AI designs new antibodies that pass blinded laboratory tests

Researchers from UTHealth Houston, competing under the name Novamab AI, secured a spot in the top five of the international AIntibody Challenge, a rigorous experimental test of AI-designed therapeutic antibodies. The competition, published in Nature Biotechnology, aimed to evaluate artificial intelligence platforms for designing antibodies through laboratory synthesis and testing.

Unlike previous computational benchmarks that utilized historical data, the AIntibody Challenge required teams to create entirely new antibody sequences, which were then independently synthesized and experimentally assessed for binding affinity and developability—critical properties for potential drug candidates.

Novamab AI distinguished itself by fine-tuning a protein language model using experimental preference data, focusing on candidates with proven biological activity and developability rather than solely on theoretical sequence scores. Out of 166 participants submitting 527 designs, Novamab AI's five submissions ranked among the top five performers.

This success marks the first independent, experimental validation that machine learning can accurately predict viable therapeutic antibody candidates before expensive laboratory screening.

The AIntibody Challenge, led by Andrew Bradbury, MD, Ph.D., provided a first-of-its-kind prospective benchmark for assessing AI's potential in antibody discovery. By testing AI predictions and designs under blinded experimental conditions, the challenge set a high standard for future AI approaches in therapeutic drug development. The study, authored by M. Frank Erasmus et al., underscores the promising role of artificial intelligence in accelerating the discovery of reliable and effective therapeutic antibodies.