Strategic Advisor | Ex-FDA | Helping Leaders Navigate Complex Decisions From Innovation to Impact | Development, Regulatory, CMC & Organizational Strategy
{This article was initially published on LinkedIn on September 7, 2026: Development Strategy Series — Article 3: Evidence Architecture | LinkedIn}
Designing evidence around the decisions a development program needs to make
Drug development teams are very good at designing studies.
They define objectives, endpoints, populations, assays, statistical plans, timelines, operational requirements and regulatory deliverables.
The problem is that a drug or device development program is really more than a collection of studies. It is a sequence of decisions made under uncertainty.
That frame points us to a different strategic question.
Rather Than Asking: What study should we run next?
A More Useful Development Question Is: What do we need to know next, why does it matter, and which decision will that knowledge change?
That is the idea behind evidence architecture.
The goal is not to generate the most evidence.
It is to generate the right evidence, in the right sequence, before the decisions that depend on it.
Start with the decision question, not the study
Development planning often begins with an activity.
We need another cohort. We need additional stability data. We need a toxicology study. We need a larger efficacy trial.
All of those may be appropriate. But evidence does not become strategically valuable simply because it adds data. Its value comes from what it allows the program to decide.
What decision are we trying to make?
What uncertainty prevents us from making it confidently?
What evidence would materially change that uncertainty?
The decision might involve dose selection, formulation, manufacturing scale-up, indication sequencing, regulatory strategy, financing, partnering or whether to proceed at all.
The key point is that the decision question comes before the evidence-generation activity. Otherwise, programs can accumulate large amounts of data without necessarily becoming more decision-ready.
Not all uncertainty needs to be resolved now
Every development program contains uncertainty.
Some must be reduced quickly. Some can safely be carried. Some becomes more dangerous with time. And some may never justify the cost of resolving it.
This is why evidence architecture is not the same as gap analysis.
A gap analysis can make every unknown look like a problem requiring another study.
Which uncertainties matter enough to the next consequential decision that we should pay to reduce them now?
There is a cost to learning. Every experiment, cohort, assay, engineering run and follow-up period consumes capital and calendar time.
But there is also a cost to not learning.
Insufficient evidence may later appear as a protocol amendment, regulatory delay, manufacturing constraint, weak diligence position or evidence package that no longer answers the questions future stakeholders care about.
So the objective is not maximum certainty. It is enough confidence to make the next decision while preserving the options that still matter.
Good evidence architecture is partly an exercise in deciding what not to learn yet.
Evidence has more than one audience
Evidence is also rarely consumed by only one decision-maker.
Regulators, investors, partners, clinicians, patients, payers and health systems may all look at the same program through different lenses.
That does not mean every early study should satisfy every future stakeholder. It means teams should understand what today’s evidence choices may make easier—or harder—to demonstrate later.
A trial population may establish efficacy while limiting generalizability. An endpoint may support approval while leaving uncertainty around clinical use. A manufacturing strategy may support early supply but create significant comparability challenges later. A formulation choice may work scientifically but constrain distribution or administration.
Does this evidence choice preserve important future options or quietly close them?
That is where evidence architecture moves beyond study planning and becomes development strategy.
Sequence learning around dependencies
Evidence also has dependencies.
A biomarker strategy may require analytical validation before it becomes clinically useful. Manufacturing scale-up may depend on sufficient process understanding. Comparability requirements may depend heavily on when a manufacturing change occurs. Early clinical findings may alter later nonclinical or CMC questions.
A development program is therefore not simply a list of evidence packages. It is a learning sequence.
When the sequence is wrong, teams may generate evidence before anyone can act on it—or generate the right evidence after the decision it was meant to inform has already been made.
That is why a highly detailed development plan can still contain significant strategic fragility. The activities may be scheduled. The learning dependencies may not be.
A practical Evidence Architecture Canvas
For each consequential evidence package, I would ask seven questions:
1. Decision — What decision must be made, and by when?
2. Uncertainty — What do we not know that could materially change that decision?
3. Evidence — What evidence would meaningfully reduce that uncertainty?
4. Dependency — What must be known, validated or completed first?
5. Leverage — Which other decisions or stakeholders could this evidence inform?
6. Optionality — What happens if we defer this evidence—or if the underlying assumption is wrong?
7. Cost and timing — What does generating this evidence consume in capital, calendar time and program risk?
Taken together, these questions turn an evidence plan into a map of decisions, uncertainties, dependencies and options.
And a historical comparison from immuno-oncology helps illustrate why that matters.
Case Study: Keytruda and Opdivo — same target, different evidence roles
Merck’s pembrolizumab (Keytruda) and Bristol Myers Squibb’s nivolumab (Opdivo) both targeted the PD-1 pathway and generated extensive evidence around PD-L1.
But PD-L1 did not play the same role in the two development strategies.
That is what makes the comparison useful.
Keytruda: evidence that progressively sharpened patient selection
For Keytruda, early clinical activity first addressed a fundamental question: was the signal strong enough to justify accelerated development?
FDA granted accelerated approval for pembrolizumab in melanoma in September 2014. In October 2015, it received accelerated approval in previously treated metastatic NSCLC whose tumours expressed PD-L1 as determined by an FDA-approved test.
As the program evolved, PD-L1 became increasingly important to patient selection in several Keytruda indications.
The evidence questions became progressively more specific:
- Is the drug active?
- How durable is that activity?
- Which patients benefit most?
- Can a biomarker improve patient selection?
- Which indications justify broader or biomarker-defined development?
The evidence architecture changed as the decisions changed.
Opdivo: learning from a biomarker without necessarily making it a gate
Opdivo provides a useful contrast.
BMS also generated extensive PD-L1 evidence, but in non-squamous NSCLC the 2015 approval of the PD-L1 IHC 28-8 pharmDx assay stated that PD-L1 expression may be associated with enhanced survival from nivolumab. The assay provided clinically meaningful information without initially acting as a mandatory treatment-selection gate.
That preserved a different kind of optionality: PD-L1 could inform understanding of treatment effect while broader nivolumab development continued.
The role of the biomarker later evolved.
In 2020, the 28-8 assay became a companion diagnostic for nivolumab plus ipilimumab in NSCLC at a PD-L1 tumour-cell expression threshold of at least 1%.
The important difference is not which strategy was better
The important lesson is not that one company had the better biomarker strategy.
It is that the same category of evidence can serve different strategic purposes at different points in development.
For one program, biomarker evidence may become a treatment-selection tool. For another, it may initially characterize differential benefit, preserve broader eligibility and later become decision-critical in a specific combination or indication.
A Typical Question: Do we need a biomarker?
A More Useful Question: What role does this evidence need to play in the next decision?
That is evidence architecture.
Evidence architecture should remain dynamic
The Keytruda and Opdivo examples also show why evidence architecture cannot be fixed once at the beginning of development.
The environment changes. Clinical signals emerge. Regulators provide feedback. Assays mature. Competitors generate new evidence. Manufacturing experience changes feasibility. Standards change. Capital availability changes.
The evidence that looked most valuable six months ago may no longer be the evidence the program most needs today.
A Typical Governance Question: Are the studies on track?
A More Useful Development Question: Are we still learning what the program most needs to know?
That shifts governance from activity monitoring toward decision quality.
From evidence plans to decision architecture
The strongest programs are not necessarily those with the largest evidence packages.
They are the ones where evidence arrives when it can change a decision.
Where important uncertainties are surfaced early. Where dependencies are understood. Where the team knows what must be learned now and what can remain open. And where evidence choices preserve enough optionality to respond when the development landscape changes.
That matters particularly in biotechnology, where scientific uncertainty is high, capital is finite and the pathway can change quickly.
A program cannot investigate everything. It has to choose what to learn. And those choices become part of the development strategy.
What decision is waiting for this evidence—and what will we do differently when we have it?
That is the shift from generating data to designing a development pathway that learns deliberately.
And ultimately, from an evidence plan to a decision architecture.
Selected source links for the case study
FDA: Verified Clinical Benefit – Cancer Accelerated Approvals: https://www.fda.gov/drugs/resources-information-approved-drugs/verified-clinical-benefit-cancer-accelerated-approvals
FDA: List of FDA Authorized Companion Diagnostic Devices: https://www.fda.gov/medical-devices/in-vitro-diagnostics/list-fda-authorized-companion-diagnostic-devices-in-vitro-and-imaging-tools
FDA: PD-L1 IHC 28-8 pharmDx (P150025): https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpma/pma.cfm?id=P150025
FDA: PD-L1 IHC 28-8 pharmDx supplement (P150025S013): https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfpma/pma.cfm?id=P150025S013