Efficient and accurate prediction of protein structure using RoseTTAFold2

AlphaFold2 and RoseTTAFold predict protein structures with very high accuracy despite substantial architecture differences. We sought to develop an improved method combining features of both. The resulting method, RoseTTAFold2, extends the original three-track architecture of RoseTTAFold over the full network, incorporating the concepts of Frame-aligned point error, recycling during training, and the use of a distillation set from AlphaFold2. We also took from AlphaFold2 the idea of structurally coherent attention in updating pair features, but using a more computationally efficient structure-biased attention as opposed to triangle attention. The resulting model has the accuracy of AlphaFold2 on monomers, and AlphaFold2-multimer on complexes, with better computational scaling for large proteins and complexes. This excellent performance is achieved without hallmark features of AlphaFold2, invariant point attention and triangle attention, indicating that these are not essential for high accuracy prediction. Almost all recent work on protein structure prediction has re-used the basic AlphaFold2 architecture; our results show that excellent performance can be achieved with a broader class of models, opening the door for further exploration.