TLsub: A transfer learning based enhancement to accurately detect mutations with wide-spectrum sub-clonal proportion

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Mutation detection is crucial for sequencing data analysis.
• Subclonal mutations, linked to tumor recurrence and metastasis, are often missed by current tools due to low signal and complex interactions.
• Existing methods struggle with ambiguous sample purities and clonal proportions, leading to false positives when thresholds are lowered.

Purpose of the Study:

• To develop a novel machine learning approach for high-specificity subclonal mutation detection.
• To accurately identify mutations across a wide spectrum of subclonal proportions, even with ambiguous sample purities.
• To filter false positive mutation calls in next-generation sequencing data.

Main Methods:

• Proposed a novel machine learning model for filtering false positive mutation calls.
• Tested the method on both simulated and real sequencing datasets.
• Compared performance against state-of-the-art tools: freebayes, MuTect2, Sentieon, and SiNVICT.

Main Results:

• The proposed method demonstrates adaptability to various diluted sequencing signals.
• Significantly reduces false positives in subclonal mutation detection compared to existing approaches.
• Maintains high specificity even with ambiguous sample purities or low clonal proportions.

Conclusions:

• The developed machine learning approach accurately detects subclonal mutations with high specificity.
• This method offers a significant improvement for analyzing sequencing data with low clonal proportions.
• The tool is available for academic use to advance cancer research.