Top Highlights

• we modify the losses to take into account permutation symmetry among identical chains, pair the MSA alignments between individual chains to surface cross-chain genetic information, introduce a new way of selecting subsets of residues for training, and make various small adjustments to the structure losses and the model architecture.
• trained specifically for multimeric inputs of known stoichiometry
• To summarize briefly, it combines information from the amino acid sequence, multiple sequence alignments and homologous structures in order to predict the structure of individual protein chains.
• These two representations are mixed and processed by a collection of neural network modules.
• The pair representation can be thought of as containing information about the relative positions of amino acids in the chain.

Domain