Boltz-2 run 100 million times: Recursion researcher builds a minimal virtual cell
HannesStaerk · x · 2026-09-14
Recursion's Dominique Beaini details a 12-month experiment: using preemptible GPUs, he ran 100 million AlphaFold Multimer / Boltz-2 co-foldings with affinity predictions across 9,000 unique targets and 500k ligands, building the 'affinity-prints' dataset.
- The hard part wasn't compute but identifying which predicted interactions matter inside a cell—leveraging Recursion's large-scale phenomics data to link affinity-prints with cell-based assays.
- Key results: built a minimal Virtual Cell with as few as 6 parameters, estimating phenomics similarity maps by coupling dense affinity-prints with Recursion's biological maps; also created Lig2Cell, a data-driven ligand scoring method.
- A representative AI-for-Science attempt at scaling structure-prediction models toward cell biology simulation.
More from AGI Musings
- Researcher argues catching up to frontier AI costs less than proposed compute budgets — eliebakouch · 2026-09-14
- Even an AI disaster would just push progress behind closed doors, argues thread — yeastsplainer · 2026-09-14
- LeCun called out as 'not very smart': critics say JEPA can't deliver superintelligence at 1000 TPS — teortaxesTex · 2026-09-14
- The hidden tax of multimodal AI: lung cancer study questions cross-validation gains — bravo_abad · 2026-09-14
- New series "Mathematics in the age of AI" asks if math is in a crisis — elsleightholm · 2026-09-14
- New Handbook Chapter Analyzes Sociodigital Exploitation of Africa — ChinasaTOkolo · 2026-09-14