(T-096) Enabling Earlier Predictions of Hepatotoxic Risk with Quantitative Structure-Activity Relationship-Machine Learning (QSAR-ML) and Quantitative Systems Toxicology (QST) Models

Christina Battista – Simulations Plus; James Beaudoin – Simulations Plus; Michael Lawless – Simulations Plus; Lisl Shoda – Simulations Plus

Objectives: Although in vitro and animal models can help identify compounds with liver liabilities, many liabilities only surface in clinical trials, causing costly delays or even halting development. Commercially available software has previously been used to mechanistically understand and predict clinical liver injury signals and mitigate future risk of hepatotoxicity. This established and successful approach requires more data than what is typically available during early drug development. To facilitate earlier and more affordable hepatotoxicity predictions, we have developed novel QSAR models utilizing ML techniques to provide qualitative and quantitative predictions for active hepatotoxicity mechanisms. These novel models can be used in conjunction with a QST model of DILI to provide further insight into hepatotoxic risk during early drug development.

Methods: In vitro data were collected describing compound effects on hepatotoxicity mechanisms: mitochondrial dysfunction, oxidative stress, and bile acid transporter inhibition, namely inhibition of the bile salt export pump (BSEP). QSAR-ML models were developed for each mechanism. For each mechanism, a classification model was developed to identify whether a given compound is predicted to have an effect on that toxicity mechanism. Quantitative models were also created to predict the minimum effective concentration (MEC) and the half-maximal activity concentration (AC50) for mitochondrial dysfunction and oxidative stress. For BSEP, a QSAR-ML model was developed that predicts the half-maximal inhibitory concentration (IC50). The QSAR-ML models were used to predict and quantify active hepatotoxicity mechanisms for compounds previously assessed with the DILI QST model. Retrospective assessment of two classes of compounds (i.e., CGRP receptor antagonists and macrolide antibiotics) was performed, as well as head-to-head comparisons of compounds with known hepatotoxic risk to those with no-to-minimal hepatotoxic risk. In assessing the QSAR-ML+QST approach, only compound structures and in vitro efficacy measures were used, mimicking the amount of data typically available during early drug development.

Results: Sensitivity and specificity for the three QSAR-ML mechanism classification models were ≥ 75.0% and the mean absolute error for each quantitative model was < 0.67. The QSAR-ML+QST approach predicted clarithromycin and solithromycin to have the highest potential for hepatotoxicity of the macrolide antibiotics tested and telcagepant and rimegepant to have the highest potential for hepatotoxicity of the CGRP receptor antagonists tested. These results aligned with previous QST predictions [1,2]. Notably, the previous QST predictions required extensive collection of in vitro data and knowledge of in vivo exposure to assess hepatotoxicity, whereas the QSAR-ML+QST approach required minimal data.

Conclusions: The QSAR-ML+QST approach exhibits great potential to rank hepatotoxic risk for groups of compounds. The results of these analyses suggest this approach has the potential to inform hepatotoxic risk assessment of new compounds in early drug development, which could result in a significant time and cost savings during overall drug development.

Citations: 1. Woodhead JL, Yang K, Oldach D, MacLauchlin C, Fernandes P, Watkins PB, Siler SQ, Howell BA. Analyzing the Mechanisms Behind Macrolide Antibiotic-Induced Liver Injury Using Quantitative Systems Toxicology Modeling. Pharm Res. 2019 Feb 7;36(3):48. doi: 10.1007/s11095-019-2582-y. PMID: 30734107; PMCID: PMC6373306.
2. Woodhead JL, Siler SQ, Howell BA, Watkins PB, Conway C. Comparing the Liver Safety Profiles of 4 Next-Generation CGRP Receptor Antagonists to the Hepatotoxic CGRP Inhibitor Telcagepant Using Quantitative Systems Toxicology Modeling. Toxicol Sci. 2022 Jun 28;188(1):108-116. doi: 10.1093/toxsci/kfac051. PMID: 35556143; PMCID: PMC9237996.

Keywords: quantitative structure-activity relationship-machine learning (QSAR-ML), drug-induced liver injury (DILI), quantitative systems toxicology (QST)