LRensemble

LRensemble predicts protein subcellular localization by aggregating multiple existing localization predictors to classify proteins into cytosol, mitochondrion, nucleus, and secretory pathways for genome-wide analyses in yeast and human proteomes.

Key Features:

• Ensemble Aggregation: Aggregates outputs from up to six individual localization predictors and integrates them via logistic regression.
• Minimalist Ensemble Design: Employs a minimalist design that reduces computational complexity by selecting a reduced subset of predictors while retaining predictive accuracy.
• Feature Selection and Contribution Scores: Uses a feature selection-based filter and contribution scores to identify and retain the most complementary predictors, minimizing redundancy and consensus errors.
• Logistic Regression Classifier: Implements logistic regression as the ensemble algorithm and final classifier for prediction integration.
• Performance Improvement: Demonstrates improved AUC on genome-wide datasets, increasing yeast AUC from 0.558 to 0.707 and human AUC from 0.628 to 0.646 relative to individual predictors and traditional ensembles.
• Computational Efficiency: Operates using approximately one-third to one-half of the predictors typically used in ensembles, reducing running time and computational demands for large-scale analyses.
• Meta-Predictor Advantage: Combines diverse feature sets from individual predictors to ensure comprehensive feature coverage and enhanced predictive performance.

Scientific Applications:

• Protein Function Inference: Supports elucidation of protein function through accurate subcellular localization predictions.
• Genome-wide Localization Mapping: Enables mapping of protein localization across yeast and human proteomes at genome scale.
• Genomics and Proteomics Research: Provides localization-aware predictions to support genomics and proteomics studies and analyses.

Methodology:

Aggregates outputs from up to six individual localization predictors, applies a feature selection-based filter and contribution-score ranking to select complementary predictors, and integrates selected predictors using logistic regression as the ensemble algorithm and classifier; experimental validation used AUC comparisons on yeast and human genome-wide datasets.