Two small corrections during structure generation helped Boltz-2 and OpenFold-3 achieve a 100% best-of-five pass rate on the paper’s physical-validity checks across five benchmark sets, according to an Oct. 4 arXiv preprint. For every target, at least one of five generated structures passed; the score does not mean every output passed.
AI structure models predict how proteins and other molecules fit together in three dimensions. A prediction can look plausible overall while containing local errors such as overlapping atoms or unrealistic chemical bonds. A ligand, the small molecule bound to a protein in these tests, can also be distorted even when its predicted position looks reasonable.
The researchers added two small corrections near the end of diffusion sampling, the repeated denoising steps that turn random coordinates into a structure. One nudges apart atoms from separate chains that overlap. The other adjusts a ligand’s distorted bonds, angles, internal spacing and ring shape. The corrections cap how far atoms move and use coordinates the sampler already has, so they add no neural-network pass and do not change model weights.
Across CASP15, CASP16, the PoseBusters monomer and complex sets, and the Boltz physical-validity test set, both corrected models reached a 100% best-of-five rate. In the CASP16 comparison, the uncorrected rates were 53.8% for Boltz-2 and 69.2% for OpenFold-3. The paper reports little change in structural-accuracy and ligand-placement metrics; runtime stayed within 3.5% of each model’s baseline, with no measurable memory increase.
The checks cover the geometry targeted by the corrections, not all physical realism. They do not assess torsional strain, protein-backbone geometry or hydrogen placement. The preprint reports no wet-lab test, so it does not establish whether a corrected prediction binds as expected in a biological experiment.
Sources
Primary study: Inference-Time Projection for Physically Valid Biomolecular Diffusion Models
Full paper, methods, results and limitations
Figure 1 from the study, used under CC BY 4.0.