A breakthrough in AI-driven crop disease detection is set to reduce harvest losses and chemical-related health risks, thanks to a first-of-its-kind tomato leaf dataset comprising almost 9,000 images.

The Sri Lankan In-Field Tomato (SLIF-Tomato) dataset—built by academics from Charles Darwin University (CDU), the University of Peradeniya (UoP) and others—is the first complete in-field dataset comprising class labels and bounding box annotations collected under diverse real-world conditions.

The dataset includes 890 raw photos of tomato leaves featuring eight classes: healthy and seven diseases—bacterial spot, early blight, mosaic, powdery mildew, septoria, wilt and late blight—which were then augmented through two phases to produce a total of 8,934 images.

The images were assessed using an AI algorithm called the Inverted Residual Convolutional Block Attention Module (IR-CBAM), which tells AI scanners where to focus, reducing image complexity.

The paper is published in the journal Neural Computing and Applications.

Research supervisor Dr. Thuseethan Selvarajah, a CDU lecturer in information technology, said the technology taught AI-based disease recognition models to focus like a trained eye, rather than scanning every part of an image equally.

Teaching AI where to look

“First, it figures out what kind of clues matter most, such as color, texture and edges, and turns up the focus on those,” Selvarajah said.

“Then, it figures out where in the photo to look, zooming in on the actual diseased spot and ignoring soil, shadows and background leaves. It does this in a lightweight, efficient way so it can still run fast on things like phones.

“This matters because field photos are messy. Unlike clean lab images, they’re full of background clutter. Teaching the model to filter out the noise and focus on the disease itself is exactly why accuracy improved significantly.”

The SLIF-Tomato dataset, built by academics from CDU and UoP, contains 8,934 images broken into eight classes. Pictured are examples of their bounding box annotations.

From lab conditions to farms

The dataset allows farmers to detect diseases in their crops early and with more than 99% accuracy. Previous studies indicated this was only achievable in controlled settings.

Sri Lanka’s tomato yield averaged 18.9 metric tons per hectare as of 2018, contributing to national income and export revenue. It also generates employment, boosts household income and supports national nutrition.

UoP Ph.D. candidate and study lead author Romiyal George said undetected or misdiagnosed tomato leaf diseases cause substantial yield losses, while excessive chemical use poses risks to human health and the environment.

George said traditional disease detection often depended on manual inspection by farmers or agricultural experts, a time-consuming process that could delay treatment decisions and result in financial losses.

“The lightweight AI models developed in this research can enable rapid disease identification using resource-constrained devices such as mobile phones or embedded systems,” he said.

“This can reduce dependency on continuous expert support, minimize crop losses through early detection, and help farmers apply treatments more efficiently.

“In the long term, this technology can contribute to precision agriculture by enabling data-driven decision-making and more sustainable use of agricultural resources.”

From research setting to deployment
George said the next step was to take their findings from a research setting to the real world for deployment and expand the available datasets to include a broader range of crops and diseases.

He said AI had a major role to play in the future of agriculture.

“This research is a step toward bridging the gap between advanced AI technologies and practical agricultural applications,” George said.

“By combining lightweight deep learning models with realistic field data, we can move closer to creating affordable and sustainable solutions that benefit farming communities, particularly in regions where access to agricultural expertise is limited.”

The research supports Australia’s agricultural innovation priorities through AI-based crop disease detection, improved productivity and enhanced biosecurity.

Tiny viruses that infect bacteria could one day help us tackle infections that antibiotics can no longer treat. But first, scientists need to understand how these viruses work, and artificial intelligence could provide them with a powerful way to do that.

Known as bacteriophages, or simply phages, these viruses infect bacteria and use them to make more copies of themselves. There are thought to be more phages on Earth than any other type of biological entity, and they have spent billions of years evolving ways to find bacteria, invade them and overcome their defenses.

This extraordinary diversity could be a boon for medicine. Some phages are already being investigated as treatments for bacterial infections, but finding the right virus for the right infection is not straightforward. Scientists need to understand why one phage can infect a particular bacterium while another cannot, why some can overcome bacterial defenses and why some work better than others in the complex environment of the human body.

The idea is no longer purely theoretical. In a new study, published in the journal Science, researchers at Stanford University used AI to design the DNA of a complete phage. When the synthetic DNA was built in the laboratory, it produced a functioning virus—an important demonstration that AI can learn enough about a biological system to design something that works in the real world.

What comes next?

For scientists who study phages, though, the more interesting question is what comes next. If AI can learn enough to build a phage, could it help us understand how these remarkable viruses work?

The Stanford researchers started with ΦX174, a bacteriophage that infects the bacterium E. coli. It is one of the smallest, simplest and best-studied phages known to science.

That makes ΦX174 a good test case for showing that an AI-designed genome can be built and used to create a viable phage. But it’s a long way from the much more complicated phages that scientists are interested in developing as medicines.

Some potentially useful therapeutic phages are five to 50 times larger than ΦX174. They have double-stranded DNA genomes and can contain hundreds of genes. They also have sophisticated molecular machinery that allows them to recognize particular bacteria, reproduce inside them and overcome the many defense systems bacteria have evolved to fight them.

This complexity matters because a phage being developed as a medicine needs to do more than simply reproduce. It needs to find and infect the right bacteria, work effectively in conditions inside a patient and, ideally, remain useful as bacteria evolve resistance.

However, much of phages’ diversity remains a mystery, and this is where AI could become particularly useful.

How AI can help

We are already seeing AI help us look deeper into phage biology. Software such as AlphaFold can predict the 3D structure of proteins from their amino acid sequences. This is particularly valuable for phage proteins whose functions are still unknown. A gene that once appeared to be little more than a mysterious stretch of DNA can now give us clues about the shape of the protein it produces and therefore what that protein might do.

AI does not replace experiments, but it can give us new ideas to test.

At the Becky Mayer Centre for Phage Research at the University of Leicester, much of our work starts with the enormous diversity of phages found in nature. We isolate, sequence and study phages that infect bacteria including E. coli, Klebsiella and Pseudomonas, and many other harmful bacteria that are hard to treat.

Again and again, when we study these phages in detail, we find biology we did not expect. Different phages recognize different parts of bacterial cells. They use different proteins and molecular tricks to overcome bacterial defenses. They can also behave differently when combined with antibiotics or placed in conditions that more closely resemble those inside the human body.

Collections of these naturally occurring phages could provide AI with something extremely valuable: a huge library of real biological solutions. Instead of simply asking AI to design a new phage genome, we could use it to connect three things: DNA sequence, protein structure and what the phage actually does.

For example, could AI identify the genetic features that determine which bacteria a phage can infect? Could it reveal which proteins help a phage defeat bacterial defenses? Could it explain why some phages work particularly well alongside particular antibiotics?

Ultimately, we want to understand what separates a phage that looks promising in a lab experiment from one that actually works in the much more complicated environment of a living patient.

AI might be able to spot patterns in this vast biological diversity that are difficult for humans to see. It could help predict which naturally occurring phages are most promising, identify proteins whose functions we have overlooked, suggest where we might find useful phages and, eventually, propose changes that could make them more effective. But every prediction would still need to be tested in the real world.

The most exciting part

That back and forth between computers and experiments could be the most exciting part. AI can suggest a biological rule or make a prediction, scientists can test it, and the results can then be fed back into the next round of analysis.

The Stanford study raises an important question: Has AI learned general rules about how phage genomes work, or has it simply learned enough about one unusually small and well-understood phage to reproduce successfully? We don’t know yet. The way to find out is to give AI much more biology to work with.

Fortunately, phages provide an almost unimaginably rich source of information. Nature has spent billions of years experimenting with different ways for viruses to infect bacteria, reproduce, evade defenses and survive in different environments. We now have an opportunity to use AI to help us make sense of that enormous natural experiment.

The ultimate goal is not simply to get AI to design new phages, but to use AI to uncover the rules hidden in the incredible diversity of phage biology—and then use those rules to help turn both natural and engineered phages into better treatments for bacterial infections.

Humans love dogs, but petting them can leave some people sneezing and wheezing. About 15% of the world’s population lives with a dog allergy, which can increase the risk of hay fever and asthma. The main culprit is Can f 1, a protein found mainly in dog saliva. Right now, staying away from furry friends or getting regular immunotherapy shots or tablets is the only way to deal with the symptoms.

In a recent study published in The CRISPR Journal, researchers wanted to explore whether it is possible to eliminate dog allergies by targeting them directly at their biological source. Rather than focus on treating human symptoms, they wanted to see whether dogs could be genetically edited to stop producing the allergen altogether.

They used CRISPR-Cas9 to make a tiny, highly precise cut in the Can f 1 gene in dog skin cells and inserted a single extra DNA building block at the site. That small change scrambled the gene’s instructions, switching it off so it could no longer produce the Can f 1 allergen protein. The edited DNA was then introduced into mature eggs from donor beagles. The result was two female beagle puppies, Alfie and Bailey, born from cloned embryos.

When researchers tested their saliva, hair and dander, they found no detectable trace of the Can f 1 allergen protein.

Editing the allergen out

Dog allergies should never be taken lightly. Symptoms can start small—a runny nose, itchy eyes or skin hives—but they can quickly escalate into serious asthma attacks and breathing difficulties that require medical attention. Keeping the home clean and limiting contact with dogs can help, but it rarely solves the problem.

Dog allergens spread easily and stick around on furniture, clothes and in the air, so avoiding them completely is nearly impossible. Therefore, real solutions are needed, not just workarounds, for anyone who lives with a dog or regularly visits family or friends with one.

Most current approaches to dog allergies work by adjusting the human immune response. This study took a different route, going straight to the source of the problem. Researchers started with ordinary skin cells, called fibroblasts, taken from a beagle, then used CRISPR-Cas9 to edit the Can f 1 gene directly.

To turn the edited cells into living dogs, they used a cloning technique called somatic cell nuclear transfer (SCNT). It involved removing the original genetic material from mature dog eggs and replacing it with genetic material from the new Can f 1-mutant cells. Twenty-five embryos were transferred to a surrogate mother dog, which gave birth to pups with the edited genetic change.

Whole-genome and targeted Can f 1 sequencing found that the gene-editing process did not cause any unintended genetic damage to the puppies’ DNA. Because this was a proof-of-concept study aimed at silencing Can f 1 in dogs, the researchers could not be certain whether the genetic change might have any negative effects on the dogs’ health. As a precaution, the puppies were kept under strict veterinary observation from birth.

The team ran biochemical tests on saliva, hair and dander samples from the gene-edited dogs, a normal beagle, a standard poodle and a goldendoodle to directly detect the allergen protein. While the protein was easily detected in the other dogs, including the poodles and goldendoodles advertised as hypoallergenic breeds, it was completely absent in the edited ones.

Then came the real test: finding out whether the beagles caused a reaction. Using a standard skin prick device, the team applied protein extracts from each dog to a volunteer’s arm and observed any reaction.

The results were striking. The volunteer broke out in allergic bumps after exposure to the normal beagle and even the supposed hypoallergenic breeds, but the gene-edited dogs triggered no reaction.

This study demonstrates that it is technically possible to directly shut down the major dog allergen in a living animal. However, before this approach can be used commercially, researchers will need to establish its long-term safety, ensure careful oversight of animal welfare and meet all regulatory requirements.

Designing an effective antibody drug is like searching for the right key in a warehouse of locks. Scientists may begin with millions—or even billions—of antibody candidates, but only a tiny fraction will recognize and bind tightly to the disease target. Identifying those rare candidates has long been one of the biggest challenges in developing antibody medicines.

Boston University researchers have now developed an antibody-specific AI framework that dramatically narrows that search. Rather than building a larger AI model, the team redesigned how AI learns, focusing it on the small regions of antibodies that recognize disease targets.

“Instead of treating antibodies like generic proteins, we designed an antibody-specific language model that learns the fundamental patterns in the regions responsible for antigen binding,” says Diane Joseph-McCarthy, Ph.D., the study’s principal investigator and executive director of Boston University’s Bioengineering Technology & Entrepreneurship Center. “That focused approach helps researchers identify the most promising therapeutic candidates before they ever enter the laboratory.”

The study was published today in the journal Communications AI & Computing. Researchers found that the approach improved predictions of antibody binding strength, known as binding affinity, by as much as 27% while requiring far fewer computational resources than many existing antibody AI models.

Why antibodies challenge AI

Artificial intelligence has transformed biology by identifying patterns across millions of protein sequences. Like ChatGPT predicts missing words, protein language models predict masked amino acids to learn the “language” of proteins.

For most proteins, randomly hiding amino acids throughout a sequence is an effective training strategy because biologically important information is distributed across the molecule. Reconstructing the missing pieces helps the model discover the patterns that determine protein structure and function.

Antibodies, however, are different.

The regions responsible for recognizing viruses, bacteria and cancer cells continually evolve so the immune system can adapt to new threats. That diversity makes antibodies extraordinarily powerful—and more difficult for AI to model.

Most of an antibody serves as a structural scaffold. The information that determines what an antibody recognizes and how tightly it binds is concentrated within six tiny loops called complementarity-determining regions, or CDRs.

“Think of an antibody like a screwdriver,” says John Misasi, M.D., a study co-author and assistant professor of virology, immunology and microbiology at Boston University’s Chobanian and Avedisian School of Medicine. “It doesn’t matter whether it’s long or short—the shape of the tip determines what kind of screw it fits. The CDRs are like that tip: they determine which target the antibody recognizes.”

Teaching AI the biology that matters

Recognizing that antibodies break many of the assumptions behind general protein language models, the researchers redesigned the training process around antibody biology instead of treating every amino acid as equally important.

The model focused its learning on the CDRs—the regions directly responsible for recognizing disease targets—and was trained using more than 1.6 million naturally paired antibody heavy and light chains that together form the binding site. During training, the researchers deliberately masked up to half of the amino acids within the CDRs while leaving most of the surrounding antibody structure intact, repeatedly challenging the AI to reconstruct the regions most critical for binding.

“A lot of AI research has focused on building larger models,” says Ioannis (Yannis) Paschalidis, Ph.D., a co-author of the study and director of Boston University’s Hariri Institute for Computing. “We asked a different question: How can we teach the model the biology that matters most? That turned out to be a much more effective strategy.”

The result was a smaller, more focused model containing about 600 million parameters that matched or outperformed much larger antibody language models on multiple benchmark tests. It improved binding affinity prediction by as much as 27% across datasets containing more than 90,000 engineered antibody variants targeting six different antigens. By training on millions rather than billions of antibody sequences, it required substantially less computational effort.

“One of the exciting findings is that we didn’t need a larger model or vastly more data,” says Paschalidis. “That’s similar to what’s been observed with human-language AI models, where smaller domain-specific models trained on high-quality data can often outperform much larger, more general ones.”

The approach emerged from a convergent research effort that brought together expertise in artificial intelligence, immunology, structural biology and experimental science—not simply to apply AI to biology, but to redesign how AI learns using biological knowledge.

From prediction to prioritization

The study addresses one of the biggest bottlenecks in antibody discovery: deciding which candidates to test. Even small changes to an antibody’s sequence can create trillions of possible variants—far more than laboratories can realistically evaluate experimentally. The model helps narrow those possibilities before laboratory testing, reducing unnecessary experiments and accelerating antibody optimization.

“Knowing not just whether an antibody binds, but how strongly it binds, gives researchers a much better starting point for deciding which candidates to move forward,” says Misasi, core faculty at BU’s National Emerging Infectious Diseases Laboratories (NEIDL). “If a computer can narrow millions of possibilities down to the few hundred most promising candidates, that saves an enormous amount of time, labor and cost in the laboratory.”

Beyond selecting antibody candidates for testing, the approach could improve antibody engineering. Researchers could use the model to predict which sequence changes are most likely to strengthen existing antibodies, helping optimize therapies against evolving viruses or other disease targets before moving those designs into experimental testing.

“If another infectious disease outbreak occurs, we’d like to identify promising antibody candidates as quickly as possible,” says Misasi. “Computational tools like this could help us find those candidates sooner and even suggest how existing antibodies might be adapted as viruses change over time.”

Looking ahead

The findings suggest that biologically informed AI may offer a more effective path for antibody discovery. Beyond this study, the researchers believe biologically informed AI could ultimately transform therapeutic antibody development by helping scientists better understand antibody-antigen recognition, prioritize candidates for laboratory testing and design more effective therapies.

“Understanding how antibodies recognize their targets is fundamental to developing better antibody therapies, diagnostic tests and vaccines,” says Joseph-McCarthy. “By teaching AI the biology that matters most, we hope to give researchers better tools to discover, optimize and ultimately design the next generation of antibody therapies.”

Like a delivery driver navigating crowded city streets, a gene-therapy-toting lipid nanoparticle faces a gauntlet of potential detours on its journey toward a cell’s nucleus. First, there’s entering the cell’s plasma membrane; then navigating around organelles like the Golgi apparatus, mitochondria and endoplasmic reticulum—all destinations that can errantly absorb the particle’s payload, rendering it ineffective at best or harmful at worst. And that’s all before the particle even enters the nucleus and successfully makes a genetic change.

As part of an effort to improve nonviral genetic editing technologies, with particular interest in treating genetic eye diseases, a group of University of Wisconsin–Madison researchers developed a platform to screen nearly all genes in the human genome—more than 19,000 genes—linking disruption of individual genes to improved genetic outcomes in human cells.

The work could provide a roadmap for identifying potential detours in different cell types when gene editors like CRISPR-Cas9 are applied to cells in the lab or to tissues in the body. Such knowledge could inform the design of new strategies, such as pretreatments to temporarily mute those genes, to enhance the effectiveness of gene therapies, mRNA vaccines and immunotherapies.

The researchers, led by biomedical engineering professor Krishanu Saha and representing the Wisconsin Institute of Discovery, the McPherson Eye Research Institute, the School of Medicine and Public Health, the Waisman Center and the School of Pharmacy, published their results in a paper in the journal Nature Communications.

From genome-wide screen to six genes

“We essentially knock out each gene one by one in different cells. It’s a massive undertaking if you were to try to do that in 19,144 different cell cultures, for instance,” says Saha. “Here we do it in essentially a few dishes of cells that have 100 million cells in them. Then we use a precise sequencing strategy to find which genes matter the most, essentially the needles in this genomic haystack.”

Over the course of the six-year project, the researchers culled a list of 19,144 genes down to 26 that, when knocked out, increased the editing efficiency by the Cas9 gene-editing protein. They further trimmed their list to six, then validated the top two and tested the effects of knocking them out on editing efficiency in human model cell systems.

Depleting either of the two genes (GJB2 and BET1L) improved editing efficiency by more than six times using base editors, so-called “CRISPR 2.0,” which chemically change individual bases in DNA rather than making cuts in the genome like the traditional CRISPR-Cas9 editor does.

How and why these particular genes impede editing remains an open question—and a larger one for research groups across the field to explore.

“After we add Cas9, if we see more of the proper edits in a cell population where a candidate gene has been knocked down, then that means the gene is somehow blocking the gene editing process,” says Saha lab alumna Shivani Saxena (Ph.D. BME), now a postdoctoral researcher at the University of California, San Francisco and one of two first authors on the paper.

Meha Kabra, a research scientist in the UW–Madison School of Medicine and Public Health’s Department of Pediatrics, is the other.

Testing the findings in retinal cells

When testing their approach on patient-derived retinal cells carrying the genetic eye disease Leber congenital amaurosis (LCA), the researchers found that inhibiting the top two genes from their screening increased editing efficiency by more than three-and-a-half times.

Saha and fellow study authors David Gamm (professor of ophthalmology and visual sciences), Bikash Pattnaik (associate professor of pediatrics) and Shaoqin “Sarah” Gong (Vilas Distinguished Professor of ophthalmology and visual sciences) are the UW–Madison project leads of the CRISPR Vision Program, applying nonviral gene-editing technologies to treating inherited eye diseases like LCA.

This consortium developed a personalized CRISPR drug on demand in 2025.

Broader clues for nanoparticle delivery

The other genes flagged by the screen could be potentially more relevant to other cell types and diseases, Saha notes, while the group’s broader approach could inform delivery strategies for emerging mRNA cancer vaccines, CAR T-cell therapies and more.

“My hope is that the field picks up the methodology and helps understand what cellular mechanisms, pathways, are important for the delivery of really anything that could be packaged into lipid nanoparticles,” says Saha. “And in that way, I believe the impacts could be beyond just genome editing.”

Scientists have crossed an important line in biological engineering. In a recent study, researchers used artificial intelligence to design complete sets of genetic instructions for bacteriophages, viruses that infect bacteria. Some of those computer-generated designs produced working viruses when they were built and tested in the laboratory.

These viruses were designed to infect bacteria rather than people. The researchers worked with bacteriophages that target E. coli, and one of the AI models they used, Evo 2, was developed with safety restrictions. Its developers excluded viruses that infect organisms such as animals, plants and humans from the data used to train it.

Even so, the experiment raises an important question: If AI becomes increasingly able to help design biological systems, are our safeguards developing fast enough?

As engineers working in vaccine manufacturing and public health security, we see good reasons to continue this research. Bacteriophages could potentially be used to treat bacterial infections, including those that are becoming harder to treat with antibiotics. But advances in biological design also create new challenges for biosecurity.

Genome language models work in a broadly similar way to AI language models that learn patterns in text. Instead of learning from words and sentences, they learn patterns in genetic code. Evo 2, for example, was trained on trillions of DNA building blocks taken from many different forms of life.

This gives researchers a powerful new way to study biology and generate possible DNA sequences. It also means that part of biological design can now happen on a computer, before anything has been made in a laboratory.

An AI-generated DNA sequence is still several steps away from becoming a real biological threat. A digital design has to be turned into physical genetic material, assembled correctly and tested under the right laboratory conditions. Each of those stages also creates an opportunity to reduce risk.

One safeguard can be built into the AI model itself. The developers of Evo 2 deliberately removed viruses that infect humans and other similar organisms from its training data for safety reasons. Their research found that the model performed poorly when tested on proteins from viruses that infect humans.

Such precautions will need to be tested and strengthened as AI tools for biology become more capable. Researchers have already argued that AI and biosecurity need to be considered together, so that safeguards develop alongside new capabilities.

Another checkpoint comes when digital DNA is turned into physical DNA. Companies that manufacture DNA to order can check both the requested genetic sequence and the person or organization ordering it for potential security concerns. Researchers working on DNA synthesis screening have argued for common international approaches as this technology becomes more widely available.

Screening may also need to become more sophisticated. A system designed mainly to recognize DNA that resembles known dangerous pathogens may struggle if AI produces something genuinely new. Recent research on identifying “sequences of concern” has therefore considered what a genetic sequence might do, as well as whether it resembles one already linked to a dangerous organism.

Laboratories and research institutions provide another layer of protection. Potentially risky research can be assessed before experiments begin, including questions about how biological material will be contained, who will have access to it and whether the expected benefits justify the risks.

Research funders can also influence how security risks are managed. UK Research and Innovation (UKRI), for example, has established a Trusted Research and Innovation team to help researchers and institutions identify and manage security risks in collaborative research.

Public health preparedness belongs in this discussion too. If biological design becomes faster and more accessible, health authorities will need to be able to detect and investigate unusual outbreaks quickly. The Metagenomics Surveillance Collaboration and Analysis Program (mSCAPE), led by the UK Health Security Agency, analyzes genetic material from samples to help detect and track emerging pathogens.

The program is not specifically designed to find AI-created viruses. But its approach is relevant because scientists do not have to know exactly which pathogen they are looking for in advance. Systems capable of detecting unusual or previously unseen organisms could become increasingly valuable as biological technologies develop. Preparedness also means being able to develop diagnostic tests, treatments and vaccines quickly when a new threat appears.

In our view, one of the biggest weaknesses in global biosecurity is that countries are not equally prepared. Some already have established systems for identifying and responding to biological risks, while others are still developing them. We see a similar gap in vaccine regulation, where countries vary considerably in their ability to assess new products. Biological threats can cross borders, so effective preparedness will depend on strengthening these capabilities internationally.

There are difficult trade-offs. Heavy restrictions can slow useful research, including work aimed at understanding disease or developing treatments. Weak safeguards can allow scientific capability to develop faster than the systems intended to prevent misuse.

Open science makes this harder. Evo 2 was released openly, including the code and technical information needed for other researchers to use and develop it. Openness can accelerate discovery and widen access to powerful research tools. But once such tools are widely distributed, controlling how they are used becomes harder.

No single safeguard can deal with all of these risks. Biosecurity will need to operate at several stages, from how biological AI models are developed and released to DNA screening, laboratory oversight and public health preparedness.

The recent bacteriophage study does not show that AI can design a pandemic virus. What it shows is significant enough: AI can already help design complete viral genomes that work when they are physically created.

We have an opportunity to decide what responsible safeguards should look like while this technology is still developing. Waiting for more dangerous capabilities to emerge would leave scientists, governments and public health systems trying to develop protections after they are already needed.

AI is beginning to change what humans can design in biology. Our ability to govern that power needs to develop alongside it.

Last year, we described Agrilo, an affordable “colorimetric” soil-testing system developed through research at MacEwan University. Using a smartphone camera to interpret color changes, Agrilo measures soil nutrients and can provide farmers with timely guidance to support fertilizer decisions.

Our goal was to make practical soil information more accessible—so farmers could use fertilizer more efficiently, improve crop productivity and reduce unnecessary environmental impacts.

But affordability is only one part of successful agricultural innovation. For a technology to create value across different regions and farming conditions, users must also be able to learn its procedures confidently and apply them consistently.

So we developed Agrilo VR, a virtual reality training platform designed to support the wider adoption and standardized use of the soil-testing system.

Quality assurance on the farm

Agrilo uses a structured seven-stage testing protocol that guides users through soil preparation, reagent mixing, timing and color measurement. As with laboratory and field-based analytical methods generally, following a consistent procedure helps ensure we can compare results reliably across users and locations.

This is especially important when testing is decentralized. Commercial laboratories achieve consistency through standardized procedures, trained personnel and controlled instruments. Agrilo is designed to bring similar principles of guided testing and quality assurance closer to the farm, allowing soil information to be generated where and when it is needed.

Agrilo VR allows farmers, students and field operators to practice the complete protocol before conducting a field test. The platform reinforces the sequence, timing and decision points within the procedure, helping users build confidence and consistency through repetition.

The Agrilo VR platform is an international collaboration between our team at MacEwan University and the research group led by Dr. Abel Méndez Porras at Tecnológico de Costa Rica’s San Carlos campus.

The barrier was never only cost

A 2025 survey of small farmers in Kentucky found that cost was the most-cited barrier to adopting the tools of precision agriculture.

However, complexity, doubts about profitability, privacy concerns, low confidence and time demands were also important. Only about one-quarter of respondents had adopted any precision tool.

A meta-analysis published in Precision Agriculture reached a similar conclusion: Technical literacy and advisory support are among the strongest predictors of adoption.

Reducing the cost of soil testing to approximately $10 can make routine testing considerably more accessible. Combining affordability with structured training, standardized procedures and ongoing user support can address several additional barriers to adoption at the same time.

The benefits of virtual training

Agrilo VR runs on a Meta Quest headset and guides trainees through the soil-testing protocol in the correct sequence. Each stage must be completed before the trainee advances, reinforcing the same structured workflow used in the Agrilo system.

A custom liquid simulation shows reagent mixing and color development as they would appear on a workbench. This is particularly valuable for colorimetric analysis, where recognizing the timing and appearance of a reaction is part of developing practical proficiency.

With users’ consent, the platform’s cloud-based component can collect aggregated training data showing where users request additional guidance or repeat an activity. These data can help educators identify training needs while allowing the team to continuously improve both the learning experience and the supporting protocol.

Virtual training also provides a resource-efficient environment for practice. Trainees can repeat the procedure as often as necessary without consuming materials.

A further benefit is standardization. The platform is being evaluated and demonstrated across locations that include KwaZulu-Natal in South Africa, Medicine Hat in Alberta, Costa Rica’s San Carlos and Nairobi. This helps support consistent testing practices and more reliable comparison of results across sites.

What can simulation contribute?

Simulation has a long record in procedural training, including aviation, health care and technical education. Research in these fields suggests that well-designed simulation can support knowledge development, procedural confidence and repeated practice.

We do not claim that virtual training is universally superior to conventional instruction. Its biggest advantage is scalability: The platform can deliver the same structured learning experience across multiple locations while reducing the need to send an experienced trainer to every site.

Our next step is to evaluate how effectively this approach supports learning, user confidence and consistent performance when trainees move from the virtual environment to soil testing in the field.

Virtual training, field experience

As co-founders of the company developing Agrilo, we recognize the importance of evaluating the technology transparently—distinguishing demonstrated benefits from those that require further study.

Research on virtual reality in agricultural training is growing, although the evidence base remains less developed than in fields such as aviation and health care. This creates an important opportunity to evaluate how virtual practice contributes to user confidence, procedural consistency and performance under field conditions.

Agrilo VR is designed to complement, rather than replace, practical field experience. Virtual training can introduce the testing sequence, reinforce key decisions and allow users to repeat procedures in a controlled setting. Field experience then adds the realities of local soils, weather, lighting, cropping systems and other environmental conditions.

Combining these approaches may offer the strongest training model: Users first develop familiarity and confidence through structured simulation and then apply and refine those skills in real agricultural settings.

Building the capacity to adopt agricultural technology

Canada’s productivity challenge is partly a technology adoption challenge. Developing a useful and affordable tool is essential, but successful adoption also depends on whether people can understand it, trust it and integrate it confidently into their work.

Agrilo was developed to make actionable soil information more accessible. Agrilo VR extends that objective by creating a scalable way to introduce the procedure, reinforce consistent practices and prepare users for field application.

Through our collaboration with Porras and his research group, we are exploring a broader model for agricultural innovation: combine affordable diagnostics with accessible training, field experience and continuous improvement.

The instrument and the user should not be viewed separately. When affordable technology is supported by well-designed training, farmers and field operators are better positioned to generate useful information and translate it into informed agricultural decisions.

Researchers at the Department of Energy’s (DOE) Oak Ridge National Laboratory (ORNL) and the University of Tennessee (UT), Knoxville, have made an advance by using a natural substance from fungi called chitin to create hydrogels that are significantly stronger and tougher. This new material could lead to safer medical devices, longer-lasting coatings and other innovations that improve everyday technology and health care.

Fungal chitin is a natural, fibrous substance found in the cell walls of fungi. Hydrogels are flexible, water-rich materials used in medical implants and contact lenses, as well as soft robots that can bend, flex and even self-heal in ways that mimic natural, living systems.

The study, published in the International Journal of Biological Macromolecules, offers a fresh approach that could transform those types of materials used in medicine and engineering.

How the team refined fungal chitin

Yue Yuan of the Center for Nanophase Materials Sciences (CNMS), a DOE Office of Science user facility at ORNL, said the research project built on previous work she completed as an intern at ORNL, including a 2021 paper. As an ORNL Laboratory Directed Research Development (LDRD) Distinguished Staff Fellow, Yuan said funding, early-career development programs and the laboratory’s collegial research environment enabled her to convene experts from multiple fields to assess fungal chitin as a promising alternative to traditional crustacean sources.

The researchers extracted chitin from three types of fungi using a gentle treatment that preserves the tiny particles’ structure. Unlike chitin from shrimp or crabs, fungal chitin has a much lower allergenic potential and does not require extensive processing to remove unwanted chemicals and heavy metals.

The new method uses a mild alkaline pretreatment and a ball-milling process to break down the material into a fine powder. This approach improves the material’s performance without the extra energy use or harsh chemicals typical of traditional techniques.

Yuan brought together a multidisciplinary team that included Tomás Rush and Yunqiao Pu from the ORNL Biosciences Division; John Lasseter from CNMS; and Toby Nelson, Kehao Ren, Lu Wang and L.P. Tharika Nirmani from UT.

Reflecting on the team’s collaborative approach, Rush, an ORNL fungal biologist, said, “Fungi offer an untapped reservoir of bioproduct potential, a perspective that guided our exploration beyond the usual model species.”

Nelson, a key figure in advanced materials design at the UT-Oak Ridge Innovation Institute (UT-ORII), added, “This collaboration integrates biosourced fiber processing with advanced materials design, demonstrating the power of a well-connected research community.”

Imaging confirms the material’s structure

To validate the material’s structure, the team used advanced electron microscopy techniques. Lasseter said, “Our ability to image these delicate biopolymers at low voltages without metallic coatings has been crucial in capturing their true shape and structure and confirming the chitin’s integrity.”

Mechanochemistry drives material innovations

The research also draws attention to the benefits of mechanochemistry in modern materials engineering. Yuan said using a ball-milling process to produce submicron- to nanoscale particles demonstrates how desired material properties can be achieved without relying on severe chemical reactions or high temperatures.

“When materials are designed for applications such as contact lenses, water-cleanup systems or wound dressings, every improvement in durability and functionality is crucial,” Yuan said. The ability to tailor the molecular architecture of fungal chitin by selecting different species opens the door to a vast range of material properties.

Yuan said the choice to pursue fungal sources over traditional crustacean sources was motivated by several factors. The fungi produce unique chemical profiles that provide a platform for designing materials from the ground up.

The availability of a variety of fungi enables researchers to tap into a warehouse of molecules to address different engineering and medical challenges. The consistent fermentation process facilitates fundamental research, including sample characterization using ORNL’s unique neutron scattering tools.

In the Midwestern summer, humans and plants have to breathe through the heat and humidity. Researchers hope that a retooled crop plant—one with an improved ventilation system within its leaves—could thrive while avoiding drought stress by reducing “crop sweat.” Thanks to next-generation imaging technology powered by Argonne National Laboratory’s particle collider beamline, a team of scientists now has precise schematics of the leaf’s interior and can refine approaches to breeding hardier crop plants.

The research, published in Plant Physiology, was led by postdoctoral researcher James Fischer in the laboratory of plant biology and crop sciences professor Andrew Leakey at the University of Illinois Urbana-Champaign. Their work represents the first detailed look at how the pores on the leaf surfaces of sorghum, a highly productive and resilient grass crop, connect to the air pathways, photosynthetic centers and veins beneath.

“There are connections between each component of the leaf . . . it’s a highly organized system,” Fischer said. “We are really defining the leaf beyond just carbon dioxide goes in, water comes out.”

The cost of crop sweat

The motivation for the study came from a fundamental trade-off that every land plant faces. During photosynthesis, plants take in carbon dioxide through tiny pores in their leaves called stomata. For each molecule of CO₂ that enters the plant through the stomata to be captured by photosynthesis, 300–400 water molecules escape from the interior of the leaf through the stomata to the atmosphere.

Mapping the leaf in 3D

A hot, bright day is good for photosynthesis but terrible for water loss as corn and other crops “sweat it out.” This loss of water is why crops only grow in times and places with adequate rainfall. One of the research team’s primary goals is to develop crop plants that handle this trade-off better.

“We’re trying to minimize how many water molecules escape while CO₂ is going into the leaf to be captured by photosynthesis,” Leakey said. Leakey is also the director of the Center for Advanced Bioenergy and Bioproducts Innovation, which developed plants with fewer stomata that were examined in the new study. “We’ve engineered plants that have fewer stomata and demonstrated proof of concept for that goal . . . But now we’re really interested in quantifying how easily a CO₂ molecule can work its way through the interior ventilation system of the leaf to the location where it actually gets captured by photosynthesis.”

To answer this question, the team needed what any engineer needs: schematics of how the system currently works. For most plants, such detailed knowledge of the internal anatomy of a leaf simply doesn’t exist. The structures inside can only be truly understood in three dimensions, so traditional two-dimensional images from microscopes are not ideal.

Fischer and Leakey seized the chance to apply a high-resolution three-dimensional method called micro-computed tomography that could reveal the path of airflow through an intact, living sorghum leaf. They formed a collaboration with Yale University plant physiological ecology professor Craig R. Brodersen and Guillaume Théroux-Rancourt, assistant director of the agronomy research company Biopterre. Brodersen and Théroux-Rancourt had previously helped innovate microCT imaging for plant tissue samples.

MicroCT works on the same principle as the CT scan one might get as part of a medical diagnostic process. A series of X-ray images are taken and then combined to form a three-dimensional view of a structure or tissue. Imaging smaller structures requires stronger and more focused X-ray beams than most CT machines can generate. However, these types of X-rays are emitted as a byproduct—an incredibly valuable one—of particle accelerator facilities used for atomic physics research. To image their sorghum samples, the research team obtained time at Argonne National Laboratory’s Advanced Photon Source.

An unexpected internal adjustment

Once the images were acquired and assembled, with the help of machine learning, into high-resolution three-dimensional models, the researchers were surprised by what they found when comparing plants with more or fewer stomata.

“We thought it was something that would be a challenge for us from the engineering perspective . . . reducing the number of stomata was going to mean that there’s a longer, more tortuous path to get CO₂ where it’s going,” Leakey said.

Instead, they found that in leaves with fewer stomata, air spaces beneath them were larger, resulting in equivalent conductance of CO₂ through the leaf. This meant that water loss could be reduced without a parallel loss of CO₂ distribution to photosynthetic cells. This is good news for further development of drought stress–avoiding plants.

“There are loads of compensatory mechanisms where we’re doing this kind of engineering that make our life difficult,” Leakey said. “This was the first one that made our life easier.”

Clues from the upper surface

Another surprising result was that stomata on the upper surface were positioned over veins rather than between them, as usually occurs to allow more space for internal air spaces to channel CO₂ to photosynthetic cells.

“This finding opened our eyes to how the two leaf surfaces may contribute differently to uptake of CO₂ for photosynthesis and cooling of the leaf through water loss,” Leakey said. “So we want to know, why is it organized like that and how can we leverage that knowledge to make more water use–efficient crops that can avoid drought-induced yield losses? That’s the exciting next step.”

Fischer and Leakey are considering how future studies could answer these and other questions, bringing crop development closer to water-use efficiency.