ProtBFF: biophysical-score embedding caches and trained models for reproducing protein-protein ddG benchmarks

Precomputed score caches (biophysical-score-scaled protein embeddings) and trained ProtBFF model checkpoints for reproducing the benchmarks in the ProtBFF paper (Feldman, Maechler, Wang, Shakhnovich). SKEMPI2 caches for the ProSST, ESM2, ESM3, ESM-C and SaProt encoders; SARS-CoV-2 deep-mutational-scanning caches for ACE2, 9LYP and 7KMG; and the trained model checkpoints. Together with the code and homology-controlled (MVA) cross-validation splits at https://github.com/Dianzhuo-Wang/ProtBFF, these reproduce the paper's results. Each *.npz cache holds forward/reverse pooled features (Xf, Xr), ddG targets (y), and complex ids.