Alphabet-owned research firm DeepMind has introduced AlphaFold-Multimer, a model that can predict the structure of multi-chain protein complexes. The new model significantly increases the accuracy of predicted multimeric interfaces over input-adapted single-chain AlphaFold while maintaining high intra-chain accuracy.
A majority of well-structured single protein chains could be easily predicted using the previous AlphaFold model, but the prediction of multi-chain protein complexes remained a challenge in many cases, which the AlphaFold-Multimer addresses readily.
AlphaFold-Multimer analyzes multiple chains during both training and inference, with native support for multi-chain featurization and symmetry handling. Multiple changes to the previous AlphaFold system were made to adapt it to training on protein complex. The AlphaFold-Multimer introduces a new way of selecting subsets of residues for training and makes various small adjustments to the structure losses and the model architecture.
AlphaFold-Multimer was tested on a benchmark dataset of 17 heterodimer proteins without templates, where it achieved at least medium accuracy on 14 targets and high accuracy on 6 targets, compared to 9 targets of at least medium accuracy and 4 of high accuracy for the previous state of the art system.
The new model shows that the confidence metrics provided by the model correlate well with the true accuracy, something that is vital for the useability of a structure prediction model.
Deepmind says that this method will enable biologists to further accelerate the recent progress in structural bioinformatics and act as a stepping stone towards executing on more complex folds, such as RNA & DNA molecules.
