Interpretable Machine Learning for Decoding Cell Dynamics - Interpretable Machine Learning for Decoding Cell Dynamics Abstract Cells constantly change in shape, location, and molecular state, reflecting their dynamic status across time and space. Decoding these cellular dynamics is critical for understanding disease mechanisms and guiding therapeutic strategies. Recent advances in single-cell and spatial omics have generated massive datasets profiling cellular states across tissues, conditions, and disease contexts. Existing methods for analyzing single- cell omics data typically summarize high-dimensional measurements into a vector space, enabling clustering, data integration, or predictive modeling. However, the resulting representations are often entangled, with multiple unrelated biological factors intertwined within the same vector representation. This limits interpretability of how distinct factors, such as regulatory state, genetic perturbations, or tissue microenvironment, influence cellular behaviors. In response, we propose a suite of interpretable machine learning models designed to disentangle these underlying factors and represent them in distinct, biologically meaningful vector spaces. Our approach enables both interpretable and generative modeling of cellular dynamics across diverse single-cell study contexts. First, we will design multimodal models that integrate transcriptomics profiles, imaging data, and text-based knowledge. These models will disentangle gene programs associated with different morphological features, helping to identify drivers of abnormal cellular behaviors and enabling more controlled generation of synthetic spatial transcriptomics data. Second, we will leverage perturbation datasets to model how genetic perturbations, cell intrinsic properties, and environmental changes, as well as their interactions, influence cell-state transitions, which facilitates large scale therapeutic screening. Finally, we will create disentangled cell representation that capture rare or poorly understood cell populations in cell dynamics by isolating subtle signals of interest from major sources of variation. This representation will identify gene programs that define these rare cell population, revealing their functional significance in aging, or disease. Together, this work will establish a new methodological foundation for interpretable modeling of cellular dynamics, advancing our understanding of disease mechanisms and therapeutic intervention. All software and data resources generated will be made publicly available to support the broader scientific community.
