AI and the World's Hard Problems · Life sciences · arXiv · 2026-10-23
Escaping the Phylogenetic Trap: Persistent Homology Enables Cross-Species Ion Channel Classification via Multimodal Graph Neural Networks
Random-split evaluation of protein classifiers can exploit taxon-specific correlations and overestimate performance on unseen taxa, a problem especially acute for venom compound classification by ion-channel target. The authors compare five structural inductive biases in a multimodal graph neural network under leave-one-taxon-out evaluation on 1,373 venom compounds. Under taxonomic shift, their persistent-homology model PH-GNN reaches 63.4% accuracy, over 8 percentage points above the next-best method, while a sequence-only baseline drops to about 30%.