Realizing the promise of computational prediction in toxicology applications.

Abstract

Toxicant screening is only as efficient and effective as the underlying methods. Unfortunately, chemical toxicity screening is predominated by slow, laborious, and costly methods some of which raise ethical concerns. Certain methods involve animal testing, and others involve tedious in vitro work. Accurate toxicity prediction is needed to enable regulatory decision making and for accelerating the drug development process. Current standard wet laboratory methods cannot keep pace with the increasingly varied panoply of potential toxicants that both human beings and fellow wildlife are bathed in. Other fields have benefitted from faster compute times as well as algorithmic advances in artificial intelligence. The increased computational power, improvement in computational methods, and increasing availability of databases have empowered a new age of toxicology prediction. Many computational predictive tools recognize the potential toxicants far faster and for less cost than an in vitro or in vivo assay possibly can, while still providing mechanistic insights. In this issue of JESH-C, we published two reviews on the newest advanced algorithms for toxicity prediction. Tang et al. focused on deep learning and detailed how the advent of this novel computational method combined with recent massive datasets enables increasingly accurate prediction. The authors reviewed big data sources relevant to the reader looking to feed a toxicology-centered deep learning algorithm and outlined the use of neural networks as a tool to construct quantitative structure–activity relationship (QSAR) models. Building on this, Idakwo et al. zoomed in on machine learning applications for the toxicity prediction field. Data cleaning is absolutely critical in any computational prediction method, and the authors provided a very helpful overview on this topic. Concentrating on a specific toxicological aspect, Sakkiah et al. detailed the utility of computational methods for predicting endocrine disrupting chemicals. Here instead of predicting general toxicity, the focus was on predicting chemicals which could bind to the estrogen receptor, the androgen receptor, alpha-phetoprotein, or other specific endocrine targets. The models reviewed in this paper could likely be applied to other toxicology prediction cases where a short list of targets of concern can readily be generated. As mentioned by Tang et al., Idakwo et al., and Sakkiah et al., current computational methods show great promise but are faced by a number of challenges. In this issue, Li et al. presented a novel computational toolkit called Target-specific Toxicity Knowledgebase (TsTKb) to address shortcomings of previous works. They curated various molecular descriptors from more than 100,000 chemicals across datasets and conducted molecular modeling to determine protein–ligand interactions. Building on such a rich compendium of datasets, they outperformed traditional QSAR modeling. Similarly, Chierici et al. presented ML4Tox, a framework that enables deep learning