Modeling Cell Fate and Perturbation Dynamics with Generative AI using Single-cell and Spatial Omics Data - Project Summary/Abstract Biological systems are inherently dynamic, spatially structured, and governed by complex nonlinear gene regulatory programs. Traditional computational models often struggle to capture the full extent of this complexity, particularly the context-dependent ways in which cells interact, differentiate, and respond to perturbations over time and space. Exploring disease mechanisms and optimizing therapeutic interventions typically requires labor- intensive, large-scale experimental screens. I hypothesize that conditional generative AI models, specifically diffusion-based architectures that incorporate molecular, spatial, and temporal information from high-throughput screening technologies, can learn biologically grounded latent representations of gene regulatory dynamics to accurately forecast cell state transitions over time and across perturbative contexts. These models will enable the discovery of causal molecular drivers of fate decisions and therapeutic responses, generating testable predictions for development, disease, and regenerative processes. Building on my PhD work in the labs of Dr. Elham Azizi and Dr. Kam Leong at Columbia and my postdoctoral work with Dr. James Zou and Dr. Stephen Quake at Stanford, I leverage interdisciplinary expertise in disease modeling, computational biology, and artificial intelligence. I have recently developed a conditional diffusion-based generative AI model for predicting cellular transcriptomics. In my K99 mentored phase, I will extend this work to build a generative time-series virtual cell foundation model (Aim 1) that generates transcriptomic trajectories and predicts cell fate transitions across species and developmental systems, in collaboration with Dr. Zhi Huang (University of Pennsylvania) and Dr. Xiao Yang (Johns Hopkins). In the R00 independent phase, I will extend this platform to incorporate spatial and multimodal data, including spatial transcriptomics and histology, to simulate virtual cell behavior within tissue contexts (Aim 2). Applications will include modeling spatiotemporal brain aging and generating spatial transcriptome conditioned on histopathology features in cancer microenvironments, in collaboration with Dr. Fan Rong (Yale), Dr. Alexander Y. Rudensky and Dr. George Plitas (MSK), and faculty at my future institution. To ground the models in biological realism and translational relevance, I will apply this comprehensive platform to predict genetic response to therapeutic perturbations, identifying key molecular mechanisms underlying drug sensitivity and resistance (Aim 3), in collaboration with Dr. Dhaval Shah (Buffalo) and clinical partners, using preclinical and patient-derived datasets for model validation. We anticipate that this work will yield novel insights into cell fate decisions, tissue remodeling, and support therapeutic development. During the K99 phase, I will be supported by an outstanding interdisciplinary mentorship team, strong institutional support from Stanford Biomedical Data Science, with formal coursework and training. I will bridge my knowledge gap in AI applications for regenerative and precision medicine and gain the communication and leadership skills essential for a successful transition to independence.
