Comparing brains by cell type, not electrode number
Every neural recording captures a different set of neurons, which is why models trained on one animal usually fail on another: there is no correspondence to transfer along. Han and colleagues at KAIST route neurons into latent groups defined by their biological identity, transcriptomic, anatomical and electrophysiological features that are shared across individuals, and analyze dynamics at group level. The model then decodes behavior in completely held-out animals with parameters frozen and no neuron matching, reaching correlations of 0.79 in whole-brain zebrafish imaging against 0.30 or below for prior methods, and it produces manifolds that stay geometrically coherent where standard methods fragment.
Source: Neuronal Identity as an Organizational Basis for Analyzing Neural Population Dynamics, arXiv:2610.36745v1, 27 October 2026. Primary source. Read: the full PDF, including the routing framework, Tables 1 and 2, the held-out transfer protocol, and the identity-feature ablations.
What the work claims
This is a primary results paper, and its claim is organisational rather than architectural: the field has been treating neurons as interchangeable indexed units, when each neuron carries a stable multimodal biological profile, and that profile is a better coordinate system for comparing dynamics across subjects than any learned activity-based embedding.1 The framework is permutation-invariant by construction. Each neuron is described by identity features, encoded by per-modality networks and softly routed into a small set of latent groups; every neuron's activity window is encoded by a shared dynamics network and aggregated into those groups by routing weight; a group-interaction module then produces the population state. Because group membership depends on biological properties shared across individuals, the entire learned organisation can be applied to a new animal with no retraining and no neuron-wise correspondence.
The evaluation is unusually clean for a transfer claim: all methods are trained on source animals or sessions only, parameters are frozen, and the held-out recording is analysed whole. On the WARP dataset, whole-brain calcium imaging of roughly 77,000 neurons per larval zebrafish jointly with single-cell transcriptomics, the framework decodes fictive-swimming behavior at a Pearson correlation of 0.793 plus or minus 0.015 in held-out fish, where the best baseline, POYO, reaches 0.304 and DeepSets 0.160. Stimulus classification shows the same ordering, balanced accuracy 0.835 against 0.195 for the strongest baseline. On the Allen Neuropixels visual-cortex datasets, over 80,000 units across 51 sessions, held-out behavior decoding reaches 0.734 and 0.694 against baseline maxima of 0.615 and 0.588. Giving the baselines the very same identity features improves them substantially, POYO moves from 0.304 to 0.489 on zebrafish behavior, yet still lands well below the routing framework, which is the paper's sharpest point: identity features help, but organising the population by them is what transfers.
How it works
The routing step replaces the global sum or mean that standard permutation-invariant models use. A global mean answers "what is the population doing" by averaging over whatever neurons happen to be recorded, discarding the fact that a parvalbumin-positive fast-spiking interneuron and a layer 5 pyramidal cell contribute to circuit computation in categorically different ways. Instead, each neuron's identity vector is mapped to a soft membership over latent groups, so neurons with overlapping biological profiles pool their activity into a shared group representation, and population dynamics are read out as the interactions among those groups. Permutation invariance survives because summing over neuron index within each group is unchanged by reordering the input set.
Identity features span modalities: transcriptomic expression, three-dimensional position, anatomical region, and electrophysiological descriptors. A key practical finding is that the framework degrades gracefully when the expensive jointly-acquired transcriptomics are unavailable. On zebrafish, substituting independently measured atlas gene-expression data assigned by anatomical coordinates still works, and ablating the transcriptomic modality entirely leaves a model that decodes behavior at 0.727 and stimulus at 0.711 balanced accuracy, still above every baseline. On the mouse data, where only position and regional features are available, transfer still beats all baselines. The manifold results are the qualitative complement: on held-out animals, latent trajectories keep temporal neighbourhoods intact, temporal-distance correlation 0.633 on zebrafish against 0.151 for the best baseline, while CEBRA, MARBLE, LFADS and joint PCA fragment or collapse.
Where a skeptic should push
The most load-bearing assumption is that neuronal identity is stable enough, and measurable accurately enough, to serve as the transfer coordinate. The authors are honest about the weak points: jointly recorded activity-plus-transcriptome data remain rare, atlas-based assignment injects registration noise, and identity-feature quality shapes everything downstream. In the zebrafish experiments the gold standard held, expression measured in the same animal, but the more impressive and more deployable results lean on noisier proxies. A skeptic should also weigh sample sizes at the transfer level: three fish, yielding three held-out combinations, and fixed session splits for the mouse data. The per-cell data volumes are enormous, which is exactly what deep group-routing networks need to shine, but the number of independent biological individuals is small, and cross-animal transfer claims with n of three deserve replication across labs before being treated as settled.
Second, the baselines were adapted rather than re-implemented from their authors' pipelines, and the manifold comparison includes the courtesy of testing 1,000 random neuron orderings for baseline methods. The decoding gaps are so large, two to five fold in correlation, that modest baseline handicapping cannot explain them away, but the identity-augmented baselines, not the originals, are the honest comparator and the paper leads with those internally even though tables headline the original forms. Third, everything here is offline analysis. Nothing in the framework speaks to causal manipulation, and a routing group is a computational convenience, not a verified circuit motif.
A cell-type coordinate for organoid recordings
The organoid field has the correspondence problem in a far crueler form. Two brain organoids from the same cell line, grown weeks apart, share no units, unstable cell-type proportions, shifting ratios of progenitor-like to neuronal populations, and rhythms that mature on different schedules. Every MEA experiment is effectively a new animal with a worse atlas, and the field's response has been to average across organoids and hope. This paper proposes the missing coordinate: route recorded units into groups by their biological identity, whatever is measurable, electrophysiological signature, morphology, or matched single-cell expression from sister tissue, and compare organoids at the level of group dynamics rather than unit indices. That would convert organoid-to-organoid variability from a nuisance that dilutes statistics into an addressable structure: batch differences would show up as differences in group composition and group-level dynamics, not as unexplained noise across incompatible electrode maps.
The mechanism opportunity is concrete. If an organoid computing assay trains a decoder on one organoid, today's version dies on the next; identity-routing is precisely the trick that lets the trained model be carried across biological replicates with frozen parameters. For biological computing, where every substrate is a unique living population, a shared identity coordinate is arguably the difference between a one-off demonstration and a reproducible platform. The same machinery is a quality-control instrument: an organoid whose group composition falls outside the training ensemble's identity manifold is detectable before any task is run.
The threats deserve equal weight. Real brains have atlases, stable laminar geometry and conserved cell types; unguided cortical organoids have none of that reliably. Their cell types are immature and partly off-target, so identity features are noisier than anything in this paper, and registration proxies that work for zebrafish brains do not exist for organoid tissue. There is a real risk of baking a false biological prior into the analysis: routing units into groups presumes the groups mean something, and in a disorganised organoid they may not. Worst case, identity-routing launders batch effects into apparent structure. The paper's own control, identity-augmented baselines that improve but do not match routing, at least shows the features carry signal even when the organisation is wrong; that is the property an organoid adaptation would lean on hardest.
The bottom line
Established: organising population analysis by multimodal neuronal identity enables direct, frozen-parameter transfer to unseen animals and sessions across two species and two modalities, beating strong activity-only baselines by wide margins, and the advantage comes from the organisation, not merely from access to the features. Not established: robustness to badly degraded identity features, generality beyond these three datasets, or anything causal about what the latent groups are. The confirming result for organoid intelligence would be a demonstration that an identity-routed decoder trained on one organoid transfers to a fresh organoid from the same protocol with less retraining than index-based models. The breaking result would be identity-routing collapsing to baseline performance once identity features are drawn from noisy organoid proxies, which would tell the field its coordinate system is not there yet.
Frequently asked questions
Why can't models just transfer across animals already?
Each recording samples a different, unmatched set of neurons, and existing models anchor their representations to unit indices or session-specific embeddings. When the neuron population changes, those anchors vanish. Identity-routing anchors instead to biological features that are shared across individuals of a species.
What counts as a neuronal identity feature?
In this paper, transcriptomic expression, three-dimensional position, anatomical region and electrophysiological descriptors. The authors show the framework still works with independently measured atlas expression substituted for joint transcriptomics, and with transcriptomics removed entirely.
Is the gain just from adding better inputs?
No. When the same identity features are given to POYO, NEDS and DeepSets, all three improve, but all remain below the routing framework. The paper's argument is that how identity organises the population matters more than whether the model can see it.
How big are the datasets behind the claims?
Whole-brain calcium imaging of about 77,000 neurons in each of three larval zebrafish with paired single-cell transcriptomics, and more than 80,000 electrophysiological units across 51 Neuropixels sessions in mouse visual cortex. Transfer is evaluated at the level of whole held-out animals or sessions.
Why is this relevant to organoid intelligence?
Organoid recordings share no units across batches and lack stable anatomy, so cross-organoid comparison is the field's core unsolved nuisance. Identity-routing offers a way to compare dynamics across organoids in a biological coordinate rather than electrode index, and to carry trained decoders across biological replicates.
What is the main risk of applying it to organoids?
Organoid cell types are immature and noisy, and no atlas exists to stabilise identity features. Routing into groups presumes the groups are biologically meaningful; if they are not, the method can manufacture apparent structure from batch effects rather than reveal it.
References
- S Han, J Y You, S Kim, M Eom, J Park, S Park, B Lee, Y-G Yoon. Neuronal Identity as an Organizational Basis for Analyzing Neural Population Dynamics. arXiv:2610.36745. 2026. http://arxiv.org/abs/2610.36745v1. Accessed 2026-11-01.