Classification of Cancer Types Using Graph Convolutional Neural Networks

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Cancer remains a leading cause of death, necessitating improved early detection and understanding of its genomic drivers.
• Machine learning and deep learning approaches have been increasingly applied to cancer classification.
• Accurate cancer prediction is vital for improving treatment outcomes and survival rates.

Purpose of the Study:

• To develop and evaluate novel graph convolutional neural network (GCNN) models for classifying 33 cancer types and normal tissue.
• To identify cancer-specific marker genes driving accurate classification.
• To provide publicly available models and code for cancer diagnosis and research.

Main Methods:

• Four GCNN models were designed using gene expression data and graph structures (co-expression, PPI, with singleton features).
• Models were trained and tested on The Cancer Genome Atlas (TCGA) dataset (10,340 cancer samples, 731 normal samples).
• In silico gene-perturbation experiments were conducted for model interpretation.

Main Results:

• GCNN models achieved high prediction accuracies ranging from 89.9% to 94.7% across 34 classes (33 cancer types + normal).
• The co-expression GCNN model identified 428 marker genes crucial for classifying cancer types and normal tissue.
• Identified markers demonstrated cancer-specificity rather than tissue-specificity.

Conclusions:

• Novel GCNN models accurately predict cancer types and normal tissue from gene expression profiles.
• The models achieved over 94% accuracy on the TCGA dataset, utilizing cancer-specific marker genes.
• Publicly available models and source codes facilitate adaptation for cancer diagnosis and other disease research.