Preprint shows AI agents improving protein binders through repeated lab tests

AI agents improved the share of molecular candidates that bound target proteins over three rounds of design and laboratory testing, researchers report in a preprint. The experiment used aptamers, molecules designed to bind proteins, against α-thrombin and myoglobin. Bead-corrected on-target hit rates for new α-thrombin candidates rose from 0.0% to 26.0% and then 42.1%; the final rate was 37.1% excluding descendants of published motifs. Myoglobin hit rates moved from 10.2% to 3.7% and then 33.5%. The agents also used experimental feedback to diagnose cross-reactivity and restore myoglobin specificity from 19.1% to 72.4%. However, newly designed scaffolds showed weaker binding than agent-engineered versions of published motifs in α-thrombin assays.