AlphaFold3 identified up to 53% of known antibody and antigen pairs and classified 3% of control pairs as plausible

AlphaFold3 identified up to 53% of known antibody and antigen pairs and classified 3% of control pairs as plausible

On 31 July, Arnav Solanki’s team published an evaluation of AlphaFold3 using 3 401 experimentally established antibody and antigen complexes and 23 798 negative pairs. After 100 runs per pair, the model identified up to 53% of the known interactions. At the selected thresholds, it also assigned positive scores to 848 negative pairs.

An antibody recognizes an antigen, meaning a target protein, through a small region on its surface called an epitope. By binding the intended target, a therapeutic antibody can alter the protein’s function or deliver a drug to a cell. The search begins with many combinations of antibodies and targets. Researchers then measure binding experimentally for a short list of candidates.

In the preprint by Solanki and colleagues, AlphaFold3 receives the antibody and protein sequences and predicts their joint three-dimensional arrangement. The model also reports confidence scores for the interface, the region where the two proteins contact each other. The authors selected known complexes from the SAbDab database and created a control set from the same antibodies paired with proteins for which SAbDab records no binding interaction.

At the selected interface score thresholds, AlphaFold3 flagged 848 of 23 798 negative pairs. Each prediction resembles contact between an antibody and the surface of a protein, even though the control set consisted of pairs with no binding interaction recorded in SAbDab. The model generates a structural hypothesis, whereas a laboratory experiment determines whether binding occurs.

In approximately one third of the false negatives, the model placed the antibody at the correct epitope, although the molecular arrangement did not match the experimental structure. Identifying the target region and predicting the precise arrangement within that region are separate tasks. The authors suggest testing these candidates separately and refining their geometry when necessary.

AlphaFold3 can therefore help prioritize candidates for testing, but it cannot replace that testing. A laboratory experiment is required to determine whether a given pair forms a complex.