New method improves how venom compounds are classified across different species

A topological approach helps models identify ion channel targets even when testing on unseen taxa.

Researchers found a way to better classify venom compounds by their ion-channel target, such as Ca2+Ca^{2+}, K+K^{+}, and Na+Na^{+}. This method uses persistent homology to improve performance when moving between different species.

The study used a benchmark of 1,373 ColabFold-predicted venom compounds spanning five taxa. They compared five structural inductive biases, including spectral, Riemannian, SE(3)-invariant, statistical-manifold, and persistent-homology representations.

Persistent homology summarizes backbone geometry through topological descriptors. These descriptors are robust to local structural perturbations. A shared multimodal backbone fuses ESM-2 sequence embeddings with a disulfide-aware graph attention network.

Standard random train/test splits cause phylogenetic leakage. This allows models to exploit taxon-specific correlations and overestimate out-of-distribution performance. Using a leave-one-taxon-out evaluation protocol shows the true ability to generalize.

Under taxonomic shift, a sequence-only baseline drops to approximately 30%. The persistent-homology model, or PH-GNN, achieves 63.4%. This outperforms the next-best method by over 8 percentage points.

Key facts

  • The PH-GNN model achieved 63.4% accuracy under taxonomic shift.
  • The study used 1,373 ColabFold-predicted venom compounds spanning five taxa.
  • Persistent homology provides an inductive bias toward phylogenetically invariant learning.

All sections

Edition