# Decoding brain structure-function dynamics in health and in psychosis via an autoencoder

## TL;DR

This paper uses a deep-learning model to better understand how brain structure relates to function in health and psychosis, showing stronger connections in certain brain activity bands.

## Contribution

A deep-learning framework using a Graph Multi-Head Attention AutoEncoder outperforms analytical models in predicting functional connectivity from structural brain data.

## Key findings

- The deep-learning model achieved mean correlation coefficients above 0.8 in alpha and beta bands for healthy participants.
- Structure-function coupling in psychosis differs significantly from healthy individuals, with pronounced changes in the model's predictions.
- Analytical models better capture delta and theta band dynamics in psychosis, while deep-learning models are needed for alpha and beta bands.

## Abstract

Understanding the intricate relationship between brain structure and function is a cornerstone challenge in neuroscience, critical for deciphering the mechanisms that underlie healthy and pathological brain function. In this work, we present a comprehensive framework for mapping structural connectivity measured via diffusion-MRI to resting-state functional connectivity measured via magnetoencephalography, utilizing a deep-learning model based on a Graph Multi-Head Attention AutoEncoder. We compare the results to those from an analytical model that utilizes shortest-path-length and search-information communication mechanisms. The deep-learning model outperformed the analytical model in predicting functional connectivity in healthy participants at the individual level, achieving mean correlation coefficients higher than 0.8 in the alpha and beta frequency bands, in comparison to 0.45 for the analytical model. Our results imply that human brain structural connectivity and electrophysiological functional connectivity are tightly coupled. The two models suggested distinct structure-function coupling in people with psychosis compared to healthy participants (\documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$p < 2\times 10^{-4}$$\end{document} for the deep-learning model, \documentclass[12pt]{minimal}
\usepackage{amsmath}
\usepackage{wasysym}
\usepackage{amsfonts}
\usepackage{amssymb}
\usepackage{amsbsy}
\usepackage{mathrsfs}
\usepackage{upgreek}
\setlength{\oddsidemargin}{-69pt}
\begin{document}$$p < 3\times 10^{-3}$$\end{document} in the delta band for the analytical model). Importantly, the alterations in the structure-function relationship were much more pronounced than any structure-specific or function-specific alterations observed in the psychosis participants. The findings demonstrate that analytical algorithms effectively model communication between brain areas in psychosis patients within the delta and theta bands, whereas more sophisticated models are necessary to capture the dynamics in the alpha and beta band.