Multi-Instance Metric Transfer Learning for Genome-Wide Protein Function Prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Multi-Instance (MI) learning is effective for genome-wide protein function prediction.
• Existing methods often assume identical data distributions between training (source domain) and testing (target domain).
• This assumption is frequently violated in real-world applications, limiting prediction performance.

Purpose of the Study:

• To propose a novel Multi-Instance Metric Transfer Learning (MIMTL) approach.
• To address the challenge of domain shift in genome-wide protein function prediction.
• To enhance the robustness and accuracy of protein function prediction models.

Main Methods:

• MIMTL transfers source domain distribution to the target domain using bag weights.
• A distance metric learning method is constructed with reweighted bags.
• An alternative optimization scheme is developed for MIMTL.

Main Results:

• The proposed MIMTL approach demonstrates effectiveness in genome-wide protein function prediction.
• Experiments on seven real-world organisms show superior performance compared to state-of-the-art methods.
• MIMTL proves efficient and robust in handling domain shift.

Conclusions:

• MIMTL is a promising approach for genome-wide protein function prediction, especially when domain distributions differ.
• The method effectively bridges the gap between source and target domains.
• This work advances the field of protein function prediction through advanced machine learning techniques.