MEL-MP

MEL-MP predicts moonlighting proteins (MPs) de novo using a multimodal deep ensemble learning architecture that integrates sequence-based and structural features for accurate identification and analysis.

Key Features:

• Multimodal deep ensemble learning: Implements a multimodal deep ensemble learning architecture combining multiple classifiers to improve prediction accuracy.
• Input features: Extracts primary protein sequences, evolutionary information, physical and chemical properties, and secondary protein structures.
• Classifier selection: Selects specific classifiers tailored to each feature type to match classifier strengths to the nature of the data.
• Ensemble integration: Integrates outputs from individual classifiers using a stacked ensemble method for model fusion.
• Feature fusion comparison: Compares direct combination of features with a multimodal deep auto-encoder for feature fusion.
• Performance evaluation: Reports model selection and cross-validation results with an F-score of 0.891 versus 0.784 for MPFit.
• Downstream analyses: Enables analysis of predicted MP distribution across chromosomes, evolutionary trajectories, disease associations, and functional enrichment patterns.

Scientific Applications:

• De novo MP prediction in the human proteome: Identifies potential moonlighting proteins within the human proteome.
• Functional and evolutionary studies: Supports analysis of evolutionary trajectories and functional enrichment of predicted MPs.
• Disease association analysis: Facilitates investigation of associations between predicted MPs and diseases.
• Chromosomal distribution studies: Enables examination of the chromosomal distribution of predicted MPs.

Methodology:

Extracts sequence-based features (primary sequences, evolutionary information, physical and chemical properties, and secondary structures), applies specific classifiers to each feature type, integrates classifier outputs via a stacked ensemble, and compares direct feature combination with a multimodal deep auto-encoder using model selection and cross-validation; performance was compared to MPFit (F-score 0.891 vs 0.784).