[Paper Review] Interpretable AI for relating brain structural and functional connectomes

Haoming Yang, Steven L. Winter+2 more|arXiv (Cornell University)|Oct 10, 2022

Functional Brain Connectivity Studies4 citations

TL;DR

This paper proposes Staf-GATE, a graph auto-encoder that jointly models structural (SC) and functional (FC) connectomes using a variational autoencoder framework with graph k-nearest neighbor layers to incorporate topological structure. It achieves state-of-the-art performance in predicting FC from SC and introduces a masking-based perturbation method to extract interpretable insights into SC-FC coupling, revealing sex-differentiated subnetworks critical for functional connectivity.

ABSTRACT

One of the central problems in neuroscience is understanding how brain structure relates to function. Naively one can relate the direct connections of white matter fiber tracts between brain regions of interest (ROIs) to the increased co-activation in the same pair of ROIs, but the link between structural and functional connectomes (SCs and FCs) has proven to be much more complex. To learn a realistic generative model characterizing population variation in SCs, FCs, and the SC-FC coupling, we develop a graph auto-encoder that we refer to as Staf-GATE. We trained Staf-GATE with data from the Human Connectome Project (HCP) and show state-of-the-art performance in predicting FC and joint generation of SC and FC. In addition, as a crucial component of the proposed approach, we provide a masking-based algorithm to extract interpretable inferences about SC-FC coupling. Our interpretation methods identified important SC subnetworks for FC coupling and relating SC and FC with sex.

Motivation & Objective

• To develop a generative model that captures the joint distribution of structural and functional connectomes across a population.
• To improve FC prediction from SC by integrating graph topology features through a k-nearest neighbor layer architecture.
• To provide interpretable inference into SC-FC coupling using a perturbation-based masking algorithm.
• To investigate sex differences in SC-FC coupling by identifying subnetworks that are most predictive of functional connectivity.
• To enable probabilistic inference and simulation of paired SC-FC data for downstream neuroscience applications.

Proposed method

• Staf-GATE employs a variational autoencoder with an encoder that maps individual SCs to latent variables, capturing population-level variation.
• The decoder uses graph k-nearest neighbor layers to reconstruct SCs from latent variables, modeling the conditional probability P(SC|z) via a Poisson latent space.
• A predictive generator infers FC from the same latent representation, enabling joint generation of SC and FC.
• A masking-based perturbation algorithm is applied to identify critical SC edges by measuring their impact on FC prediction performance.
• The method uses bootstrapping to generate null distributions for statistical significance testing of identified subnetworks.
• The model is trained on Human Connectome Project (HCP) data, leveraging large-scale population connectome data for robust learning.

Experimental results

Research questions

• RQ1 Which subnetworks of the structural connectome are most predictive of functional connectivity, and how do they vary across individuals?
• RQ2 How does the inclusion of graph topology features improve the accuracy of functional connectome prediction from structural connectomes?
• RQ3 What are the sex-specific differences in SC-FC coupling, and which structural connections are most critical for predicting functional connectivity in males versus females?
• RQ4 Can a deep generative model like Staf-GATE realistically simulate joint SC-FC data while preserving network topology and population-level variation?
• RQ5 How can interpretable inference be systematically extracted from complex, non-linear deep learning models in connectomics?

Key findings

• Staf-GATE achieves state-of-the-art performance in predicting functional connectivity from structural connectomes, with a group-average correlation of 0.9 and individual-average correlation of 0.55.
• The masking-based interpretation method identifies specific subnetworks of the structural connectome that are highly influential in predicting functional connectivity, revealing non-linear SC-FC relationships.
• The method detects significant sex differences in SC-FC coupling, identifying distinct sets of structural connections that are most predictive of functional connectivity in males and females.
• The model successfully generates high-fidelity SC and FC pairs that preserve the topological structure of the training data, enabling realistic simulation of connectome data.
• The perturbation-based interpretation algorithm is generalizable and can be applied to other deep learning models to analyze SC-FC coupling outcomes.
• Despite strong performance, the model shows limited ability to predict individual-level FC from SC, suggesting SC alone may not contain sufficient information for precise individual prediction.