Converge Bio

One prompt, four antibody seeds, an eight-hour run, ten variants per seed, four new biobetters. Last month, we reported the creation of a cetuximab biobetter using a single ConvergeAB prompt (biorxiv.org/content/10.64898…). That result naturally raises a question: does the same approach generalize to other targets? To find out, we designed an antibody optimization campaign to stress-test the platform. We selected four starting antibodies spanning the full optimization continuum: * A discovery-stage PD-L1 binder (22.5 nM) * A validated Spike binder from a relatively young antibody class (7.5 nM) * A HER2 binder from a class with decades of antibody engineering (23.9 nM) * Tafasitamab, an FDA-approved CD19 therapeutic already at sub-nanomolar affinity (0.56 nM) Each represents a different potential failure mode for an antibody foundation model: limited affinity headroom, sparse historical data, or overfitting to specific antigens. The experimental design was simple: For each seed antibody, ConvergeAB was prompted with the antibody sequence and its matching antigen sequence. The run takes approximately eight hours. The result is a ranked list of candidates. We selected the top ten variants (twelve for CD19, because we had some extra room in the plate..) without additional curation, and sent them directly for experimental validation. In the lab, the variants and their parent antibodies were expressed and purified side-by-side. Binding affinity was measured by SPR, and biophysical properties were evaluated across the panel. Results We identified a higher-affinity binder in every campaign: *PD-L1: 22.5 nM → 0.04 nM (556× improvement in KD, 6 edits) Spike: 7.5 nM → 0.29 nM (26×, 9 edits) HER2: 23.9 nM → 9.28 nM (2.6×, 6 edits) CD19 (tafasitamab): 0.56 nM → 0.21 nM (2.7×, 10 edits) Additional highlights: *Every target produced multiple variants that outperformed the parent antibody (see figure below). *For all four campaigns, the lead variant matched or improved upon the developability profile of the starting antibody. *The sequence edits were not random. Our model was at libert to edit any position in the antibody, and yet roughly one-third of the mutations in the best-performing variants occurred in Vernier-zone residues, the framework positions that shape and orient CDR loops, representing approximately a 5-fold enrichment over chance. The full case study, including methods, sequences, number of failures, and additional analysis is available on our website (converge-bio.com/case-studie…). Talk to me if you'd like to evaluate ConvergeAB on your own antibody optimization challenge (one prompt, an eight-hour run, and ten variants to test in the lab).