NeuroPred-FRL

NeuroPred-FRL predicts neuropeptides (NPs) at large scale using a machine-learning meta-predictor that applies feature representation learning to identify NPs implicated in immune systems and in regulation of central anxious hormones.

Key Features:

• Meta-predictor framework: Integrates predicted probability scores from multiple baseline models into a meta-model for NP prediction.
• Feature representation learning: Employs feature representation learning to enhance predictive accuracy.
• Baseline models: Constructs 66 optimal baseline models using a combination of 11 different encodings and six classifiers.
• Two-step feature selection: Applies a two-step feature selection approach to refine baseline models.
• Probability feature vector: Combines predicted probability scores from the 66 baseline models into a 66-dimensional input feature vector.
• Second-round feature selection: Performs a second round of feature selection on the 66-dimensional probability feature vector.
• Random forest meta-model: Trains a random forest classifier on the refined features to construct the final meta-model.
• Benchmarking: Evaluates performance using cross-validation and independent tests.
• Interpretability: Applies the SHapley Additive exPlanation (SHAP) algorithm to interpret model predictions.

Scientific Applications:

• Large-scale NP identification: Enables large-scale identification and cataloging of neuropeptides from sequence-derived features.
• Immunoinformatics and drug development: Supports basic research and drug development investigations of neuropeptides in immune systems.
• Characterization and translational applications: Facilitates characterization of NP roles in biological systems and supports potential applications in clinical therapies.
• Mechanistic insight: Uses SHAP-based interpretability to reveal model mechanisms and provide insights into neuropeptide functional mechanisms.

Methodology:

Generate 66 optimal baseline models using 11 encodings and six classifiers; combine their predicted probability scores into a 66-dimensional feature vector; apply a two-step feature selection including a second-round selection on the probability vector; train a random forest classifier as the final meta-model; evaluate with cross-validation and independent tests; interpret predictions with SHAP.