DeepIII: Predicting Isoform-Isoform Interactions by Deep Neural Networks and Data Fusion

Area of Science:

• Genomics
• Proteomics
• Bioinformatics

Background:

• Alternative splicing allows a single gene to produce multiple protein isoforms with distinct biological functions.
• Isoform-isoform interactions (IIIs) offer a higher-resolution view of cellular mechanisms compared to traditional protein-protein interactions (PPIs).
• Existing knowledge of IIIs is limited, hindering pathway mapping and understanding of protein complexity.

Purpose of the Study:

• To develop a systematic, genome-wide prediction method for isoform-isoform interactions (IIIs).
• To leverage diverse biological data and deep learning for enhanced interaction prediction.
• To improve the understanding of cellular processes and disease mechanisms through detailed interactome analysis.

Main Methods:

• Proposed DeepIII, a deep learning-based approach for predicting IIIs.
• Integrated multiple data sources: RNA-seq, exon array, domain-domain interactions (DDIs), nucleotide, and amino acid sequences.
• Utilized a four-layer deep neural network for representation learning and binary classification of isoform pairs.

Main Results:

• DeepIII demonstrated superior prediction performance compared to existing state-of-the-art methods.
• The III network generated by DeepIII improved the accuracy of isoform function prediction.
• Case studies confirmed DeepIII's ability to distinguish specific interaction partners for different isoforms from the same gene.

Conclusions:

• DeepIII provides a powerful tool for systematic genome-wide prediction of isoform-isoform interactions.
• The method enhances the resolution of interactome studies, crucial for understanding complex biological functions and diseases.
• Accurate prediction of IIIs contributes to a deeper comprehension of protein diversity and cellular pathways.