Mechanistic bridges import structural model error
Statement
Mechanistic and pharmacometric models can link treatments across a gap through exposure, biomarker response, disease progression, or pathway inhibition, and may be more credible than statistical exchangeability when the mechanism is well established, provided the treatment-to-mechanism mapping, the structural parameters, and the transport assumptions are made explicit. Such a model generates a cross-gap contrast but does not restore randomized identification, and it imports structural model error from the mechanistic model itself. The structural model is not wholly untestable, since observable parameters and predictions can be calibrated or falsified; what cannot be validated internally is the unobserved target contrast. The literature auditor found no published analysis using a fully mechanistic PK/PD, biomarker, or pathway model as the identifying bridge across a disconnected evidence network, and set confidence low because absence cannot be established by not finding one.
Why it is open
A mechanistic bridge replaces an untestable statistical assumption with a structural one whose parameters, functional forms, and biological simplifications were fitted for a different purpose, and the evidence synthesis cannot diagnose failure on the cross-gap contrast because that contrast is unobserved by construction. Calibrating the observable parts of the model constrains it but does not bound the extrapolation that carries the answer. The claim of no established practice here is an absence claim neither auditor could verify positively, so it stands as an unfilled niche rather than a proven void.
What has been tried
| Work | What it contributes |
|---|---|
| Pedder et al. 2021 | the nearest existing case of a semi-mechanistic bridge benchmarked against purely empirical NMA, comparing disconnected-network MBNMA estimates with NMA on augmented networks |
Probable solution or research direction
Where a connected analogue can be constructed, calibrate a mechanistic bridge prospectively and compare it directly against purely empirical alternatives on the same network, following the augmented-network benchmarking Pedder et al. applied to the semi-mechanistic dose-response bridge. Require the treatment-to-mechanism mapping and the transport assumptions to be stated as identifying assumptions, and report which parts of the structural model were calibrated against observable data and which were assumed. Treat fully mechanistic bridges as a high-risk, high-reward research direction rather than as a routine option.
Verification
Verdict. Confirmed open. Auditors agreed the problem is open and no counterevidence surfaced.
The source said. “Mechanistic and pharmacometric models can link treatments across a gap through exposure, biomarker response, disease progression, or pathway inhibition, and may be more credible than statistical exchangeability when the mechanism is well established. They also risk importing structural model error from the mechanistic model itself.”
What was wrong with it. The source said mechanistic bridges are not calibrated prospectively or compared with purely empirical alternatives; that narrows to fully mechanistic bridges, because the semi-mechanistic dose-response bridge was benchmarked against NMA on augmented networks by Pedder et al. 2021 (doi:10.1177/0272989X20983315), and model-based meta-analysis routinely compares treatments never studied head-to-head. GPT-5.6 Sol also corrected the claim that nothing diagnoses failure of the mechanistic model: observable parts of it can be calibrated or falsified, and only the cross-gap contrast is beyond reach.
- Literature and prior-art check
- open / supported-with-caveat. prior_work is empty and I could not retrieve a paper that uses a fully mechanistic PK/PD, biomarker or pathway model as the identifying bridge across a disconnected evidence network, so the problem is genuinely open and the risk of importing structural model error is uncontroversial. The caveat is claim 4, that mechanistic bridges are not currently calibrated prospectively or compared with purely empirical alternatives. The closest relative, the semi-mechanistic dose-response bridge, was benchmarked against purely empirical NMA on the same data by Pedder et al. 2021, agreeing in 15 of 18 augmented-network comparisons. Model-based meta-analysis in pharmacometrics is a mature field in which exposure-response models are routinely used to compare treatments never compared head-to-head. I did not read that literature closely enough to assert it contains a formal head-to-head calibration study, so I set confidence low and flag claim 4 as needing narrowing to fully mechanistic bridges rather than pharmacometric bridges in general.
- GPT-5.6 Sol, technical and source-code lens
- open / supported-with-caveat. C1-C3 are technically defensible only when the treatment-to-mechanism mapping, structural parameters, and transport assumptions are explicit. Such a model can generate a cross-gap contrast, but it does not restore randomized identification. The structural model is not wholly untestable because observable parameters and predictions may be calibrated or falsified, although an unobserved target contrast cannot be validated internally. C4 is an uncited, field-wide practice claim that this technical audit cannot establish.
- Decision path
-
R5:default-open → R11:caveated(n=2)
Source
Derived from C:L1138-1140, C:L2498 of the reviewed corpus.