A Machine Learning Crowdsourcing Initiative to Accelerate DEL-ML Hit Finding in Drug Discovery

Rapid progress in artificial intelligence (AI) and machine learning (ML) is expected to dramatically accelerate the way therapeutic strategies and drugs are discovered. One of the first steps in a drug discovery campaign is to identify bioactive “hit” molecules that can be optimized into drugs. DNA-encoded libraries coupled with ML (DEL-ML) is a novel hit-finding technique where ML models are trained on ultra-large DEL screening data and used to predict hits from catalogs of readily available molecules. However, best strategies for data representation, model architecture, or hit selection are not yet clearly established. In this first open DEL-ML Target 2035 drug discovery challenge, AI experts combined their ML models with computational chemistry approaches to discover hits for the target protein WDR91, both retrospectively and prospectively. A great variety of computational strategies and tools were deployed, and novel WDR91 ligands were discovered. While a clear differentiating factor was the choice in data to train ML models, variations in fingerprint representation or model architecture had comparatively little impact on the outcome, emphasizing the need for impactful AI-driven technological improvement.