Multi-modal Differentiable Unsupervised Feature Selection.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• High-throughput biological data offers vast scientific potential but poses significant computational challenges.
• Multi-modal measurements, where samples are observed by multiple sensors, often contain irrelevant variables (nuisance features).
• Identifying informative features is crucial for extracting meaningful biological insights from complex datasets.

Purpose of the Study:

• To develop a multi-modal unsupervised feature selection framework.
• To identify informative variables from coupled high-dimensional measurements.
• To distinguish between shared and differential latent structures within the data.

Main Methods:

• Proposed a novel framework for multi-modal unsupervised feature selection.
• Introduced two Laplacian-based scoring operators to identify informative features.
• Incorporated differentiable gates to mask nuisance features and enhance graph Laplacian structure accuracy.

Main Results:

• Successfully identified informative features associated with shared and differential latent structures.
• Demonstrated the framework's performance on synthetic and real-world datasets.
• Applied the method to a single-cell multi-omics dataset, showcasing its biological relevance.

Conclusions:

• The proposed framework effectively addresses the computational challenges of multi-modal biological data.
• It enables accurate identification of biologically relevant features from complex, high-dimensional datasets.
• This approach enhances the analysis of single-cell multi-omics and other multi-modal biological measurements.