Evidence checked October 31, 2026. Regulatory announcements are distinguished from completed validation. Numerical examples below are hypothetical.
1. A change in permission is not a result in a patient
On September 21, 2026, the US Food and Drug Administration announced a direct final rule updating terminology in its regulations to recognize non-animal methods where appropriate in nonclinical testing. The announcement described a change in the available routes to evidence, not a declaration that any particular organoid could establish a drug's safety. That distinction matters whenever a photograph of living human tissue in a dish appears beside a prediction that drug development is about to become faster or more reliable. The FDA announcement creates an immediate reason to scrutinize those predictions without treating regulatory flexibility as experimental proof.
Organoids are compelling partly because their cells can be human and their organization can preserve features missing from a flat layer of cells. But a laboratory model does not become a small patient merely by acquiring three dimensions. A patient has interacting tissues, circulating substances, histories of exposure, a changing immune system, and a life outside the assay. A model contains a deliberately selected subset of those features. Its scientific usefulness often comes from that selection: it makes a specific question observable and manipulable.
The difficult question is therefore not whether an organoid looks impressively biological. It is whether the features retained are the features needed for a particular inference. A system useful for detecting one mechanism of injury may be poor at predicting another. A disease model that reveals how a mutation changes a cell may not rank treatments correctly. Calling both systems human-relevant leaves the central work unfinished. Relevance has to be attached to an outcome, an exposure, and a population.
2. The smallest useful claim
Consider three hypothetical claims about the same intestinal model. The first says a compound changes a measured cellular pathway. The second says the compound damages an intestinal barrier at clinically relevant exposure. The third says administering the compound to patients will produce unacceptable gastrointestinal harm. These statements require progressively larger bridges between experiment and decision. The first may be addressed within the model. The second needs evidence about the barrier measurement and exposure. The third also requires a connection to the people, treatment schedule, competing processes, and severity threshold involved.
A strong development program can begin with the smallest claim rather than advertising the largest. It might establish that the model detects a defined class of barrier injuries under specified conditions. That is a valuable capability even if the system cannot summarize every gastrointestinal risk. A narrower claim can make a model more useful because another researcher knows exactly where it belongs in an evidence package.
The FDA's April 10, 2025 announcement encouraged new approach methodologies, including organoids and computational approaches, in drug development. Its promised benefits were prospective policy objectives; they were not a randomized comparison showing that every such method improves development outcomes. Keeping that distinction visible preserves the significance of the 2025 roadmap announcement without converting its aspirations into measured results.
This is also how to read a successful demonstration. Ask what was classified, compared, or explained; then ask what was not. A model can achieve an important advance without supporting the strongest interpretation available to a press release. The distance between those interpretations is where validation work belongs.
3. An organoid, a chip, and a question
Organoid and organ-on-a-chip are sometimes used as if they name one technology. For practical evaluation, the distinction should be made explicit. A three-dimensional cellular structure and an engineered platform that controls aspects of a tissue environment may be combined, but their contributions are different. The cells contribute a biological response; the platform contributes conditions under which that response is observed. Neither label alone tells the reader which cell types, geometries, measurements, or interactions are present.
A kidney model, for example, might be designed around one transport function rather than the entire kidney. A brain model might reproduce a developmental process rather than an adult brain's organization. An intestinal model might preserve an epithelial feature while excluding important surrounding components. These are illustrative possibilities, not a catalog of defects shared by every system. Their significance depends on the intended question.
An NCATS fact sheet published in fall 2023 described several distinct uses of three-dimensional tissue models, including work on kidney cyst formation and models supporting movement toward clinical trials. The agency's examples show why a single success category is too crude: discovering a mechanism, prioritizing a candidate, and informing a trial are different achievements.
The reporting consequence is straightforward. Name the model's biological components and its job before celebrating its resemblance to an organ. A reader should be able to finish the sentence: this system represents these features well enough to answer this question under these conditions. If that sentence cannot yet be written, the work may still be promising discovery research. It should not be described as a validated substitute for a much broader biological system.
4. A beautiful image can be a weak endpoint
Images make tissue models understandable to people who do not work with them. They also create a temptation to equate visual similarity with functional adequacy. A structure can have an appealing shape without performing the specific operation relevant to a drug's effect. Conversely, a less visually elaborate preparation might measure the needed function with greater consistency. A picture answers a question about appearance unless additional evidence connects that appearance to something else.
Imagine a hypothetical model intended to detect impaired transport across a barrier. Investigators show that its cells form a continuous-looking surface. That is useful characterization, but it does not establish how substances cross the surface, how the model responds to a known disruption, or whether its baseline properties remain stable during exposure. The inference requires measurements tied to the proposed mechanism, not a collection of representative images chosen after the experiment.
There is a related problem in selecting the image itself. A single striking organoid may conceal a broad distribution of less organized structures. If the image is intended to illustrate the preparation, the publication should explain how representative it is. If it is an exceptional specimen used to demonstrate a possibility, it should be labeled accordingly. Neither use is inherently illegitimate; confusing them is.
For editorial purposes, the most informative figure might therefore be a distribution rather than a microscope photograph: the range of a functional measurement across independently generated batches, with failed preparations included. That figure is less immediately dramatic, but it reveals whether the system is becoming a dependable instrument. A model that works predictably can be scientifically more consequential than one that occasionally produces a spectacular structure.
5. The donor is part of the experiment
Human-derived does not mean representative of all humans. A model begins with particular material from a particular source. The donor's biological background and the history of the sample may matter for the outcome being studied. A result from a single source establishes something about that preparation under those conditions; expanding the claim requires additional evidence, not simply more wells containing descendants of the same cells.
The ISSCR's standards for stem cell-based models emphasize documenting starting material, considering donor characteristics, establishing quality criteria, and comparing models with relevant tissue features. They also distinguish technical and biological variability. These model-system recommendations provide a compact foundation for a broader analytical point: the source of the cells belongs in the explanation of the result, not merely in a methods appendix that readers can safely ignore.
Suppose a hypothetical study compares organoids from two affected donors with those from two unaffected donors. Each donor contributes fifty structures. There are two hundred structures, but only four donor backgrounds. Treating every structure as an independent representative of a human population would exaggerate the information available about differences between people. Repeated structures are useful for measuring repeatability within a donor; they do not manufacture additional donors.
The distinction changes investment decisions. If the uncertainty is mainly about laboratory consistency, more technical replication may help. If it is about whether a response generalizes across patients, adding more structures from the same source may leave the main question untouched. Good study design identifies which level of uncertainty needs reducing before buying more measurements. That is a statistical judgment and a biological one.
6. Precision within a batch can hide fragility across batches
A model may behave consistently on one day and differently when generated again. The first observation concerns within-batch precision; the second concerns reproducibility across preparations. Both matter, but they answer different questions. A tight cluster of measurements can give a misleading impression of reliability if every member of the cluster shares the same unexamined source of variation.
Take a hypothetical experiment with two candidate drugs. Every sample for drug A comes from the first preparation, while every sample for drug B comes from a later preparation. If the preparations differ in baseline behavior, the drug comparison is entangled with a batch comparison. Measuring each sample more precisely will not disentangle those explanations. The problem lies in how the experiment was organized.
An informative design distributes the relevant comparisons across independent preparations and records the batch structure in the analysis. The exact arrangement depends on practical constraints, but the principle is general: avoid making the factor of interest travel together with a factor that could produce the same result. When that separation is impossible, the uncertainty should be acknowledged rather than hidden beneath a small statistical error bar.
Batch robustness also has a commercial meaning. A laboratory method that requires one exceptional operator or one unusually favorable production run may be an interesting discovery tool without being ready for widespread deployment. Moving toward routine use means learning which variations are tolerable, which are consequential, and how failure is recognized. That learning can be slow precisely because the system is biological. Standardization cannot make living material identical, but it can make the remaining variation visible and interpretable.
7. Cell quality is a moving property
Characterization at the beginning of a project does not automatically characterize every later experiment. A biological preparation has a history. Storage, expansion, selection, and repeated handling can change which properties are relevant to its interpretation. The key editorial question is whether the checks supporting a claim describe the material actually used, rather than a related material examined at an earlier stage.
The 2023 paper introducing the ISSCR standards framed material integrity and clear reporting as foundations of reproducible stem cell research. It described recommendations covering basic characterization, monitoring of genetic properties, model systems, and reporting. The standards paper supports a modest but important claim: quality is something to demonstrate and document, not an attribute permanently acquired when a cell line receives a name.
Consider a hypothetical longitudinal study in which a laboratory compares drug responses measured months apart. If the preparation has changed during that interval, an apparent change in drug activity might reflect the material rather than the compound. There may be ways to investigate that possibility using retained reference material or repeated benchmarks, but those checks need to be considered before the original material is unavailable.
This does not mean every experiment requires every conceivable characterization test. More testing can consume scarce material and produce information unrelated to the question. The useful approach is risk-based in the ordinary scientific sense: identify which changes could invalidate the inference and measure the properties needed to detect them. The resulting quality plan should be specific enough that another laboratory can understand why a sample was accepted, rejected, or treated cautiously.
8. The matrix and the missing surroundings
Cells do not experience only the drug added to a model. They experience the conditions created by the system around them. For analytical purposes, this environment should be treated as part of the model rather than neutral packaging. If a result changes when the surrounding conditions change, that does not necessarily make it false. It identifies a dependency that may be central to interpreting the result.
A hypothetical tumor model might show different growth behavior in two environments. One interpretation is that one environment is a technical nuisance. Another is that the environments reveal different biological constraints. Distinguishing these possibilities requires a question about what the model is supposed to represent. If the clinical phenomenon depends on an interaction absent from the preparation, reproducing the cells alone may be insufficient.
The reverse problem is also possible: adding complexity can create additional variation without improving the prediction of interest. A developer might include more components because a richer system appears more realistic. Yet if those components make the assay less stable and contribute little to the endpoint, the result may be a less useful test. Complexity should earn its place through demonstrated relevance.
A helpful comparison is between two models aimed at the same decision. Does the more elaborate version improve discrimination of meaningful outcomes, reveal a previously missed mechanism, or support a wider range of exposures? If the answer is unknown, the added complexity remains a hypothesis about value. This framing encourages innovation while resisting the assumption that more organ-like always means more decision-useful. Models should be assessed by the work they can reliably perform, not by a competition to include the greatest number of biological features.
9. A concentration on the label is not necessarily an exposure
Drug testing claims often compress exposure into a single number. For interpretation, one must ask what that number represents: the amount introduced into the system, the amount available near the cells, or a measurement over time. These quantities need not be interchangeable. A model may respond differently because the biological mechanism differs, because the delivered exposure differs, or because both differ.
This is particularly important when comparing platforms. If two systems receive the same nominal concentration but deliver different effective exposure to the relevant cells, a difference in response cannot immediately be attributed to superior biological realism. The experiment may instead be comparing the behavior of the compounds in two physical environments. That behavior is itself relevant, but it needs to be distinguished from the intended biological inference.
The TEX-VAL consortium's 2022 account of microphysiological-system evaluation discussed practical problems in technology transfer, cell sourcing, device complexity, and interpreting exposure. Its first-hand evaluation report is valuable because it describes the work required to make platforms usable rather than relying only on their potential. The report concerns microphysiological systems broadly; its observations should not be attributed indiscriminately to every organoid preparation.
For a hypothetical drug-safety comparison, the appropriate question is whether the tested exposure has a defensible connection to the clinical scenario being modeled. That connection may require several measurements and assumptions. A transparent publication identifies both. Without it, a finding of no injury could mean the model is insensitive, the drug is not harmful under the tested conditions, or the relevant exposure was never achieved. Those possibilities lead to very different conclusions.
10. Time changes the question
A brief experiment and a prolonged exposure study can ask fundamentally different biological questions. A model observed at one time point may reveal an immediate response while missing a delayed one. Conversely, a late measurement may combine initial injury, adaptation, selection, and recovery in a way that obscures the original sequence. The number of days is therefore not merely a scheduling detail.
Imagine a hypothetical experiment in which a treatment reduces a growth-related signal after twenty-four hours. That observation alone cannot establish whether the cells are transiently slowing, permanently impaired, or changing composition. Additional observations might distinguish these explanations, but the original endpoint cannot be silently expanded into all of them. A model's usefulness depends partly on whether its observation window matches the phenomenon of interest.
Timing also complicates comparisons with people. A laboratory exposure pattern may not correspond to the changing concentrations produced by an actual treatment schedule. A model may be suitable for a mechanistic test under sustained exposure without reproducing clinical dosing. That is not a flaw if the purpose is stated correctly. It becomes a problem when the mechanistic result is presented as a direct forecast of a patient's experience.
The same issue arises with the model's own developmental state. A system can change over time before any drug is introduced. If the baseline is moving, treatment comparisons must account for that movement. An apparent effect might depend on when the model was examined. Reporting the relevant timing makes the finding more useful, because another investigator can test whether the response persists, disappears, or reverses outside that window. A time-limited result can be valid and important without being timeless.
11. What counts as the right answer?
Validation requires something against which a prediction can be judged. In drug development, choosing that reference can be difficult. Clinical outcomes may be uncommon, inconsistently recorded, dependent on dose, or entangled with other conditions. A label such as toxic or safe can conceal substantial variation in the circumstances that produced it. The quality of the reference therefore limits how confidently the model's performance can be interpreted.
Suppose a hypothetical benchmark includes twenty compounds classified as harmful and twenty classified as not harmful. A model distinguishes the groups well. Before calling that a general safety test, ask how the compounds were selected, whether they represent the chemical and biological range of future candidates, and whether the classification was known during assay development. A balanced, curated benchmark is not necessarily representative of the next hundred compounds a company will encounter.
There is also an outcome-matching problem. If the model measures a particular type of cellular injury while the reference label combines many unrelated adverse effects, disagreement may be hard to interpret. The model could be accurate for its intended mechanism yet appear weak against an overly broad label. Alternatively, a broad association could conceal failure on the mechanism that matters most for a planned use.
A useful validation study therefore defines the reference with the same care it defines the model. What event, at what exposure, in which population, and with what degree of certainty? Where the reference itself is uncertain, the analysis should not pretend that every disagreement is a model error. Good evaluation distinguishes uncertainty in the instrument from uncertainty in the answer key.
12. The test set must remain a test
Developers learn from their mistakes. That is how models improve. But once a benchmark has repeatedly guided adjustments, performance on that benchmark no longer represents an untouched test of the final system. The issue is not misconduct; it is the ordinary consequence of using feedback. A genuinely independent evaluation asks whether the accumulated improvements extend beyond the examples that shaped them.
A hypothetical organoid assay might initially miss several known harmful compounds. The team changes an endpoint and improves detection. That is progress in development. If the same compounds are then used to report final predictive accuracy, the publication should explain their role in tuning the assay. A new set of compounds, tested under a prespecified procedure, provides a different kind of evidence.
Blinding can help protect that distinction. If operators or analysts know the expected classification, choices about exclusions, image selection, or borderline results may unintentionally favor agreement. A locked analysis plan does not eliminate every source of judgment, but it makes consequential decisions visible before the answers are known. This is especially important when the readout involves a complex combination of measurements rather than a single unambiguous event.
Independence also has a laboratory dimension. A method that succeeds only in its originating laboratory has not yet demonstrated portability. An outside team may encounter undocumented assumptions that the developers perform automatically. Resolving those differences can improve the method substantially. A failed transfer need not invalidate the original science; it may reveal the distance between a skilled demonstration and a reproducible tool. That distance should be measured rather than concealed.
13. A false reassurance and a false alarm have different costs
A drug-safety model can miss a harmful effect or flag a compound that would not cause the relevant harm in people. Both errors matter, but their consequences depend on where the model is used. An early discovery screen may tolerate extra false alarms if follow-up tests can recover promising candidates. A decision used to omit another line of safety evidence may demand a different balance.
Consider a hypothetical system that identifies nine of ten harmful compounds and flags twenty of ninety other compounds. It has detected most of the harmful examples, but twenty-nine compounds receive a positive signal and only nine belong to the harmful group. Those numbers do not make the system useless. They show why the next action matters. If positives receive a careful secondary investigation, the model may be useful. If positives are automatically abandoned, the cost of false alarms becomes much larger.
The same numerical performance can therefore have different value in different workflows. It should not be summarized with one adjective such as accurate or reliable. Readers need the error pattern and the decision rule. They also need uncertainty around the estimates, particularly when only a small number of relevant adverse examples were tested.
For organoid research, this is a way to move beyond a contest between animal and non-animal methods. The question becomes which combination of evidence produces the most defensible decision for the particular risk. A model can reduce uncertainty without eliminating it. Its contribution should be evaluated against the alternatives actually available, including the consequences of acting, delaying, or seeking additional evidence.
14. Patient-derived does not automatically mean patient-predictive
The phrase patient-derived is powerful because it connects a model to a person rather than an abstract cell population. But a sample taken from a patient does not necessarily preserve every feature responsible for that patient's response to treatment. The sample represents a location, a time, and the biological material successfully carried into the model. Its relationship to the whole clinical situation has to be tested.
A hypothetical tumor sample might contain a responsive cellular population while another part of the patient's disease behaves differently. Alternatively, the model might reproduce an important response faithfully but omit the constraints that determine whether a useful exposure can be achieved in the body. In both cases, the laboratory result can be informative without serving as a treatment instruction.
To establish patient-predictive performance, an evaluation needs a clear sequence. The model produces a prediction using a defined procedure. The relevant clinical outcome is then assessed independently. The study includes the cases in which the model could not be established or could not return a result in time. Leaving those cases out can make the method look more broadly applicable than it is.
A further step is needed to establish clinical usefulness. Even a prediction that correlates with outcomes may not improve decisions over existing information. A study of whether using the prediction changes outcomes asks a different question from a study of whether the prediction is associated with them. This article does not recommend organoid testing for individual care. Its point is methodological: the patient's name attached to the source material cannot substitute for evidence about what the resulting test can do for patients.
15. Failure to build the model is part of performance
A laboratory platform has an operating denominator. How many samples entered the workflow, how many produced usable models, how many passed quality checks, and how many returned interpretable results? Reporting only the successful final experiments hides information that may be essential to the intended application. A perfect prediction from a small, selected fraction of samples is not the same service as a dependable result for nearly everyone.
Imagine a hypothetical study beginning with one hundred samples. Sixty yield usable models, fifty produce timely readouts, and forty-five of those agree with a clinical reference. Reporting ninety percent agreement among completed tests is arithmetically correct. It is incomplete if readers infer that the method delivers an accurate answer for ninety percent of the original population. The overall workflow and the conditional test result need separate reporting.
Failures may also be informative rather than random. If certain sample types are less likely to survive the process or meet quality criteria, the successful models could represent a selected biological subset. That selection can change the meaning of the performance estimate. A method validated in easily modeled samples may not transfer to cases where creating the model is difficult.
Turnaround time belongs in the same discussion. A result that arrives after the relevant research or clinical decision has been made may have little operational value even if it is scientifically sound. Different uses permit different delays, so there is no universal acceptable timeline. The requirement is to evaluate the workflow in the conditions of its proposed use rather than treating every completed assay as equally actionable.
16. More measurements do not necessarily mean more information
Modern tissue models can produce images, molecular profiles, electrical signals, and other readouts. Combining them can reveal a richer response than any one measurement alone. It can also create a large space in which an apparently strong pattern may be discovered by chance. The number of available measurements should therefore sharpen the need for a clear analysis strategy rather than relax it.
A hypothetical experiment might measure hundreds of features in a small number of independent donor models. A flexible algorithm can find a combination that separates those donors exceptionally well. The central question is whether the combination captures a generalizable response or the peculiarities of the observed set. Testing on new donors is more informative for that question than splitting closely related measurements from the same donor between training and test groups.
There is a second issue: interpretability of failure. If a composite score changes, researchers should understand which components drove the change and whether those components remain technically reliable. A score that depends heavily on an unstable imaging feature may behave well until the microscope, sample presentation, or analysis software changes. The biological model and the computational analysis form one measurement system for purposes of validation.
This does not argue against complex analysis. It argues for preserving the chain from raw observation to conclusion. A useful publication explains how features were selected, how missing measurements were handled, where the analysis was fixed, and what remained exploratory. Rich data are most valuable when they support a testable explanation or a reproducible prediction. Their volume alone cannot compensate for too few independent biological examples.
17. Qualification belongs to a use, not a technology category
A successful regulatory interaction involving one model does not qualify every model that shares its label. Nor does permission to submit non-animal evidence mean that the evidence will answer every question in an application. The scientific claim and the regulatory claim should be stated at the same level of specificity: a defined method, a defined purpose, and a defined evidentiary role.
In July 2024, NCATS announced four translational centers working with FDA to develop tissue-chip technologies toward use as drug-development tools. The announcement described projects and goals, including work on liver, kidney, barrier, and pregnancy-related models. It did not establish that all the proposed tools had completed qualification or that they could replace every relevant study.
The FDA's NAMs page, checked for this article, also distinguishes policy developments from specific draft guidance. It states that the September 2026 terminology update does not require a particular method or change evidentiary standards. That agency description is a useful constraint on headlines: greater methodological flexibility is not a lower burden of showing that evidence is relevant and credible.
For readers following the field, the practical questions are concrete. What decision is the tool intended to support? Which substances and exposure conditions fall within its evaluated range? What additional evidence remains necessary? What would count as a result outside the model's capability? Answers to those questions reveal more than the broad statement that a technology has regulatory interest. They identify the circumstances in which the model can actually change scientific practice.
18. Reproducibility has an economic structure
A model's cost is not just the price of one successful assay. It includes failed preparations, quality checks, operator time, equipment, analysis, and the work required to interpret ambiguous results. Those costs may be justified by better information. But claims of efficiency should compare complete workflows rather than an idealized laboratory step with an entire existing development process.
A hypothetical platform could be more expensive per experiment yet reduce the number of uninformative follow-up studies. Another could be inexpensive per well but require so much rework that its apparent advantage disappears. The useful economic outcome is connected to a decision: fewer missed hazards, better candidate selection, faster resolution of uncertainty, or a defensible reduction in other testing. Counting samples alone will not establish any of those benefits.
Access also affects the evidence base. If only a few highly specialized laboratories can operate a system, published performance may reflect expertise that is difficult to reproduce elsewhere. A simpler version with slightly less elaborate biology might ultimately generate more reliable comparative evidence because more groups can use it consistently. This is a hypothesis to evaluate, not a reason to prefer simplicity in every case.
The strongest commercialization claim is therefore operational as well as biological. It describes what a receiving laboratory needs, what support is required, how performance is checked after transfer, and what happens when a result fails acceptance criteria. These details may appear mundane beside images of living tissue. They are the details that determine whether a scientific advance remains a demonstration or becomes an instrument other people can trust.
19. Consent and interpretation travel with the cells
Human-derived models raise questions about the conditions under which material was obtained and how associated information is handled. A model can carry biological information relevant beyond the original experiment. The appropriate handling depends on the consent, institutional oversight, applicable requirements, and the nature of the data. Broad claims that every sample is anonymous or that every donor has agreed to every future use should not be assumed.
For scientific reporting, provenance serves two purposes that can pull in different directions. Rich information about the source can help explain variation and assess generalizability. Restricting access to identifying or sensitive information can protect the people who contributed material. A thoughtful publication can explain what characteristics were considered and what access arrangements exist without exposing personal details.
There is also an obligation to avoid misleading interpretation of unusual models. A neural organoid is not made equivalent to a person's experience by being called a miniature brain. A tumor-derived preparation is not a complete copy of a person's cancer. Language that suggests identity between model and person can exaggerate both capability and ethical conclusions. Precise descriptions support more serious discussion of the actual questions raised by each system.
For donors and readers alike, the most respectful account is specific about what the research uses, measures, and claims. It neither trivializes the human source nor attributes person-level capacities to the model without evidence. Consent and scientific validity are distinct matters, but both benefit from resisting the same rhetorical shortcut: treating a biological resemblance as proof of equivalence. The material deserves careful handling, and the people represented by it deserve accurate explanations.
20. The next convincing result will have boundaries
The strongest next-stage organoid studies may be less sweeping than the headlines surrounding them. They may show that a defined model reliably detects a particular injury mechanism, that an independent laboratory can reproduce a response, or that a prediction retains value in a new set of donors. Each result would narrow a real uncertainty. None needs to establish that a dish contains an entire patient.
For a reader evaluating a new announcement, the first task is to identify the claim's level. Is the work about creating a structure, explaining a mechanism, predicting an outcome, or improving a decision? Then inspect the evidence at that level. A visually convincing structure does not settle prediction. A predictive association does not settle clinical usefulness. A policy announcement does not settle experimental validity. These are connected stages, not interchangeable achievements.
The next task is to find the boundaries: donors, batches, exposures, timing, endpoints, failed attempts, and independent tests. Boundaries are not merely caveats added to protect the authors. They define the conditions under which a model can be used intelligently. A tool with well-understood limits can contribute more to science than a vaguely universal tool whose failures are difficult to recognize.
The regulatory developments verified for this article make that discipline more urgent, because the field is being asked to supply evidence for consequential decisions. Organoids can offer access to human biology that is difficult to obtain in other ways. Realizing that value requires treating them as models with testable capabilities. Their promise does not depend on pretending they are miniature patients. It depends on showing exactly when a carefully constructed part of biology can answer a question about the whole.