Refining Drug-Induced Cholestasis Prediction: An Explainable Consensus Model Integrating Chemical and Biological Fingerprints

Abstract

Effective drug safety assessment, guided by the 3R principle (Replacement, Reduction, Refinement) to minimize animal testing, is critical in early drug development. Drug-induced liver injury (DILI), particularly drug-induced cholestasis (DIC), remains a major challenge. This study introduces a computational method for predicting DIC by integrating PubChem substructure fingerprints with biological data from liver-expressed targets and pathways, alongside nine hepatic transporter inhibition models. To address class imbalance in the public cholestasis data set, we employed undersampling, a technique that constructs a small and robust consensus model by evaluating distinct subsets. The most effective baseline model, which combined PubChem substructure fingerprints, pathway data and hepatic transporter inhibition predictions, achieved a Matthews correlation coefficient (MCC) of 0.29 and a sensitivity of 0.79, as validated through 10-fold cross-validation. Subsequently, target prediction using four publicly available tools was employed to enrich the sparse compound-target interaction matrix. Although this approach showed lower sensitivity compared to experimentally derived targets and pathways, it highlighted the value of incorporating specific systems biology related information. Feature importance analysis identified albumin as a potential target linked to cholestasis within our predictive model, suggesting a connection worth further investigation. By employing an expanded consensus model and applying probability range filtering, the refined method achieved an MCC of 0.38 and a sensitivity of 0.80, thereby enhancing decision-making confidence. This approach advances DIC prediction by integrating biological and chemical descriptors, offering a reliable and explainable model.

1

Depiction of the workflow used to generate the models in this study. SMILES were standardized for the generation of transporter inhibition models and PubChem fingerprints, which are steps of our in-house pipeline for chemical fingerprint calculation in modeling tasks. For the retrieval of ChEMBL IDs, drug names were used when available; otherwise, InChiKeys generated from preprocessed SMILES were utilized. Predicted and known targets were merged for the retrieval of pathway fingerprints in the retrained predicted target models. Subsequently, all fingerprints were converted into binary tables, that served as an input for the machine learning models.

2

Model selection according to the highest balanced accuracy.

3

Individual and consensus model building.

4

Illustration of the fingerprint length for both the baseline and target prediction models in comparison to the maximum attainable fingerprint length.

5

Visualization of the feature space of the baseline, +all, +2-C and +3-C models combining all features. (A) Feature space of the baseline data set. (B) +all. (C) +2-C. (D) +3-C. Green indicates cholestasis-negative compounds, red indicates cholestasis-positive compounds.

6

Bar chart showing percent of albumin interacting compounds in the cholestasis and DILI rank data sets in high, medium, low and inconclusive or no evidence plasma protein binding level categories.