Research map / Probabilistic and scientific modeling / Dynamical systems
Accurate Identification of Communication Between Multiple Interacting Neural Populations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MR-LFADS, a multi-region sequential VAE that disentangles inter-regional communication, unobserved inputs, and local dynamics, and shows better communication recovery than RRR, mp-srSLDS and MR-SDS on synthetic multi-region RNNs and predicts held-out photoinhibition effects in mouse electrophysiology.
Problem
Existing data-driven communication models struggle to separate inter-regional communication from unrecorded-region inputs and local population dynamics, yielding inaccurate portraits of communication even when reconstruction is good.
Approach
Stack per-region LFADS-style sequential VAEs with communication messages computed from reconstructed firing rates (data-constrained), automatically inferred unobserved inputs, and a structured KL bottleneck that penalizes unobserved inputs more heavily (klu = 10 klm) so communication is preferred when it can explain the data.
Key results
- On the memory network, MR-LFADS-R infers the most accurate effectome (S_cos reported in Experiment 1) while other models infer no communication or lagged communication.
- In Experiment 3 across 35 random multi-region networks, MR-LFADS-R was the only model performing best on both message R^2 directions and matched top effectome accuracy with MR-LFADS-G.
- Applied to mouse Neuropixels data, MR-LFADS predicted photoinhibition effects on held-out trials, with Thal(A) most affected and MRN least affected; message content was more consistent across seeds for -R than -G.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Automatically infer unobserved inputs instead of manually specifying all stimulus signals (ablation) | MR-LFADS · Memory network | effectome cosine similarity | Helped |
| Use manually specified external inputs instead of automatically inferred inputs (ablation) | MR-LFADS · Memory network | effectome cosine similarity | Hurt |
| Constrain communication to reconstructed firing rates rather than factors or generator states (ablation) | MR-LFADS · Pass-decision network | effectome cosine similarity | Helped |
| Use factor-based communication instead of firing-rate-based communication (ablation) | MR-LFADS · Pass-decision network | effectome cosine similarity | Hurt |
| Use generator-based communication instead of firing-rate-based communication (ablation) | MR-LFADS · Pass-decision network | effectome cosine similarity | Hurt |
| Use firing-rate-based communication rather than generator-based communication (ablation) | MR-LFADS · Random multi-region networks | message-content consistency across seeds | Helped |
| Use firing-rate-based communication rather than generator-based communication (ablation) | MR-LFADS · Random multi-region networks | effectome cosine similarity | No clear effect |
| Use data-constrained communication inferred from reconstructed firing rates (ablation) | MR-LFADS · 35 randomly generated multi-region networks | message-content decoding R² | Helped |
| Ablate outgoing messages from ALM to simulate photoinhibition | MR-LFADS(R) · Multi-region electrophysiology, held-out ALM photoinhibition trials | prediction of brain-wide perturbation effects | Helped |
| Use MR-LFADS(R) rather than MR-LFADS(G) (ablation) | MR-LFADS · Multi-region electrophysiology | message-content consistency across seeds | Helped |
Surprising findings
- Manually specifying external inputs can cause models to forgo communication entirely, relying on the specified inputs instead.
Limitations
- Sensitive to hyperparameter settings, requiring significant compute for tuning
- Cannot resolve correlation-causation confounds from passively observed activity when an unrecorded region drives multiple recorded regions
- Gaussian prior and posterior assumptions may mismatch non-Gaussian inputs
Experimental setup
Datasets: 37 synthetic multi-region datasets from task-trained RNN data-generating networks multi-region Neuropixels recordings from mice performing decision-making task (ALM, MRN, SC, thalamus)
Benchmarks: 37 synthetic multi-region datasets (memory, pass-decision, random multi-region networks) held-out photoinhibition effects in mouse ALM electrophysiology
Metrics: R^2 of held-out neuron reconstruction cosine similarity of effectomes (S_cos) cross-validated R^2 of message and input recovery correlation of inferred message norms across seeds
Baselines: RRR mp-srSLDS MR-SDS MR-LFADS-F MR-LFADS-G MR-LFADS-S
Builds on: LFADS reduced-rank regression communication subspace mp-srSLDS MR-SDS
Closest papers by meaning
- Multi-Region Markovian Gaussian Process: An Efficient Method to Discover Directional Communications Across Multiple Brain Regions (ICML 2024 · similarity 0.88)
- Identifying interactions across brain areas while accounting for individual-neuron dynamics with a Transformer-based variational autoencoder (NeurIPS 2025 · similarity 0.88)
- Learning Time-Varying Multi-Region Brain Communications via Scalable Markovian Gaussian Processes (ICML 2025 · similarity 0.88)
- Inpainting the Neural Picture: Inferring Unrecorded Brain Area Dynamics from Multi-Animal Datasets (NeurIPS 2025 · similarity 0.87)
- CREIMBO: Cross-Regional Ensemble Interactions in Multi-view Brain Observations (ICLR 2025 · similarity 0.86)
- eXponential FAmily Dynamical Systems (XFADS): Large-scale nonlinear Gaussian state-space modeling (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.