Reproduce “Protein Representation Learning by Geometric Structure Pretraining” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zuobai Zhang, Minghao Xu, Arian Jamasb, Vijil Chenthamarakshan, Aurelie Lozano, Payel Das, Jian Tang
GearNet combines relational residue graphs, edge-level geometric message passing, and protein-specific self-supervision to learn transferable representations from 805,000 AlphaFold structures. Its multiview contrastive pretraining improves enzyme, Gene Ontology, fold, and reaction prediction, matching or exceeding sequence-pretrained models on several tasks despite using far fewer structural examples.
Key insights
- Protein-specific structural views are effective contrastive augmentations: partial sequence segments and spatial neighborhoods can provide compatible representations of the same protein while retaining functional geometric information.
- Explicit edge-level geometry improves over ordinary residue-level graph aggregation, with especially strong benefits for fold-related classification.
- Sequence and structure representations are complementary rather than interchangeable; combining ESM-1b with GearNet yields further gains on function prediction.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.