PaperlayerView exact revision on arXiv

Source-linked AI summary

Connectome-to-Function: Conditional Generative Latent Representations for Reservoir Computing

Zhuolin Yu, Xingyu Liu, Yuanhao Jia, Yunhang Xiao, Hairuo Xue, Feihan Sun, Guozhang Chen

arXiv:2609.06093v1cs.LG

TL;DR

Mapping connectome structure to computation remains difficult, and existing approaches do not provide a generative representation that also organizes functional variation. This paper learns a conditional latent representation from biological connectivity and finds that its coordinates predict reservoir performance, with task-specific structural associations.

  • Problem

    It remains difficult to represent connectomes in a generative form that also relates biological wiring to computational function.

    Source: Introduction · p. 1

  • Method

    A conditional generative framework encodes connectome structure using neuronal cell type and spatial organization as conditions for reconstruction and generation.

    Source: Introduction · p. 2, Conditional VAE for Connectome · p. 3

  • Results

    The latent space predicts reservoir-performance variation across memory, dynamic prediction, and classification, with memory associated with reciprocal connectivity and prediction and classification with spectral properties.

    Source: Conclusion · p. 8

  • Takeaways & Limitations

    Sparse connectome graphs can be organized into a latent space connecting circuit topology with measured computational behavior.

    Source: Conclusion · p. 8

    Experiments use local cortical microcircuits from MICrONS, and generalization to other brain regions, species, or spatial scales remains unestablished.

    Source: Conclusion · p. 8

Abstract

from arXiv · show

Connectomes, graph-level maps of neurons and their synaptic connections, provide a structural basis for understanding how brain circuits support function and computation. However, mapping connectome structure to computation remains difficult because these graphs are high-dimensional, sparse, and sensitive to local structural variation. Existing approaches often depend on hand-crafted structural descriptors or task-specific predictors, which limits their ability to represent connectomes in a form that is both generative and functionally meaningful. We propose a conditional generative latent framework that encodes connectome graphs into a compact structural space while using available node-level conditions to guide reconstruction and generation. From this space, the model can reconstruct observed connectivity with a mean edge-reconstruction AUC up to 0.910 and generate new candidate connectomes, enabling a unified analysis of graph structure and computational behavior. Using connectome-derived graphs as recurrent computational substrates, we found that the learned latent space captures functional variation across reservoir-computing experiments, with cross-validated $R^2$ values up to approximately 0.87. Interpretability analysis further revealed task-specific structural mechanisms: in our examples, memory performance is associated with reciprocal recurrent connectivity, whereas prediction and classification are more strongly associated with spectral properties of the recurrent network. These findings suggest an AI-for-science approach to linking neural connectivity to computation and provide a generative and interpretable basis for studying how distinct structural mechanisms shape computational capacity.

Introduction

The paper asks whether measured connectome structure can be organized into a continuous representation that helps explain variation in circuit computation. It proposes a conditional generative latent framework and evaluates whether its coordinates relate to reservoir performance across tasks.

Introduction

  • A key gap is a generative representation learned from biological connectivity alone that can reveal computational organization without using function to train the representation.Prior work either emphasizes structural reproduction, task-trained artificial networks, or fixed connectomes and rewired variants.Source: Introduction · p. 2
  • The framework conditions connectome generation on neuronal cell type and spatial organization, supporting reconstruction and controlled sampling of candidate connectomes.It encodes local circuits into compact structural coordinates.Source: Introduction · p. 2
  • The learned latent coordinates predict reservoir performance across memory, dynamic prediction, and classification tasks.The reported cross-validated R2 values are approximately 0.46–0.87.Source: Introduction · p. 2
  • Interpretability analysis associates memory with reciprocal recurrent connectivity, while prediction and classification are more strongly associated with recurrent-network spectral properties.These task-specific mechanisms link functional directions in the learned space to distinct structural descriptors.Source: Introduction · p. 2

Related Work

The paper situates its approach between descriptive graph statistics, mechanistic graph generators, and connectome-based reservoir studies. It then describes a conditional graph representation and tests its structural fidelity and relation to reservoir function.

Related work

  • Existing descriptors are predefined summaries, mechanistic generators lack continuous data-driven latent spaces, and general graph generators are difficult to apply to sparse neural microcircuits with spatial and cell-identity features.Prior connectome-reservoir studies mainly evaluate fixed empirical networks or simple rewired variants rather than learning a generative latent space.Source: Related Work · p. 2

Problem setup

  • The model represents each local circuit as an attributed directed graph, with node annotations comprising neuronal soma coordinates and one-hot neuron types.The latent variable encodes graph structure conditioned on observed node features.Source: Conditional VAE for Connectome · p. 3

Architecture

  • The encoder combines spatial and node attributes with adjacency information, then uses a permutation-invariant graph-level representation to parameterize the latent distribution.PointNet++ fuses geometric features, a GAT integrates adjacency, and a Transformer-based encoder uses a learned CLS token.Source: Conditional VAE for Connectome · p. 4
  • The decoder combines node-condition embeddings with the latent code and predicts directed edge probabilities that are sampled into binary adjacency matrices.Decoded representations are used to generate candidate connectomes under node-level conditions.Source: Conditional VAE for Connectome · p. 4

Connectome-based reservoir computing

  • Connectome graphs are evaluated as fixed Dale-signed reservoirs, with recurrent connectivity normalized to a fixed spectral radius and only the readout fitted to task targets.The evaluation uses recurrent state updates and linear readout; the recurrent matrix and random input weights remain fixed during training.Source: Linking Latent Structure to Function · p. 4

Latent-to-function regression

  • To test latent-to-function prediction, the analysis decodes latent codes under shared node conditions and evaluates generated graphs as reservoirs; reliable cross-validated prediction is taken as evidence that the representation contains function-informative structural variation.The shared condition template is repeated across 10 choices, and the functional score averages 10 generated adjacency matrices across 3 random seeds per latent code.Source: Linking Latent Structure to Function · p. 4

Reconstruction and latent organization

  • Figure 3 presents representative original, naive-VAE, and full-CVAE adjacency matrices alongside a PCA view of latent vectors colored by graph descriptors.Table 1 compares full-CVAE and naive-VAE reconstruction using edge-level AUC and graph-topology relative errors; the supplied caption does not give cell values.Source: Figure 3 · p. 5, Table 1 · p. 5

Experiments

Experiments test how connectome graph structure is represented and how latent variation relates to reservoir performance across three tasks. They report topology preservation and task-specific structural associations.

Conclusion

The framework links compact connectome representations to reservoir-computing performance and task-specific structural associations. The experiments are limited to MICrONS local cortical microcircuits and reservoir computing as a proxy for circuit computation.

Conclusion

  • The latent space predicts functional variation across memory, dynamic prediction, and classification, with reciprocal-core associations for memory and Dale-signed spectral associations for prediction and classification.The memory-associated cores are mixed excitatory–inhibitory; the spectral properties are global.Source: Conclusion · p. 8
  • The framework learns compact connectome latent representations conditioned on neuronal cell type and spatial organization.Generated connectomes are used as recurrent reservoirs to evaluate computational behavior.Source: Conclusion · p. 8
  • The experiments use MICrONS local cortical microcircuits, and generalization to other brain regions, species, or spatial scales remains unestablished.Reservoir computing is a controlled, interpretable proxy that does not capture the full range of biological neural dynamics or learning mechanisms; synthesized circuits have not been experimentally validated.Source: Conclusion · p. 8
Loading 2609.06093v1…