DEL-Ranking framework for ranking-aware denoising of DNA-encoded library screens for improved affinity prediction

DNA-encoded library (DEL) screening links small molecules to unique DNA barcodes, enabling massively parallel measurements of compound-protein interactions in which sequencing read counts serve as surrogates for binding affinity. Converting these counts into reliable affinity estimates remains challenging because stochastic Poisson noise dominates low-abundance compounds, while synthesis, selection, and amplification biases can decouple observed enrichment from true thermodynamic binding strength. Here, we present DEL-Ranking, an integrated computational framework that learns stable relative orderings through complementary pairwise and listwise ranking objectives and incorporates a chemical-referenced correction (CRC) module to recalibrate count predictions using structure-activity priors. Across five carbonic anhydrase datasets, DEL-Ranking consistently improves agreement with experimental K i values, including a 0.304-unit gain in Spearman correlation under high-noise overexpression conditions, and independently prioritizes pyrimidine sulfonamide—a known bioisostere of benzene sulfonamide—without explicit supervision for that scaffold. These results show that coupling ranking-aware denoising with chemically informed correction yields more reliable, interpretable, and validation-ready DEL hit lists.