DPAS: disease-associated peptide anomaly score for identifying pathogenic peptides via one-class learning.

Area of Science:

• Bioinformatics and computational biology
• Peptide analysis and disease association
• Machine learning in biological data analysis

Background:

• Predicting disease-associated peptides is crucial but challenging due to unreliable negative datasets, leading to biased models.
• Traditional binary classification methods struggle with imbalanced or incomplete negative data in peptide analysis.
• Developing robust methods for identifying disease-associated peptides is essential for biomarker discovery.

Purpose of the Study:

• To propose and evaluate a one-class classification approach for predicting disease-associated peptides using only positive-labeled data.
• To compare the performance of One-Class Support Vector Machines (OCSVM), Isolation Forest, and Autoencoders for this task.
• To introduce a novel scoring metric, Disease Peptide Anomaly Score (DPAS), for enhanced interpretability and peptide prioritization.

Main Methods:

• Utilized a one-class classification framework focusing exclusively on positive-labeled disease-associated peptide data.
• Implemented and compared three machine learning classifiers: OCSVM, Isolation Forest, and Autoencoders.
• Integrated sequence-based features and employed mean reconstruction errors for Autoencoder evaluation.
• Developed the Disease Peptide Anomaly Score (DPAS) by combining model anomaly scores with SHAP feature importance.

Main Results:

• Autoencoders demonstrated superior performance in classifying disease-associated peptides compared to OCSVM and Isolation Forest.
• The Autoencoder model effectively differentiated inliers (disease-associated peptides) from outliers using mean reconstruction errors.
• The proposed DPAS metric successfully combined anomaly detection with feature interpretability for ranking peptides.

Conclusions:

• The one-class classification approach offers an efficient and robust alternative to traditional binary methods for predicting disease-associated peptides.
• Autoencoders show significant promise for identifying disease-associated peptides, particularly when negative datasets are scarce or unreliable.
• The DPAS metric provides an interpretable and effective tool for prioritizing potential peptide biomarkers in disease research.