“Protein complex prediction with AlphaFold-Multimer,” a DeepMind preprint first posted to bioRxiv in October 2021 by Richard Evans, Michael O’Neill, Alexander Pritzel, and colleagues, extended AlphaFold from single proteins to assemblies of several protein chains. Most proteins do their work not in isolation but bound together into complexes, so predicting how chains fit one another is a distinct and important problem.
The original AlphaFold 2 predicted the structure of a single folded chain. AlphaFold-Multimer is a version of the model trained specifically on multi-chain inputs of known composition. The paper showed that it substantially improved the accuracy of the predicted interfaces between chains, where the proteins actually touch, while keeping the high within-chain accuracy that made AlphaFold famous. Its confidence scores also tracked true accuracy well, which matters because users need to know when to trust a prediction.
This capability fed directly into later work. AlphaFold 3, released in 2024, generalized the idea further to predict not just protein complexes but interactions with DNA, RNA, and small molecules.
For a general reader, AlphaFold-Multimer marks the point where AI structure prediction moved from individual proteins toward the messier reality of molecular machines built from many parts, a necessary step for understanding how cells actually function and for designing drugs that target protein interactions.