Machine Learning-Assisted Iterative Screening for Efficient Detection of Drug Discovery Starting Points
Alex T. Müller, Markus Hierl, Dominik Heer, Paul Westwood, Philippe Hartz, Bigna Wörsdörfer, Christian Krämer, Wolfgang Haap, Doris Roth, Michael Reutlinger
Abstract
High-throughput screening (HTS) remains central to small molecule lead discovery, but increasing assay complexity challenges the screening of large compound libraries. While retrospective studies have assessed active-learning-guided screening, extensive prospective validations are rare. Here, we report the first prospective evaluation of machine learning (ML)-assisted iterative HTS in a large-scale drug discovery project. Using a mass spectrometry-based assay for salt-inducible kinase 2, we screened just 5.9% of a two million-compound library in three batches and recovered 43.3% of all primary actives identified in a parallel full HTS─including all but one compound series selected by medicinal chemists. This demonstrates that ML-guided iterative screening can significantly reduce the experimental cost while maintaining hit discovery quality. Retrospective benchmarks further showed that the ML approach outperforms similarity-based methods in hit recovery and chemical space coverage. In summary, this study highlights the potential of ML-driven iterative HTS to improve efficiency also in large-scale drug discovery projects.