Improved functional prediction of proteins by learning kernel combinations in multilabel settings.

Area of Science:

• Bioinformatics
• Computational Biology
• Proteomics

Background:

• Developing computational models to predict protein function is crucial for understanding biological systems.
• Existing methods often struggle with the inherent complexity of protein function, which can involve multiple roles or labels.
• Kernel methods offer a powerful framework for integrating diverse biological data to infer protein properties.

Purpose of the Study:

• To develop an advanced probabilistic model for combining kernel matrices to predict protein function.
• To extend existing approaches by explicitly handling multiple labels associated with protein functions.
• To improve the accuracy and interpretability of protein function prediction models.

Main Methods:

• A probabilistic model was developed to combine multiple kernel matrices.
• The model was extended to explicitly handle multiple labels (multilabels) in protein function prediction.
• Simultaneous prediction of subcellular localization and combination of pairwise classifiers were employed to enhance model consistency and interpretability.

Main Results:

• Explicit modeling of multilabels significantly improved protein function prediction from multiple kernels.
• Simultaneous prediction of subcellular localization enhanced model performance and interpretability.
• Combining pairwise classifiers led to consistent class membership estimates, improving overall prediction reliability.

Conclusions:

• Multilabel information is valuable for protein functional prediction and should be integrated into classifier training.
• Co-predicting subcellular localization alongside functional categories improves learning.
• Pairwise separation rules offer detailed insights into the relevance of various data types (sequence, structure, interactions, expression) for protein function prediction.