Fig. 1: Hit optimization workflow. | Nature Communications

Fig. 1: Hit optimization workflow.

From: Expediting hit-to-lead progression in drug discovery through reaction prediction and multi-dimensional optimization

Fig. 1: Hit optimization workflow.The alternative text for this image may have been generated using AI.

A set of 125 monoacylglycerol lipase (MAGL)-inhibiting “hit" molecules (starting points) containing N-heterocycles are combined with 211 commercially available carboxylic acids to generate 26,375 hypothetical products. These virtual molecules are passed through a series of filters to identify potent, synthetically accessible molecules with favorable physicochemical properties ("machine learning funnel"). Multi-template docking and a machine learning-based potency scoring function are used to identify candidate molecules. Synthetic accessibility is predicted using graph neural networks (GNNs) trained on a novel reaction data set encompassing 13,490 Minisci-type alkylation reactions. Physicochemical properties are determined using various, readily available machine learning-based predictions. Consequently, from 212 potentially suitable compounds, 34 molecule-acid combinations were manually selected and screened using miniaturized high-throughput experimentation (HTE) to identify favorable alkylation reaction conditions. Based on building block availability, up-scaling of a selected number of reactions successfully delivered 14 compounds with improved profiles.

Back to article page