netDx

netDx constructs interpretable patient classifiers by integrating heterogeneous multi-omic and clinical data into patient similarity networks for precision medicine applications.

Key Features:

• Multi-Omic Data Integration: Integrates diverse patient datasets, including clinical records and genomic profiles, to build multi-modal classifiers.
• Patient Similarity Networks: Converts patient data into networks that encode pairwise patient similarity for classification based on profile similarity.
• Interpretability and Mechanistic Insights: Groups genes into pathways to provide pathway-level, mechanistic interpretation of predictive features.
• Performance in Cancer Survival Prediction: Shows superior performance compared to many machine learning methods in binary cancer survival prediction and handles missing data without requiring imputation.
• Bioconductor Package and Workflows: R/Bioconductor implementation provides workflows for predictor generation, feature selection across multiple train/test splits, customizable similarity metrics, and parallel execution.
• Model Performance Metrics: Computes AUROC (Area Under the Receiver Operating Characteristic curve), AUPR (Area Under the Precision-Recall curve), and accuracy for model evaluation.
• Visualization Tools: Uses RCy3 to export top-scoring pathways and integrated patient networks for visualization in Cytoscape.
• Pathway-Based Classification Workflow: Provides a workflow for pathway-based patient classification from sparse genetic data.

Scientific Applications:

• Oncology — cancer survival prediction: Applied to cancer datasets where it outperforms many traditional machine learning methods for predicting survival outcomes.
• Precision medicine — patient stratification and personalized treatment: Integrates multi-omic and clinical data to support patient stratification and to inform personalized treatment strategies.

Methodology:

Integrates multi-omic and clinical data into patient similarity networks; groups genes into pathways for feature definition; performs feature selection across multiple train/test splits using customizable similarity metrics; computes AUROC, AUPR, and accuracy for evaluation; handles missing data without imputation; exports visualizations via RCy3 for Cytoscape.