ProtDML: label-aware representation learning for broad-spectrum protein function prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Protein function prediction is vital for understanding biological processes, yet experimental annotations lag behind sequence data growth.
• Existing protein language models (pLMs) often fail to capture complex correlations among functional labels, limiting their predictive power.
• There is a need for scalable and accurate methods to predict protein function, especially considering protein multifunctionality.

Purpose of the Study:

• To develop ProtDML, a framework that improves protein function prediction by explicitly modeling label correlations.
• To enhance the modeling of protein multifunctionality through a novel distance metric learning approach.
• To integrate complementary protein representations for improved sequence-based prediction.

Main Methods:

• ProtDML utilizes distance metric learning with a similarity-weighted pull-push objective to model label co-occurrence patterns.
• A multi-view strategy integrates diverse sequence-derived protein representations.
• The framework learns a task-specific distance metric, creating a label-aware feature space.

Main Results:

• ProtDML significantly outperforms sequence-based baselines on Pfam multi-label classification by creating distinct, separable clusters.
• The method demonstrates structural generalization capabilities on SCOP benchmarks without explicit structural input.
• ProtDML achieves state-of-the-art performance on the CAFA3 benchmark, showing robustness in large-scale, imbalanced prediction tasks.

Conclusions:

• ProtDML offers an accurate and scalable solution for sequence-based protein function prediction by effectively modeling label correlations.
• The framework is model-agnostic and complements existing protein language model embeddings.
• ProtDML shows promise for real-world applications, including viral data analysis.