The phase model is the bottleneck
We’ve all heard this a lot by now. Bringing a new drug or biologic to patients takes about ten years, costs $1–2 billion, and fails more than 90% of the time.
But what most don't realize is that a lot of this cost is not science. It is the start-stop structure of clinical development itself, a Phase 1 trial, then a pause, then a Phase 2 with a new protocol, new sites, new contracts, a new data system, and then the whole thing again for Phase 3.
Every one of those restarts rebuilds infrastructure that already existed a few months earlier. Control participants are recruited fresh for each study. Lessons learned by statisticians, regulatory teams and site coordinators are rarely carried forward in any systematic way. Promising candidates get shelved not because the biology failed but because the economics of running three disconnected trials no longer pencil out.
That model is now being challenged at the federal level. On 30 September 2026, ARPA-H's Proactive Health Office released a solicitation called SURPASS (Simulation-augmented, Real-time Platform Adaptive Seamless Trials) that describes clinical development as a continuous learning system rather than a sequence of trials. It is one of the clearest articulations we have seen of where trial design, in silico modeling and AI-enabled operations are heading together, and it is worth reading closely even if you never plan to bid.
This post walks through the three ideas at its core and what this actually means for anyone who runs trials.
The trial becomes the platform
Instead of one trial per drug per phase, a platform trial is a standing piece of infrastructure that studies a single disease and evaluates many interventions at once. Arms enter, stop or graduate; the platform stays.
Three design features make that work:
- A master protocol sets up platform-level processes, data structures and decision rules once, before launch. Each new drug or biologic is added as a modular appendix rather than a new protocol from scratch.
- Shared controls mean a single, continuously enrolled control population serves every investigational arm. SURPASS calls shared controls a key source of platform efficiency and asks proposers to design for them explicitly, down to choosing placebos and eligibility criteria so that control data can be pooled.
- A seamless (or phaseless) design carries each arm from first-in-patient safety through preliminary efficacy, and ideally through confirmatory evaluation, without a formal stop between phases. The objectives of Phase 1, 2 and 3 are still met but the restarts are gone.
This is where the clinical trial management system has to change. A conventional CTMS is built around a single study with a fixed start, a fixed site list and a fixed set of documents. A platform needs a CTMS that treats the study as the long-lived object and intervention arms as things that attach to it: arm-specific contracts, IRB amendments, supply chains and safety reporting all layered onto a shared operational core. SURPASS puts a number on the ambition. Its final-stage targets include executing the contract for a new arm in under a week, securing IRB approval for that arm in under a week, and moving from FDA's "may proceed" to first subject dosed in under a month. None of those are achievable with a system that assumes every study starts from a blank page.
Governance matters just as much as software. A multi-sponsor platform needs clear decision rights for adding or stopping arms, a way to resolve conflicts between competing sponsors, rules for allocating shared versus arm-specific costs, and a plan for who runs the platform after initial funding ends. The solicitation prefers a disease-expert clinical institution as the long-term operator, which tells you something about how durable these structures are meant to be.
Digital twins move from the slide deck into the protocol
The second shift is that predictive models stop being a research curiosity and become part of how the trial is designed and run. SURPASS groups these under the term in silico capabilities: computer models that predict or simulate patients, disease and trial processes. The list is broad: digital twins and synthetic controls that predict individual outcomes, generative disease-progression models, multimodal longitudinal models that fuse clinical, imaging, biomarker and wearable data, PK/PD models, predictive safety models, and even models of recruitment and retention.
The important move is that each model must have a defined job. The solicitation is explicit that a model needs a stated role in the platform, a decision it informs, the data it requires, and a compute time compatible with operational timelines. A PK model that selects the next dose automatically. A biomarker model that triggers a futility stop. A digital twin that augments the control arm in the primary analysis. A model that cannot produce its output fast enough to inform the decision it was built for does not count.
Two things stand out for anyone building these models.
Simulation becomes the design tool. SURPASS asks for a design engine that can represent the entire seamless platform, run heterogeneous models through common APIs, and estimate operating characteristics such as sample size, time-to-decision, power and bias across a range of plausible scenarios. The target is for a non-programmer trialist to build a simulator in a month and re-simulate the design in under five hours after changing a parameter. That is a very different relationship to trial design than a biostatistician running a one-off power calculation.
Control replacement is graduated and earned. Digital twins may start in an exploratory role that does not touch the primary analysis. As validation evidence accumulates, they are promoted to formal use, with the goal of reducing human control allocation by a third in year 3, half in year 4 and 80% in year 5. Every step up must preserve the statistical and regulatory validity of the treatment-effect estimate.