SNMF: Integrated Learning of Mutational Signatures and Prediction of DNA Repair Deficiencies

Area of Science:

• Genomics
• Cancer Biology
• Bioinformatics

Background:

• Tumors often exhibit DNA damage response (DDR) deficiencies, impacting cancer development and treatment.
• Identifying these deficiencies is crucial for predicting patient response to DDR-targeting therapies.
• Current methods for detecting DDR deficiencies, like mutational signatures, are often identified independently and may not be optimized for specific pathway or gene defects.

Purpose of the Study:

• To develop a novel method, supervised non-negative matrix factorization (SNMF), for jointly learning and optimizing mutational signatures predictive of DDR deficiencies.
• To improve the accuracy and completeness of DDR deficiency signatures compared to unsupervised methods.
• To discern distinct molecular mechanisms within DDR pathways.

Main Methods:

• Applied SNMF to mutation profiles from human induced pluripotent stem cell lines with gene knockouts in three DDR pathways.
• Trained the SNMF model to jointly optimize for signatures shared across samples and predictive of DDR deficiency.
• Validated SNMF-derived signatures against known COSMIC signatures and predicted DDR pathway deficiencies in TCGA tumor data.

Main Results:

• The SNMF model achieved high predictive accuracy (0.971) for DDR deficiency.
• SNMF learned more comprehensive signatures of a sample's DDR status, distinguishing between different mechanisms within pathways.
• Signatures derived from cell line data successfully recapitulated tumor-derived COSMIC signatures and accurately predicted DDR pathway deficiencies in TCGA tumors.

Conclusions:

• SNMF is an effective supervised method for identifying DDR deficiency signatures.
• This approach can leverage induced DDR deficiencies in cell lines to decipher complex signatures in patient tumors.
• SNMF-based models hold promise for improving patient stratification for DDR-targeting cancer therapies.