Reproduce “Automating AI R&D for Molecular Discovery” 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
Intology
Intology evaluates Locus, its autonomous AI R&D system, on three molecular discovery tasks: proprietary molecular property prediction, adapting Boltz-2 for small-molecule potency prediction, and designing antibody workflows around BoltzGen/Boltz-2 and ESMC/ESMFold2. Under matched compute, Locus outperforms Claude Code, Codex, and baselines from Boltz and BioHub, generalizing better from limited data and finding more sample-efficient solutions.
Key insights
- Scaling synthetic data was decisive: Codex generated none and Claude Code reached about 100k molecules, while Locus scaled to nearly 1M.
- More diverse exploration of adaptation approaches correlated with stronger final performance, and Locus explored the most diverse set.
- Learned filtering criteria retain more strong binders than BoltzGen's expensive generate-then-strictly-filter workflow.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.