MarkerScout: A Disease-Agnostic Machine Learning Framework for Biomarker Prediction from Multi-Scale Mechanistic Models
Identifying robust biomarkers from high-dimensional biomedical data is a central challenge in translational research, but candidate rankings produced by any single feature-selection or classification method depend on algorithmic choices and rarely reproduce across pipelines. We present a disease-agnostic machine-learning framework that addresses this dependence by systematically benchmarking 25 (feature-selection x classifier) pipelines under five-fold stratified cross-validation, aggregating per-feature evidence by two independent methods (a weighted-selection consensus score and Robust Rank Aggregation), and characterizing the direction of each candidate using Cohens d. We demonstrate the framework on immune-response measurements from two clinical phases: SARS-CoV-2 hospitalization and intensive-care admission; obtaining cross-validated mean F1 above 0.99 with balanced classification errors and producing tiered, direction-aware biomarker lists per phase. Interleukin-18 (IL-18) reached the strongest tier in both phases with consistent direction. The framework generalizes to any binary clinical classification problem and supports principled, reproducible biomarker prioritization.