Microphysiological Systems

Established A microphysiological system is a small volume of human cells held under flow, under mechanical load and in a defined geometry, so that it behaves less like a well plate and more like a tissue. The family covers organ chips (microfluidic devices, usually two channels separated by a porous membrane, lined with human cells and perfused), organoids (self-organising three-dimensional cultures grown from pluripotent stem cells or patient tissue), perfused spheroid and tissue-slice platforms, and multi-organ assemblies marketed as a body-on-a-chip. The shared claim is specific: these systems carry enough human physiology to answer one question at a time — will this compound injure a human liver, does this patient’s tumour respond to this drug — more accurately than a rat does.

Established This is a different enterprise from building organs for transplant, and the two are routinely confused. A transplantable organ must survive diffusion limits, immune rejection and a surgical anastomosis; an assay must only be predictive. The Lab-Grown Organs brief tracks the first problem, where a human liver needs on the order of 160 to 360 billion cells against 10^4 to 10^6 in an organoid. None of that arithmetic binds here. An organ chip needs milligrams of tissue, and its only performance metric is how often it is right.

Frontier The field’s position is that of a technology with one convincing validation dataset, a permissive statute, an agency roadmap, and no completed qualification. The FDA Modernization Act 2.0 (December 2022) removed the statutory requirement for animal testing; the FDA’s 2025 roadmap for reducing animal use in preclinical safety studies set out to make animal studies the exception rather than the default, starting with monoclonal antibodies. Neither event made any chip a qualified drug development tool, and no organ chip holds an OECD test guideline. The gap between permission and acceptance is where this subject lives.

Established The science ran ahead of the validation architecture, and the validation architecture is an institutional product nobody has funded. The most-cited predictive result in the field was produced and paid for by a vendor, retrospectively, on compounds whose human outcomes were already known. That is a legitimate first step and a poor last one. The experiment that would settle the matter — a pre-registered, blinded, multi-site prediction of human toxicity before the humans are dosed — needs no new hardware and has never been run.

Established A note on sourcing. This brief was commissioned in September 2026 from the Institute’s research base. Reading-list entries without links are cited from the bibliographic record rather than re-fetched, and claims are dated no later than early 2026 unless carried by a linked source.

Established The statute changed in December 2022 and changed less than the coverage suggested. The FDA Modernization Act 2.0, enacted inside the Consolidated Appropriations Act, 2023, struck the language requiring animal testing from the drug-approval provisions of the Federal Food, Drug, and Cosmetic Act and replaced it with a list of permitted nonclinical tests naming cell-based assays, organ chips, microphysiological systems, computer models and bioprinted tissue. It is permissive; it obliges the agency to accept nothing. A follow-on bill styled FDA Modernization Act 3.0, which would have directed a rewrite of the implementing regulations, had not been enacted as of early 2026. A sponsor who submits chip data in place of a repeat-dose animal study is therefore making a request, not exercising a right.

Frontier The 2025 FDA roadmap is the strongest demand signal the field has had, and it is a plan rather than a result. The agency’s published roadmap opens with monoclonal antibodies, proposes pilot arrangements in which sponsors may submit new-approach-methodology data in place of some animal work, and sets a multi-year horizon over which animal studies become the exception. It was announced alongside a parallel NIH commitment to prioritise human-based research technologies. The choice of antibodies is not arbitrary: they are usually tested in one relevant species, that species is usually a macaque, and macaques are the most expensive and most supply-constrained input in preclinical safety.

Frontier The liver-chip drug-induced-liver-injury study is the field’s best evidence and it is a vendor study. The 2022 performance assessment of a human Liver-Chip, published in Communications Medicine by authors from the chip’s manufacturer and its academic originators, tested twenty-seven small molecules assembled as matched pairs — one member known to cause human drug-induced liver injury despite clearing animal testing, the other a close analogue that did not. The chip identified 87 per cent of the toxic compounds and produced no false positives on the non-toxic members, a specificity of 100 per cent on that set. The paper’s economic model put the value of broad adoption at roughly $3 billion a year for the industry, almost all of it from avoided late-stage attrition rather than cheaper assays.

Established What that study is not: prospective, independent, or a test of idiosyncratic injury. The compounds were chosen because their human outcomes were known, which makes the design retrospective however carefully operators were blinded. The matched-pair construction is hostile to the chip’s sensitivity and favourable to its specificity, because the comparators are compounds that also failed to injure humans. And the injuries are predominantly intrinsic, dose-related hepatotoxicity. Idiosyncratic liver injury — roughly one exposure in ten thousand, months after first dose, through an adaptive immune mechanism — is invisible to a device containing no adaptive immune system.

Established The incumbent’s accuracy is worse than most readers assume, and this is the comparison that matters. The canonical industry survey of concordance between animal and human toxicity, assembled from around 150 compounds contributed by a dozen companies, reported that animal studies as a whole anticipated roughly 70 per cent of human toxicities, non-rodents alone about 63 per cent, and rodents alone about 43 per cent. Rodent-to-rodent concordance in two-year carcinogenicity bioassays has been reported near 57 per cent. A method right 87 per cent of the time on a curated panel is not obviously worse. Nobody has run the incumbent through the same panel, so the field’s central comparison is still made across studies rather than within one.

Frontier Cross-laboratory reproducibility is where the field is weakest and where the evidence is most specific. A reproducibility characterisation of a commercial liver microphysiological system by FDA investigators, published in Clinical and Translational Science in 2021, found that basic functional readouts — albumin and urea output, compound metabolism and accumulation — varied enough between experiments and operators that tight protocol standardisation was a precondition of decision-grade use rather than a refinement of it. The NIH tissue-chip testing centres exist for exactly this: independent laboratories whose job is to receive a developer’s platform and try to reproduce the developer’s result. That programme is the right institutional shape and it is small.

Established Where organoids predict, they predict for individual patients, not for compound libraries. The cleanest case is cystic fibrosis: rectal organoids from a patient swell in response to forskolin in proportion to CFTR function, the assay tracks that patient’s response to modulator drugs, and it has been used to support access decisions for patients whose genotypes are too rare to appear in a registration trial. In oncology, patient-derived organoids from metastatic gastrointestinal cancers reproduced patient responses to targeted agents with about 88 per cent positive predictive value and 100 per cent negative predictive value in the founding study — on a few dozen patients.

Frontier Where they do not predict is equally documented and less quoted. A follow-on colorectal study found organoids forecast response to irinotecan-based regimens and failed for combinations built on 5-fluorouracil and oxaliplatin — agents whose clinical behaviour involves stromal, vascular and systemic factors an organoid does not contain. The rule of thumb the evidence supports is that organoids predict cell-autonomous drug effects and miss everything mediated by the tissue around the cells.

Established Three structural deficits explain most failures and none is a manufacturing problem. First, metabolic competence: stem-cell-derived hepatocytes typically express the major cytochrome P450 enzymes one to two orders of magnitude below adult primary human hepatocytes, so a compound toxic only through a metabolite can be missed. Second, immunity: almost no platform contains a functioning adaptive immune compartment, which excludes the hypersensitivity mechanisms behind a large share of withdrawals. Third, transport: oxygen and nutrients penetrate roughly 200 to 400 µm from a perfused surface, and in static organoids more than half of cells sitting 150 to 600 µm deep become hypoxic or necrotic by day 60 — which is why perfusion is not a luxury feature but the thing that makes a chip tissue rather than a necrotic ball.

Established The sharpest fact in the brief: there is no qualified organ chip. International acceptance of an alternative method runs through the OECD test guideline programme, which requires documented within-laboratory and between-laboratory reproducibility, usually a ring trial across at least three laboratories, against a defined reference chemical set. The alternatives that completed that route are comparatively simple: reconstructed human epidermis for skin corrosion and irritation, tissue assays for eye irritation, and the 2021 guideline on defined approaches to skin sensitisation, the first formal combination of non-animal methods to replace an animal test outright. No organ chip and no organoid assay has an equivalent. In the United States, ICCVAM — the interagency committee formalised by statute in 2000 and supported by NICEATM — publishes a strategic roadmap and coordinates agencies; it does not itself qualify a device for a sponsor’s submission. The FDA’s route for tools fitting no existing category, the ISTAND pilot opened in 2020, qualifies on a stated context of use, and as of early 2026 no microphysiological system had completed it.

Frontier The money says the same thing twice. The founding public investment — roughly $70 million each from NIH and DARPA from 2012, with the FDA as partner — built the platforms that exist today. The preclinical services industry it is meant to displace turns over billions of dollars a year, dominated by a few contract research organisations. Organ-chip vendor revenue across the whole field is best estimated in the tens of millions annually: a rounding error in the same market, and one of the best-known chip companies cut staff in 2023 while holding the field’s flagship dataset. No counted set of investigational new drug applications in which chip data replaced an animal study appears anywhere in the published literature, and that absence is the most informative fact about where this field stands. The technology is not held back by scepticism about the biology but by a buyer who pays per assay and captures the value of avoided attrition years later, in a different budget.

Frontier What is the prospective predictive value of an organ chip? Every headline number here is retrospective. A prospective number requires registering the chip’s call before the human data exist, which is possible today for any compound entering first-in-human studies and has not been organised.

Frontier Does predictivity transfer with the platform, or does it live in the laboratory that built it? The reproducibility studies suggest a substantial operator component. If 87 per cent falls to 60 per cent in three naive laboratories, the technology is a specialist service; if it holds, it is a product.

Frontier Can a chip be given an immune system that behaves? Adding monocytes, Kupffer cells or peripheral blood mononuclear cells to a liver or lung chip is routine; getting antigen-specific adaptive responses without spontaneous activation is not. Idiosyncratic toxicity is also the largest category of what animals miss, so a genuine solution here would be the field’s first claim to superiority rather than parity.

Frontier Is in-vitro-to-in-vivo extrapolation solvable at chip scale? A chip reports a response at a medium concentration; a regulator needs a dose. Bridging them requires free concentration in the device, the tissue-to-medium ratio, and a physiologically based pharmacokinetic model. Free concentration is rarely measured, and in devices moulded from polydimethylsiloxane it is frequently not the nominal concentration at all.

Established Polydimethylsiloxane absorbs the molecules it is supposed to be testing. The elastomer that makes chips cheap, transparent and gas-permeable is a sink for small hydrophobic compounds, which partition into the walls and leave the cells exposed to an unknown fraction of the nominal dose. It is measured, compound-dependent, and worst for exactly the lipophilic small molecules that dominate toxicology panels. Thermoplastic and glass devices avoid it and are harder to prototype and to oxygenate. Any chip paper reporting a nominal concentration without a measured free concentration is reporting an uncertain exposure.

Established Cell sourcing sets the ceiling on what the device can detect. Primary human hepatocytes are metabolically competent, scarce, donor-variable and lose function within days; stem-cell-derived hepatocytes are unlimited, reproducible and fetal in enzyme expression. No source is simultaneously abundant, mature and consistent, which is why much of the practical debate in this field is an argument about cells wearing an argument about devices.

Frontier Throughput is two to three orders of magnitude below the incumbent screening format. A screening campaign runs hundreds of thousands of wells; a perfused chip experiment runs tens to hundreds of chips with hands-on operator time each. That confines chips structurally to late-stage triage of small compound sets — a defensible role, not the one the popular framing promises.

Established This subject waits on maturation biology it does not fund. The gap between a fetal-like stem-cell-derived hepatocyte and an adult one is a developmental problem being worked on for regenerative-medicine reasons, as the Whole Organ Regeneration and Lab-Grown Organs briefs describe. Chips are a downstream beneficiary of every advance in directed differentiation, and get none of the budget.

Frontier It waits on physiologically based pharmacokinetic modelling credible enough to carry a chip result across to a dose. The modelling community is mature; the coupling to microphysiological data is not standardised. Without it a chip produces a hazard statement where a regulator needs a risk statement.

Established It waits on a reference dataset of human outcomes. Predictivity can only be measured against a truth set, and the truth set is clinical: post-marketing injury reports, withdrawals, and the failure reasons behind terminated programmes, most of which are commercially confidential. The largest single improvement available is not a better chip but a better register of what actually went wrong in humans.

Frontier It does not wait on computing. Quantitative structure-activity models, read-across and machine-learned toxicity predictors are cheap and improving, and they are trained largely on animal outcomes, which caps them at the incumbent’s accuracy. Human mechanistic data are the thing that could raise that ceiling, which makes the two approaches complements for at least a decade.

Frontier The decisive experiment is a pre-registered, blinded, multi-laboratory prospective predictivity trial, and it requires no new hardware. Take twenty to forty compounds entering first-in-human studies, distribute them under code to at least three independent laboratories running the same liver-chip or multi-organ protocol, require each laboratory to deposit a public, time-stamped call — hepatotoxic or not, at what exposure — before any clinical data exist, and score the calls against what the humans do. That single study would produce the field’s first prospective sensitivity and specificity, its first between-laboratory reproducibility on the endpoint that matters, and the evidence base a qualification decision needs. Nobody has funded it.

Established The second experiment costs almost nothing: run the incumbent through the same panel. Every chip performance study should carry, alongside its own numbers, the rat and dog calls on the same compounds, scored the same way. The historical concordance literature suggests the incumbent would land somewhere in the forties to seventies depending on species, but that is a cross-study comparison, and cross-study comparisons are how fields fool themselves.

Frontier Third, and already running: the monoclonal-antibody pilot is a natural experiment with a publishable endpoint. If the FDA roadmap produces a cohort of antibody programmes whose safety packages were reduced or replaced by new-approach methods, comparing their clinical safety records against matched conventionally tested programmes is the real-world trial the field could not organise for itself. It requires the agency to publish the cohort.

Speculative Fourth: a patient-level randomised test of organoid-guided therapy. Cystic fibrosis assays and oncology patient-derived organoids both predict at the individual level in small series. Randomising patients to organoid-guided versus standard selection, with survival or lung function as the endpoint, would convert a correlation into a clinical claim. Precision Medicine records how easily assay-guided assignment produces impressive match rates and disappointing response rates; that lesson applies directly.

Established The device requirements are well understood and rarely met in one product. A decision-grade chip needs an optically clear, non-absorbing material; controlled, measurable flow at physiological shear; a tissue-to-medium ratio close enough to the body’s that concentrations mean something; gas control; in-line sensing of at least oxygen and pH; and enough parallelism to run a dose-response with replicates in one operator-day.

Frontier The material choice is a genuine engineering tie. Polydimethylsiloxane gives gas permeability, clarity and rapid prototyping, and absorbs lipophilic compounds. Injection-moulded thermoplastics such as cyclic olefin copolymer eliminate absorption and complicate oxygenation. Declaring the tie: for hydrophilic biologics the elastomer is fine and for small-molecule toxicology it is not, so the field needs two product lines rather than one winner.

Frontier Automation is the difference between a demonstration and a method. Robotic handling of coupled chips has been demonstrated at up to ten linked organ compartments with quantitative pharmacokinetic readouts, and the operator dependence found in reproducibility studies is precisely the variable automation removes. This is a capital-cost problem with a known solution — the most tractable class of problem in the brief.

Established Perfusion is doing physiological work, not plumbing. The diffusion ceiling that limits static culture is why a chip can hold denser, thicker and longer-lived tissue than a well. The vascularisation literature is explicit that no in-vitro method has produced a vessel network matching the in vivo standard, so engineered perfusion is not a stopgap on the way to grown vasculature; for assay purposes it is the answer.

Established The nearest neighbour is organ manufacturing, and the relationship is one-directional. Lab-Grown Organs supplies the cell biology, matrices and vascularisation science this field consumes; this field supplies revenue that keeps those laboratories funded, because toxicology buyers exist today and transplant buyers do not. The honest description of the organoid economy in 2026 is that it is a toxicology and disease-modelling economy wearing the language of transplantation.

Frontier Patient-specific assays connect this subject to Precision Medicine and inherit its scepticism. A functional assay on a patient’s own cells reports phenotype rather than genotype, which is precisely what genomics-led precision oncology was criticised for missing. The counter-lesson from that brief’s record of large basket trials is that the bottleneck is rarely the assay and usually the availability of a drug that works once the assay has spoken.

Established Regulatory acceptance is a document, not an opinion, and the document has a standard shape. An OECD test guideline requires a defined method, a reference chemical set, demonstrated within-laboratory repeatability and between-laboratory reproducibility, and a predictive-capacity analysis. Producing that package for one endpoint costs a multi-laboratory ring trial and several years. It is not technically hard. It is nobody’s job.

Frontier The missing institution is a public validation utility with a blinded compound bank. The closest existing structures are ICCVAM and NICEATM, the European reference laboratory function, and the NIH tissue-chip testing centres. None holds a coded library of compounds with known human outcomes, runs platforms against it under contract, and publishes the scores regardless of vendor. Every mature measurement field has such a body; toxicology, for the methods that matter most, does not.

Established Context-of-use qualification is the right mechanism and it fragments the evidence. The FDA’s drug-development-tool pathways, including ISTAND, qualify a method for a stated purpose, so a liver chip qualified for hepatotoxicity screening of small molecules is qualified for nothing else. That is scientifically correct and commercially punishing: each context needs its own dossier, and no vendor can amortise one across a product line.

Frontier Liability is unallocated and this is a live disincentive. If an animal study misses a toxicity, the sponsor followed the standard of care; if a chip misses one, the sponsor chose an unconventional method. Until a regulator states that accepted new-approach data constitute the standard of care for a stated context, every rational sponsor runs the animal study as well — converting the technology from a replacement into an added cost. This is the clearest explanation for slow adoption, and it is institutional rather than scientific.

Frontier The stronger ethical argument is about humans. If a method that better predicts human toxicity exists and is not used, the cost falls on first-in-human volunteers and early-phase patients. The severe hepatotoxicity, clinical holds and one death reported in a large in-vivo gene-editing programme are a reminder that liver injury still emerges in humans after a conventional preclinical package. Whether a chip would have caught that case is unknowable retrospectively, which is exactly why prospective registration of predictions matters.

Frontier Representation in cell panels will become a health-equity issue exactly as it did in genomics. The evidence that polygenic risk scores built on European-ancestry cohorts transfer poorly and can widen disparities is the template. A toxicology panel built on a handful of convenient donor lines encodes the same bias into safety decisions, and nobody currently reports donor ancestry as a variable in chip performance.

Frontier Neural organoids attract public concern the rest of the field does not. A 2026 survey of 1,238 nationally representative United States respondents found average attribution of consciousness to embodied brain organoids at 2.83 on a five-point scale, with 30.7 per cent agreeing such systems could be conscious. That is a governance variable whether or not the attribution is warranted, and a liver chip is likely to be regulated in the shadow of a brain organoid.

Frontier The prize is attrition, and attrition is where drug development spends its money. Most compounds entering clinical development fail, and a meaningful share fail for safety reasons that appear after substantial investment. A method that moves even a fraction of those failures earlier compounds across a portfolio, which is how a device costing a few hundred dollars reaches an economic model measured in billions a year.

Speculative If predictive toxicology gets cheap, the class of diseases worth treating expands. Rare-disease programmes fail on fixed costs as much as on science; a safety package costing six figures instead of eight changes which indications a sponsor can rationally pursue. This is the strongest civilizational argument for the technology and it is entirely contingent on qualification, because an unqualified method is an additional cost rather than a substitute one.

Handwave The claim that microphysiological systems will make animal research obsolete belongs here as a claim, not a forecast. It requires solving immunity, whole-body integration, behaviour and reproduction in vitro — four problems on no funded roadmap — and it is asserted most confidently by parties whose advocacy or commercial interest depends on it.

These horizons track results that would change what can honestly be written about this subject, not predictions of market adoption.

• 10 yr: Frontier The first context-of-use qualification of a microphysiological system by a major regulator is the defining near-term event; its absence a decade from now would be a stronger verdict than any vendor dataset. Frontier The monoclonal-antibody pilot either produces a published cohort whose safety record can be compared with conventionally tested programmes, or quietly does not. Speculative A prospective, blinded, multi-laboratory predictivity trial is cheap enough to run inside this window and would settle the central question; it has no sponsor. Established Routine internal use of chips in discovery continues regardless, because it needs nobody’s permission.
• 25 yr: Speculative Non-animal methods become the default first-line safety package for at least one product class, most plausibly antibodies or chemicals rather than small molecules. Speculative Immunocompetent chips either demonstrate detection of immune-mediated toxicity or the field formally concedes that endpoint to animals and to pharmacovigilance. Speculative Patient-derived organoid assays are reimbursed for at least one indication on randomised evidence, or they remain a research service.
• 50 yr: Speculative Human-cell-based assessment of the industrial chemical inventory becomes technically routine, which would be the largest public-health consequence of this technology and the one least discussed. Handwave Whole-body in vitro pharmacology accurate enough to substitute for a first-in-human dose escalation is asserted by advocates on this horizon and rests on integration problems nobody is funding.
• 100 / 250+ yr: Handwave Claims that all animal experimentation ends, or that simulation absorbs the whole enterprise, are statements about values and computing power rather than about any measured trend in this field.

• Depends on cell biology produced by neighbours for other reasons: the differentiation and maturation work tracked in Lab-Grown Organs, whose diffusion and reproducibility numbers this brief uses directly, and the functional-assay logic developed in Precision Medicine, whose basket-trial record is the best available warning about assay-guided decision-making. It does not depend on the vascularisation breakthrough those briefs await, because engineered perfusion already solves transport at assay scale.
• Requires (not on this map) first, a regulator stating in a public decision that a named method is acceptable for a named purpose, which converts an unconventional choice into the standard of care and reallocates liability. Second, a prospective trial in which independent laboratories register predictions before human data exist — the missing measurement on which every performance claim in this field rests. Third, a consumable manufactured to lot-release specification in a material that does not absorb the test article, which is an injection-moulding and surface-chemistry problem rather than a biological one. Fourth, a supply of metabolically adult human hepatocytes and matched immune cells at certificate-of-analysis quality, which no current source provides. Fifth, a purchaser whose budget captures avoided late-stage failure, since the value of the method appears years later and in a different cost centre from the one that buys the chip.
• Enables cheaper safety packages for rare-disease and small-market drugs; human-cell-based assessment of chemicals that will never see an animal study; safety testing for human cell and gene therapies that animals cannot host; and patient-level functional assays where a genomic marker cannot decide.
• Adjacent to Whole Organ Regeneration, which shares the cell-sourcing bottleneck from the other side, and to Biological Computing, which uses the same long-term culture methods for an entirely different purpose and drags this field into its governance debates.

Established “The FDA has banned animal testing.” It has not, and the claim survives because the 2022 statute was widely misreported. The FDA Modernization Act 2.0 removed a requirement; it created no prohibition and no obligation to accept any particular alternative. The 2025 roadmap is a planning document with a multi-year horizon, beginning with one product class. As of early 2026, a sponsor omitting animal studies without agency agreement is taking a regulatory risk.

Frontier “Organ chips are 87 per cent accurate, so they beat animals.” The number is real and the inference is not yet earned. It comes from one vendor-authored study, on twenty-seven retrospectively selected compounds, in matched pairs, for intrinsic hepatotoxicity, in the laboratory that built the device. What would earn the inference is the same panel run through animals and through three naive laboratories, scored together. That study is cheap and absent.

Established “Animal testing is scientifically worthless anyway.” The concordance data do not support this, and overclaiming here damages the case for alternatives. Animal studies anticipate a majority of human toxicities in the industry survey record, are strongest for non-rodent species, and remain the only routinely used system with intact immunity, endocrine feedback, behaviour and reproduction. The accurate statement is that they are mediocre and expensive, not useless.

Established “Organoids are miniature organs.” They are fetal-like, avascular, immune-free tissue fragments, and their own literature says so. Reviewers state that organoids remain fetal-like after prolonged culture, that vascularisation has never matched the in vivo standard in vitro, and that reproducibility is a central problem. The naming convention did the field a marketing favour and a scientific disservice.

Frontier “A body-on-a-chip models whole-organism pharmacology.” Coupled systems have reproduced the pharmacokinetics of specific drugs; that is a narrower claim. Robotic assemblies of up to ten linked organ chips produced quantitative pharmacokinetic predictions for individual compounds. Scaling rules between compartments, immune trafficking and neuroendocrine feedback are unsolved, and no coupled system has been shown to predict a toxicity its component chips missed.

Handwave “Regulatory acceptance is just a matter of time.” The abandoned 2035 deadline for eliminating mammal studies at the Environmental Protection Agency is the counterexample. An announced phase-out without a validated replacement behind it can be relaxed as easily as it was announced. Time does not produce ring trials, reference compound banks or context-of-use dossiers; budgets do, and no budget line currently exists for them.