Patient Stratification
Identify molecularly distinct subgroups within a clinical cohort and characterize what makes each one different. Cortex performs dimensionality reduction, unsupervised clustering, and subgroup-specific differential analysis to reveal stratification that may not be visible from clinical variables alone. Each stratum is characterized by its distinguishing molecular features, enriched pathways, and, critically, whether those molecular differences translate to outcome differences.
Deliverables
Subgroup definitions with molecular features
PCA-based dimensionality reduction reveals ancestry-driven sample separation in the TCGA-BRCA cohort. Each subgroup is characterized by its distinguishing genomic and transcriptomic features: African American patients show higher TP53 mutation rates (46% vs 31%), elevated 16p loss (OR 3.87), and enriched IFNα response pathways (NES +1.97). European American patients show higher PIK3CA rates (37% vs 21%) and different pathway profiles.
Survival analysis across subgroups
Kaplan-Meier and Cox regression analysis tests whether molecular subgroups correspond to outcome differences. For the ancestry comparison, the result is a null survival effect after clinical adjustment (HR 1.046, p=0.837), a critically important finding. This demonstrates that molecular stratification must be interpreted carefully: not every molecular difference drives clinical outcomes. The platform reports this with the same rigor as positive findings.
Distinguishing features per stratum
Immune checkpoint markers show large effect sizes between groups: OX40/TNFRSF4 (Cohen's d=0.82), PD-1, CTLA-4, and LAG-3 all elevated in African American patients. The prognostic signature overlap between groups is remarkably low, only 1.2% (2 shared genes out of 167), suggesting that population-specific biomarker panels may be needed rather than one-size-fits-all approaches.
Define inclusion criteria for enrichment trials or identify subpopulations for targeted therapies. Use molecular stratification to design more precise clinical trials, but only when molecular differences translate to meaningful outcome differences.
Breast cancer ancestry stratification: PCA, genomics, survival
| Group | Top Pathway | TP53 | PIK3CA |
|---|---|---|---|
| African American | IFNα response (NES +1.97) | 46% | 21% |
| European American | Protein secretion (NES -1.93) | 31% | 37% |
| Feature | Effect | Direction | Type |
|---|---|---|---|
| 16p loss | OR 3.87 | AA > EA | CNA |
| TP53 mutation | 46% vs 31% | AA > EA | Mutation |
| PIK3CA mutation | 21% vs 37% | EA > AA | Mutation |
| Marker | Cohen's d | Direction |
|---|---|---|
| OX40/TNFRSF4 | +0.82 | AA > EA |
| PD-1 | +0.65 | AA > EA |
| CTLA-4 | +0.58 | AA > EA |
| LAG-3 | +0.54 | AA > EA |
HR 1.046 (p=0.837): null survival difference after clinical adjustment. Molecular differences do not always translate to outcome differences. Prognostic signature overlap: 1.2% (2 shared genes out of 167).
ATOPIC DERMATITIS PATIENT SUBTYPES (GSE157194)
| Cluster | Label | Patients | Key Feature |
|---|---|---|---|
| C1 | Low-inflammation | 19 (34%) | Minimal barrier disruption |
| C2 | Moderate | 23 (41%) | Mixed Th2/Th17 signature |
| C3 | High-inflammation | 14 (25%) | Strong S100A/SPRR upregulation |
C1 (low-inflammation, 34%) represents a potential non-responder subgroup. These patients may not benefit from aggressive immunosuppression.
ATG7 MICROGLIA BIFURCATION (GSE286094-GSE286095)
| Cell Type | Change | Odds Ratio | Direction |
|---|---|---|---|
| Neutrophil | -45% | 0.55 | decreased |
| T cell | +76% | 1.76 | increased |
ATG7 knockout drives HM1 (MAPK/PI3K hyperactive) depletion (logFC -2.00, 806 cells lost) with compensatory HM2 (p53/NFκB senescent) enrichment (+1.21, 541 cells gained). Peripheral immune: neutrophil loss (-45%, OR 0.55) paired with T cell infiltration (+76%, OR 1.76), a shift from innate to adaptive immunity.
Methodology
Dimensionality reduction
PCA on normalized expression data reveals the major axes of variation in the dataset. For the 1,084-patient TCGA-BRCA cohort, the first two principal components captured 12.3% and 8.1% of variance respectively, with ancestry as a visible separation axis.
Unsupervised clustering
Clustering algorithms identify molecularly distinct groups without relying on known labels. The resulting clusters are compared against clinical annotations (ancestry, subtype, stage) to assess what drives the molecular separation.
Subgroup characterization
Each subgroup is profiled for differentially expressed genes, enriched pathways, mutation frequencies, copy number alterations, and immune marker levels. This produces a comprehensive molecular fingerprint for each stratum.
Survival analysis with covariate adjustment
Cox regression models test whether subgroup membership predicts outcomes after adjusting for clinical covariates (stage, grade, treatment). This separates molecular associations from clinical confounders.
Feature importance ranking
Features that best distinguish each stratum are ranked by effect size, statistical significance, and biological interpretability. This prioritized list guides downstream validation experiments and biomarker panel design.
Target personas
Clinical trial designer
Use molecular stratification to define enrichment criteria and design more precise inclusion/exclusion criteria.
Precision medicine lead
Identify subpopulations with distinct molecular profiles that may respond differently to targeted therapies.
Biostatistician
Assess whether molecular subgroups translate to outcome differences with rigorous covariate-adjusted survival analysis.
Explore more
Ready to see your own data analyzed?
Tell us what you're working on. We can show you what the output would look like.