MR-LFADS model disentangles brain region communication with 10x accuracy

A team of researchers (Belle Liu, Jacob Sacks, Matthew D. Golub) has published a paper on arXiv introducing MR-LFADS (Multi-Region Latent Factor Analysis via Dynamical Systems), a novel sequential variational autoencoder designed to untangle the complex communication between multiple interacting neural populations. Modern neural recording technologies can simultaneously capture activity from many brain regions, but existing statistical models struggle to separate the influences that drive that activity—such as direct inter-regional signals, local dynamics, and inputs from unobserved areas. MR-LFADS addresses this by explicitly modeling each component, resulting in a disentangled representation of brain-wide interactions.

In extensive simulations of task-trained multi-region networks, MR-LFADS consistently outperformed existing approaches at identifying true communication patterns. The researchers also validated the model on large-scale electrophysiology data, where it accurately predicted the brain-wide effects of circuit perturbations that were not included during training. These results, presented at the Forty-second International Conference on Machine Learning (2025), position MR-LFADS as a promising tool for discovering principles of neural communication. The work has implications for understanding brain disorders, designing brain-computer interfaces, and building more biologically realistic AI systems.