Protein complex prediction with AlphaFold-Multimer

ABSTRACT

While AlphaFold 2 achieved unprecedented accuracy in predicting the structure of single protein chains, biological functions are primarily carried out by multi-protein complexes. We present AlphaFold-Multimer, an extension of the AlphaFold architecture specifically trained to predict the structure of multimeric protein complexes. By modifying the multiple sequence alignment (MSA) pairing logic and retraining the model on known complex structures from the Protein Data Bank, AlphaFold-Multimer significantly outperforms existing docking methods. The model provides highly accurate predictions for both homo- and hetero-omeric interfaces, expanding the utility of deep learning to the interactome.