MarkerScout: A Disease-Agnostic Machine Learning Framework for Biomarker Prediction from Multi-Scale Mechanistic...

Area of Science:

• Biomedical data analysis
• Machine learning in healthcare
• Translational research

Background:

• Identifying robust biomarkers from high-dimensional data is challenging due to pipeline-dependent rankings.
• Existing methods lack reproducibility across different feature-selection and classification algorithms.

Purpose of the Study:

• To develop a disease-agnostic machine-learning framework for reproducible biomarker discovery.
• To systematically benchmark multiple computational pipelines for robust candidate prioritization.

Main Methods:

• Benchmarking 25 feature-selection and classifier pipelines using five-fold cross-validation.
• Aggregating feature evidence via consensus scoring and Robust Rank Aggregation.
• Characterizing biomarker directionality using Cohen's d.

Main Results:

• Achieved cross-validated mean F1 scores above 0.99 for SARS-CoV-2 hospitalization and intensive care admission.
• Generated tiered, direction-aware biomarker lists with balanced classification errors.
• Identified Interleukin-18 (IL-18) as a top-tier biomarker in both clinical phases.

Conclusions:

• The proposed framework enables principled and reproducible biomarker prioritization for binary clinical classification.
• The methodology is generalizable to various biomedical datasets and clinical problems.
• Consistent identification of IL-18 underscores its potential as a critical immune response biomarker.