DMIL-IsoFun: predicting isoform function using deep multi-instance learning.

Area of Science:

• Computational Biology
• Genomics
• Proteomics

Background:

• Alternative splicing generates extensive proteomic diversity from a limited genome.
• Understanding gene functions at the isoform level is crucial for finer biological insights.
• Existing computational methods for isoform function differentiation face limitations due to data imbalance and annotation scarcity.

Purpose of the Study:

• To propose DMIL-IsoFun, a deep multi-instance learning framework to accurately differentiate gene isoform functions.
• To overcome challenges in isoform-level annotation and gene-isoform relation modeling.
• To improve the accuracy and reliability of functional predictions for alternatively spliced isoforms.

Main Methods:

• Developed a deep multi-instance learning framework (DMIL-IsoFun).
• Employed a convolutional neural network (CNN) with isoform sequences and gene-level annotations for feature extraction and initial annotation.
• Utilized a class-imbalanced Graph Convolutional Network (GCN) to refine isoform annotations using co-expression networks and extracted features.

Main Results:

• DMIL-IsoFun significantly improved performance metrics (Smin and Fmax) compared to state-of-the-art methods, with increases of at least 29.6% and 40.8%, respectively.
• The framework demonstrated effectiveness on human multiple-isoform genes.
• Validated on maize isoforms involved in photosynthesis, confirming its broad applicability.

Conclusions:

• DMIL-IsoFun provides a robust computational approach for differentiating gene isoform functions.
• The method effectively handles data imbalance and complex gene-isoform relationships.
• This advancement facilitates a more granular understanding of gene function and proteomic complexity.