PULSE

PULSE applies positive-unlabeled (PU) learning to predict which alternative splicing events generate stable protein isoforms and to link those predictions to protein-level evidence from mass spectrometry (MS).

Key Features:

• Semi-supervised learning: Employs positive unlabeled learning to leverage labeled (positive) and unlabeled data for splicing-event prediction.
• Feature set: Incorporates 48 diverse features spanning multiple categories to capture the complexity of alternative splicing.
• Validation and accuracy: Validated on bona fide protein isoforms and directly on mass spectrometry (MS) spectra, achieving an overall AU-ROC of 0.85.
• Robustness to limited negatives: Designed to operate effectively under low experimental coverage and in the absence of comprehensive negative data.
• Prediction target: Predicts the fraction of exon skipping events that produce stable proteins.

Scientific Applications:

• Protein isoform prediction: Estimates that approximately 32% of exon-skipping alternative splicing events result in stable protein isoforms.
• Functional insights: Characterizes the distribution of positive isoforms across functional classes to inform effects of alternative splicing on protein function.
• Structural analysis: Provides information on structural effects of alternative splicing relevant to protein stability and folding.

Methodology:

Applies positive unlabeled (semi-supervised) learning using 48 features and validates predictions with bona fide protein isoforms and mass spectrometry (MS) spectra to accommodate low coverage and lack of negative data.