Multiclass Cancer Prediction Based on Copy Number Variation Using Deep Learning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• DNA copy number variation (CNV) is linked to human diseases.
• Traditional machine learning for cancer classification relies on manual feature engineering.
• Existing deep learning approaches also incorporate feature extraction steps.

Purpose of the Study:

• To develop and compare end-to-end deep learning models for cancer classification using CNV data.
• To leverage representation learning for automatic feature extraction from genomic data.
• To identify the optimal deep learning architecture for classifying cancer types based on CNV profiles.

Main Methods:

• Developed three end-to-end deep learning models: DNN, CNN, and RNN.
• Utilized CNV data from 24,174 genes for classifying six cancer types.
• Compared model performance against state-of-the-art techniques.

Main Results:

• The best performing model achieved 92% accuracy and an ROC of 0.99.
• Proposed deep learning models outperformed existing state-of-the-art methods in accuracy, precision, and ROC.
• Demonstrated the effectiveness of end-to-end deep learning for CNV-based cancer classification.

Conclusions:

• End-to-end deep learning models offer a powerful approach for cancer classification using CNV data.
• Automatic feature extraction via representation learning enhances classification performance.
• The developed models show significant potential for genomic data analysis in oncology.