Here we explore the use of the recently developed deep learning method, AlphaFold, to predict structures of transient interactions between proteins from their sequences. With a benchmark of 152 heterodimeric protein complexes of various classes, including enzyme-inhibitor and antibody-antigen interactions, and an additional set of 20 antibody-antigen complexes, multiple implementations and parameters of AlphaFold were tested for accuracy.
AlphaFold predictions of heterodimeric protein complexes in BM5.5 (download predictions) and additional antibody-antigen complexes (download predictions).
AlphaFold predictions generated with alternative parameters (e.g., paired MSA, larger ensembling iterations (Nensemble) and larger recycling interations (Ncycle)) (download predictions).
Alternatively, you may download all (file size: 130M).
Yin R, Feng BY, Varshney A, Pierce BG. Benchmarking AlphaFold for protein complex modeling reveals accuracy determinants. bioRxiv 2021.10.23.465575 bioRxiv preprint