PPSBoost

PPSBoost predicts phosphorylation sites in protein sequences using a LightGBM-based gradient boosting classifier to identify phospho-sites for studies of post-translational modification and signaling.

Key Features:

• Machine Learning Framework: Uses a LightGBM-based gradient boosting model employing tree-based learning algorithms.
• Feature Engineering: Incorporates evolutionary data, geometric properties, sequence environment context, and amino acid-specific characteristics as manually engineered features.
• Interpretability: Based on decision tree classifiers, enabling inspection of rules derived from trees to interpret predictions.
• Performance Metrics: Evaluated on the Phospho.ELM benchmark (2429 protein sequences from 11 organisms) with F1 = 0.504 and ROC AUC = 0.836.
• Efficiency: Achieves faster processing time compared to recent deep learning–based frameworks.
• Enhanced Predictive Power: Incorporating the output probability from existing deep learning models as an additional feature improves performance to F1 = 0.546 and ROC AUC = 0.849.
• Structural Analysis Validation: Structural analysis on selected protein sequences confirms predictions encompass all phosphorylation sites listed in the Phospho.ELM dataset.

Scientific Applications:

• Phosphorylation site annotation: Provides interpretable predictions of phosphorylation sites to support studies of post-translational modification, protein function, signaling pathways, and disease mechanisms.

Methodology:

LightGBM-based gradient boosting using manually engineered features (evolutionary data, geometric properties, sequence environment context, amino acid-specific characteristics); inclusion of output probabilities from existing deep learning models as an additional feature; problem transformation approach to tune precision–recall balance; evaluated on Phospho.ELM (2429 sequences, 11 organisms) with reported F1 and ROC AUC metrics and validated by structural analysis on selected proteins.