henhaobio
Sign in

Archives of ToxicologyPharmacology & ToxicologyPeer-reviewed journal article

Integration of In Vitro and In Silico Approaches Enables Prediction of Drug-Induced Liver Injury

René GeciiD, Ahenk Zeynep SayiniD, Stephan SchalleriD, Lars KuepferiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Drug-induced liver injury (DILI) is a major cause of drug attrition and poses a significant threat to patient safety. However, current preclinical prediction methods, including heuristic screening rules, in vitro assays, machine learning models and animal testing, have serious limitations. Here, we demonstrate that combining in vitro toxicity data (cytotoxicity, mitochondrial toxicity, bile salt export pump (BSEP) inhibition) with pharmacokinetic information enables high DILI predictivity. In a retrospective analysis of 241 drugs, we show that the ratio of their in vivo maximum plasma concentration (Cmax) to their lowest in vitro toxicity strongly correlates with clinical DILI risks, with ROC AUC up to 96%. Then, we show that comparable predictivity (ROC AUC up to 91%) is achievable prospectively when Cmax values are predicted in silico by high-throughput physiologically based kinetic modelling. Dynamic simulations of bile acid perturbations further identify drugs potentially causing DILI specifically through BSEP inhibition, providing additional mechanistic insights. This integrative, mechanistic approach shows enhanced DILI predictivity and interpretability, offering an animal-free alternative for early drug development.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (27 Oct 2025)

Dividing Cmax by lowest in vitro toxicity potency separates drugs by liver-injury (DILI) risk, AUC up to 96%; PBK-predicted Cmax without clinical data reaches 91%.

The paper describes in vitro hepatotoxicity readouts (cytotoxicity, mitochondrial toxicity, bile salt export pump inhibition) as usually analysed without pharmacokinetic exposure, with limited predictivity for drug-induced liver injury (DILI). It combines them with maximum plasma concentration (Cmax), observed or predicted by high-throughput physiologically based kinetic (HT-PBK) modelling, as an interpretable, animal-free alternative to black-box machine learning.

Key results

  1. ShownSimple heuristics predict DILI poorly. LogP did not separate DILIrank classes, and in vitro toxicity values alone did not beat the heuristics. Observed Cmax was the best single heuristic.

    In vivo Cmax alone separated No-DILI from Most-DILI plus Clinical Development Failure drugs with specificity 89.6% and sensitivity 82.1%. Lowest in vitro potency per compound still did not outperform simple heuristic rules (SI-Table 1). · Evaluation 1: Simple heuristic screening rules…; Tables

  2. ShownThe ratio of observed Cmax to the lowest functional in vitro toxicity value (cytotoxicity, mitochondrial toxicity) tracked DILIrank severity. Ratios rose from No-DILI through Less-DILI to Most-DILI.

    ROC AUC 96% for No-DILI vs Most-DILI plus Clinical Development Failures, 90% for No-DILI vs Less-, Most-DILI and Failures. The mechanistic 25% threshold gave specificity 100% and sensitivity 51%; the Youden-optimal 1.1% gave 82% and 97%. · Evaluation 2: Cmax to in vitro toxicity ratios…; Figure legends

  3. SuggestedReplacing observed Cmax with HT-PBK-predicted Cmax kept most of the predictivity, so prospective use is feasible. It added some false positives and its errors skew towards overprediction.

    90% of drugs had Cmax predicted within 10-fold. AUC was 86% instead of 90%, or 91% instead of 96%. DILI classification changed for 29 of 208 drugs (13.94%). 17 drugs were overpredicted more than ten-fold and 7 underpredicted. · Evaluation 3: In silico predicted Cmax enables…

  4. SuggestedAdding BSEP inhibition (Cmax to BSEP IC50 ratio) also predicted DILI and mostly raised ratios of Less-DILI and cholestatic compounds. Two apparent false positives, mifepristone and primidone, were argued to be DILIrank misclassifications using later case reports.

    ROC AUC up to 91% for Cmax to BSEP IC50 ratios. Mifepristone and primidone were No-DILI in DILIrank but had Cmax above BSEP IC50. The authors cite three 2019 to 2023 mifepristone cholestatic DILI case reports and one primidone case report. · Evaluation 4: Inclusion of BSEP inhibition…; Tables

  5. SuggestedDynamic HT-PBK bile acid simulations largely agreed with the static 25% ratio threshold, but exposed discrepancies. They also yielded a short list of drugs potentially causing DILI only through BSEP inhibition.

    Troglitazone had a Cmax to BSEP IC50 ratio above 1 but negligible simulated bile acid accumulation; bicalutamide had a low ratio but relevant accumulation. Filtering identified eight drugs from 241 (Table 1), all DILI-positive in DILIrank except mifepristone. · Evaluation 5: Dynamic modelling reveals in vivo…; Tables

Methods, data and reproducibility

Seventeen literature in vitro hepatotoxicity datasets were merged with DILIrank labels and manually curated oral dose and Cmax. Ratios of Cmax (observed or HT-PBK-predicted in PK-Sim) to the lowest toxicity potency were evaluated by ROC; a physiologically based bile acid model simulated BSEP inhibition dynamically.

  • Integrated in vitro hepatotoxicity dataset from 17 literature datasets (cytotoxicity, mitochondrial toxicity, BSEP IC50) with DILIrank labels, oral dose and Cmax · human (clinical DILI labels and Cmax) · n = 487 compounds, 355 with potency, 241 final drugs (208 in one analysis)
  • Curated cholestatic vs hepatocellular DILI pattern annotations from SIDER and publications · human
Methods & tools
PK-Sim 11.1.137, R 4.2.2, ADMETLab, ADMET Predictor v12, PKsmart, PubChem, DILIrank, SIDER, Physiologically based bile acid (PBBA) model, LiverTox
Software & environment
PK-Sim version 11.1.137 parameterised from R version 4.2.2; ADMET Predictor v12 used for fasted-state simulated intestinal fluid solubility.
Limitations (5)
  • The in vitro toxicity values span a much narrower range (4 orders of magnitude) than Cmax (7 orders). Highly potent hepatotoxins and very low-potency compounds are absent, which inflates the apparent power of dose or Cmax alone. · stated by the authors
  • The data are sparse and heterogeneous, and not all compounds were tested across all in vitro systems. A single fixed 25% threshold may not be optimal, and dosing schedules, transporters, tissue exposure and synergistic mechanisms were not modelled. · stated by the authors
  • DILI-positive compounds are over-represented, so performance with in silico Cmax may look inflated. In silico Cmax tends to be overestimated. · stated by the authors
  • False negatives such as estradiol and tacrine reflect mechanisms (gene expression changes, metabolite formation) that the in vitro assays do not capture. · stated by the authors
  • The 25% threshold was chosen after manually reviewing false negatives in the same dataset, and there is no external validation set. The reported performance in SI-Table 1 treats mifepristone and primidone as DILI-positive, which departs from DILIrank. · noted in this summary

Read this if you work on animal-free hepatotoxicity prediction and want to see exposure-normalised in vitro toxicity data, including BSEP inhibition, benchmarked against heuristics.

Citations
5
FWCI
32.81
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISSeal '24Zhang '26Zhang '21Talari '24Katritsis '22Umemori '24
84% semantic matchBioinformatics
Srijit Seal, Dominic P. Williams, Layla Hosseini-Gerami+4 more·bioRxiv·Jan 10, 2024· 7 citations