Last November, while most of the tech world fixated on OpenAI's boardroom drama, the FDA published something arguably more consequential: final guidance on computational modeling and simulation for medical device submissions. Few noticed. Three months later, HeartFlow—a company that simulates blood flow through coronary arteries using nothing but CT scans—announced it had analyzed 250,000 patients. By February, Unlearn, a startup creating AI-powered "digital twins" to replace control groups in clinical trials, closed a $50 million Series C.
The pattern isn't coincidental. Regulators are publishing frameworks. Payers are issuing reimbursement codes. And capital—serious capital—is flowing toward infrastructure that treats human physiology not as mystery, but as computable system.
Consider the language buried in the FDA Modernization Act 2.0, signed in December 2022 with little fanfare. It replaced the requirement for "animal tests" with "nonclinical tests," explicitly permitting in vitro, in silico, in chemico, and non-human in vivo alternatives for investigational new drug and new drug applications. Not encouraged as a nice-to-have. Permitted as regulatory evidence. Digital twins—virtual replicas of human organs, disease progression, even entire hospital systems—have crossed from speculative science into the machinery of drug and device approval.
Which raises a question: Are we witnessing the early industrialization of biology itself?
Defining the Indefinable
The National Academies took a stab at definition in 2024. A digital twin, they said, is a virtual construct that mimics a system's structure, context, and behavior; gets dynamically updated with data from its physical counterpart; has predictive capability; and informs decisions through bidirectional interaction.
That's broad enough to include everything from physics-based heart models to machine learning outcome predictors to hospital capacity simulators. Perhaps too broad. Market estimates reflect the confusion. Coherent Market Insights pegs healthcare digital twins at $1.37 billion in 2025, growing to $6.80 billion by 2032—a 25.7% compound annual growth rate. Mordor Intelligence sees $2.81 billion this year, climbing to $14.12 billion by 2031. One outlier, MarketResearchFuture, projects $234 billion by 2035. The spread tells you as much about definitional chaos as it does about genuine growth potential.
Yet beneath the fuzzy terminology, something concrete is happening.
HeartFlow's FFRct analysis—computational fluid dynamics that derives coronary physiology from standard CT angiography—has now processed more than a quarter-million patients. The company secured CPT Category I reimbursement codes in 2024. Medicare Administrative Contractors and Cigna expanded coverage in 2025. This isn't a research tool anymore. It's a line item in cardiology workflows, competing with invasive catheterization.
Certara's Simcyp, a physiologically-based pharmacokinetic platform used to predict drug-drug interactions, is licensed by 11 regulators globally. In August 2025, it received an EMA qualification opinion—essentially regulatory stamp of approval for a specific use case. Unlearn, after raising more than $130 million total, has deployed its TwinRCT methodology with Merck KGaA for immunology programs. These aren't prototypes in academic labs. They're revenue-generating products navigating the same FDA pathways as any diagnostic device or therapeutic.
Three Converging Forces
Talk to people building in this space—founders, regulators, pharma executives—and three accelerants come up repeatedly.
First: regulatory clarity, which arrived faster than many expected. The FDA's November 2023 guidance on computational modeling and simulation for medical devices, paired with the ASME V&V40 standard, provides what one regulatory consultant called "the first real playbook for model credibility." It's risk-informed, emphasizing verification, validation, and uncertainty quantification tied to context of use. For drug development, the FDA formalized its Model-Informed Drug Development Paired Meeting Program under PDUFA VII, running through fiscal year 2027. This creates structured interactions for exposure models, biological models, and statistical simulations used in dose selection and trial design.
The ICH M15 guideline on model-informed drug development principles reached Step 2b in 2024–2025, harmonizing assessment across global regulators. The EMA published a concept paper in early 2025 outlining forthcoming guidance on mechanistic models—PBPK, physiologically-based biopharmaceutics models, and quantitative systems pharmacology—for assessment and reporting. Translation: the fog is lifting.
Europe's broader data infrastructure is also enabling twin development, though not without friction. The European Health Data Space regulation, adopted in January 2025, mandates interoperable cross-border electronic health record exchange and secondary use. That creates data pipelines that can feed longitudinal models. The EU AI Act, finalized in May 2024, classifies most medical and diagnostic AI as high-risk, adding transparency and testing requirements. It's both obstacle and opportunity—more compliance overhead, yes, but also a formalized pathway where none existed.
Second: technological convergence. GPU-accelerated simulation is collapsing time-to-insight in ways that weren't possible five years ago. Ansys introduced PyAnsys-Heart at NVIDIA's GTC conference in March 2025. NVIDIA's Omniverse platform now offers real-time physics digital twin blueprints with partners including Siemens and Cadence. Dassault Systèmes' Living Heart project—a physics-based cardiovascular model used in FDA collaboration—has expanded to lungs, brain, and liver. These models can now run interactively rather than requiring overnight batch jobs. That shift makes pre-surgical planning and iterative device design genuinely feasible for clinical teams, not just research engineers.
Third: brutal economic pressure. Bringing a drug to market costs roughly $2.6 billion and takes over a decade. Medical device timelines stretch two to four years just for regulatory validation. Any technology that compresses those timelines or reduces the failure rate carries immediate return on investment.
Twin Health—a company using whole-body digital twins for diabetes and hypertension management—published peer-reviewed randomized controlled trial results in NEJM Catalyst in 2025 showing significant outcome improvements and reduced medication reliance. That kind of evidence opens employer and payer adoption, not just venture funding. When your business case shifts from "faster drug development" to "employers will pay us to reverse diabetes," the economics change entirely.
From Coronary Arteries to Clinical Trials
HeartFlow offers perhaps the clearest commercialization signal. Its FFRct analysis uses computational fluid dynamics to simulate blood flow and pressure through coronary arteries from standard CT scans, potentially replacing invasive catheterization for certain patients. The clinical logic is straightforward: Why thread a catheter through someone's femoral artery if you can derive the same physiological information from imaging data you already have?
After crossing 250,000 patient analyses, securing CPT Category I reimbursement codes, and expanding payer coverage through 2024 and 2025, HeartFlow is no longer a niche academic tool. It's embedded in cardiology decision-making.
Unlearn took a fundamentally different approach: augmenting clinical trial control arms with AI-generated digital twins. The company's TwinRCT methodology creates synthetic control patients trained on historical trial data, potentially reducing enrollment requirements and trial duration. After the February 2024 Series C, Unlearn partnered with Merck KGaA for immunology trials. Press reports reference EMA qualification engagement, though the exact scope and timing remain somewhat opaque.
The model architecture, detailed in a 2024 arXiv paper, uses prognostic scores and counterfactual prediction to forecast outcomes for patients who would have been randomized to control. It's not replacing an invasive procedure—it's replacing human enrollment. But it's using the same "digital twin" language and navigating similar regulatory evidence standards.
In hospital operations, GE HealthCare and Siemens Healthineers are deploying operational twins for capacity planning and radiology workflow optimization. GE's ActExcell Operational Twin has been profiled at Children's Mercy Hospital, where it simulates surge scenarios and resource allocation. Siemens' radiology twin models scanner utilization and patient flow.
These aren't clinical models of human physiology. They're discrete-event simulations of hospital systems. Yet they share data integration challenges and predictive validation requirements with patient-specific twins. The definitional boundaries, already fuzzy, keep blurring.
TCS, the IT services giant, built a "Digital Twin Heart" for elite and mass-participation runners, using wearable data and physiological models to predict cardiovascular load during marathons. The program supported Des Linden's training and was deployed at the 2024 Sydney Marathon. It's a sports application, not a medical one—but the boundary is dissolving as wearable-derived insights inform clinical risk stratification.
Then there's Mantis Biotechnology, a Y Combinator Winter 2026 company based in New York. Founder Georgia Witchel describes it as "the infrastructure powering human-in-computer models"—a positioning that spans data integration (unifying electronic data capture, clinical trial management systems, labs, omics, and manufacturing systems), simulation, and regulatory workflows.
On its Wellfound hiring page, Mantis describes itself as a "unified biomedical testing and regulatory platform" linking CAD to simulation to bench testing to FDA Q-Submissions, calling the approach "Palantir for the biomedical industry." Witchel previously co-founded Louiza Labs, a Y Combinator Summer 2025 company focused on physics-driven human simulations for device testing and robotic training, which raised roughly $5 million according to university sources.
The Mantis thesis: fragmented tooling and manual evidence assembly are bottlenecks. A canonical, versioned dataset layer can unify downstream analytics, AI training, and regulatory packages. The company's website emphasizes HIPAA-compliant workflows and forward-deployed engineering for adoption—signaling enterprise sales motion more than self-serve SaaS. Whether that's the right wedge remains to be seen. But the diagnosis of fragmentation isn't wrong.
What Happens Next
Predicting the future is a fool's errand. But three shifts seem likely over the next five years, barring regulatory reversal or catastrophic model failure.
First: hybrid physics-ML architectures will become standard. Pure physics models are interpretable but computationally expensive. Pure machine learning models are fast but lack mechanistic grounding and regulatory trust. Research groups are already publishing physics-informed neural networks for medical twins, and vendors like NVIDIA are offering blueprints that blend real-time GPU-accelerated physics with learned surrogates. That convergence enables both interactive clinical planning and scaled virtual cohorts for trial design. The question isn't whether this happens, but who captures the value.
Second: regulatory acceptance will expand beyond established niches. PBPK models for drug-drug interactions are already routine; Certara's Simcyp is licensed by 11 global regulators and cited in numerous FDA labels. Device computational models are advancing under the FDA's November 2023 guidance and ASME V&V40 framework. The frontier is model-augmented clinical endpoints—using digital twins to forecast outcomes or support control-arm generation. Unlearn's traction with pharma partners and EMA engagement suggests that pathway is opening, but evidence standards remain high. Rightfully so, perhaps.
The FDA's Model-Integrated Evidence pilot for generics and Model Master Files (Type V DMF) program are adjacent proof points. Regulators are signaling openness. But "openness" and "routine acceptance" are different things, and the gap between them is filled with validation studies, post-market surveillance requirements, and a healthy dose of regulatory skepticism.
Third: infrastructure layers will consolidate. Right now, companies are stitching together CAD tools, simulation software, data warehouses, evidence management systems, and regulatory document generators. Mantis Biotechnology is betting that a unified platform—data canonicalization through verification reporting—can capture value. Certara, Dassault Systèmes, Ansys, and Siemens all have pieces of this stack, but they've historically served simulation engineers, not regulatory affairs or clinical operations teams.
The wedge for a startup might be regulatory fluency: automating ASME V&V40 credibility plans, FDA Q-Submission packages, or ICH M15 model reporting templates. That's less sexy than real-time rendering but perhaps more defensible. When your competitive advantage is "we understand how to document uncertainty quantification for the FDA," you're not building a viral consumer product. But you might be building a genuinely valuable business.
Employers and payers are watching too. Twin Health's randomized controlled trial data showing diabetes and hypertension reversal attracted employer adoption. If that model extends to other chronic conditions, it shifts the business case from cost-avoidance to revenue generation. Medicare's willingness to reimburse HeartFlow suggests payers will cover simulation-derived diagnostics when evidence meets traditional benchmarks—no special pleading required.
The Constraints Are Real
None of this is frictionless.
Data interoperability remains fragmented despite the European Health Data Space and similar U.S. initiatives. Privacy and consent frameworks lag technical capability—most patients don't understand what it means for their imaging and lab data to train a predictive model, let alone how to consent to it meaningfully. The EU AI Act's high-risk classification for medical AI adds compliance overhead. And black-box machine learning credibility remains contentious; the FDA's computational modeling guidance emphasizes mechanistic models, leaving uncertainty about how purely data-driven twins will be assessed.
There's also the definitional problem we started with. The phrase "digital twin" spans too many use cases to be analytically useful. A hospital capacity simulator shares almost nothing with a patient-specific cardiac model beyond vague metaphor. Yet they're both called twins. That linguistic inflation might seem trivial, but it creates real confusion in capital allocation, regulatory pathway selection, and commercial positioning.
Biology as Software
Still, the underlying trend seems undeniable: computational models of human biology can now generate regulatory-grade evidence. Not in every case, not without validation, but increasingly and systematically. That's new.
For founders building in this space, the implications are straightforward if not simple. Focus on specific contexts of use, not broad platforms. Partner early with regulators through pre-submission pathways—the FDA's Q-Submission process exists for exactly this reason. And recognize that the infrastructure play may be less about simulation algorithms and more about evidence management and workflow orchestration. The science is maturing faster than the translation layer.
The market estimates will keep varying wildly—$6 billion, $14 billion, $234 billion, pick your number. The definitional debates will continue. But somewhere beneath the noise, a fundamental shift is underway. Human biology is being treated less as irreducible complexity and more as system that can be modeled, simulated, and—perhaps most importantly—regulated.
Whether that's promise or peril depends largely on execution. The bottleneck, as one founder put it, isn't the science anymore. It's everything that comes after.
