A novel modality contribution confidence-enhanced multimodal deep learning framework for multiomics data.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Multimodal learning is crucial for bioinformatics classification.
• Existing methods often overlook biases by assuming equal modality contribution.
• There is a need for advanced frameworks to handle inherent biases in multimodal data.

Purpose of the Study:

• To present a novel deep learning framework that enhances multimodal learning by considering modality contribution confidence.
• To improve the fusion space and classification performance on multiomics data.
• To address the limitations of current approaches that overlook biases in multimodal data.

Main Methods:

• Utilized a non-parametric Gaussian Process to assess unimodal confidence and learn within-modality features.
• Employed Kullback-Leibler divergence for multi-modality alignment and cross-modality feature learning.
• Developed a modality contribution confidence-enhanced deep learning framework.

Main Results:

• The proposed framework significantly improved classification performance on multiomics datasets.
• Demonstrated enhanced fusion space through effective modality alignment and feature learning.
• Validated effectiveness across four diverse multiomics datasets (static, DNA, mRNA, miRNA, protein).

Conclusions:

• The modality contribution confidence-enhanced deep learning framework effectively addresses biases in multimodal learning.
• The method offers improved classification accuracy and a more robust fusion space for multiomics data.
• The framework shows practical utility, as evidenced by a case study on blister recovery.