be possible with a simpler model. We then study the consequences of de- and re-myelination in a spiking neural network model. Again, we could not use a simpler model (e.g. a firing rate attractor model) without making gross assumptions about how demyelination affects circuit function. In sum, we believe that our models are relatively simple but comprehensive given the phenomenon that we are studying.
The reviewer is correct in that there exist “known other age-related effects in PFC circuits”. These are reviewed in the introduction and we discuss future extensions of our model that would incorporate those effects as well. It is important to note that this is the first comprehensive study of demyelination effects in aging PFC, demonstrating that myelin changes alone predict working memory changes associated with aging.
The specific issues about modeling choices and interpretation of the results are discussed below.
Both for the de/re myelination the spatial patterns are fully random. Why is this justified?
We agree that myelin dystrophy during aging could be non-random, that is, localized to certain regions of an axon. Our collaborators (Drs Jennifer Luebke, Maya Medalla, and Patrick Hof) are currently addressing this question using 3D electron microscopy and immunohistochemistry on axons of individual neurons and their associated myelin, but results are not available yet. Early on in this study we examined how the location of myelin alterations affected AP propagation. Focusing demyelination along a section of axon led to more AP slowing and failure than when spatially randomized. Likewise, remyelination of such spatially localized dystrophy led to greater recovery, as there were fewer transitions between long and short internodes (Supplemental Figure 4). Since otherwise the effects in the localized cases were largely similar to those in the spatially random case (see Author response image 3 below), for brevity in this paper we assumed myelin alterations were randomly distributed. Our next paper, extending this study to collateralized axons and which was presented as a poster at the 2023 Society for Neuroscience meeting, will include an examination of localized myelin dystrophy.
Effect of localized myelin alterations on CV change.
Myelin alterations were either focused on the third of myelinated segments closest to the initial segment (‘proximally clustered’), the third of myelinated segments furthest from the initial segment (‘distally clustered’), or distributed according to a uniform distribution as in the current study. For demyelination, all lamellae were removed from 25% of myelinated segments (showing mean +/- SEM of all 50 cohort models, 30 randomized trials each). For remyelination, affected segments were replaced by two shorter segments with 75% of the original lamellae thickness and a node in between.
We have added two sentences in Methods to justify this assumption more clearly (line 510): “Evidence suggests that aging affects oligodendrocytes in several ways, including the ability for oligodendrocyte precursor cells to mature (Dimovasili et al., 2022). Knowing that individual oligodendrocytes myelinate axons of many different neurons, but without data quantifying how oligodendrocyte dystrophy affects myelination in individual axons, we assumed that myelin alterations were randomly distributed.”
We have also added a sentence in the Discussion alluding to our upcoming study (line 434): “Our model can also be extended to explore interactions between spatially localized myelin perturbations (such as those seen in multiple sclerosis) and axon collateralization (Sengupta et al., 2023), which would affect the distance-dependence of AP failures.”
Similarly, to model the myelin parameters were drawn from uniform distributions, Table 1 (I guess). Again, why is this reasonable?
The reviewer is correct that our initial Latin hypercube sample generated a uniform distribution. However, parameters of the random sample of models selected as biologically feasible were not uniformly distributed. We have added a new figure (Supplementary Figure 1A) to illustrate the parameter distributions, and have added two sentences in Methods (starting on line 596):
“Of the 1600 simulated models, 138 met these criteria; for the present study, we randomly selected 50 models to comprise the young, control model cohort. Along most dimensions, the chosen cohort was approximately normally distributed (Supplementary Figure 1). The g-ratio (ratio of axon to fiber diameter) among models in the cohort was 0.71 ± 0.02, with total axon lengths of 1.2 ± 0.1 cm.”
Distribution of parameters and conduction velocities in the single neuron model cohort.
(A) Histograms of axon morphology parameters of models selected for the single neuron cohort. Top: axon diameter: middle, length of unperturbed myelin segments; bottom: total myelin thickness in unperturbed segments, computed as the product of lamella thickness and number of lamellae. (B) Histograms of the CV for the 50 axons of the unperturbed model cohort (top), and representative demyelination and remyelination perturbations: mild demyelination (removing 25% of lamellae from 25% of the myelinated segments, second row); severe demyelination (removing all lamellae from 75% of the myelinated segments, third row); and complete (100%) remyelination (where the demyelinated segments from the third row were remyelinated by two shorter segments with 75% of lamellae). CVs averaged over 30 trials in each case. (C) Changes in CV (measured in %) in response to demyelination and remyelination versus the magnitude of current clamp step (+180, +280, or +380 pA). Shown are mean +/- SEM for demyelinating 50% of myelinated segments (removing all lamellae), and subsequent remyelination of those segments by shorter segments with 75% of lamellae.
The focus of most analysis is on the conduction velocity but in the end, this has no effect on WM, so the discussion of CV remains sterile.
CV delays likely do affect brain functions that rely on neuronal oscillations and synchrony, as mentioned in the Discussion. As such, we feel that our single neuron model results on CV delays as well as AP failures are valuable for the scientific community. Yet, given the results of our network models here, the reviewer has a valid point. We have clarified in the introduction that AP failures but not CV delays affected the network output (line 115):
“Higher degrees of demyelination led to slower propagation and eventual failure of APs along the axons of the multicompartment models. In the network models, an increase in AP failure rate resulted in progressive working memory impairment, whereas slower conduction velocities, in the range observed in the multicompartment models, had a negligible effect.”
We have also revised the single neuron section of the Results throughout, to better highlight the effects of myelin dystrophy on AP failures. Revisions to address this in the demyelination section start on line 148:
“AP propagation was progressively impaired as demyelination increased (Figure 3): CV became slower, eventually leading to AP failure. Removing 25% of lamellae had a negligible effect on CV, regardless of how many segments were affected. However, when all lamellae were removed, CV slowed drastically – by 38 ± 10% even when just 25% of the segments were demyelinated in this way, and 35 ± 13% of APs failed. When 75% of segments lost all their lamellae, CV slowed by 72 ± 8% and 45 ± 13% of APs failed.”
Similiarly, we have added several sentences about AP failures that remain after remyelination of the single neuron model (starting on line 190):
“Results for the percentage of AP failures (Figure 4C,F) were consistent with those for CV recovery. Remyelinating all previously demyelinated segments, even adding just 10% of lamellae, brought AP failure rates down to 14.6 ± 5.1%. Remyelinating all affected segments with 75% of lamellae (the maximal amount of remyelination) nearly eliminated AP failures (1.8 ± 1.1%). Incomplete remyelination, where some segments were still demyelinated, still had relatively high AP failure rates. For example, when one eighth of segments were remyelinated with the maximal amount of lamellae and one eighth were left bare, 25.7 ± 11.5% of APs failed across the cohort (Figure 4C, red dashed line and arrow). AP failure rates were slightly lower when starting with partial demyelination: 10.6 ± 7.6% of APs failed in the analogous paradigm (Figure 4F, red dashed line and arrow). In short: combinations of demyelinated and remyelinated segments often led to sizable CV delays and AP failures.”
The more important effect of de/re myelination is on failure. However, the failure is, AFAIK, just characterized by a constant current injection of 380pA. From Fig 2 it seems however that the first spike is particularly susceptible to failure. In other words, it has not been justified that it is fine to use the failure rates from this artificial protocol in the I&F model. I would expect the temporal current trace to affect whether the propagation fails or not.
In general, we did not find the first spike to be more susceptible to failure than latter spikes; the trace in Figure 2 is a representative snapshot intended to illustrate CV slowdown, AP failure, and recovery. Regarding the constant current injection: while the reviewer is correct that neurons do not receive such inputs in vivo, the applied current injections were designed to match in vitro current clamp protocols for these rhesus monkey neurons. While our future studies will include responses to more realistic synaptic inputs, we focused on somatic current injections here. We have added a new panel (C) to Supplementary Figure 1 (see previous response above) showing that the current step magnitude had little effect on the CV change after myelin perturbations; there was little effect on AP failure rates too. We now also state this finding more explicitly in Methods (starting on line 561):
“As done during in vitro electrophysiological experiments (Chang et al., 2005; Ibanez et al., 2020) and past modeling studies (Coskren et al., 2015; Rumbell et al., 2016), we first applied a holding current to stabilize the somatic membrane potential at -70 mV, then injected a current step into the somatic compartment for 2 seconds. …The CV changes in response to myelin alterations were relatively insensitive to variations in the magnitude of suprathreshold somatic current steps (Supplementary Figure 1C), and whether the current was constant or included Gaussian noise. Therefore, here we quantified CV changes and AP failures from responses to constant +380 pA current steps only.”
I don't know if there are many axon-collaterals in the WM circuits and or distance dependence in the connectivity, but if so, then the current implementation of failure would be questionable.
We agree that axon collaterals may affect our results; our unpublished morphological analyses of individual neuron axons indicate that there is a high degree of local axon collateralization in Layer 3 pyramidal neurons in LPFC. In this first study from our group on myelin perturbations, we chose to focus here on unbranched axons. There was some distance dependence of AP failure along the length of the axon. For example, in our most extreme demyelination case (75% of segments losing all their lamellae), about 14% of the axons showed more AP failure at their distal ends relative to the middle (mean difference 6.33%). We are examining this distance dependence more broadly in our next study, now cited in the Discussion (line 434): “Our model can also be extended to explore interactions between spatially localized myelin perturbations (such as those seen in multiple sclerosis) and axon collateralization (Sengupta et al., 2023), which would affect the distance-dependence of AP failures.”
I would also advise against thresholding at 75% failure in Fig3C. Why don't the authors not simply plot the failure rate?
We thank the reviewer for this suggestion, and have made this change. As suggested by the reviewer, we now show the AP failure rate in Figure 3 and Figure 4. The trends shown are nearly identical to those from the high failure trials.
Regarding the presentation, there are a number of dead-end results that are not used further on. The paper is rather extensive, and it would be clearer if written up in half the space. In addition, much information is really supplementary. The issue of the CV I already mentioned, also the Lasso regression for instance remains unused.
We understand the reviewer’s perspective, and we do value brevity when possible. During the revision process we examined the paper carefully, and made things more concise when it was feasible. As mentioned above, reporting CV results is important, though these revisions increased emphasis on results for AP failures in our revision. We combined the two Supplementary Figures about remyelination in the single neuron model into one (Supplementary Figure 3). We also moved the Lasso figure and associated methods to the Supplementary Material (Supplementary Figure 2), and have separated the Lasso results for demyelination and remyelination into their respective paragraphs (lines 154-160 and lines 200-204 respectively). While we do not use the Lasso explicitly later in Results, we cite them in the Discussion when comparing our findings to previous work (starting on line 417):
“Since our single neuron cohort sampled a wide range of parameter space, we used Lasso regression to identify which of the complex, interacting parameters contributed most to CV delays (which preceded AP failures). Parameters including axon diameter, node length, length of myelinated segments, and nodal ion channel densities predicted how our models responded to demyelination and remyelination; these findings are consistent with past modeling studies over more limited parameter ranges (e.g., Goldman and Albus, 1968; Moore et al., 1978; Babbs and Shi, 2013; Young et al., 2013; Schmidt and Knösche, 2019).”
We hope that our revision has struck an appropriate balance between clear and concise writing, and addressing concerns from both reviewers. We greatly value the time you have given to help us to improve our manuscript.
Response to Recommendations for the Authors:
Reviewer #1 (Recommendations for the Authors):
As I mentioned above, I consider that this study is well designed and it offers very interesting results. I have detailed below some of the issues that should be addressed to improve its potential impact in the field:
(1) Across the manuscript, it is not entirely clear how the results of the multicompartmental model compare to existing modeling results on demyelination and CV changes (such as in the papers cited by the authors). Is this section confirming previous results with a new (more accurate) computational model, or are there any new insights previously unreported? A new paragraph in the Discussion putting these results in context would be very useful for the reader.
We thank the reviewer for this suggestion. We have added two new subheadings to organize the Discussion better, and have expanded the single neuron section to three paragraphs. We feel this now clarifies how our model fits in with previous work while stating its novelty more explicitly. Starting on line 391:
“Myelin changes affect AP propagation in a cohort of model neurons
The novelty of our neuron model lies in its systematic exploration of a combination of different myelin perturbation types known to occur in myelin dystrophies, across a wide range of biologically feasible models. Our single neuron model assumed that age-related myelin dystrophies (e.g., Figure 1) alter the insulative properties of lamellae analogously to demyelination, and examined interactions between demyelination and remyelination. Past studies of myelin dystrophy examined how either demyelination or remyelination of all segments affected AP propagation for a few representative axon morphologies. For example, Scurfield and Latimer (2018) explored how remyelination affected CV delays, finding that axons with more transitions between long and short myelinated segments had slower CV (Supplementary Figure 4), and was first to explore how remyelination interacts with tight junctions. However, their study did not couple remyelination and demyelination together or examine AP failures. Other basic findings from our single neuron cohort are consistent with past modeling studies, including that demyelination caused CV slowing and eventual AP failures (Stephanovaetal., 2005; StephanovaandDaskalova,2008; NaudandLongtin,2019), and, separately, that remyelination with shorter and thinner myelinated segments led to CV slowing (Lasiene et al., 2008; Powers et al., 2012; Scurfield and Latimer, 2018). However, by assuming that some previously demyelinated segments were remyelinated while others were not, we found that models could have much higher AP failure rates than previously reported. Such a scenario, in which individual axons have some segments that are normal, some demyelinated, and some remyelinated, is likely to occur. We also found a few neurons in our cohort showing a CV increase after remyelination, which has not generally been reported before and is likely due to an interplay between ion channels in the new nodes and altered electrotonic lengths in the perturbed myelinated segments (e.g., Waxman, 1978; Naud and Longtin, 2019).
Since our single neuron cohort sampled a wide range of parameter space, we used Lasso regression to identify which of the complex, interacting parameters contributed most to CV delays (which preceded AP failures). Parameters including axon diameter, node length, length of myelinated segments, and nodal ion channel densities predicted how our models responded to demyelination and remyelination; these findings are consistent with past modeling studies over more limited parameter ranges (e.g., Goldman and Albus, 1968; Moore et al., 1978; BabbsandShi,2013;Youngetal.,2013;SchmidtandKnösche,2019). Better empirical measurements of these parameters in monkey dlPFC, for example from 3-dimensional electron microscopy studies or single neuron axon studies combined with markers for myelin, would help predict the extent to which myelin dystrophy and remyelination along individual axons with aging affect AP propagation.
Another important feature of our multicompartment model is that it was constrained by morphologic and physiological data in rhesus monkey dlPFC —an extremely valuable dataset from an animal model with many similarities to humans (Upright and Baxter, 2021; Tarantal et al., 2022). While beyond the scope of the current study, this computational infrastructure –with a detailed axon, initial segment, soma, and apical and basal dendrites– enables simultaneous investigations of signal propagation through the dendritic arbor and axon. Our model can also be extended to explore interactions between spatially localized myelin perturbations (such as those seen in multiple sclerosis) and axon collateralization (Sengupta et al., 2023), which would affect the distance-dependence of AP failures. Integrating such results from single neuron models into network models of working memory, as we have done here, is a powerful way to connect empirical data across multiple scales.”
(2) Although the authors provide a well-designed study for the multi-compartmental model, it would be useful to add more details about how an unperturbed model and a completely remyelinated model differ in practice, perhaps right before the first results on the single cell model are presented. Are the new myelin sheaths covering the same % of axon as in the original case? Are there the same number of nodes? It is hard to distinguish which of these results are due to a compensation by the new myelin sheaths and which ones are just the model coming back to its original (and mathematically equivalent) starting point.
A ‘fully remyelinated’ axon is not mathematically equivalent to the unperturbed axon. Newly remyelinated segments had at most 75% of the original number of myelin wraps, with a new node in between, consistent with empirical observations in rhesus monkey dlPFC. Our manuscript changes in response to this recommendation are described in detail above in our response to the public review of the same reviewer.
(3) The authors observe a directed component in the bias that is known to be caused by heterogeneities in network connectivity, as stated in the text. It occurs to me that similar effects could be also caused by an heterogeneous demyelination in parts of the network. Inducing these biases could be another potential effect of demyelination in practice, and could be easily revealed by the author's current model (and displayed in a supplementary figure).
As suggested by the reviewer, we have tested heterogeneous demyelination in parts of the network and the results confirm the reviewer’s intuition. We have included these new results as new Supplementary Figure 7 (see below) and we have added the following sentences in the Legend of Figure 5, line 1265: “When demyelination is restricted to a part of the network, diffusion only increases in the perturbed zone (Supplementary Figure 7).” and in the Discussion (line 457): “In addition to age-related changes in memory duration and precision, our network model predicts an age-related increase in systematic errors (bias) due to an increased drift of the activity bump (Supplementary Figure 11). Moreover, if demyelination is spatially localized in a part of the network, the model predicts a repulsive bias away from the memories encoded in the affected zone (Supplementary Figure 7).”
Effect of spatially heterogeneous demyelination of the model neurons according to their preferred angle.
We also tested working memory performance in the network when demyelination affects only parts of the network. The figure shows the decoded bump center position during the cue and delay period for the eight possible cue directions when a fraction of neurons was perturbed and the rest of the neurons in the circuit were unaltered (Figure 5B). We perturbed 10% of the neurons around the neuron with preferred direction 90° (left panel), 25% of the neurons around -90° (middle panel), and 50% of the neurons around 180° (right panel). Bump traces for cues that lie inside the perturbed portion of the circuit are shown in blue. Network perturbation in the three cases consisted in demyelinating 25% of the segments along the axons of model neurons, by removing 70% of the myelin lamellae. In each case, 280 trials were simulated for one network. These simulations show an increased drift and diffusion inside the perturbed zone, consistent with the increased drift and diffusion when perturbing the entire network (Figure 6B and Supplementary Figure 11). In particular, spatially heterogeneous demyelination in our network leads to a bias away from the affected zone and to increased trial-to-trial variability. Note that this is a model prediction, but we are not aware of empirical data showing heterogeneous demyelination with aging. Further, note that while our network model has a topological ring structure, neurons in PFC are not anatomically arranged depending on their preferred features. Thus, spatially heterogeneous demyelination would likely affect neurons with different feature preferences (i.e., neurons throughout our ring model).
(4) The bump attractor model of WM relies on a continuous attractor dynamics to encode the information stored in memory --a fixed point dynamics that can only vary via the slow noise-driven drift. This means, as the authors mention, that changes in CV won't affect the performance of WM in their model. This seems to be a limitation of the model, or at least an effect which is highly dependent on the modeler's choice, rather than an accurate prediction. While testing the effects of oscillations (as the authors argue in the Discussion) might be out of the scope of this work, there are other WM models which are more sensitive to temporal differences in activity. The authors should test whether the same (lack of) effects are also found in other WM models. A silent WM model seems to be the ideal candidate for this, as the authors already have the key dynamics of that model incorporated in their computational framework (namely, short-term synaptic facilitation in excitatory synapses).
We fully agree that considering the effects of demyelination in networks with alternative mechanisms would strengthen our manuscript. As suggested by the reviewer, we have simulated demyelination effects (AP failures and changes in CV) in an activity silent working memory model. The results are described in detail above in our response to the public review of the same reviewer.
We also would like to mention that we have now also tested larger conduction delays in the bump attractor model, revealing additional working memory errors. This is shown in the revised version of Supplementary Figure 6 (see below). However, those delays are unrealistically large and thus the main effect in both the bump attractor and the activity-silent model is due to AP failures.
Effect of propagation delays on control and perturbed networks.
(A) Memory strength (left panels) and diffusion (right panels) for the young, control networks with zero propagation delays (blue solid line), as in Figure 5, and with propagation delays from a uniform distribution with a range between 0 and 100 ms (yellow dashed line). (B) Memory strength and diffusion for perturbed networks when demyelinating 50% of the segments along the axons of model neurons, by removing 60% of the myelin lamellae without delays (red solid line), and with delays from a uniform distribution with a range between 0 and 40 ms (gray dashed line) and between 0 and 85 ms (black dash-dotted line). The measures of working memory performance were calculated by averaging across 20 networks and 280 trials for each network. Shaded areas indicate SEM for each case. For the young, control networks, there was no difference with and without propagation delays, even though the delays used in the network simulations were much larger than the delays quantified in the single neuron model (the longest delays found for the most extreme perturbation condition –demyelination of 75% of the segments by removing 100% of the myelin lamellae– were of 49.9 ms on average; A). Working memory performance was also unaffected in the perturbed network with AP failures for delays ranging between 0 and 40 ms, also larger than the ones quantified in the single neuron model (for the case of 50% of the segments demyelinated by removing 60% of the myelin lamellae, the average delay in the cohort was 4.6 ms and the maximum delay was 15.7 ms; B). However, including extremely long delays of up to 85 ms did further impair memory compared to the impairment level introduced by AP failures alone (B).
(5) Impact of demyelination and remyelination on working memory: Could the authors explain here how these biologically detailed alterations are implemented in the bump attractor model? Is the CV and AP failure rate adjusted to the values produced by the multicompartment neuron model with these myelin alterations?
Yes, the reviewer is right, the CV and AP failure rate have been adjusted to the values produced by the multicompartment neuron model. To clarify this in the manuscript, we have restated the text as follows:
Lines 243 - 249 (Results):
To investigate how myelin alterations affect working memory maintenance, we explored in the network model the same demyelination and remyelination conditions as we did in the single neuron model. Because our network model consists of point neurons (i.e., without detailed axons), we incorporated CV slowing as an effective increase in synaptic transmission delays (see Methods). To simulate AP failures, we adjusted the AP failure rate to the values given by the single neuron model, by creating a probabilistic model of spike transmission from the excitatory presynaptic neurons to both the excitatory and inhibitory postsynaptic neurons (see Methods).
Lines 722 - 747 (Methods):
Modeling action potential propagation failures in the network. The network model is composed of point neurons without an explicit model of the axon. To effectively model the action potential failures at the distal end of the axons quantified with the single neuron model under the different demyelination and remyelination conditions, the AP failure rate was adjusted to the values produced by the single neuron model. To do this, we perturbed the 10 control networks by designing a probabilistic model of spike transmission from the excitatory presynaptic neurons to both the excitatory and inhibitory postsynaptic neurons. From the single neuron model, for each demyelination/remyelination condition, we quantified the probability of AP failure for each of the neurons in the control cohort, as well as the percentage of those neurons that shared the same probabilities of failure. That is, the percentage of neurons that had probability of failure = 0, probability of failure = 1 or any other probability. Then, we computed the probability of transmission, , and we specified for the corresponding percentages of excitatory neurons in the networks. Thus, in the network model, we took into account the heterogeneity observed in the single neuron model under each demyelination/remyelination condition.
Modeling conduction velocity slowing in the network. To explore the effect of CV slowing along the axons of model neurons, we simulated 20 young, control networks and 20 perturbed networks with AP failure rates adjusted for the case of single model neurons with 50% of the segments demyelinated along the axons by removing 60% of the myelin lamellae (we ran 280 trials for each network). Then, we added random delays uniformly distributed with a minimum value of 0 ms in both cases, a maximum value of 100 ms in the control networks, and a maximum values of 40 ms and 85 ms in the perturbed networks, in both the AMPA and NMDA excitatory connections to both E and I neurons (Supplementary Figure 6). These large values were chosen because we wanted to illustrate the potential effect of CV slowing in our network and smaller, more realistic, values did not have any effect.
(6) "We also sought to reveal the effect on working memory performance of more biologically realistic network models with AP transmission probabilities matched to both axons with intact and with altered myelin sheaths, as likely occurs in the aging brain (Figure 1). Thus, we ran network model simulations combining AP failure probabilities corresponding to groups of neurons containing intact axons and axons presenting different degrees of demyelination." I fail to see the difference with respect to the results in previous sections. Is it that now we have subnetworks in which axons are intact and subnetworks with significant AP failures, while before there was no topological separation between both cases? Please clarify.
In Figures 5 and 6 the AP failure rate of the neural population in the network simulations was matched to the AP failure rate of the cohort of single model neurons for each demyelination/remyelination condition. Since not all model neurons have equal features, a given condition produces different levels of impairment in its neuron. Thus, we quantified the probability of AP failure for each neuron in the control cohort, as well as the percentage of those neurons that shared the same probabilities of failure. Then, we computed the probability of AP transmission for the corresponding percentages of excitatory neurons in the networks. Thus, in the network model, we took into account the heterogeneity observed in the single neuron model under each demyelination/remyelination condition.
However, In Figures 7 and 8, we consider additional heterogeneity due to a different degree of demylination/remyelination of different neurons. Here, excitatory neurons in the network model are not perturbed according to a single demyelination/remyelination condition. Instead, we allowed that different percentages of excitatory neurons had AP failure rates corresponding to different demyelination/remyelination conditions: some were unperturbed, while others had different degrees of demyelination (Figure 7) and different degrees of remyelination (Figure 8). We have modified the text for clarification in several places.
First, when we describe the impact of demyelination on working memory, we already mention that (line 271): “In each of the 10 networks, we set the AP failure rate of the excitatory neurons according to the distribution of failure probabilities of the neurons in the single neuron cohort for the given demyelination or remyelination condition. Thus, we took into account the heterogeneity of demyelination and remyelination effects from our single neuron cohort (Figure 3A; Supplementary Figure 3). Note that this heterogeneity originates from differences in axon properties, but probabilities of failure for all neurons in the network correspond to the same degree of demyelination (Figure 6). We will also consider networks that contain different combinations of axons with either intact or perturbed myelin (Figure 7 and Figure 8).”
Second, we have combined the text describing Figures 7 and 8 under a single section title, which reads “Simulated heterogenous myelin alterations match empirical data” (line 334) and start this section with (line 337): “Up to this point we have studied network models with AP failure probabilities corresponding to a single degree of myelin alterations (i.e., with all excitatory neurons in the network having AP failure rates matched to those of the single neuron cohort for one particular demyelination or remyelination condition). Next, we sought to reveal the effect on working memory performance of more biologically realistic network models, where excitatory neurons in the networks were perturbed according to a combination of different demyelination or remyelination conditions. That is, we simulated networks with excitatory neurons having AP failure probabilities matched to both neuronal axons with intact and with altered myelin sheaths in different degrees, as likely occurs in the aging brain (Figure 1).”
(7) "Unexpectedly, our model indicates that compared to the performance of networks composed of neurons possessing axons with intact myelin sheaths, both demyelination and remyelination leads to an impaired performance." This conclusion is quite interesting, but I lack intuition from the paper as of why it is happening. In fact, the authors say in the Discussion that "complete remyelination of all the previously demyelinated segments with sufficient myelin, with fewer transitions between long and short segments, recovered working memory function." Would we then see a minimum and then an increase in memory duration in Figure 9B if we extended the X-axis until we hit 100% of new myelin sheaths?
This is a very important question that we have carefully addressed in Results and Discussion. We distinguish between two remyelination cases in the models. Complete remyelination: when all (100%) the previously demyelinated segments have been subsequently remyelinated, and incomplete remyelination: when less than 100% (25%, 50% or 75%) of the demyelinated segments have been remyelinated. Figure 6 (middle and right columns) shows the two cases (black lines for any percentage of lamellae added vs. colored lines): for 100% of the segments remyelinated, the network performance is nearly or completely (when enough lamellae are added) recovered to the young network performance. In fact, with the single neuron model we observe that (lines 192 - 193 in Results): “Remyelinating all affected segments with 75% of lamellae (the maximal amount of remyelination) nearly eliminated AP failures (1.8 ± 1.1%)”. However, incomplete remyelination recovers the performance compared to demyelination (middle and right columns in Figure 6 vs left column), but this performance is worse than the performance of the young networks. The single neuron model shows that (lines 194 - 197 in Results): “Incomplete remyelination, where some segments were still demyelinated, still had relatively high AP failure rates. For example, when one eighth of segments were remyelinated with the maximal amount of lamellae and one eighth were left bare, 25.7 ± 11.5% of APs failed across the cohort (Figure 4C, red dashed line and arrow).”
In Figure 9B (now Figure 8B), we combine intact axons with axons that are only partially remyelinated (i.e., incomplete remyelination). Extending the X-axis in Figure 8B until 100% of new myelin sheaths would not imply a minimum and a subsequent increase, but a continuous impairment: the more axons we perturb (remyelinate) the higher is the impairment compared to the young cases where all the axons are intact.
The sentence "Unexpectedly, our model indicates that compared to the performance of networks composed of neurons possessing axons with intact myelin sheaths, both demyelination and remyelination leads to an impaired performance.", now reads as (lines 379 380 in Results): “Therefore, both demyelination and incomplete remyelination lead to impaired performance in our networks, compared to networks with intact myelin sheaths”. We have also rewritten the corresponding section in Discussion (lines 486 - 489) as follows: “Therefore, it is reasonable to assume that ineffective remyelination may lead to working memory impairment. In fact, complete remyelination of all previously demyelinated segments with sufficient myelin, with fewer transitions between long and short segments, led to full recovery of working memory function.”
(8) [minor] "Our recent network model found that age-related changes in firing rates and synapse numbers in individual neurons can lead to working memory impairment (Ibañez et al., 2020), but did not consider myelin dystrophy." Could you be more precise about which age-related changes were studied in Ibanez et al. 2020? From the paper it seems like it was mostly cellular excitability and synaptic density, so this should be added here for more context.
To clarify this, we have added the following sentences in the Introduccion (line 105):
“Our recent network model revealed that the empirically observed age-related increase in AP firing rates in prefrontal pyramidal neurons (modeled through an increased slope of the f-I curve) and loss of up to 30% of both excitatory and inhibitory synapses (modeled as a decrease in connectivity strength) can lead to working memory impairment (Ibañez et al., 2020), but this model did not incorporate the known changes to myelin structure that occur during normal
aging.”
(9) [minor] "Recurrent excitatory synapses are facilitating, which promotes robust and reliable persistent activity despite spatial heterogeneities in the connectivity or in the intrinsic properties of the neurons." It would be great to add a reference here to justify the inclusion of this type of plasticity in the excitatory circuit (for example Wang, Markram et al. Nat Neuro 2006).
We have added the references suggested by the reviewer and a further one in the Results (line 216):
“Recurrent excitatory synapses are facilitating, as has been empirically observed in PFC (Hempel et al., 2000; Wang et al., 2006), which promotes robust and reliable persistent activity despite spatial heterogeneities in the connectivity or in the intrinsic properties of the neurons.”
References:
Hempel, C. M., Hartman, K. H., Wang, X. J., Turrigiano, G. G., and Nelson, S. B. (2000). Multiple forms of short-term plasticity at excitatory synapses in rat medial prefrontal cortex. J. Neurophysiol. 83, 3031–3041. doi: 10.1152/jn.2000.83.5.3031
Wang, Y., Markram, H., Goodman, P. H., Berger, T. K., Ma, J., and Goldman- Rakic, P. S.(2006). Heterogeneity in the pyramidal network of the medial prefrontal cortex. Nat.Neurosci. 9, 534–542. doi: 10.1038/nn1670