An In Silico Model for Predicting Drug-Induced Hepatotoxicity

Abstract

As one of the leading causes of drug failure in clinical trials, drug-induced liver injury (DILI) seriously impeded the development of new drugs. Assessing the DILI risk of drug candidates in advance has been considered as an effective strategy to decrease the rate of attrition in drug discovery. Recently, there have been continuous attempts in the prediction of DILI. However, it indeed remains a huge challenge to predict DILI successfully. There is an urgent need to develop a quantitative structure–activity relationship (QSAR) model for predicting DILI with satisfactory performance. In this work, we reported a high-quality QSAR model for predicting the DILI risk of xenobiotics by incorporating the use of eight effective classifiers and molecular descriptors provided by Marvin. In model development, a large-scale and diverse dataset consisting of 1254 compounds for DILI was built through a comprehensive literature retrieval. The optimal model was attained by an ensemble method, averaging the probabilities from eight classifiers, with accuracy (ACC) of 0.783, sensitivity (SE) of 0.818, specificity (SP) of 0.748, and area under the receiver operating characteristic curve (AUC) of 0.859. For further validation, three external test sets and a large negative dataset were utilized. Consequently, both the internal and external validation indicated that our model outperformed prior studies significantly. Data provided by the current study will also be a valuable source for modeling/data mining in the future.

Computational Models Using Multiple Machine Learning Algorithms for Predicting Drug Hepatotoxicity with the DILIrank Dataset

“…Predicting liver toxicity from chemical structure has been a preoccupation of over two decades, first starting somewhat timidly with local models [44,45], to move later also to global models [46,47]. Many of the previous models used relatively small size datasets (less than 400 compounds) [48,49]; studies based on datasets larger than DILIrank have also been published [11,17,[50][51][52], but the authors of DILIrank used a methodology that, in theory at least, was superior and more consistent with the totality of available clinical data. The total DILIrank dataset includes 1036 compounds, but 254 were classified as of "ambiguous DILI-concern", because the causality evidence was limited; besides, it took great care in defining DILI negatives, which varied among four large sources previously studied, labeling some compounds previously considered of no DILI concern as of less DILI concern.…”

“…A review found, in 2014, that in models published up to that time point, the external validation datasets had been "quite small" (20-50 drugs), and that for the larger external datasets the model performance seemed to be "less favorable" [53]. One of the best-performing ensemble models recently published used three external validation datasets and had a pooled balanced accuracy of 71.60% [17]. Although based on a different dataset (DILIrank), our model compares favourably, with a balanced accuracy in the nested cross-validation (i.e., an average of 10 external datasets) that is slightly superior for several meta-models attempted and higher than 74% (as shown in the results section).…”

“…A first annotated dataset on DILI originated at the FDA and was limited to a small number of 287 active substances; more recently (2016), a group of researchers from the FDA published an improved annotated list of medicines with respect to their DILI risk, constituting "the largest reference drug list ranked by the risk for developing drug-induced liver injury in humans" (DILIrank) [13]. Few Quantitative Structure-Activity Relationship (QSAR) studies have focused exclusively on the DILIrank dataset up to date [14]; published studies have either used other datasets [15], only a subset of the DILIrank [16], or have pooled DILIrank with other data sources [17]. This latter approach may have advantages (increasing the dataset and allowing the development of more robust models), but it also has shortcomings (misclassification bias due to different criteria in annotating drugs from different lists).…”

Prediction and mechanistic analysis of Drug-Induced Liver Injury (DILI) based on chemical structure

“…Major challenges associated with the prediction of DILI using in vitro approaches lie in identifying relevant assays (5) and extrapolating from assay concentrations to in vivo blood concentrations associated with a hepatotoxic risk (6). Numerous in silico models have been generated based on molecular structure (7)(8)(9)(10)(11)(12) and in vitro readouts, such as bioactivity (13) and gene expression (14) in cell culture, which are able to predict DILI better than random, but with a…”

“…In the case of computational predictions, DILI is often simplified to a classification problem, i.e. separating compounds with or without this annotation in a data set (7)(8)(9)11,13). These labels, however, do not provide information on important factors such as dose-dependency or affected patient population, and consequently the practical applicability of such models is limited.…”