Background : Predicting toxicological outcomes is crucial for human health and safety. Hight hroughput transcriptomics provides vast gene expression data requiring advanced analytical approaches for meaningful toxicity prediction.
Methods: We leveraged autoencoders, unsupervised deep learning models, to analyze transcriptomics data. These encoder-decoder architectures compress gene expression profiles into latent representations, learning normal patterns and identifying reconstruction errors indicating potential toxicity.
Results: Autoencoders effectively managed batch effects and dataset variability, improving predictive robustness. The approach distinguished minimally annotated compounds and detected subtle gene expression alterations, particularly for liver toxicity before clinical symptoms appear.
Conclusion: Autoencoders demonstrate transformative potential in toxicology research, outperforming conventional methods by capturing complex relationships and enabling precise, early toxicity detection for improved safety assessments.
