As federal policymakers weigh potential changes to how biomedical research is funded and regulated in the United States, a Virginia Tech scientist highlights the importance of preserving the nation’s ability to turn discovery into life-saving therapies.

In a commentary published this week in Nature Biotechnology, Robert Gourdie, professor at the Fralin Biomedical Research Institute at VTC, notes that well-intentioned but overly restrictive policies could inadvertently undermine the technology-transfer ecosystem that has driven decades of U.S. leadership in biomedical innovation.

He emphasizes that the key to making the process work is strong transparency and careful oversight of conflicts.

“The United States didn’t become a global leader in medical innovation by accident,” Gourdie said. “It happened because we built systems that allow discoveries made with public funding to move efficiently from academic laboratories into the clinic and the marketplace, where they can benefit patients.”

Biomedical research funded by taxpayers is being examined more closely by federal agencies, policymakers, and the broader scientific community, with growing attention on research integrity, reproducibility, and scientists’ financial ties to industry—concerns that have prompted calls for tighter federal oversight.

At the same time, unclear signals about future National Institutes of Health (NIH) funding and policy direction have raised questions about whether reforms could unintentionally weaken the research system they aim to protect.

Together, these pressures are fueling a national conversation about how to safeguard scientific integrity without undermining the innovation pipeline that turns discovery into patient care.

Gourdie emphasized that technology transfer—the process through which universities license inventions and collaborate with private-sector partners—is an essential but often misunderstood component of the NIH’s mission to improve health, lengthen life, and reduce disability.

“Conflicts of interest are an inherent part of innovation,” Gourdie said. “The question is not whether they exist, but whether they are disclosed, overseen, and managed in ways that protect scientific integrity.”

In the article, Gourdie pointed to China’s rapid expansion of its biomedical innovation ecosystem, with policies modeled in part on the U.S. system, pairing strong government investment in research with incentives for commercialization and business creation. International intellectual property data show that China now leads the world in patent filings, highlighting what Gourdie describes as an increasingly competitive global landscape.

“Other nations are not retreating from translation—they are accelerating it,” Gourdie said. “If the U.S. weakens the mechanisms that connect discovery to deployment, we risk ceding both economic and biomedical leadership.”

Under current federal rules, researchers disclose significant financial interests, while universities implement oversight and management plans designed to ensure objectivity in the design, conduct, and reporting of research.

Gourdie argues that this framework has enabled productive academic-industry partnerships while maintaining public trust, and that abandoning it could have unintended consequences.

Gourdie also challenges the assumption that commercialization erodes research rigor. Instead, he said, translational work often introduces additional layers of scrutiny through investors, regulators, and independent validation by contract research organizations.

“In many cases, translational science subjects data to more external review, not less,” Gourdie said.

Gourdie is transparent about his own role in academic entrepreneurship. He is a co-founder and shareholder of several biotechnology companies—including The Tiny Cargo Company, Acomhal Research, and Xequel Bio—that are developing therapies originating from his academic laboratory, with related intellectual property licensed through universities.

Gourdie frames technology transfer as an ethical obligation to the public. Taxpayers, he said, support biomedical research with the expectation that discoveries will ultimately improve health and save lives.

Looking ahead, Gourdie proposed a constructive way to reinforce U.S. leadership in biomedical innovation without sacrificing public trust.

He suggests dedicating a small fraction of royalties from federally funded patents to a national “sovereign wealth”-style fund that would be reinvested in basic and translational research. By routing 1% to 3% of licensing income back into a transparent, publicly governed fund, he said policymakers could strengthen long-term support for NIH-funded science while preserving the incentive structure that has made U.S. technology transfer so effective.

“The greater ethical failure is allowing promising discoveries to languish in academic journals,” said Gourdie, who is also a professor in the Department of Biomedical Engineering at Virginia Tech. “If we have the ability to move knowledge into real-world use responsibly, we have a duty to do so.”

Rising day and night temperatures are threatening rice, wheat, and maize production by disrupting plant growth, grain filling, and grain quality, putting global food security at risk. Precision breeding and genome editing offer ways to reprogram plant clocks, optimize flowering and panicle architecture, and protect grain quality under heat stress.

The world’s “cereal bowl,” or the production of rice, wheat, and maize, is under the dual challenges of a surging human population and a rapidly warming climate. As global temperatures rise, agricultural yields are failing to keep pace with demand, with scientists estimating that the rate of yield increase for these three staple crops must rise by a staggering 37% to ensure food security by 2050.

A particularly insidious and overlooked threat is the rise in high night temperatures, which are increasing nearly twice as fast as daytime temperatures. This nocturnal heat disrupts the delicate internal rhythms of plants, causing “source-sink” imbalances where the energy produced during the day is wasted through excessive respiration at night, ultimately leading to stunted grains and lower grain quality.

In a recently published review in Trends in Plant Science, scientists from the International Rice Research Institute and the Max Planck Institute of Molecular Plant Physiology analyzed how understanding the genetic regulation of flowering, plant architecture, and grain filling can provide a roadmap for developing climate-resilient varieties with sustained yield and grain quality. The authors argue that while the Green Revolution of the 20th century relied on stable, cooler climates, the current era requires precision breeding strategies to overcome the stagnation of crop production observed in low-income, food-deficit regions.

Reprogramming the plant’s biological clock

One of the most innovative solutions discussed in the review involves manipulating the plant’s circadian rhythm to help crops escape the worst of the heat. By identifying and tweaking “thermometer genes,” scientists can develop varieties that bloom earlier in the morning before temperatures peak.

In rice, for example, the gene OsMADS51 has been identified as a key factor in conferring thermotolerance during the critical heading and grain-filling stages. Similarly, in maize, researchers are targeting the “evening complex,” a group of genes including ZmELF3 and ZmLUX, which coordinates flowering and adaptation across different latitudes. By modifying these clock genes, breeders can ensure that the delicate process of flowering is promoted under heat stress.

Building a more efficient panicle architecture

Beyond timing, the inflorescence architecture of the plant can be re-engineered to maximize efficiency of grain number per panicle. The review highlights the potential of genes like DEP1 in rice, which produces dense, erect panicles that create a more favorable microclimate for the plant. These erect structures allow for better light distribution and photosynthetic rates, even under heat stress.

Furthermore, scientists are investigating the vascular highways of the plant, the bundles that transport sucrose to the developing grains. By identifying genes like SPIKE, GIF1, SPL14, and APO1-HI1, which increase the number of primary branches and vascular bundles, researchers can improve the “sink strength” of the grain, ensuring that nutrients are delivered effectively even when high temperatures threaten to disrupt the flow.

Studying the shape of tissues and organs is critical to understanding how they are formed. Embryonic development happens in three dimensions, but many studies are limited by the use of two-dimensional approaches and images to describe three-dimensional processes. To overcome this challenge, researchers at EMBL Barcelona have created LimbLab—an open-source pipeline made for three-dimensional visualization and analysis of growing limb buds.

A new tool for 3D limb research

The platform was primarily designed to study mouse limb development, but the concept can be useful for any researcher working with complicated volumetric imaging data. The researchers describe the platform in a new study published in the journal BMC Bioinformatics.

“We developed LimbLab because we realized that current tools miss crucial aspects of embryonic development and are not designed specifically for developmental biology,” said Laura Aviñó-Esteban, first author of the work and Ph.D. student in EMBL’s Sharpe Group.

Researchers studying limb development need specific software tools for their work. For instance, researchers may need to assign a developmental age to certain samples based on their visible features, or they might have to align or morph images of samples to allow accurate and consistent comparisons.

These tools exist in 2D, like eMOSS and LimbNET, but not in 3D. LimbLab bridges this gap by enabling 3D visualization of gene expression patterns. It achieves this through a modular workflow in Vedo, an open-source Python library developed at the Sharpe Lab that enables seamless 3D analysis and high-fidelity and aesthetic rendering of meshes and volumes.

Inside the LimbLab pipeline workflow

First, the pipeline cleans the raw volumetric data obtained from a microscope, removing noise and artifacts. Then it extracts information about tissue surfaces to build computational structures called “meshes,” which are computationally efficient to work with.

After this, the pipeline analyzes the sample to give it a developmental age and aligns or morphs the sample with a reference model. LimbLab also provides advanced visualization methods that help researchers to explore and present gene expression in full 3D. Each step is logged and standardized, which improves reproducibility.

Broader impact beyond limb development

The importance of this work goes beyond limb development. Many areas of biological research, like those using organoids, tumors, or engineered tissues, depend on 3D volumetric imaging. LimbLab shows how a specialized pipeline can organize messy 3D data and make it clear and aligned.

It also underscores the importance of reproducibility in imaging research, where small differences in processing can result in very different biological interpretations. LimbLab is also a proof-of-concept: it shows how special computational tools can change the way we analyze 3D biological data.

LimbLab is not only a technical improvement, it is a step towards making developmental biology research more quantitative, reproducible, and accessible. The pipeline is open source, easy to install, and has full documentation. While the present version is optimized for mouse limb buds, the researchers plan to adapt it for other species and tissues to help answer bigger questions about regeneration and evolution.

Plant-based food as an alternative to meat is high on the agenda today, and mycoprotein (fungal protein) in particular has come into focus in recent years. A new doctoral thesis from the University of Borås in Sweden, has investigated how mycoprotein and its minerals are digested in the body.

What happens to the food we eat as it passes through our digestive system? How well are mycoproteins, a type of fungal protein from edible filamentous fungi, digested? Are their nutrients really accessible to the body? And what is the importance of different cultivation methods and media for cultivation?

Researcher Ricky Wang delved into these issues and recently presented his findings in his doctoral thesis “In vitro Gastrointestinal Fate of Edible Filamentous Fungi: Protein and Mineral Digestibility for Food Applications.”

In his project, Wang focused on two main issues:

• Can mycoprotein of filamentous fungi be digested as effectively as traditional food proteins such as fish and chicken?
• Are minerals more easily accessible when we eat fungal protein compared to other plant-based ingredients, which often contain antinutrients such as phytate, a substance that often limits mineral absorption in other plant-based ingredients, and which is found naturally in seeds, nuts, legumes and cereals, for example?

Mycoproteins contain amino acids that meet human nutritional needs. In addition, mycoprotein is digested as efficiently as chicken and fish.

“Filamentous fungi have great potential as a sustainable protein source. With the right cultivation techniques, they could become a key to more resilient and climate-smart food production systems,” explained Wang.

He also found that the iron present in mycoprotein can be absorbed more easily, due to the absence of phytate.

Simulated digestion

In the project, filamentous fungi were grown in a bioreactor. Wang then analyzed the nutritional values using a standardized method, INFOGEST, that simulates digestion in the mouth, stomach, and intestines under laboratory conditions.

The results provide an important insight: although filamentous fungi have great potential as a sustainable source of protein, their nutritional performance is strongly influenced by fungal variety and cultivation strategy.

“For example, for fungi that are cultivated on a winery side stream, their nutritional value could be lower compared to if the medium consisted of only sugar. Optimizing these conditions could unlock their full potential, paving the way for more resilient and sustainable food systems,” said Wang.

The results of the project should be of interest to the food industry, especially to actors focusing on mycoprotein as an alternative in food production.

“Research on mycoprotein for food production is relatively new, and more research is needed to fully understand its potential,” according to Wang.

The project supports the UN Sustainable Development Goals, in particular Goal 2: Zero hunger and Goal 3: Good health and well-being.

Implementing novel management practices in dairy farming, one of the commonwealth’s major agricultural industries, could help alleviate a large source of both nutrient pollution and greenhouse gas emissions, according to a multidisciplinary team led by researchers at Penn State. Those practices include continuous cover—keeping fields covered with vegetation year-round—and anaerobic digestion—a microbial process that converts manure and plant organic matter, called biomass, into biogas—a combustible fuel consisting mostly of methane.

Simulating a modern Pennsylvania dairy

To evaluate the effectiveness of these practices, which the team refers to as Grass2Gas when they are combined, the researchers conducted a study involving the simulation of a typical large Pennsylvania dairy farm. The team compared the environmental footprint of the farm employing different scenarios of Grass2Gas with that of a dairy farm under traditional management.

Using life cycle assessment, which accounts for every practice and resource used by a farm from origination to disposal, the researchers found that using anaerobic digestion of manure with grassy biomass could reduce the carbon footprint of milk production by over 20% on average, compared with a typical dairy farm. But their modeling suggested that for the Grass2Gas approach to diminish overall water and air pollution, adjustments and trade-offs may be necessary.

They published their findings in Environmental Science & Technology. The work is the latest from a multi-institutional project, also titled Grass2Gas, that includes collaborators from Penn State, Iowa State University and Roeslein Alternative Energy.

Balancing sustainability promises and complexity

“It has been suggested that promoting perennial plant species for nutrient management and converting manure and biomass into biogas with anaerobic digestion, which can be upgraded to renewable natural gas or directly burned in a generator to produce electricity and heat, can support sustainability on Pennsylvania dairy farms—and we wanted to see if that is true,” said study senior author Christine Costello, assistant professor of agricultural and biological engineering in the College of Agricultural Sciences at Penn State.

“We found that it could be true, but our research highlights the complexities of integrating anaerobic digestion into farm systems, including the impact on soil biogeochemistry and nutrient balances.”

Mixed results for water quality impacts

Surprisingly, the researchers said, the practices’ effects on water pollution in the simulation were mixed. Continuous cover reduces nutrient runoff on the farm, which should reduce eutrophication—excessive enrichment of water bodies with nitrogen and phosphorus that triggers algae growth, depletes oxygen and kills aquatic life. But growing more vegetation for anaerobic digestion also increased the need for off-farm imports of feed for the cows in most scenarios. More feed imports results in more environmental impact from producing that feed somewhere else, the researchers reported, offsetting many of the water quality benefits from a life cycle perspective.

In their modeling, the researchers found that reducing herd size to match available feed resulted in small milk losses, comparable to typical levels of milk wasted in the dairy supply chain.

“If the farm avoids feed imports by reducing herd size, the reduced milk production was less than current estimates of milk wasted in the dairy supply chain,” Costello said, explaining the team compared their modeled result to the reported food loss and waste rates calculated by the U.S. Department of Agriculture. “In other words, if waste could be avoided, reducing herd size wouldn’t significantly hurt overall milk availability.”

How digestate changes farm emissions

They estimated the potential for multiple impacts—global warming, marine eutrophication and acidification, which are linked to emissions that contribute to acid rain—across multiple scenarios. They used data from previous studies to estimate how winter cover crops, manure handling and management of the digestate—the nutrient-rich fertilizer left after the anaerobic digestion process—affect emissions and resource use.

“The digestion process changes the chemistry of manure, especially how nitrogen is stored and released,” she said. “That means when farmers spread digestate on fields, it behaves differently in soil and air than undigested manure. That influences emissions considerably, especially nitrogen, including ammonia, nitrous oxide and nitrate emissions, which impair climate, air quality and water quality.”

Seeing farms as connected systems

This study highlights the complex balance between milk production, energy generation from biogas, manure and digestate management, food waste and environmental impacts such as climate change and eutrophication, Costello said. Integrated crop-livestock-digester systems can reduce climate impacts but may unintentionally shift nutrient pollution elsewhere unless carefully managed.

“The main takeaway from the study is that agricultural and environmental scientists, engineers and policymakers should think about crop and livestock production and energy technology as one interconnected system,” Costello said. “When we think about adding an energy-production technology—in this case, anaerobic digestion—to a farm, we really need to think about how the residual materials—in this case, digestate—will be handled.”

Outcomes of this project and the work of others in the region working on anaerobic digestion will be covered in the Anaerobic Digestion on the Farm Conference at Penn State June 9–11.

Watching proteins move as they drive the chemical reactions that sustain life is one of the grand challenges of modern biology. In recent years, X-ray free-electron lasers, or XFELs, have begun to meet that challenge, capturing ultrafast snapshots of molecules as they shift shape during a reaction—effectively creating molecular slow-motion movies.

But the technique comes at a cost: These experiments typically consume enormous amounts of precious protein samples, putting many studies out of reach.

Now, researchers at Arizona State University and their international colleagues, including scientists from the Consejo Superior de Investigaciones Científicas, have developed a device that cuts sample consumption by as much as 97% while still producing high-quality structural data. The device, called a microfluidic droplet injector, delivers protein crystals to an XFEL beam in tiny, precisely timed packets rather than as a continuous stream.

The work is published in the journal Communications Chemistry.

The technology could accelerate drug discovery by showing how medicines interact with their protein targets in real time and help engineers design better enzymes for industry and biotechnology. It may also enable deeper insights into disease, enable the study of rare proteins that are difficult to produce, and unlock the full potential of next-generation X-ray laser facilities without excessive sample waste.

“Seeing proteins react in real time is incredibly powerful, but the sample demands to unravel dynamic protein behavior with X-ray crystallography have been a major limitation,” says Alexandra Ros, lead author of the new study. “Our droplet approach dramatically reduces that burden, which is exciting because many more labs can now ask dynamic questions that were previously too costly or impractical.”

Professor Ros is a researcher with the Biodesign Center for Applied Structural Discovery and the School of Molecular Sciences at Arizona State University. She is joined by a team of ASU colleagues and international researchers, including collaborators from CSIC.

Freezing motion at the atomic scale

XFELs effectively freeze motion by firing ultrashort pulses lasting just femtoseconds—one quadrillionth of a second. In that tiny slice of time, light itself travels less than the width of a single strand of human hair.

In these experiments, proteins are first grown into microscopic crystals and then exposed to the XFEL pulses. This produces diffraction patterns that researchers use to reconstruct atomic-scale snapshots of the molecules in action.

In conventional XFEL experiments, protein crystals are sprayed continuously across the path of the X-ray beam. Only a small fraction of that material is actually hit by an X-ray pulse, meaning most of the protein goes to waste. The new system generates a train of microscopic droplets, each carrying a small amount of sample. These droplets are synchronized with the laser’s pulse timing, so the sample arrives only when it is needed.

Working at the European X-ray free-electron laser, the researchers used time-resolved serial femtosecond crystallography, which can capture proteins as they change shape during reactions. With this powerful technique, they studied a human enzyme called NQO1, which plays an important role in cellular detoxification and protection against oxidative stress.

Big gains from tiny droplets

The researchers were able to capture early snapshots of how the enzyme begins interacting with its cofactor—a helper molecule called NADH—while using up to 97% less protein than traditional approaches. Understanding exactly how NADH binds and moves inside the enzyme could help scientists better understand how the enzyme functions and how it might be targeted in disease.

Although the biological findings are preliminary, the technological innovation is a significant advancement. By drastically reducing the amount of protein required, the method makes it feasible to study proteins that are rare, fragile, or difficult to produce in large quantities. This opens the door to many experiments that were previously inaccessible.

The injector itself is built using high-resolution 3D printing and integrates tiny channels that mix solutions and form droplets on demand. The design is compatible with the extremely rapid pulse structure of modern XFEL facilities, allowing researchers to fully exploit the capabilities of these billion-dollar instruments without wasting valuable samples.

It could also complement emerging efforts to make X-ray laser technology more compact and accessible, including the compact X-ray free-electron laser currently under development at ASU.

In biology, many RNA molecules act as sophisticated microscopic machines. Among them, riboswitches function as tiny biological sensors, changing their 3D shape upon binding to a specific metabolite. This shape-change acts as a switch, often turning a downstream gene “on” or “off.” The ability to design artificial switches from scratch would hold immense promise for synthetic biology, drug design, and new diagnostic tools. However, designing a sequence that can stably fold into two different shapes and switch between them is an extremely difficult challenge.

A collaboration involving CNRS researchers at École Normale Supérieure Paris, Université Paris Cité, École Polytechnique, and the Institut de Physique Théorique, have successfully used a machine learning approach to design entirely new, functional RNA switches. The team applied a model known as Restricted Boltzmann machines (RBM), to learn the “design rules” of the aptamer domain of a natural riboswitch family. By training the RBM on thousands of natural sequences, the model learned to capture the complex interaction patterns leading to the formation of secondary and tertiary contacts and essential for the structural switch, which simpler models miss.

The paper is published in the journal Nature Communications.

This work is a significant leap forward in the rational design of allosteric biomolecules. It demonstrates how generative models, born from the intersection of statistical physics and AI, can not only understand the complex language of biology but also use it to design new, functional molecules, paving the way for custom-designed molecular tools.

The trained RBM was then used as a “generator” to create 476 novel RNA sequences, some differing by up to 40% from known natural sequences. The team synthesized and tested these artificial sequences using high-throughput chemical probing (SHAPE and DMS). The results were a remarkable success: approximately one-third of the high-scoring designed sequences functioned as effective switches, changing their conformation in response to the target metabolite (SAM) just as their natural counterparts do.

One of the most enduring goals in regenerative medicine is deceptively simple: replace a person’s damaged or dying cells with healthy new ones grown in the laboratory.

Researchers at Harvard Medical School and around the world have made striking progress toward that goal, learning how to guide stem cells to become muscle, nerve, and other specialized cell types. In principle, those lab-grown cells could one day be used to repair injured tissues or slow the progression of disease.

In practice, however, only a small fraction of these advances has moved beyond the lab.

The difficulty lies in how hard it is to control the process by which stem cells develop. As cells mature, they respond to a sequence of chemical signals that tell them what to become. Those signals must arrive at the right moment and in the right amount. Small missteps can cause cells to stray from their intended path, leaving them immature, inconsistent, or unsuitable for use as therapy. Even protocols that work well in one lab can be difficult to reproduce elsewhere.

Researchers in the Blavatnik Institute at HMS co-founded the company Cellular Intelligence to address that challenge.

Building on large experimental datasets from developmental, systems, and computational biology, the company aims to build a foundation model—a large machine-learning system trained on experimental data—to improve two requirements for turning cell replacement therapies into viable treatments: predictability and scalability.

“For many cell therapies, the biology works but not robustly enough,” said scientific co-founder Allon Klein, professor of systems biology at HMS. “You can get the right cells once, but reproducing that result reliably and at scale is a very different problem.”

Once trained and validated, the machine-learning tool could reveal underlying rules that guide cell development—rules that researchers can use to predict how cells will behave under new conditions.

“Developmental biology already has an internal logic,” Klein said. “What we’re trying to do is understand that logic well enough to guide it.”

Exploring translation, step by step

The origins of Cellular Intelligence, formerly known as Somite AI, lie in decades of basic research.

Much of the early work took place at HMS before the team ever envisioned a company. At that stage, the focus was on building tools and asking whether they could shed light on fundamental questions about how cells develop.

As the scientific picture sharpened, Klein and colleagues—including scientific co-founders Olivier Pourquié, the HMS Frank Burr Mallory Professor of Pathology at Brigham and Women’s Hospital and professor of genetics at HMS, and Clifford Tabin, the George Jacob and Jacqueline Hazel Leder Professor of Genetics and head of the Department of Genetics at HMS—began to consider what it would take to develop the work further.

Some questions, such as about reproducibility, scale, and integration of experimental and computational approaches, were difficult to address within academic labs.

The Blavatnik Harvard Life Lab Longwood offered a practical next step. The incubator provides shared laboratory space and infrastructure for early-stage, Harvard-connected life science startups, allowing teams to organize work outside the academic setting while remaining close to ongoing research.

For Cellular Intelligence, that proximity mattered. In its earliest phase, frequent interaction between the scientific founders and the growing team helped clarify which ideas could translate and which would need to be rethought.

“For something this complex, being able to easily move back and forth mattered,” Klein said. “It helped us test ideas quickly.”

Gene regulation is far more predictable than previously believed, scientists conclude after developing the deep learning model PARM. This might bring an end to a scientific mystery: how genes know when to switch on or off.

Scientists have published in Nature about their relentless back-and-forth between lab experiments and computation that enabled them to build this lightweight model. Scientists around the world can now start using this tool for reading these genetic instructions, creating leads for new cancer diagnostics, patient stratification, and future therapies.

“The classical genetic code explains how genes in our DNA encode proteins,” explains Bas van Steensel, group leader at the Netherlands Cancer Institute (NKI) and Oncode Institute and co-corresponding author on the paper.

“But for most genes, we honestly didn’t understand how they are regulated. We know that the DNA between our genes contains regulatory elements such as promotors. However, the language of this control system that decides whether a gene turns on or off, in which cell, and how strongly was largely unknown.”

Bold mission

At the same time, most cancer-related mutations are located in the non-coding part of our genome, illustrating the immense relevance of this unsolved issue. Until now, interpreting such mutations has been extremely difficult. With the PARM model, this becomes possible.

Starting on a bold mission to decode the genome operating system, seven research groups joined forces in Oncode Institute’s PERICODE project. A technology developed in the Bas van Steensel lab at the NKI enabled measuring gene regulation at an unprecedented scale. Millions of carefully controlled measurements captured how short DNA sequences influence gene activity.

But data alone is not insight. That is where Jeroen de Ridder’s research group from UMC Utrecht and Oncode Institute entered the picture. The volume of data specifically targeted to gene regulation enabled training AI models that truly captured the biological rules underlying gene activation.

“Most AI models learn from whatever data happens to exist,” de Ridder explains. “Here, the measurements and the AI were designed together. This allowed us to make super-efficient models for specific cell types that could be applied at a scale previously unthinkable.”

Rigorous testing

The new model enabled the team to predict how gene regulation differs between cell types and how it changes when cells are exposed to stimuli such as specific drugs. Moreover, the model revealed in extreme detail what the architecture of the “on and off buttons” of each gene is. Crucially, the team did not stop at prediction. Every model output was subjected to rigorous experimental testing to make sure that these predictions were indeed correct.

“We can now actually read the language of the gene control system,” says Van Steensel. “Our PARM model allows us to uncover these rules at scale, so we can now understand, and even predict, how regulatory DNA controls gene activity.”

Despite notable progress in the field, the existing AI models were either too heavy to be applied to the vast numbers of mutations that exist or are too generic and do not adequately capture cell type variability. The PARM model changes that. It allows researchers to predict the functional impact of regulatory mutations in specific cell types and under specific conditions, such as drug treatments, opening new paths for cancer diagnostics, patient stratification, and future therapies.

Recently, Google’s Deepmind published in Nature an article about their model AlphaGenome, also aimed at understanding gene regulation. “This is a great model,” says Van Steensel.

“However, PARM is more flexible and it is experimentally and computationally lightweight. The tool requires around 1,000 times less computing power than AlphaGenome, making it far more feasible for academic researchers around the world. With this model, you only need one petridish of cells and one day of computing to see in detail how a particular cell type, such as a tumor cell, uses its DNA code to respond to a signal such as a hormone, nutrient or drug.”

The PARM model was developed within the PERICODE project, initiated by Oncode Institute. Seven research groups collaborated within the project: Bas van Steensel (NKI), Jeroen de Ridder (UMCU), Emile Voest (NKI), Michiel Vermeulen (NKI), Lude Franke (UMCG), Sarah Derks (Amsterdam UMC), Wilbert Zwart (NKI).