Drug development has a famously unforgiving arithmetic: more than a decade of work, somewhere between $1 billion and $2 billion per approved medicine, and a failure rate above 90 percent. On September 30, the U.S. Department of Health and Human Services decided to attack that arithmetic with software. Through its research agency ARPA-H, it launched a five-year program called SURPASS, built around simulation, real-time statistics, and automation, plus three companion projects aimed at the plumbing around clinical trials.
What Happened
SURPASS stands for Simulation-augmented, Real-time Platform Adaptive Seamless Trials. According to the HHS announcement, the program will fund technology that lets researchers evaluate drugs and biologics faster, at lower cost, and with fewer participants. Health Secretary Robert F. Kennedy Jr. framed the goal in national terms: "America should lead the world in turning medical breakthroughs into treatments."
The program is split into three technical areas. The first is a "phaseless design engine," which uses digital twins and predictive models to simulate how a trial is likely to behave before a single patient is enrolled, and which is meant to replace the rigid Phase 1, 2, and 3 sequence with something closer to one continuous study. The second is a continuous inference engine that analyzes data as it accumulates, so a trial can adapt midstream and may need smaller control groups while still producing rigorous evidence. The third is an agentic operations layer that automates trial startup, data collection, and data cleaning, work that today consumes enormous amounts of staff time.
ARPA-H also announced three complementary projects. STACK uses AI to speed up clinical site activation. COMMONS aims to build a nationwide, privacy-by-design infrastructure for sharing trial data. CINCH, led by program manager Erika Kim, is meant to make it easier for patients to contribute real-world data and navigate their care. The agency is seeking cross-disciplinary teams spanning statistics, AI, trial design, operations, and regulation, and BioSpace reports that applications for a related effort, Operation TrialBlazer, are open until October 30.
Why It Matters
Most of the attention around AI in medicine has gone to discovery: models that propose molecules or read scans. SURPASS targets a later and arguably costlier bottleneck. A candidate drug can look excellent in the lab and still spend years crawling through trials that are slow to open, slow to enroll, and slow to analyze. If the leading estimates are right, the trial stage is where much of the time and money disappears.
The ambition is large. According to BioPharma Dive, program managers suggest the innovations could cut development timelines in half and reduce costs roughly tenfold. Those are aspirational figures from planning documents, not results, and nothing has been built yet. Still, the design philosophy is notable: instead of a sequence of disconnected trials, the stated aim is a continuous learning system in which every data point feeds the next decision.
The technical bet is also worth separating from the hype. Digital twins and adaptive statistics are not new ideas, and regulators have accepted adaptive designs in limited settings for years. What SURPASS adds is the attempt to make them standard infrastructure, with shared, publicly available regulatory frameworks so that the whole ecosystem, not just one sponsor, can adopt them.
Reaction
The political framing came through louder than the technical one. Officials explicitly tied the launch to competition with China, and the numbers they cite explain why. China passed the United States in its share of Phase 1 trials in 2021 and registered more total clinical studies than the U.S. in 2024, the first time that had happened.
The comparison is not only with China. Industry coverage notes that Australia can open trials six to twelve months faster than the United States at roughly half the cost, which is a pointed benchmark for a country that sees itself as the home of biotech. China, meanwhile, offers lower costs and a lighter regulatory burden. In that light, SURPASS reads less like a moonshot and more like an attempt to close a widening operational gap.
Early reporting also flagged what was missing. STAT noted that the announcement did not specify funding amounts, and it said little about participant incentives or how resources would be allocated. For a program that depends on outside teams stepping forward, those details will shape how much talent it actually attracts.
What's Next
SURPASS is expected to get going this fall, with proposals solicited across its three technical areas. The first real test will not be a trial result but a set of open questions: who wins the awards, how much money is behind them, and how quickly the program can turn simulation outputs into evidence a regulator will trust.
The regulatory piece may matter most. A phaseless design only helps if the people approving drugs accept the evidence it produces, which is why the publicly available frameworks are central to the plan rather than a footnote. Watch also for how COMMONS handles consent and privacy, because a national data layer is only as useful as the public's willingness to contribute to it.
For readers following applied AI, the pattern is worth tracking. Earlier waves focused on models that were smarter. This one is about infrastructure, rules, and operations, the unglamorous parts that decide whether a technology changes outcomes outside a demo.
Closing Thoughts
It is easy to read SURPASS as a story about speed, and speed is a real prize. A patient with a serious illness does not experience a trial timeline as an abstraction. But the more durable idea in the announcement may be the one about participants: fewer people needed, less burden on each, and an easier way to contribute data. If those promises hold, faster trials and kinder trials could turn out to be the same thing.
.jpg){kind=link}
The caution is equally plain. Simulations are only as good as the data behind them, and a program of this scale will be judged by evidence, not announcements. If SURPASS delivers even a fraction of its stated targets, it will be remembered less for the algorithms than for the quiet fact that a decade-long process finally started to learn as it went.
한글 요약
미국 보건복지부(HHS)와 ARPA-H는 9월 30일 임상시험을 AI와 시뮬레이션으로 혁신하는 5년짜리 프로그램 SURPASS를 발표했습니다. 디지털 트윈으로 시험 결과를 미리 예측하는 '단계 없는 설계 엔진', 데이터를 실시간 분석하는 추론 엔진, 시험 개시와 데이터 정리를 자동화하는 에이전트 운영 계층이 핵심입니다. STACK, COMMONS, CINCH 등 세 가지 보조 프로젝트도 함께 공개됐습니다.
신약 개발은 현재 10년 이상, 최대 20억 달러가 들고 실패율이 90%를 넘습니다. 프로그램 책임자들은 기간을 절반으로, 비용을 10분의 1로 줄일 수 있다고 제시했지만, 이는 아직 목표치일 뿐 검증된 결과가 아닙니다. 중국이 2021년 임상 1상 점유율에서, 2024년 전체 임상시험 등록 건수에서 미국을 앞선 점이 이번 발표의 배경으로 거론됩니다.
다만 구체적인 예산 규모와 참가자 지원 방안은 공개되지 않았습니다. 앞으로는 가을 중 제안 공모, 규제 당국이 시뮬레이션 기반 증거를 받아들일지 여부, 국가 데이터 인프라의 개인정보 보호가 핵심 관전 포인트입니다. 참고: HHS 보도자료, BioPharma Dive, STAT, BioSpace.