Research analysis · Biocomputing

PULSE: one latent space for organoid dynamics, disease, and closed-loop control

A foundation model pretrained on 914 hours of microelectrode array recordings claims to give every neural preparation - organoid, slice, mouse, human, diseased or healthy - a common coordinate system. The most consequential result for biological computing is not the disease biology: it is that a closed-loop task engages dynamics the network already had.

Source: A foundation model for brain organoid network dynamics uncovers intrinsic computation and convergent disease phenotypes, bioRxiv preprint, 2026-10-13. Primary source. Read the full text of the preprint including figure legends and the limitations section.

What the work claims

PULSE (Population Unified Latent Space Embedding) is a self-supervised foundation model built on the Mamba-2 state-space architecture, about 12 million parameters, trained on 3,200 high-density MEA recordings totaling 914.3 hours of spiking activity from mouse stem-cell-derived forebrain organoids, 2D dissociated cultures, and ex vivo cortical slices1. This is a preprint reporting primary experimental and modeling work: real recordings, modest sample sizes (5 to 13 organoids per comparison), careful explicit statistics, no peer review yet.

The paper makes three claims. First, the frozen model predicts population spiking across preparations better than a causal exponentially weighted moving average baseline, and a linear readout fitted on two minutes of a new recording restores calibration on never-seen cultures. Second, two neurodevelopmental disorder genotypes - a 16p11.2 deletion and SHANK3 deficiency - leave distinguishable, partly convergent signatures in the latent space, including accelerated maturation in the 16p11.2 line. Third, and the claim this analysis leans on hardest: when organoid networks are embedded in a closed-loop cartpole task, the activity they engage lies almost entirely inside the subspace their spontaneous activity already occupied. Selection, not instruction.

How it works

The model input is spike counts per unit binned at 10 ms, concatenated with the electrical stimulation delivered, the number of active units, and a learned embedding identifying the spike sorter. A convolution downsamples to 20 ms, six Mamba-2 layers compress the history into a 512-dimensional latent state - fixed dimensionality regardless of how many units a recording contains - and a Poisson head predicts the next 20 ms of population firing. The fixed latent size is the whole trick: recordings with 30 units and recordings with thousands become points in the same space, so their statistics can be compared directly.

Trained without biological labels, the model is then used as a measuring instrument. A ridge regression from latent-state covariance onto chronological age, fitted on one control line alone, acts as a developmental clock that transfers without refitting. A label-free principal component across all 327 spontaneous recordings acts as a phenotype axis. A linear discriminant on the latent state decodes which electrode was stimulated, or whether a closed-loop episode will end in success or failure.

The headline numbers hold up on inspection. On validation organoids the reference model reaches a per-neuron median of 0.339 bits per spike against 0.042 for the EWMA baseline. On nine held-out mouse cultures it beats the baseline on correlation metrics but not on calibration, and a linear head fitted on 120 seconds of data lifts per-neuron median bits per spike from 0.335 to 0.359, improving all nine cultures (Wilcoxon P = 0.0039). Stimulated-site identity is decodable from the latent state at a median 4.14 times chance with site identity hidden from the model, against 1.74 times for binned firing rate over the same window (n = 17 organoids).

On the human side - four iPSC lines held out from pretraining entirely, a genuine cross-species transfer test - the 16p11.2-deletion line (14765-X2) is predicted significantly worse than its donor control on all three per-neuron metrics (per-neuron Pearson r medians 0.219 versus 0.302, Cohen's d = 2.39, P = 0.0043), while the isogenic SHANK3 heterozygous line is not (all P ≥ 0.24). The developmental clock, anchored on the unaffected-donor line, orders 14765-X2 recordings by age (mean Spearman ρ = +0.70, P = 1 × 10-4), and 14765-X2 experiments project ahead of their chronological age by a median of +9.7 days against +0.8 for the control, with all six deletion organoids ahead of all six controls - the smallest P value attainable at that sample size. The label-free PC2 axis shifts in the same direction for both disorder lines; the shift is significant for 16p11.2 (Cliff's δ = +0.411, P = 0.037), not for SHANK3 alone (P = 0.26), and significant in the joint stratified permutation (P = 0.022). Across 28 classical spiking features, the two mutations move 15 of 15 bursting, topology, and temporal-structure features in a common direction, but only 2 of 8 rate-and-yield features.

Where a skeptic should push

The most load-bearing assumption is that a latent space trained to predict spike counts is a neutral coordinate system rather than a model-specific filter. Every downstream claim - the developmental clock, the convergence axis, the selection result - is a statement about PULSE's latent geometry, and the authors acknowledge that a transformer or hybrid backbone might carve the space differently. The subspace-overlap controls partly address this: the selection result replicates in binned firing rates with no model in the path (median 86 to 88 percent of variance captured against 50 percent chance, 10 of 10 organoids), so the geometry is not purely a Mamba-2 artifact. That is the right control and it passes.

Second, the convergence claim is statistically thinner than the abstract implies. With 5 and 6 organoids per line, the SHANK3-alone comparison cannot reach conventional significance, and the joint P = 0.022 is doing real work. The 16p11.2 comparison is donor-versus-donor, not isogenic, so part of the maturation acceleration may be genetic background rather than the deletion; the authors state this plainly. The isogenic SHANK3 pair is the cleaner experiment, and it is the one that did not individually clear the bar.

Third, keep the closed-loop result in proportion. The paper analyzes 13 organoids across spontaneous, open-loop, and closed-loop cartpole regimes, where stimulation encodes the pole angle and decoded firing applies force to the cart. Closed-loop control shifts the network's average state furthest from spontaneous activity (median 15.5 times within-recording movement against 4.4 for open loop, n = 13, P ≤ 0.047), yet 85.7 percent of closed-loop variance still falls inside the spontaneous subspace, and the balance-versus-failure direction lies 62.7 percent within it against 11.2 percent chance. A latent-state decoder separates sustained balance from quick failure at median AUC 0.790 (significant in 10 of 12 organoids). None of this demonstrates that the organoid learns the task; the learning claim belongs to earlier work the paper cites. What is shown is that task-relevant variation lives in the intrinsic repertoire, a statement about representation, not causation.

One result deserves more attention than it gets: at 8 Hz stimulation the networks generate a subharmonic at half the stimulation frequency (median power 3.17 times off-harmonic controls, 7 of 8 organoids, P = 0.016), absent at 4 Hz and abolished by memoryless Poisson surrogates. A subharmonic requires memory outlasting one stimulus cycle - direct evidence the tissue performs a nonlinear transformation with temporal memory, not passive relay.

What PULSE changes for biological computing

For anyone building computers out of living tissue, this paper quietly rearranges the engineering checklist.

The first consequence is a design constraint: you recruit, you do not write. If closed-loop task activity is confined to the spontaneous subspace - 85.7 percent of the spontaneous ceiling here - then the set of computations a given organoid can support is bounded by the repertoire its spontaneous dynamics already span. Training a biological computer is not programming; it is selecting among pre-existing modes. That inverts the usual procurement logic: instead of asking how to teach an organoid a task, you screen organoids for a spontaneous repertoire that already contains the task's dynamics, using the latent space as the assay. The spontaneous-subspace capture fraction becomes a capability metric you can measure before wiring anything up.

The second consequence is a quality-control metric hiding in plain sight. The model predicts 16p11.2 organoids worse than controls, in a way that survives fine-tuning. For biocomputing, predictability is calibration stability: a substrate whose next 20 ms of firing cannot be predicted is one whose decoders will silently degrade. The two-minute linear-head recovery curve is, read uncharitably, the most useful number in the paper for our field - it quantifies how much calibration data a new tissue batch costs you. A foundation model like this turns substrate QC into a pass-fail measurement rather than a morphology inspection: grow, record two minutes, fit the head, read off the bits-per-spike recovery.

The third consequence is about drift, the field's quiet enemy. Across multi-day recordings (median 49.7 hours, up to 73.5), the latent state wandered in an ordered, bounded way - pseudotime recovered true chronological order at median Spearman ρ = 0.96, and centroid displacement grew sublinearly with lag (MSD exponent α = 0.77). Most drift was global (median 76 percent shared across all electrodes), and per-pulse baseline subtraction stripped nearly all of it, keeping the electrode-specific code readable at 1.86 to 3.43 times chance even 7 to 12 cycles (median 32 hours) apart, where the uncorrected readout collapsed. The practical rule: drift is slow, mostly shared, and correctable with per-pulse baselining - and the structure you pay for, the electrode-specific response, is the most durable part.

The threat is the mirror image. If selection dominates, organoid computing has a capability ceiling set by development, and no closed-loop cleverness writes new dynamics into adult-ish tissue. Disease or protocol variability that perturbs spontaneous dynamics - the coordination phenotype shared here by two autism-risk genotypes - directly rewrites what the substrate can ever compute. And if selection is even partly wrong, if instruction contributes in regimes this study did not probe, the screening logic above becomes conservative in a way the field should know it is choosing. Either way, the paper hands biological computing its first serious common ruler.

The bottom line

Established here: a 12-million-parameter model trained on organoid and slice electrophysiology transfers across preparations, species, and genotypes with two minutes of calibration; two disorder genotypes leave convergent coordination phenotypes (jointly significant, individually thin); and closed-loop task activity is confined to the spontaneous subspace, with a genuine nonlinear memory signature at 8 Hz. Not established: that organoids learn the cartpole task, that the SHANK3 phenotype replicates at adequate power, or that the selection account holds outside cortical organoids on planar arrays. What would break the claim: a backbone-independent failure of the subspace-overlap result, or a closed-loop regime where evoked activity escapes the spontaneous subspace. Watch for the peer-reviewed version; the drift and subharmonic results are the ones to reproduce first.

Frequently asked questions

What is PULSE in one sentence?

A Mamba-2 foundation model trained on 914 hours of MEA recordings that maps any neural preparation's spiking activity into a shared 512-dimensional latent space, enabling comparison across genotypes, protocols, and tasks.

Does this show organoids learn the cartpole task?

No. The paper analyzes activity during closed-loop control but does not demonstrate learning; the learning claim belongs to earlier cited work. What it shows is that task-relevant network states lie inside the spontaneous repertoire.

What does selection versus instruction mean?

Selection says external input recruits dynamics a network already generates; instruction says input writes new dynamics. PULSE finds closed-loop and even 8 Hz stimulation activity stays almost entirely within the spontaneous subspace, favoring selection.

Why does the 16p11.2 result matter for computing?

Beyond the disease biology, the deletion line is intrinsically less predictable even after calibration. For biocomputing, predictability is decoder stability, so genotype and protocol variability directly change substrate quality.

How bad is drift for closed-loop experiments?

Bounded and mostly global. Latent-state drift was ordered (recoverable as pseudotime at Spearman ρ = 0.96) and confined, and per-pulse baseline subtraction kept the electrode-specific code decodable roughly 32 hours later.

What is the strongest evidence of real computation?

The subharmonic. At 8 Hz stimulation the networks produced a frequency component at half the stimulation rate, which requires memory spanning more than one stimulus cycle, and memoryless surrogates abolished it.

References

  1. Schweiger, H. E., Robbins, A., Hernandez, S., Moreno-Ochando, A., et al. A foundation model for brain organoid network dynamics uncovers intrinsic computation and convergent disease phenotypes. bioRxiv. 2026. doi:10.64898/2026.10.11.758231. https://www.biorxiv.org/content/10.64898/2026.10.11.758231. Accessed 2026-11-12.