(7/7) We see MR-LFADS as a promising new tool for discovering principles of brain-wide information processing from large-scale neural datasets. Feedback, questions, or ideas are welcome!

And here’s a great blog post overview by Kristine White @uwcse.bsky.social : tinyurl.com/3h9yd6uu

(6/7) 🧪 In real electrophysiology data (Neuropixels in mice), MR-LFADS predicts brain-wide effects of circuit perturbations (ALM photoinhibition) that it had never seen during training!

(5/7) 🔍 We evaluate on 37 synthetic multi-region datasets, covering challenging and diverse scenarios. MR-LFADS reliably recovers both communication pathways (“effectomes”) and content, outperforming prior methods.

(4/7)
Communication from data-constrained inferred firing rates, rather than from overly flexible latent factors
Unsupervised inference of unobserved influences
Region-specific nonlinear dynamics modules
Structured KL bottlenecks support disentangling of all above

(3/7) Our contribution: Multi-Region LFADS (MR-LFADS), a sequential-VAE built on the powerful LFADS framework (Nature Methods, 2018), with 4 key design features that support accurate identification of single-trial communication:

(2/7) When modeling multi-region neural recordings, we found it difficult yet critically important to disentangle communication between recorded regions, unobserved influences (e.g., unrecorded regions), and local-region neural population dynamics.