TranSiGen is a VAE-based model that learns three distributions simultaneously: basal profiles without perturbation, perturbational profiles, and the mapping relationship between them. It utilizes self-supervised representation learning to mitigate noise effects in transcriptional profiles and uncover perturbation signals. The model uses transcriptional profiles from the CMAP LINCS 2020 dataset, consisting of 978 landmark genes per profile. TranSiGen processes the data to ensure one pair of basal and perturbational profiles per perturbation on each cell line. The model consists of two VAE models encoding basal and perturbational profiles and learns to map from basal profiles and perturbation representation to perturbational profiles. TranSiGen demonstrates excellent performance in reconstructing profiles and inferring differential expression genes. It effectively learns cellular and compound representations from data.
TranSiGen outperforms baseline models in predicting DEGs for unseen compounds and cell lines. It demonstrates superior generalizability across cell lines and effectively leverages basal cell profiles. TranSiGen-derived representation is effective in ligand-based virtual screening, drug response prediction, and phenotype-based drug repurposing for pancreatic cancer. It can differentiate between sensitive and resistant compounds, prioritize compounds for disease treatment, and identify potent candidate compounds for pancreatic cancer. TranSiGen’s phenotype-based strategies outperform structure-based methods and successfully identify active compounds with unique mechanisms of action. The model enhances drug discovery efficiency and minimizes costs by expanding the range of compounds that can be screened with predicted perturbational profiles. Overall, TranSiGen’s self-supervised representation learning approach is effective in transcriptional profiling and has broad applications in drug discovery and disease treatment.
Source link
Source link: https://www.nature.com/articles/s41467-024-49620-3