FFixR: A Machine Learning Framework for Accurate Somatic Mutation Calling from FFPE RNA-Seq Data in Cancer.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Formalin-fixed paraffin-embedded (FFPE) tissues are crucial for research and clinical use.
• RNA sequencing (RNA-seq) from FFPE tissues is challenged by artefacts like cytosine deamination and strand-specific damage.
• Existing tools for DNA sequencing are unsuitable for filtering FFPE artefacts in RNA-seq data.

Purpose of the Study:

• To develop a machine learning framework, FFixR, for filtering FFPE-induced artefacts in RNA-seq data.
• To enable accurate somatic variant calling from FFPE RNA-seq without matched-normal samples.

Main Methods:

• FFixR utilizes a machine learning approach trained on FFPE melanoma samples with matched DNA.
• The framework incorporates allele-specific read counts, variant features, and mutational signature probabilities.
• FFixR was evaluated on independent cohorts to assess its performance.

Main Results:

• FFixR effectively removed up to 98% of artefactual mutations while retaining approximately 92% of true variants.
• SHAP analysis identified key features influencing the model's decision-making process.
• Application to independent cohorts restored the correlation between RNA- and DNA-derived tumor mutational burden (R2 = 0.881) and recovered meaningful mutational signatures.

Conclusions:

• FFixR provides an accurate method for somatic variant calling from FFPE RNA-seq data.
• The tool enhances the utility of archival FFPE samples for both research and clinical applications.
• FFixR is freely available, facilitating its adoption in the scientific community.