Enhancing cancer classification accuracy with a self-attention network using panel capture sequencing data

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Cancer classification is crucial for precision oncology but challenged by tumor molecular heterogeneity.
• Traditional methods often fall short in accurately classifying diverse cancer types.
• Panel capture sequencing is increasingly used in clinical settings for cancer analysis.

Purpose of the Study:

• To develop and validate a novel machine learning network for precise cancer classification using panel capture sequencing data.
• To improve diagnostic accuracy and identify key molecular drivers across different cancer types.
• To explore the potential clinical utility of advanced machine learning in oncology.

Main Methods:

• Development of a self-attention based Conv1D machine learning network.
• Integration of clinical panel capture sequencing data with The Cancer Genome Atlas (TCGA) data.
• Performance evaluation based on overall accuracy, precision, and recall rates for various cancers.

Main Results:

• Achieved overall cancer classification accuracy exceeding 90%.
• Demonstrated high precision rates (100%) for cervical and gastric cancers.
• Reported robust recall rates, with gastric cancer at 95.79% and cervical cancer at 77.46%.
• Identified key genes (C3orf36, JHY, TASP1) with significant mutation count differences.
• Highlighted critical pathways (acute myeloid leukemia, adipocytokine signaling) via gene enrichment analysis.

Conclusions:

• The developed machine learning approach significantly enhances cancer classification precision.
• The model shows strong potential for clinical application in oncology.
• The study provides deeper insights into cancer biology and identifies novel gene targets and pathways.