Abstract
Subclonal populations shape progression and therapy response but are difficult to resolve in bulk transcriptomes, where rare signals are diluted by dominant clones. Existing methods often treat samples as static mixtures, infer lineages without expression-defined states, or rank genes without specifying when to evaluate, yielding unstable signatures. We propose magicSubclonal, a physics-informed framework that embeds driver-gene dynamics using a Chemical Master Equation. From gene expression, we estimate decay, burst initiation, and burst size; automatically select the time of maximal rare-state separation; and identify subclones whose extremes lie within calibrated predictive envelopes. Driver-timed states are then linked to non-driver genes with False Discovery Rate control, and clinical risk is assigned using bootstrap-stabilized Cox/logistic models. Applied to ovarian, lung, Ductal Carcinoma in Situ, and breast cohorts, fits yielded plausible half-lives and well-calibrated predictions. We introduce the Subclone Driver Relevance Score (SDRS), a bounded metric quantifying outcome alignment under a uniform operating point. Across cohorts, magicSubclonal achieved the highest SDRS and superior ROC and precision-recall performance, outperforming variant-allele clustering (sciClone), expression factorization (NMF), and two expression-only baselines: single-sample deconvolution (ss-Deconv) and a diagonal log-Gaussian mixture (MM). Gains were most pronounced at low false-positive rates and early-recall regions, reflecting added signal beyond static mixtures or unsupervised clustering. By integrating stochastic driver dynamics with population-level heterogeneity and outcome-anchored evaluation, magicSubclonal provides interpretable, reproducible subclone discovery and robust risk labeling. Sensitivity analyses confirmed that predictions are driven by burst initiation and size, with decay increasingly influential over longer horizons, supporting cross-cohort consistency under identical preprocessing.