Convolutional neural network models for cancer type prediction based on gene expression.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Accurate cancer type prediction is crucial for diagnosis and treatment.
• Identifying cancer marker genes is a key challenge in cancer research.
• Existing machine learning models may be biased by the tissue of origin.

Purpose of the Study:

• To develop and evaluate Convolutional Neural Network (CNN) models for predicting cancer types from gene expression profiles.
• To identify cancer marker genes by interpreting the CNN models, accounting for tissue of origin.
• To assess the models' performance in classifying tumor vs. non-tumor samples and predicting cancer subtypes.

Main Methods:

• Implemented three CNN models (1D-CNN, 2D-Vanilla-CNN, 2D-Hybrid-CNN) using gene expression data.
• Trained and tested models on 10,340 samples from The Cancer Genome Atlas (TCGA) across 33 cancer types and normal tissues.
• Utilized a guided saliency technique for model interpretation to identify cancer markers.

Main Results:

• Achieved high prediction accuracies (93.9-95.0%) for classifying 34 classes (33 cancers + normal).
• Identified 2090 potential cancer markers (average 108 per class) using the 1D-CNN model.
• Successfully predicted breast cancer subtypes with 88.42% accuracy and identified known markers like GATA3 and ESR1.

Conclusions:

• Novel CNN designs enable accurate, simultaneous prediction of cancer/normal status and cancer types.
• The model interpretation scheme effectively identifies biologically relevant cancer markers, mitigating tissue-of-origin bias.
• The proposed models are adaptable for future cancer diagnosis applications due to their efficient training.