First, we wondered what specific type of PCD is responsible for sculpting VNC lineages in D. melanogaster, reasoning that only by gaining insight into the exact developmental process involved can we understand its role in nervous system evolution. The majority of studies on neuronal PCD in insects have focused on its role at metamorphic transitions, where death eliminates fully differentiated neurons either at puparium formation (Truman et al., 1994) or in adults post-eclosion (Draizen et al., 1999; Kimura and Truman, 1990). Both of these remodelling events are gated by ecdysteroids. However, our previous observations in Drosophila (Truman et al., 2010), together with studies in the fly brain (Nordlander and Edwards, 1968; Kumar et al., 2009; Lin et al., 2010; Lovick et al., 2017), made us consider that hemilineage-specific PCD takes place early, in newly born neurons. So far, the dynamics of cell death has been difficult to evaluate on a cell-by-cell basis within a complex nervous system.

To interrogate postembryonic neuronal death, we have built a novel genetically encoded effector caspase probe called SR4VH (Figure 2A,B). SR4VH consists of a membrane-bound red fluorescent protein (Src::RFP) and a yellow fluorescent protein with a strong nuclear localisation signal from histone H2B (Venus::H2B) separated by four tandem repeats of the amino acid sequence DEVD. When effector caspases cleave the DEVD site, Venus accumulates in the nucleus while RFP remains bound to the cell membrane (Figure 2B). This reporter is similar in design to Apoliner (Bardet et al., 2008), but has different subcellular localisation signals as well as four tandem caspase cleavage sites instead of one. We also found that tethering the probe to the membrane with the myristoylation signal from Src means that there is no excess signal accumulation in the Golgi apparatus (Mukherjee et al., in preparation). The nuclear localisation signal from H2B allows for highly efficient sequestration of cleaved Venus in the nucleus even in late stages of apoptosis, when the nuclear membrane is likely compromised.

Using the GAL4/UAS system and the NB driver Worniu-GAL4, we found we could visualise postembryonic neurogenesis and label up to 20 of the most recently born progeny from a single NB (this is due to GAL4 and reporter perdurance). The number of progeny we can detect at any one time using Worniu-GAL4 varies from 10 to 20, most likely as a result of differential proliferation rates across lineages. We confirmed that SR4VH is reliable as a reporter for cell death in larvae by analysing its expression pattern in all lineages of postembryonic neurons in the thoracic VNC and comparing it to our previous work on MARCM homozygous mutant clones of the initiator caspase Dronc (Truman et al., 2010; Figure 2C,D,E,F and Figure 2—figure supplement 1). We found dying cells associated with lineages in the brain and VNC throughout the whole of postembryonic neurogenesis (Figure 2C,D,E,F), which lasts for 3.5 days, from mid-2^nd instar (L2) to 12 hr after pupariation. As previously suggested (Truman et al., 2010), the time course of PCD indicates that cells die early – very soon after they are born – often before they have even extended a neuritic process. This death appears to be unlike the ‘trophic’ PCD found in vertebrates, where a neuron extends a process, interacts with its target cell and dies in the absence of appropriate survival signals. In support of an early onset of PCD, we were able to see sequential stages of cell death, dependent on the distance from the NB (Figure 2G,H,I and Figure 2—figure supplement 1B). Older cells located further away from the NB appear to be at a more advanced stage of PCD indicated by the complete translocation of Venus from the membrane to the nucleus (Cell three in Figure 2G,H) and by the accumulation of RFP-positive dead cell membranes close to the lineage bundle (arrowheads in Figure 2E,F).

The number of dying cells within a doomed lineage varied from 1 to 8, with most lineages containing 1–2 dying cells from a total of 10–20 cells labelled with Worniu-GAL4 (dying cells/lineage: 1.3 ± 1.5 given as average ± standard deviation; n = 444 doomed lineages from 5 VNCs). From a total of 444 doomed lineages, 243 harboured more than one dying cell, of which 148 displayed a progression of cell death (Figure 2G,H). Truman and Bate, 1988 approximated the cell cycle of an NB to 55 min and that of a GMC to 6.5 hr, with 7 GMCs present in a proliferating lineage at all times. Therefore, after subtracting the NB and GMCs from clusters of 10–20 Worniu-GAL4-labelled cells, 2–12 will be neurons which resulted from 1 to 6 divisions, each separated in time by 55 min. This means that PCD was initiated early, at some time between 0 and 5.5 hr after neurons were born.

To look at death specifically during the development of the MNB lineage (lineage 0) we imaged SR4VH in wandering L3 larvae and used molecular markers to identify members of hemilineage 0A and 0B. The transcription factors Engrailed/Invected (En/Inv) are known to be expressed in immature and fully differentiated interneurons of hemilineage A (Allen et al., 2020; Lacin et al., 2019; Lacin et al., 2014; Truman et al., 2004). As previously reported, the mature differentiated octopaminergic neurons found in hemilineage B express the transcription factor Vestigial (Vg) (Landgraf et al., 2003). Here, we find that a small number of immature postembryonic neurons (about 3–5) in close proximity to the MNB also express Vg (Figure 2J). Within these immature neurons the expression of Vg and En are mutually exclusive. Using Worniu-GAL4 to drive SR4VH we found that only the engrailed-negative cells are undergoing apoptosis (Figure 2K), i.e. the same small number of cells that express Vg. Their proximity to the MNB suggests that Vg-positive B cells (i.e. immature octopaminergic neurons) undergo an early death, very soon after they are born.

After observing the extent of early PCD during development, we wondered if, by reducing PCD, we could generate novel functional expansions of a hemilineage. To explore this, we made use of the powerful genetic tools available in Drosophila to block PCD in the MNB lineage to determine if ‘undead’ cells survive into adulthood, elaborate their neurites and acquire a distinctive neurotransmitter identity.

From our previous work (Truman et al., 2010), we know that during postembryonic neurogenesis MNB hemilineage A survives, expresses Engrailed (Truman et al., 2004) and differentiates into GABAergic interneurons (Lacin et al., 2019). Because during embryonic development the MNB hemilineage B produces a small number of octopaminergic neurons, we hypothesised that preventing PCD would generate additional octopaminergic neurons in the later postembryonic phase of neurogenesis. In postembryonic nomenclature, all the neurons generated by the MNB are collectively called lineage 0 and therefore we will refer to octopaminergic neurons generated by the MNB as hemilineage 0B.

Using the octopaminergic neuron driver, TDC2-GAL4, we observed a 4- to 9-fold increase in the number of octopaminergic neurons in the thoracic VNC of H99/XR38 adult flies deficient for proapoptotic genes (hid^+/-, grim^+/-, rpr^-/- and skl^+/-) (Peterson et al., 2002; White et al., 1994) compared with wild-type control animals (Figure 3A,B,C,D), (T1: 20.9 ± 2.3, Mann-Whitney U = 0, p = 0.0002; T2: 26.3 ± 4.4, Mann-Whitney U = 0, p=0.0004; T3: 27.5 ± 3.5, Mann-Whitney U = 0, p = 0.0004; n = 11 each). These ‘undead’ neurons also express the vesicular glutamate transporter VGlut (Figure 3E,F), just like wild-type octopaminergic neurons (Greer et al., 2005).

Ideally, to label and manipulate dying neurons from hemilineage 0B, we require a specific driver line expressed only in the newly born doomed neurons. To test if we could use TDC2-GAL4 to label and manipulate dying neurons from hemilineage 0B during their development, we performed a timeline of expression in wild-type and H99/XR38 flies (Figure 3—figure supplement 1). Unfortunately, even though undead neurons are generated from L2 onwards in H99/XR38 flies, the TDC2-GAL4 is only active in the undead cells days later. Gradually, in pupae, TDC2-GAL4 expression reveals the remaining undead B cells (Figure 3—figure supplement 1C,D). We concluded that the TDC2 driver line cannot be used to visualise and manipulate newly born postembryonic ‘doomed cells’. Instead, TDC2-GAL4 allowed us to accurately reveal ‘undead’ hemilineage 0B neurons but only in the adult (see cartoon Figure 3G).

To ensure sparse labelling and the precise manipulation of only doomed cells from hemilineage 0B, we generated postembryonic TDC2-GAL4-expressing MARCM clones homozygous for the loss-of-function allele Dronc^ΔA8 (in which PCD is inhibited) (Kondo et al., 2006; Truman et al., 2010). This strategy guarantees that, even though cell death can be rescued in other lineages, it is only within the TDC2-positive postembryonic neurons that UAS-based tools are expressed.

Analysis of the projection patterns of undead neurons revealed that they display both common and distinct features compared to their wild-type embryonically born counterparts. Similar to wild-type octopaminergic cells (Monastirioti et al., 1995), the undead neurons have cell bodies located ventrally at the midline, at the posterior border of the thoracic segment (Figure 3H,I,J), project a primary neurite in the dorsal-most region of the neuropil, the tectulum (Court et al., 2020) and join thoracic nerves (Pauls et al., 2018; Figure 3H,I,J, Figure 3—figure supplement 2 and 3). Unlike wild-type cells which bifurcate and branch extensively in the tectulum, the primary neurite of undead neurons fails to bifurcate, branches in both dorsal and ventral regions of the neuropil and sends projections to neighbouring segments (Figure 3J and Figure 3—figure supplement 3). As we describe in Figure 3—figure supplements 1 and 2, a few wild-type octopaminergic neurons are produced in all thoracic segments during postembryonic neurogenesis in lineage 0. We propose that the very few bilateral projecting neurons we encounter in our clones are wild-type cells (Figure 3—figure supplement 2E,F,G). To avoid any uncertainty when performing our behavioural experiments (below), we excluded flies which contained a bifurcating neuron in undead MARCM clones. Thus, using MARCM clonal approaches we show that undead neurons in hemilineage 0B become octopaminergic, elaborate complex neurites and join thoracic nerves.

We next asked if these differentiated undead neurons are functional. To address this, we tested if activating undead neurons with the warm temperature-gated ion channel TrpA1 in headless adult Drosophila could elicit behaviours (Figure 4 and Video 1). For this purpose, we deployed the same MARCM-based technique detailed above which ensured that only postembryonic octopaminergic neurons expressed CD8::GFP and TrpA1. The stochastic nature of MARCM allowed for generating both controls and flies with undead neurons in one mating cross using the same genotype, rearing and heat-shock conditions, i.e. there would be animals that would have experienced the heat- shock but have no octopaminergic neurons labelled. This further meant that behavioural experiments were performed blindly and each fly was matched to its control or undead neuron group only following dissection and imaging of the VNC. As mentioned above, we excluded flies with MARCM clones containing bilaterally symmetric neurons from our analysis, as these may be wild-type (see Figure 3—figure supplement 1 and Figure 3—figure supplement 2E,F,G). Additionally, we examined the effects of heat exposure in negative control flies with the genotype UAS-TrpA1 and the effects of heat-activation of wild-type octopaminergic neurons in positive controls expressing UAS-TrpA1 driven by TDC2-GAL4.

In TDC2-GAL4 positive controls (expressing TrpA1 in wild-type embryonic-born octopaminergic neurons), thermogenetic stimulation induced long bouts of locomotion (Figure 4C,D and Video 1) in 13/18 flies. Importantly, we found UAS-TrpA1 negative controls (i.e. an absence of a GAL4) and MARCM control flies (containing no GAL4-positive clones; see Figure 3—figure supplement 2A) did not walk in response to temperature elevation (Figure 4B,C,D and Video 1) (1/19 negative control, 2/22 MARCM control). We found that the activation of undead neurons expressing TrpA1 caused decapitated males to walk in 9/17 samples (Figure 4B,C,D,E and Video 1; also see Figure 3—figure supplement 2B,C,D for examples of MARCM clones of undead neurons) (negative control versus positive control, χ² = 17.6, p = 0.0002; negative control versus MARCM control, χ² = 0.2 p = 6; negative control versus MARCM undead neurons, χ² = 10.2, p = 0.0135; MARCM control versus positive control, χ² = 16.8, p = 0.0002; MARCM control versus MARCM undead neurons, χ² = 9.1, p = 0.0242; MARCM undead neurons versus positive control, χ² = 1.4, p = 1.4283; All comparisons were performed using Pearson’s chi-squared and p values were adjusted using a Bonferroni correction). A further analysis of the occurrence of walking after splitting the MARCM undead neuron group into the three anatomical subgroups T1, T2 and T3 according to the location of undead neurons in the pro-, meso- or metathoracic segment, yielded no significant differences: 4/6 T1, 3/4 T2 and 2/7 T3 (χ² = 2.9, p = 0.262855, Pearson’s chi-squared) (Figure 4E).

As previously reported, decapitated flies walked slowly by moving their limbs in a seemingly erratic manner, without having a tripod gait (Harris et al., 2015; Yellman et al., 1997). Flies were considered to be walking if they covered a distance of at least one body length during recordings and if they moved their legs in the order T3-T2-T1 at least once on each side, as evidence of intersegmental coordination (Strauss and Heisenberg, 1990). The direction of walking was either forward, sideways or backwards and most flies turned or walked in circles (Figure 4C), probably caused by variation in step size (Yellman et al., 1997; Harris et al., 2015).

Our data are consistent with the observation that octopamine applied to the exposed anterior notum of decapitated flies causes walking (Yellman et al., 1997) and suggests that undead neurons are functional and capable of releasing neurotransmitters in the CNS. The extent of walking was greater in positive control flies than in the undead MARCM condition (compare panels in Figure 4C and Video 1). This is likely because, alongside activating thoracic octopaminergic neurons in the VNC, in the positive controls we also stimulate the severed axons of octopaminergic cells in the brain which send descending projections to the VNC. These are not present in our TDC2-GAL4 MARCM flies.

To determine if undead neurons are integrated into thoracic motor circuits, we recorded the activity of mixed undead and wild-type octopaminergic neuron populations expressing GCaMP6s (an activity reporter) and tdTomato (an anatomical fiduciary) in intact H99/XR38 flies during tethered behaviour on a spherical treadmill (Chen et al., 2018; Figure 5A,B). The complexity of our calcium imaging experiment (see Figure 5A), together with a rate of success for obtaining MARCM clones with undead neurons of 15%, prompted us to approach our question whether undead neurons are functional during natural walking by using H99/XR38 flies. Keeping in mind that, instead of bifurcating, undead neurons collectively take a turn (see Figure 3I,J and Figure 3—figure supplement 2B,C,D), we interpret activity from the thickest bundle in the bifurcation as belonging to both undead and wild-type neurons. In these animals we observed conspicuous increases in neural activity during air-puff-induced walking in both wild-type controls (Figure 5—figure supplement 1) and in H99/XR38 flies (Figure 5C,D,E and Videos 2, 3 and 4). Because undead neurons outnumber their wild-type counterparts by a ratio of 6.5 to 1 (see Figure 3A,B,C,D), these results imply that both neuronal types are active in H99/XR38 flies. Supporting this, we observed an increase in GCaMP6s fluorescence across all subregions along the width of the thickest primary neurite bundle in the bifurcation, which contains all undead neurons together with three wild-type cells (Figure 5—figure supplement 2 and Video 4). Taken together, these data reveal that ‘undead’ neurons in the adult fly are functional and can integrate into motor networks.

Our observation that undead neurons functionally integrate into the CNS of adult flies strongly supports the possibility that PCD could be leveraged to modify neural circuits over the course of evolution. Alongside walking, octopaminergic neurons in the MNB lineage have a well-known function in flight (Roeder, 2005), and differences in neuron numbers correlate well with varying degrees of flight performance (Figure 6). Proficient fliers such as locusts and most flies have more octopaminergic neurons in winged thoracic segments (highlighted in yellow in Figure 6), while clumsy fliers such as cockroaches and crickets have similar numbers of neurons across thoracic segments. Figure 6 reviews our current knowledge of the number of GABAergic (hemilineage A) and octopaminergic (hemilineage B) neurons in the MNB of insects compiled from multiple studies spanning decades of research (see references in Figure 6). The lack of data for one population or the other is indicated with a question mark. Alongside our own data from swift louseflies (see Figure 7), here we also include our unpublished observations in the horse lousefly (for 0B) and the bee lousefly (for 0A). Having limited samples and antibody, we successfully labelled one preparation each and therefore resort to depicting these as a cartoon in Figure 6.

Because octopaminergic population size reflects flight ability across insects, and dipterans generally display larger populations in the mesothorax, we wondered if flies which have lost flight during evolution show reduced numbers of octopaminergic neurons in this segment. To this end, we described the MNB lineage in the flightless dipteran Crataerina pallida, the swift lousefly (Figure 7A), a viviparous haematophagous ectoparasite of the swift Apus apus (Bequaert, 1952; Hagan, 1951; Hutson, 1984; Walker and Rotherham, 2010a; Walker and Rotherham, 2010b). We labelled lousefly octopaminergic neurons from hemilineage 0B using antibodies for tyramine β-hydroxylase (Monastirioti et al., 1996) and by comparing the ratio of octopaminergic cells in the mesothorax (winged segment) and the prothorax (lacks wings), we found that, unlike flying dipterans, the swift lousefly has lost segment-specific variability of cell numbers (fruit fly [2.2 ± 0.2, n = 7] versus swift lousefly [1.1 ± 0.2, n = 3], p=0.012, Mann-Whitney U = 0, Mann- Whitney t-test) (Figure 7B,C,D,E,F). In the sister hemilineage 0A we found a considerably larger number of GABAergic neurons (Figure 7F), suggesting that PCD may be responsible for the selective elimination of the octopaminergic hemilineage.

We next wondered if we can find evidence of early PCD in hemilineages in the swift lousefly. Swift louseflies are viviparous, with only one progeny being produced and carried by the female at any one time (see Figure 7A, lower panel). Larvae hatch and remain inside the uterus for their entire larval life, feeding on lipid-rich secretions from milk glands until pupariation, when they are deposited in the swift nests and pupate. Adults only emerge the following spring, when swifts return from North Africa (Hagan, 1951; Bequaert, 1952). To capture postembryonic development, we dissected both larvae from inside female abdomens and pupae staged from Day 0 onwards, indicating days passed since laying (Figure 7G). Similar to the tsetse fly (Truman, 1990), neurodevelopment is significantly delayed compared to ‘typical’ dipteran flies - the nervous system only acquires dipteran larval-like features many days after pupariation (Figure 7H). Using EdU to label proliferating cells and immunostaining for active Dcp-1 (Figure 7H,I,J), we found dying cells located close to NBs throughout the 24 days of pupal neurogenesis (Pop et al., in preparation). In the MNB lineage, which is easily identified by its medial position and projection pattern in the neuropil, we found cell death in thoracic segments at all time points examined, from Day 4 after pupariation to Day 23 (Figure 7I,J).

Because cell death is present during neurogenesis in winged insects: selectively eliminating hemilineages in the fruit fly (Truman et al., 2010), killing off immature octopaminergic neurons produced by the MNB in the grasshopper (Jia and Siegler, 2002), appearing to sculpt neural networks in swift louseflies (see Figure 7H,I,J); we wondered if PCD also occurs during CNS development in a ‘primitive’ wingless insect. Using TUNEL labelling in the firebrat Thermobia domestica (Figure 8A,B), we found dying cells close to many NBs in all thoracic neuromeres at 50–55% of embryonic development (Figure 8C,D,E).

Similar to what we see in Drosophila (see Figure 2 and Figure 2—figure supplement 1), our observation that dying cells are found close to NBs in firebrats and louseflies suggests that this early PCD may be a universal and ancestral feature that sculpts the nervous system of all insects. To further explore if changes in PCD may have been deployed during evolution to accommodate adaptive modifications to behaviour, we next searched for evidence of increased PCD in other flight hemilineages of flightless dipterans.

To explore the possibility that changes in the pattern and/or extent of PCD are adaptive, we looked for evidence of evolutionary modifications in the VNC of yet another species, the bee lousefly Braula coeca (Figure 9A,B,C,D). Braula, a close relative of drosophilids, is wingless, lacks halteres and has an extremely reduced thorax (Figure 9A). Bee louseflies spend their entire adult life as kleptoparasites on the honeybee Apis mellifera (Imms, 1942; McAlister, 2018). We specifically asked whether lineages known to function in flight circuitry might be modified in flightless insects. In Drosophila, a thorough anatomical study recently described the pattern of innervation for each lineage into known functional domains of the adult neuropil and categorised them accordingly as being involved in leg, wing and both leg and wing control (Shepherd et al., 2019). In addition, the functional role of most hemilineages was previously assessed by thermogenetic activation in headless flies and those involved in flight-associated behaviours, such as wing wave, wing buzz or take-off, were identified (Harris et al., 2015). Together, these studies provide an excellent starting point for anatomical comparisons of homologous hemilineages which may have served an ancestral role in flight.

Using antibodies for Neuroglian, we compared homologous hemilineages involved in the flight circuits of the two flightless dipterans, the swift lousefly Crataerina pallida and the bee lousefly Braula coeca. Comparing the Neuroglian-labelled axon bundle width of homologous hemilineages between flightless and flying species, we found that hemilineages 3B, 7B, 11B and 12A, which produce wing waving, wing buzzing or take-off in the fruit fly (Harris et al., 2015), hemilineages 6A and 19B, which innervate the wing neuropil, and 5B, which innervates both leg and wing neuropils (Shepherd et al., 2019), are reduced in bee louseflies, but not in swift louseflies (Figure 9B,C,D). Among the latter, hemilineages 5B and 6A are known to contribute to leg movements and changes in posture, while the role for 19B is yet to be determined (Harris et al., 2015). Importantly, 3B, 6A, 11B, 12A and 19B belong to lineages in which both hemilineages survive in fruit flies (see schematic in Figure 1E). A difference in axon bundle diameter between sister hemilineages could indicate a difference in cell number, possibly established by PCD during development. We chose to quantify the ratio of axon bundle diameters in lineage three because we expected a reduction in hemilineage 3B, which innervates the wing neuropil, but no change in hemilineage 3A, which projects into the leg neuropil (Shepherd et al., 2019). The fibre tracts of 3A and 3B originate as a common bundle and split only in the intermediate neuropil. After they split, the sister fibre tracts sit in the same plane before they defasciculate, making them easy to trace and compare (Figure 9E,F,G). We found that the ratio between sister hemilineages A and B is significantly higher in bee louseflies compared to fruit flies, indicating that hemilineage 3B, which controls flight-related behaviours, is severely reduced in these flightless flies (T1: fruit fly [0.6 ± 0.1, n = 7] versus bee louse [1.8 ± 0.4, n = 8], p < 0.0001, t = −8.084, independent samples t-test; T2: fruit fly [0.7 ± 0.1, n = 7] versus bee louse [1.7 ± 0.4, n = 8], p=0.0001, F = 42.22, Welch’s t-test; T3: fruit fly [1.1 ± 0.4, n = 7] versus bee louse [1.7 ± 0.2, n = 8], p = 0.0044, F = 14.84, Welch’s t-test). Even though we cannot make a precise inference of cell numbers in each hemilineage, as the fibre tract of hemilineage 3B appears frayed while 3A is more compact in both species, our results clearly show that 3B is greatly reduced, associated with the loss of flight machinery.

To help us understand more about the patterning of PCD, we designed and used a new effector caspase probe SR4VH, which allowed us to interrogate the extent and dynamics of hemilineage-specific cell death. It shows us that an early onset PCD is responsible for the elimination of postembryonic neurons in the fly VNC, that this happens throughout the entire 3.5 days of postembryonic neurogenesis, and is hemilineage-specific. Although PCD has been reported as a fate within lineages in the embryo (Karcavich and Doe, 2005; Rogulja-Ortmann et al., 2007), the impact of this ‘early’ and hemilineage-specific PCD on the construction of the adult network has yet to be fully appreciated. This type of PCD is responsible for removing almost half of all postembryonic neurons that are born in the fly (Truman et al., 2010). Until now, the most frequently reported type of neuronal death described in insects has been the hormonally regulated PCD that removes mature neurons during the narrow developmental windows at the beginning of metamorphosis and within the first day after adult eclosion (Draizen et al., 1999; Lee et al., 2013). We now know from our data that these make up only a small fraction, compared to the total number of neuronal deaths in the fly.

Our SR4VH probe allows us to see that newly born neurons initiate cell death within the first 5.5 hr after birth. We can capture different stages of cell death: with young cells at very initial stages of PCD located closer to the NB, while older cells at more advanced stages of PCD with RFP labelled cell membranes found close to the lineage bundle. Importantly, death happens before neurites have extended, strongly suggesting that this PCD is not an analogue of neurotrophic death, found in vertebrates - where neuron-target interactions play a major role in the decisions of cell survival (Dekkers and Barde, 2013). In Thermobia domestica and Crataerina pallida, where the use of sophisticated genetic reporters such as SR4VH was not yet possible, EdU incorporation to mark dividing cells and immunolabelling for the active effector caspase Dcp-1 as a proxy for cell death revealed dying cells close to sites of division. Dying cells were found in the proximity of NBs (identifiable in all insects by their large size and position in the outermost layer of the CNS cortex) and far removed from mature neurons which congregate in the innermost layer of the cortex, adjacent to the neuropil. Therefore, we speculate that these cells are immature neurons and propose that they too undergo a death with rapid onset following division, similar to the early hemilineage-specific cell death we see with SR4VH in fruit flies.

The critical question that these data bring into focus is how early onset PCD is orchestrated, especially that an early intrinsically determined mode of cell death seems to be widespread across animals, from C. elegans to mice (Fricker et al., 2018; Southwell et al., 2012). Early PCD likely involves a combination of intrinsic patterning and cell-cell interactions between sibling neurons which ultimately deploy the activity of proapoptotic genes rpr, hid, grim and/or skl. Previously, patterning genes such as Ubx have been shown to contribute to the survival of hemilineages in the thoracic VNC in a parasegment-specific manner (Marin et al., 2012), while the transcription factor Unc-4 has been recently demonstrated to provide neural identity to hemilineages involved in flight (Lacin et al., 2020). Such spatial patterning is also required to establish NB identity which, in turn, can determine which hemilineage is maintained and which dies. In the developing optic lobe Bertet et al., 2014 have shown that the temporal sequence of transcription factors expressed in NBs (inherited by the GMC and newly born neurons) produces a switch in the selective survival of one hemilineage over the other. In this manner, the changes to NB identity which have been documented in other insects (Biffar and Stollewerk, 2014), despite a conserved NB array and progeny, could in turn influence the pattern of PCD. Alongside, cell-cell interactions between newly born neurons could influence fate choices. The requirement of interactions between newly born siblings in determining asymmetric fates has been shown in the grasshopper VNC (Doe et al., 1985; Kuwada and Goodman, 1985), although this exact same mechanism has not been demonstrated in Drosophila.

The spatio-temporal pattern of death revealed with SR4VH shows us that understanding the molecular control of hemilineage-based death is the key question going forward and is likely to provide insight into how networks evolve.

Here we blocked PCD within the MNB lineage (called lineage 0 in the postembryonic literature) and found that ‘doomed’ neurons become octopaminergic, generate arborisations and target the tectulum neuropil (Court et al., 2020). Previous work has shown that octopamine can induce and maintain walking in locusts (Sombati and Hoyle, 1984) and decapitated fruit flies (Yellman et al., 1997). Consistently, we find that our ‘undead’ hemilineage 0B cells can induce walking when activated thermogenetically in headless flies and show calcium activity during naturalistic bouts of locomotion. Thus, by blocking death, we have ‘resurrected’ functional neurons that are able to integrate into thoracic motor networks. Although this ‘dialling up’ of cell numbers, in our system, is artificial, it reveals how doomed neurons possess cryptic cellular phenotypes that can emerge when death is blocked, advocating for the evolvability of such a hemilineage-based system.

While these undead hemilineage 0B octopaminergic neurons share many conserved features with wild-type cells, the variations we see in their morphology could act as a substrate for evolutionary change. A recent study has linked structural changes in the VNC with changes in behaviour between strains of Drosophila melanogaster (Mellert et al., 2016). Mellert et al., show that hemilineage 12A in the mesothorax has variable bundle morphologies and that these correlate well with the time of flight initiation. Flight is used as an escape response and can be instrumental for predator evasion, one of the major evolutionary forces which have selected for flight in insects in the first place (Dudley, 2002). With this in mind, it seems plausible that changes in either neuron number and/or innovations in ‘undead’ neuron structure could affect adult behaviour and be ultimately adaptive. Recently it has been shown that undead sensory neurons that are functional, integrate and appear to be tuned to specific odours (Prieto-Godino et al., 2020). Importantly, Prieto-Godino et al. also show that there is cell number variation in this neuronal population across drosophilids and that blocking PCD in melanogaster results in the survival of mosquito-like CO₂-sensing neurons in the maxillary palps. This suggests that both the central and peripheral nervous system may use similar modes of early PCD to sculpt circuits during the evolution of true flies.

Our data show that PCD is extensive and widespread during neurogenesis in the CNS of insects, from the primitive firebrats Thermobia domestica, to true flies Drosophila melanogaster and the swift lousefly Crataerina pallida.

We wondered whether alterations in cell death may have contributed to adaptations in the VNC of flightless dipterans and found that a greater extent of PCD in the MNB lineage may be responsible for abolishing the segment-specific difference in octopaminergic cell numbers in the swift lousefly Crataerina pallida – which has lost flight and gained adaptations to a parasitic lifestyle (Bequaert, 1952; Hagan, 1951; Hutson, 1984; Lehane, 2008; Petersen et al., 2018; Walker and Rotherham, 2010b). As octopaminergic neurons are involved in flight-related behaviours (Brembs et al., 2007; Duch and Pflüger, 1999; Roeder, 2005), we suggest that PCD has been co-opted in this lineage during the evolution of flightlessness in the swift louse. We believe that this PCD takes place early, in newly born neurons, as we have seen dying cells close to the MNB in all of our 20 pupae, from Day 4 to Day 23. Therefore, the decrease in octopaminergic cell number in the mesothoracic segment of swift louseflies is likely the result of increased hemilineage-specific PCD during evolution. Alternatively, an increase in the number of neurons in the non-flying prothorax and metathorax could lead to a uniform population size across thoracic segments in swift louseflies. This could be achieved either by additional MNB divisions, or by reduced PCD. If changes in MNB proliferation play a role remains to be determined.

The midline neurons within the VNC of insects have long been a source of interest because they are homologous across species, yet show a diversity in cell numbers correlated with body form and function (Lacin et al., 2019; Stevenson and Spörhase-Eichmann, 1995; Witten and Truman, 1998). For example, flying insects have greater numbers of midline octopaminergic neurons within segments that control wings (Campbell et al., 1995; Eckert et al., 1992; Jia and Siegler, 2002; Konings et al., 1988; Monastirioti et al., 1995; Schlurmann and Hausen, 2003; Siegler and Pankhaniya, 1997; Siegler et al., 2001; Siegler et al., 1991; Spörhase-Eichmann et al., 1992; Stevenson et al., 1992; Thompson and Siegler, 1993), while grasshoppers have more GABAergic neurons in the metathoracic/abdominal ganglia, which receives auditory input (Witten and Truman, 1998; Thompson and Siegler, 1991). These two midline neuronal populations are derived from the same NB, the MNB, which buds off multiple GMCs, each dividing once to generate a GABAergic (A cell) and an octopaminergic (B cell) neuron. Both A and B neurons are generated in equal numbers but, in all cases, the numbers of GABAergic (‘hemilineage A’) and octopaminergic (‘hemilineage B’) neurons within one segment are never the same. The greater numbers of GABAergic cells within each segment has been shown in grasshoppers (Jia and Siegler, 2002) and fruit flies (Truman et al., 2010) to be the result of removal by PCD of large numbers of cells from hemilineage B.

As suggested by our previous work, the ‘hemilineage’ emerges as a discrete developmental unit that shows common features of gene expression and function (Harris et al., 2015; Lacin et al., 2019; Lacin and Truman, 2016; Shepherd et al., 2019; Truman et al., 2010). Therefore, we refer here to the PCD found in the MNB lineage of grasshoppers as ‘hemilineage-specific PCD’. Following our observations of PCD during development in other insects, together with a vast body of knowledge on homology in insect nervous systems (Kutsch and Breidbach, 1994; Thomas et al., 1984), we suggest that variations in neural circuits between species is very likely set up by modifying hemilineages, with PCD playing a major role.

In the wingless bee lousefly Braula coeca, we found clear reductions in the thickness of fibre tracts in several hemilineage bundles which in Drosophila are associated with flight-related behaviours (Harris et al., 2015). It remains for us to determine if early PCD takes place in these specific bee lice lineages during development and causes the reduction in bundle diameter that we see in flight hemilineages. As with most parasitic insects, bee and swift louseflies are impossible to maintain in the laboratory in the absence of their hosts and procuring them is a challenge (e.g. bee louseflies are now only found on two islands in the UK, while collecting swift louseflies is restricted to the summer months due to Apus apus migrations). Nonetheless, our observations that PCD is widespread across insects complements our findings in bee louseflies, strongly suggesting that an extensive PCD in flight hemilineages may have accompanied the loss of flight during evolution. Interestingly, the reduction in fibre diameter we see in bee louseflies was not evident in swift louseflies. This difference is likely due to the more significant changes to body plan in bee louseflies, i.e. a complete loss of the flight apparatus during evolution. Swift louseflies however still maintain vestigial wings and halteres (Walker and Rotherham, 2010b), whereas bee louseflies have a severely reduced thorax, completely lacking wings, halteres and flight muscles (McAlister, 2018).

Here we have shown that undead neurons elaborate complex arborisations, express distinct transmitter identities and function. We find that ‘early’ PCD is widespread during the development of the CNS of insects from the primitive firebrats, to most derived true flies. Early cell death appears to be a specific subtype of PCD present across animals. Understanding how early PCD is specified across species should help us elucidate how nervous systems are built and evolve. Our exploration of homologous lineages in flightless dipterans shows that changes in body plan may accompany changes in the extent and pattern of PCD. As the evolutionary changes seen in neural networks ultimately result from heritable differences in developmental processes, our future endeavours will be directed towards elucidating how genetic programs are deployed to establish the pattern of PCD. The cellular leitmotif of hemilineage-based cell death, we present here, provides us with something tangible that we can search for. Thus, we suggest that viewing the evolution of insect nervous systems through the lens of the ‘hemilineage’ will be critical for understanding how development brings about adaptive changes in neural network motifs.

We used the following Drosophila melanogaster stocks: Worniu-GAL4; Dr/TM3, Ubx-LacZ, Sb (BDSC_56553), TDC2-GAL4 (BDSC_9313), OK371-GAL4 (BDSC_26160), UAS-SR4VH (described here), UAS-CD8::GFP (BDSC_5137), UAS-tdTomato-p2A-GCaMP6s (Chen et al., 2018) (kind gift from M. Dickinson), H99/TM3, Sb (BDSC_1576), XR38/TM3 (Peterson et al., 2002), Sb, If/CyO; dronc^∆A8, FRT2A/TM6β, Tb, Hu (Kondo et al., 2006) and hs-flp;; TubP-GAL80, FRT2A/TM3, Sb (Truman et al., 2010).

Firebrat adults of Thermobia domestica were obtained from Buzzard Reptile and Aquatics (buzzardreptile.co.uk) and reared on a diet of fish flakes and wholemeal bran at 40°C in darkness inside a humid plastic container. A staging series was calculated by time to hatching.

Crataerina pallida swift lousefly adults were collected from swift (Apus apus) nesting boxes fitted behind the louvres of belfry windows from churches in Cambridgeshire and Suffolk (UK) with the help of local conservationists Simon Evans, Richard Newell and Bill Murrells. Swift louseflies were kept at 20°C on a 12 hr dark:12 hr light cycle until dissected. Pregnant females, recognised by their enlarged and translucent abdomen through which larvae or prepupae could be detected, were kept separately and checked daily for pupa ejection. The day in which a pupa was laid was defined as ‘Day 0’ of external development (outside the mother’s abdomen).

Braula coeca bee lousefly adults were obtained from a black bee (Apis mellifera mellifera) colony on the Isle of Colonsay, UK (kind gift from A. Abrahams). Bee louseflies were shipped by post in small cages containing worker bees feeding on bee fondant. The black bees and bee louseflies were anaesthetised by placing the cage on a CO₂ pad and the bee louseflies were removed for dissection.

SR4VH was constructed by standard molecular biology procedures. It comprises the myristylation signal of Drosophila Src64B (amino acids 1–95), a monomeric red fluorescent protein mRFP1 (Campbell et al., 2002), a linker that contains four DEVD sites, a yellow florescent protein Venus (Nagai et al., 2002), and a nuclear localisation signal of Drosophila histone H2B (amino acids 1–51). While the design is similar to the previously reported caspase probe Apoliner (Bardet et al., 2008), the Src64B myristylation signal and the H2B NLS offers better membrane and nuclear localisation, respectively, and four DEVD sites are expected to provide higher sensitivity. The probe was cloned in pUAST (Brand and Perrimon, 1993) and introduced into the Drosophila genome by P element-mediated transformation.

Drosophila larvae were dissected in PBS without anaesthesia. Firebrat embryos were removed from their chorion and dissected using minuten pins. Drosophila, swift lousefly and bee lousefly adults were anaesthetised on ice, briefly submerged in absolute ethanol and dissected in PBS. Swift lousefly pupae were immobilised on double sided sticky tape, removed from their pupal case using forceps and dissected in PBS without anaesthesia. Samples were fixed in 3.6% paraformaldehyde in PBS for 30 min (larvae and pupae) or 1 hr (adults), washed 3 times in 0.3% PBST (0.3% Triton-X100 in PBS, Sigma-Aldrich), blocked in 5% goat serum (Sigma-Aldrich) in PBST for 1 hr and incubated with primary antibodies in block for 1–3 days at 4°C (Drosophila, bee louseflies, swift lousefly pupae), room temperature (firebrats) or 37°C to increase antibody penetration (swift lousefly adults; block supplemented with 0.02% NaN₃ to prevent microbial growth). Samples were then washed four times throughout the day in PBST and incubated with secondary antibodies in block for a further 1–3 days, followed by final washes in PBST and PBS. Brains and VNCs were mounted on poly-L-lysine-coated coverslip, dehydrated in increasing serial concentrations of ethanol (15%, 30%, 70%, 80%, 90% and twice in 100%) for 5 min each, dipped once in xylene, then incubated twice for 5 min in fresh xylene. A droplet of DePeX (EMS) was added on top of the mounted sample and the coverslip was placed face-down on a glass slide.

We used the following primary antibodies: chicken anti-GFP (1:500; ab13970, Abcam), mouse anti-Neuroglian (1:50; BP 104, Developmental Studies Hybridoma Bank), rabbit anti-cleaved Drosophila Dcp1 (1:100; 9578, Cell Signaling), guinea pig anti-Syncrip (1:100; kind gift from I. Davis; to label NBs and early progeny in lineages - JW Truman, personal communication, January 2019), mouse anti-Engrailed/Invected (1:2; 4D9, Developmental Studies Hybridoma Bank), rabbit anti-DVGLUT C-terminus (Mahr and Aberle, 2006) (1:5000; AB_2490071, kind gift from H. Aberle), rat anti-tyramine β-hydroxylase (Monastirioti et al., 1996) (1:50; AB_2315520, kind gift from M. Monastirioti), rabbit anti-vestigial (1:400; kind gift from Sean Carroll and Kirsten Gruss) and rabbit anti-GABA (1:100; 20094, ImunoStar).

Secondary antibodies were Alexa Fluor 488-conjugated goat anti-chicken (1:500; A11039, Invitrogen, Thermo Fisher Scientific), Alexa Fluor 488-conjugated goat anti-rabbit (1:500; A11070, Invitrogen, Thermo Fisher Scientific), Cy3-conjugated donkey anti-rabbit (1:500; 711-006-152, Jackson ImmunoResearch), Cy5-conjugated donkey anti-mouse (1:500; 715-006-151, Jackson ImmunoResearch), Alexa Fluor 488-conjugated donkey anti-rat cross-adsorbed against mouse (1:100; 712-545-153, Jackson ImmunoResearch), Alexa Fluor 488-conjugated donkey anti-guinea pig (1:500; A11073, Invitrogen, Thermo Fisher Scientific).

In firebrat embryos we detected dying cells using the Click-iT Plus TUNEL assay kit (C10618, life technologies). To stain cell nuclei and neuropil, firebrat samples were incubated with DAPI (1 µg/mL; D9542, Sigma-Aldrich) and Phalloidin-488 (1:100, A12379, life technologies) in PBST for 30 min at room temperature. Incubations were carried out following secondary antibody treatment. Samples were then washed in PBST and PBS, and mounted.

To label proliferating cells and their progeny we used the Click-iT EdU imaging Kit (C10337, life technologies). Freshly dissected nervous systems from swift lousefly pupae were incubated in EdU 1:1000 in PBS at room temperature for 1–3 hr on a shaker, rinsed with PBS and fixed in cold buffered formaldehyde 3.6% in PBS for 30 min. Samples were then stained using the immunohistochemistry protocol described above. The colour reaction for EdU was carried out as instructed by the vendor after the secondary antibodies were washed out.

To induce mitotic clones of undead neurons, rescued from PCD, we used the mosaic analysis with a repressible cell marker technique (Lee and Luo, 1999). 0–4 hr first instar larvae resulting from crossing females of the genotype hs-flp;; TubP-GAL80, FRT2A/TM3, Sb with; TDC2-GAL4, UAS-CD8::GFP, UAS-TrpA1; dronc^∆A8, FRT2A/TM6β, Tb, Hu males were heat-shocked at 37°C in a plastic food vial placed in a water bath for either 1 hr or 45 min, followed by 45 min at room temperature and a second incubation period at 37°C for 30 min. After heat-shock, larvae were immediately returned to 23°C or 25°C. Cell death was blocked in clones homozygous for the loss-of-function allele of the initiator caspase Dronc. Because we used the octopaminergic driver line TDC2-GAL4 to induce the expression of CD8::GFP and TrpA1, we were able to visualise and thermogenetically activate only postembryonic neurons of hemilineage 0B. A small number of wild-type octopaminergic neurons are born during postembryonic neurogenesis (one in T1 and T3, 4–5 in T2, see Figure 3—figure supplement 1). To ensure the characterisation of undead neurons only, MARCM clones including a bilaterally symmetrical primary neurite were excluded from analysis.

Prior to recordings, 2–6 day-old males of Drosophila melanogaster (hs-flp/+; TDC2-GAL4, UAS-CD8::GFP, UAS-dTRPA1/+; dronc^∆A8, FRT2A/TubP-GAL80, FRT2A) reared at 23°C or 25°C in a 12 hr:12 hr light:dark cycle were anaesthetised on ice and decapitated using a pair of micro spring scissors in under 3 min. We used males as we found they are more responsive to octopamine release by thermogenetic activation than females (data not shown). The headless flies were brushed back into a food vial placed on its side and left to recover for at least 1 hr. To generate the heat ramp required to thermogenetically activate undead neurons, we used a 12V thermoelectric Peltier plate (model: TEC1-12706, size: 40 mm x 40 mm x 3.6 mm) connected to a DC power supply (HY3005D, Rapid Electronics) set at a constant current of 0.46A, with a variable voltage, calibrated using an infrared laser thermometer (N92FX, Maplin). These settings generated a temperature ramp which lasted 70 s from 22°C to 34°C. Videos were recorded at 25 fps using a Sony NEX-5N digital camera (kindly provided by Ian Wynne) mounted to a stereo microscope. A piece of graph paper was used for spatial calibration. To match the presence of undead neurons with behaviour, each decapitated fly used for thermogenetic activation was indexed and prepared for dissection and immunostaining.

The method for in vivo two-photon imaging of the VNC in behaving adult Drosophila is described in Chen et al., 2018. Briefly, flies were anaesthetised through cooling and then mounted onto custom imaging stages. The dorsal thoracic cuticle was removed and indirect flight muscles were left to degrade over the course of 1 hr. Subsequently, the proventriculus and salivary glands were resected to gain optical access to the VNC.

Horizontal sections of the T1 leg ganglion were imaged using galvo-galvo scanning. For control animals, the bifurcation point of TDC-positive neurites were imaged to circumvent ROI disappearances caused by movement. For animals harbouring undead TDC2-GAL4-positive neurons, the thickest branch of the axonal bifurcation was chosen because they were most likely to contain undead neurites. Image dimensions ranged between 512 × 512 and 320 × 320, resulting in 1.6 to 3.4 fps data acquisition. Imaging areas ranged between 92 × 92 µm and 149 × 149 µm. Laser power was held at ~8 mW.

Python scripts (modified from Chen et al., 2018) were used to extract ROI fluorescence traces and to compute spherical treadmill ball rotations. Walking epochs were determined by placing a threshold on ball rotations, which were first converted into anterior-posterior (v_forward) and medial-lateral (v_side) speeds (one rot s⁻¹ = 31.42 mm s⁻¹) and into degrees s⁻¹ (1 rot s⁻¹ = 360° s⁻¹) for yaw (v_rotation) movements. Thresholds were 0.12 mm, 0.12 mm and 5°, respectively. Periods below these thresholds were considered ‘resting’ while other periods were considered ‘walking’. Fluorescence traces for epochs with the same behaviour were aligned by start point to compute average %∆R/R traces for specific actions.

To calculate fluorescence traces for small subregions-of-interest across neuritic bundles containing both undead and wild-type neurites, images were registered using an optic flow method described in Chen et al., 2018. This registration served to minimise motion artefacts. Analysis was limited to a period with no warping artefacts and no ROI disappearance. Subregions were manually selected as small circular ROIs across the neuritic bundle of the registered image. Fluorescence values were then computed from each sub-ROI.

Images were acquired using a Zeiss LSM 510 or a Zeiss LSM 800 confocal microscope at a magnification of 20x or 40x with optical sections taken at 1 µm intervals. The resulting images were examined and processed using Fiji (https://imagej.net/Fiji). Some images were manually cropped using the Freehand Selection tool to remove debris or to cut out neuronal lineages in Worniu-GAL4, UAS-SR4VH samples.

To generate fluorescence intensity along Line plots, we used the Plot Profile tool in Fiji to extract raw fluorescence intensity values for the RFP and Venus channels. The values were imported into MATLAB (R2018a, MathWorks) and normalised by dividing all fluorescence intensity values to the maximum value encountered along each Line. In this manner, all fluorescence intensity along Line plots have a common scale from 0 to 1, with one being the highest value encountered along that Line.

Decapitated flies were considered to be walking if they covered a distance greater than one body length and moved their legs in a coordinated sequence from T3 to T2 to T1 at least once on each side (Harris et al., 2015). Forward, backward and sideway movements were all interpreted as walking when both aforementioned conditions were respected. To generate fly body traces video recordings were imported in MATLAB (R2018a, MathWorks) and the centroid of the decapitated fly (located on the scutellum) was extracted from each frame using a custom-written script which can be found at github.com/snznpp/undead-walking (Pop, 2020; copy archived at https://github.com/elifesciences-publications/undead-walking). Each frame was converted into a greyscale image, its contrast enhanced using contrast-limited adaptive histogram equalisation, filtered using a Gaussian smoothing kernel with a standard deviation of 4, binarised using a custom threshold and the geometric centre of the fly body automatically extracted and stored in an array. To confirm that the centroid detection was accurate, a red dot with the centroid coordinates was superimposed onto each frame of the original recording and the annotated movie was saved for manual inspection.

For calculating 3A/3B hemilineage bundle diameter ratios in fruit flies and bee louseflies, we generated transverse rendered maximum intensity projections of inverted greyscale confocal stacks for the pro- and mesothorax (T1 and T2) and frontal projections for the metathorax (T3). Optical sections were selected to include the common lineage bundle and the individual hemilineage bundles after their split. Diameter measurements were taken at the widest point within 5 µm of the bundle split using the Straight Line tool in Fiji and ratios were calculated by dividing the diameter of hemilineage 3A to that of 3B.

For comparing neuron numbers, 3A/3B bundle diameter and T2/T1 number of neurons, data were tested for normal distribution using the Kolmogorov-Smirnov test and visualisation of Normal Q-Q plots. Differences between groups were analysed using either the independent samples t-test for normally distributed data, Welch’s test if data failed to meet the homogeneity of variances assumption or Mann-Whitney t-tests if data failed to meet the normality and homogeneity of variances assumptions of the independent samples t-test.

For comparing the number of flies which walked in each experimental group, we performed a Pearson chi-squared test and interpreted the resulting exact significance if the minimum expected count was greater than 5, or the Fisher’s Exact Test 2-sided significance if the minimum expected count was lower than five in at least one cell of the contingency table. To correct for multiple comparisons we performed a Bonferroni correction (i.e. p values were multiplied by 6, the total number of pairwise tests).

All statistical tests were performed in SPSS Statistics 23 (IBM) with an α set at 0.05. In all figures, bars represent means ± standard deviation; *p < 0.05, ***p < 0.001, ^nsp = not significant.