Authors: John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, et al. (DeepMind)
Jointly embeds Multiple Sequence Alignments (MSAs) and pairwise residue representations inside the 48-block Evoformer using triangular multiplicative updates and Invariant Point Attention (IPA) in SE(3) equivariant space.
Key Quantitative & Theoretical Takeaway: Triangular multiplicative updates enforce geometric triangle inequality consistency across residue pair representations before the Structure Module predicts 3D backbone frames via Frame Aligned Point Error (FAPE).
Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known.
Pair representation z_ij is updated by aggregating information over all intermediate nodes k along edges ik and jk, enforcing triangle inequality consistency across pairwise residue distances.
IPA computes attention weights in local residue backbone frames T_i in SE(3) so that transforming the global protein coordinates by any rigid rotation or translation leaves the internal attention affinities unchanged.
Replace arxiv.org with nujan.app on any arXiv URL (for example, https://nujan.app/abs/2006.11239) to open the PDF and AI research partner side-by-side.