Ran Boltz-2 100 million times across 9,000 targets to test if AI can simulate cell biology
HannesStaerk · x · 2026-11-14
Recursion researcher Dominique Beaini spent 12 months running AlphaFold Multimer and Boltz-2 co-folding with affinity predictions on preemptible GPUs, producing 100 million co-foldings across 9,000 unique protein targets and 500k ligands — a dataset he calls affinity-prints.
Key findings:
- The hard part was identifying which predicted interactions matter inside cells, done by coupling affinity-prints with Recursion's large-scale phenomics maps for sparse screening.
- He built the smallest Virtual Cell with just 6 parameters, estimating phenomics similarity maps by combining dense affinity-prints with proprietary gene expression and gene-gene relationship data.
- He created Lig2Cell, a simple data-driven ligand score assessing cellular effectiveness.
The post shows a scalable path for AI for Science: mass structure prediction coupled with cell-based assay data.
More from Research
- Researchers claim 1,000× faster reasoning training with 80s-era energy models — alec_helbling · 2026-11-14
- Ai2's XYBench Finds Five Top LLMs Fail to Correct Users' Misguided Questions — allen_ai · 2026-11-14
- Ai2 Asks Whether Prompting or Training Can Make LLMs Give More Useful Advice — allen_ai · 2026-11-14
- Ai2's XYBench Wins EMNLP Best Resource Award, Exposing LLMs' XY Problem Blind Spot — allen_ai · 2026-11-14
- G-STAR paper: 32–64x token compression gives 20x faster autoregressive image generation — _akhaliq · 2026-11-14
- OpenCLIP 4.2 ships with CLAP2 variable-length audio and NaFlexViT support — wightmanr · 2026-11-14