Deep learning for cancer type classification and driver gene identification.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Genetic information is increasingly vital for cancer type and subtype prediction.
• Current classification methods often rely on somatic mutations, limiting their scope and power.
• A need exists for novel methods to integrate diverse genetic variants for enhanced cancer prediction.

Purpose of the Study:

• To develop a novel deep learning model, DeepCues, for cancer type prediction.
• To explore the utility of raw DNA sequencing data, including germline variants and small insertions/deletions.
• To improve feature extraction from genetic data for more accurate cancer classification.

Main Methods:

• Proposed DeepCues, a deep learning model employing convolutional neural networks.
• Utilized raw whole-exome sequencing data as input features.
• Amalgamated germline variants and somatic mutations (including insertions and deletions) for feature generation.

Main Results:

• Achieved 77.6% overall accuracy in classifying seven major cancer types using TCGA data.
• Demonstrated significant improvement over conventional methods (p < 0.001).
• Identified top 20 breast cancer relevant genes with 40% overlap with known driver genes.

Conclusions:

• DeepCues enhances the representational resolution of DNA sequencing data.
• The model effectively derives features from raw sequences for cancer type prediction.
• DeepCues shows promise in discovering novel cancer-relevant genes.