DARKIN: a zero-shot benchmark for phosphosite-dark kinase association using protein language models

Area of Science:

• Computational biology
• Bioinformatics
• Proteomics

Background:

• Protein language models (pLMs) are advanced tools for analyzing protein sequences.
• Evaluating pLM performance across diverse biological tasks requires robust benchmarks.
• Identifying kinases responsible for specific phosphorylation events is a significant challenge in cell signaling.

Purpose of the Study:

• Introduce DARKIN, a novel zero-shot classification benchmark.
• Assess the efficacy of various pLMs in assigning phosphosites to understudied kinases (dark kinases).
• Provide a biologically relevant test bed for exploring dark kinases.

Main Methods:

• Developed a zero-shot classification benchmark (DARKIN) with stratified data splits.
• Evaluated multiple protein language models using k-NN and bilinear classifiers.
• Ensured data splits respected the zero-shot nature and kinase grouping.

Main Results:

• ESM, ProtT5-XL, and SaProt demonstrated superior performance in phosphosite assignment to dark kinases.
• The developed zero-shot classifiers effectively evaluated pLM capabilities.
• DARKIN serves as a challenging benchmark for pLM assessment.

Conclusions:

• DARKIN provides a valuable resource for evaluating pLMs on understudied kinases.
• The benchmark facilitates deeper investigation into dark kinase functions.
• pLMs show significant potential for advancing phosphosite-kinase assignment.