taqo: free-energy perturbation quality congeneric ranking from a single interaction energy

Relative binding free-energy (RBFE) methods such as free-energy perturbation (FEP) set the accuracy standard for ranking congeneric ligands, but their cost limits throughput. We ask whether a single machine-learned force-field (MLFF) interaction energy with no alchemical sampling, explicit solvent, or affinity training, can recover a useful fraction of the RBFE signal. We introduce \taqo, a model-agnostic static score, and benchmark ten MLFFs across six families on 22 public congeneric-series benchmarks against published OpenFE and FEP+ references on matched ligand sets. The top-performing models are statistically indistinguishable and all significantly outrank a classical force field, all at a few seconds per ligand on one GPU. Aggregated over the benchmark, \taqo matches OpenFE and trails FEP+ by a small margin when considering mean Kendall $\tau$. On the congeneric JACS set it is on par with OpenFE and approaches FEP+, at roughly three orders of magnitude lower cost. On identical poses it outranks classical docking and is statistically indistinguishable from deep learning based methods, without any affinity-specific training. Together these establish MLFF interaction-energy scoring as a practical high-throughput approach for lead optimization.