and in the few cases of truly novel complexes, success was often explained by pairing common secondary structure elements together. e.g., AF3 prediction of PDB 8HDJ (5/5)

while the target MSA was almost always necessary, we were able to rule out inter-chain covariation as important for good performance by shuffling sequence pairing. e.g., AF2 prediction of PDB 3TRS (4/5)

we even found that this bias might even mislead predictions to use wrong but common binding sites. e.g., Boltz-1 prediction of PDB 8V8E (3/5)

we found some concerning evidence of memorization for both pre-cutoff structures and structures homologous to pre-cutoff structures 🤔 (2/5)

RLA has been tested on several benchmark sets, including several design libraries of miniprotein designs for a variety of protein targets. For all but two targets, filtering with RLA results in a higher success rate after subsequent AF2-based filtering. 2/3