AI-designed antibodies: a blinded competition has 500 candidates synthesized and tested in the laboratory

Published in Nature Biotechnology, the AIntibody challenge put 29 organizations in competition on criteria measured in the benchtop. The best variants gain up to 75 times in affinity.

A CASP for antibodies

Antibody design rankings are too often based on calculated metrics. The AIntibody challenge, the results of which were published on September 22 in Nature Biotechnology, took the opposite approach, inspired by the CASP competition that structured the prediction of protein structures: participants submit their sequences blindly, without knowing the comparison targets, and it's the bench that decides.

Twenty-nine organizations from academia, biotechnology, and the pharmaceutical industry played. More than 500 antibodies optimized or entirely engineered by machine learning were synthesized, expressed, and then subjected to standardized tests of affinity and "developability" — the latter of which includes the properties that determine whether a molecule will survive the scale-up.

Among the results, variants reach an affinity up to 75 times higher than that of the parent molecule, with bindings of the order of picomolar. Zymo Research is among the first four teams in the affinity maturation component, ahead of most pharmaceutical competitors. The interest of the protocol lies above all in its rigor: the announced gains were measured, not predicted.