CELL INTELLIGENCE · 肿瘤细胞治疗研究

Identifying patient-specific cancer driver genes and their upstream modulators remains challenging due to temporal heterogeneity and the lack of matched multi-omics data. Existing methods often rely on large cohorts with DNA or perturbation assays, treat gene expression as static correlation rather than directed influence, or recover broad regulatory modules without resolving causal, time-resolved driver-modulator relationships. To overcome these limitations, we introduce CancerTrace , a time-aware computational framework that operates directly on single-cell RNA sequencing (scRNA-seq) data to recover dynamic, patient-specific regulatory mechanisms. CancerTrace integrates Transfer Entropy and sparse conditional structure within a variational Bayesian model to (i) isolate the malignant compartment, (ii) reconstruct stage-resolved expression dynamics, and (iii) map directed influence from modulators to drivers. It yields ranked drivers, driver-centered influence networks, and in-silico perturbation readouts without requiring matched DNA or external perturbations. Applied to nine longitudinal scRNA-seq libraries from three lung adenocarcinoma (LUAD) patients, CancerTrace identified an EpCAM -anchored epithelial compartment and recovered both canonical and novel drivers. Over half of the top epithelial drivers corresponded to known oncogenes or tumor suppressors, while VPS37D and ATP11 AUN emerged as new candidates. Driver coefficients for TP53INP1 , CA12 , and CCNL1 aligned with known biology, and epithelial drivers exerted measurable influence on NK cells, consistent with their stage-wise decline. By leveraging temporal single-cell structure, CancerTrace transcends the static and cohort-dependent limitations of existing tools, inferring causal, time-directed driver-modulator relationships that advance mechanistic understanding and precision oncology.