Abstract
In vitro to in vivo extrapolation (IVIVE) of toxicogenomics (TGx) data is essential for enhancing mechanism-based toxicity evaluations and minimizing animal use. However, translating in vitro findings to in vivo responses remains challenging. Generative adversarial networks (GANs) show potential in synthesizing gene expression data but often miss subtle, toxicologically relevant signals. We developed AIVIVE (artificial intelligence-aided IVIVE), a novel framework integrating GANs with local optimizers guided by biologically relevant gene modules to improve prediction accuracy. AIVIVE was trained using rat liver in vitro and in vivo transcriptomic data from the Open TG-GATEs (Toxicogenomics Project-Genomics-Assisted Toxicity Evaluation System) database. AIVIVE was evaluated using cosine similarity, root mean squared error (RMSE), and mean absolute percentage error (MAPE), demonstrating synthetic profiles comparable to real biological replicates. Notably, the model showed high overlap with differentially expressed genes, including Cytochrome P450 enzymes, which are often underrepresented in vitro. AIVIVE recapitulated in vivo CYP expression patterns, overcoming in vitro limitations. Further analysis revealed that AIVIVE captured liver-related pathways like bile secretion, steroid hormone biosynthesis, hepatitis C, and chemical carcinogenesis. It also captured gene expression changes linked to liver-specific adverse outcome pathways, such as Cyp2e1 upregulation in non-alcoholic fatty liver disease. Additionally, AIVIVE slightly outperformed real data in necrosis classification tasks, suggesting its potential for advancing toxicology predictions. These findings support AIVIVE as a tool for generating biologically relevant, in vivo-like profiles from in vitro data to enhance risk assessment, drug safety, and the 3Rs (reduce, replace, refine) principle.