Occlusion enhanced pan-cancer classification via deep learning.

Area of Science:

• Computational biology and bioinformatics
• Genomics and transcriptomics
• Machine learning in healthcare

Background:

• RNA sequencing (RNA-Seq) offers quantitative RNA expression measurement, surpassing conventional microscopy for cancer diagnosis.
• Current RNA-Seq cancer studies focus on sample classification and marker gene discovery, often using statistical comparisons.
• Traditional methods may miss subtle marker genes and are susceptible to experimental variability.

Purpose of the Study:

• To introduce GENESO, a novel framework for pan-cancer classification and marker gene discovery.
• To leverage deep learning and the occlusion method for improved accuracy and identification of non-differentially expressed marker genes.
• To develop a more efficient and robust approach to cancer subtyping and biomarker identification.

Main Methods:

• Trained a baseline deep Long Short-Term Memory (LSTM) neural network for pan-cancer classification using RNA-Seq data.
• Developed a novel 'Symmetrical Occlusion (SO)' method to quantitatively assess gene importance by simulating gene function gain/loss.
• Utilized identified key genes to train reduced-set LSTM models for enhanced classification performance.

Main Results:

• The baseline LSTM network achieved 96.59% accuracy in pan-cancer classification.
• The GENESO framework with SO improved accuracy to 98.30% while utilizing 67% fewer genes.
• The method successfully identified important marker genes with low expression level differences, validated on single-cell RNA-Seq data.

Conclusions:

• GENESO provides a powerful and efficient framework for pan-cancer classification and marker gene discovery.
• The Symmetrical Occlusion method effectively identifies critical genes, improving model performance and reducing feature dimensionality.
• This approach offers a significant advancement in leveraging RNA-Seq data for precision oncology and biomarker identification.