Gaussian Embedding for Large-scale Gene Set Analysis.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in biology

Background:

• High-throughput biological data has led to a proliferation of gene sets (e.g., protein complexes, signaling pathways).
• Extracting biological insights from gene sets requires computational methods for machine learning compatibility.
• Existing methods often represent gene sets as single points, potentially losing network information.

Purpose of the Study:

• To develop a novel network-based gene set embedding approach compatible with machine learning models.
• To represent gene sets as multivariate Gaussian distributions reflecting gene proximity in protein-protein interaction networks.
• To demonstrate the utility of this approach for biological discovery and clinical applications.

Main Methods:

• Introduced Set2Gaussian, a network-based embedding method.
• Represented gene sets as multivariate Gaussian distributions based on gene proximity in protein-protein interaction networks.
• Evaluated Set2Gaussian on gene set member identification, tumor stratification, and gene set enrichment analysis.

Main Results:

• Set2Gaussian improved gene set member identification.
• The method accurately stratified tumors.
• It identified concise gene sets for enrichment analysis.
• Set2Gaussian uncovered a novel NEFM-centric clinical prognostic and predictive subnetwork in sarcoma.
• This subnetwork was validated in independent cohorts.

Conclusions:

• Set2Gaussian provides a powerful new method for embedding gene sets into machine learning frameworks.
• The approach enhances biological discovery by leveraging network information.
• Set2Gaussian has potential clinical applications, as demonstrated by the identification and validation of a novel sarcoma subnetwork.