John C. Faver , Flora Sundersingh , Lauren A. Viarengo-Baker , Ying-Chu Chen , Katelyn Billings , Patrick F. Riley , Ching-Hsuan Tsai , Christopher S. Kollmann

Journal of Medicinal Chemistry

DOI: 10.1021/acs.jmedchem.5c02259

Abstract

DNA-encoded chemical libraries (DELs) enable the highly efficient screening of billions of small molecules for binding to a target of interest and provide valuable training data for machine learning models for virtual screening. However, DEL screening data are notoriously noisy due in large part to significant variance in the synthetic yield of library members. Here, we show an analysis from a split-sample DEL screening strategy against Bruton’s tyrosine kinase (BTK), which includes a panel of affinity selections against the target at varying concentrations and a probabilistic model to estimate the binding affinity and relative input concentrations of library members. We compared model predictions to SPR measurements of resynthesized DNA-conjugated compounds and found that this methodology yielded an improved ranking of library members by binding affinity compared to enrichment metrics. Additionally, the method successfully recovered a library member with a potent binding affinity that would not have been detected in our standard DEL selection.