RFcaller: a machine learning approach combined with read-level features to detect somatic mutations

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning in Oncology

Background:

• Decreasing sequencing costs and comprehensive cancer genomics are driving wider adoption of tumor sequencing in research and clinics.
• Existing somatic mutation identification pipelines show variable results, necessitating multi-caller approaches that are computationally intensive and imperfect.
• Expert review of mutation calls is crucial for clinical applications but is time-consuming, highlighting the need for automated, accurate solutions.

Purpose of the Study:

• To introduce RFcaller, a machine learning-based pipeline for somatic mutation detection in tumor-normal paired samples.
• To provide an efficient and accurate alternative to existing, resource-intensive mutation callers.
• To improve the reliability of somatic mutation identification for research and clinical use.

Main Methods:

• Development of RFcaller, a pipeline utilizing machine learning algorithms for somatic mutation detection.
• Application of RFcaller to tumor-normal paired samples for identifying substitutions and insertions/deletions (indels).
• Validation of RFcaller's performance against deep sequencing and Sanger sequencing.

Main Results:

• RFcaller demonstrates high accuracy in detecting substitutions and indels from whole genome and exome data.
• The pipeline successfully identifies mutations in driver genes that are often missed by other methods.
• RFcaller operates without requiring substantial computational resources.

Conclusions:

• RFcaller offers a computationally efficient and accurate machine learning approach for somatic mutation detection.
• The pipeline enhances the ability to identify critical driver gene mutations, supporting clinical diagnosis.
• RFcaller represents a significant advancement in the reliable genomic characterization of cancers.