Abstract
<jats:p>Generative protein design can produce thousands to hundreds of thousands of candidates for a single target, making accurate experimental prioritization a central bottleneck. Existing protein Bayesian optimization and active-learning workflows generally require initial measurements from the current target and therefore retain a cold-start problem. Here, we introduce Active Selection with Prior-Informed Ranking for Experiments (ASPIRE), which combines a transferable cross-target ranking prior with sparse within-target preference updates. ASPIRE learns target-binder ranking from multi-target interaction data and assigns candidate priorities before target-specific measurements are available. Experimental outcomes are converted into preference relations within the same target and assay method, enabling subsequent ranking updates. Across four parallel protein landscapes containing 2,000-149,361 candidates, first-round ASPIRE selection identified candidates within the top 0.055%-3.54% of the true landscape using batches of four or ten. Iterative updates further improved candidate ranks and outperformed random-initialized Gaussian-process, prior-initialized Gaussian-process, and random-initialized Bradley-Terry/Laplace controls. In a prospective ASPIRE-guided VISTA-binding peptide campaign, four of eight synthesized candidates produced clear binding signals. A ninth lead advanced to kinetic validation yielded an affinity range of 1.62-2.31 pM. ASPIRE therefore improves protein-binder prioritization from the first experimental round, with reduced experimental burden emerging as a practical consequence of more accurate selection.</jats:p>