The digital twin revolution in healthcare is gaining momentum. San Francisco-based AI startup Unlearn has been working to deploy digital twins—virtual patients generated by artificial intelligence—as external comparators in clinical trials, with regulatory engagement from both the FDA and Europe's EMA. The company's approach represents a broader shift in how computational evidence is being integrated into drug development.
It was the kind of milestone that might have passed unnoticed outside a narrow circle of computational biologists and clinical trial specialists. But for those paying attention, developments like these signal something larger: after years of pilot studies, academic white papers, and cautious regulatory dances, the infrastructure for virtual patients is no longer theoretical. It's becoming operational.
And that shift—quiet, technical, decidedly unglamorous—is driving a wave of investment that dwarfs the breathless hype around consumer-facing generative AI. Market analysts at SNS Insider project the digital twin healthcare market will swell from $2.22 billion in 2023 to $69.67 billion by 2033, according to estimates published in February 2024. Whether those numbers hold up is anyone's guess. What's harder to dismiss is the underlying convergence: regulators are formalizing credibility standards, interoperability mandates are forcing clinical systems to open their data vaults, and foundation models are providing the computational scaffolding that makes physics-enhanced simulations tractable at enterprise scale.
The question isn't whether virtual patients will reshape drug development. It's whether the industry can build credible, reproducible infrastructure fast enough to justify the capital flooding in.
A Market Still Finding Its Shape
The digital twin market in healthcare is, to put it mildly, fragmented. Definitions remain slippery. Is a digital twin a mechanistic model? A clinical decision support algorithm? A sophisticated data integration platform that happens to spit out predictions?
Depending on who's counting—and what they're counting—valuations swing wildly. Mordor Intelligence pegged software platforms at roughly 55% of market share in January 2026, with drug discovery and preclinical applications claiming 26.6% of deployments. Fortune Business Insights valued the sector at $4.61 billion in 2025. Precedence Research came in with $6.93 billion for AI in drug discovery, projecting a more conservative 9.9% compound annual growth rate through 2035.
The variance isn't just methodological noise. It reflects genuine ambiguity about what's being sold and to whom. Patient-level analytics—using AI twins to predict how individuals might respond to treatments—is forecast to grow fastest through 2031. But enterprise buyers, for now, are concentrating their bets on operational efficiency: speeding up trial design, reducing patient recruitment timelines, streamlining regulatory submissions.
A Deloitte survey of biopharma and medtech executives, released in February 2026, found 75% expressing cautious-to-enthusiastic optimism about the sector. Forty-eight percent cited accelerated digital transformation as a major driver, up from the prior year. IQVIA's 2025 R&D trends report underscores the shift in starker terms: the company quietly reclassified $674 million of its real-world evidence offerings into its clinical research segment, a tacit acknowledgment that late-phase data products are now operationally inseparable from trial businesses.
Perhaps more telling than the forecasts is what's happening in the data layer. TEFCA—the Trusted Exchange Framework and Common Agreement, a nationwide health information exchange initiative—reported 12,130 participating organizations and 474 million documents exchanged as of June 2025, according to the Sequoia Project. The Office of the National Coordinator's HTI-1 Final Rule set a December 31, 2025 baseline for USCDI v3 adoption, with enforcement discretion stretching to March 1, 2026 for certain criteria.
These aren't abstractions. They're mandates forcing electronic health record systems to expose standardized APIs and semantic mappings—precisely the infrastructure needed to train and validate digital twins at scale. The plumbing, in other words, is finally getting installed.
When Regulators Stop Saying No
Two forces are colliding, one regulatory and one technical, and their intersection is creating a narrow window of opportunity.
Start with the regulators. The FDA's November 2023 final guidance on computational modeling and simulation for medical device submissions remains the anchor. It established risk-informed credibility standards—verification, validation, uncertainty quantification—that sponsors are now adapting for drug development contexts. The European Medicines Agency launched parallel efforts in 2025, issuing a concept paper (EMA/5875/2025) and convening a multi-stakeholder workshop in October to formalize reporting standards for mechanistic models, including physiologically-based pharmacokinetic (PBPK) and quantitative systems pharmacology (QSP) frameworks.
Draft ICH M15 guidance on model-informed drug development, posted by the FDA in 2025, provides general principles across modalities. CDER and CBER's MIDD paired meeting program—a structured pathway for sponsors to socialize computational evidence before pivotal submissions—has been extended through fiscal year 2027. In August 2025, Certara's Simcyp Simulator became the first software platform to receive an EMA Qualification Opinion for PBPK modeling, a milestone that validates regulator-endorsed infrastructure for mechanistic simulations.
None of this guarantees smooth sailing. Regulatory acceptance of AI patient-level twins for pivotal decisions will proceed cautiously, likely through pilot pathways in specific disease contexts—rare diseases, pediatric populations, conditions where external controls offer material ethical or logistical advantages. But the scaffolding is in place. Regulators have stopped saying "show us how this could work" and started saying "here's how we'll evaluate it."
On the technical side, foundation models are changing the economics. AlphaFold 3, released for non-commercial use in May 2024, has become a structural prior for in silico workflows. Isomorphic Labs—Google DeepMind's drug discovery spinout—raised $600 million on March 31, 2025, following partnerships with Novartis and Eli Lilly worth up to $3 billion. NVIDIA's BioNeMo platform is enabling GPU-accelerated training and inference across discovery workloads, with adoption announcements proliferating through 2025 and early 2026.
The data infrastructure, meanwhile, is standardizing around cloud lakehouses. Medidata detailed its modern lakehouse architecture on AWS in a November 26, 2025 blog post, describing how its Clinical Data Studio and Data Connect serve thousands of trials with lineage-preserved, queryable datasets. Databricks showcased an NHS federated data platform in June 2025, using Unity Catalog and Clean Rooms to enable de-identified access across institutions.
What these platforms provide isn't just storage. It's semantic harmonization. OMOP Common Data Model adoption is accelerating—the UK's Clinical Practice Research Datalink released its OMOP-mapped primary care dataset in March 2025—creating something like a lingua franca for observational research. HL7 FHIR US Core v6.x and USCDI v3 are setting floors for data semantics and API interoperability, reducing the friction of feeding heterogeneous sources into twin construction workflows.
It's infrastructure, all the way down. Unsexy. Essential.
Who's Building What
The commercial landscape spans established medtech giants, clinical-stage startups, and infrastructure bets that assume data plumbing must precede scale simulation.
Dassault Systèmes announced the next phase of its Living Heart Project on February 26, 2025, introducing AI-powered, customizable patient and population twins. The update positions the cardiac simulation framework—originally built on finite element methods—as a foundation for broader mechanistic modeling. Siemens Healthineers highlighted operational twins at RSNA 2025, including features for department optimization and interventional planning. These are less about individual patient modeling than systems-level decision support, but they share a common thread: physics-informed prediction.
HeartFlow's FFRCT service, while not always marketed as a "digital twin," represents a mature physics-based workflow. A Nature Medicine analysis published May 6, 2025 examined over 90,000 patients in NHS England and found that coronary CT angiography plus FFRCT was associated with improved outcomes and resource utilization compared to standard pathways. U.S. reimbursement improved in 2024, accelerating adoption.
Twin Health raised $53 million in August 2025 after publishing results from a randomized, Cleveland Clinic-led study in NEJM Catalyst. The company's metabolic digital twin—built on continuous glucose monitors, wearables, and predictive algorithms—demonstrated significant improvements in Type 2 diabetes outcomes and reduced medication reliance. It gives regulators and payers tangible evidence that personalized simulation can drive measurable clinical benefit, not just cost savings on paper.
Which brings us back to Unlearn. The company's work on AI patient-level twins represents perhaps the most direct test of this technology in regulatory decision-making. Unlearn's generative models create participant-level external comparators for clinical trials, with the company citing regulatory engagement with both the EMA and FDA. While these applications are still evolving—often starting with exploratory endpoints rather than pivotal approval mechanisms—the regulatory engagement is real, and it sets a precedent for how virtual controls might be integrated into trial designs for diseases where recruiting sufficient patient cohorts is logistically or ethically fraught.
Against this backdrop, early-stage infrastructure plays are emerging. Mantis Biotechnology, a Y Combinator Winter 2026 company, positions itself as "Databricks for Biomedical and Clinical Data." Founder Georgia Witchel, who previously founded Louiza Labs (focused on digital twins for surgical robotics), describes the startup as building "full-stack search across clinical trial systems" with domain-aware canonical datasets that preserve lineage and enable single-query science.
The pitch addresses a chronic pain point: according to Mantis's launch materials, 80% of trials face delays due to data quality issues, costing an average of $15 million per trial. Whether those figures are precise or illustrative hardly matters—anyone who's worked in clinical operations knows the problem is real. Mantis's stated focus—combining large language models with high-fidelity physics simulations to convert rare human behavior data into predictive models—reflects the broader shift toward physics-enhanced synthetic data generation.
It's an infrastructure-first approach that assumes you can't scale simulation without first fixing the plumbing. The company's three-person team is building in New York, and its positioning against established lakehouse providers will depend on depth of biomedical ontologies and integrated credibility tooling mapped to FDA and EMA standards. That's a tall order for a seed-stage startup. Then again, Databricks was once three people, too.
Other players occupy adjacent niches. Novadiscovery (rebranded as Nova In Silico with its Jinkō platform) focuses on mechanistic modeling for trial design. InSilico Trials, an EU-based company, has announced FDA collaborations and projects through 2025. Evinova, AstraZeneca's digital trial platform, continues its go-to-market buildout with AI-enabled portfolio and study design modules, partnering with contract research organizations to embed capabilities across sponsor workflows.
The Plumbing Problem No One Wants to Talk About
Beneath the aspirational vision of "virtual patients" lies a reality that's less glamorous and more obstinate: most pharmaceutical and biotech organizations still struggle with basic data integration.
Clinical trials generate data across electronic data capture systems, clinical trial management systems, central labs, imaging vendors, and increasingly, wearables and molecular profiling platforms. Each source has its own schema, versioning conventions, and quality control processes. Harmonizing these streams—let alone linking them to real-world evidence from electronic health records or claims databases—requires semantic mapping, lineage tracking, and governance frameworks that few organizations have productized at scale.
This is where the current wave of lakehouse adoption becomes relevant. Medidata's migration to AWS, detailed in its November 2025 case study, illustrates the shift from point solutions to unified platforms. The company now serves thousands of trials through a single lakehouse architecture, enabling cross-study queries and reusable data models. Databricks's NHS implementation, presented in June 2025, demonstrates federated access with privacy-preserving Clean Rooms—a model that could extend to multi-sponsor consortia or real-world data partnerships.
The regulatory mandates are forcing convergence, whether enterprises are ready or not. HTI-1's requirement for standardized APIs and decision-support transparency means that health IT systems must expose not just data but also the logic used to transform and analyze it. TEFCA's scaling of nationwide exchange creates the potential for longitudinal, cross-institutional datasets that can serve as training corpora for digital twins. OMOP's adoption in observational research provides a common data model that reduces the need for bespoke ETL pipelines.
But standardization alone isn't sufficient. Building credible digital twins—especially for regulatory submission—requires verification and validation workflows that most data platforms don't natively support. ASME's V&V 40-2018 standard, developed for medical device computational modeling, provides a template. The FDA's November 2023 guidance operationalizes it with risk-informed credibility assessments. The EMA's 2025 mechanistic modeling initiative extends similar principles to PBPK and systems pharmacology contexts.
Few vendors currently integrate credibility tooling with data infrastructure. This gap creates an opportunity for platforms that collapse ingestion, harmonization, lineage, and V&V into a single workflow—precisely the positioning that Mantis and similar infrastructure startups are targeting. Whether they can execute at enterprise scale before incumbent platforms build or acquire similar capabilities remains an open question.
The Next Two Years
The near-term trajectory is less about radical breakthroughs than systematic formalization. Regulatory clarity for mechanistic models will continue to improve as the EMA finalizes its guidance and ICH M15 advances through the harmonization process. Formal acceptance of AI patient-level twins for pivotal decisions will proceed cautiously, one disease context at a time.
Foundation models in biology will increasingly backstop simulation priors. AlphaFold-derived structural insights, combined with BioNeMo toolchains for generative molecular design, will enable physics-enhanced synthetic datasets that calibrate multi-scale models faster and with less reliance on large observational cohorts. This matters especially for low-resource settings—pediatric populations, rare genetic variants, diseases with limited natural history data—where traditional statistical approaches struggle.
The commercial landscape will likely consolidate around platform plays. Enterprises are moving away from point solutions toward governed lakehouses that unify R&D, clinical, and real-world data with lineage and audit trails as procurement criteria. Unity Catalog, Clean Rooms, OMOP and FHIR alignment, and integrated credibility frameworks will become table stakes for vendors selling into pharma IT. The companies that win will be those that can demonstrate not just technical capability but also regulatory credibility and ecosystem interoperability.
That said, validation gaps remain substantial. Academic research on LLM-driven "human digital twins" for behavioral and attitudinal simulation—published as recently as late 2025—highlights significant challenges in ensuring that generative models accurately represent human decision-making and physiological responses. Physics-informed architectures may help, but the field is still working through how to balance mechanistic rigor with the data efficiency that makes AI approaches attractive in the first place.
For founders and executives evaluating infrastructure investments, the calculus comes down to platform versus point solution. The regulatory tailwinds—HTI-1, TEFCA scaling, EMA MIDD initiatives—create demand for data plumbing that can support both current analytics needs and future digital twin workflows. Buyer organizations will prioritize interoperable, lineage-rich, governed data layers in 2026 and 2027, with an eye toward enabling credible simulations at scale once regulatory pathways mature.
The $70 billion question isn't whether digital twins will transform drug development. Most people close to the technology assume they will, eventually. The real question is whether the industry can build the infrastructure to make them credible, reproducible, and economically viable before the current investment cycle runs out of patience.
The regulatory scaffolding is in place. The technical components are maturing. What remains is execution. And that, as always, comes down to data—unglamorous, obstinate, essential data.
