Neural Network Models Constrained by Multiscale Data to Infer Minimal Functional Motifs in the Brain

Abstract State-of-the-art computational approaches offer powerful tools to model the biological brain, but the models are typically ‘single-scale’, or built to match a fixed level of abstraction. Such single-scale models have limited ability to generalize to other levels or scales, preventing the inference of mechanisms that align with real biology. There is a critical need for models that bridge data across spatiotemporal scales and levels of description to enable mechanistic understandings of brain function in health and disease, but producing such models requires innovative technical and conceptual developments to incorporate multiple scales or levels. Recurrent neural network (RNN) modeling offers a powerful approach to overcome past technical limitations to fill this need. In prior work, we developed data-constrained RNNs built on neural dynamics data from multiple brain regions to successfully predict behaviors. Now, preliminary findings suggest that we can further adapt and build on these models to achieve models capable of incorporating data at multiple levels or scales. Indeed, our preliminary work indicates that RNN models with multiple constraints yield more consistent mechanistic inferences across individuals and datasets, are more robust to noise, and a smaller solution space relative to single-level models. These advantages suggest that we can significantly advance our RNNs by simultaneously constraining them by data across multiple scales to develop multiscale RNN models. In response to RFA-EB-20-002, we will apply these new models to decision-making as a use case, and we expect this work to generate a new integrative theory to infer mechanisms across scales and study their collective contributions to adaptive and maladaptive behaviors. Further, we propose that adding multiple biological constraints could enable models to capture structures that are conserved across individuals and species during the same behavior. With this new conceptual framework, we will test whether multiscale RNN models can uncover functional motifs: ensembles of active neurons, independent of brain areas, that mediate key behaviors. Leveraging existing datasets including connectomics/inter-area circuitry, neural dynamics, and behavioral data across species with small, densely sampled nervous systems and those with large, sparsely sampled nervous systems, our aims are to: 1) Build multiscale RNN models with biologically realistic priors based on connectivity, neural dynamics, and behavior; 2) Deploy unsupervised methods to extract functional motifs from data-constrained RNNs; and 3) Disseminate our models and methods. Scientific results from the proposed aims will position multiscale RNNs as a novel and innovative class of models for mechanism discovery in neuroscience and functional motifs as a fundamental mechanistic principle linking brain structure and function. We will work closely with our growing network of collaborative end-users to deploy and test our pipeline to facilitate wide adoption of the resultant tools, methods, and frameworks for further discovery.