Research map / Probabilistic and scientific modeling / Dynamical systems
Disentangling the Roles of Distinct Cell Classes with Cell-Type Dynamical Systems
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 cell-type dynamical systems (CTDS), a latent linear dynamical system with separate E and I latents under Dale's-law-style sign constraints, fit by constrained EM, and applies it to rat FOF/ADS decision-making recordings to reproduce optogenetic perturbation effects.
Problem
Latent dynamical system models of neural population activity ignore cell types, which limits their ability to capture the functional roles of excitatory and inhibitory classes and to predict cell-specific perturbation effects on behavior.
Approach
Extend latent LDS with separate latent sets per cell class, constraining dynamics columns to obey Dale's law and emissions to be non-negative and block-diagonal; fit via EM with constrained quadratic programs, initialized by NNMF of an E-I RNN-derived connectivity matrix, with a multi-region extension and an approach to infer unknown cell types.
Key results
- CTDS and multi-region CTDS achieved higher held-out log-likelihood than standard LDS across latent dimensionalities on FOF-ADS data.
- In-silico silencing of FOF E latents and ADS I latents in CTDS reproduced the qualitative pattern of experimental behavioral bias, whereas the LDS with random latent silencing did not.
- CTDS could infer cell identities for unknown neurons with accuracy above chance when up to 50 of 109 neurons were masked.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add cell-type-specific latents and Dale’s-law constraints (CTDS instead of LDS) | Neural activity modeling on rat FOF and ADS recordings · Auditory decision-making task, held-out trials | log-likelihood | Helped |
| Add region-specific latent structure to CTDS | CTDS · Auditory decision-making task, held-out trials | log-likelihood | Helped |
| Use separate excitatory and inhibitory latents | Multi-region CTDS · Rat FOF and ADS, choice decoding | choice information | Helped |
| Inactivate FOF excitatory latents during the early rather than late trial half | Multi-region CTDS · FOF in-silico perturbations | behavioral bias | Mixed |
| Inactivate ADS inhibitory latents during either trial half | Multi-region CTDS · ADS in-silico perturbations | behavioral bias | Hurt |
| Replace cell-type-specific CTDS with a multi-region LDS lacking explicit cell-type structure | Perturbation modeling on rat FOF and ADS · In-silico FOF and ADS perturbations | replication of experimental behavioral effects | Hurt |
| Infer unknown cell classes from neural activity using CTDS | Multi-region CTDS · Rat FOF and ADS, masked cell identities | cell-type identification accuracy: upto 50% | Helped |
| Use CTDS rather than LDS to recover connectivity in simulated E-I networks | Connectivity recovery from simulated neural activity · Simulated E-I networks with 100 and 200 units | root mean squared error | Helped |
Surprising findings
- Standard LDS with random latent clamping recovered quickly after ADS perturbations, failing to reproduce experimental deficits.
Limitations
- Assumes linear dynamics; non-linear extension left to future work
- Only two broad cell classes (E and I) modeled, not finer cell types
- Experimental data from a single rat cannot be released
Experimental setup
Datasets: rat FOF and ADS Neuropixels recordings during auditory Poisson clicks task (Brunton et al. 2013) in-vivo optogenetic inactivation results from Hanks et al. and Yartsev et al. used for comparison
Benchmarks: held-out trial log-likelihood on FOF-ADS recordings choice decoding accuracy in-silico optogenetic perturbation behavioral bias
Metrics: test log-likelihood choice classification accuracy ipsilateral behavioral bias RMSE of recovered connectivity cell-type identification accuracy
Baselines: standard linear dynamical system (LDS) CTDS without region structure
Builds on: latent linear dynamical systems low-rank recurrent neural networks expectation maximization
Closest papers by meaning
- Modeling Neural Activity with Conditionally Linear Dynamical Systems (NeurIPS 2025 · similarity 0.87)
- Modeling Dynamic Neural Activity by combining Naturalistic Video Stimuli and Stimulus-independent Latent Factors (NeurIPS 2025 · similarity 0.87)
- BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.86)
- Learning dynamic representations of the functional connectome in neurobiological networks (ICLR 2024 · similarity 0.86)
- Probabilistic Decomposed Linear Dynamical Systems for Robust Discovery of Latent Neural Dynamics (NeurIPS 2024 · similarity 0.86)
- Meta-Dynamical State Space Models for Integrative Neural Data Analysis (ICLR 2025 · similarity 0.85)
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.