Ab-initio Binding Affinity, No Force Field, No Docking Score
A derived physics readout scores absolute binding across four independent public benchmarks, including the broadest held-out set we have. We report every panel, strongest and weakest.
Rank agreement across four benchmarks
| Benchmark | n | Rank agreement (Spearman ρ) | ΔG MAE (kcal/mol) |
|---|---|---|---|
| CASF-81 core set | 81 | 0.9182 | 1.17 |
| CASF-285 | 285 | 0.8800 | 1.29 |
| Blind held-out (2019) | n/a | 0.7712 | 1.36 |
| PDBbind broad held-out | n/a | 0.7078 | 1.47 |
Rank agreement (Spearman ρ) is consistent across independent data splits. We report our weakest panel too: on the broadest, most chemically diverse held-out set (PDBbind general), rank agreement is ρ ≈ 0.71, reported not hidden. ΔG MAE is from a monotone calibrator fit on training data only and applied to the held-out panels. Ranking (Spearman ρ) is the primary metric; the calibrated error is reported, not optimized.
How it compares
| Method | CASF-2016 correlation | Speed |
|---|---|---|
| LatticeZero | 0.8800 (Spearman ρ) | 1000s lig/sec |
| RF-Score v3 | 0.75 (Pearson R) | ~1,000 lig/sec |
| FEP+ (AB-FEP) | 0.69 (Pearson R) | hours per molecule |
| AutoDock Vina | 0.60 (Pearson R) | ~50 lig/sec |
| Glide SP | 0.57 (Pearson R) | ~100 lig/sec |
LatticeZero is Spearman ρ on CASF-285 (the CASF-2016 core set); competitor values are published Pearson R on CASF-2016. Physics-first scoring runs thousands of ligands per second, while FEP+ runs hours per molecule.
What this shows
Binding affinity ranked from derived physical laws, with no force field and no docking function, reaches the rank accuracy of established scoring functions on curated benchmarks and holds up on the broadest, most diverse held-out set we have. Every panel is reported, strongest and weakest. Rank agreement (Spearman ρ) is the primary metric; the calibrated absolute-ΔG error (kcal/mol) is reported alongside, fit on training data only and not optimized to the panels.