Last year, the Food and Drug Administration did something quietly radical: it approved a generic drug using nothing but math.
No patients enrolled. No head-to-head clinical trials pitting brand against generic. Just a physiologically-based pharmacokinetic model—essentially, a sophisticated computer simulation—that convinced regulators a topical diclofenac gel would behave identically to the original in real human skin. The decision barely registered in the press. But for those paying attention, it marked a turning point, perhaps more than the agency itself realized at the time.
The diclofenac approval wasn't an outlier or a one-off experiment. It was a signal. Across drug development, medical device design, and clinical research, the long-held assumption that every hypothesis must be tested in living tissue is starting to fracture. The technical term—"in-silico trials"—lacks poetry, though the industry can't quite settle on what to call this shift. Digital human models. Virtual patients. Computational physiology. Human digital twins. The jargon varies, but the underlying promise remains constant: replace expensive, slow, and ethically complicated animal studies with physics-based simulations that predict how drugs, devices, and treatments will perform in actual people.
It's a vision that's lived in academic labs and engineering departments for decades, the kind of thing researchers presented at conferences to polite skepticism. What's different now? Regulators are ready to accept the evidence. Big Pharma is deploying these models at commercial scale. And a scrappy new wave of startups is racing to build the infrastructure that makes it all work—or at least, trying to.
Money Is Paying Attention
The healthcare digital twin market is experiencing the kind of growth trajectory that makes venture capitalists forget about their other meetings. Mordor Intelligence estimates the sector hit $2.81 billion in 2025 and will climb to $14.12 billion by 2031—a compound annual growth rate north of 30 percent. Grand View Research, typically more conservative, pegs 2024 at $902.6 million and forecasts $3.55 billion by decade's end. Custom Market Insights splits the difference: $2.16 billion in 2023, ballooning to $15.13 billion by 2033.
The wide spread tells you something. This space is still nascent, fragmented, with research firms essentially guessing at the contours of an industry that's only beginning to define itself. But the directional consensus is unmistakable: sustained, double-digit growth for years to come.
The drivers won't surprise anyone tracking healthcare innovation. Personalized medicine demands models capable of simulating individual patient variability—no two diabetics respond to metformin quite the same way. Clinical trial costs keep climbing, a problem one Y Combinator-backed startup, Mantis, claims is exacerbated by the fact that 80 percent of trials face delays from data inaccuracies, costing an average $15 million per study. Animal models, for all their historical utility, routinely fail to predict human outcomes; oncology is littered with compounds that cured mice but devastated people.
Then there's the regulatory shift. The FDA Modernization Act 2.0, signed in December 2022, explicitly permits non-animal alternatives—in vitro systems, microphysiological systems, computational models—to satisfy certain preclinical testing requirements. It's not a ban on animals. But it cracks open a door that was once firmly shut, signaling that methods once considered too exotic for regulatory submission packages are now fair game.
The Unglamorous Truth: It's a Data Problem
Building a credible digital human isn't fundamentally a modeling challenge. It's a data engineering nightmare.
You need motion capture from biomechanics labs. Imaging data from MRIs and CT scans. Electronic health records from clinical trials. Omics datasets from sequencing runs. Wearable sensor streams. Training logs. Lab results spanning years. Each of those sources speaks a different dialect, lives in a different institutional silo, comes with its own compliance headaches. HIPAA here, GDPR there, consent forms that may or may not have anticipated this use case.
Mantis Biotechnology, a three-person team out of Y Combinator's Winter 2026 batch, is betting the data layer is the real bottleneck. Founded in 2025 by Georgia Witchel, the company positions itself as "infrastructure powering human-in-computer models"—which is consultant-speak for building the plumbing no one else wants to touch. On their website, Mantis describes a "domain-aware data platform" that encodes biological and clinical meaning into reusable datasets, integrates with electronic data capture systems, clinical trial management software, labs, and omics platforms, while maintaining full lineage and auditability. Think Databricks, but for biomedical and clinical data. Mantis makes the comparison explicitly in their YC launch post.
Their pitch echoes a broader industry realization: before you can run a validated simulation, you need clean, semantically rich, version-controlled datasets that regulators will actually trust. The stakes aren't abstract. Mantis points to the Simufilam (PTI-125) controversy, where traceability issues around clinical data raised serious questions about trial integrity. If a pharma company can't quickly trace back through its data lineage to prove where a single number originated—especially when billions in market cap hang on that number—that's not just an operational headache. That's an existential risk to an entire drug program.
Regulators Are Moving Faster Than You Think
The FDA's Model-Informed Drug Development Paired Meeting Program, part of the PDUFA VII commitment running through 2027, offers quarterly slots for sponsors to discuss how computational models can inform dose selection, clinical trial design, and safety assessments. In December 2024, the agency released a draft of ICH M15, titled "General Principles for Model-Informed Drug Development," signaling international harmonization on what regulators actually expect. The European Medicines Agency followed in 2025 with a concept paper toward a guideline on mechanistic models—PBPK, physiologically-based biopharmaceutics modeling, quantitative systems pharmacology.
For medical devices, the FDA published final guidance in 2023 on computational model credibility, aligning with the ASME V&V40 risk-based framework. The message is unambiguous: demonstrate that your model is fit for its intended use, with appropriate verification and validation, and regulators will consider it as evidence.
Perhaps the clearest proof? Numbers. A 2025 review in a peer-reviewed journal found that 65 out of 245 new drug applications and biologics license applications submitted between 2020 and 2024—roughly 26.5 percent—included physiologically-based pharmacokinetic modeling. Certara, whose Simcyp PBPK platform dominates the market, claims its software was used in over 120 FDA novel drug approvals and supported more than 300 label claims "in lieu of" some clinical studies. Oncology leads the pack, but use cases now span drug-drug interactions, organ impairment, pediatric extrapolation, and pregnancy dosing—areas where running traditional trials is either prohibitively expensive or ethically fraught.
Real-World Deployments
Dassault Systèmes has been at this for years with its Living Heart Project, a consortium effort to build AI-powered, configurable virtual hearts. The company just extended its FDA collaboration for another five years, explicitly focused on using virtual patients and in-silico trials to support device reviews. The project has already informed pacemaker lead testing and valve evaluations—actual products on the market, not just research prototypes.
In January 2025, the American Heart Association reported on digital heart twins that predicted arrhythmia substrates with 80 percent accuracy, potentially guiding ablation procedures faster than traditional electrophysiological mapping. Eighty percent. Not perfect, but good enough to change how cardiologists plan interventions.
Unlearn takes a different angle. Instead of replacing trials entirely, the company's platform creates digital versions of patients in control arms, using historical data to predict disease progression. That allows sponsors to shrink control groups or boost statistical power without breaking randomization. Unlearn has partnerships with Merck KGaA, Johnson & Johnson, and AbbVie. In late 2024, it announced a collaboration with VectorY to support an ALS Phase 1/2 trial—a disease where recruiting patients is agonizingly difficult. Results presented at the 2024 Alzheimer's Association International Conference showed meaningful sample size reductions, which in trial economics translates to millions saved.
Twin Health, which raised $50 million in 2023, built what it calls a "whole-body digital twin" for metabolic disease management. The system ingests data from continuous glucose monitors, wearables, and nutrition logs to simulate metabolic responses and personalize interventions. Randomized controlled trial data cited by the company shows A1c reductions and deprescribing of diabetes medications. Though the model here leans more toward individualized care optimization than regulatory submission—it's consumer health tech masquerading as precision medicine, depending on your level of generosity.
The Infrastructure Land Grab
NVIDIA is making a characteristically aggressive bet on this space. In 2025, the company announced Isaac for Healthcare, a platform layering MONAI imaging tools, Omniverse simulation capabilities, and Holoscan edge computing to create physics-based digital twins of anatomy, sensors, and clinical environments. A collaboration with GE HealthCare aims to advance autonomous diagnostic imaging, using synthetic data generated from simulations to train AI models—a workaround for the perpetual scarcity of labeled medical images.
Separately, NVIDIA partnered with Mayo Clinic on foundation models built atop enormous pathology datasets, a step toward truly longitudinal, multi-modal "human digital twins" that span imaging, genomics, and clinical outcomes over decades.
The data infrastructure layer is heating up in parallel. Databricks introduced a Lakehouse for Healthcare and Life Sciences, positioning it as a unified platform to make imaging, omics, and EHR data "AI-ready"—whatever that means in practice. Parexel and Palantir expanded their collaboration to bring Palantir's AIP to clinical trial data, promising to unify fragmented sources and accelerate execution. Flywheel built an imaging data platform for trials, devices, and AI development with built-in 21 CFR Part 11 and HIPAA compliance baked in.
The common thread: pharma and medtech companies are realizing that before they can deploy digital twins at commercial scale, they need canonical, lineage-rich datasets that satisfy both AI training workflows and regulatory scrutiny. You can't fake provenance when the FDA comes knocking.
Standards matter here, even if they're boring. The OMOP Common Data Model, maintained by the OHDSI collaborative, has become a de facto standard for observational health data; the UK's CPRD released its primary care data in OMOP format in 2024. CDISC's SDTM and ADaM standards have been required for FDA submissions since 2017. These aren't technical footnotes. They're the semantic foundation that makes cross-study analysis and model validation tractable.
What Comes Next
The next few years will likely see more PBPK and QSP submissions, more devices cleared with computational evidence alone, and more clinical trials augmented—if not partially replaced—by digital twins. The FDA's MIDD program runs through 2027, giving sponsors structured pathways to socialize novel modeling approaches before putting them in formal submissions. The EMA's forthcoming mechanistic modeling guideline will codify expectations across Europe. Standards like ASME V&V40 are maturing from niche engineering documents into regulatory lingua franca.
But this isn't a straight path. Data fragmentation remains a massive hurdle; mapping legacy trial data to OMOP or CDISC is the kind of heavy lifting that makes data engineers weep quietly in Slack channels. Validation benchmarks for many organs and disease states simply don't exist yet. The verification and validation burden is real—regulatory acceptance hinges on context-of-use credibility, not blanket trust in a black-box model.
The EU AI Act, which entered into force in August 2024, imposes new compliance requirements on high-risk AI systems, including medical applications. Companies will need to demonstrate data quality, transparency, and human oversight—vague mandates that will likely spawn a cottage industry of consultants interpreting what "transparency" actually means when your model has 50 million parameters.
Culturally, there's conservatism to overcome. No one—not even the most zealous advocates—is suggesting every drug should skip Phase I entirely and launch based on simulation alone. The diclofenac approval worked because topical pharmacokinetics are well-understood, the compound has decades of safety data, and the model was rigorously validated against known outcomes. Extending that logic to first-in-human oncology trials or rare disease biologics? That's a profoundly different proposition, one that will require not just better models but a fundamental shift in how regulators think about evidence.
Still, the direction of travel seems unmistakable. Digital evidence is moving from fringe to mainstream, from academic curiosity to commercial infrastructure, from regulatory skepticism to cautious acceptance. The companies that figure out the data layer—how to ingest, harmonize, validate, and maintain the inputs that digital twins fundamentally depend on—stand to capture outsize value in a market that credible analysts believe could top $15 billion within a decade.
Whether that's optimism or hype remains to be seen. But when the FDA approves a drug using nothing but a simulation, it's worth paying attention. The future, it turns out, might not need as many lab rats as we thought.
