Joint Representation Learning for Retrieval and Annotation of Genomic Interval Sets

Area of Science:

• Bioinformatics
• Computational Biology
• Genomic Data Analysis

Background:

• Genomic interval data is rapidly increasing, necessitating efficient search systems.
• Current methods like string matching and region overlap analysis have limitations in accuracy and computational cost.
• There is a need for advanced methods to query large, complex genomic interval databases.

Purpose of the Study:

• To develop a novel genomic interval search system using representation learning.
• To capture similarities between genomic region sets and their metadata in a low-dimensional space.
• To enable fast, flexible, and accurate information retrieval for genomic region data.

Main Methods:

• Developed a system that trains numerical embeddings for region sets and metadata simultaneously.
• Utilized representation learning to create co-embeddings in a low-dimensional space.
• Employed embedding distance computations for information retrieval tasks.

Main Results:

• The system successfully performs three key information retrieval tasks: query string matching, label suggestion, and similar region set retrieval.
• Jointly learned representations of region sets and metadata demonstrate effectiveness.
• Achieved fast, flexible, and accurate genomic region information retrieval.

Conclusions:

• Representation learning offers a promising approach for enhancing genomic interval database searches.
• The developed system addresses limitations of existing methods for querying large-scale genomic data.
• This method facilitates more efficient and accurate analysis of genomic region information.