Biological fingerprints beat chemical structure in predicting drug-induced steatosis

Machine-learning models that describe drugs by the proteins and pathways they touch predict which compounds cause fatty liver better than models built from chemical structure alone, according to a new open-access study published on 5 October 2026 in Discover Pharmaceutical Sciences.

Drug-induced liver injury is one of the most common and serious toxicities in drug development, contributing to clinical trial failures and drug withdrawals. Steatosis, an early pathological event that can precede inflammation, arises when drug exposure disrupts mitochondrial function, fatty acid oxidation, lipid synthesis or endoplasmic reticulum homeostasis. Current laboratory assays typically measure lipid accumulation in hepatocyte models such as HepG2 cells or hiPSC-derived hepatocytes, capturing only selected aspects of the underlying biology. Computational approaches complement these assays because machine learning can detect multivariate and nonlinear patterns across large chemical spaces.

To build their dataset, the researchers curated steatosis-positive drugs from the US Food and Drug Administration Adverse Event Reporting System, using three hepatic reaction terms and filtering reports to those submitted by healthcare professionals with a single reaction term and a single suspected active ingredient. Steatosis-negative compounds were drawn from the DILIst dataset, on the assumption that drugs classified as DILI-negative are also steatosis-negative, an approximation the authors acknowledge and evaluated by manually reviewing LiverTox entries. Only small, organic molecules were retained, and glucocorticoids and tetracyclines with steatosis evidence only under specific exposure conditions were kept as negatives because the classification does not account for dose, duration or route.

The final dataset contained 372 unique compounds, 189 negative and 183 positive, a nearly balanced split. Each compound was described by three types of fingerprints: the PubChem chemical fingerprint of 881 structural keys, a target fingerprint covering human proteins with at least medium expression in the liver, and a pathway fingerprint spanning liver pathways from the Reactome database. Compound-target interactions were gathered from ChEMBL, DrugBank, PharmGKB, TTD, IUPHAR/BPS and PubChem Bioassays, with an activity threshold corresponding to an effective concentration of 10 micromolar or lower, and an interactome layer added first-degree protein interaction partners.

Random forest and gradient boosted tree models were trained on each fingerprint type and evaluated with 5×2 repeated cross-validation. Biological fingerprints significantly outperformed the chemical fingerprint in pairwise Wilcoxon signed-rank tests, and the consensus classifier, which combined all three models by majority voting, reached a balanced accuracy of 0.72 and a sensitivity of 0.69, comparable to the individual biological models. A permutation test in which labels were randomly shuffled 50 times produced a mean balanced accuracy of 0.50 against 0.73 for the best model, confirming the performance was well above chance.

To stress-test generalizability, the team withheld the 10 percent most dissimilar compounds, selected by the lowest maximum Tanimoto similarity to their nearest neighbors across the combined fingerprints, as an external test set of 37 compounds. On this challenging split, sensitivity, precision, Matthews correlation coefficient and balanced accuracy all decreased while specificity remained stable. The models were also evaluated on an independent external steatosis dataset from Cotterill and colleagues, reduced to 19 pharmaceutical compounds not present in the training data, and predictive performance was poor for both individual and consensus models, even though the external compounds were not structurally underrepresented in the feature space. The authors note that the relatively limited dataset size of 372 compounds is itself a limitation for assessing generalizability.

Feature importance analysis pointed to transporter proteins and cytochrome P450 enzymes as leading contributors, with the top-ranked pathways all tied to absorption, distribution, metabolism and excretion processes. Hepatic transporters such as ABCB1 and ABCG2 ranked highly, consistent with prior work identifying CYP1A1 as a predictive gene in steatosis models and with reports of stage-dependent changes in ADME gene expression across human fatty liver disease. The highest-ranked molecular initiating event was the nuclear receptor pregnane X receptor, appearing at position 22 with a mean scaled importance of 0.097, though the authors caution that their fingerprints record interactions without distinguishing activation from inhibition, which can dilute the discriminative power of nuclear receptor features.

Beyond the models themselves, the study uncovered a strikingly simple signal: the number of recorded interactions with liver-expressed targets separated steatosis-positive from steatosis-negative compounds, with a median of 14 interactions for positives versus 5 for negatives and an area under the ROC curve above 0.7. This separation persisted when the analysis was restricted to targets shared by both classes, controlling for measurement bias, and it also appeared in the DILIrank dataset. In contrast, discrimination weakened for a cholestasis dataset, for steatosis compounds with targets outside the liver filter, and dropped to near-random for developmental neurotoxicity targets, suggesting the effect is tied to the liver-expressed feature space rather than generic coverage differences. The authors frame this as a polypharmacology signal consistent with the multi-mechanism nature of steatosis, and offer it as a heuristic metric worth exploring for other complex toxicity endpoints.

As a demonstration of practical utility, the framework was applied to compounds with sparse FAERS evidence, fewer than 50 reported cases and no measured interactions with liver-expressed targets. From an initial pool of 2,038 compounds, 60 met the criteria, and predicted target interactions were generated with four computational tools. Three compounds, the antiviral elsulfavirine, the antibiotic fosfomycin and the analgesic clonixin, were flagged as positive across all three assessment layers: consensus predictions within a nuclear receptor molecular initiating event space, the machine-learning models, and interaction counts. Elsulfavirine showed 53 predicted interactions with liver-expressed targets, a figure the authors themselves flag as possibly an artefact of similarity-based target prediction rather than true pharmacology. Fosfomycin does have independent clinical signals, including a monograph recommending hepatic monitoring and a case report of acute hepatitis with focal steatosis after a single 3-gram dose, but the authors stress that all three compounds are hypothesis-generating candidates rather than confirmed steatosis inducers, lacking experimental validation and comprehensive clinical evidence.

The authors are candid about the limitations of their label source. Spontaneous reporting systems such as FAERS suffer from underreporting, uneven drug utilization and uncertain causality in patients taking multiple medications, so FAERS-derived labels should be read as phenotype-associated signals rather than definitive proof of causality, and residual misclassification may shape the patterns the models learn. Nevertheless, the study’s central conclusion stands: integrating chemical, target and pathway information in a consensus framework delivers robust steatosis prediction while offering interpretable features, and biological interaction profiles carry more predictive weight for this toxicity endpoint than chemical structure alone. The KNIME workflows and datasets behind the study are publicly available, allowing other groups to test whether the interaction-count signal and the consensus approach extend to further complex toxicity endpoints.

Subject of Research: Computational prediction of drug-induced steatosis using chemical, target and pathway fingerprints

Article Title: Biological interaction fingerprints improve prediction of drug-induced steatosis

Article References: Helmke, P. S., Oehler, F., & Ecker, G. F. (2026). Biological interaction fingerprints improve prediction of drug-induced steatosis. Discover Pharmaceutical Sciences, 2(1), Article 20. https://doi.org/10.1007/s44395-026-00057-1

Image Credits: AI Generated

DOI: 10.1007/s44395-026-00057-1

Keywords: drug-induced steatosis, drug-induced liver injury, machine learning, hepatic steatosis, FAERS, molecular fingerprints, polypharmacology, toxicology, cytochrome P450, nuclear receptors, ADME, consensus modeling