Single-Cell Drug Perturbations Prediction Using Machine Learning

Human biology may be complicated, due in part to the function and interplay of the body's approximately 37 trillion cells, which are organized into tissues, organs, and systems. Recent developments in single-cell technology have offered unmatched insight into the function of cells and tissues at the DNA, RNA, and protein levels. However, utilizing single-cell approaches to generate medications necessitates the mapping of causal linkages between chemical perturbations and the downstream influence on cell state. We leverage the Drug Response Variational Autoencoder (Dr. VAE), Single-Cell Generative(ScGen), Compositional Perturbation Autoencoder(CPA) and Chemical Compositional Perturbation Autoencoder (ChemCPA) prediction because of their adaptability, scalability, and spectacular performance in other domains. Experimental results demonstrate that Dr. VAE and ChemCPA perform better than ScGen and CPA showing that the combined pre-training and the perturbation network produce State-Of-The-Art (SOTA) performance. Notably, the base version of chemCPA outperforms both CPA and ScGen, demonstrating the positive impact of the extra regularisation provided by the perturbation networks on single-cell perturbation models.