Around 11,000 years ago, communities in southwest Asia began one of the most consequential experiments in human history: growing their own food. Wild cereals, legumes and fruit-bearing plants were gradually brought under human management, setting in motion the transition from mobile hunter-gatherer lifeways to farming societies. For decades, archaeologists have assumed that one of the earliest signs of this transformation was an increase in cereal grain size. A new study, however, suggests that some of the biggest grains found at the dawn of agriculture may not have been genetically altered at all. They may simply have grown in wetter conditions.

The findings, published in the Proceedings of the National Academy of Sciences, challenge a central assumption about how cereal domestication began. Researchers studying charred barley and emmer wheat grains from the Early Neolithic sites of el-Hemmeh and Sharara in the southern Levant found that the plants still possessed fully wild characteristics during the earliest phases of the Pre-Pottery Neolithic. Their ears had brittle rachises, the structures that naturally break apart when mature, allowing seeds to disperse in the wild. There was no evidence of the tough, non-shattering rachis that later became one of the defining features of domesticated cereals.

The absence of non-shattering forms is significant because a tough rachis makes harvesting far more efficient. In wild cereals, the rachis fractures as the plant ripens, scattering seeds before people can collect them. A mutation that keeps the grains attached would have offered major advantages to human harvesters and could then have spread through repeated cultivation and replanting. Researchers have therefore often treated the appearance of non-shattering cereals as a key marker of domestication. Grain size, by contrast, has frequently been interpreted as an earlier signal that people were already selecting plants through cultivation.

At el-Hemmeh and Sharara, however, grain size told a more complicated story. Barley and emmer wheat remained morphologically wild, yet their grains varied dramatically. Some were as small as typical wild cereals, while others fell within the size range commonly associated with domesticated plants. If large grains were the product of genetic selection caused by systematic tillage, the researchers expected to find other evidence of intensive cultivation. Instead, the ecological evidence pointed toward relatively low-disturbance environments.

The team examined the weed species preserved alongside the cereal remains. Weeds are valuable archaeological indicators because different plants thrive under different environmental conditions. Species associated with heavily tilled or frequently disturbed soils were not dominant in the assemblages from these early sites. This weakened the argument that repeated ploughing or intensive soil preparation had already created strong selection for larger seeds. The researchers began looking for another explanation for the striking variation in grain size.

Their answer came from stable carbon isotope analysis. Carbon exists in several naturally occurring forms, or isotopes, and plants incorporate these isotopes in ways that are influenced by their growing conditions. In cereals, the stable carbon isotopic composition of a grain can reflect the balance between carbon uptake and water loss during photosynthesis. When plants experience water stress, changes in gas exchange through their leaves alter the carbon isotope signal recorded in their tissues. By measuring this signal in ancient grains, scientists can reconstruct aspects of the moisture conditions under which the plants grew.

The results showed a strong relationship between grain size and water availability. Barley grains that reached domestic-sized dimensions had carbon isotope values indicating that they developed under wetter conditions. Smaller, wild-sized grains carried signatures consistent with considerably drier growth environments. The difference suggests that environmental moisture, rather than inherited genetic change, was responsible for much of the early increase in cereal grain size. In other words, the same wild plant population could produce grains of very different sizes depending on the conditions it experienced during development.

This phenomenon is known as developmental plasticity: the ability of a single genotype to produce different physical traits in response to the environment. It is widespread in plants and can affect characteristics such as height, leaf area, flowering time and seed production. In the Early Holocene southern Levant, developmental plasticity may have created the visual impression that cereals were already becoming domesticated. People collecting and replanting the largest grains might eventually have generated genetic selection, but the initial size increase itself appears to have been ecological rather than evolutionary.

“Environmental conditions provided a much better explanation for the observed variation in grain size than genetic selection under tillage,” said Dr Jade Whitlam of the University of Oxford, who led the study. The researchers propose that genetic selection for larger grains may have occurred later than previously thought, perhaps after selection for non-shattering rachises had already begun. This reverses the sequence often assumed in models of domestication and places greater emphasis on the complex interaction between climate, plant biology and human harvesting practices.

The study also reframes the origins of agriculture as a gradual process of niche construction rather than a sudden decision to domesticate plants. Early communities altered their relationship with wild vegetation through harvesting, protection, transport and repeated use of particular landscapes. These practices may have changed plant populations long before crops displayed unmistakable domestication traits. By combining archaeobotany, weed ecology and stable isotope evidence, the researchers argue that the first steps toward agriculture were not necessarily driven by deliberate genetic engineering of crops, but by prolonged human management operating within a variable environment. The result is a more nuanced picture of farming’s origins—one in which climate could temporarily make wild cereals look domesticated before their genomes caught up.

Subject of Research: The environmental and biological processes that influenced cereal grain size and plant domestication in the Early Holocene southern Levant.

Article Title: Developmental plasticity under human management shaped cereal evolution prior to domestication in the Early Holocene southern Levant

News Publication Date: 4-Aug-2026

Web References: https://www.eurekalert.org/multimedia/1146086

References: Proceedings of the National Academy of Sciences. DOI: 10.1073/pnas.2535274123

Image Credits: Jerome Poulalier, www.jerome-poulalier-photography.com

Keywords: Archaeology, archaeobotany, cereal domestication, agriculture, Neolithic, Early Holocene, southern Levant, barley, emmer wheat, stable carbon isotopes, developmental plasticity, plant evolution, climate, hunter-gatherers

Cities may soon be able to identify their hottest walking routes with the same precision used to map traffic or public transport. HeiGIT, the Heidelberg Institute for Geoinformation Technology, has added a tree-shade indicator to its free Climate Action Navigator, a digital platform designed to help municipalities and civil society organizations understand how urban environments affect everyday life. The new tool provides a street-level view of where pedestrian paths are protected by tree canopies and where people are exposed to direct sunlight, offering planners a detailed basis for decisions about urban greening, heat adaptation and safer walking routes.

The need for such information is becoming increasingly urgent as cities experience more frequent and intense heat waves. A citywide estimate of tree cover can suggest that an urban area is relatively green while concealing major differences between neighborhoods. In one district, a large park or wooded hillside may raise the average, even though the streets used daily by residents remain almost entirely unshaded. The Climate Action Navigator is designed to reveal these local contrasts by measuring shade along pedestrian routes rather than relying only on broad land-cover statistics.

In Innsbruck, Austria, tree cover extends across approximately 30 percent of the city’s relevant areas, but the distribution is far from even. Much of the shaded pedestrian network lies outside the central urban area, while heavily used streets in the city center can remain exposed to full sun. This distinction matters because people experience heat at the scale of individual journeys: walking to school, commuting to work, visiting shops or reaching public transport. Two routes separated by only a few streets may therefore offer very different levels of thermal comfort and protection from solar radiation.

The indicator is also intended to connect physical infrastructure with human behavior and well-being. Shaded paths can make walking more attractive during hot weather, potentially influencing whether people choose to travel on foot, wait for public transport or avoid certain destinations altogether. Claudia Schmiedeberg of the GLEN Study said that understanding the social effects of climate change requires examining how local surroundings shape daily decisions over time. Combining the tree-shade data with longitudinal survey information could help researchers investigate whether residents in greener neighborhoods adapt differently to heat, and whether unequal access to shade contributes to wider social disparities.

Technically, the tool identifies pedestrian walkways that directly overlap with tree canopies at least two meters high. The height threshold is designed to distinguish trees capable of producing meaningful overhead shade from low shrubs, grass and ground vegetation. The analysis uses high-resolution global canopy datasets and applies the same processing method across different cities. This allows a small town, a regional center and a major capital to be compared using a consistent measure rather than relying on incompatible local mapping systems.

The calculation also models a high-noon scenario, when the sun is strongest and shadows are generally shortest. By estimating shade under these demanding conditions, the indicator focuses on a daily period when heat exposure may be especially severe. It measures the availability of shade at what the developers describe as the weakest point of the day, creating a conservative benchmark for planning. The result is not a complete simulation of human thermal comfort, which also depends on air temperature, humidity, wind, pavement materials and building geometry, but it provides a clear spatial signal of where tree canopy is present above pedestrian infrastructure.

Early comparisons show how sharply shade levels can differ between cities. In the historic city of Maribor, Slovenia, approximately 35 percent of the pedestrian network is covered by trees. In Tempe, Arizona, a university city located in a hot and generally arid region, the corresponding figure is just above 7 percent. Such numbers do not by themselves determine whether a city is comfortable or unsafe, but they provide a comparable starting point for identifying exposure. When combined with information about schools, care facilities, public transport stops or tourist destinations, the data can show which groups and routes may require priority attention.

Urban planners could use the indicator to decide where new trees would provide the greatest benefit, while also recognizing streets where existing canopy already supports walking during hot periods. Linking shade maps with school catchment areas, for example, could reveal whether children have protected routes to school or must travel through long stretches of direct sunlight. Tourism agencies and local authorities could also use the information to recommend cooler walking routes to major attractions. In this way, the platform is intended not only to document an environmental problem but also to support targeted interventions, from tree planting and maintenance to short-term route guidance.

Jim Walker of the Walk21 Foundation said the indicator offers comparable, street-by-street evidence that can help administrations prioritize infrastructure investment during the hottest months. Its developers argue that global canopy datasets make it possible to overcome the local data shortages that often prevent smaller municipalities from conducting detailed climate assessments. Dr. Kirsten von Elverfeldt, community engagement manager at HeiGIT, said the goal was to make granular diagnostic information freely available so that greening initiatives can be directed toward the places where they are most needed.

The Climate Action Navigator is open to the public, allowing users to explore how their municipality performs and compare different urban areas. The platform’s tree-shade indicator does not replace on-the-ground assessments, but it can help transform an abstract climate challenge into a visible pattern of shaded and exposed streets. As heat increasingly shapes how people move through cities, that ability to see the difference at walking level could make urban climate planning more precise, more equitable and more immediately actionable.

Subject of Research: Not applicable

Article Title: New Global Tool Maps Tree Shade Street by Street to Reveal Cities’ Heat Exposure

News Publication Date: August 11, 2026

Web References: Climate Action Navigator: https://heigit.org/ ; Tree shade indicator video: https://www.youtube.com/watch?v=vAQxEnAgM44 ; Innsbruck example: plugin/hiwalk?share-id=f50386c3-4ea6-45b8-946a-6fedfae38dc8 ; Maribor example: plugin/hiwalk?share-id=f675a5bd-f15f-429c-81f4-80cac6c7051a ; Tempe example: plugin/hiwalk?share-id=31880ebb-1c6d-40e1-89cf-ffac9bceb607

Image Credits: Climate Action Navigator

Keywords: urban heat, tree shade, climate adaptation, pedestrian routes, urban planning, tree canopy, heat exposure, Climate Action Navigator, HeiGIT, sustainable cities

Lipid nanoparticles (LNPs) have become one of the most important delivery technologies in modern medicine, helping transport messenger RNA into cells for vaccines, protein replacement therapies and experimental treatments. Yet a formulation that performs well in mice can behave very differently in humans. This species gap makes nonhuman primate studies an important step in drug development, but conventional methods for comparing multiple nanoparticles often require euthanasia and extensive tissue collection. A new study reports a way to measure the functional activity of several LNPs in the same animal using only a small blood sample.

Published in Nature Biotechnology, the work introduces a non-invasive nanoparticle barcoding platform called snapCodes. The method is designed to determine which LNP formulation successfully delivers its mRNA cargo and prompts cells to produce the encoded protein. Rather than relying on tissue harvesting, the researchers identify delivery events through DNA barcodes that can be recovered from serum and analyzed by sequencing. The approach could allow scientists to compare many candidate nanoparticles in a single nonhuman primate while reducing both animal use and the logistical burden of downstream analysis.

The central challenge addressed by the study is that nanoparticle delivery is not adequately described by measuring where an LNP travels. An LNP may reach an organ or circulate in the bloodstream without releasing its mRNA in a form that cells can translate. The new system instead records functional mRNA delivery. Each nanoparticle is associated with a distinctive DNA sequence, and the mRNA cargo is engineered to encode a fusion protein containing nanoluciferase and a SNAP-tag. Nanoluciferase is a compact enzyme that produces light in the presence of its substrate, while the SNAP-tag is a protein module capable of forming a covalent bond with specific chemical groups.

For the barcode to become a permanent molecular record of delivery, the researchers modified the DNA tags with benzylguanine. When a cell receives an mRNA molecule encoding the nanoluciferase–SNAP-tag fusion, the resulting protein can react with a benzylguanine-bearing snapCode. This creates a stable DNA–protein complex. The complexes are then exported from cells into the circulation, where they can be isolated from a small serum sample. Sequencing the recovered DNA reveals which barcoded LNPs produced the fusion protein and therefore achieved functional mRNA delivery.

The design turns transient protein expression into a measurable signal that can be collected without sacrificing the animal. This is particularly valuable in studies involving nonhuman primates, where tissue-based nanoparticle profiling can require euthanasia and can limit the number of formulations tested in each subject. A blood-based assay also makes it possible to monitor delivery over time, potentially showing how the activity of a formulation changes after dosing. The study used just 30 microliters of serum, a volume small enough to support repeated sampling and longitudinal measurements.

Before comparing nanoparticle performance, the researchers validated the molecular components of the system, including the activity of the snapCode chemistry and the ability of the fusion protein to participate in the intended covalent reaction. These steps were necessary to establish that the barcode signal reflected genuine expression of the delivered mRNA rather than nonspecific interactions. Once the platform had been characterized, the team used it to evaluate multiple chemically distinct LNPs in living animals.

The investigators intravenously administered six snapCoded LNP formulations to mice and nonhuman primates. Each formulation carried its own molecular identity, allowing all six candidates to be assessed in parallel. Serum collected after administration was analyzed for the DNA–fusion protein complexes, and sequencing was used to quantify the relative delivery activity associated with each barcode. By testing the same barcoded set across species, the researchers could directly examine whether patterns observed in mice were reproduced in nonhuman primates.

Such cross-species comparisons are crucial because LNP composition strongly influences biological behavior. Changes in ionizable lipids, helper lipids, cholesterol, polyethylene glycol-lipid components or particle structure can affect stability, organ distribution, cellular uptake and endosomal escape. These properties may differ substantially between rodents and primates because of differences in blood proteins, immune responses, tissue physiology and cellular trafficking. A method that measures expression from several formulations in the same animal may therefore reveal species-specific delivery patterns more efficiently than traditional one-formulation-per-animal experiments.

The snapCode strategy does not replace every form of nanoparticle analysis. It reports successful expression of a particular engineered mRNA and depends on the recovery and sequencing of the associated molecular complexes. It does not, by itself, provide a complete map of tissue distribution, long-term safety or therapeutic efficacy. Nevertheless, it adds a functional layer to LNP evaluation: scientists can ask not only where particles go, but which particles actually release mRNA that cells translate. In early development, that distinction can help prioritize candidates before more resource-intensive studies begin.

By combining multiplexed dosing with low-volume serum sampling, the researchers aim to make nonhuman primate testing more informative while using fewer animals. The platform could be adapted for screening delivery systems designed for different organs or therapeutic proteins, provided that the encoded reporter and barcode chemistry remain compatible with the intended experiment. As mRNA medicines expand beyond vaccines into gene editing, immunotherapy and protein replacement, tools that connect nanoparticle chemistry to real biological output may become increasingly important. Non-invasive barcoding offers a way to track those differences in living subjects and could help narrow the path from promising LNP design to clinically relevant drug candidate.

Subject of Research: A non-invasive method for comparing functional mRNA delivery by multiple lipid nanoparticles in mice and nonhuman primates using serum-based DNA barcodes.

Article Title: Non-invasive nanoparticle barcoding in nonhuman primates

Article References: Jang, B., Zenhausern, R., Lian, L. et al. Non-invasive nanoparticle barcoding in nonhuman primates. Nat Biotechnol (2026). https://doi.org/10.1038/s41587-026-03262-6

Image Credits: AI Generated

DOI: https://doi.org/10.1038/s41587-026-03262-6

Keywords: lipid nanoparticles, mRNA delivery, nanoparticle barcoding, snapCodes, SNAP-tag, nanoluciferase, nonhuman primates, serum sequencing, drug delivery, RNA therapeutics

A new study is drawing attention to a hidden problem in mental-health research: scientists may be using the same words to describe first episode psychosis (FEP) while measuring entirely different things. In an analysis published in Schizophrenia, researchers Seiber, Sridhar, Hasenstab and colleagues examine how U.S. insurance claims data are used to estimate the incidence of FEP—the number of new cases appearing in a population over a defined period—and argue that greater standardization is urgently needed.

First episode psychosis refers to the point at which a person experiences psychotic symptoms for the first time in their life. These symptoms can include hallucinations, delusions, disorganized thinking or major changes in behavior and functioning. FEP is not a single diagnosis. It can occur in schizophrenia-spectrum disorders, bipolar disorder, severe depression and several medical or substance-related conditions. Because early treatment is linked to better long-term outcomes, knowing how often FEP occurs is essential for planning specialized clinics, staffing early-intervention programs and directing public-health resources.

The researchers focus on U.S. claims data, enormous digital records generated when patients receive medical care and providers submit bills to insurers. These databases can include diagnostic codes, procedure codes, prescription information, dates of service and details about emergency, inpatient and outpatient treatment. Unlike traditional epidemiological surveys, claims datasets can cover millions of people and make it possible to study patterns across large geographic areas. Their scale, however, does not automatically make their estimates precise.

The central challenge is that a billing code is not the same as a clinical assessment. A code for psychosis may indicate a confirmed diagnosis, a suspected condition, a historical problem or a provisional label assigned during an emergency evaluation. Claims data also rarely contain the complete clinical narrative needed to determine whether symptoms truly represent a person’s first psychotic episode. A patient may have received care in another health system, paid out of pocket, changed insurers or carried an earlier diagnosis that is invisible in the database being studied.

Even the definition of “incidence” can vary. Some researchers count a first-ever psychosis-related diagnosis, while others count a first hospitalization, a first emergency-department visit or the first appearance of a relevant code after a period without documented care. These approaches can produce different numbers from the same underlying population. A study requiring a long “clean period”—a span of time with no previous psychosis-related claims—may reduce the likelihood of counting established cases as new, but it can also exclude people whose earlier treatment occurred outside the available records.

Age, sex, insurance type and geography can further influence the apparent rate of FEP. Young adults may be more likely to encounter emergency services, while people in rural areas may face limited access to psychiatrists and specialized programs. Differences in insurance coverage can determine which services generate observable claims. If researchers compare estimates without accounting for these factors, they may mistake variations in detection or access for genuine differences in disease occurrence.

The study’s call for standardization points toward a more consistent technical framework. Researchers need to specify which diagnostic codes qualify as psychosis, whether substance-induced and medically caused episodes are included, how long a person must remain free of prior claims to be considered a new case, and which healthcare settings are examined. They also need to define the population denominator—the number of people considered at risk—and explain how individuals with incomplete enrollment or interrupted insurance coverage are handled.

Validation is another crucial step. Claims-based algorithms should be compared with detailed chart reviews, clinical registries or structured assessments to determine how accurately they identify genuine FEP. Two statistical measures are especially important: sensitivity, or the ability to capture true cases, and positive predictive value, or the proportion of algorithm-identified cases that are confirmed after review. An algorithm with high sensitivity may find more possible cases but also produce more false positives; one with high specificity may miss people who received vague or inconsistent coding.

Standardized methods could make FEP estimates more comparable across states, insurers and research teams. That would help scientists detect real trends rather than methodological noise and could reveal whether early psychosis services are reaching the populations most in need. More reliable estimates could also strengthen efforts to investigate racial, socioeconomic and regional disparities, provided that claims data are interpreted carefully and not treated as a complete substitute for clinical information.

The message from the research is not that U.S. claims databases are unusable. On the contrary, their breadth makes them one of the most powerful tools available for studying mental-health care at scale. But their value depends on transparent definitions, validated case-finding methods and clear reporting standards. As health systems increasingly rely on administrative data and automated analytics, agreeing on what counts as a new case of psychosis may be the difference between a map that reveals a public-health crisis and one that merely reflects the quirks of the billing system.

Subject of Research: Estimating first episode psychosis incidence rates using U.S. claims data

Article Title: Estimating first episode psychosis (FEP) incidence rates using U.S. claims data: a need for standardization

Article References: Seiber, E.E., Sridhar, S., Hasenstab, K.A. et al. “Estimating first episode psychosis (FEP) incidence rates using U.S. claims data: a need for standardization.” Schizophrenia (2026). https://doi.org/10.1038/s41537-026-00793-4

Image Credits: AI Generated

DOI: 10.1038/s41537-026-00793-4

Keywords: First episode psychosis, FEP, incidence rates, U.S. claims data, schizophrenia, mental health, epidemiology, healthcare data, standardization

Expansion microscopy has spent the past decade changing the rules of super-resolution imaging. Instead of relying solely on increasingly sophisticated optics, the technique physically enlarges biological specimens so that molecules separated by nanometers become easier to distinguish with ordinary fluorescence microscopes. Now, researchers have introduced a method designed to push that concept toward one of the most difficult goals in microscopy: seeing the shapes of individual proteins directly.

The approach, described by Ahmed H. Shaib, Mahmoud M. Alawieh and Stefan O. Rizzoli in Nature Protocols, combines one-step nanoscale expansion microscopy with fluctuation-based super-resolution analysis. The resulting workflow, called ONE microscopy, is intended to make molecular-scale imaging more accessible to laboratories that do not have access to cryo-electron microscopes or specialized high-end optical systems.

Traditional light microscopy is limited by diffraction, a physical effect that causes light from two nearby objects to blur together when they are too close. Even modern super-resolution methods can require complex instruments, intense labeling strategies or extensive computational processing. Expansion microscopy takes a different route. Researchers anchor biological molecules within a swellable polymer network and then expand the material, increasing the physical distance between fluorescent labels.

The expansion step does not automatically reveal every molecular detail. A protein’s structure can still be represented by only a small number of fluorescent signals, and conventional images may contain noise, background fluorescence and motion-related fluctuations. ONE microscopy addresses these limitations by analyzing changes in fluorescence intensity over time. These fluctuations contain information about the presence, position and behavior of labeled molecules that may not be obvious in a single frame.

In practical terms, the method links chemical preparation, physical enlargement and computational analysis into a single workflow. Samples are embedded in a polymer gel, labeled to identify the proteins of interest and then expanded. After expansion, researchers acquire image sequences using conventional fluorescence equipment. Specialized software analyzes the temporal variation in the recorded signals, extracting spatial information beyond what a standard diffraction-limited image would provide.

This combination is significant because it shifts the focus from simply locating a protein to examining its overall shape. For many biological questions, knowing that a protein is present is not enough. Its size, orientation and structural organization can determine how it interacts with membranes, vesicles, organelles or neighboring proteins. Directly observing those features could help researchers investigate molecular machines in their native cellular environments rather than relying exclusively on purified samples or averaged structural models.

The protocol is designed to work across a broad range of biological materials, including purified proteins, cultured cells and tissues. That flexibility could make the technique useful for researchers studying protein organization at multiple scales. A purified protein might provide a controlled test of shape reconstruction, while cells and tissues could reveal how the same protein is arranged amid the crowded and complex environment of living biology.

A major part of the reported advance is the accompanying software package. Fluctuation-based imaging can be powerful, but its usefulness depends on reliable data processing, and computational analysis has often been a barrier for non-specialist users. The authors present the software as stable and user-friendly, with the goal of making the analysis more efficient, reproducible and practical for laboratories using standard fluorescence microscopes.

ONE microscopy does not replace cryo-electron microscopy, which remains capable of resolving structures at atomic or near-atomic scales under appropriate conditions. Nor does it eliminate the challenges associated with labeling, gel chemistry, image quality and sample preparation. Expansion can introduce distortions, and successful imaging depends on preserving the relationship between the fluorescent labels and the underlying structures. Nevertheless, the method offers a complementary strategy: rather than averaging thousands or millions of molecules, it aims to examine individual protein shapes through expanded, fluorescence-labeled specimens.

The researchers describe the workflow as a practical framework for protein imaging on conventional equipment. By combining physical separation of fluorophores with information extracted from fluorescence fluctuations, the method brings nanoscale structural analysis closer to routine biological imaging. Its broader impact may come not from replacing existing forms of super-resolution, but from making a previously specialized capability more reproducible and attainable for laboratories investigating how individual proteins operate inside cells and tissues.

Subject of Research: One-step nanoscale expansion microscopy for visualizing individual protein shapes using conventional fluorescence microscopes.

Article Title: ONE microscopy.

Article References: Shaib, A.H., Alawieh, M.M. & Rizzoli, S.O. ONE microscopy. Nature Protocols (2026). https://doi.org/10.1038/s41596-026-01399-x

Image Credits: AI Generated

DOI: https://doi.org/10.1038/s41596-026-01399-x

Keywords: expansion microscopy, ExM, nanoscale imaging, super-resolution microscopy, fluctuation-based analysis, protein structure, fluorescence microscopy, single-protein imaging, cryo-electron microscopy, biological imaging

Foundation models are promising to become the “universal translators” of biomedical data—but a new perspective argues that medicine should resist treating them as all-knowing digital doctors. In a review published in Nature Biomedical Engineering, researchers describe how large, adaptable artificial intelligence systems are reshaping biomedical imaging while warning that impressive benchmark scores can conceal serious weaknesses in clinical environments. Their central message is direct: foundation models are most likely to improve healthcare by augmenting specialists, not replacing them.

Foundation models are designed to learn broad representations from enormous and diverse datasets before being adapted to specific tasks. In biomedical imaging, those data may include magnetic resonance imaging, computed tomography, X-rays, ultrasound, digital pathology slides and ophthalmic images. The same underlying model could theoretically identify tumors, segment organs, estimate disease risk, retrieve similar cases and generate clinical reports. The vision becomes even broader when imaging is combined with pathology, electronic health records and genomic information, creating a composite system intended to analyze a patient across multiple biological scales.

That ambition reflects a major change from traditional medical AI. Earlier systems were generally built for one narrowly defined task, such as detecting pneumonia on chest radiographs or outlining a brain lesion on MRI scans. Foundation models instead attempt to create a reusable backbone that can be transferred across diseases, hospitals and imaging technologies. Technically, they learn statistical patterns in high-dimensional data, often through self-supervised training in which the system predicts missing, transformed or associated information. Afterward, developers fine-tune or prompt the model for particular clinical applications.

But biomedical imaging is not a single, uniform world. Images differ according to scanner manufacturer, acquisition protocol, patient population, disease prevalence and local clinical practice. A model trained primarily on data from large academic hospitals may perform very differently in community clinics, rural settings or countries with limited equipment. This problem, known as domain shift, can occur even when images appear visually similar. Small changes in image quality, contrast, hardware or patient demographics may alter the statistical distribution learned by the model and undermine its predictions.

The authors introduce a framework called real-world evaluation and assessment of foundation models, or REAL-FM, to examine whether these systems are ready for practical use. Rather than focusing only on accuracy scores, REAL-FM considers several dimensions at once, including the quality and representativeness of training data, technical readiness, clinical value, integration into existing workflows and responsible artificial intelligence. The framework is intended to help clinicians interpret claims about new models and to push developers toward evaluations that resemble actual medical practice.

One of the sharpest distinctions in the perspective is between pattern recognition and causal reasoning. Foundation models can become extraordinarily skilled at recognizing visual associations—for example, linking a particular texture or anatomical feature with a diagnosis present in their training data. Yet association is not the same as understanding why a disease occurs, how it will progress or whether a treatment will benefit an individual patient. A model may identify a correlation that is valid in one hospital but reflects a hidden confounder, such as a scanner type, reporting convention or patient-selection pattern, rather than a biological signal.

This limitation becomes especially dangerous when models are moved beyond simplified benchmarks. Many benchmark datasets provide carefully curated images, clear labels and narrowly defined tasks. Clinical care is messier: scans can be incomplete, diagnoses can be uncertain, records may contain contradictory information and several conditions may coexist. The perspective highlights a lack of verified generalization across such conditions, along with a shortage of prospective, outcome-based validation. In a prospective study, an AI system would be evaluated while care is actually being delivered, with researchers measuring whether it improves diagnostic accuracy, treatment decisions, patient outcomes or workflow efficiency.

Data scarcity is another obstacle. The largest foundation models require vast quantities of data, but medical information is difficult to collect, standardize and share. Patient records are fragmented across institutions, imaging data may lack reliable annotations and rare diseases are inherently underrepresented. Privacy regulations and governance requirements further complicate the creation of centralized datasets. As a result, a model may appear broadly capable while remaining poorly tested in the populations and clinical circumstances where errors could have the greatest consequences.

The researchers argue that human oversight therefore remains indispensable. In practical terms, foundation models may be most useful as clinical assistants that prioritize images for review, highlight suspicious regions, summarize longitudinal records or offer a second opinion that specialists can interrogate. Safe deployment will require transparent reporting of uncertainty, monitoring for performance drift, mechanisms for correcting errors and clear accountability when recommendations influence care. The future envisioned by the authors is not a single monolithic medical oracle, but a coordinated ecosystem of specialized AI systems, each evaluated for a defined clinical role and connected to expert-led workflows.

The perspective arrives as the biomedical AI field races toward increasingly general-purpose systems. Its warning is not that foundation models lack value, but that technical scale alone cannot establish clinical reliability. Before these models can be trusted with consequential decisions, they must demonstrate robustness across domains, usefulness in real workflows, safety under unexpected conditions and measurable benefits for patients. The proposed REAL-FM framework offers a way to separate viral demonstrations from durable medical progress—by asking not only what a model can recognize, but where it works, why it works and whether it makes care better.

Subject of Research: Foundation models and their real-world evaluation in biomedical imaging and clinical medicine.

Article Title: Foundation models in biomedical imaging: turning hype into reality

Article References: Muneer, A., Zhang, K., Hamdi, I. et al. Foundation models in biomedical imaging: turning hype into reality. Nature Biomedical Engineering 10, 1557–1575 (2026). https://doi.org/10.1038/s41551-026-01762-z

Image Credits: AI Generated

DOI: 10.1038/s41551-026-01762-z

Keywords: foundation models, biomedical imaging, medical artificial intelligence, clinical AI, multimodal AI, domain shift, causal reasoning, clinical validation, responsible AI, REAL-FM

Recently, a man I was rock climbing with told me about how he'd used AI to make a motivational poster for himself, which he'd hung on his bedroom wall: a bear, walking a slackline over a canyon, holding a sign that said, "Do cool shit." I made what I hoped was a polite noise. What reaction was this man looking for? He probably hoped someone, maybe me, would say it was cool or inspiring. It seemed I was not the first person he'd told about this poster. "You know, it's funny," he said to me. "No one ever says anything about it."

I have been thinking about this now for several days. There's a sort of turducken of things that I find off-puttin …

Read the full story at The Verge.