Taxonomy-aware, disorder-matched benchmarking of phase-separating protein predictors

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Genomics

Background:

• Liquid-liquid phase separation (LLPS) drives cellular organization.
• Predicting phase-separating proteins (PSPs) computationally is crucial for proteome-scale analysis.
• Existing benchmarks for PSP predictors suffer from biases in taxonomy and protein disorder.

Purpose of the Study:

• To develop a more reliable benchmark for evaluating PSP predictors.
• To identify and mitigate biases in current PSP prediction benchmarks.
• To understand taxon-specific and disorder-related variations in PSP prediction.

Main Methods:

• Constructed a taxonomy-aware, disorder-matched benchmark for PSPs.
• Analyzed sequence and biophysical features of PSPs across different taxa.
• Benchmarked nineteen PSP predictors using the new framework.
• Performed disorder-stratified evaluation of PSP predictors.

Main Results:

• The new benchmark minimizes biases from taxonomic origin and intrinsic disorder.
• PSPs exhibit distinct sequence and biophysical features across taxa, but LLPS-associated shifts are conserved.
• PSP predictor performance varies significantly across taxa.
• Proteins lacking intrinsically disordered regions present a greater challenge for PSP prediction methods.

Conclusions:

• The developed benchmark framework reduces shortcut-driven biases in PSP prediction.
• It enables more accurate and interpretable evaluation of PSP predictors.
• The findings guide the development of models that capture transferable LLPS signals, avoiding dataset-specific artifacts.