Inferring ongoing cancer evolution from single tumour biopsies using synthetic supervised learning.

Area of Science:

• Computational biology
• Cancer genomics
• Machine learning

Background:

• Variant allele frequencies (VAF) reflect tumor evolution and subclonal selection.
• Existing VAF analysis methods are often slow, computationally intensive, or inaccurate.

Purpose of the Study:

• To develop a novel supervised learning method, TumE, for inferring cancer evolution from bulk-sequenced tumor biopsies.
• To improve the accuracy and efficiency of detecting positive selection, deconvoluting subclonal populations, and estimating subclone frequencies.

Main Methods:

• TumE integrates simulated cancer evolution models with Bayesian neural networks.
• The method utilizes a synthetic supervised learning approach for training.

Main Results:

• TumE demonstrates significant improvements in accuracy and inference time for detecting positive selection and estimating subclone frequencies.
• Analyses in both synthetic and patient tumors validate the method's performance.
• Transfer learning capabilities within TumE reduce data and computational requirements for related tasks.

Conclusions:

• TumE offers a powerful and efficient framework for cancer evolutionary inference using VAF data.
• The developed library of recyclable deep learning models supports the cancer evolution research community.
• This work lays the foundation for advanced computational methods in cancer genomics to benefit patients.