Interpretable deep learning methods for multiview learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Technological advancements facilitate the integration of diverse data types (genomics, proteomics, metabolomics).
• Multiview learning research holds significant potential for novel biomedical discoveries.

Purpose of the Study:

• To introduce iDeepViewLearn, an interpretable deep learning method for multiview learning.
• To enable learning of nonlinear relationships and feature selection across multiple data views.

Main Methods:

• iDeepViewLearn employs deep neural networks for view-independent low-dimensional embedding.
• It utilizes an optimization problem minimizing data reconstruction error with regularization.
• Normalized Laplacian is used for feature selection by modeling variable relationships within each view.

Main Results:

• iDeepViewLearn demonstrated competitive classification performance against state-of-the-art methods.
• Clustering analysis identified molecular subtypes in breast cancer associated with survival rates.
• The method successfully reconstructed handwritten images from minimal pixel data.
• iDeepViewLearn shows potential for small-sample size problems in multiview learning.

Conclusions:

• iDeepViewLearn is an innovative deep learning model for capturing nonlinear relationships in multiview data.
• The model effectively performs feature selection, enhancing interpretability.
• It is an open-source tool available for broader research application.