Cancer is a complex disease affected by heterogeneous genetic alterations, the cell type of origin, and interactions with the tumor microenvironment. Understanding how molecular perturbations propagate through cellular networks is therefore central to enable precision cancer medicine. Our research group develops mechanistic artificial intelligence frameworks (Mechanistic AI) that integrate biological knowledge with deep learning to model intracellular signaling, metabolism, and gene regulation. These models are designed to generate biologically interpretable predictions using large-scale multi-omics datasets, including transcriptomics, proteomics, metabolomics, and functional genomics data. Our long-term goal is to develop virtual cell models capable of predicting systems-level responses to mutations, drugs, and environmental perturbations. This project hosts activities related to developing scalable and biologically meaningful representations of drugs, proteins, and genomic perturbations. We will investigate how embeddings of drugs and proteins from different foundation models can be used with deep learning models to predict drug–protein and protein–protein interactions. We will also investigate efficient sequence-to-function modeling to account for mutations and other genomic variation. Recent deep learning models have demonstrated strong performance in predicting functional genomic signals directly from DNA sequence, including chromatin accessibility and transcriptional activity. However, these approaches are computationally expensive and are therefore difficult to utilize as building blocks in system-level frameworks with end-to-end learning. We will investigate novel architectures that improve computational efficiency while preserving most of the predictive accuracy. Together, these activities will contribute key components to mechanistic AI models of cancer by improving how molecular entities, interactions, and genomic perturbations are represented. The requested GPU resources are essential for running foundation-models to generate embeddings, and for training and benchmarking large-scale deep learning models. Ultimately, this work aims to enable computer-aided design of precision cancer medicine (CAD-PCM).