Biotech Advanced Biomarker Discovery Simulation

Real biomarker discovery rarely fails because a candidate looks flat — it fails because a candidate that looks separated in a raw scatterplot turns out to carry no predictive signal once tested rigorously. This simulator generates a synthetic disease-vs-healthy patient cohort across three candidate biomarkers (one strong, one weak, one statistical decoy), trains a bagged ensemble of decision-tree stumps to classify patients from their expression values, and then runs permutation feature importance — repeatedly scrambling one biomarker at a time and measuring the resulting drop in held-out accuracy — to work out which candidates the model actually relies on. Every patient renders as a point in 3D biomarker-expression space, colour-coded by disease status and by whether the ensemble classified it correctly; the biomarker axes light up as their importance clears your selection threshold, turning the discovery pipeline itself into something you can watch happen.