The AIntibody challenge: benchmarking the use of AI/ML in antibody discovery

Does AI actually improve antibody discovery? The AIntibody competition set out to answer that with a blinded, prospective benchmark rather than retrospective claims. In this High-Affinity Talks session, Andrew Bradbury (Chief Scientific Officer, Specifica) presents the design and final results of the inaugural challenge, alongside a background to modern in vitro antibody discovery. Despite extensive publicity, the real-world value of AI in antibody discovery remains unclear. AIntibody was launched to measure it directly through a blinded, prospective experimental design. In the inaugural challenge, two Specifica NGS datasets from a Generation 3 selection output were provided to 29 participating organizations, which submitted 527 antibody sequences across three competitions. Each competition asked participants to design or identify antibodies with high affinity and developability: 1. Given NGS datasets from three affinity-maturation selection outputs with binding diversity in HCDR1/2, LCDR1/2, or LCDR3, design high-affinity, developable antibodies. (Experimentally, these outputs are usually combined and sorted to generate affinity-matured variants.) 2. Identify the highest-affinity developable antibody in each of three HCDR3-clustered selection outputs, comprising 400–3,500 different sequences. 3. Given the full NGS selection output of challenge 2, comprising over 30,000 different sequences, design high-affinity out-of-dataset antibodies. All sequences were expressed as full-length IgGs and experimentally tested for binding affinity by surface plasmon resonance (SPR) and for developability. The highest-affinity antibodies were further validated by KinExA. The session walks through the competition design and the final results. Methods and concepts covered: blinded prospective benchmarking of AI antibody design, NGS selection-output analysis, affinity maturation, HCDR3 clustering, full-length IgG expression, affinity measurement by SPR, KinExA affinity validation, and developability assessment. Hosted by MiLaboratories, the creator of Platforma — the discovery decision platform that takes teams from NGS data to developable, diverse lead candidates, built on the open-source MiXCR toolkit. Platforma is free for academic research. Learn more: https://platforma.bio Part of the High-Affinity Talks series by MiLaboratories.