JuliaCon 2025

Dynamic Causal Modeling (DCM) is a framework for inferring neural connectivity from neuroimaging data based on variational Bayes estimation techniques. Among the most widely used techniques is spectral DCM, which provides a fast parameter estimation technique in the spectral domain, allowing to fit biophysically detailed models.Here we present a novel implementation in the Julia programming language of spectral DCM, leveraging its high-performance computational capabilities. We validate and benchmark our implementation using simulated and empirical fMRI data and compare our implementation against the DCM implementation of Statistical Parametric Mapping (SPM), which is MATLAB based and the most commonly used DCM software.
In particular we employed automatic differentiation to gain up to tenfold speed enhancements over the SPM implementation while maintaining parameter estimation accuracy. Since computational cost is considered a practical limitation of the application of DCM, this improvement makes the application of this method amenable to the inference of connectivities among more measured brain regions. We further employ Julia's ModelingToolkit for modular model assembly which allows a much more flexible modeling environment as the one provided by SPM whose extension to non pre-implemented models is laborious and error prone.
By enabling flexible model specification and extending spectral DCM's applicability through substantial speed improvements, this work provides an open-source platform for advancing neural connectivity research and parameter fitting in neuroscience.

Paper:
David Hofmann, Anthony G. Chesebro, Chris Rackauckas, Lilianne R. Mujica-Parodi, Karl J. Friston, Alan Edelman, and Helmut H. Strey. “Leveraging Julia’s Automated Differentiation and Symbolic Computation to Increase Spectral DCM Flexibility and Speed.” bioRxiv: The Preprint Server for Biology, 2023. https://doi.org/10.1101/2023.10.27.564407.