Skip to main content
Back

MIT's Boltz-2 Ranks 2x More Drug Hits With Just 300 Labels

A small dose of assay-specific labels teaches Boltz-2's affinity head to rank real hits far better, without retraining the giant structure model.

Paper, Full Access
  • Fine-tuning only Boltz-2's affinity heads with 300 binary labels lifted average precision 2.14x on eight MF-PCBA targets.
  • Actives in the top 1% rose by a geometric mean of 1.77x, improving early enrichment for hit discovery.
  • Head fine-tuning beat both DrugCLIP projection-head tuning and LightGBM trained on the same labels.
  • Gains held for compounds with scaffolds unseen during fine-tuning, suggesting real generalization.
  • Rescoring only the top ~10% of the library preserved hit recovery, cutting expensive Boltz-2 inference cost.
  • Built on the open-source Boltz-2 foundation model from MIT and Recursion.

A new preprint reports that target-specific fine-tuning can improve Boltz-2 rankings with 300 binary assay labels. Across eight MF-PCBA targets, updating only the model’s affinity heads increased the number of actives in the top 1% by a geometric mean of 1.77 times and improved average precision by 2.14 times relative to the unmodified model.

Virtual-screening campaigns often rank millions of compounds while experimental budgets cover only a small fraction of the library. Assay-specific data can therefore matter more than broad benchmark performance: modest ranking changes determine how many confirmed actives reach the test plate.

A broad model meets a narrow assay

Boltz-2, developed by MIT’s Jameel Clinic and Recursion, jointly predicts protein-ligand complex structure and binding affinity. Its affinity module was trained on about 5 million measurements, according to the model overview.

On the retrospective MF-PCBA benchmark, the Boltz-2 paper reports roughly twice the average precision of earlier machine-learning and docking baselines, plus an enrichment factor of 18.4 in the top 0.5%. An enrichment factor of 18.4 means that this narrow slice contains 18.4 times as many actives as random selection would yield.

Benchmark scores average performance across targets, while production screens depend on one protein construct, one assay protocol, and one chemical library. The new study tests whether a small set of measurements from that exact assay can improve Boltz-2 without retraining its costly structural core.

The trunk stays frozen

The adaptation keeps Boltz-2’s structure trunk fixed and updates only its affinity heads. The trunk generates learned representations of the protein-ligand complex; the heads convert those representations into binding scores. Freezing the trunk avoids updating most model parameters, although uncached compounds still require a forward pass to generate their representations.