Reproduce “Emyx: Fast and efficient all-atom protein generation” 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
Nicholas J. Williams, Ward Haddadin, Matteo P. Ferla, Constantin Schneider, Nicholas B. Woodall, Ruby Sedgwick, Christian D. Madsen, Andrew L. Hopkins and 2 more
Emyx is a 140-million-parameter all-atom protein generator designed for catalytic-motif and ligand scaffolding with sparse attention and lightweight conditioning. On the 41-target AME benchmark, it reaches 13.4% strict self-consistency success, exceeding Proteína-Complexa at 8.8% and RFdiffusion3 at 6.7%, while using about one quarter of RFdiffusion3’s reported training compute.
Key insights
- Strict evaluation matters: Emyx’s success falls from 35.8% under heavy-atom sc-RMSD to 13.4% when global backbone recovery, motif-tip accuracy, and ligand-clash avoidance are required together.
- Sparse, lightweight generation can remain competitive with larger systems while producing more structurally diverse outputs; successful Emyx designs form 441 unique Foldseek clusters versus 292 for Proteína-Complexa and 242 for RFdiffusion3.
- EDM sampling outperforms deterministic Euler and stochastic SDE sampling on the tested six-target subset, suggesting that the sampling procedure materially affects flow-matching protein generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.