Deep Learning for Biomarker Discovery in Cancer Genomes

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genomic data analysis for precision oncology faces challenges with complex bioinformatics pipelines and information loss from feature engineering.
• Traditional methods limit the potential of next-generation sequencing (NGS) data for biomarker extraction and discovery.

Purpose of the Study:

• To develop an end-to-end deep learning (DL) framework for analyzing NGS data.
• To integrate somatic mutation sequences for predicting microsatellite instability (MSI) and homologous recombination deficiency (HRD) biomarkers.

Main Methods:

• A multiple instance learning DL framework was developed.
• Data from 3,184 cancer patients from TCGA and CPTAC databases were utilized.
• The framework integrates somatic mutation sequences to predict MSI and HRD status.

Main Results:

• The DL model achieved high accuracy for MSI (0.98) and HRD (0.80) prediction on an external validation cohort.
• The DL approach significantly outperformed traditional machine learning methods for both MSI and HRD prediction.
• Explainability techniques confirmed that predictions are based on biologically meaningful features.

Conclusions:

• Deep learning can identify patterns in unfiltered somatic mutations without manual feature extraction.
• This approach enhances the detection of actionable targets in precision oncology.
• It enables the development of NGS-based biomarkers using minimally processed data.