Ligand preparation#
Ligand structures are validated and canonicalised in RDKit, so incompatible compounds drop out before expensive GPU computation.
What we read out#
- Affinity, reported by the model as a log IC50. In this campaign we converted it to a concentration scale and treated it as an operational affinity signal, even though IC50 is not physically equivalent to a dissociation constant without a model of the assay conditions. The conversion will be corrected in the next version.
- Structural confidence measures: ipTM and pLDDT.
Physical plausibility of the pose#
We do not report a single aggregate confidence number. The physical plausibility check (PoseBusters) runs only on the benchmark path, so poses are not filtered on that basis before molecular dynamics.
Strict PoseBusters checking on 848 BindingDB complexes passed no pose at all: in 100% of cases the distance and overlap tests against the organic cofactor fail, which is expected for a binding site adjacent to NADPH. With those checks disabled, 68 poses pass validation (8.02%).
Cost#
Boltz-2 is the most expensive stage in the pipeline: a measured median of about 26.5 seconds per molecule, roughly 62% of the cost of a reference 50-molecule run.