Five of the most cited names in research integrity — John Ioannidis (2005), Amgen's Begley & Ellis (2012), Bayer's Prinz, Schlange & Asadullah (2011), Freedman, Cockburn & Simcoe (2015), and Schneider & Kilicoglu (HHS/ORI-funded, ongoing) — have independently documented the same structural failure in biomedical research.
Pharma companies cannot reproduce 65–89% of the published findings they buy programs on, the U.S. spends roughly $28 billion a year on preclinical work that cannot be repeated, and 25.4% of medical articles contain citation errors where the cited paper does not support what the citing paper claims.
The problem is documented at industry scale. The process that produces $100M-to-$10B biotech deals has not been fixed to address it.
| Finding | Source | |
|---|---|---|
| 89% | Amgen could not reproduce 47 of 53 landmark cancer biology papers, with cooperation from the original authors. | Begley & Ellis, Nature, 2012 |
| 65% | Bayer could not reproduce 65% of 67 published drug-target findings they used to source internal programs. | Prinz et al., Nat Rev Drug Discov, 2011 |
| 25.4% | Medical articles contain citation errors — the cited paper does not support what the citing paper claims. | Schneider & Kilicoglu, HHS/ORI (ongoing) |
| $28B/yr | U.S. spend on preclinical research that cannot be reproduced by anyone else. | Freedman et al., PLOS Biology, 2015 |
| 0 | Firms offering systematized, named citation-chain forensics as a service for biotech M&A diligence. | White space confirmed (04-2026) |
AbbVie paid $5.8 billion for Stemcentrx to get Rova-T, an antibody-drug conjugate targeting DLL3 on cancer stem cells. The deal rested on a foundational assumption: that the cancer stem cell model, functionally validated in acute myeloid leukemia and breast cancer, applied to small cell lung cancer.
It had never been functionally validated in SCLC. The standard assay (in vivo limiting dilution) was never published for that tumor type pre-acquisition. Two peer-reviewed papers (Hidalgo 2014, Ehm 2015) directly challenged the assumption and were available in the literature before close.
Phase III halted December 2018 — patients on Rova-T died sooner than on standard chemotherapy. Program terminated 2019. AbbVie took a $4 billion impairment charge.
A forensic audit tracing the citation chain backward from Bonnet & Dick (1997) to Saunders (2015) would have surfaced the missing validation in days. AbbVie could have demanded the assay before closing, restructured payments as milestones contingent on functional validation, or walked away. Without the flag, none of those conversations happen.
The same forensic method applies across the biotech M&A record. Four more citation-chain audits follow the same format: trace the evidence backward, flag where validation is missing, show what was available before close.
The twist case. The hypothesis was right. The Phase II data was real. Manufacturing scaled up for Phase III, phosphorothioate stereochemistry drifted, and Smad7 knockdown collapsed from 60–85% to 0–15%. A $10,000-per-batch head-to-head knockdown comparison — the same assay used in Phase II — would have caught the drift before the trial enrolled a single patient. Root cause not surfaced for six years.
Emraclidine is an M4-selective muscarinic agonist for schizophrenia. The confirmatory evidence cited at deal close (Karuna's KarXT Phase II data) was generated with xanomeline — an M1/M4 co-activator, not M4-selective. M1 receptor knock-out mouse data showing independent M1 contribution to antipsychotic-like effects was published before close. Ten days of forensics would have surfaced the gap. Phase II trial failed weeks after close.
Pfizer torcetrapib, Roche dalcetrapib, Merck anacetrapib. The same HDL-causation hypothesis run three times. The first failure (torcetrapib, 2006) was where the hypothesis broke. Voight et al. Mendelian randomization data — showing HDL levels are not causally linked to cardiovascular disease — was available four years before dal-OUTCOMES and well before REVEAL. Attempts two and three were avoidable.
Twenty years removing plaques the hypothesis said were causing the disease. DIAN familial-Alzheimer's studies showing tau pathology (not amyloid) better predicts cognitive decline, Jack et al. biomarker trajectory models, and longitudinal cohorts of amyloid-positive cognitively normal patients were all published before the major late-stage failures. The hypothesis predicted those cognitively normal patients should not exist. They did.
An independent forensic audit of a biotech target's scientific evidence chain — tracing every citation link behind the mechanism-of-action claim backward to primary sources, checking whether cited studies actually support the claims built on them, and flagging where standard validation is missing. The audit does not run wet-lab experiments. It produces an evidence-chain map the acquirer can use to restructure the deal before close, not after.
→ Read the full product spec (methodology, process, ten case studies, pricing anchor)