Modeling and predicting cancer clonal evolution with reinforcement learning.

Area of Science:

• Computational Biology
• Cancer Research
• Evolutionary Biology

Background:

• Cancer evolves through a clonal process, generating tumors with diverse mutations.
• Accurate modeling of cancer evolution is crucial for prediction and understanding.
• Existing models often overfit and support limited prediction tasks.

Purpose of the Study:

• To introduce CloMu, a flexible, low-parameter generative model for cancer evolution.
• To enable diverse prediction tasks including evolutionary trajectories, mutation causality, and fitness.
• To overcome limitations of previous methods, particularly overfitting and restricted functionality.

Main Methods:

• Developed CloMu, a two-layer neural network model trained with reinforcement learning.
• Model infers mutation probabilities based on existing mutations within a clone.
• Evaluated CloMu using simulations and real-world breast cancer and leukemia datasets.

Main Results:

• CloMu matches or surpasses current methods in various prediction tasks.
• Effectively uncovers causal mutation relationships, especially with interchangeable mutations.
• Accurately determines mutation similarities, causal links, and fitness in cancer cohorts.
• Validated mutation fitness predictions against independent leukemia clonal proportion data.

Conclusions:

• CloMu provides a powerful and flexible framework for modeling cancer evolution.
• Its low-parameter approach enhances predictive accuracy and avoids overfitting.
• Enables comprehensive analysis of cancer evolutionary dynamics using cohort data.