Lung cancer subtype diagnosis using weakly-paired multi-omics data.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate cancer subtype diagnosis is essential for effective, personalized treatment strategies.
• Multiomics data fusion enhances diagnostic accuracy but is challenged by incomplete and weakly paired datasets.
• Existing incomplete multiview learning methods often overlook data individuality and lack interpretability.

Purpose of the Study:

• To develop an interpretable and flexible solution for lung cancer subtype diagnosis using weakly paired multiomics data.
• To address limitations of current methods by focusing on both shared and individual omics information.
• To improve diagnostic accuracy and feature interpretability in lung cancer.

Main Methods:

• LungDWM utilizes an attention-based encoder for each omics to identify key diagnostic features and extract shared/complementary information.
• An individual loss function is proposed to capture the unique information within each omics dataset.
• Generative adversarial learning imputes missing omics data, followed by feature fusion for subtype diagnosis.

Main Results:

• LungDWM demonstrates superior performance compared to existing competitive methods on benchmark datasets.
• The method achieves high authenticity in its diagnostic predictions.
• LungDWM provides good interpretability of the extracted features for precise diagnosis.

Conclusions:

• LungDWM offers an effective and interpretable approach for lung cancer subtype diagnosis with weakly paired multiomics data.
• The method successfully integrates shared and individual omics information, outperforming previous techniques.
• The developed solution enhances diagnostic precision and provides interpretable features crucial for clinical application.