The history of cancer biomarkers is largely a history of failure.
Not because the science was bad — though sometimes it was — but because the questions were poorly shaped. We asked "is this gene associated with survival?" when we should have asked "under what biological conditions would this association be meaningful, reproducible, and actionable?"
The seduction of significance
A p-value below 0.05 is a terrible reason to believe something. Yet in biomarker discovery, it remains the primary gatekeeper. Thousands of genes tested against dozens of endpoints across hundreds of samples produce thousands of nominally significant associations. Most are noise.
The problem is not statistical — it is epistemological. We confuse the ability to detect a pattern with the existence of a meaningful pattern.
Survival endpoints are blunt instruments
Overall survival conflates disease biology with treatment history, comorbidities, age, and socioeconomic factors. Progression-free survival is cleaner but still reflects clinical decisions as much as tumor biology.
When we correlate gene expression with these endpoints, we are not measuring what we think we are measuring.
What would better look like?
Better biomarker research would:
- Pre-specify biological mechanisms, not just statistical tests
- Require replication before publication, not after
- Report effect sizes and confidence intervals, not just p-values
- Distinguish prognostic from predictive biomarkers explicitly
- Acknowledge when a finding is exploratory rather than confirmatory
A biomarker that cannot be falsified is not a biomarker.
It is a story..
The uncomfortable truth
Most published biomarkers will not survive contact with independent datasets. This is not a scandal — it is the expected outcome of a field that rewards discovery over replication, novelty over rigor.
The question is whether we have the institutional courage to change the incentives.