Mentor Name: Nai-Kong Cheung

Therapy resistance and relapse remain the primary causes of mortality in high-risk pediatric neuroblastoma. Increasing evidence suggests that these outcomes are not driven solely by new genetic mutations, but by reversible cell-state transitions between adrenergic (ADR) and mesenchymal (MES) phenotypes. MES-like tumor cells are associated with chemotherapy resistance, immune evasion, and relapse, yet the mechanisms that enable tumor cells to switch between states remain poorly understood. Current computational tools can describe tumor heterogeneity but assume linear, irreversible trajectories and therefore fail to capture the cyclic, feedback-driven dynamics that characterize cancer plasticity. This project will develop and apply a novel computational framework to identify regulators of reversible ADR–MES transitions in pediatric neuroblastoma. Using publicly available and institutional single-cell RNA sequencing datasets, we will first estimate RNA velocity to infer local transcriptional change. Rather than relying on traditional pseudotime methods that assume unidirectional progression, we will construct a continuous-time Markov chain (CTMC) model that explicitly represents bidirectional and cyclic state transitions across tumor cell populations. This probabilistic framework enables modeling of dynamic plasticity rather than static cell states.

Building upon this transition graph, we will implement a graph recurrent neural network (GRNN) to identify genes that drive state switching. By integrating state-level gene expression with transition probabilities, the GRNN will learn regulatory patterns associated with ADR?MES and MES?ADR transitions. Candidate regulators will be prioritized based on model importance scores and expression contrasts between source and target states. Known markers such as PHOX2B and PRRX1 will serve as internal validation, while novel high-ranking genes will represent testable hypotheses for future functional studies. This work addresses a critical unmet need in pediatric oncology: understanding non-genetic mechanisms of treatment resistance. By moving beyond static tumor classification toward quantitative modeling of plasticity, this project aims to identify regulatory drivers that may be therapeutically targetable. In collaboration with clinical investigators, findings from this study will inform future validation in patient-derived organoids and xenograft models. Ultimately, this framework has the potential to improve risk stratification and guide rational combination therapies designed to prevent relapse in children with neuroblastoma.