Expanding biological space coverage enhances the prediction of drug adverse effects in human using in vitro activity profiles

Abstract

In vitro assay data have recently emerged as a potential alternative to traditional animal toxicity studies to aid in the prediction of adverse effects of chemicals on humans. Here we evaluate the data generated from a battery of quantitative high-throughput screening (qHTS) assays applied to a large and diverse collection of chemicals, including approved drugs, for their capacity in predicting human toxicity. Models were built with animal in vivo toxicity data, in vitro human cell-based assay data, as well as in combination with chemical structure and/or drug-target information to predict adverse effects observed for drugs in humans. Interestingly, we found that the models built with the human cell-based assay data performed close to those of the models based on animal in vivo toxicity data. Furthermore, expanding the biological space coverage of assays by including additional drug-target annotations was shown to significantly improve model performance. We identified a small set of targets, which, when added to the current suite of in vitro human cell-based assay data, result in models that greatly outperform those built with the existing animal toxicity data. Assays can be developed for this set of targets to screen compounds for construction of robust models for human toxicity prediction.

Bioactivity Signatures of Drugs vs. Environmental Chemicals Revealed by Tox21 High-Throughput Screening Assays

“…In vitro Assay and Structure Data qHTS data generated from the Tox21 10K collection up to the end of 2018 were used for modeling, including 70 assays with 213 readouts ( Supplementary Table 1) (Attene-Ramos et al, 2013;Huang et al, 2016Huang et al, , 2018. All data and detailed descriptions of these assays with target annotations are publicly available through the NCATS website (https://tripod.nih.gov/ tox21/assays/) and PubChem (Wang et al, 2012;PubChem, 2016).…”

Molecular Image-Based Prediction Models of Nuclear Receptor Agonists and Antagonists Using the DeepSnap-Deep Learning Approach with the Tox21 10K Library

“…These bioassay data consisted of quantitative HTS (qHTS) data derived from two cell-based reporter gene assays, including beta-lactamase or luciferase reporter genes. The activity of these reporter genes is controlled by the binding of transcriptional factors induced or suppressed by an agonist/antagonist with response elements ( The chemicals were derived from the Tox21 10K library, which contains approximately 8900 unique compounds gathered from commercial sources, such as pesticides, industrial and environmental chemicals, natural dietary supplement products, food additives, and drugs, by the NTP, the National Center for Advancing Translational Sciences (NCATS), and the EPA (Table 1) [71][72][73][74][75][76][77][78][79][80][81][82]. These compounds were dissolved in dimethyl sulfoxide (DMSO) as stock solutions, and compound plates with the different concentrations were prepared in the 1536-well plate format [71][72][73]80,83].…”

“…The cells were dispensed at 1500 to 5000 cells/5 (for antagonist mode) or 6 (for agonist mode) microL/well in 1536-well black wall/clear bottom plates [72,73,75,[78][79][80]. After the cells were incubated at 37 • C for 5 to 6 h depending on the particular NR cell line to allow for cell attachment, 23 nL of the compounds at different concentrations were transferred to the assay plates.…”

“…For the antagonist mode assay, the known agonist for each NR was added into the assay plates. For the agonist and antagonist mode assays, positive control compounds were dispensed into each other's wells on the plates (Table 1) [72,73,75,[78][79][80]. The plates were incubated for 16 to 18 h at 37 • C depending on the particular NR cell line.…”

Exploring the Chemical Space of Cytochrome P450 Inhibitors Using Integrated Physicochemical Parameters, Drug Efficiency Metrics and Decision Tree Models

“…Furthermore, probing toxicological profiles of new chemical entities remains an important cornerstone of the drug development process which is experimentally expensive and the translation of animal model results to humans are also challenging. Therefore, a number of in silico models based on machine learning techniques using different combinations of data types have been developed for toxicity prediction of thousands of NCEs/drugs yet with their own strengths and weaknesses [125]. Herein, we presented a data set of 6999 inhibitors against five CYP isoforms mainly CYP1A2,2C9,2C19,2D6, and 3A4 with known activity values.…”