Huge shoutout to @dikshagup.bsky.social who collected this data and whose work motivated us to go down this route in the first place. Check out her recent work which excitingly aligns with our findings about ADS and FOF communication: www.biorxiv.org/content/10.1...
Stay tuned for more on this!

As a bonus: we also show theoretical equivalence under special cases with an EI-RNN, and that our model can be used to infer cell-type identity!

However, ADS silencing derails dynamics of the FOF, during both early and late stages explaining the observed biased behavior. This finding also implicates a necessary role for the feedforward connection from ADS to FOF in evidence accumulation.

This is particularly exciting as we could then use CTDS to visualize latent dynamics of both regions / cell-types and see how they change under these perturbations. Early FOF silencing has short-lived effect on dynamics of both regions, thus causing no behavioral deficit.

Remarkably, in both cases (silencing E neurons in FOF, and I neurons in ADS), predictions of our model aligned with previous experimental results (from Hanks et al. and Yartsev et al), while that of a standard LDS did not!
(Note that we did not fit the model on perturbed data.)

We next used fitted CTDS model to perform in-silico cell-specific perturbations in both regions. The biologically constrained dynamics allows us to model perturbations easily, by shutting down latents corresponding to the perturbed cell-type.

When applied to simultaneous recordings from FOF and ADS (two regions involved in evidence accumulation) in rodents during an auditory clicks task, CTDS dynamics show that both regions communicate bidirectionally, and that both E and I cells encode choice information.

Our model builds up on linear dynamical systems to contain distinct latent variables for distinct cell classes, with biologically inspired constraints on both dynamics and emission—this allows for disentangling cell-specific dynamics.
We also extend it to multi-region settings!

Paper: www.biorxiv.org/content/10.1...
Code: github.com/97aditi/Cell...