MOCapsNet: Multiomics Data Integration for Cancer Subtype Analysis Based on Dynamic Self-Attention Learning and...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Integrating multiomics data aids precise disease treatment but faces challenges like feature redundancy and noise in high-dimensional datasets.
• Simple data concatenation may miss crucial correlations between omics layers, hindering effective information capture.
• Deep neural networks often lack interpretability due to complex structures and numerous parameters.

Purpose of the Study:

• To develop an interpretable method for multiomics data integration specifically for cancer subtype classification.
• To address limitations of existing methods in handling high-dimensional omics data, feature redundancy, and inter-omics correlations.
• To improve the accuracy and interpretability of cancer classification using integrated multiomics data.

Main Methods:

• Proposed MOCapsNet, an interpretable multiomics integration method utilizing self-attention and capsule networks.
• Implemented a self-attention confidence learning module to weight and integrate feature information across different omics data.
• Employed capsule networks for the final cancer classification task, enhancing feature representation.

Main Results:

• Achieved 87.8% accuracy on the BRCA multiclassification dataset.
• Attained 83.6% accuracy and an 88.8% AUC on the LGG dataset.
• Demonstrated consistent effectiveness in integrating multiomics data for improved classification accuracy.

Conclusions:

• The MOCapsNet framework effectively integrates multiomics data, enhancing cancer classification accuracy.
• The method provides improved interpretability of results by leveraging feature information comprehensively.
• The approach shows significant promise for precise disease treatment through advanced data integration techniques.