Toxicity problems cause around 40% of drug candidates to be discontinued, often after significant resources have been invested. Furthermore, drug-induced liver injury is the most common reason why already approved drugs are withdrawn. Frequently, drug toxicity is due to electrophilic drugs (and drug metabolites) that covalently bind to nucleophilic sites within biological macromolecules, including DNA and proteins. DNA conjugation can cause cancer, and protein conjugation can cause a toxic immune response. Nucleophiles conjugate to specific atoms within reactive molecules: their sites of reactivity. Identification of a molecule’s site of reactivity suggests the mechanism of its toxicity. Sites of reactivity encompass both harder electrophiles (such as saturated aldehydes) and softer electrophiles (such as alpha, beta-unsaturated aldehydes). Hard electrophiles generally react with hard nucleophiles, such as purine and pyrimidine bases in DNA, whereas soft electrophiles tend to react with GSH and cysteine residues in proteins. In vitro techniques to detect hard and soft electrophiles, including incubation with the nucleophiles cyanide and GSH, are very expensive to perform for thousands of drug candidates. Furthermore, many of these assays do not identify sites of reactivity, which can be sometimes be modified to make drugs safer. This work offers a computational alternative that can be used to quickly screen molecules for both hard and soft reactivity. From a literature-derived database, we extracted around 2000 reactions of drug-like molecules with cyanide, DNA, GSH, and protein. For each of these reactions, we labeled the exact site of reactivity on the starting molecule. This data was inputted into a single deep neural network that predicts these experimentally-observed sites of reactivity for cyanide, DNA, GSH, and protein with cross-validated accuracies of 96.9%, 90.2%, 93.4%, 94.4%, respectively. This approach outperformed individual models for each type of conjugation, due to the machine learning concept of transference learning. Furthermore, the model calculates a molecule reactivity score (MRS) that across a structurally diverse dataset of aldehydes, benzoquinones, and esters predicted the rate of glutathione reactivity with an R-squared of 0.86. Site of reactivity data is an underutilized resource that can be used to not only predict if molecules are reactive, but also how they might be modified to reduce toxicity while retaining efficacy.

Modeling reactivity to soft, hard, and biological targets with a deep learning network

251st National Meeting (2016)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.