Multi-region Network of Networks Recurrent Neural Network Models of Adaptive and Maladaptive Learning

Back to Main BRAIN TMM page

PI: Rajan, Kanaka


Institution: Icahn School of Medicine at Mount Sinai

Title: Multi-region Network of Networks Recurrent Neural Network Models of Adaptive and Maladaptive Learning

Grant #: EB028166 

Status: Active


We in the Rajan lab at Mount Sinai design neural network models constrained by experimental data, and reverse engineer them to figure out how brain circuits function in health and disease. Here are two scalable, flexible, and robust tools we have developed that will find wide adoption across to the broader neuroscience community.

1) The first tool we have developed is named Current-based Decomposition (CURBD). This powerful new theory-based framework is intended for “in-vivo tract tracing” from multi-regional neural activity collected experimentally.  CURBD employs recurrent neural networks (RNNs) directly constrained, from the outset, by time series measurements acquired experimentally, such as Ca2+ imaging or electrophysiological data. Once trained, these data-constrained RNNs let us infer matrices quantifying the interactions between all pairs of modeled units. Such model-derived “directed interaction matrices” can then be used to separately compute excitatory and inhibitory input currents that drive a given neuron from all other neurons. Therefore, different current sources can be de-mixed – either within the same region or from other regions, potentially brain-wide – which collectively give rise to the population dynamics observed experimentally. Source de-mixed currents obtained through CURBD allow an unprecedented view into multi-region mechanisms inaccessible from measurements alone. We have applied this method successfully to several types of neural data from our experimental collaborators, e.g., zebrafish (Deisseroth lab, Stanford), mice (Harvey lab, Harvard), monkeys (Rudebeck lab, Sinai), and humans (Rutishauser lab, Cedars Sinai), where we have discovered both directed interactions brain wide and inter-area currents during different types of behaviors. With this powerful framework based on data-constrained multi-region RNNs and CURBD, we ask if there are conserved multi-region mechanisms across different species, as well as identify key divergences.

Perich, M. G., Arlt, C., Soares, S., Young, M. E., Mosher, C. P., Minxha, J., Rutishauser, U., Rudebeck, P. H., Harvey, C. D., Rajan, K., (2020) Untangling brain-wide current flow using data-constrained recurrent neural network models, bioRxiv, DOI:

Young, M. E., Spencer-Salmon, C., Mosher, C., Tamang, S., Rajan, K., Rudebeck, P. H., (2020) Temporally-specific sequences of neural activity across interconnected corticolimbic structures during reward anticipation, preprint submitted, bioRxiv, DOI:

Pinto, L., Rajan, K.*, DePasquale, B., Thiberge, S. Y., Tank, D. W., Brody, C. D., (2019) Task-dependent changes in the large-scale dynamics and necessity of cortical regions, Neuron, 2019 Nov 20;104(4):810-824.e9.

2) We have additionally developed a second data-inspired RNN-based tool, termed TRAKR. This tool enables us to detect state transitions from time series data. Using TRAKR, we can identify neural state transitions that are predictive of animal behavior under different experimental and experiential conditions. For example, we can pick up behavioral, neural, and circuit mechanistic changes in decision making, those that occur during shifts in attention or task engagement in multi-task paradigms, those accompanying learning of such tasks in the lab, etc. The current implementation of TRAKR works with time series data such as behavioral monitoring from pose detection methodologies, EEGs, ECoGs, LFPs, spiking neural data, and calcium fluorescence signals. This is currently a human-in-the-loop method that detects state transitions after correlation with behavior, so behavioral recordings are helpful but not necessary. There is no threshold for sampling frequency, but larger datasets are desired since RNNs are more robust when supplied with greater amounts of data. This tool is agnostic to the species, brain areas, or cell types. It is similarly oblivious to modality and can work equally well with large amounts of EEG, LFP, and even imaging time series.

Link to Data/Model Reuse abstract, [LINK] 


2021 Brain PI Meeting


In collaboration with Karl Deisseroth’s lab at Stanford University, the Rajan lab successfully constructed and trained a recurrent neural network to investigate mechanisms underlying circuit dynamics in larval zebrafish responding to an inescapable stressor (Andalman et al., 2019). Recurrent neural network models were trained on activity recorded from fish that had been exposed to a stressor or baseline activity recorded prior to the behavioral challenge. In this use case, our analysis revealed significant and specific changes in intra-habenular and raphe-to-habenular connectivity patterns as a result of the behavioral challenge. Connectivity mechanisms revealed by our computational modeling are also consistent with other prior experimental work demonstrating habenular hyperactivity in depression-like states and linking feedback from serotonergic raphe neurons to the lateral habenula with a pattern of habenular hyperactivity and depression. Our work demonstrates that recurrent neural network modeling can accurately replicate experimental data and uncover network connectivity dynamics that are not accessible experimentally.

We are also expanding this collaboration to build novel multi-region network models that go beyond neural activity to also consider the role of non-neuronal cells and tissues in cognition (Benster et al., in prep). Specifically, we are focusing on glial-neuronal interactions in a small, highly accessible nervous system–the larval zebrafish–in a unique time-extended behavioral paradigm. Our multi-scale approach will incorporate vasculature as an intermediate readout between the scales of neuronal and glial operation (milliseconds- to seconds-long) and that of behaviors (operating over minutes and hours, even days) thereby combining both experimental advances in the Deisseroth lab–simultaneously imaged neural and glial dynamics, and vasculature in the whole brain at high spatiotemporal resolution–and computational innovations in the Rajan lab–unifying theory based on a new class of modular and interpretable network models constrained directly by these diverse multiscale data. Analyzing the trained models in conjunction with data enables us to infer circuit mechanisms inaccessible experimentally. The range of spatiotemporal scales at which glial-neuronal interactions operate and interface with vascular readouts could provide the key ‘missing links’ encountered while trying to mechanistically bridge dynamics with behavior, or connectivity with either dynamics or behavior.

Benster, T., Andalman, A., Deisseroth, K.*, Rajan, K.*, (2021) Current-based decomposition reveals differential roles of brainwide glial and neuronal populations in controlling adaptive and maladaptive behavioral state transitions, in preparation

Andalman, A. S., Burns, V. M., Lovett-Barron, M., Broxton, M., Poole, B., Yang, S. J., Grosenick, L., Lerner, T. N., Chen, R., Benster, T., Mourrain, P., Levoy, M., Rajan, K., & Deisseroth, K. (2019). Neuronal Dynamics Regulating Brain and Behavioral State Transitions. Cell, 177(4), 970–985.e20.

Link to Poster: [no poster]

Demo: [no demo]


Back to Main BRAIN TMM page

Table sorting checkbox