Physics-informed AI with chemical master equation dynamics for driver-gene subclone detection and risk labeling

Area of Science:

• Computational Biology
• Genomics
• Systems Biology

Background:

• Subclonal populations significantly influence cancer progression and treatment outcomes.
• Resolving rare subclones within bulk transcriptomes is challenging due to signal dilution.
• Existing methods often lack dynamic modeling or expression-defined states, leading to unstable signatures.

Purpose of the Study:

• To introduce magicSubclonal, a physics-informed framework for subclone discovery and risk prediction.
• To integrate driver gene dynamics using a Chemical Master Equation for improved subclone resolution.
• To develop a robust method for linking subclone states to clinical outcomes.

Main Methods:

• Utilized a Chemical Master Equation to model driver-gene dynamics, estimating decay, burst initiation, and burst size from gene expression.
• Developed an automated method for selecting optimal time points for rare-state separation.
• Integrated driver-timed states with non-driver genes using False Discovery Rate control and assigned clinical risk via stabilized Cox/logistic models.

Main Results:

• magicSubclonal demonstrated plausible parameter estimations (e.g., half-lives) and well-calibrated predictions across multiple cancer cohorts (ovarian, lung, breast).
• Achieved superior performance in identifying subclone driver relevance (SDRS) and predictive accuracy (ROC, precision-recall) compared to existing methods like sciClone, NMF, ss-Deconv, and MM.
• Performance gains were particularly notable at low false-positive rates and early recall, indicating enhanced signal detection beyond static mixture models.

Conclusions:

• magicSubclonal offers interpretable and reproducible subclone discovery by integrating stochastic driver dynamics with population heterogeneity.
• The framework provides robust risk labeling by anchoring evaluation to clinical outcomes.
• Sensitivity analyses confirm the model's reliance on burst initiation and size for short-term predictions, with decay becoming more critical over longer time scales.