Binding affinity within 1 kcal/mol for novel scaffolds across protein families

• Physics-based relative FEP. On the curated FEP+ benchmark, FEP+ achieves pairwise RMSE of about 1.25 kcal/mol, against 0.91 kcal/mol for experimental reproducibility in public data (Ross et al., Communications Chemistry 2023). It performs well for congeneric series on well-behaved targets, much less well for charge-changing perturbations, scaffold hops, flexible loops, membrane proteins, and nucleic-acid targets (about 1.39 kcal/mol RMSE for DNA/RNA binders in a vendor-sponsored preprint). Open implementations (OpenFE) are somewhat less accurate than FEP+ in a 59-dataset comparison by OMSF, with atom-mapping failures producing outliers.
• CASP16 (2024). In the first large blind affinity test in the CASP series, 30 groups predicted 140 affinities on two targets: the best Kendall's $\tau$ was 0.42, below the noise ceiling of about 0.73; providing experimental structures did not help, which the assessors take as evidence that scoring functions are the limiting factor.
• Learned affinity heads. Boltz-2 (preprint, June 2025) reports average Pearson 0.66 on a four-target FEP+ subset and approaches OpenFE on 876 measurements at more than 1,000 times lower cost, outperforming all CASP16 participants out of the box; on eight blinded internal industrial assays it exceeded Pearson 0.55 on only three, and the authors state that performance "varies strongly between assays". These are developer-reported numbers without independent prospective replication in sources reviewed.
• Hit discovery. The same preprint reports an enrichment factor of 18.4 at the top 0.5% in a retrospective screen (MF-PCBA); this measures ranking against decoys, not accurate ΔG on scaffolds.
• Not shown. No method meets the benchmark above prospectively across families without target-specific information.

Accurate scaffold-agnostic affinity prediction would let medicinal chemists rank virtual compounds before synthesis, cut the number of design-make-test cycles in lead optimisation, make large virtual screens reliable, and allow design of selectivity (predicting off-target binding). It would also be the scoring function needed for generative chemistry.

A prospective, pre-registered, blinded evaluation (a D3R/CASP-style or industry-sponsored challenge with unpublished measured affinities) in which one method, with no target-specific data, meets the RMSE, unsigned-error and Kendall's $\tau$ thresholds above on at least 10 protein families and at least 200 compounds with novel scaffolds, with the analysis published in a refereed journal and the method accessible to independent groups.