Unveiling and Steering Connectome Organization with Interpretable Latent Variables

TL;DR

This work addresses how to extract compact, interpretable representations of complex brain connectomes and use them to reconstruct and controllably generate subgraphs. It combines functionally guided subgraph sampling with a graph variational autoencoder, an SHAP-based explainability module, and optimization tools (dynamic programming and CMA-ES) to map latent variables to biologically meaningful graph features and to synthesize connectomes with predefined properties. The approach achieves accurate graph reconstruction, reveals interpretable latent factors linked to topology (e.g., edge-count, reciprocity, betweenness, non-neuronal content), and demonstrates controllable generation of target subgraphs on FlyWire, offering insights into neural circuit design and a blueprint for bio-inspired AI. The framework advances connectome analysis by enabling low-dimensional reasoning about structure–function principles and provides practical pathways for designing artificial networks with brain-inspired architectures.

Abstract

The brain's intricate connectome, a blueprint for its function, presents immense complexity, yet it arises from a compact genetic code, hinting at underlying low-dimensional organizational principles. This work bridges connectomics and representation learning to uncover these principles. We propose a framework that combines subgraph extraction from the Drosophila connectome, FlyWire, with a generative model to derive interpretable low-dimensional representations of neural circuitry. Crucially, an explainability module links these latent dimensions to specific structural features, offering insights into their functional relevance. We validate our approach by demonstrating effective graph reconstruction and, significantly, the ability to manipulate these latent codes to controllably generate connectome subgraphs with predefined properties. This research offers a novel tool for understanding brain architecture and a potential avenue for designing bio-inspired artificial neural networks.

Figures (9)

• Figure 1: Model architecture: (a) shows the encoder module which takes the graph adjacency matrix $G$ and node categories $C$ as input and outputs a first-order low-dimensional representation vector; (b) presents the decoder that reconstructs the graph $(\hat{G}, \hat{C})$ from the latent representation;(c) illustrates the surrogate model designed for SHAP analysis, which takes graph $G$ as input and outputs the approximated feature $y'$.
• Figure 2: Generation results on FlyWire in Adjacency Matrix Format. (a) Original: Shows the sampled graphs from the Drosophila brain connectome with three distinct divergence levels. (b) Ours: Our model's generation results, also exhibiting three divergence levels. (c)-(d) EDGE and GruM results: While performing well on high-divergence graphs, they lack medium and low-divergence samples, indicating limited diversity. (e)-(f) GDSS and DisCo results: Their main limitations are insufficient edge generation and similarly constrained diversity.
• Figure 3: Analysis of Edge Count.SHAP value results are shown in (a), where the y-axis represents latent dimensions arranged from top (most influential) to bottom (least influential). Point colors indicate magnitude - red for larger values and blue for smaller values. The x-axis shows the direction and strength of impact, with positive values indicating positive influence and negative values indicating negative influence. (b) displays the relationship between each dimension's values and the statistical features of generated graphs. (c) presents the dynamic programming results, with the x-axis showing desired target values and the y-axis showing the actual statistical values of graphs generated from the DP-optimized low-dimensional representations.
• Figure 4: Reconstruction Capability Versus N.Figure (a) demonstrates the variation of edge reconstruction metrics with changing values of n, while figure (b) shows the corresponding changes in node category recovery performance metrics as n changes.
• Figure 5: Model Reconstruction Capacity under Latent Code Variation.Figure (a) illustrates the variation of edge reconstruction metrics with respect to changes in latent code dimensionality, while figure (b) displays the corresponding changes in node category restoration metrics across different latent code dimensionality.