Neuronal Identity as an Organizational Basis for Analyzing Neural Population Dynamics

Seungjae Han; Joshua Yedam You; Soi Kim; Minho Eom; Jinho Park; Seunghee Park; Byeongwook Lee; Young-Gyu Yoon

q-bio.NCOct 27, 2026

Neuronal Identity as an Organizational Basis for Analyzing Neural Population Dynamics

Authors: Seungjae Han, Joshua Yedam You, Soi Kim, Minho Eom, Jinho Park, Seunghee Park, Byeongwook Lee, Young-Gyu Yoon

Organizations: School of Electrical Engineering, KAIST, Daejeon, Republic of Korea · Department of Brain and Cognitive Sciences, KAIST, Daejeon, Republic of Korea · Graduate School of Data Science, KAIST, Daejeon, Republic of Korea · Department of AX, KAIST, Daejeon, Republic of Korea ·

Abstract

Individual neurons exhibit diverse molecular, morphological, and electrophysiological properties that shape their distinct roles in neural circuits. While these multimodal features are largely shared across individuals, they remain an underutilized source of information in analyzing neural population dynamics. We posit that these shared biological profiles provide a common reference for relating neural dynamics across subjects and recording sessions. Our framework leverages this structure by routing neurons into latent groups based on their biological profiles and analyzing population dynamics through group-level representations. Because these assignments are anchored in biological properties that are consistent across subjects, the learned organization supports transfer to unseen animals without retraining or neuron-wise correspondence. We evaluate our framework on whole-brain calcium imaging in larval zebrafish and large-scale electrophysiology in the mouse visual cortex. Across both species and recording modalities, the framework delivers state-of-the-art performance in both neural decoding and manifold analysis: it achieves superior decoding of behavior and stimulus labels while better preserving the temporal consistency and local geometric structure of latent trajectories on unseen data. Our experiments demonstrate that not only does providing biological features improve baseline performance, but our proposed framework achieves even stronger results by using these properties to explicitly structure the population rather than treating them as auxiliary inputs. Together, these results suggest that organizing neural populations by their multimodal biological features offers a promising framework for generalizing population dynamics across subjects and recording sessions.

Figures & tables

Figure 1: Overview of identity-guided neural population analysis. (a) Conceptual overview. Multimodal biological features organize neurons into corresponding groups across sessions, even when exact individual neurons cannot be matched across subjects. Population dynamics are described at the level of these identity-defined groups and their interactions. (b) Identity-guided routing framework. Identity-related features are encoded to softly route neuron-level activity tokens into latent group tokens, which are processed into a population embedding. Manifold learning, stimulus classification, and behavior decoding are evaluated as separate downstream objectives.

Dataset	Method	Behavior (PCC) $\uparrow$	Stimulus (Balanced Accuracy) $\uparrow$	Stimulus (Accuracy) $\uparrow$
WARP	Ours	$\mathbf{0.793\pm 0.015}$	$\mathbf{0.835\pm 0.029}$	$\mathbf{0.839\pm 0.029}$
	POYO* [ 9 ]	${0.304\pm 0.120}$	${0.153\pm 0.032}$	${0.156\pm 0.034}$
	NEDS* [ 19 ]	${0.248\pm 0.207}$	${0.195\pm 0.034}$	$0.196\pm 0.033$
	DeepSets [ 18 ]	$0.160\pm 0.081$	$0.161\pm 0.011$	$0.160\pm 0.011$
Allen Brain Observatory	Ours	$\mathbf{0.734\pm 0.008}$	$\mathbf{0.893\pm 0.012}$	$\mathbf{0.899\pm 0.010}$
	POYO [ 9 ]	${0.526\pm 0.043}$	${0.220\pm 0.022}$	$0.258\pm 0.114$

Table 1: Behavior and stimulus decoding under direct held-out transfer across three datasets (mean

\pm

std). Bold indicates the best result per column within each dataset. * denotes models adapted to calcium signals.

Dataset	Method	Temporal Distance Correlation $\uparrow$	Normalized Latent-Neighbor Time Gap $\downarrow$	Local Trajectory Linearity $\uparrow$
WARP	Ours	$\mathbf{0.633\pm 0.078}$	$\mathbf{0.145\pm 0.009}$	$\mathbf{0.911\pm 0.006}$
	CEBRA [ 3 ]	$0.086\pm 0.010$	$0.328\pm 0.003$	$0.684\pm 0.010$
	MARBLE [ 4 ]	$0.045\pm 0.017$	$0.332\pm 0.002$	$0.512\pm 0.006$
	LFADS [ 26 ]	$0.141\pm 0.020$	$0.297\pm 0.012$	$0.666\pm 0.015$
	Joint PCA	$0.151\pm 0.045$	$0.312\pm 0.007$	$0.887\pm 0.032$
Allen Brain Observatory	Ours	$\mathbf{0.355\pm 0.033}$	$\mathbf{0.063\pm 0.021}$	$\mathbf{0.896\pm 0.005}$

Table 2: Geometric structure of latent trajectories evaluated on held-out animals or sessions across three datasets (mean

\pm

std). Bold indicates the best result per column.

Figure 2: Latent trajectories for two source animals (Fish 59 and Fish 63) and held-out animals across all methods. Points are colored by time within each recording. Our method preserves structured temporal trajectories in held-out animals, whereas baseline methods yield fragmented or collapsed trajectories under direct held-out transfer.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: Data from three fish in the WARP dataset. Neurons are shown in the registered zebrafish brain coordinate space and colored by transcriptomic marker identity (top), anatomical region label (middle), and selected marker-gene categories (bottom).

Figure A2: Representative Allen visual cortex sessions shown in a shared anatomical coordinate space and colored by visual area label.

Figure A3: Decoding targets used in WARP and Allen experiments. For WARP, we show a representative fish with neural activity from selected neurons, the swim-power behavioral target, and the visual-stimulus classification target. For Allen, we show representative sessions from Allen Brain Observatory dataset and Allen Functional Connectivity dataset with binned spike counts from selected units, running-speed behavioral targets, and grouped stimulus-family labels. Behavior decoding is evaluated with Pearson correlation coefficient, and stimulus decoding is evaluated with accuracy.

Feature Group	Description	WARP	Allen
Transcriptomics	41-dimensional gene-expression marker vector	✓
Spatial position	3D registered coordinate in reference space	✓	✓
Anatomical region	Multihot vector over brain-region labels	✓	✓
HCR expression	236-dimensional atlas-derived gene-expression marker vector as documented in Appendix A.3	✓
Electrophysiological features	Features from observed dynamics including mean, standard deviation, extrema, quartiles, positive fraction, mean squared value, and the standard deviation of first differences	✓	✓

Appendix

Table A1: Identity feature groups for each dataset.

Parameter	Value
Latent groups $K$	256
Activity encoder hidden size	128
Group activity dimension	384
Identity hidden size	192
Group mixer hidden size	64
Routing temperature	0.35

Appendix

Table A2: Hyperparameters for supervised decoding with the proposed method.

Parameter	Value
Output dimension	3
Latent groups $K$	32
Activity encoder hidden size	16
Group activity dimension	32
Identity hidden size	16
Group mixer hidden size	32

Appendix

Table A3: Hyperparameters for manifold learning with the proposed method.

Figure A4: Latent trajectories for two source fish (Fish 59 and Fish 63) and held-out fish across all methods. Points are colored by behavior label within each recording.

Figure A5: Latent trajectories for two source fish (Fish 59 and Fish 63) and held-out fish across all methods. Points are colored by stimulus label within each recording.

Dataset	Method	Behavior (PCC) $\uparrow$	Stimulus (Balanced accuracy) $\uparrow$
WARP	POYO* $+$ (T,S,H,R,P)	${0.489\pm 0.096}$	${0.164\pm 0.042}$
	POYO* [ 9 ]	${0.304\pm 0.120}$	${0.153\pm 0.032}$
	NEDS* $+$ (T,S,H,R,P)	${0.574\pm 0.108}$	${0.374\pm 0.074}$
	NEDS* [ 19 ]	${0.248\pm 0.207}$	${0.195\pm 0.034}$
	DeepSets $+$ (T,S,H,R,P)	$0.466\pm 0.051$	$0.163\pm 0.023$
	DeepSets [ 18 ]	$0.160\pm 0.081$	$0.161\pm 0.011$

Appendix

Table A4: Behavior and stimulus decoding under direct held-out transfer across three datasets (mean

\pm

std). * denotes models adapted to calcium signals.

Method	Behavior (PCC) $\uparrow$	Stimulus (Balanced accuracy) $\uparrow$
Ours (T,S,H,R,P)	$0.793\pm 0.015$	$0.835\pm 0.029$
Ours (S,H,R,P), w/o $T$	$0.727\pm 0.013$	$0.711\pm 0.016$
Ours (T,H,R,P), w/o $S$	$0.721\pm 0.014$	$0.573\pm 0.010$
Ours (T,S,R,P), w/o $H$	$0.717\pm 0.029$	$0.693\pm 0.032$
Ours (T,S,H,P), w/o $R$	$0.716\pm 0.020$	$0.682\pm 0.030$
Ours (T,S,H,R), w/o $P$	$0.704\pm 0.026$	$0.722\pm 0.025$

Appendix

Table A5: Modality ablation on WARP (mean

\pm

std over 5 seeds). Each row removes one modality from the full identity set. We denote static identity features by

T

for transcriptomic profile,

S

for spatial position,

H

for atlas-derived HCR gene markers, and

R

for anatomical region. We denote the electrophysiological features by

P

Method	Behavior (PCC) $\uparrow$	Stimulus (Balanced accuracy) $\uparrow$
Ours (S,R,P)	$0.734\pm 0.008$	$0.893\pm 0.012$
Ours (R,P), w/o $S$	$0.686\pm 0.037$	$0.786\pm 0.066$
Ours (S,P), w/o $R$	$0.711\pm 0.007$	$0.800\pm 0.040$
Ours (S,R), w/o $P$	$0.696\pm 0.004$	$0.787\pm 0.007$

Appendix

Table A6: Modality ablation on the Allen Brain Observatory dataset (mean

\pm

std over 3 seeds). Each row removes one modality from the full identity set. Here

S

denotes spatial position,

R

anatomical region, and

P

electrophysiology features.

Dataset	Configuration	Changed component	Behavior (PCC) $\uparrow$
WARP	Full model	Soft routing	$\mathbf{0.793\pm 0.015}$
WARP	Hard top-1 routing	Soft routing removed	$0.708\pm 0.022$
Allen Brain Observatory	Full model	Soft routing	$\mathbf{0.734\pm 0.008}$
Allen Brain Observatory	Hard top-1 routing	Soft routing removed	$0.678\pm 0.003$
Allen Functional Connectivity	Full model	Soft routing	$\mathbf{0.694\pm 0.058}$
Allen Functional Connectivity	Hard top-1 routing	Soft routing removed	$0.626\pm 0.078$

Appendix

Table A7: Architectural ablations on behavior decoding (mean

\pm

std over 5 seeds for WARP and 3 seeds for Allen). Each row varies one design choice from the full model. Bold indicates the best result within each dataset.

Figure A6: Target-ordering analysis for temporal distance correlation. Each histogram shows baseline performance over $1000$ sampled target-neuron orderings. The vertical green line indicates our method and the vertical red dash indicates the best performance among the sampled orderings.

Figure A7: Target-ordering analysis for normalized latent-neighbor time gap. Each histogram shows baseline performance over $1000$ sampled target-neuron orderings. The vertical green line indicates our method and the vertical red dash indicates the best performance among the sampled orderings.

Figure A8: Target-ordering analysis for local trajectory linearity. Each histogram shows baseline performance over $1000$ sampled target-neuron orderings. The vertical green line indicates our method and the vertical red dash indicates the best performance among the sampled orderings.

Experiment group	GPUs used	Wall-clock time	Approx. GPU-hours
WARP supervised decoding experiments	3 RTX 3090 GPUs	1.5 h	4.6
Allen supervised decoding experiments	3 RTX 3090 GPUs	7.8 h	23.3
WARP latent trajectory experiments	3 RTX 3090 GPUs	0.5 h	1.5
Allen latent trajectory experiments	3 RTX 3090 GPUs	6.4 h	19.2
Total for reported experiments	–	–	$\approx 49$ GPU-hours

Appendix

Table A8: Compute required to reproduce the reported experiments. GPU-hours are estimated as wall-clock time multiplied by the number of GPUs used by the launcher.

Neural Decoding

Neural decoding maps brain activity or error-correction syndromes to stimuli, behavior, language, or corrective actions.

See all 184 papers

Oct 2, 2026

Neural Data Needs Semantic Tokenization: Behavioral Events as Boundaries of Session-Transferable Tokens

Extracellular electrophysiology records a different set of neurons in every session. Neural foundation models embed each neuron and each session into their tokens, so every new session is an input they have never seen, and they fail to generalize to it. A tokenizer for new sessions needs a unit that every session shares and that carries behavior. Population activity offers such a unit. It evolves on a low-dimensional manifold that persists across neuronal turnover and across animals once sessions are aligned. This manifold changes regime at task events such as stimulus onset and movement onset. Within each regime the population occupies a state, the part of the manifold it spans between two events, and each state carries its own behavioral meaning. We propose Tokenization with States (TWS), which segments each trial, one repetition of the task, at these events and converts every state into tokens of population geometry, with no neuron or session embedding. On held-out International Brain Laboratory (IBL) sessions, TWS decodes movement even from regime boundaries that carry no information about the target, while a per-neuron foundation model pretrained on those sessions decodes at chance. Frozen after training on mice alone, TWS transfers to macaques and Utah arrays with only a linear probe. On reach direction, a target that its boundaries do not define, it achieves a Matthews correlation of 0.23, where the event time alone achieves

0.01

. For cross-session generalization, the token matters more than the model on top.

Sangyoon Bae, Jiook Cha

Interdisciplinary Program in Artificial Intelligence Seoul National University Seoul, 08826, South Korea · Department of Psychology Seoul National University Seoul, 08826, South Korea

Neural Decoding Transfer Learning

Ask PDF

Jun 9, 2026

Machine Learning Methods for Studying Latent Neural Activity Dynamics

Recent developments in brain recording are driving a demand for machine learning tools capable of decoding the latent structure of large populations of neurons. In this paper, we provide a comprehensive survey that outlines the trajectory of Latent Variable Models (LVMs) from early state-space models to more recent deep generative models. We organize the literature into three closely related domains: (1) Single-Region Latent Dynamics, which includes models such as linear dynamical systems to more complex dynamics represented by Recurrent Neural Networks (RNNs) and Neural Ordinary Differential Equations (ODEs); (2) Multi-Region Communication, which employs probabilistic as well as subspace methods to study how information is transferred across different brain areas considering synaptic propagation delays and network connectivity; and (3) Behavior-Aligned Modeling, which seeks to disentangle neural activity related to task performance from other internal states via supervised or contrastive learning. This survey also includes large-scale neural foundation models, such as Transformers and diffusion models, that rely on large-scale pre-training for optimal performance across subjects. Finally, we conclude and discuss benchmarks, evaluation criteria, and open challenges, such as the ability to identify causal links or directionality of communication, to facilitate future research for bridging interpretable brain dynamics with reliable neural decoding.

Shufeng Kong, Fumei Deng, Xinyi Dong +7

School of Software Engineering, Sun Yat-sen University, Zhuhai, China · School of Computer Science and Artificial Intelligence, Foshan University, Foshan, China · Department of Ecology and Evolutionary Biology, Cornell University, Ithaca, NY, USA +2

Dynamical Systems Latent Variable Models Latent Dynamics Modeling Neural Processes Neural Decoding

Ask PDF

Mar 11, 2026

JEDI: Jointly Embedded Inference of Neural Dynamics

Animal brains flexibly and efficiently achieve many behavioral tasks with a single neural network. A core goal in modern neuroscience is to map the mechanisms of the brain's flexibility onto the dynamics underlying neural populations. However, identifying task-specific dynamical rules from limited, noisy, and high-dimensional experimental neural recordings remains a major challenge, as experimental data often provide only partial access to brain states and dynamical mechanisms. While recurrent neural networks (RNNs) directly constrained neural data have been effective in inferring underlying dynamical mechanisms, they are typically limited to single-task domains and struggle to generalize across behavioral conditions. Here, we introduce JEDI, a hierarchical model that captures neural dynamics across tasks and contexts by learning a shared embedding space over RNN weights. This model recapitulates individual samples of neural dynamics while scaling to arbitrarily large and complex datasets, uncovering shared structure across conditions in a single, unified model. Using simulated RNN datasets, we demonstrate that JEDI accurately learns robust, generalizable, condition-specific embeddings. By reverse-engineering the weights learned by JEDI, we show that it recovers ground truth fixed point structures and unveils key features of the underlying neural dynamics in the eigenspectra. Finally, we apply JEDI to motor cortex recordings during monkey reaching to extract mechanistic insight into the neural dynamics of motor control. Our work shows that joint learning of contextual embeddings and recurrent weights provides scalable and generalizable inference of brain dynamics from recordings alone.

Anirudh Jamkhandi, Ali Korojy, Olivier Codol +2

Mila - Quebec AI Institute · Université de Montréal

Latent Dynamics Modeling Dynamical Systems Recurrent Neural Networks System Identification

Ask PDF

Abstract

Neural Decoding

Explore similar work

Neural Data Needs Semantic Tokenization: Behavioral Events as Boundaries of Session-Transferable Tokens

Machine Learning Methods for Studying Latent Neural Activity Dynamics

JEDI: Jointly Embedded Inference of Neural Dynamics