Biomarker Discovery: Random-Forest Feature Importance

Patient Cohort

Signal strength 1.50

Assay noise σ 1.00

Discovery Ensemble

Trees in ensemble 30

Selection threshold 0.050

Results

Test accuracy

Test AUC

Biomarkers selected into panel

0 / 3

BM-A0.000

BM-B0.000

BM-C0.000

Disease, correctly classified Healthy, correctly classified Misclassified by the ensemble

How it works

Each of the 150 simulated patients carries a value for three candidate biomarkers, and each patient is disease-positive or healthy. In this simulation two biomarkers really do shift with disease status and one is a pure decoy — but which is which is not shown to the algorithm:

feature_k(patient) = effect_k · signal · (±1 for disease/healthy) + σ · noise_k
effect = [1.00, 0.55, 0.00]   (BM-A strong, BM-B weak, BM-C decoy)

The discovery ensemble is a bagged forest of single-split decision stumps: each tree draws a bootstrap resample of the training patients, picks whichever biomarker and threshold best separates the two classes by Gini impurity, and votes. The ensemble's prediction is the majority vote across trees; its continuous score (fraction of trees voting "disease") gives the ROC/AUC on held-out test patients.

Permutation importance — the actual discovery step — measures how much test accuracy drops when one biomarker's values are randomly shuffled across patients while the label stays fixed:

importance_k = accuracy(original test set) − accuracy(test set with biomarker k permuted)

A biomarker the ensemble actually relies on causes a real accuracy drop when scrambled; a decoy the model never learned to use causes almost none, even if it happened to differ between the classes in the raw scatter. Raise the selection threshold to keep only biomarkers whose importance clears the bar — that shrinking/growing set is the discovered biomarker panel, and its matching axis lights up in the 3D view.

BM-A axis BM-B axis BM-C axis

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.

Biotech Advanced Biomarker Discovery Simulation