Late last January, a 200-page document landed in the inboxes of pharmaceutical regulatory affairs teams worldwide. ICH M15, as it's known in the shorthand of drug development bureaucracy, did something that would have been unthinkable a decade ago: it explicitly recognized artificial intelligence and machine learning as legitimate modeling approaches for regulatory submissions.
Most pharma executives probably filed it away without a second thought. Technical guidance documents rarely make for gripping reading.
But ICH M15—the first global harmonized framework for model-informed drug development—signals something larger than the sum of its carefully negotiated paragraphs. Digital twins and predictive human simulations, once relegated to academic research labs and speculative conference presentations, are being woven into the regulatory infrastructure itself.
Which means someone thinks they're ready for prime time. Or at least close enough.
The Market That Can't Quite Define Itself
The numbers are flying around, though they don't always agree with each other. One market research firm estimated the global healthcare digital twins market at $2.22 billion in 2025, projecting it will balloon to $69.67 billion by 2035—a compound annual growth rate exceeding 41%. Another pegged last year's market at roughly $1.06 billion.
The disparities matter less than what they reveal: this space remains nascent and definitionally fuzzy. A scoping review published in npj Digital Medicine last October surveyed 149 studies spanning 2017 to 2024 and noted what it diplomatically termed "pervasive misuse of terminology" around what actually constitutes a human digital twin.
Is it a patient-specific computational model of your cardiovascular system? A population-level simulation trained on clinical trial data? A 3D rendering of your tumor derived from MRI scans? The answer, frustratingly, depends on who you ask.
That ambiguity is starting to resolve—not in the definition, which remains contentious, but in the infrastructure required to make digital twins operationally useful. And that's where things get interesting.
The Unglamorous Truth: It's a Data Plumbing Problem
Georgia Witchel has a particular way of describing the challenge. In a launch post earlier this year for her YC-backed startup Mantis—part of the Winter 2026 batch—she framed the product as "Databricks for biomedical and clinical data."
That comparison is revealing. Databricks, for the uninitiated, is an $43 billion data analytics company built on solving the messy reality of enterprise data infrastructure. Witchel, who previously ran a digital twin simulation company for medical device testing, is betting that healthcare faces the same fundamental problem at an even more acute level.
Her company's pitch: 80% of clinical trials face delays due to data inaccuracies, costing an average of $15 million per trial. Mantis describes itself as building "infrastructure for predictive human analytics," combining large language models with physics simulations to convert rare human behavior data into predictive models. The platform aims to integrate electronic data capture systems, clinical trial management platforms, lab results, and omics data with full lineage tracking.
It's not the kind of thing that makes for compelling investor pitch decks. Data plumbing rarely is.
But it speaks to a broader realization percolating through the industry: digital twins in healthcare aren't primarily modeling problems. They're data problems. A 2025 survey from the CHIME Digital Health Most Wired program found that 27% of healthcare organizations reported digital-twin modeling for facility or operational use either live or in pilot. Yet governance gaps persist. Between 67% and 74% of surveyed organizations had deployed at least one AI clinical decision support tool, but many lack the data infrastructure to feed those tools reliably across trial phases or care delivery contexts.
The models, in other words, are getting ahead of the pipes.
Green Lights from the Gatekeepers
ICH M15 isn't arriving in a vacuum. The regulatory winds have been shifting for a while now, though the changes come in the incremental, jargon-laden language that regulators favor.
Back in May 2023, the FDA finalized guidance on covariate adjustment in randomized trials, explicitly encouraging the use of baseline covariates to improve statistical precision and power. To the casual observer, this reads like bureaucratic fine-tuning. To the digital twin developers, it opened the door for prognostic indices—computational models that predict disease progression based on patient data—to be baked directly into trial designs.
Unlearn.ai, one of the earliest movers in this space, has traced an instructive path. The company secured an EMA qualification opinion for its PROCOVA methodology back in 2022. By early 2024, Unlearn blogged that FDA feedback indicated PROCOVA "does not deviate from guidance." A $50 million Series C followed in February 2024.
Then, in March of this year, the company announced it would support SOLA Biosciences' ALS clinical study using its digital twin platform. From regulatory qualification to real-world deployment in less than four years—a timeline that would have seemed wildly optimistic in 2020.
The FDA's Project Optimus, finalized in August 2024, pushed oncology dose optimization toward mechanistic exposure-response modeling and quantitative systems pharmacology. That same month, the agency continued its Model-Integrated Evidence pilot program for generic drugs, initially launched in October 2023, which uses physiologically-based pharmacokinetic models to establish bioequivalence in complex formulations like inhalables and topicals.
In Europe, the Health Data Space Regulation entered into force last March, standardizing electronic health record access and enabling lawful secondary use of data for research across borders. It's the kind of infrastructure that matters less for what it allows today and more for what it will enable when data pipelines mature.
Three years from now, maybe.
The Picks-and-Shovels Players
If digital twins are the gold rush, a cluster of companies are selling shovels and surveying equipment.
Certara's Simcyp PBPK platform, updated as recently as last April, has become a workhorse for physiologically-based pharmacokinetic modeling in regulatory submissions. The Open Systems Pharmacology suite—open-source tools like PK-Sim and MoBi—released version 12 in 2025 and continues to anchor translational modeling workflows that feed into regulatory packages.
These aren't digital twins in the sense of patient-specific organ replicas. They're population models, virtual cohorts, mechanistic simulations. Yet they share a critical dependency: high-quality, longitudinal, multi-modal data that can be queried, versioned, and validated.
Mantis frames its platform as domain-aware, meaning it understands the structure and semantics of biomedical datasets in ways general-purpose data platforms supposedly don't. Whether that's a defensible moat remains to be seen—Databricks itself has healthcare customers, after all—but the framing speaks to a conviction that healthcare data is sufficiently distinct to warrant specialized tooling.
Perhaps it is. Or perhaps the market will eventually consolidate around a handful of horizontal platforms that get good enough at healthcare. The jury's out.
Who's Already in Production
The clinical applications are starting to cluster in predictable places.
HeartFlow received FDA 510(k) clearance last September for its next-generation plaque analysis platform, which uses CT-derived fractional flow reserve models. Dassault Systèmes announced in late February that its Living Heart Project was entering a new phase with AI-powered, customizable virtual hearts in beta testing. SimBioSys, which builds tumor digital models from dynamic contrast-enhanced MRI, entered a strategic collaboration with Mayo Clinic in early 2024.
Mayo itself expanded a partnership with Siemens Healthineers in February, applying digital-twin technologies to surgical pathways in neurodegenerative disease and oncology. Separately, Mayo's neurosurgery department published a case series last August describing how 3D personalized models enhance surgical planning—practical applications of digital-twin concepts already changing workflows, even when they don't involve machine learning.
These examples cluster around imaging-rich specialties: cardiology, oncology, neurosurgery. That's no accident. Digital twins require dense, structured data. Radiology provides it in abundance.
The question is whether the technology can jump the gap into specialties with messier, sparser data. Primary care, for instance. Psychiatry. Chronic disease management outside major academic medical centers.
That's a much harder problem.
Where This Goes
The market will likely bifurcate, as nascent markets tend to do when they mature.
One segment will chase the high-fidelity, patient-specific simulation: the computational model of your heart that predicts how you'll respond to a specific stent design. The other will chase the infrastructure—the platforms that make it possible to build, validate, and deploy those models at scale across trial sponsors, hospitals, and regulatory bodies.
ICH M15's adoption means every major regulatory region—FDA, EMA, Japan's PMDA—will soon require standardized evidence for model-informed submissions, including explicit model risk assessments and context-of-use statements. That formalization will accelerate adoption.
It will also expose which players have real validation data and which have compelling demos.
The governance question lingers. The same CHIME survey that showed 27% digital-twin adoption also revealed persistent gaps in AI governance maturity. If digital twins are to move from pilot projects to core infrastructure, the data provenance, model versioning, and auditability that companies like Mantis are pitching will matter as much as the physics or the algorithms.
Perhaps more. Though nobody wants to lead with "we've solved your metadata problem" when pitching investors.
Still, unglamorous infrastructure has a way of mattering more than the flashy applications built on top of it. Ask anyone who built a business on AWS.
