Statistics on the frequencies of pi interactions in folded protein structures enable successful prediction of intrinsically disordered protein phase separation, with clear implications for a physical understanding of cellular organization.

A new method of protein structure prediction that incorporates residue–residue co-evolution information into the Rosetta structure prediction program was used to develop models for 58 large protein families that had no previous structural information.

Co-evolving residue pairs in the different components of a protein complex almost always make contact across the protein–protein interface, thus providing powerful restraints for the modeling of protein complexes.

Alternative conformations of membrane protein structures can be predicted to high accuracy with AlphaFold2 by reducing the depth of the multiple sequence alignments used for modeling.

Experimental determination of residue contacts from mutational data allows model discrimination and identification of in vivo functional conformations of proteins.

Integration of language model and geometric deep learning enables accurate and efficient genome-scale annotation of comprehensive protein-ligand binding sites.

A new machine-learning toolbox unveils coevolutionary protein motifs related to structure, function, and phylogeny from sequence information only.

The AlphaFold protein structure prediction network can be specialized for T cell receptor docking, leading to improved models of ternary complexes and some ability to discriminate correct from incorrect peptide epitopes.

The ability to accurately predict a protein's structure gives us an entirely new way to annotate the human genome.

Integrating multiple protein language models using protein geometric graphs can dramatically improve the model performance for predicting the contacting residue pairs between interacting proteins.