The Apollo implant consists of two parts, the payload and the docking modules, inspired by previous designs (van Daal et al., 2021; Figure 1). Both parts can be 3D-printed in a variety of materials, although we typically used a combination of Nylon PA12 and Formlabs Rigid Resin. The Neuropixels 1.0, 2.0α, and 2.0 implants weigh 1.7, 1.3, and 0.9 g (Table 1). Payload modules can accommodate up to two parallel probes, with the second probe adding a further 0.4, 0.2, and 0.2 g. The Apollo implant is therefore 40% lighter than the only published Neuropixels 2.0α solution, which recommends animals are at least 25 g (van Daal et al., 2021; Steinmetz et al., 2021). This allows for use in smaller animals (e.g. female and water-restricted mice). The electronics are protected by lids with slots to accommodate the flex cables when not connected to the headstage (Figure 1A). To ensure the implant is maximally compact, flex cables can be folded into the cavity beneath the lids (Figure 1B). This minimizes implant height (29, 21, and 17 mm for a Neuropixels 1.0, 2.0α, and 2.0), reducing the moment of inertia above the head. The implant can be 3D-printed for $10 ($3 for each disposable module).

The implant is flexible and recoverable, allowing for different configurations, and the same Neuropixels probe(s) to be used multiple times (Figure 1C). Once the payload module is constructed, the distance between the two probes remains fixed. The docking module is connected to the payload module via small screws, which makes it easy to assemble, and disassemble upon explantation. Only the docking module is cemented to the animal’s skull, and it is covered with merlons to increase contact with the cement and therefore the stability of the implant. To facilitate different implantation depths and angles with the same payload module, both the length and base-angle of the docking module can be adjusted. The base of the implant thus remains parallel to the skull (Figure 1C) which improves stability and reduces weight by minimizing implant height and the quantity of cement required. All adjustments can be achieved by inexpert CAD users with preset parameters supplied in the editable files to change distance between probes (1.8–6.5 mm—beyond 6.5 mm, two implants can be used), implantation depth (2–6.5 mm), or angle of insertion (up to 20 degrees) (Video 1). As the fully editable files are provided, users can (and have) adjusted the implants to exceed these default boundaries, or create their own custom modifications which are also available online (see Methods).

To help combine the payload and docking modules, we designed a dedicated constructor (Figure 1D). The docking holder, containing a new docking module, slides onto the constructor posts, and the payload holder, containing the payload module, is fixed to the end. The two modules are thus coaxial, and the docking module can slide into position and be secured to the payload module without risk of damaging the probes. The constructor comprises 3D-printed parts and Thorlabs 6 mm poles for a one-time cost of $25.

A comprehensive protocol for assembly and implantation, including variations employed across labs, is provided in Methods. Payload modules are assembled with one or two Neuropixels probes. After probe-sharpening (see Methods, Figure 2—figure supplement 1A), an empty payload module was positioned on adhesive putty and coated with a thin layer of epoxy. The probe(s) can then be affixed and aligned to the payload module, either by eye or using graph paper, before covering the base and electronics with epoxy or dental cement (Figure 2A). The flex cable was folded and inserted into the lid and lids were then glued to the payload module (Figure 2B).

The payload module (new or previously used) was combined with a new docking module for each experiment. Docking modules were adjusted to match experimental requirements (e.g. insertion depth, angle, etc.). The docking module was secured in its holder and slid onto the arms of the constructor. The payload module was secured in its holder and attached to the end of the constructor (Figure 2C, D). The docking module holder was then slid along the constructor arms, and the two modules were secured with screws (Figure 2D). Before each experiment, any gaps in the assembled implant were filled (Figure 2E). Prior to each implantation, probes were typically coated with fluorescent dye for post-experiment trajectory tracking (Figure 2—figure supplement 1B).

Craniotomies were performed on the same day as the implantation, but this could be any time after assembly (Figure 2F). The implant was held using the 3D-printed payload holder and positioned using a micromanipulator. The eight shanks (in the case of a dual 4-shank implant) are positioned at the surface of the brain (Figure 2G). Care is taken to avoid large blood vessels, and the implant can be rotated and repositioned. If the vessel is not completely avoidable, the shanks can be positioned on each side of the blood vessel. Probes were inserted to the desired depth at a slow speed (3–5 µm/s). Finally, to complete the implantation, the docking module was cemented to the skull (Figure 2H).

Explantations were performed with a payload holder attached to a micromanipulator. The holder was aligned to the payload module, slid into place, and secured with a screw. The screws between the payload and the docking modules were then removed, and the payload module extracted (Figure 2—figure supplement 1B). Probes were cleaned with a Tergazyme solution, occasionally followed by a silicone cleaning solvent if Dural-Gel stuck to the probe. The payload module was combined with a new docking module for subsequent experiments.

Across laboratories, 97% of probes were recovered without any broken shanks (61/63 explanted probes, Supplementary file 1). In only two cases were probes damaged, and in one of those cases the skull integrity was compromised by infection (a rare occurrence) and the probe was likely broken before explantation. On six further occasions, probes stopped working due to connection errors (typically revealed by a ‘shift register’ error in SpikeGLX). The recovery rate is therefore 86% when including all connection errors. However, as this type of error is also observed with acute probe use, and there was no observable damage to the chronic probes, these failures may reflect long-term wear rather than any issue with the implant. Consistent with this, probes that failed with this error had typically been used for several months (Supplementary file 1). Outside of the originating laboratory (UCL), 95% of probes (19/20) were recovered without any broken shanks (90% when including all connection errors) demonstrating the ease with which new users adopt this design.

We tested the stability of the Apollo implant with Neuropixels 1.0, 2.0α, and 2.0 probes (Figure 3). We implanted 48 mice using 4-shank Neuropixels 2.0α implants (20 mice with a single-probe implant and 13 mice with a dual implant), single-probe Neuropixels 2.0 implants (7 mice), and Neuropixels 1.0 implants (7 mice with a single-probe implant and one mouse with a dual), as well as 3 rats with a single Neuropixels 1.0 implant (Supplementary file 1). In many cases, the same implants were reused (up to six times) and remained fully functional across different animals (Supplementary file 1). Recordings were performed over a period of days to months. The probes were typically inserted 5–6 mm inside the brain, traversing multiple brain regions (Figure 3A, B). Because only 384 of the 5120 channels, termed a ‘bank’ of channels, can be recorded simultaneously on each 4-shank 2.0 probe, multiple recording sessions were often used to cover all recording sites located in the brain. Compared with acute recordings, this strategy dramatically reduces experimental setup time and complexity, and is especially beneficial for whole brain approaches. The raw signal quality did not seem to change across days (Figure 3C, D), allowing us to identify single spikes reliably for months. The spiking patterns on each probe were similar across days (Figure 3E, F), suggesting that the same populations of neurons were being tracked.

The number of recorded neurons was reasonably stable across weeks (Figure 4, Figure 4—figure supplements 1–3). For each session, we quantified the number of well-isolated single units for each individual channel bank (Figure 4A). Units were selected based on stringent criteria including amplitude, percentage of missing spikes, and refractory period violations (Fabre et al., 2023; van Beest et al., 2024). The number of single units for each probe is the sum of units across all banks within the brain (Figure 4A). Unit numbers could remain stable for more than 50 days, and we observed comparable stability in most of mice (Figure 4B). As previously described (Luo et al., 2020), we often observed an initial fast decrease in the number of units, but this was not systematic. Indeed, in some animals, unit number increased slowly across days until reaching a peak. The mean decrease in unit count per day was 3% (median 2%), within the range previously observed for chronic Neuropixels implants (Steinmetz et al., 2021). Although implants with more rapid unit loss were not suited for long-term recordings, others remained stable for months. Across all banks, the average number of recorded neurons on each bank was 85 ± 6 during the first 10 days (n = 59 probes, mean ± SEM), 65 ± 7 during days 10–50 (n = 50 probes), 54 ± 15 during days 50–100 (n = 6 probes), and 44 ± 25 beyond (n = 2 probes) (Figure 4—figure supplement 2). The initial number of units did not depend on the number of times the probe was reimplanted (p > 0.25, linear mixed-effects model, Figure 4C) or the insertion coordinates of the probe (p > 0.94, linear mixed-effects model, Figure 4D). The rate of unit loss was also independent of these two variables (p > 0.29 and p > 0.31 for probe reuse and AP position, linear mixed-effects model, Figure 4E, F). However, implant quality was more variable in posterior brain regions, with instances of rapidly decreasing neuron counts, as previously described (Luo et al., 2020). Stability was qualitatively similar across different laboratories (Figure 4—figure supplement 1). Surgical optimizations are ongoing, and protecting the craniotomy with silicon may significantly increase recording stability (Figure 4—figure supplement 3, Melin et al., 2024).

The overall quality of the signal remained high throughout days and probe reuses. We quantified the overall noise present in the recordings by computing the root-mean-square (RMS) value of the raw signal (Figure 4G–L). The RMS values were stable across days, across all mice (Figure 4G, H). Both the average RMS value and its changes over time were independent of the number of times the probe had been used (Figure 4I, K). We observed a significant effect of AP position on the RMS value, but not on its changes over time (Figure 4J, L). Similarly, the median unit amplitude was stable and unaffected by probe reuse (Figure 4—figure supplement 4).

Individual neurons could be tracked across days and months (Figure 5). We used the tracking software UnitMatch to track the same units across days, based on their waveforms (van Beest et al., 2024). In a mouse recorded for 100 days, a significant fraction of units could be tracked for months (Figure 5A, B). Tracked neurons had stable waveforms over days, as expected from the matching procedure, but also stable inter-spike intervals histograms (ISIHs) (Figure 5C, D). These ISIHs were not used to match neurons across days, and their stability therefore strongly suggests the same units were tracked over months. The proportion of units tracked between two recordings decreased as a function of time between the recordings (Figure 5E). For two recordings on the same day, 50% of neurons were matched, suggesting an upper limit in neuron tracking with the method we used, likely due to variability in neural activity or conservative choices in software parameters. The proportion of tracked neurons typically decreased to 10% after 32 days, but the rate of decay varied across recordings: with some cases where 20% of neurons were tracked for 64 days, and others dropping to 0% after a week. In all cases, tracked units had consistent ISIHs, as measured by the area under the receiver operating characteristic curve (AUC)—comparing the similarity of ISIHs for tracked versus different units (Figure 5F). This indicates the tracking algorithm remains accurate for large intervals between recordings.

To test whether the Apollo implant could be used in more naturalistic conditions, we recorded from freely behaving mice and rats in various configurations (Figure 6). First, to minimize the weight, we recorded from two freely moving mice using either a dual Neuropixels 2.0α implant or a Neuropixels 1.0 implant, with the headstage suspended by its connection cable (Figure 6A, B). The mice explored their home cage and exhibited normal behaviors, such as grooming, running, and sleeping, suggesting that the implant did not impair basic movements. The recordings yielded high-quality, well-isolated single units for weeks (Figure 6C, Figure 4—figure supplement 1). The distributions of the RMS values (Figure 6D) and spike amplitudes (Figure 6E) were similar to the head-fixed conditions, suggesting an equivalent quality of recording despite differences in conditions, and labs. It can also be more convenient to secure the position of the headstage in each recording, or permanently attach the headstage to the implant. We thus designed a headstage holder, which we tested with Neuropixels 1.0 (Figure 6F–J). To further reduce the weight on the mouse, we also designed a 1-probe version of the implant for Neuropixels 2.0, with a minimal headstage holder (Figure 6K–O), inserted at an angle (16 or 25 degrees), at the back of the brain. In rats (Figure 6P–T), the implant was inserted in the center of a 3D casing, that afforded extra protection, and neural data was recorded wirelessly using SpikeGadgets.

To quantify the effect of implantation on behavior, we compared the performance of mice on a complex behavioral task before and after implantation with a Neuropixels 1.0 probe—the heaviest version of the Apollo Implant (Table 1). This implant was modified to allow the headstage to be permanently attached from the first recording session (see Methods). Mice were placed in a large octagonal arena (80 cm diameter). On each trial, mice were required to respond to visual stimuli projected onto the floor of the chamber and perform a nose poke in one of the ports located around the perimeter of the chamber. Thus, during a typical session of 100 trials, mice typically traversed tens of meters (Figure 7A). We compared mouse hit rate, trial number, and reaction times in sessions before implantation (when the mouse moved entirely freely) and after implantation (when the mouse was tethered). After implantation, mice continued to perform trials and fully explore the chamber (Figure 7B). We observed an initial reduction in hit rate and trial number, and an increase in reaction time immediately after implantation (Figure 7C–E). The first two measures recovered within five sessions (Figure 7C, D), but reaction times did not recover to pre-implantation levels, indicating that implantation may impact mobility in physically demanding tasks. Although we cannot disambiguate whether these changes are due to tethering, or the implant, it represents the maximal impact of the implantation, particularly as the heaviest Apollo implant was used. Therefore, the consistent hit rate and trial number, in a complex task requiring exploration in a large arena, demonstrates that the implant is well-suited to extended recordings from freely moving mice.

All parts of the implant (except the constructor probes, Thorlabs) were designed using Autodesk Inventor Professional 2023 software, acquired free of charge through the renewable education plan. Parts were 3D-printed by external companies (primarily SGD 3D, https://sgd3d.co.uk/), or at the SWC FabLab. Stereolithography (SLA, using Rigid4000 resin, Formlabs) was typically used for the payload and docking modules, the docking holder, and the constructor head. Selective laser sintering (using Nylon PA12) was typically used for the payload module lids and payload holder. Brass threaded inserts were manually added to the payload module, payload holder, and docking holder using a soldering iron after printing. For parts (e.g. the payload and docking modules) where strength and inflexibility were advantageous, we used Rigid4000 resin, although this material is denser than Nylon PA12. With this combination, the Neuropixels 1.0, 2.0α, and 2.0 implants weigh 1.7, 1.3, and 0.9 g. The weight of the implants can be further reduced to 1.5, 1.1, and 0.8 g if all parts are printed with Nylon PA12. The full-PA12 implants have been successfully used with 1.0 probes, but remains untested with the 2.0 versions. The miniaturized Neuropixels 2.0 implant with headstage holder weighed 0.6 g by itself or 1.1 g with the probe epoxied and ground attached. All probes used, and any resulting issues/breakages are detailed in Supplementary file 1. Damage resulting from historical procedural steps that are no-longer used (e.g. manually separating the shanks of a 4-shank probe with a needle, now achieved by de-ionized water or strong solvant) or carelessness during probe handling outside of mounting, implantation and explantation are not indicated in the table as they are independent of the implant itself.

In addition to the 3D-printed implant, the following materials are required (due to variable supply, up-to-date links are provided in the GitHub repository):

What follows is the protocol used by the originating laboratory with some minor variants. This is the most thoroughly tested and recommended approach. Methods employed by each individual lab are detailed in a later section.

Sessions were automatically spike-sorted using pyKilosort (Banga et al., 2022), python port of Kilosort (Pachitariu et al., 2016; version 2.0), and automatically curated using Bombcell (Fabre et al., 2023). One mouse (GB012) was spike-sorted using Kilosort 4 (Pachitariu et al., 2024).

A variety of parameters were used to select high-quality units, based either on their waveform and their spiking properties.

The template waveform-based criteria were: (1) a maximum of two peaks and one trough, (2) a spatial decay slope below −3 µV·µm⁻¹, defined as the slope of a linear fit (using the MATLAB polyfit function) between the maximum absolute amplitude of the peak channel and nearest five channels along the length of the probe (i.e. 75 µm away for Neuropixels 2.0), (3) a duration between 0.1 and 0.8 ms (Deligkaris et al., 2016), and (4) fluctuations during baseline not exceeding 30% of the peak amplitude. The raw waveform-based criteria, computed using at least 1000 randomly sampled spikes, were: (1) a minimum mean amplitude of 20 µV (only 1% of units had amplitudes below 30 µV, and increasing this threshold to 50 µV did not affect the results), and (2) a minimum mean signal-to-noise ratio of 0.1, defined as the absolute maximum value divided by the baseline variance. Both somatic and non-somatic spikes (Deligkaris et al., 2016) were kept.

The spiking properties-based criteria were: (1) a minimum of 300 spikes, (2) less than 20% of spikes missing, estimated by fitting a Gaussian to the spike amplitude distribution with an additional cut-off parameter below which no spikes are present, (3) a maximum of 10% refractory period violations, using a previously published approach (Hill et al., 2011), defining the censored period as 0.1 ms and estimating the refractory period using a window between 0.5 and 10 ms, and (4) a minimum presence ratio of 0.2, defined as the fraction of 1 min bins with at least one spike.

Raw traces (Figures 3C, D, 6C) were obtained by bandpass filtering each channel from spikeGLX between 400 and 9000 Hz (using the MATLAB bandpass function) and subtracting the median across channels. The RMS value was computed on the processed signal, and the median across all channels was used to summarize each recording.

To obtain the total number of spikes per second at each depth along the probe (Figure 3E, F), we summed spikes across all units present within each 20 µm depth bin.

In the case of 4-shank probes, we estimated the number of recorded units for a probe on a given day (Figure 4A, B) by summing units from a complete set of independently recorded banks—a set of channel banks that tiled the entirety of the implanted probe—when available. Because each such set was recorded across at least 2 days, we used the closest recordings within a window of 5 days, centered on the day of interest. Days were excluded if it was not possible to form a complete set of banks within a 5-day window.

To obtain P the percentage change in unit count Nd across days d (4C–F), we fit an exponential decay function to the number of units detected on each bank across days and extracted the decay parameter τ:Nd=N010τd

P was defined as:P=100×(Nd+1Nd−1)=100×(10τ−1)

P was averaged across all longitudinally recorded banks from each implantation to obtain a single value. Only banks with at least three recordings were included.

The median of the units’ amplitude and the median RMS values were fitted using a linear fit. Similarly, a single value for each implantation was obtained by taking the variable’s mean value across all banks.

To estimate the effect of repeated probe use, U, and antero-posterior and medio-lateral insertion coordinates (Y and X) on a variable of interest (e.g. the percentage change in unit count, or the RMS values) across days, we used a linear mixed-effects model (using the MATLAB fitlme function):P∼1+Y+X+U+(1|probeID)+(U−1|probeID)

Here, P is the response variable, Y,X, and U are fixed effect terms, and probeID (the probe identity) is the single grouping variable. We then assessed whether Y,X, or U had a significant effect on the response variable.

To track neurons across days (Figure 5), we used UnitMatch, which uses only the neurons’ waveform (van Beest et al., 2024). We identified cell presence across days using the intermediate algorithms, which have been shown to maximize the probability of tracking neurons while preserving a low false-positive rate. We then identified the neurons that were tracked across 6 arbitrary days spanning the whole recording period. The inter-spike intervals were computed as the distribution of the times between consecutive spikes, binned on a logarithmic scale from 0 to 5 s. As in van Beest et al., 2024, to compute the probability of a unit being tracked, we looked at each unit across all recordings and computed the probability of this unit being tracked in previous or subsequent recordings. These probabilities were then averaged across all the units from each animal, and averaged across animals.

To quantify the amount of information present in the distributions of the correlations of the fingerprints, we computed the ROC curve for different populations of pairs (from putative matched units or nonmatched units) across days. We then computed the area under the ROC curve (AUC) to quantify this difference between distributions. Only sessions with at least 20 matched units were considered. Units that had a match within recordings were excluded from this analysis. For each mouse, the AUCs were then averaged across recording locations.

To compute the distributions of RMS values and unit amplitudes (Figure 6), we computed these distributions first for single sessions (across channels and units, respectively), and then computed the average distribution across sessions and animals.