On November 16, 2010, a retraction notice appeared in the Journal of Clinical Oncology that should have forced a reckoning across the entire field of omics-based drug development. Anil Potti’s genomic predictor studies at Duke University had directed cancer patients into clinical trials based on biomarker signatures that were, at root, artifacts of statistical mismanagement. The trials were halted. The papers were retracted. And the question that nobody wanted to answer publicly was this: how many other omics programs, running right now, are built on the same fragile foundation?

A new taxonomy published in Nature Biotechnology offers the most coherent answer the field has produced. The paper proposes a unified, practical framework for false discovery rate (FDR) control in omics research. The timing matters. So does the architecture.

Why FDR Chaos Is a Regulatory Crisis in Disguise

The false discovery rate was formally introduced in 1995 by Benjamini and Hochberg as a principled solution to the multiple comparisons problem: when you test thousands of hypotheses simultaneously, as every genomics experiment does, the probability of calling something significant purely by chance scales with the number of tests. FDR control sets a bound on the expected proportion of false positives among all declared discoveries. It was always meant as an improvement over family-wise error rate (FWER) control, which is so conservative that it drives statistical power to zero in high-dimensional settings. The B-H procedure gained near-universal adoption in omics research because it offered power. What investigators failed to standardize was the specific flavor of FDR being controlled, the dependency structure of the data, and the inferential question actually being asked.

That ambiguity now lives inside regulatory submissions.

When a sponsor files an NDA with a companion biomarker claim, or attaches genomic stratification endpoints to a Phase II/III adaptive design, the statistical reviewers at CDER or CDRH face a document that may cite FDR control without specifying which FDR, controlled at which level, under which assumptions about feature correlation. A 2025 paper published in PubMed under the title “Beware of counter-intuitive levels of false discoveries in datasets with strong intra-correlations” put the operational problem bluntly: in omics datasets where features are strongly correlated, as they are in gene expression panels and proteomics arrays, the Benjamini-Hochberg procedure systematically underestimates the true false discovery proportion. The method that everyone trusts most reliably misleads when the data structure it was designed for doesn’t match the data structure it encounters. For a sponsor who pre-specified B-H correction in their statistical analysis plan without accounting for intra-correlation, that is not just a methodological footnote. It is a potential Complete Response Letter.

The Tour Through the Taxonomy

To understand what the Nature Biotechnology framework actually does, start with the problem it names: The paper identifies multiple conceptually distinct FDR quantities that researchers routinely conflate, mapping different inferential goals onto different FDR procedures with guidance on when each applies. Each of these answers a different scientific question. Using the wrong one doesn’t just inflate false positives; it can deflate them, causing investigators to miss real biological signals that a companion diagnostic might have captured.

The schizophrenia literature illustrates the downstream cost. A multiomics study examining expression quantitative trait loci found that 55.2% of eSNP discoveries from 2014-era studies nominally replicated at P < 0.05 in 2018-era studies when FDR was controlled at alpha 0.05. That sounds promising until you invert the number: 44.8% of discoveries did not replicate. Nearly half. In a psychiatric drug program where biomarker-stratified enrichment determines which patients enter a Phase III, a 44.8% false discovery proportion in the biomarker selection phase corrupts every downstream endpoint. The trial becomes an experiment designed around noise.

The taxonomy’s organizing insight is that FDR control choices should be matched to the inferential goal, mapping distinct scientific goals onto distinct FDR procedures with guidance on when each applies. The taxonomy provides explicit guidance matching each inferential goal to the appropriate FDR procedure, with particular relevance for sponsors developing predictive classifiers for regulatory submission.

That third category is where the clinical trial community should focus hardest. Consider a sponsor developing a 200-gene expression panel as a companion diagnostic for a targeted oncology therapy, filing a PMA with a planned adaptive enrichment design. The statistical analysis plan specifies FDR control at 0.05 using B-H. The panel was developed in a tissue type with high inter-gene correlation. Under those conditions, the 2025 intra-correlation paper cited above demonstrates that the actual false discovery proportion can substantially exceed 0.05, meaning the classifier admits more noise than the sponsor disclosed. When the Phase III enriches on that classifier and the primary endpoint misses, the sponsor and the FDA both lack the statistical machinery to determine whether the drug failed or the biomarker failed. That ambiguity doesn’t resolve itself in a Type A meeting.

The taxonomy also addresses a dependency structure problem that the hepatocellular carcinoma literature crystallized. In a prognostic marker study for HCC, investigators applied both a P-value threshold below 0.05 and an FDR threshold below 0.05 to screen differentially expressed genes. Layering two thresholds sounds conservative. But if the gene expression matrix contains strong covariance structure, the two thresholds are not independent guards: they can both be satisfied simultaneously by correlated false positives. The Nature Biotechnology taxonomy explicitly categorizes this scenario and recommends conditional FDR procedures that incorporate covariate information about feature dependencies to restore the nominal error guarantee. For a regulatory submission, that distinction between nominal and actual error control is exactly what a statistical reviewer should be testing, and rarely does because the vocabulary to demand it hasn’t existed in unified form until now.

What the FDA Hasn’t Said Yet

The FDA’s existing framework for omics-based submissions pulls from several guidance documents, including the 2013 guidance on Qualification of Biomarker, and the agency’s ongoing framework for pharmacogenomics data submissions. That silence has been productive for sponsors in the short term: vague standards are easy to satisfy on paper. But it has been catastrophic for the literature, where the Potti affair remains the most visible example of what happens when FDR methodology is treated as a compliance checkbox rather than a scientific commitment.

The Nature Biotechnology taxonomy gives FDA reviewers, for the first time, a shared vocabulary to interrogate submissions with precision. A reviewer who can point to the taxonomy and ask “which FDR are you controlling, and does your choice match your inferential goal?” transforms the review from a documentation exercise into a scientific audit. That shift creates near-term pressure for sponsors whose statistical analysis plans are built on the ambient assumption that FDR control means B-H at 0.05.

Industry’s standard response to new methodological frameworks is to wait for agency guidance before revising SAPs. That logic has a flaw here. The taxonomy doesn’t require FDA action to become operationally relevant: it becomes relevant the moment a statistical reviewer cites it in a deficiency letter. And given that the paper appears in Nature Biotechnology, a journal FDA statisticians read, the window between publication and citation is probably shorter than the typical Phase III enrollment timeline.

The Potti trials were halted in 2010. Years later, the field has yet to converge on a standardized answer to the question his case exposed: which FDR, controlled how, under what data assumptions, for what inferential purpose? The Nature Biotechnology taxonomy answers all four parts. The question now is whether sponsors update their statistical analysis plans before the FDA starts asking why they haven’t.

References

  1. Nature Biotechnology, “A unified and practical taxonomy for false discovery rate control in omics research”
  2. Columbia University Mailman School of Public Health, “False Discovery Rate”
  3. PMC / NCBI, “A selective inference approach for false discovery rate control using multiomics covariates yields insights into schizophrenia”
  4. PMC / NCBI, Prognostic molecular marker study for hepatocellular carcinoma using FDR-adjusted gene screening
  5. PubMed, “Beware of counter-intuitive levels of false discoveries in datasets with strong intra-correlations”
  6. Cancer History Project, Anil Potti retraction, Journal of Clinical Oncology, November 16, 2010
Website | + posts

Moe Alsumidaie, MBA, MSF, is founder and Chief Editor of Vanguard Publications, which publishes Clinical Trial Vanguard, Pharma Vanguard and BullScope, and Head of Research at CliniBiz. He has two decades in clinical trial operations and data science, with earlier roles at Genentech, Abbott Vascular and Stanford University Medical Center, and is a guest lecturer in clinical trial sciences at Rutgers University.