DIFFUSE: predicting isoform functions from sequences and expression profiles via deep learning

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Alternative splicing generates diverse protein isoforms from a single gene, increasing genomic functional complexity.
• Understanding specific isoform functions is essential but challenging due to limited isoform-level annotations and data integration difficulties.

Purpose of the Study:

• To develop a novel computational approach for accurately predicting isoform functions by integrating sequence and expression data.
• To address the scarcity of isoform-specific functional annotations using a semi-supervised learning strategy.

Main Methods:

• Introduced DIFFUSE (Deep learning-based prediction of IsoForm FUnctions from Sequences and Expression), a hybrid deep neural network (DNN) and conditional random field (CRF) framework.
• Employed an iterative semi-supervised learning algorithm to train the DNN and CRF models, overcoming the lack of ground truth labels.
• Integrated genomic sequences and isoform co-expression relationships for function prediction.

Main Results:

• DIFFUSE achieved high performance in predicting isoform and gene functions, with an average Area Under the Receiver Operating Characteristics Curve (AUC) of 0.840 and Area Under the Precision-Recall Curve (AUPRC) of 0.581 across 4184 Gene Ontology (GO) categories.
• Demonstrated significantly superior performance compared to state-of-the-art methods.
• Validated prediction accuracy through analysis of functional, sequence, expression, and structural similarities, and consistency with known isoform features.

Conclusions:

• DIFFUSE provides an effective computational solution for predicting isoform functions, leveraging diverse data types.
• The developed semi-supervised approach successfully addresses the challenge of limited isoform-level functional annotations.
• The findings advance our ability to understand the functional landscape of alternative splicing and its contribution to genomic diversity.