Prediction of plant phase-separating proteins using positive-unlabeled learning

Area of Science:

• Biochemistry
• Molecular Biology
• Plant Science

Background:

• Liquid-liquid phase separation (LLPS) is crucial for biological processes, forming dynamic condensates.
• The number of experimentally validated phase-separating proteins in plants is limited, hindering predictive capabilities.
• Understanding LLPS in plants is vital for deciphering cellular organization and function.

Purpose of the Study:

• To develop a predictive model for identifying plant proteins capable of liquid-liquid phase separation.
• To overcome the challenge of limited experimentally validated data in plant LLPS research.
• To provide a comprehensive resource for future studies on plant phase separation.

Main Methods:

• Applied positive-unlabeled learning, a semi-supervised machine learning approach suitable for imbalanced datasets.
• Utilized a dataset of 6,559 known plant phase-separating proteins from eight species.
• Integrated sequence and structural features to train a predictive model.

Main Results:

• Successfully predicted 174,656 high-confidence candidate phase-separating proteins across 14 plant species.
• Experimental validation confirmed LLPS in 67.9% of candidate proteins from Arabidopsis, rice, and maize.
• The developed positive-unlabeled framework demonstrated strong predictive accuracy.

Conclusions:

• The novel computational framework significantly expands the repertoire of known plant phase-separating proteins.
• This study provides a valuable, open-access resource for advancing plant LLPS research.
• The findings highlight the potential of machine learning in addressing biological prediction challenges.