This project consists of three phases. Phase I We will analyze large-scale pharmacogenomic datasets to characterize how intrinsic cellular variation influences drug responses. Using transcriptomic data, we will define continuous measures of cellular state and stratify cell lines into distinct groups for comparative analysis of drug sensitivity and transcriptional response patterns. Phase II We will develop deep generative models, including variational autoencoders, to learn low-dimensional representations of cellular states. These models will be extended to capture how perturbations influence cellular transitions in a state-dependent manner. To account for non-linear and stochastic dynamics, we will incorporate advanced generative approaches capable of modelling complex trajectories in high-dimensional gene expression space. Phase III We will extend this framework to large-scale single-cell and perturbation datasets in primary cell systems [1]. This requires aligning representations learned from bulk and single-cell data and applying trained models to predict perturbation-induced changes in cellular states. Experimental systems and perturbation datasets will be used for validation. The scale of the data and the need to train high-capacity generative models across multiple datasets motivate our request for substantial GPU and storage resources. [1] DeMeo, Benjamin, et al. "Active learning framework leveraging transcriptomics identifies modulators of disease phenotypes." Science 390.6776 (2025): eadi8577.