A surgeon at Boston Children's Hospital needs to repair a congenital heart defect in a four-year-old. Before touching a scalpel, she consults a computer simulation—a digital replica of the child's anatomy that predicts how blood will flow after the procedure. Across the Atlantic, a drug company planning a diabetes trial cuts its control group by half, filling the gap with synthetic patients generated from historical data. In both cases, regulators are signing off.
This isn't speculative. The regulatory infrastructure for computational medicine arrived faster than most people noticed.
In December 2024, the FDA and its international counterparts quietly published draft guidance for Model-Informed Drug Development—the culmination of three decades of pharmaceutical industry lobbying. Three months after that, Eli Lilly and NVIDIA announced they'd pump up to $1 billion into a lab dedicated to "in silico biology." Then, this past August, European regulators granted their first formal qualification to a physiologically based pharmacokinetic platform, blessing it for use in actual drug submissions.
What's happening goes beyond regulatory catch-up. A new infrastructure layer is taking shape—one built to validate, standardize, and operationalize human-in-computer models across drug discovery, medical device testing, and clinical trials. The infrastructure race is on. And the winners will fundamentally reshape how medicines and devices reach patients.
When the Market Numbers Tell Half the Story
Healthcare digital twins represented a $902.6 million market in 2024, at least according to Grand View Research, with North America claiming nearly 47% of the action. By 2030, analysts project $3.55 billion, a compound annual growth rate hovering around 26%. Other forecasters run hotter: Custom Market Insights pegged the 2023 baseline at $2.16 billion and sees $15.13 billion by 2033.
The in-silico clinical trials segment alone could hit $6.39 billion by 2033, per a November 2025 analysis. Device simulation accounted for roughly 28% of activity in 2024 and 2025. Software dominates the stack—nearly 79% of the market—reflecting that compute and modeling platforms, not just consulting services, are becoming products regulators will actually approve.
But market sizing, however eye-catching, tells only part of the story. The real shift is regulatory formalization—the moment abstract simulation becomes accepted evidence.
On August 4, 2025, Certara's Simcyp Simulator became the first PBPK platform to receive a European Medicines Agency Qualification Opinion for specified contexts of use, including cytochrome P450-mediated drug-drug interaction risk assessment. That milestone matters more than the press release suggested. Sponsors can now reference a pre-qualified tool in submissions rather than re-validating models from scratch every single time.
In the United States, the FDA's Model-Informed Drug Development Paired Meeting Program—launched under the latest PDUFA reauthorization—now offers sponsors quarterly slots to align on exposure-based, biology-based, and statistics-based models before pivotal trials begin. The agency also introduced Model Master Files in 2025, enabling reusable evidence packages for generic drug applications. And the December 2024 ICH M15 guidance on general principles for MIDD created, for the first time, a harmonized international assessment framework.
It's the kind of regulatory plumbing that doesn't make headlines but changes how billion-dollar bets get made.
Physics Meets Machine Learning at Industrial Scale
Traditional digital twin platforms leaned heavily on physics-based simulation: computational fluid dynamics for blood flow, finite element analysis for device mechanics, physiologically based pharmacokinetics for drug absorption and distribution. Rigorous, yes. But computationally expensive and difficult to personalize at scale.
The convergence happening now pairs physics solvers with machine learning—a marriage that's proving more fruitful than many expected. Researchers are embedding physics constraints directly into neural networks, creating so-called physics-informed neural networks (PINNs) that accelerate simulations while preserving biological plausibility. A 2025 preprint demonstrated inverse PINNs for brain PBPK modeling. Another 2026 paper proposed "Physiologically Informed Deep Learning" frameworks for next-generation PBPK tools.
Dassault Systèmes signaled this shift in February 2025 when it announced a beta test of an AI-powered, parametric whole-heart digital twin that can be configured to individual patients or virtual populations. The company's Living Heart Project—which began as an FDA collaboration to validate cardiac device simulations—has since sprawled into the Living Brain, Living Lungs, and Living Liver. Steve Levine, who founded the Living Heart Project, argued in a Fortune commentary that virtual twins will eventually extend to whole-body models capable of supporting "virtual human subjects" in early-phase trials.
Perhaps he's getting ahead of himself. Or perhaps not.
Meanwhile, NVIDIA's BioNeMo platform—a GPU-accelerated stack for generative biology—is being adopted across biopharma. The January 2026 announcement of a co-innovation lab with Eli Lilly, valued at up to $1 billion over five years, underscored just how compute-intensive this work has become. In silico trials at population scale demand high-performance computing resources once reserved for climate modeling. The industry is effectively simulating weather systems, except the atmosphere is a human body.
Real-World Validation Stops Being Theoretical
HeartFlow's fractional flow reserve computed from CT angiography—FFRCT in the industry shorthand—has become the poster child for clinically validated simulation. A Nature Medicine study published in May 2025 tracked more than 90,000 NHS patients over two years and found that CTCA plus FFRCT reduced unnecessary invasive coronary angiograms compared to CTCA alone. A separate five-year peripheral artery disease cohort study, published in the Journal of Vascular Surgery in May 2024, reported a greater than 60% mortality reduction with FFRCT-guided care. HeartFlow now cites over 600 peer-reviewed publications supporting its non-invasive coronary physiology models.
The company has reimbursement, clinical adoption, and the kind of evidence base that makes skeptical cardiologists pause.
FEops received FDA De Novo authorization in October 2021 for its HeartGuide platform, which uses patient-specific digital twin simulations to plan left atrial appendage occlusion procedures. The company expanded U.S. availability in July 2023, backed by randomized controlled trial evidence from the PREDICT-LAA study.
Twin Health's whole-body metabolic digital twin produced striking clinical trial data in a Cleveland Clinic-sponsored RCT. Results published in NEJM Catalyst showed 71% of participants achieved HbA1c below 6.5% versus just 2.4% in usual care, with significant medication de-escalation to boot. In August 2025, Twin Health closed a $53 million Series E following those trial results, demonstrating payer and employer appetite for outcomes-based digital twin models. When self-insured employers start paying, skepticism fades fast.
On the drug development side, Unlearn's TwinRCT platform—built to generate synthetic control patients using prognostic scores from historical trial data—earned a draft EMA qualification opinion for its PROCOVA method. The approach enables sponsors to reduce control arm enrollment by 25% to 50% in certain continuous-endpoint Phase 2 and 3 trials, according to CEO Charles Fisher. AbbVie used Unlearn in an Alzheimer's disease trial, and the company's regulatory footprint is expanding.
Aitia has built what it calls "Gemini disease twins," combining causal AI with multi-omics data. The company signed multiple collaborations with Servier across multiple myeloma (2022), pancreatic cancer (2023), Parkinson's disease (January 2024), and gliomas (late 2024). A November 2023 partnership with Charles River Laboratories applied Aitia's Logica platform to patient-derived xenograft digital twins in oncology and neurodegeneration.
The proof points are stacking up. Not uniformly, not without controversy, but faster than regulators anticipated.
The Regulatory Infrastructure Takes Shape
The FDA Modernization Act 2.0, signed December 29, 2022, removed statutory language requiring "animal tests" for nonclinical safety evaluation, permitting "nonclinical tests" that explicitly include in silico methods. That legislative green light—buried in end-of-year spending bills—accelerated agency pilot programs more than the law's sponsors probably expected.
The FDA's Center for Devices and Radiological Health now formally recognizes ASME V&V 40-2018, a risk-based credibility framework for computational models used as evidence in device submissions. MDIC—the Medical Device Innovation Consortium—has convened recurring symposiums on computational modeling, simulation, and what they're calling "digital evidence." The 2024 and 2025 symposiums featured case studies on congenital heart surgery planning pipelines at Boston Children's Hospital and end-to-end applications of V&V 40 principles.
European regulators moved in parallel. EMA's PBPK reporting guidance took effect July 1, 2019, and a February 2025 concept paper outlined plans for broader mechanistic MIDD guidance covering physiologically based biopharmaceutics modeling (PBBM) and quantitative systems pharmacology (QSP). The EU AI Act, which entered into force August 1, 2024, classifies most medical software as high-risk, with obligations for data quality, human oversight, and robustness phasing in from August 2026 through 2027.
FDA finalized its Predetermined Change Control Plan guidance in December 2024, part of a multi-year AI/ML software as a medical device framework. Additional draft lifecycle guidance for AI-enabled devices followed in 2025, alongside transparency principles published in June 2024. The September 2025 rare disease evidence principles statement signaled regulatory openness to flexible evidentiary approaches—including synthetic controls and MIDD methods—where traditional trial designs prove impractical.
Industry-led standardization efforts are maturing in parallel. The 2024 publication "Toward Good Simulation Practice," led by the In Silico World consortium and Virtual Physiological Human Institute with FDA contributors, proposed community-endorsed best practices for computational modeling in regulatory processes. The Avicenna Alliance continues policy engagement on MDR and IVDR revisions to formalize in silico data acceptance across Europe.
The infrastructure is becoming real. Concrete. Enforceable.
The Data Backbone Arrives, Finally
Interoperability has been healthcare AI's Achilles' heel for years, but infrastructure is finally arriving. FHIR—Fast Healthcare Interoperability Resources—is now mandatory in U.S. federal programs via ONC certification and the TEFCA roadmap for nationwide health information exchange. Oracle Health became a designated Qualified Health Information Network (QHIN) in November 2025, joining a growing list of entities piloting QHIN-to-QHIN FHIR exchange.
Cloud FHIR stores with built-in clinical natural language processing and SQL-on-FHIR capabilities, such as AWS HealthLake, enable ingestion from EHRs, claims systems, imaging archives (DICOM), wearables, genomics platforms, and lab instruments. Data harmonized to the OMOP Common Data Model—widely adopted for real-world data analytics—can feed population-level twins and synthetic cohort generators. The UK's Clinical Practice Research Datalink deployed OMOP in 2024, and the OHDSI community continues refining version 5.4 while version 6.0 remains under development.
For mechanistic model repositories, the IUPS Physiome project maintains curated multi-scale models in CellML and FieldML formats. The NIH IMAG Multi-scale Modeling Consortium and JSim collections provide additional public libraries. Open Systems Pharmacology (PK-Sim/MoBi) has built an active regulator-industry community around open-source PBPK tools, with a 2025 conference summary published in Clinical Pharmacology & Therapeutics: Pharmacometrics & Systems Pharmacology.
Privacy-preserving collaboration is advancing through federated learning, which enables multi-institutional model training without centralizing protected health information. Systematic reviews in 2024 and 2025 documented federated learning applications in healthcare—a critical enabler for population-level twins that learn continuously from distributed data sources.
None of this sounds sexy. But without it, nothing else scales.
Who's Building the Stack
The infrastructure layer is fragmenting into specialized niches, as infrastructure layers tend to do.
Certara (Simcyp, now EMA-qualified) and Simulations Plus (GastroPlus, with FDA-funded projects on inhaled and long-acting injectable products) dominate PBPK and model-informed drug development services. Their platforms appear in over 120 FDA approvals combined. Eleven regulatory agencies worldwide license Simcyp.
Unlearn, Novadiscovery (Jinkō platform), and Aitia are carving out clinical trial augmentation and synthetic control territory. Each offers a different modeling philosophy: Unlearn's prognostic-score approach emphasizes unbiased integration into randomized controlled trials; Novadiscovery curates mechanistic disease models for virtual patient cohorts; Aitia layers causal AI onto multi-omics to identify responder subtypes.
Device planning and diagnostics have clearer commercialization paths. HeartFlow's FFRCT technology is reimbursed and integrated into cardiology workflows. FEops expanded into U.S. structural heart procedures. Boston Children's Hospital publishes peer-reviewed case studies on CFD-based congenital surgery planning, and ELEM Biotech is advancing whole-organ multiphysics twins through Horizon Europe's VITAL clinical studies.
Patient management digital twins remain nascent but promising. Twin Health's metabolic twin model, with its strong RCT data, is gaining traction among self-insured employers and health systems willing to contract on outcomes. It's a model that either scales dramatically or dies quietly; there's not much middle ground.
Policy and standards bodies—MDIC, the Avicenna Alliance, the Virtual Physiological Human Institute—serve as neutral conveners, ensuring that credibility frameworks like ASME V&V 40 and Good Simulation Practice gain cross-industry adoption.
Then there's a stealth-stage player called Mantis Biotech, which has appeared in early job listings describing "infrastructure validating human-in-computer models" and a "unified biomedical testing and regulatory platform" linking simulation and bench data to FDA Q-Subs. Public information is sparse—Wellfound lists the company as recently funded with one to ten employees—but the positioning suggests an attempt to productize credibility infrastructure itself. Whether that's visionary or vaporware remains to be seen.
What Happens Next
The National Academies published a 2024 consensus study on digital twins that defined the field, identified research gaps, and emphasized verification, validation, and uncertainty quantification needs. The report underscored what industry insiders already knew: moving from academic prototypes to regulatory-grade tools requires formalized model provenance, traceable datasets, and auditable workflows.
Expect ICH M15 to finalize and propagate consistent MIDD review templates globally. The EMA's forthcoming mechanistic models guideline will expand beyond PBPK to encompass PBBM and QSP. Acceptance of simulation-derived "digital evidence" in device and drug submissions will accelerate as platforms earn qualification opinions and Model Master Files proliferate.
Technically, the industry is shifting from single-organ models to multi-organ and whole-body twins. Dassault's expansion from heart to lungs, brain, and liver reflects this trajectory. Configurable virtual patient populations will train and validate AI systems—and run in silico clinical trials at scale.
Clinical trial sponsors will increasingly adopt prognostic score augmentation methods like PROCOVA to reduce sample sizes and timelines. Regulators will formalize transparent model validation and reporting norms. Hybrid real-world data-borrowing approaches will continue evolving to control bias. Pharma's move toward lab-in-the-loop, agentic laboratories, and foundation models for biology—supported by GPU and HPC backbones—will support the computational intensity required for synthetic cohorts and population twins.
TEFCA-enabled FHIR exchange across QHINs, combined with payer and provider APIs, will reduce what one venture capitalist recently called the "data wrangling tax." High-risk AI obligations under the EU AI Act will push vendors to embed MLOps with audit trails. FDA's AI lifecycle guidance will gain enforcement teeth, probably sooner than vendors would prefer.
The most critical near-term challenge isn't scientific—it's productizing credibility. Platforms must bake risk-based context-of-use mapping, traceable datasets with FHIR and OMOP lineage, and GxP-like Good Simulation Practice checklists into their core architecture. Auto-assembled "evidence objects" aligned with EMA PBPK and FDA reporting expectations, reusable across submissions via Model Master File constructs, will separate winners from vaporware.
The pharmaceutical industry spent decades perfecting chemical synthesis and biologics manufacturing. The plants, the clean rooms, the quality systems—all built to prove that what comes out of the reactor is safe and consistent.
The next infrastructure battle is computational: who can validate, at scale, that a model of a human being is trustworthy enough to guide a billion-dollar trial decision or a surgeon's scalpel. The companies solving that problem won't just participate in drug development. They'll redefine it.
