PrismEXP: gene annotation prediction from stratified gene-gene co-expression matrices

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Gene-gene co-expression correlations from RNA-sequencing (RNA-seq) can predict gene annotations.
• Prior work showed uniformly aligned RNA-seq data predicts gene annotations and protein-protein interactions.
• Prediction accuracy varies based on tissue-specific versus agnostic annotations; tissue-specific data offers potential for higher accuracy but optimal partitioning is challenging.

Purpose of the Study:

• Introduce and validate PrismEXP (PRediction of gene Insights from Stratified Mammalian gene co-EXPression) for enhanced gene annotation predictions.
• Improve accuracy by leveraging stratified gene-gene co-expression data.
• Facilitate understanding of understudied genes and proteins.

Main Methods:

• Utilized uniformly aligned RNA-seq data from the ARCHS4 database.
• Applied the PrismEXP approach to stratify gene-gene co-expression data.
• Predicted gene annotations including pathway membership, Gene Ontology terms, and human/mouse phenotypes.

Main Results:

• PrismEXP predictions outperformed predictions from a global cross-tissue co-expression correlation matrix across all tested domains.
• Training on one annotation domain enabled predictions in other domains.
• Demonstrated utility of PrismEXP predictions in multiple use cases.

Conclusions:

• PrismEXP enhances unsupervised machine learning for gene and protein function discovery.
• The tool is accessible via a user-friendly web interface, Python package, and Appyter.
• PrismEXP provides a valuable resource for genomic and functional annotation.