PULPS

PULPS predicts scaffold proteins that drive liquid-liquid phase separation (LLPS) using positive unlabeled learning to prioritize candidate scaffolds for experimental validation.

Key Features:

• Positive Unlabeled Learning Framework: PULPS employs a positive unlabeled (PU) learning approach to handle extreme class imbalance between limited experimentally identified scaffolds and numerous unlabeled candidates.
• Hybrid multimodal feature modeling: The method leverages the joint distribution of hybrid multimodal features to enhance prediction accuracy.
• ProbTagging and Penalty Logistic Regression (PLR): PULPS integrates ProbTagging with penalty logistic regression to optimize profiling of scaffold propensities.
• Performance metrics: Validation studies reported an area under the receiver operating characteristic curve (AUC) of 0.8353 and an area under the lift curve (AUL) of 0.8339.
• Comparative superiority: On recent experimentally verified scaffolds AUL increased from 0.8339 to 0.8577, representing a 45.7% improvement over PLR alone and an 8.2% improvement versus other existing tools.

Scientific Applications:

• Scaffold discovery in the human proteome: PULPS prioritizes candidate scaffold proteins that potentially co-drive LLPS across the human proteome.
• Candidate generation for experimental validation: The predictions produce ranked scaffold candidates for downstream experimental testing.
• Study of phase separation and disease: PULPS supports investigations into phase separation states and their implications for cellular organization and disease pathology.

Methodology:

PULPS applies positive unlabeled learning using the joint distribution of hybrid multimodal features, integrates ProbTagging with penalty logistic regression (PLR), and evaluates performance using AUC and AUL metrics.