Predictive biomarker discovery

Train a cross-validated model that predicts a clinical outcome — responder vs non-responder, a subtype, a survival group, or a continuous score — from tumor expression features, with honest out-of-sample performance and a top-biomarker ranking.

How this differs from the other use cases

Mutation to outcomes and the cohort-comparison layer answer which features associate with an outcome (descriptive statistics). This use case answers how well features predict an outcome out-of-sample — the standard biomarker-discovery question — with cross-validation and a permutation-test significance check, not just association p-values.

Research question

Can expression features predict a clinical target (treatment response, subtype, survival group, or a continuous score) out-of-sample, and which genes carry the signal?

Who this is for

Translational oncology and precision-medicine teams building predictive signatures
Pharma biomarker groups screening for response/outcome predictors
Computational oncology labs benchmarking expression-based classifiers

Data requirements

Data	Required	Purpose
Expression matrix (genes × samples)	Yes	Per-sample feature vectors
Clinical table with a target column	Yes	Prediction target (label or continuous value)

Sample identifiers in the expression matrix and the clinical table must match. Classification needs ≥ 2 classes with ≥ 2 samples each; both tasks need ≥ 8 labeled samples.

Workflow

Register expression + clinical datasets
  → Predictive biomarker modeling (choose target column + task)
  → Review CV performance, permutation p, and top biomarkers
  → AI interpretation

Step 1 — Train the model

On the Analyze page, open the Predictive biomarker modeling panel, pick the expression and clinical datasets, enter the target column, and choose classification or regression.

The pipeline selects the most-variable genes, fits a cross-validated random forest, and runs a label-permutation null.

Step 2 — Review performance and biomarkers

Classification: ROC AUC, balanced accuracy, sensitivity/specificity.
Regression: R², MAE, Pearson r.
Both: a permutation-test p-value (chance-corrected) and a top-biomarker importance ranking.

Step 3 — Interpretation

Include the predictive-biomarker run in an interpret job for a grounded narrative that cites the model metrics and top biomarkers.

Expected outputs

Out-of-sample performance with a permutation p-value
Top-biomarker (gene) importance ranking
Per-sample cross-validated predictions
Run manifest and parameters for reproducibility

Typical analyses

Analysis	Target	Question
Response prediction	responder vs non-responder	Do expression features predict checkpoint-inhibitor benefit?
Subtype classification	molecular subtype	Can a signature assign subtype out-of-sample?
Risk-group prediction	long vs short survival	Does expression separate prognostic groups?
Continuous score	TMB, a pathway score	Can expression predict a continuous endpoint (regression)?