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

Area of Science:

• Computational biology
• Bioinformatics
• Network science

Background:

• Understanding protein function necessitates integrating diverse biological data, but current protein embedding methods often lack task-specific evaluation.
• Existing methods may overestimate real-world utility by focusing on well-studied proteins and using inappropriate data partitioning strategies.

Purpose of the Study:

• To introduce Gatsbi, a graph attention-based framework for generating context-aware protein embeddings.
• To evaluate Gatsbi using task-aligned protocols that reflect specific biological prediction tasks.
• To improve predictions for both well-studied and understudied proteins.

Main Methods:

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

Main Results:

• Gatsbi consistently outperforms existing pretrained embeddings across interaction, function, and functional set prediction tasks.
• Significant performance gains were observed for understudied proteins, particularly under inductive node-held-out evaluation.
• Task-aligned evaluation demonstrated markedly better generalization compared to standard protocols.

Conclusions:

• Gatsbi provides improved protein embeddings by integrating diverse data and employing appropriate evaluation strategies.
• The framework shows particular promise for advancing research on understudied proteins.
• Learned embeddings are available for broader application in protein prediction tasks.