RoseTTAFold2

RoseTTAFold2 (Baek et al., bioRxiv 2023) is the University of Washington upgrade of the original RoseTTAFold three-track network. It extends the 1D/2D/3D three-track architecture across the full network and borrows several AlphaFold2 ideas, including frame-aligned point error, recycling during training, a distillation set from AlphaFold2, and structurally coherent attention for updating pair features. Notably it skips AlphaFold2's hallmark invariant point attention and triangle attention, using a cheaper structure-biased attention instead, and still reaches comparable accuracy.

The authors report AlphaFold2-level accuracy on monomers and AlphaFold2-multimer-level accuracy on complexes, with better computational scaling for large proteins and complexes. The network matters beyond its own predictions: it is the documented backbone that RFdiffusion, RF2NA, and RoseTTAFold All-Atom were built on. It remains MSA-dependent, so accuracy drops on orphan or low-MSA sequences, and it outputs a single static structure rather than conformational ensembles. Code and weights are openly available.

Author-reported (Baek et al., bioRxiv 2023): accuracy of AlphaFold2 on monomers and of AlphaFold2-multimer on complexes, with better computational scaling for large proteins and complexes; no independent blind-benchmark audit published