The breakthrough cardiac simulation models can predict arrhythmia circuits with 80% accuracy. European drug regulators have begun issuing qualification opinions for computational modeling methodologies in drug development. And yet Georgia Witchel, who runs a Y Combinator-backed startup called Mantis Biotechnology, insists the real problem strangling computational medicine has nothing to do with physics or algorithms.
It's plumbing.
More precisely: fragmented datasets locked in incompatible systems, stripped clean of the biological context that makes them useful, impossible to trace back to their origins. Witchel's bet—announced when Mantis emerged from stealth in February 2026—is that fixing this unglamorous infrastructure problem matters more than building ever-more-sophisticated models. Her pitch sounds almost mundane in an industry obsessed with AI breakthroughs: a "domain-aware data platform" that unifies electronic data capture systems, clinical trial management software, lab results, genomics pipelines, and manufacturing logs into clean, canonical datasets ready for analytics, machine learning, and regulatory submissions.
Mundane, perhaps. But if she's right, Mantis is positioning itself at a chokepoint—the infrastructure layer that could determine which digital twin companies actually scale and which remain expensive research projects.
The timing isn't accidental.
When Market Projections Scatter, Watch the Direction
The numbers from market research firms tell a story of inflection, though they scatter across a surprisingly wide range. SNS Insider pegged the global digital twins in healthcare market at $2.22 billion in 2025, projecting $69.67 billion by 2035—a compound annual growth rate north of 41%. Research and Markets took a more conservative stance in January 2026: $1.76 billion in 2025, growing to just $2.02 billion in 2026. Grand View Research, working from a 2024 baseline of $902.6 million, saw the market hitting $3.55 billion by 2030.
The variance matters less than the direction everyone agrees on. Every projection shows double-digit growth, driven by a consistent set of forces: regulatory agencies opening doors for computational evidence, payers starting to reimburse AI-driven precision tools, and pharmaceutical sponsors desperate—truly desperate—to shorten trial timelines and cut costs.
And the established players? They're moving. Fast.
Dassault Systèmes, whose Living Heart Project became something of a poster child for computational cardiology, unveiled what it calls "AI companions" and an "orchestrated intelligence" vision in March 2026, explicitly targeting the fusion of virtual twins with clinical data infrastructure. Certara's Simcyp PBPK Simulator secured an EMA qualification opinion in August 2025—the first and only software platform to achieve that status for physiologically-based pharmacokinetic modeling in specified contexts. That qualification isn't just a badge. It means drug sponsors can cite Simcyp outputs in EU submissions without running additional validation studies in certain scenarios. Compliance infrastructure, in other words, not just nice-to-have software.
HeartFlow saw its reimbursement landscape shift when CMS raised payment rates effective January 2025—FFR-CT analysis reimbursed at roughly $1,017, AI-QPA at around $950. Not trivial sums. FEops, which builds digital twins to plan transcatheter aortic valve implantations, published real-world validation data in the European Heart Journal Digital Health in December 2025 and signed a commercialization deal with TeraRecon the same year.
Clinical-grade applications are stacking up in ways that feel less like research novelties and more like the beginnings of a real market. Johns Hopkins researchers won a 2025 award for cardiac digital twin work predicting ablation targets in atrial fibrillation and ventricular tachycardia. Twin Health, operating what it calls a whole-body metabolic digital twin for Type 2 diabetes and obesity, published outcomes in NEJM Catalyst showing medication reduction and improved glycemic control; the Validation Institute certified those findings for payers. Unlearn, which generates synthetic control arms to shrink trial sizes, applied its Alzheimer's disease twin—trained on more than 26,000 participants across major research cohorts—in partnership with remynd throughout 2025.
The adjacent in silico clinical trials market is overlapping and expanding. Mordor Intelligence noted in January 2026 that pharmaceutical companies hold roughly 60.6% of that market, with the medical device segment growing. Multiple vendor estimates place the 2025 global in silico trial market around $3.8 billion.
Three Forces, One Moment
What makes this moment distinct—what separates it from earlier waves of hype around computational medicine—is the convergence of three forces that have finally started moving in sync.
The regulatory tailwinds have shifted from tolerance to something closer to encouragement. The FDA published draft frameworks in late 2024 to advance the credibility of AI models in drug and biological product submissions, noting that more than 500 submissions have included AI components since 2016. The agency's 2016 guidance on computational modeling in device submissions remains foundational, but a 2024 webinar on credibility assessment signaled evolving standards aligned with ASME V&V 40—the risk-based framework for computational model verification and validation. The FDA's Model-Informed Drug Development Paired Meeting Program, running through fiscal year 2027 under commitments made in the user fee agreement, gives sponsors a formal channel to negotiate what computational evidence will actually be accepted.
Europe is running a parallel track, perhaps moving even faster. The EMA held a workshop in late 2025 on external controls in evidence generation, exploring when digital twins or real-world data arms can substitute for traditional control groups. In March 2025, the agency issued a qualification opinion on an AI-based development methodology—a signal of receptivity. The European Health Data Space Regulation, adopted January 21, 2025, creates legal frameworks for secondary use of health data across the EU, smoothing access to the training and validation datasets digital twins require. The EU AI Act, which entered force in August 2024, sets staged compliance requirements for high-risk AI in medical devices, with major obligations kicking in around 2027.
Reimbursement decisions are following regulatory precedent, which is how these things usually work. The HeartFlow payment update wasn't just about one company—it signaled CMS willingness to pay for computationally derived diagnostics at scale. A green light, in other words, for an entire category. When FEops published validation data showing their TAVI planning twin improved procedural outcomes, it wasn't just academic publishing; it was groundwork for coverage policy.
Technical maturity is catching up to the ambition that's been floating around for years. Physics-informed neural networks, operator learning frameworks, emulators for computational fluid dynamics and finite element analysis—these approaches are collapsing the time required to calibrate patient-specific models. Research published in 2025 and early 2026 demonstrates pipelines that take ECG signals and generate 4D cardiac digital twins, graph neural networks that accelerate arrhythmia simulations, ensemble Kalman filters that tune electrophysiology models in minutes rather than hours.
NVIDIA's Modulus and PhysicsNeMo frameworks, adopted across industries for digital twin development, are being repurposed for bio-physiological emulation. UK Biobank is scaling toward 100,000 imaging scans, with projects in 2025 explicitly focused on "AI-empowered digital twining of the heart" using MRI data from roughly 55,000 subjects. That scale enables something different: population-level twins, generative models that capture biological variability across demographics, not just individual patient reconstructions.
And—perhaps most important for Mantis' thesis—the data infrastructure is finally standardizing. HL7 FHIR became mandatory in U.S. certified health IT as of January 1, 2026, under the ONC HTI-1 final rule. (Enforcement discretion runs to March 1, 2026, because these things are never as clean as regulatory timelines suggest.) FHIR US Core Implementation Guides tied to USCDI v3 set baseline interoperability expectations. A 2025 HL7 ballot advanced a FHIR-to-OMOP Implementation Guide, bridging clinical data exchange with the OMOP common data model favored by observational researchers; OHDSI reported progress in December 2025 on pipelines that unify real-world data into model training and validation workflows.
TEFCA Qualified Health Information Networks are expanding—Oracle Health secured QHIN designation in November 2025—creating what could become national-scale data highways feeding the multimodal inputs digital twins demand: EHR notes, imaging studies, lab results, wearable sensor streams, genomic panels.
Enterprise demand is shifting from curiosity to procurement. Pharma sponsors aren't asking whether computational models can reduce trial costs anymore. They're asking which vendors' platforms will actually pass regulatory scrutiny. Device companies are embedding simulation into verification and validation workflows because FDA reviewers increasingly expect to see it. Hospitals and health systems are evaluating digital twin platforms not as research tools but as clinical decision support infrastructure they might actually deploy.
A November 2025 scoping review in npj Digital Medicine mapped the landscape of human digital twin applications, clarifying taxonomy (digital model versus digital shadow versus true twin with bidirectional feedback) and surveying clinical exemplars. NIH's National Heart, Lung, and Blood Institute ran a September 2025 workshop on digital twins, explicitly addressing infrastructure needs—data pipelines, compute resources, workflow standards. EDITH, a European consortium, published a Virtual Human Twin Roadmap in October 2025 outlining policy and technical requirements for integrated multiscale, multi-organ models in European health systems.
The Data Quality Thesis
Mantis Biotechnology enters this landscape with a specific argument: the digital twin economy will be gated by data quality, not model sophistication.
It's a contrarian bet, in a sense. Most of the attention—and most of the venture capital—flows to companies building better models, more accurate simulations, faster solvers. Witchel is betting that those efforts hit a wall without clean inputs. Her platform claims to unify electronic data capture, clinical trial management systems, lab results, omics pipelines, and manufacturing logs into canonical datasets with "full lineage to raw source systems," HIPAA-compliant and ready for regulatory reporting.
On its recruiting page, Mantis describes itself as "infrastructure validating human-in-computer models," with applications in professional sports and medical device development. It pitches a "unified biomedical testing & regulatory platform linking CAD→simulation→bench testing→verification→FDA Q-Subs," calling itself "Palantir for the biomedical industry." The pitch, distilled: fragmented data—stripped of context, impossible to trace, semantically inconsistent—is the binding constraint on computational medicine. Fix the plumbing, and the models will follow.
Witchel's background crosses several worlds: elite athletics (her Y Combinator profile mentions connections to performance testing), computer science at Harvey Mudd, previous YC founder experience with a venture called Louiza Labs, and bioengineering graduate work. That combination suggests awareness of both wet-lab complexity and software infrastructure demands—the gap between bench science, clinical operations, and regulatory compliance.
The company participated in Y Combinator's Winter 2026 batch and currently lists approximately three employees. YC congratulated Witchel on the launch via LinkedIn in late February 2026. The company's funding status remains unclear; while references to a seed round have circulated in secondary sources, no verified announcement from the company or major financial press has confirmed details as of mid-March 2026.
Mantis is betting on a moment when data lineage and semantic encoding become competitive differentiators rather than technical implementation details. Regulatory guidance increasingly emphasizes context-of-use credibility and traceability—ASME V&V 40's risk-based framework, FDA's evolving credibility messaging, EMA's qualification pathways. Platforms that embed audit trails, encode biological meaning, and output submission-ready documentation might find themselves becoming infrastructure rather than just tooling.
Other infrastructure plays are emerging in parallel. InSilicoTrials joined Microsoft's Pegasus program in 2025, positioning itself around "digital regulatory-grade evidence" generation. FHIR-to-OMOP bridge implementations are maturing. Oracle Health's TEFCA QHIN designation expands data corridors. These are pieces of the same puzzle: making multimodal biomedical data computationally legible, traceable, and compliant.
Validation in the Wild
The device digital twin ecosystem offers clearer validation loops than drug development, probably because the physics is more tractable and the regulatory pathways more established. FEops' December 2025 study compared its TAVI planning twin's predictions against real-world procedural outcomes, published the results in a European cardiology journal, and scaled commercialization with TeraRecon. HeartFlow's FFR-CT workflow, now reimbursed at higher rates, demonstrates payer willingness to fund computationally derived diagnostics when clinical evidence is robust enough.
Drug development moves slower, but the applications are stacking up. Unlearn's digital twin generators reduce placebo arm enrollment requirements by creating synthetic controls matched to real patients. The company's Alzheimer's twin, trained on datasets spanning ADNI, CPAD, NACC, and EPAD, was applied with remynd in 2025—a sign of growing sponsor comfort with external control methodologies. Certara's Simcyp qualification by the EMA in August 2025 set a precedent that matters: certain PBPK modeling contexts no longer require additional validation in European submissions, making the software a de facto regulatory tool.
Whole-body metabolic twins are hitting clinical adoption, though the business models remain somewhat unclear. Twin Health's platform, validated in an August 2025 NEJM Catalyst study led by Cleveland Clinic researchers, showed outcomes improvements and medication reductions in Type 2 diabetes and obesity. The Validation Institute certified the findings for payers in 2025, signaling that digital twin interventions are moving from pilot studies to coverage consideration. Whether they'll achieve broad adoption is another question.
The Hard Questions Ahead
The next three years will separate infrastructure providers from research projects that raised venture rounds. Several questions will sort winners from noise, and the answers aren't obvious yet.
Who owns credibility? Regulatory bodies are formalizing what "credible" actually means for computational models—context-of-use specificity, verification and validation rigor, traceability from raw inputs through intermediate transformations to final predictions. Platforms that embed these requirements as native data structures, not as post-hoc reporting exercises, will have an advantage. The ASME V&V 40 framework, FDA's credibility guidance for AI in drugs and biologics, and EMA's evolving standards for external controls are converging on similar principles: computational evidence must be as auditable as bench data.
Mantis' emphasis on lineage and semantic encoding aligns with these trends. But execution is everything in enterprise software. If the platform can demonstrate that its data pipelines meet FDA computational modeling reporting requirements or EMA's external control evidence standards without manual rework, it becomes infrastructure. If it's just another ETL tool with biomedical branding, it's noise, and there's plenty of noise in this space already.
Can data rails standardize fast enough? FHIR mandates, TEFCA QHIN expansion, FHIR-to-OMOP Implementation Guides, the European Health Data Space Regulation—they're all moving toward interoperable, traceable, semantically rich health data. But enforcement is phased, adoption is uneven, and legacy systems remain deeply entrenched in ways that make CIOs weep. The ONC HTI-1 rule set a January 1, 2026 compliance date with enforcement discretion to March 1; real-world adoption will lag regulatory timelines by months if not years.
Infrastructure providers that can bridge legacy formats—HL7 v2, C-CDA, proprietary feeds that are genuinely baroque—with emerging standards will capture the transition period. Platforms that require clean FHIR-only inputs will struggle until data providers catch up, which could take a while. Mantis' positioning as a unifier of "EDC, CTMS, labs, omics, manufacturing" suggests awareness of this heterogeneity. Whether that translates into scalable product or endless custom integration work for each customer will determine their trajectory.
What does "domain-aware" actually mean in practice? The hardest challenge in biomedical data infrastructure isn't storage or compute—cloud providers solved those problems. It's encoding the biological semantics that make data useful for models. A glucose measurement isn't just a number; it carries context about timing relative to meals, medication dosing, sensor calibration, patient fasting state. An imaging study isn't just pixels; it encodes anatomy, pathology, acquisition protocol, contrast agent presence.
If "domain-aware" means the platform understands these semantics and preserves them through transformations, that's genuinely valuable and potentially defensible. If it means the platform stores metadata fields that users must manually curate, it's less differentiated from existing solutions. The test will be whether Mantis can deliver datasets that feed straight into model pipelines—physics-informed neural networks, digital twin emulators, statistical learning models—without requiring data scientists to write custom cleaning scripts for each project.
Which use cases scale first? The digital twin market isn't monolithic, despite how market research firms aggregate it. Device companies need simulation-to-submission workflows linking CAD models, finite element analysis outputs, bench testing, and regulatory documentation. Pharma sponsors need digital control arms and PBPK models to reduce trial size and duration. Hospitals need clinical decision support twins for procedure planning and patient stratification.
Adjacent problems, maybe. But distinct buyers, budgets, procurement processes, and success metrics.
Mantis' positioning touches all three verticals—validating human-in-computer models for professional sports and medical devices, unifying trial and manufacturing data, supporting analytics and AI. That breadth could be strategic diversification. Or it could be diffused focus, which is a common failure mode for early-stage infrastructure companies trying to serve every constituency at once. Mantis will need to pick a wedge, probably sooner than they'd like.
Device companies might be the clearest beachhead. Regulatory pathways for computational modeling in devices are more mature than in drugs—FDA's 2016 guidance on computational modeling submissions, the 2024 credibility assessment webinar, industry adoption of ASME V&V 40. A defined market exists. If Mantis can deliver a CAD-to-Q-Sub workflow that satisfies FDA reviewers in the Center for Devices and Radiological Health, that's a product with clear value. The question—and it's not a small one—is whether a three-person startup can execute integration and support at the scale device companies require. Enterprise sales cycles are brutal.
What happens when everyone has data infrastructure? If Witchel's thesis is correct—that data quality, not model sophistication, is the binding constraint—then data platform infrastructure will become table stakes eventually. Dassault Systèmes, with its 3DEXPERIENCE platform and Medidata clinical data assets, is already building toward unified virtual twin and trial data stacks. Oracle's TEFCA QHIN designation positions it to control data pipelines. Palantir itself, which Mantis cites as a comparison point, has signaled interest in healthcare and life sciences applications.
Early movers in infrastructure often win not by building the best technology but by establishing themselves as the integration layer before incumbents notice the opportunity. Mantis has a narrow window—18 months, maybe 24 at the outside—to become embedded in enough customer workflows that switching costs become a moat. After that, larger vendors with established sales teams, support organizations, and regulatory expertise will replicate the core functionality. They always do.
The Plumbing Problem
The convergence is real: regulatory acceptance, technical maturity, enterprise demand. The digital twin market will reach tens of billions by the mid-2030s, though the exact figures remain contested and probably depend on how you define the category. Infrastructure providers that solve data quality, traceability, and semantic encoding will capture a slice of that growth—how large a slice depends on execution and timing.
Whether Mant
