Mapping the diverse topologies of protein-protein interaction fitness landscapes

De novo binder discovery is unpredictable and inefficient due to a lack of quantitative understanding of protein-protein interaction (PPI) sequence-function landscapes. Here, we use our PANCS-Binder technology to perform >1,300 independent selections of various library sizes and compositions of a randomized small protein to identify binders to a panel of 96 distinct target proteins. For successful selections, we discovered reproducible fitness landscapes that group into a few, target-specific, clusters. Each cluster defines a minimal binding motif whose frequency is inversely proportional to the number of specified amino acids (∼2–8) and determines selection success, which is quantifiable by the density of binders to the target within a theoretical sequence space. We leverage these data to develop a supervised contrastive learning approach that discriminates binders from non-binders and demonstrates generalization beyond a threshold amount of data. Together, this framework renders PPI landscapes measurable and predictive, accelerating de novo binder discovery and optimization.