Similarity-driven multi-view embeddings from high-dimensional biomedical data.

Area of Science:

• Computational Biology
• Bioinformatics
• Data Science

Background:

• Large-scale scientific data, including multi-omics and neuroimaging, offers opportunities for integrative hypothesis testing.
• Analyzing high-dimensional, multi-modal data presents significant computational challenges.

Purpose of the Study:

• To introduce Similarity-driven multi-view linear reconstruction (SiMLR), an algorithm for dimensionality reduction of large, diverse scientific datasets.
• To develop a method that exploits inter-modality relationships for joint signal estimation and interpretable low-dimensional representations.

Main Methods:

• SiMLR employs an objective function to identify joint signals across different data modalities.
• It utilizes regularization with sparse matrices to incorporate prior within-modality relationships.
• The implementation supports the joint reduction of large data matrices.

Main Results:

• SiMLR demonstrated superior performance compared to related methods in simulation data.
• The algorithm achieved strong results on a multi-omics cancer survival prediction task.
• SiMLR proved effective on multiple neuroimaging datasets with diverse modalities.

Conclusions:

• SiMLR is a powerful tool for joint signal estimation from disparate scientific data modalities.
• The algorithm provides practically useful results across various application domains, enhancing data interpretability and analytical power.