Bayesian Inference of Dynamical Models of Connectomes

Recent progress in connectomics has opened new frontiers for understanding the underlying principles of neural circuits. By leveraging high-resolution maps of synaptic connections, computational models can simulate neural dynamics with unprecedented detail. However, satisfying, compact models of the brain that integrate circuit activity data with connectomic information remain elusive. We propose Bayesian inference on biophysical models of neural dynamics as a principled method for bringing to bear existing data, enabling uncertainty quantification for inferring parameters of interest, as well as for predicted circuit outputs. To demonstrate this approach, we implement a connectome-scale spiking neuron model with leaky-integrate-and-fire dynamics, and use published firing rate data to perform inference on a compact biophysical model of neuronal voltage. We evaluate how models with varying levels of biological detail fit experimental data and examine how training on different subsets of data influences model predictions.