Benchmark measures how well AlphaFold predicts antibody-antigen complex structures

What this research was about

Antibodies are immune proteins that latch onto a target, called an antigen, and the grip depends on shape. The two molecules fit together at a small patch of contact, and knowing the exact arrangement of atoms there tells researchers how the antibody recognises its target. Working that arrangement out experimentally is slow and demanding work. Predicting it from the sequence of amino acids alone is harder still for antibodies than for most proteins, because the loops that do the binding vary enormously between antibodies and have few close relatives in the evolutionary record for a method to learn from.

The researchers set out to measure how often AlphaFold, a deep neural network that predicts protein structures from sequence, gets these pairings right. They assembled a non-redundant set of 429 complexes from a public antibody structure database, choosing cases whose structures were released after the model's training data stopped, so the answers could not already be known to it. They also compared the results against an older, physics-style docking approach, and examined whether the model's own confidence numbers flag which predictions to trust.

Where AI came in

AlphaFold did the prediction itself. Given antibody and antigen sequences, two released versions of AlphaFold-Multimer, run off the shelf on a local cluster of GPUs, produced ranked three-dimensional models of each complex. The same sequences were also run through ColabFold, a wrapper around the model, adjusted to save a prediction at every internal refinement pass. AlphaFold additionally modelled the antibody and antigen separately, and those single-molecule models were fed into the ZDOCK docking program as the comparison baseline. In the record's terms, the model stood in for a physical experiment: the structures it output take the place of ones that would otherwise be determined in a laboratory.

Everything else was conventional software. The predicted shapes were compared with the experimentally determined structures using DockQ and the CAPRI accuracy classes, which are standard yardsticks for how close a predicted complex is to the real one. The model's own per-atom confidence values, averaged over the residues at the contact patch, were turned into a score for judging which predictions were likely to be accurate. Top-ranked predictions reached Acceptable or better accuracy in 25 per cent of the 429 cases and Medium or better in 18 per cent, while the docking baseline reached Medium or better in 1 per cent of the 390 cases it was run on.

Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.