mims-harvard/SubGNN

Encodes subgraph-level predictions by aggregating three complementary neural channels—neighborhood, structure, and position—each capturing distinct topological patterns within induced subgraphs. Built on PyTorch with modular config-driven training, it supports both real-world biological networks (HPO, PPI) and synthetic benchmarks, with hyperparameter search and multi-seed evaluation pipelines.

202 stars. No commits in the last 6 months.