Single-Cell Transcriptomic Signatures are Predictive of Organ-Specific Drug Toxicity

Drug-induced toxicity accounts for approximately 30% of drug development failures1,2. Hepatotoxicity and cardiotoxicity are some of the most common causes, driving clinical trial failures and post-market withdrawals3–5. Existing drug structure-based machine learning models achieve reasonable predictive performance (AUC 0.75–0.83)6,7 but provide no insight into which cell types and tissues are vulnerable to toxic drug effects. Recent single-cell RNA sequencing atlases such as Tabula Sapiens8 have mapped cellular expression across millions of cells from multiple human organs, yet no reproducible framework uses normal tissue single-cell expression profiles to derive cell-type resolved associations with organ-specific drug toxicity. To address this gap, we present a multi-organ single-cell computational framework for drug toxicity risk classification that identifies cell types whose baseline target expression profiles are statistically associated with hepatotoxic and cardiotoxic drugs. We compiled drug toxicity annotations from DILIrank9 (n=217), and DICTrank10 (n=187), with targets retrieved from ChEMBL11 and DrugBank12, and single-cell expression data from Tabula Sapiens across liver (22,214 cells, 24 cell types), and heart (25,832 cells, 22 cell types). For each drug–cell type pair, we computed Drug2Cell13 scores representing aggregate target gene expression, creating organ-specific matrices with log-transformed scores. Differences between toxic and non-toxic drugs were tested by cell type using t-tests with FDR correction. To address non-independence of drugs within therapeutic families, we employed linear mixed-effects models with ATC Level 3 classification as random effects, and trained organ-specific drug toxicity classifiers using stratified splits with oversampling and cross-validation. 14/24 liver cell types remained significant after mixed-effects correction (FDR < 0.05); Hepatocytes (FDR = 0.044) and hepatic stellate cells (FDR = 0.034) were significant, indicating that liver susceptibility is independent of drug class confounding. These findings mirror clinical drug-induced liver injury patterns of hepatocellular injury and fibrotic response. All significant heart cell types were eliminated by mixed-effects correction, indicating that these associations reflected drug-class-specific patterns. Predictive performance of the best models was similar across liver (XGBoost AUC = 0.699 [0.584–0.804]) and heart (Random Forest AUC = 0.680 [0.564–0.788]). Our framework demonstrates that single-cell transcriptomics enable mechanistically interpretable drug toxicity prediction. Although predictive accuracy is modestly lower than structure-based approaches, the value lies in identifying specific cellular vulnerabilities and injury mechanisms that enable hypothesis-driven biomarker development.