ARPA-H Launches IGoR to Fix the Replication Crisis With AI. Solution Summaries Are Due June 25.

On May 5, 2026, the Advanced Research Projects Agency for Health quietly posted Notice ID ARPA-H-SOL-26-155 on SAM.gov and announced the launch of the Intelligent Generator of Research program — IGoR. The press cycle treated it as another AI-in-medicine headline. It is something considerably more ambitious. IGoR is a deliberate attempt to rebuild the operating system of biomedical research itself, and ARPA-H wants the first solution summaries on its desk in five weeks.

The agency has been telegraphing this move for the better part of a year. Senior program staff have spoken at conferences about the replication crisis — the persistent finding, going back to the Ioannidis 2005 paper and the 2015 Open Science Collaboration project, that a startling fraction of published biomedical results cannot be independently reproduced. Estimates vary depending on field, but in some areas of preclinical cancer biology, fewer than one in five published findings have replicated successfully. The economic damage runs into the tens of billions of dollars annually in dead-end downstream drug development.

IGoR is ARPA-H's structural answer to that problem. Not another reproducibility working group. Not another reporting checklist. A new ecosystem that, if it works, would generate validated biological knowledge "at least ten times faster than traditional research approaches" — the agency's words, not a marketing inflation.

For grant-funded researchers, biotech founders, AI labs, and contract research organizations, IGoR is one of the most consequential funding opportunities ARPA-H has issued. It is also one of the strangest, structurally. Here is what is actually on offer, what kinds of teams stand a chance, and why the four-week sprint to Solution Summary submission demands an organizational response that starts this week.

What ARPA-H Is Actually Buying

Most federal R&D programs procure a deliverable: a device, a drug candidate, a dataset, a manuscript. IGoR is procuring an apparatus — a closed-loop, self-improving research engine that other scientists can use indefinitely after the program ends.

The Funding Opportunity decomposes that apparatus into four interlocking technical components, and the agency expects winning teams to address all four.

Mechanistic disease models that encode causal biological relationships across scales. This is not "another machine-learning model trained on TCGA." ARPA-H is asking for executable, computational representations of disease biology that bridge molecular, cellular, tissue, and organ levels — the kind of model where you can intervene at the gene-expression layer and see a predicted phenotype at the organism layer. Think systems biology meets digital twins, applied to chronic conditions like Alzheimer's disease, type 2 diabetes, and major cancers.

An AI orchestration layer that identifies knowledge gaps and designs optimal experiments. The system must look at the current state of the disease model, identify the assertions where evidence is weakest or contradictory, and then design the single experiment that would most efficiently resolve the uncertainty. This is active learning, but at the level of an entire research program rather than a single classifier.

A protocol architecture enabling reproducible execution across laboratories. Every experiment the system designs must be specified in a way that a different lab, with different equipment and different hands, can execute and return comparable data. The agency is asking for something close to a "compilable" experimental protocol — versioned, parameterized, machine-readable, and instrument-aware.

A distributed marketplace of validated laboratories that execute standardized protocols and return gold-standard data. This is the most novel component. ARPA-H is not just funding software. It is funding the construction of a federated network of wet labs — academic cores, contract research organizations, and possibly automated cloud labs — that can be commissioned to run the designed experiments and pipe the resulting data back into the model. The validated lab tier is the part that turns IGoR from a thought experiment into an industrial-scale knowledge factory.

The four components have to talk to each other. A team that can build a brilliant disease model but cannot stand up a lab network is not competitive. A team that can run a lab network but cannot do the agentic AI orchestration is not competitive. This is the central organizational puzzle of IGoR.

Program Structure and Award Mechanics

IGoR is a 5-year, 3-phase program under ARPA-H's Proactive Health Office. The agency anticipates awarding approximately three Other Transaction (OT) agreements to teams capable of executing across all four technical areas.

The Other Transaction mechanism matters. OTs are not grants and they are not standard contracts. They give ARPA-H wider latitude to set milestones, restructure deliverables mid-program, and bring in non-traditional performers — startups, large companies, and consortia that would struggle to receive a NIH R01. The downside is that OTs often require significant cost share, especially when traditional defense or commercial firms participate, and the IP and data rights terms are negotiated rather than off-the-shelf.

The eligibility language is broad: "academic institutions, non-profit organizations, companies, or a combination of highly skilled performers across sectors." Read that as an explicit invitation for mixed consortia. ARPA-H is not looking for a single university lab to do everything. The program structure assumes a prime contractor with multiple subawardees, each contributing depth in one of the four technical areas.

Three awards across a 5-year, 3-phase program implies sizeable team-level funding — likely tens of millions of dollars per consortium across the life of the program, though ARPA-H has not published a per-award ceiling and OT amounts are negotiated. Compare this to the typical biomedical R01 ($250,000–$500,000 per year for 4–5 years) and the difference in ambition becomes clear. IGoR awardees will be running operations comparable in scale to a small biotech company.

Key Dates Compressed Into Six Weeks

The timeline is unforgiving and has to drive every team's behavior from now until late summer.

• June 9, 2026 — Proposers' Day in the Washington, D.C. metro area. Optional, but the only structured opportunity to ask the program manager direct questions and to meet potential partners.
• June 25, 2026, 12:00 PM ET — Solution Summaries due.
• August 6, 2026, 12:00 PM ET — Full proposals due (from teams invited to submit after Solution Summary review).

That gives a serious team five weeks from today, May 20, to assemble a consortium and produce a Solution Summary that articulates an approach across all four technical components. The agency has also disclosed that it is piloting secure large language model tools to assist with the initial review of Solution Summaries — a notable methodological choice. ARPA-H's program manager Paul E. Sheehan, Ph.D., has published a statement on the use of AI tools in IGoR proposal reviews, and applicants who write fuzzy, jargon-dense narratives should expect those signals to be detected at scale by the review apparatus itself. Clarity has always been a virtue in proposal writing; under LLM-assisted triage it becomes a precondition.

How Different Kinds of Teams Should Position

Established AI/ML labs at research universities. You almost certainly cannot win IGoR alone, but you bring the most credible piece: the orchestration layer. Your immediate question is partnership. Who builds the mechanistic disease models? Who operates the lab network? If you do not have those answers by the Proposers' Day, you should be at the Proposers' Day specifically to find them. Bring a one-pager.

Computational biology and systems biology groups. You own the disease-model component. The risk is that you build a beautiful in-silico model that nobody else in the consortium can interrogate. Push for a model architecture that exposes well-defined APIs — both for the AI orchestrator (to query knowledge gaps) and for the lab network (to receive experimental designs and ingest results).

Biotech and pharma R&D organizations. You may not lead a Solution Summary, but you have the disease expertise and possibly the lab infrastructure to be the indispensable industrial partner. Your value proposition to a university prime is regulatory familiarity, scaled wet-lab operations, and the institutional knowledge of which experiments actually distinguish hypotheses. Be explicit about data rights — IGoR data must flow back into the public model, which is incompatible with proprietary lock-up.

Contract research organizations and automated cloud labs. This is your strongest opportunity in years. The "distributed marketplace of validated laboratories" component is, in effect, ARPA-H subsidizing the creation of the infrastructure CROs have been building for two decades. Position yourself as the network anchor — the entity that can validate protocols, certify executing labs, and operate the quality system that makes the data trustworthy.

Small businesses and non-traditional performers. OT eligibility means you can compete. The realistic path is as a subawardee with a sharp, narrow capability — protocol formal-methods tooling, lab automation, federated data infrastructure, agentic LLM systems specifically tuned for experimental design. Reach out to academic and industrial primes who have public AI-for-science profiles. The Proposers' Day attendance list will be a who's-who of those candidates.

Strategic Risks and Why This Could Still Fall Short

IGoR's vision is exhilarating; the execution risks are real and worth naming.

The first is the lab-network problem. Federated experimental networks have been attempted before — the NCI Cancer Models Initiative, the Reproducibility Project: Cancer Biology, the Allen Institute's open-data ecosystem — with mixed results. Standardization is brutally hard. Reagents differ. Mouse colonies differ. The same antibody from the same vendor performs differently in two labs. ARPA-H will need to fund not just the network but the metrology that allows the network to be trusted.

The second is the AI-design problem. Active learning for experimental design works in narrow, well-modeled domains. Biology is not a narrow, well-modeled domain. There is a real possibility that the AI orchestration layer will produce experimental recommendations that are technically correct but biologically irrelevant — the kind of "hallucinated" suggestion that wastes lab capacity and erodes trust in the system. Teams that surface this risk in their Solution Summary, and propose human-in-the-loop validation against domain experts, will read as serious.

The third is the chronic-disease focus. Alzheimer's, diabetes, and cancer are precisely the areas where decades of well-funded research have produced disappointing translational yields. There is a legitimate critique that the bottleneck in those fields is not the speed at which experiments are run, but the conceptual frameworks scientists are working within. IGoR can accelerate the wrong questions just as efficiently as the right ones. The strongest proposals will be explicit about how the disease models are validated against clinical outcomes, not just internal coherence.

The Larger Bet

IGoR sits inside a broader 2026 federal pattern of AI-for-science initiatives. NSF's AI Institutes, DOE's Office of Science AI investments, and ARPA-H's own Intelligent Generator of Research are converging on the same proposition: that the current model of human-led, hypothesis-driven biomedical research is too slow, too irreproducible, and too expensive for the scale of disease burden the country faces. Federal funders are placing significant capital on the alternative — closed-loop, AI-orchestrated, machine-readable science.

That bet may be premature. It may be exactly right. What is no longer in doubt is that the funding pipeline is being reshaped around it. Researchers and organizations that build the muscle to operate inside agentic, reproducibility-first research systems will find themselves on the inside of the next decade of biomedical grant funding. Researchers who do not will find that the rubrics have changed underneath them.

If you intend to compete for IGoR, the next forty days matter. Convene partners by the end of next week. Have a Solution Summary outline in circulation by June 1. Plan to attend the Proposers' Day on June 9. And spend serious time on the question that ARPA-H has placed at the center of the program: not "what experiment do I want to run," but "how does my consortium prove that the experiments anyone runs will produce data that anyone else can trust."

That question is older than AI. IGoR is just the largest check anyone has written to answer it.