GOLDEN fusion: a graph-oriented learning with domain-embedding network fusion for generating super gene sets in...

Area of Science:

• Bioinformatics
• Computational Biology
• Translational Biomedical Research

Background:

• Integrative analysis of omics data is crucial for biomedical research.
• Existing gene set enrichment methods often yield redundant results due to high similarity in gene membership and descriptions.
• Pathways, annotated gene lists, and gene signatures (PAGs) are used to represent biological functions, but their integration can be challenging.

Purpose of the Study:

• To develop an integrative framework for interpreting omics data by reducing redundancy in gene set analysis.
• To enhance the utility of diverse gene sets through metadata enrichment and network construction.
• To create concise, higher-level representations of biological pathways for improved interpretability.

Main Methods:

• Introduced Graph-Oriented Learning with Domain-Embedding Network (GOLDEN) fusion.
• Leveraged connection-based embeddings from PAG-PAG relationships and semantic-based embeddings from PAG descriptions using a large language model (LLM).
• Combined embeddings via early fusion, followed by clustering to identify super-PAGs and a Connection Disparity Index to assess clusterability.

Main Results:

• GOLDEN fusion demonstrated superior performance in recovering biological pathway structures compared to connection-only or semantic-only methods on Gene Ontology Annotation benchmarks.
• Achieved consistent gains in Adjusted Rand Index and Normalized Mutual Information, indicating improved accuracy.
• Generated concise summaries for super-PAGs by synthesizing member PAG descriptions using LLMs.

Conclusions:

• GOLDEN fusion provides an effective integrated framework for interpreting omics results by creating concise, higher-level pathway representations.
• The method addresses the challenge of redundancy in gene set enrichment analysis.
• Enables more accurate and interpretable analysis of complex biological data.