How biotech diligence actually runs — and where it breaks
The workflow has a missing stage.
Last updated 10-13-2026, 05:25 PM PT
Every biotech acquisition passes through the same eight stages between first interest and close. At each stage, specific people check specific things. This page shows the full workflow in order, what each stage checks, what errors slip past, and where a missing stage — backward citation auditing — belongs. In our view, the errors are less about people doing their jobs badly than about a stage that is missing from the workflow.
How to read this page. The eight stages are our own simplified model of a typical biotech acquisition, drawn from general knowledge of the process. Real deals vary, and we have not validated the model against a survey of acquirers. The “what slips past” and “where AI could help” notes are our hypotheses and proposals, not demonstrated results.
11%
of 53 “landmark” cancer papers were confirmed in Amgen’s own attempts
Begley & Ellis, Nature, 2012 (a commentary; the authors chose papers describing something new)
20–25%
of Bayer’s 67 in-house target-validation projects had published data completely in line with Bayer’s own findings
Prinz et al., NRDD, 2011, as summarized in open-access reviews; Begley & Ellis give about 25%
25.4%
total quotation error rate in a pooled analysis of studies of medical journal articles (major errors 11.9%, minor 11.5%; minor includes secondary citations)
Jergas & Baethge, PeerJ, 2015
$28B
per year (estimate) — US preclinical research spend that is not reproducible. A lab-cost figure, not a measure of deal losses.
Freedman et al., PLOS Bio, 2015
The standard biotech M&A workflow, stage by stage
1
Target identification
Who
Business development team at the acquirer.
What they check
Strategic fit. Is this asset in a therapeutic area we care about? Does it plug a pipeline gap?
What slips past
The science is still a pitch at this stage. Citation chains are not typically audited yet.
Where AI could help
Pre-screen every public scientific claim in the target's deck against primary sources before the BD team even writes the brief.
2
Non-binding LOI and term sheet
Who
BD plus the acquirer's finance team and outside bankers.
What they check
Valuation model. Probability of success. Total addressable market. Deal structure and milestones.
What slips past
Valuation inputs are taken from the target's own projections. No independent check that the underlying science actually supports the probability numbers.
Where AI could help
Flag mechanism claims that drive PoS assumptions but have broken citation chains — so finance knows which inputs rest on unstable ground.
3
Scientific diligence
Who
Acquirer's R&D team. Senior scientific advisors. KOL calls with domain experts.
What they check
Plausibility of the mechanism. Trial data. Preclinical models. Expert opinion on whether this can work.
What slips past
They check claims for plausibility, not citation chains for truthfulness. Advisors are domain experts, not citation-forensics experts. Deal timelines are compressed. References are rarely walked backward.
Where AI could help
Walk every load-bearing citation backward three to five levels. Flag broken links, entity conflations, mechanism inversions before the KOL call even happens.
3.5
Citation-chain forensic audit
Who would do it
In the workflows we reviewed, no role owns this.
What it would check
Every load-bearing scientific claim walked backward to primary sources. Each link tagged supports, partial, contradicts, or missing.
What it would catch
Protein-name confusion. Mechanism inversions. Entity conflations. Statistical artifacts from underpowered trials. Unvalidated mechanism transfers between therapeutic areas.
Why it can exist now
Citation walking, entity matching, and primary-source retrieval can now be done systematically by software, with a person reviewing the flagged links.
4
Clinical and regulatory diligence
Who
CRO consultants. Former FDA reviewers. Regulatory affairs experts.
What they check
Trial design. Endpoint validity. Regulatory path. CMC (chemistry, manufacturing, controls).
What slips past
They assume the scientific foundation is sound. They audit trial execution, not claim lineage. If the mechanism is wrong, the trial can be flawlessly executed against the wrong target.
Where AI could help
Cross-reference the mechanism being tested against what the preclinical citation chain actually supports. Flag when the trial endpoint measures a mechanism the science doesn't back.
5
Commercial diligence
Who
Market research firms. Commercial DD consultants. Internal commercial team.
What they check
Market size. Pricing. Access and reimbursement. Competitive landscape. Prescriber adoption curves.
What slips past
They take the indication and mechanism as given. Revenue forecasts sit on scientific claims this stage is not set up to audit.
Where AI could help
Feed the forensic audit output upstream. If the citation chain flags a broken mechanism, the commercial model knows the revenue line is sitting on air.
6
Legal and IP diligence
Who
Internal IP counsel. External patent counsel. Litigation specialists.
What they check
Patent strength. Freedom to operate. IP portfolio completeness. Prior-art challenges.
What slips past
Patent validity and scientific validity are adjacent but not the same thing. Patent counsel does not walk citation chains. A patent can be strong while the underlying scientific claim is broken.
Where AI could help
Cross-reference the scientific claims supporting the patent against the forensic audit. Flag patents whose IP story doesn't match the actual science.
7
Financial and accounting diligence
Who
Big Four accounting firms. Tax counsel. Reps & warranties insurance brokers.
What they check
Financial statements. Tax structure. Working capital. Insurance coverage for known risks.
What slips past
They audit books, not biology. R&W insurance generally covers breaches of the target’s representations and excludes known issues; whether a given impairment is covered depends on the policy terms.
Where AI could help
Feed forensic audit findings into the reps & warranties process, so the acquirer can negotiate on science-specific risks, not just generic indemnities.
8
Definitive agreement and close
Who
M&A counsel on both sides. Investment bankers. Board approval.
What they check
Deal documentation. Representations and warranties. Closing conditions. Escrow and milestones.
What slips past
The reps are signed based on the target's own representations of its science. By the time the agreement is drafted, it's too late for any party to walk the citation chain — the deal team is closing, not auditing.
Where AI could help
The forensic audit becomes an exhibit to the agreement. Specific science-related representations are carved out where the audit flagged weak links.
Why humans at every stage miss the citation chain
It isn't anyone's job. In the workflows we reviewed, no role owns backward citation auditing, and each stage tends to assume the prior stage checked it.
Time pressure. Scientific diligence runs on a compressed deal clock. Walking every citation backward through a full mechanism chain is slow, focused work that the clock rarely allows.
Expertise mismatch. Senior scientific advisors are domain experts, not citation-forensics experts. Plausibility review is a different skill from lineage verification.
Entity-conflation blind spot. Same protein name across multiple organisms, multiple glycoforms, multiple research groups. Spotting this requires systematic entity matching, not expert intuition.
Paradigm-transfer blind spot. When a mechanism validated in therapeutic area A gets moved to therapeutic area B, the underlying citation chain doesn't always transfer. It is rarely checked whether it did.
Forward bias. The CSO builds the story forward (hypothesis, experiments, results). Auditing requires reading backward (claim, citation, source). The two motions aren't interchangeable.
What the missing stage would actually do
Walk every load-bearing citation backward
Take the target's scientific claims. Pull every cited paper. For each citation, verify that the cited paper actually supports what the citing paper claims. Tag each link: supports, partial, contradicts, missing.
How: citation retrieval across PubMed, Semantic Scholar and primary literature, with expert review of the flagged links.
Flag entity conflations
When a claim uses a protein name, gene name, or drug identifier, verify it refers to the same entity across every cited source. Flag when a name is doing different work in different papers (for example, closely related proteins with similar names, or glycoform differences between lab and clinical material).
How: cross-reference against public databases such as UniProt and ChEMBL, where the database covers the question.
Flag mechanism inversions
When a deck says "drug X blocks receptor Y," verify the binding data supports that mode of action. Common failure: the actual mechanism is ligand sequestration, not receptor blocking. Different kinetics, different combination logic, different biomarkers.
How: check each load-bearing mechanism claim against the primary binding data in the cited papers.
Flag cross-indication mechanism transfers
When a mechanism validated in one therapeutic area gets transferred to another (AML to small cell lung; breast cancer to prostate), verify the transfer is supported by primary data in the new context. Unsupported mechanism transfers can be costly errors.
How: indication-specific literature scan against the mechanism's original validation context.
One case where a large write-down followed
AbbVie / Stemcentrx
Announced April 2016, closed June 2016 · Rova-T (DLL3, small cell lung cancer)
The Phase 3 TAHOE trial stopped enrolling in December 2018 after an independent monitoring committee cited shorter overall survival than the comparison arm. The case is set out here, from AbbVie’s filings and the original papers.
~$4B
Net impact estimate, January 2019 8-K
This is the thesis in one paragraph
Eight stages. Billions of dollars. Smart, credentialed people doing their jobs well. And between stage 3 and stage 4, a structural hole where backward citation auditing should be. The people aren't failing. The workflow is incomplete. The missing stage is what a forensic citation-chain audit installs.