OncoBERT: Context-Aware Modeling of Somatic Mutations for Precision Oncology

Area of Science:

• Genomics
• Computational Biology
• Oncology

Background:

• Somatic mutation profiling is crucial for cancer diagnosis and treatment.
• Current methods often overlook the broader mutational context influencing tumor evolution and treatment outcomes.

Purpose of the Study:

• To introduce OncoBERT, a language model for learning contextual representations of somatic mutations.
• To uncover patient-specific mutational subtypes and their clinical significance.

Main Methods:

• Utilized large-scale clinical sequencing data from >210,000 patients across 113 cancer types.
• Developed OncoBERT to learn contextual mutation representations.
• Integrated OncoBERT outputs with clinical biomarkers (TMB, MSI) and transcriptomic data.

Main Results:

• OncoBERT identified robust, patient-specific mutational subtypes across diverse cohorts.
• These subtypes revealed patterns associated with differential responses to chemotherapy, targeted therapies, and immunotherapy.
• Integration with existing biomarkers significantly improved prediction of immunotherapy response.
• Linked mutational subtypes to distinct cancer hallmark programs and tumor microenvironment states.

Conclusions:

• OncoBERT offers a scalable framework for analyzing somatic mutational landscapes.
• This approach enables improved patient stratification for precision oncology.
• OncoBERT advances the understanding of tumor evolution and treatment response.