PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.36745v1q-bio.NCcs.LGeess.SP

TL;DR

Neural population models face limited transfer across animals and recording sessions because recorded populations differ and neuronal identity is underused as a shared reference. The paper routes neurons into identity-defined groups and analyzes their dynamics through group-level representations. Across zebrafish and mouse data, this framework reports state-of-the-art decoding and stronger preservation of latent-trajectory structure under held-out transfer.

  • Problem

    Neural recordings differ across animals and sessions, with no guaranteed neuron-wise correspondence, while shared multimodal neuronal identity features remain underused in population-dynamics analysis.

    Source: §1 · pp. 1–2

  • Method

    The framework uses multimodal neuronal identity features to route neurons into latent groups and analyze population dynamics through group-level representations.

    Source: §1 · p. 2, §6 · p. 9

  • Results

    Across zebrafish and mouse datasets, identity-guided routing achieves state-of-the-art decoding and preserves temporal consistency and local geometry in held-out animals, outperforming baselines also given identity features.

    Source: §1 · pp. 2–3, §6 · p. 9

  • Takeaways & Limitations

    Organizing population dynamics by neuronal identity offers a shared structure for transfer across animals and sessions without retraining or neuron-wise correspondence.

    Source: §6 · p. 9

    Performance depends on identity-feature availability and quality; limited joint measurements and imperfect alignment can introduce noise into group assignments.

    Source: §6 · p. 10

Abstract

from arXiv · show

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.

1 Introduction

Neural population models face a transfer challenge because recordings differ across animals and sessions, while shared neuronal identity features offer a potential common reference. The paper routes neurons into identity-defined groups and reports strong decoding and manifold results across zebrafish and mouse datasets.

1 Introduction

  • Neural recordings differ across individuals and conditions, with no guaranteed correspondence between recorded populations, making direct transfer to unseen individuals challenging.The neuron populations themselves differ across animals, and activity patterns shift with recording conditions and internal states.Source: §1 · pp. 1–2
  • Multimodal neuronal identity features are underused, yet are largely shared within a species and can provide a stable basis for relating population dynamics across individuals.The features include transcriptomic, regional, and electrophysiological properties.Source: §1 · p. 2
  • The framework routes neurons into latent groups using biological profiles and analyzes dynamics at group level, supporting transfer without retraining or neuron-wise correspondence.The method uses shared neuron-wise networks and assignment-weighted aggregation to maintain permutation invariance.Source: §1 · p. 2
  • Supplying identity features improves baseline performance, while identity-guided routing consistently outperforms baselines on cross-session decoding.The reported transfer is applied directly to held-out recordings without retraining or neuron-wise correspondence.Source: §1 · p. 2
  • Across zebrafish and mouse datasets, the framework achieves state-of-the-art behavior and stimulus decoding and preserves temporal consistency and local geometry in held-out animals.The reported evaluations include whole-brain zebrafish calcium imaging and mouse visual-cortex electrophysiology.Source: §1 · pp. 2–3

2 Method

The method uses neuronal identity to organize population activity into latent groups, addressing arbitrary neuron ordering and the lack of neuron-wise correspondence across recordings. Shared encoders and routing-weighted aggregation produce group representations that can be processed into transferable population states.

2 Method

  • The method targets transfer across sessions without neuron-wise correspondence, using neuronal identity as a shared description because recording order changes across animals and sessions.The design also requires permutation invariance and groups neurons with shared properties.Source: §2 · p. 3
  • Rather than pooling all neurons identically, the framework routes neurons by identity into latent groups and aggregates their activity into group-level representations.Group representations are then processed to obtain the final population state.Source: §2 · p. 3

2.1 Framework

  • Because group assignments depend on identity features shared across individuals, the learned organization transfers across subjects and sessions without retraining or neuron-wise correspondence.The framework processes identity–activity pairs into a population representation.Source: §2 · p. 3, §2.1 · p. 3
  • Each modality is encoded separately and combined into an identity embedding that produces soft memberships, allowing a neuron to contribute to multiple latent groups.Modality-specific encoders address heterogeneous feature types and measurement scales before projection to a shared embedding.Source: §2.1 · p. 4
  • A shared dynamics encoder maps each neuron's activity window to an embedding, and routing-weighted aggregation forms each group's activity representation.The group representations are stacked and processed by an interaction module to produce a population state for downstream objectives.Source: §2.1 · p. 4
  • Shared neuron-wise encoders and routing, followed by summation across neurons, make the group matrix and final population state invariant to neuron ordering.The passage states that a formal proof is provided in Appendix A.2.Source: §2.1 · p. 4

3 Experiments

The experiments evaluate identity-guided population analysis across two species and recording modalities, with decoding and manifold methods assessed under held-out transfer. The protocol freezes trained parameters and applies them directly to unseen recordings.

3.1 Datasets and Experimental Setup

  • The experiments span whole-brain zebrafish calcium imaging with transcriptomics and mouse visual-cortex Neuropixels electrophysiology, covering distinct species, modalities, and identity-feature availability.The zebrafish dataset includes structured visual stimuli and fictive-swimming behavior; the mouse data include visual stimuli and running speed.Source: §3 · p. 4, §3.1 · pp. 5–6

3.2 Baselines

  • Decoding comparisons include POYO, NEDS, and DeepSets, while manifold comparisons include CEBRA, MARBLE, LFADS, and Joint PCA.Decoding baselines are tested in original and identity-feature-augmented variants to separate access to features from the routing architecture.Source: §3.2 · p. 6

3.3 Evaluation metrics and protocol

  • Behavior decoding is evaluated with Pearson correlation and stimulus classification with accuracy, while manifold metrics assess temporal organization and local latent structure.The manifold measures include Temporal Distance Correlation, Normalized Latent-Neighbor Time Gap, and Local Trajectory Linearity.Source: §3.3 · p. 6
  • All methods are trained on source animals or sessions, then applied to complete held-out recordings with learned parameters frozen and held-out labels reserved for evaluation.The WARP protocol holds out one fish at a time; mouse models train on 20 sessions and evaluate on the remaining sessions.Source: §3.3 · pp. 6–7

4 Results

Under direct transfer to held-out animals or sessions, the framework leads decoding across the evaluated datasets and preserves temporal and local geometric structure in latent trajectories. Identity-feature ablations remain competitive, but the full model performs better.

4.1 Behavior and Stimulus Decoding Under Direct Held-Out Transfer

  • The framework achieves the best behavior and stimulus decoding across all three datasets under held-out transfer without retraining.Its behavior PCC is 0.793 on WARP and 0.734 and 0.694 on the Allen Brain Observatory and Functional Connectivity datasets, respectively.Source: §4.1 · p. 7
  • Providing biological features improves POYO, NEDS, and DeepSets, but their identity-augmented versions still underperform the proposed method across datasets.On WARP behavior decoding, POYO rises from 0.304 to 0.489, NEDS from 0.248 to 0.574, and DeepSets from 0.160 to 0.466.Source: §4.1 · p. 7
  • Removing transcriptomic descriptors leaves the model competitive: behavior PCC is 0.727 and stimulus balanced accuracy is 0.711, versus 0.793 and 0.835 for the full model.All tested single-modality ablations remain competitive and outperform baseline methods.Source: §4.1 · p. 7

4.2 Neural Manifold Geometry Under Direct Held-Out Transfer

  • On held-out WARP, temporal distance correlation is 0.633 ± 0.078, normalized nearest-neighbor temporal gap is 0.145 ± 0.009, and local trajectory linearity is 0.911 ± 0.006.The framework also outperforms all baselines across evaluated geometric metrics on the Allen datasets; Table 2 covers held-out recordings across three datasets.Source: §4.2 · p. 8, Table 2 · p. 8
  • The framework's transferred trajectories have more consistent temporal neighborhoods and smoother local structure, while baseline representations are tied to source neurons and lack alignment when populations change.The proposed method grounds neurons in a shared identity-defined group coordinate.Source: §4.2 · p. 8

5 Related Work

Prior approaches learn cross-recording representations from neural activity or characterize low-dimensional population manifolds. Multimodal profiling and cell atlases provide another source of shared biological structure for organizing population analyses.

Neural population dynamics analysis methods

  • POYO, POCO, and NEDS support cross-session or cross-task decoding through learned representations, but rely on observed dynamics and implicit indexing rather than explicit biological identity [9] [19] [31].Work on learning identity from population context has begun, while cross-session generalization still primarily relies on statistical regularities in neural dynamics [10].Source: §5 · p. 8

Neural manifold methods

  • Neural manifold methods use low-dimensional geometry to relate population activity to cognition, perception, and movement.Approaches include sequential variational autoencoders, contrastive alignment with behavioral or stimulus variables, and geometric deep learning over local flow fields.Source: §5 · pp. 8–9

Multimodal neuron profiling and cell-type atlases

  • Patch-seq, spatial transcriptomics, and cellular-resolution atlases measure molecular and spatial properties, while newer datasets align neural activity with transcriptomic identity.Across species, molecular and spatial identity are associated with functional circuit organization.Source: §5 · p. 9

6 Discussion

The discussion presents multimodal biological features as a shared organizational basis for transferring population representations across animals and sessions. Results support this approach across decoding and manifold analysis, while its performance depends on identity-feature quality and availability.

6 Discussion

  • Routing neurons by multimodal biological features organizes population dynamics into shared latent groups while preserving permutation invariance and enabling transfer without neuron-wise correspondence or retraining.Populations from new animals or sessions can be mapped into a common identity-defined structure.Source: §6 · p. 9
  • Across zebrafish and mouse visual cortex, the method combines strong behavior and stimulus decoding with geometrically consistent manifolds under held-out transfer.Baselines also improve with identity features, but the routing-based framework consistently outperforms them.Source: §6 · p. 9
  • The method performs comparably with independently acquired atlas data, extending its applicability to datasets without neuron-level feature measurements in the same animals.This claim is supported by the discussion of atlas-based features and data lacking jointly collected identity measurements.Source: §6 · pp. 9–10
  • Performance depends on identity-feature availability and quality; individual variability and registration imperfections can add noise to features and affect group assignments.The discussion identifies joint data acquisition and improved multimodal registration as avenues to strengthen identity-based analysis.Source: §6 · p. 10
  • The results position neuronal identity as a complementary shared biological axis for organizing population representations across individuals without exact neuron-wise correspondence.The discussion describes this as a path toward integrating data across subjects.Source: §6 · p. 10

A.1 Dataset overview

The dataset overview shows comparable spatial and regional organization across WARP fish and Allen visual cortex sessions, motivating transfer without neuron-wise correspondence. It also specifies the supervised decoding targets for each dataset.

A.1 Dataset overview

  • WARP fish share broadly similar whole-brain distributions of transcriptomic markers and brain-region labels, despite differing recorded neurons.Figure A1 maps neurons from three fish in registered brain coordinates, colored by marker identity, anatomical region, and selected marker-gene categories.Source: §A.1 · p. 14, Figure A1 · p. 14
  • Representative Allen visual cortex sessions cover comparable areas in a shared anatomical coordinate space, despite differing recorded units.Figure A2 colors the sessions by visual area label.Source: §A.1 · p. 14, Figure A2 · p. 14
  • The datasets use scalar-regression behavior targets and multi-class stimulus targets: swim power and visual-stimulus type for WARP, and running speed and visual-stimulus type for Allen.Behavior decoding is scored by Pearson correlation coefficient and stimulus decoding by accuracy; these targets are shared by the proposed method and supervised baselines.Source: §A.1 · p. 14

A.2 Proof of Permutation Invariance

The proof makes the arbitrary enumeration of the neural population explicit and establishes permutation invariance for the identity-guided population encoder when its neuron-level components are shared.

A.2 Proof of Permutation Invariance

  • The proof explicitly represents the population's arbitrary neuron enumeration and tracks how permuting inputs permutes neuron-level quantities.It then shows the effective group mass and resulting group matrix are unchanged, so the deterministic final representation is unchanged.Source: §A.2 · pp. 15–16
  • The population encoder is permutation-invariant when the identity encoder, dynamics encoder, and routing map are shared across neurons.The proposition defines the encoder through identity-guided routing, shared dynamics encoding, group-level aggregation, and a group interaction module.Source: §A.2 · p. 15

A.3 Neuronal Identity Feature Construction

The identity encoder is designed to accept whichever neuron-level feature modalities are available, with feature availability varying across datasets. For WARP, spatial locations are registered to a reference atlas and transcriptomic and atlas-derived HCR features are available; Allen lacks joint transcriptomic measurements.

A.3 Neuronal Identity Feature Construction

  • The identity encoder is designed to accommodate varying sets of neuron-level feature modalities across datasets and experimental conditions.The feature-group table summarizes which groups are used in each dataset; for WARP, spatial coordinates are registered to mapzebrain, while Allen coordinates use the Allen Common Coordinate Framework.Source: §A.3 · p. 16, Table A1 · p. 16
  • Allen lacks joint transcriptomic measurements, and its spatial coordinates are registered to the Allen Common Coordinate Framework.The available identity modalities therefore vary between the Allen and WARP datasets.Source: §A.3 · p. 16
  • For WARP, jointly acquired transcriptomic markers and atlas-derived HCR expression features complement spatial coordinates registered to mapzebrain.HCR features are sampled from 236 marker volumes at each registered neuron location using a local Gaussian-weighted average.Source: §A.3 · p. 16

A.4 Geometric Trajectory Metrics

Three metrics assess whether transferred latent trajectories preserve temporal organization and local structure: temporal distance correlation, normalized latent-neighbor time gap, and local trajectory linearity.

A.4 Geometric Trajectory Metrics

  • Each metric is computed separately for each held-out trajectory and then averaged across held-out animals or sessions.The trajectories are latent representations of held-out recordings.Source: §A.4 · p. 17
  • Temporal distance correlation measures the Spearman correlation between latent distances and temporal gaps for sampled frame pairs; higher values indicate temporal separation is reflected in latent geometry.The metric samples 20,000 valid frame pairs with offsets up to 100 frames.Source: §A.4 · p. 17
  • Normalized latent-neighbor time gap averages each frame's temporal distance to its 10 nearest latent neighbors, divided by recording length; lower values indicate closer temporal neighborhoods.The metric excludes the frame itself when finding neighbors and is scaled to [0, 1].Source: §A.4 · p. 17
  • Local trajectory linearity measures how much variance in short latent-trajectory windows lies along the first principal direction; higher values indicate more locally one-dimensional dynamics.It uses 25-frame sliding windows with a stride of 5 and averages the window scores.Source: §A.4 · p. 17

A.5 Implementation Details for the Proposed Method

The appendix specifies the proposed model and baseline implementations, held-out evaluation, and analyses of identity features, routing, and manifold transfer. Across these analyses, identity generally improves decoding baselines, while the proposed method generally outperforms the best of 1,000 sampled baseline neuron orderings on manifold metrics.

A.5 Implementation Details for the Proposed Method

  • The proposed method routes neurons using identity features, aggregates activity into latent groups, and processes group representations with an interaction module; its output dimensions and objectives vary by task.For manifold learning, it produces a three-dimensional normalized latent vector; for decoding, outputs are task-specific. Held-out evaluation uses fixed model parameters and no held-out labels for training, fine-tuning, or model selection.Source: §A.5 · pp. 17–18

A.6 Baseline methods for decoding

  • Decoding baselines include DeepSets, POYO, and NEDS; POYO and NEDS retain their original architectural designs while their inputs are adapted to the recording modality.The appendix also lists decoding-baseline hyperparameters, including batch size 64 and dataset-specific POYO and NEDS learning rates.Source: §A.6 · pp. 18–19, §A.7 · p. 20

A.8 Manifold Baseline Implementation Details

  • Manifold comparisons use three-dimensional baseline embeddings under the held-out transfer protocol, with source-trained encoders or projections applied to held-out recordings.The appendix also visualizes source and held-out fish trajectories colored by behavior and stimulus labels.Source: §A.8 · p. 20, §A.9 · p. 20

A.10 Effect of Neuronal Identity on Decoding Baselines

  • Adding neuronal identity generally improves baseline decoding across datasets, with gains also observed for NEDS and POYO in held-out transfer.Baseline models are trained on source data and evaluated on held-out recordings without fine-tuning; Table A4 reports behavior and stimulus decoding under this protocol.Source: §A.10 · p. 20, Table A4 · p. 22

A.11 Ablation Studies

  • On WARP, removing most single identity modalities leaves performance comparable to the all-modality configuration, while Allen decoding remains effective under single-modality removals.The authors interpret the Allen result as showing that the available spatial, region, and electrophysiological feature set supports transfer without molecular identity features.Source: §A.11 · pp. 21–22, Table A5 · p. 22, Table A6 · p. 22
  • Replacing soft routing with hard top-1 assignment reduces behavior-decoding performance, consistent with representing neurons' soft identity-to-group assignments.Table A7 varies one architectural choice at a time and reports behavior-decoding results.Source: §A.11 · pp. 22–23, Table A7 · p. 23

A.12 Best-Ordering Analysis for Manifold Baselines

  • Across manifold baselines and metrics, the proposed method generally outperforms the best result among 1,000 sampled target-neuron orderings.This comparison indicates that the gap is not explained by a single unfavorable ordering for the baselines.Source: §A.12 · p. 23, Figure A6 · p. 23, Figure A7 · p. 24

A.13 Compute Resources

  • The reported compute covers the main decoding and latent-trajectory experiments, including the proposed method, ablations, and evaluated baselines; exploratory development runs are excluded.The reproduction table reports wall-clock compute, with GPU-hours estimated from wall-clock time and the number of GPUs used.Source: §A.13 · p. 25, Table A8 · p. 25
Loading 2610.36745v1…