Isoform function prediction by Gene Ontology embedding.

Area of Science:

• Functional genomics
• Computational biology
• Bioinformatics

Background:

• High-resolution gene function annotation is crucial for understanding biological complexity.
• Alternative splicing generates multiple proteoforms (isoforms) from a single gene, increasing functional diversity and impacting complex diseases.
• Existing multi-instance learning (MIL) methods for gene ontology (GO) annotation have limitations in handling massive GO terms and their hierarchical structures.

Purpose of the Study:

• To develop an effective computational approach, IsofunGO, for differentiating massive functions of gene isoforms.
• To leverage GO embedding and attention mechanisms within a multi-instance learning framework for improved isoform function prediction.

Main Methods:

• Developed an attributed hierarchical network to model massive Gene Ontology (GO) terms.
• Implemented a GO network embedding strategy to learn compact GO representations and compress gene GO annotations.
• Designed an attention-based multi-instance learning (MIL) network integrating genomics and transcriptomics data for isoform function prediction.

Main Results:

• IsofunGO successfully differentiates massive isoform functions by utilizing GO embedding and an attention-based MIL network.
• The GO embedding strategy effectively preserves GO term hierarchy and reduces prediction load.
• Extensive experiments validated the efficacy, performance boost, and interpretability enhancement provided by IsofunGO's components.

Conclusions:

• IsofunGO provides an effective solution for high-resolution isoform function annotation.
• The integration of GO embedding and attention mechanisms significantly improves the prediction of isoform functions.
• This approach advances the understanding of molecular mechanisms underlying complex diseases at a finer granularity.