GATSBI: improving context-aware protein embeddings through biologically motivated data splits.

Area of Science:

• Computational biology
• Bioinformatics
• Protein science

Background:

• Understanding protein function necessitates integrating diverse biological data, considering context-specific interactions.
• Current protein embedding methods often use evaluation protocols misaligned with specific biological prediction tasks.
• Performance evaluation frequently focuses on well-studied proteins, potentially overestimating real-world utility.

Purpose of the Study:

• To introduce a novel graph attention-based framework (GATSBI) for constructing context-aware protein embeddings.
• To develop and validate task-aligned evaluation protocols for protein embedding models.
• To improve the prediction of protein interactions, functions, and functional modules, especially for understudied proteins.

Main Methods:

• Integrated diverse biological data: protein-protein interactions, co-expression, sequence representations, and tissue-specific associations.
• Employed a graph attention mechanism to generate protein embeddings.
• Utilized task-aligned evaluation protocols with biologically appropriate data partitions (e.g., node-held-out splits).

Main Results:

• GATSBI generates context-aware protein embeddings that demonstrate superior generalization performance.
• The framework consistently outperforms existing pre-trained embeddings across interaction, function, and functional set prediction tasks.
• Significant performance gains were observed for understudied proteins, particularly under inductive node-held-out evaluation.

Conclusions:

• Task-aligned evaluation is crucial for developing effective protein embedding models.
• GATSBI provides a robust method for generating protein embeddings that benefit diverse biological prediction tasks.
• The learned embeddings are publicly available to facilitate broader research applications.