bioRNNS: toward more biologically plausible neural network models of brain dynamics

Recurrent neural networks (RNNs) are powerful tools for modeling neural dynamics, but standard architectures often lack biological realism in two key aspects: their all-to-all connectivity pattern where every node processes inputs and outputs directly, and their unconstrained hidden layer formation which can lead to biologically unrealistic architectures. We address these limitations by developing biophysical RNNs (bioRNNs) with two novel constraints: (1) input/output masking that restricts inputs to "visual cortex" nodes and outputs to "default mode cortex" nodes, mirroring biological information flow, and (2) biophysical embedding based on distances along the sensory-association axis, which represents the functional gradient from sensory to higher-order cognitive regions. Our bioRNNs demonstrated more brain-like neural dynamics in their hidden layer timescales compared to standard RNNs. Furthermore, in the presence of I/O masking, bioRNNs outperformed standard RNNs across various behavioral tasks. These findings suggest that incorporating biophysical constraints in RNNs can create more realistic and interpretable brain models, enhancing their utility for both basic and clinical neuroscience research.