Comprehensive protein datasets and benchmarking for liquid-liquid phase separation studies.

Area of Science:

• Biochemistry
• Molecular Biology
• Cell Biology

Background:

• Proteins form dynamic biomolecular condensates via liquid-liquid phase separation (LLPS) in cells.
• Proteins in condensates act as drivers (forming condensates) or clients (localizing within).
• Existing LLPS databases have data inconsistencies, hindering reliable predictive model development.

Purpose of the Study:

• To create high-confidence datasets of driver and client proteins involved in LLPS.
• To establish standardized negative datasets for non-LLPS proteins.
• To provide a comprehensive benchmark for LLPS predictive algorithms.

Main Methods:

• Integrated biocuration of existing LLPS databases.
• Generation of standardized positive (client/driver) and negative (non-LLPS) protein datasets.
• Analysis of physicochemical properties and benchmarking of 16 predictive algorithms.

Main Results:

• Creation of confident datasets for client and driver proteins.
• Identification of significant physicochemical differences between LLPS and non-LLPS proteins.
• Comprehensive benchmarking revealing limitations in predictive algorithms.

Conclusions:

• The generated datasets provide a reliable resource for studying protein roles in LLPS.
• Physicochemical properties underlying LLPS are distinct among protein subsets.
• This work offers the most extensive benchmark to date for LLPS predictive algorithms.