Hit discovery is often the first and most critical step in a drug discovery campaign upon target identification and characterization. This is often achieved by virtually screening (VS), via protein-ligand docking, large chemical libraries (~10⁶) in order to narrow down a vast pool of compounds to a more manageable set of potential hits. However with the advent of make-on-demand ultra large chemical libraries the possible chemical search space has increased by orders of magnitude (~10⁹), and deployment of traditional VS methods has become prohibitively long and costly. To address this we have developed a proprietary solution called DockAI which is a fast, low cost, and highly effective AI-based method for VS of ultra large scale virtual compound databases. It is based on an active learning approach which allows to identify most promising compounds for further evaluation by docking a small fraction (typically <1%). It works by iteratively sampling the database and using the results of each docking round to guide the selection of next set of compounds to be docked, allowing ultimately to converge on the most promising candidates. DockAI is also fully compatible with molecules generated by our de novo design approach for synthesizable (virtual) molecules based on available building blocks and chemistry not constrained by any database, allowing exploration of a previously unexplored chemical space and discovering compounds with high activity and novelty. DockAI’s machine learning and active learning models outperformed Schrodinger's comparable approach on a dataset of Dopamine receptor compounds (n~140 million), retrieving 75% of hits after docking 1 million compounds. DockAI has been successfully implemented to two commercial projects with other projects under way. In its current technical implementation, DockAI is deployed on AWS enabling docking up to 10⁵ molecules in parallel, providing results of our screening campaigns within 24 hours. Drug discovery campaigns are venturing into novel modalities, drugging more challenging targets, and cost-effective and technically efficient docking tools like DockAI are expected to become a part of the solution.

Nicolas Do Huu

Appearances