DeepIsoFun: a deep domain adaptation approach to predict isoform functions

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Alternative splicing produces messenger RNA (mRNA) isoforms from the same gene locus, potentially leading to diverse functions.
• While gene functions are well-studied, the specific roles of individual mRNA isoforms remain largely unknown.
• Existing computational methods for isoform function prediction are limited by insufficient labeled training data.

Purpose of the Study:

• To develop a novel deep learning method, DeepIsoFun, for accurate prediction of mRNA isoform functions.
• To address the challenge of limited labeled data in isoform function prediction.
• To leverage domain adaptation to transfer knowledge from gene functions to isoform functions.

Main Methods:

• DeepIsoFun combines multiple instance learning with domain adaptation techniques.
• A deep neural network architecture is employed to adapt to varying expression distributions.
• Domain adaptation provides additional labeled training data by transferring gene function knowledge.

Main Results:

• DeepIsoFun demonstrated significantly superior performance on human and mouse expression datasets compared to existing methods.
• The method achieved at least a 26% improvement in the area under the receiver operating characteristics curve.
• Improvements of at least 10% in the area under the precision-recall curve were observed.

Conclusions:

• DeepIsoFun offers a robust and effective solution for predicting mRNA isoform functions.
• The approach successfully overcomes the limitations of previous methods, particularly regarding data scarcity.
• Further analysis explored the functional divergence of isoforms and the correlation between expression and function similarity.