A Predictive Ligand-Based Bayesian Model for Human Drug-Induced Liver Injury

Abstract

ABSTRACT:Drug-induced liver injury (DILI) is one of the most important reasons for drug development failure at both preapproval and postapproval stages. There has been increased interest in developing predictive in vivo, in vitro, and in silico models to identify compounds that cause idiosyncratic hepatotoxicity. In the current study, we applied machine learning, a Bayesian modeling method with extended connectivity fingerprints and other interpretable descriptors. The model that was developed and internally validated (using a training set of 295 compounds) was then applied to a large test set relative to the training set (237 compounds) for external validation. The resulting concordance of 60%, sensitivity of 56%, and specificity of 67% were comparable to results for internal validation. The Bayesian model with extended connectivity functional class fingerprints of maximum diameter 6 (ECFC_6) and interpretable descriptors suggested several substructures that are chemically reactive and may also be important for DILI-causing compounds, e.g., ketones, diols, and ␣-methyl styrene type structures. Using Smiles Arbitrary Target Specification (SMARTS) filters published by several pharmaceutical companies, we evaluated whether such reactive substructures could be readily detected by any of the published filters. It was apparent that the most stringent filters used in this study, such as the Abbott alerts, which captures thiol traps and other compounds, may be of use in identifying DILI-causing compounds (sensitivity 67%). A significant outcome of the present study is that we provide predictions for many compounds that cause DILI by using the knowledge we have available from previous studies. These computational models may represent cost-effective selection criteria before in vitro or in vivo experimental studies.

A testing strategy to predict risk for drug-induced liver injury in humans using high-content screen assays and the ‘rule-of-two’ model

“…Indeed, in silico models can assist in the prediction of safety of new drug candidates and allow the rapid screening of literately unlimited number of chemicals. Thus, several studies reported the utility of in silico models for the prediction of human hepatotoxicity (Ekins et al 2010;Greene et al 2010;Chen et al 2013b), and recently, Chen et al reported for oral medications at high daily doses and lipophilicity a significant increased risk for clinical DILI as determined by the 'rule-of-two' (RO2, i.e., daily dose ≥100 mg/day & logP ≥ 3) (Chen et al 2013a). While the RO2 model alone added value to the prediction of clinically relevant human hepatotoxicity (Kaplowitz 2013), its performance was still insufficient due to the limited sensitivity caused by a high rate of false negatives.…”

In silicomodels for drug-induced liver injury – current status

“…This trend is reflected in predictive models for DILI, most approaches being statistical using different methods (e.g. discriminant analysis [52], Bayesian models [65], artificial neural networks (ANN) [66], k-nearest neighbour quantitative structure--activity relationship (QSAR) [67,68], random forest (RF) [68] or QSAR software [69]), whereas only two published techniques involve developing structural alerts implemented in the knowledge-based expert systems (the Vertex cheminformatics platform (VERDI) [57] and Derek Nexus, formerly Derek for Windows [58]). …”

“…The elevated levels of five liver enzymes, together with a composite module, were chosen to detect hepatotoxic drugs. Another example of a DILI dataset is that of more than 300 drugs and chemicals with a classification scheme based on clinical data of hepatotoxicity [33], later expanded by Ekins et al with a further 237 compounds [65]. Compounds were defined as being DILI positive if they were i) withdrawn from the market due mainly to hepatotoxicity, ii) not marketed in the US due to hepatotoxicity, iii) received black box warnings from the US FDA due to hepatotoxicity, iv) marketed with hepatotoxicity warnings on their labels, v) had well-known associations to liver injury and had a significant number (> 10) of independent clinical reports of serious hepatotoxicity that met the Hy's Law criteria.…”

“…Table 1 summarises these models with details about methodology, data and predictive statistics. Only two in silico predictive methods used in vitro data [87,88], one approach utilised mixed data (in vitro together with in vivo) [86] and the remaining models were built using in vivo data [14,57,58,[65][66][67][68][69]. With regard to the prediction of in vivo hepatotoxicity, one model was developed based strictly on intrinsic hepatotoxicity data [14]; remaining models utilised idiosyncratic or mostly idiosyncratic data [57,58,[65][66][67][68][69].…”

“…Compounds were defined as being DILI positive if they were i) withdrawn from the market due mainly to hepatotoxicity, ii) not marketed in the US due to hepatotoxicity, iii) received black box warnings from the US FDA due to hepatotoxicity, iv) marketed with hepatotoxicity warnings on their labels, v) had well-known associations to liver injury and had a significant number (> 10) of independent clinical reports of serious hepatotoxicity that met the Hy's Law criteria. Drugs that did not meet any of these positive criteria were classified as DILI negatives [65].…”