Spatially-Informed AI to Dissect Complex Cell-State Transitions in Tissue Niches - Single-cell RNA-sequencing (scRNA-seq) and spatial transcriptomics (ST) have independently emerged as transformative technologies for studying cellular heterogeneity, tissue organization, and cell-state dynamics. Evidence strongly suggests that the transcriptional states of cells are intricately linked to their spatial context within tissues, and that cells dynamically change in response to stimuli from neighboring cells and the environment. Tools to decipher the transcriptional dynamics of cells in specific tissue niches from single timepoint data have broad potential impacts in both basic and translational biology; for example, detection of early transcriptional changes in a cell-state trajectory and their dependence on tissue location will lead to more accurate hypotheses about the initiation of homeostatic or pathogenic cell-state transitions, tissue-state biomarkers, and therapeutic targets. However, such tools are broadly lacking, due in part to the relatively low coverage or resolution of the most widely used ST platforms. While scRNA-seq has both deeper coverage and single-cell resolution, its pseudotime trajectory inference (TI) tools, including RNA velocity, estimate transitions without spatial context. Hence, there is a critical need for new quantitative tools that jointly analyze ST and scRNA-seq data to estimate and compare niche-dependent dynamics as statistical evidence. Our proposal aims to overcome critical conceptual and methodological barriers in the development of inference methods for spatial transcriptional dynamics by: (i) learning spatially aware gene programs shared across modalities (Aim 1); (ii) inferring process-specific dynamics and embedding them into sample- or context-specific transition matrices (Aim 2); and (iii) transforming those matrices into quantitative evidence -- effect sizes, confidence intervals, and reproducible group contrasts -- rather than visual streamlines alone (Aim 3). Our preliminary data demonstrates that striking spatially organized cell-state trajectories can be detected, even in low-resolution ST data, using an ad hoc combination of single-modality methods that our proposal will develop into a robust framework for joint inference. As testbeds for methods development, validation, and generalization, we will leverage high-quality publicly available datasets that cover diverse ST technologies and biological systems, as well as collect ST and scRNA-seq data from an immune-infiltrated mouse tumor model in which our team has expertise. Methodologically, our proposed methods will establish a framework for interpretable, joint ST and scRNA-seq inference that could be leveraged to further improve spatial methods, e.g., to combine spatial RNA velocity with multi-timepoint spatial TI based on optimal transport. Thus, they have the potential to uncover a highly under-explored space of dynamic, spatially dependent, cellular behaviors in a wide range of biological systems. Furthermore, the results will support more precise, translationally meaningful predictions of the outcomes of perturbations and help identify spatial biomarkers for clinically relevant tissue states.
