From Connectome to Computation: Predicting Neural Function with Machine Learning

Abstract

Connectomics maps synapse-level wiring diagrams with electron microscopy, creating detailed 3D reconstructions of information processing neural networks in the brain. However, such connectome measurements cannot directly reveal brain function – electrical dynamics of single neurons and synapses – making their utility controversial for understanding how the brain computes. While computational neuroscience has traditionally focused on either mechanistically detailed models of small circuits or models that learn patterns from neural recordings without simulating the underlying biology, recent deep learning approaches have taken a different angle by task-optimizing artificial neural networks and establishing that artificial neural network activity matches patterns in brain measurements. However, a main gap remains: Models that integrate single-neuron mechanistic detail with end-to-end task computation and accurately match brain activity. We built a differentiable neural network model from the fly visual system connectome to test whether connectome-constrained networks can accurately predict neural electrical activity. Each neuron in the model corresponds to a neuron of the motion detection pathways in the real fly. We parametrized unknown neuron and synapse properties and trained them on a motion detection task from computer vision using backpropagation through time. Despite equal connectome constraints and task, equivalently trained models generated variable predictions for responses of corresponding neurons. We therefore trained a model ensemble to characterize predictions statistically. Our ensemble accurately predicts neural tuning properties across 26 studies spanning decades of experimental research. Control experiments showed that models with full connectomic constraints and task-optimization predict neural function best. Using dimensionality reduction and clustering of model responses to naturalistic stimuli, we found model predictions for the same neuron type grouped into competing hypotheses across the ensemble, such as upward versus downward motion selectivity. We found that hypotheses from the best task-performing clusters typically matched experimental data, and that clustering combined with a few measurements of neural activity sufficed to effectively identify correct function predictions. Our work demonstrates how connectomes can be useful for predicting neural circuit function when combined with task constraints and machine learning. Our approach – parametrizing unknowns, optimizing via tasks, and distilling predictions from a model ensemble – can be broadly applied to end-to-end connectome-constrained modeling. We published pretrained models and software as research tools. In ongoing work, we develop new models and insights based on, for instance, new connectome data, more realistic synapse models, neural activity measurements, and systematic evaluations of different task optimizations.