Predicting Neural Circuit Fragility with GNNs
A self-directed summer project comparing graph neural networks against classical centrality measures for predicting critical neurons in the C. elegans connectome.
I wanted to know whether a GNN could actually learn something useful from a connectome that simple graph statistics couldn't. So I built a cascade simulator over the 297-neuron chemical connectome, computed 5 centrality baselines, and trained a 2-layer GCN against them in PyTorch Geometric. Result: the GCN (R² ≈ 0.32-0.35) didn't beat linear regression on the centralities (R² ≈ 0.51). Error analysis suggested the bottleneck was features, not architecture.
Skills: Python, PyTorch Geometric, NetworkX, scikit-learn, GCN