On global development

The Decision Is Not Yours — and It Is Not One Decision

Published 2026-09-02 · Regulatory references are identified below as final, draft, or adopted. Guidance is non-binding and does not establish legally enforceable responsibilities. PMDA and NMPA primary documents are published in Japanese and Chinese; see the note on sourcing at the end. Not legal or regulatory advice.

Where the last essay stopped

In a recent piece I argued that Bender and colleagues' distinction between model validation and process validation runs out at the regulatory interface. Process validation asks whether a model improves a decision. That test holds while the decision is yours. Some decisions aren't — the agency decides whether the evidence supports the claim, and your internal conviction is an input to that decision rather than a substitute for it.

The practical consequence was that one model output can occupy four very different regulatory positions: an operational tool, an exploratory biomarker, a measure used to make decisions inside the programme, and a measure intended to support a regulatory decision. The burdens differ by more than an order of magnitude, and an output can slide from one to the next without anyone deciding it should.

That argument had an assumption buried in it that I did not examine, and a reader in Basel or Tokyo would have spotted it immediately. It assumed there was one adjudicator.

The convergence is real, and you are right to have noticed it

Anyone planning a global programme right now has good reason to think this problem is being solved for them.

ICH M15, the guideline on model-informed drug development, reached Step 4 in January 2026. It explicitly covers artificial intelligence and machine learning alongside population PK, PBPK and quantitative systems pharmacology, and it does the thing harmonisation is supposed to do: it supplies one set of expectations for planning, evaluating and documenting model-derived evidence across ICH regions.

In the same month, FDA and EMA jointly published ten guiding principles of good AI practice in drug development — the first genuine transatlantic alignment on this, co-developed rather than merely compatible. Two of the ten are clear context of use and risk-based performance assessment, which is process validation restated in regulatory grammar.

And the two agencies people expect to be outliers are quieter than the reputation suggests. PMDA's action plan of October 2025 is about PMDA's own use of AI — document retrieval, summarisation, translation, and later in-house systems for review. It imposes no obligations on sponsors at all. NMPA's September 2025 announcement optimising the review clock for innovative-drug trials contains no AI-specific language; it is a procedural acceleration, not an evidentiary standard.

So: converging, genuinely, and faster than most commentary has caught up with. If you concluded from that you can build one credibility package and file it in four places, you have read the documents correctly.

You have still drawn the wrong conclusion, and the reason is in the frameworks themselves.

The harmonisation is at the wrong level

Every one of those documents makes credibility depend on context of use — the specific question the model is being asked, in a specific setting, with a stated consequence if it is wrong. That is the right dependency. It is also the thing that has not been harmonised, and cannot be, because context of use is a property of your development programme rather than of the framework.

The shape of a programme is set by four things: which population you enrolled, which comparator you chose, how you justified your dose, and which endpoint you are asking to carry the claim. Every one of those is adjudicated regionally. So harmonised criteria, applied to divergent contexts, produce divergent evidence requirements. The convergence is real and it is aimed one level above the place where your work actually lives.

Three axes make this concrete.

Dose. FDA's August 2024 final guidance on dosage optimisation — the regulatory expression of Project Optimus — asks for a dosage justified by comparative data. It recommends randomised designs comparing at least two dosages, it requires that optimisation happen before or alongside the establishment of safety and efficacy rather than after, and it states that expedited programmes do not exempt you: not breakthrough designation, not accelerated approval.

EMA has no Project Optimus. What it has is Section 6 of the anticancer guideline (EMA/CHMP/205/95 Rev. 6), and it would be a mistake — one I nearly made — to read the absence of a named initiative as an absence of expectation. For molecularly targeted agents, Rev. 6 asks that dose-finding extend past the maximum tolerated dose toward the optimal biologically active dose, defined as the dose at the start of the dose-response plateau where more adds no benefit. It asks for late-onset and persistent low-grade toxicities, for time-dependent toxicity assessment, for patient-reported outcomes, and for the same totality of evidence. EMA's oncology working party reviewed FDA's guidance formally in December 2024 and came away aligned on principle.

Aligned on principle, and different in procedure. FDA names the randomised two-dose comparison and closes the expedited-pathway door explicitly. Rev. 6 is case-by-case. A model informing dose selection therefore sits in position four at FDA — squarely inside something the agency will examine — while the same model, on the same data, faces a question in Europe that is asked with more discretion and less prescription. The evidence you generate to answer one is not automatically the evidence that answers the other, and the difference is not a difference in rigour. It is a difference in what gets asked.

Population. ICH E17 changed how multi-regional trials are read, and Japan implemented it in June 2018. The current expectation is not a separate Japanese bridging study; it is a consistency evaluation of the Japanese subgroup within the global dataset — same direction of effect, broadly similar magnitude, explicitly not statistical equivalence. In December 2023 MHLW went further and repealed the standing requirement for a Japanese phase 1 before joining a global trial. Japan has moved decisively toward global-first development.

But notice what the consistency evaluation actually is. It is a second question asked of the same data — not does this work but does this work here, and if the subgroup diverges, can you explain why. A stratification model that performs beautifully in aggregate can still leave that question unanswered, and no amount of harmonised credibility framework will answer it for you, because the framework is what tells you the question matters. It doesn't tell you the answer.

China asks its own version. The 2018 reform and the technical guidelines that followed sorted foreign clinical data into three tiers — fully accepted, partially accepted, not accepted — and the gate between the first two is ethnic sensitivity: whether genetic variation in metabolising enzymes, differences in disease biology, or a different standard of care create uncertainty in extrapolating to a Chinese population. Same data. A different question again.

Endpoint. FDA accelerated approval rests on a surrogate or intermediate endpoint reasonably likely to predict clinical benefit. That standard is what the pathway is built on.

EMA's conditional marketing authorisation is not built that way. It asks four things — positive benefit-risk on the data you have, likelihood that comprehensive data will follow, an unmet medical need, and a benefit of immediate availability that exceeds the risk of the remaining uncertainty. There is no surrogate requirement in that list. There is also a restriction with no FDA counterpart: conditional authorisation is available only for an initial marketing authorisation of a new product, never for a new indication of a medicine already approved. And it renews annually until the obligations are discharged, at which point it converts by an explicit mechanism rather than by drifting into permanence.

So a model output positioned to support an accelerated approval has to survive the question is this surrogate reasonably likely to predict benefit. The same output aimed at a conditional authorisation faces a different question and, if the product is already approved for something else, faces no such pathway at all.

One dataset, two answers

The cleanest public illustration is not about AI, which is why it is useful — it isolates the adjudication from the technology.

ORIENT-11 was a randomised, double-blind, placebo-controlled phase 3 trial of sintilimab with pemetrexed and platinum chemotherapy in non-squamous non-small-cell lung cancer. It enrolled 397 patients, all in China. It met its primary endpoint: progression-free survival of 8.9 months against 5.0 for chemotherapy alone, hazard ratio 0.48.

China's regulator approved the drug on that basis.

In February 2022, FDA's Oncologic Drugs Advisory Committee took up the US application and voted 14 to 1 that additional trials should be required to demonstrate the applicability of those data to the US population and US medical practice. The application did not proceed on ORIENT-11 alone.

Nothing about the dataset changed between those two outcomes. What differed was the question each regulator asked of it — and the panel's reasoning is worth reading closely, because none of it is about data quality. The trial was single-country, and the applicable US regulation requires foreign data to show applicability to the US population and US medical practice. The comparator was chemotherapy alone, which had stopped being the US standard of care. The primary endpoint was progression-free survival where recent US approvals in that setting had leaned on overall survival, and the survival data were immature. And the trial had been designed without FDA engagement, so there had been no opportunity to align on any of the above before the trial was run.

That last point is the one I would put in front of anyone planning a global programme. Three of the four objections were fixable at the design stage and expensive to nothing but a calendar. The fourth objection is the reason the first three were never raised in time.

What this changes for a programme

The last essay ended with four questions. This adds a fifth, and it belongs before the others rather than after:

In whose room does this output have to survive, and does the question change when it gets there?

Then the original four, asked once per room:

• Which decision does this model change?
• Which of the four positions does the output occupy, and which will it need to occupy in eighteen months?
• If the answer is the fourth, what is the context of use, and what would the agency have to accept?
• What evidence has to exist by the time you ask, and when does generating it have to start?

The arithmetic is unforgiving in one specific way. Evidentiary work has lead times measured in programme-years, and a second region's question discovered late is not a question you can retro-fit data to. You cannot go back and enrol a population you didn't enrol, or run a comparator you didn't run. This is the whole argument for asking early, and asking early costs almost nothing.

What I would not conclude from this

I would not conclude that harmonisation is theatre. It isn't. ICH M15 and the joint principles genuinely reduce the amount of separate work a sponsor has to do, and a team that ignored them on the grounds that regions differ anyway would be making a more expensive mistake than the one I'm describing. The claim is narrower: harmonisation of the framework does not deliver harmonisation of the requirement, because the requirement is generated where the framework meets a specific programme.

I would not conclude that the divergences are permanent. Japan's direction of travel over the last decade has been steadily toward global-first development, and China's since 2018 has been toward accepting foreign data under stated conditions. Both are moving. Neither has arrived.

And I would not overstate the AI-specific position. FDA's January 2025 document on AI supporting regulatory decision-making is a draft, which describes current thinking and does not bind. EMA's reflection paper is adopted but is a reflection paper. ICH M15 is final, and regional implementation is still working through. What these tell you reliably is how the agencies reason. That is worth a great deal when you are deciding what evidence to generate two years before you need it, and it is not the same as a rule.

Here is what would refute the argument. If sponsors routinely carried a single credibility package for a position-four use across all four regions, and it was accepted without region-specific evidence, then the distinction I have drawn would be wrong. I haven't seen that. But it is the observation to weigh this against, and it's a fair test.

A note on where this leaves you

If you are running or funding an oncology programme with a global ambition, the useful question is not which agency is strictest. It is which questions your evidence can already answer, and which it cannot — and whether the ones it cannot are the ones that decide the value of the asset.

That question does not change with jurisdiction, because it is a question about the evidence rather than about a filing. It is the question the FIH Diligence Read is built to answer.

Sources

• ICH — M15: General Principles for Model-Informed Drug Development (Step 4, 29 January 2026).
• FDA and EMA — Guiding Principles of Good AI Practice in Drug Development (January 2026).
• FDA — Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products (January 2025, draft guidance — describes current thinking, not binding).
• EMA — Reflection Paper on the Use of Artificial Intelligence in the Medicinal Product Lifecycle (adopted September 2024).
• FDA — Optimizing the Dosage of Human Prescription Drugs and Biological Products for the Treatment of Oncologic Diseases (final, August 2024).
• EMA — Guideline on the Clinical Evaluation of Anticancer Medicinal Products, EMA/CHMP/205/95 Rev. 6, Section 6.
• ICH — E17: General Principles for Planning and Design of Multi-Regional Clinical Trials (Step 4, November 2017; implemented in Japan June 2018).
• MHLW — notification of 25 December 2023 on phase 1 studies in Japanese participants prior to multi-regional clinical trials.
• NMPA — Technical Guidelines for Accepting Data from Overseas Clinical Trials of Drugs, Announcement No. 52 (2018), and the CDE technical guidelines of October 2020.
• 21 CFR Part 314 Subpart H and Part 601 Subpart E; FD&C Act §506(c) — accelerated approval.
• Regulation (EC) No 726/2004 Article 14-a and Commission Regulation (EC) No 507/2006 — conditional marketing authorisation.
• 21 CFR 312.120 — acceptance of foreign clinical studies not conducted under an IND.
• FDA Oncologic Drugs Advisory Committee, 10 February 2022, BLA 761222 (sintilimab). Pivotal trial: ORIENT-11.

A note on sourcing. PMDA and NMPA primary documents are published in Japanese and Chinese respectively. Where this piece describes their positions it relies on the governing international text where one exists — ICH E17 for multi-regional trials — together with English-language regulatory analysis. The ODAC vote and the panel's reasoning are drawn from the public record of that meeting and contemporaneous reporting, including the sponsor's own statement.

← Back to Writing  ·  The decision that is not yours to make →  ·  Where a first-in-human trial begins →  ·  The FIH Diligence Read →