Evaluation of DNA encoded library and machine learning model combinations for hit discovery
Sumaiya Iqbal, Wei Jiang, Eric R. Hansen, Tonia Aristotelous, Shuang Liu, Andrew G. Reidenbach, Cerise Raffier, Alison Leed, Chengkuan Chen, Lawrence Chung, Eric A. Sigel, Alex B. Burgin, Sandy J. J. Gould, Holly H. Soutter
Abstract
DNA-Encoded Library (DEL) technology allows the screening of millions to billions of compounds in a pooled fashion, which is faster and cheaper than traditional approaches. The massive amounts of DEL binder and not-binder data enable Machine Learning (ML) model development and virtual screening of readily accessible, drug-like libraries in an ultra-high-throughput fashion. Here, we report a comparative assessment of DEL + ML pipeline for hit discovery using three DELs and five ML models (fifteen DEL + ML combinations). Each ML model was used to identify orthosteric binders of two therapeutic targets, Casein kinase 1 α /δ (CK1 α /δ). Overall, 10% and 94% of the predicted binders and not-binders were confirmed in biophysical assays, including two nanomolar binders (187 and 69.6 nM). Our study provides insights into the DEL + ML paradigm for hit discovery: the importance of chemical diversity in training data and ML model generalizability over accuracy. We publicly shared our results for further use and similar developments.