Clarification of checkpoint pair features and pocket-sample score selection
We are benchmarking AQAffinity on published PICK1 and syntenin inhibitors before applying it to SYNJ2BP. We observed poor affinity rankings, including a PICK1 control with an accurately predicted crystal pose. Before interpreting this as a model limitation, we would appreciate clarification of two implementation details.
Which pair features does the released checkpoint expect?
The training loader appears to zero protein–protein pair features, while inference supplies the full matrix. Is this intentional, or should inference use the same masking?Which pocket-sample score should be reported?
After selecting the highest-ipTM full-complex pose, pocket inference generates five samples. Our installation saves the first pocket sample’s affinity. Is that intended, or should another selection or aggregation rule apply?