Verify whether Boltzina matches Boltz-2 on absolute affinity prediction and pose physical validity

Determine whether the Boltzina framework—an affinity prediction pipeline that inputs AutoDock Vina docking poses directly to the Boltz-2 affinity module while omitting Boltz-2’s structure prediction—can match the performance of Boltz-2’s full pipeline on (i) absolute protein–ligand binding affinity prediction and (ii) assessment of the physical validity of predicted protein–ligand docking poses across appropriate benchmarks.

Background

Boltzina is introduced as a faster alternative to Boltz-2 by bypassing the diffusion-based structure prediction stage and feeding docking poses from AutoDock Vina directly into Boltz-2’s affinity module. This approach delivers substantial speedups while retaining much of Boltz-2’s screening accuracy in MF-PCBA assays.

However, the paper focuses on screening metrics (e.g., AP, EF, ROC-AUC) and does not evaluate two key tasks where Boltz-2 is notably strong: absolute binding affinity prediction and the physical validity of predicted poses. The authors explicitly state that these evaluations were not performed and that parity with Boltz-2 on these tasks remains unverified.

Establishing whether Boltzina can match Boltz-2 on these tasks is important for determining the scope of applications where Boltzina’s computational efficiency can be adopted without sacrificing crucial accuracy or physical plausibility, particularly for cases requiring reliable absolute affinity estimates or structural validation of poses.