A robot searching for workers trapped in a partially collapsed mine shaft must rapidly generate a map of the scene and identify its location within that scene as it navigates the treacherous terrain.

Researchers have recently started building powerful machine-learning models to perform this complex task using only images from the robot’s onboard cameras, but even the best models can only process a few images at a time. In a real-world disaster where every second counts, a search-and-rescue robot would need to quickly traverse large areas and process thousands of images to complete its mission.

To overcome this problem, MIT researchers drew on ideas from both recent artificial intelligence vision models and classical computer vision to develop a new system that can process an arbitrary number of images. Their system accurately generates 3D maps of complicated scenes like a crowded office corridor in a matter of seconds. 

The AI-driven system incrementally creates and aligns smaller submaps of the scene, which it stitches together to reconstruct a full 3D map while estimating the robot’s position in real-time.

Unlike many other approaches, their technique does not require calibrated cameras or an expert to tune a complex system implementation. The simpler nature of their approach, coupled with the speed and quality of the 3D reconstructions, would make it easier to scale up for real-world applications.

Beyond helping search-and-rescue robots navigate, this method could be used to make extended reality applications for wearable devices like VR headsets or enable industrial robots to quickly find and move goods inside a warehouse.

“For robots to accomplish increasingly complex tasks, they need much more complex map representations of the world around them. But at the same time, we don’t want to make it harder to implement these maps in practice. We’ve shown that it is possible to generate an accurate 3D reconstruction in a matter of seconds with a tool that works out of the box,” says Dominic Maggio, an MIT graduate student and lead author of a paper on this method.

Maggio is joined on the paper by postdoc Hyungtae Lim and senior author Luca Carlone, associate professor in MIT’s Department of Aeronautics and Astronautics (AeroAstro), principal investigator in the Laboratory for Information and Decision Systems (LIDS), and director of the MIT SPARK Laboratory. The research will be presented at the Conference on Neural Information Processing Systems.

Mapping out a solution

For years, researchers have been grappling with an essential element of robotic navigation called simultaneous localization and mapping (SLAM). In SLAM, a robot recreates a map of its environment while orienting itself within the space.

Traditional optimization methods for this task tend to fail in challenging scenes, or they require the robot’s onboard cameras to be calibrated beforehand. To avoid these pitfalls, researchers train machine-learning models to learn this task from data.

While they are simpler to implement, even the best models can only process about 60 camera images at a time, making them infeasible for applications where a robot needs to move quickly through a varied environment while processing thousands of images.

To solve this problem, the MIT researchers designed a system that generates smaller submaps of the scene instead of the entire map. Their method “glues” these submaps together into one overall 3D reconstruction. The model is still only processing a few images at a time, but the system can recreate larger scenes much faster by stitching smaller submaps together.

“This seemed like a very simple solution, but when I first tried it, I was surprised that it didn’t work that well,” Maggio says.

Searching for an explanation, he dug into computer vision research papers from the 1980s and 1990s. Through this analysis, Maggio realized that errors in the way the machine-learning models process images made aligning submaps a more complex problem.

Traditional methods align submaps by applying rotations and translations until they line up. But these new models can introduce some ambiguity into the submaps, which makes them harder to align. For instance, a 3D submap of a one side of a room might have walls that are slightly bent or stretched. Simply rotating and translating these deformed submaps to align them doesn’t work.

“We need to make sure all the submaps are deformed in a consistent way so we can align them well with each other,” Carlone explains.

A more flexible approach

Borrowing ideas from classical computer vision, the researchers developed a more flexible, mathematical technique that can represent all the deformations in these submaps. By applying mathematical transformations to each submap, this more flexible method can align them in a way that addresses the ambiguity.

Based on input images, the system outputs a 3D reconstruction of the scene and estimates of the camera locations, which the robot would use to localize itself in the space.

“Once Dominic had the intuition to bridge these two worlds — learning-based approaches and traditional optimization methods — the implementation was fairly straightforward,” Carlone says. “Coming up with something this effective and simple has potential for a lot of applications.

Their system performed faster with less reconstruction error than other methods, without requiring special cameras or additional tools to process data. The researchers generated close-to-real-time 3D reconstructions of complex scenes like the inside of the MIT Chapel using only short videos captured on a cell phone.

The average error in these 3D reconstructions was less than 5 centimeters.

In the future, the researchers want to make their method more reliable for especially complicated scenes and work toward implementing it on real robots in challenging settings.

“Knowing about traditional geometry pays off. If you understand deeply what is going on in the model, you can get much better results and make things much more scalable,” Carlone says.

This work is supported, in part, by the U.S. National Science Foundation, U.S. Office of Naval Research, and the National Research Foundation of Korea. Carlone, currently on sabbatical as an Amazon Scholar, completed this work before he joined Amazon.

What if clinicians could place tiny electronic chips in the brain that electrically stimulate a precise target, through a simple injection in the arm? This may someday help treat deadly or debilitating brain diseases, while eliminating surgery-related risks and costs.

MIT researchers have taken a major step toward making this scenario a reality. They developed microscopic, wireless bioelectronics that could travel through the body’s circulatory system and autonomously self-implant in a target region of the brain, where they would provide focused treatment.

In a study on mice, the researchers show that after injection, these miniscule implants can identify and travel to a specific brain region without the need for human guidance. Once there, they can be wirelessly powered to provide electrical stimulation to the precise area. Such stimulation, known as neuromodulation, has shown promise as a way to treat brain tumors and diseases like Alzheimer’s and multiple sclerosis.

Moreover, because the electronic devices are integrated with living, biological cells before being injected, they are not attacked by the body’s immune system and can cross the blood-brain barrier while leaving it intact. This maintains the barrier’s crucial protection of the brain.

The researchers demonstrated the use of this technology, which they call “circulatronics,” to target brain inflammation, a major factor in the progression of many neurological diseases. They show that the implants can provide localized neuromodulation deep inside the brain achieving high precision, to within several microns around the target area.

In addition, the biocompatible implants do not damage surrounding neurons.

While brain implants usually require hundreds of thousands of dollars in medical costs and risky surgical procedures, circulatronics technology holds the potential to make therapeutic brain implants accessible to all by eliminating the need for surgery, says Deblina Sarkar, the AT&T Career Development Associate Professor in the MIT Media Lab and MIT Center for Neurobiological Engineering, head of the Nano-Cybernetic Biotrek Lab, and senior author of a study on the work.

She is joined on the paper by lead author Shubham Yadav, an MIT graduate student; as well as others at MIT, Wellesley College, and Harvard University. The research appears today in Nature Biotechnology.

Hybrid implants

The team has been working on circulatronics for more than six years. The electronic devices, each about one-billionth the length of a grain of rice, are composed of organic semiconducting polymer layers sandwiched between metallic layers to create an electronic heterostructure.

They are fabricated using CMOS-compatible processes in the MIT.nano facilities, and then integrated with living cells to create cell-electronics hybrids. To do this, the researchers lift the devices off the silicon wafer on which they are fabricated, so they are free-floating in a solution.

“The electronics worked perfectly when they were attached to the substrate, but when we originally lifted them off, they didn’t work anymore. Solving that challenge took us more than a year,” Sarkar says.

Key to their operation is the high wireless power conversion efficiency of the tiny electronics. This enables the devices to work deep inside the brain and still harness enough energy for neuromodulation.

The researchers use a chemical reaction to bond the electronic devices to cells. In the new study, they fused the electronics with a type of immune cell called monocytes, which target areas of inflammation in the body. They also applied a fluorescent dye, allowing them to trace the devices as they crossed the intact blood-brain barrier and self-implanted in the target brain region.

While they explored brain inflammation in this study, the researchers hope to use different cell types and engineer the cells to target specific regions of the brain.

“Our cell-electronics hybrid fuses the versatility of electronics with the biological transport and biochemical sensing prowess of living cells,” Sarkar says. “The living cells camouflage the electronics so that they aren’t attacked by the body’s immune system and they can travel seamlessly through the bloodstream. This also enables them to squeeze through the intact blood-brain barrier without the need to invasively open it.”

Over the course of about four years, the team tried many methods to autonomously and noninvasively cross the blood-brain barrier before they perfected this cellular integration technique.

In addition, because the circulatronics devices are so tiny, they offer much higher precision than conventional electrodes. They can self-implant, leading to millions of microscopic stimulation sites that take the exact shape of the target region.

Their small size also enables the biocompatible devices to live alongside neurons without causing harmful effects. Through a series of biocompatibility tests, the researchers found that circulatronics can safely integrate among neurons without impacting the brain processes behind cognition or motion.

After the devices have self-implanted in the target region, a clinician or researcher uses an external transmitter to provide electromagnetic waves, in the form of near-infrared light, that power the technology and enable electrical stimulation of the neurons.

Targeting deadly diseases

The Sarkar lab is currently working on developing their technology to treat multiple diseases including brain cancer, Alzheimer’s disease, and chronic pain.

The tiny size and self-implantation capabilities of circulatronics devices could make them well-suited to treat brain cancers such as glioblastoma that cause tumors at multiple locations, some of which may be too small to identify with imaging techniques. They may also provide new avenues for treating especially deadly cancers like diffuse intrinsic pontine glioma, an aggressive type of tumor found in the brain stem that usually cannot be surgically removed.

“This is a platform technology and may be employed to treat multiple brain diseases and mental illnesses,” Sarkar says. “Also, this technology is not just confined to the brain but could also be extended to other parts of the body in future.”

The researchers hope to move the technology into clinical trials within three years through the recently launched startup Cahira Technologies.

They are also exploring integration of additional nanoelectronic circuits into their devices to enable functionalities including sensing, feedback based on-chip data analysis, and capabilities such as creating synthetic electronic neurons.

“Our tiny electronic devices seamlessly integrate with the neurons and co-live and co-exist with the brain cells creating a unique brain-computer symbiosis. We are working dedicatedly to employ this technology for treating neural diseases, where drugs or standard therapies fail, for alleviating human suffering and envision a future where humans could transcend beyond diseases and biological limitations,” says Sarkar.

One of the highest-risk components of nuclear waste is iodine-129 (I-129), which stays radioactive for millions of years and accumulates in human thyroids when ingested. In the U.S., nuclear waste containing I-129 is scheduled to be disposed of in deep underground repositories, which scientists say will sufficiently isolate it.

Meanwhile, across the globe, France routinely releases low-level radioactive effluents containing iodine-129 and other radionuclides into the ocean. France recycles its spent nuclear fuel, and the reprocessing plant discharges about 153 kilograms of iodine-129 each year, under the French regulatory limit.

Is dilution a good solution? What’s the best way to handle spent nuclear fuel? A new study by MIT researchers and their collaborators at national laboratories quantifies I-129 release under three different scenarios: the U.S. approach of disposing spent fuel directly in deep underground repositories, the French approach of dilution and release, and an approach that uses filters to capture I-129 and disposes of them in shallow underground waste repositories.

The researchers found France’s current practice of reprocessing releases about 90 percent of the waste’s I-129 into the biosphere. They found low levels of I-129 in ocean water around France and the U.K.’s former reprocessing sites, including the English Channel and North Sea. Although the low level of I-129 in the water in Europe is not considered to pose health risks, the U.S. approach of deep underground disposal leads to far less I-129 being released, the researchers found.

The researchers also investigated the effect of environmental regulations and technologies related to I-129 management, to illuminate the tradeoffs associated with different approaches around the world.

“Putting these pieces together to provide a comprehensive view of Iodine-129 is important,” says MIT Assistant Professor Haruko Wainwright, a first author on the paper who holds a joint appointment in the departments of Nuclear Science and Engineering and of Civil and Environmental Engineering. “There are scientists that spend their lives trying to clean up iodine-129 at contaminated sites. These scientists are sometimes shocked to learn some countries are releasing so much iodine-129. This work also provides a life-cycle perspective. We’re not just looking at final disposal and solid waste, but also when and where release is happening. It puts all the pieces together.”

MIT graduate student Kate Whiteaker SM ’24 led many of the analyses with Wainwright. Their co-authors are Hansell Gonzalez-Raymat, Miles Denham, Ian Pegg, Daniel Kaplan, Nikolla Qafoku, David Wilson, Shelly Wilson, and Carol Eddy-Dilek. The study appears today in Nature Sustainability.

Managing waste

Iodine-129 is often a key focus for scientists and engineers as they conduct safety assessments of nuclear waste disposal sites around the world. It has a half-life of 15.7 million years, high environmental mobility, and could potentially cause cancers if ingested. The U.S. sets a strict limit on how much I-129 can be released and how much I-129 can be in drinking water — 5.66 nanograms per liter, the lowest such level of any radionuclides.

“Iodine-129 is very mobile, so it is usually the highest-dose contributor in safety assessments,” Wainwright says.

For the study, the researchers calculated the release of I-129 across three different waste management strategies by combining data from current and former reprocessing sites as well as repository assessment models and simulations.

The authors defined the environmental impact as the release of I-129 into the biosphere that humans could be exposed to, as well as its concentrations in surface water. They measured I-129 release per the total electrical energy generated by a 1-gigawatt power plant over one year, denoted as kg/GWe.y.

Under the U.S. approach of deep underground disposal with barrier systems, assuming the barrier canisters fail at 1,000 years (a conservative estimate), the researchers found 2.14 x 10–8 kg/GWe.y of I-129 would be released between 1,000 and 1 million years from today.

They estimate that 4.51 kg/GWe.y of I-129, or 91 percent of the total, would be released into the biosphere in the scenario where fuel is reprocessed and the effluents are diluted and released. About 3.3 percent of I-129 is captured by gas filters, which are then disposed of in shallow subsurfaces as low-level radioactive waste. A further 5.2 percent remains in the waste stream of the reprocessing plant, which is then disposed of as high-level radioactive waste.

If the waste is recycled with gas filters to directly capture I-129, 0.05 kg/GWe.y of the I-129 is released, while 94 percent is disposed of in the low-level disposal sites. For shallow disposal, some kind of human disruption and intrusion is assumed to occur after government or institutional control expires (typically 100-1,000 years). That results in a potential release of the disposed amount to the environment after the control period.

Overall, the current practice of recycling spent nuclear fuel releases the majority of I-129 into the environment today, while the direct disposal of spent fuel releases around 1/100,000,000 that amount over 1 million years. When the gas filters are used to capture I-129, the majority of I-129 goes to shallow underground repositories, which could be accidentally released through human intrusion down the line.

The researchers also quantified the concentration of I-129 in different surface waters near current and former fuel reprocessing facilities, including the English Channel and the North Sea near reprocessing plants in France and U.K. They also analyzed the U.S. Columbia River downstream of a site in Washington state where material for nuclear weapons was produced during the Cold War, and they studied a similar site in South Carolina. The researchers found far higher concentrations of I-129 within the South Carolina site, where the low-level radioactive effluents were released far from major rivers and hence resulted in less dilution in the environment.

“We wanted to quantify the environmental factors and the impact of dilution, which in this case affected concentrations more than discharge amounts,” Wainwright says. “Someone might take our results to say dilution still works: It’s reducing the contaminant concentration and spreading it over a large area. On the other hand, in the U.S., imperfect disposal has led to locally higher surface water concentrations. This provides a cautionary tale that disposal could concentrate contaminants, and should be carefully designed to protect local communities.”

Fuel cycles and policy

Wainwright doesn’t want her findings to dissuade countries from recycling nuclear fuel. She says countries like Japan plan to use increased filtration to capture I-129 when they reprocess spent fuel. Filters with I-129 can be disposed of as low-level waste under U.S. regulations.

“Since I-129 is an internal carcinogen without strong penetrating radiation, shallow underground disposal would be appropriate in line with other hazardous waste,” Wainwright says. “The history of environmental protection since the 1960s is shifting from waste dumping and release to isolation. But there are still industries that release waste into the air and water. We have seen that they often end up causing issues in our daily life — such as CO2, mercury, PFAS and others — especially when there are many sources or when bioaccumulation happens. The nuclear community has been leading in waste isolation strategies and technologies since the 1950s. These efforts should be further enhanced and accelerated. But at the same time, if someone does not choose nuclear energy because of waste issues, it would encourage other industries with much lower environmental standards.”

The work was supported by MIT’s Climate Fast Forward Faculty Fund and the U.S. Department of Energy.

Although heart cells and skin cells contain identical instructions for creating proteins encoded in their DNA, they’re able to fill such disparate niches because molecular machinery can cut out and stitch together different segments of those instructions to create endlessly unique combinations.

The ingenuity of using the same genes in different ways is made possible by a process called splicing and is controlled by splicing factors; which splicing factors a cell employs determines what sets of instructions that cell produces, which, in turn, gives rise to proteins that allow cells to fulfill different functions. 

In an open-access paper published today in Nature Biotechnology, researchers in the MIT Department of Biology outlined a framework for parsing the complex relationship between sequences and splicing regulation to investigate the regulatory activities of splicing factors, creating models that can be applied to interpret and predict splicing regulation across different cell types, and even different species. Called Knockdown Activity and Target Models from Additive regression Predictions, KATMAP draws on experimental data from disrupting the expression of a splicing factor and information on which sequences the splicing factor interacts with to predict its likely targets. 

Aside from the benefits of a better understanding of gene regulation, splicing mutations — either in the gene that is spliced or in the splicing factor itself — can give rise to diseases such as cancer by altering how genes are expressed, leading to the creation or accumulation of faulty or mutated proteins. This information is critical for developing therapeutic treatments for those diseases. The researchers also demonstrated that KATMAP can potentially be used to predict whether synthetic nucleic acids, a promising treatment option for disorders including a subset of muscular atrophy and epilepsy disorders, affect splicing.

Perturbing splicing 

In eukaryotic cells, including our own, splicing occurs after DNA is transcribed to produce an RNA copy of a gene, which contains both coding and non-coding regions of RNA. The noncoding intron regions are removed, and the coding exon segments are spliced back together to make a near-final blueprint, which can then be translated into a protein. 

According to first author Michael P. McGurk, a postdoc in the lab of MIT Professor Christopher Burge, previous approaches could provide an average picture of regulation, but could not necessarily predict the regulation of splicing factors at particular exons in particular genes.

KATMAP draws on RNA sequencing data generated from perturbation experiments, which alter the expression level of a regulatory factor by either overexpressing it or knocking down its levels. The consequences of overexpression or knockdown are that the genes regulated by the splicing factor should exhibit different levels of splicing after perturbation, which helps the model identify the splicing factor’s targets. 

Cells, however, are complex, interconnected systems, where one small change can cause a cascade of effects. KATMAP is also able to distinguish between direct targets from indirect, downstream impacts by incorporating known information about the sequence the splicing factor is likely to interact with, referred to as a binding site or binding motif.

“In our analyses, we identify predicted targets as exons that have binding sites for this particular factor in the regions where this model thinks they need to be to impact regulation,” McGurk says, while non-targets may be affected by perturbation but don’t have the likely appropriate binding sites nearby. 

This is especially helpful for splicing factors that aren’t as well-studied. 

“One of our goals with KATMAP was to try to make the model general enough that it can learn what it needs to assume for particular factors, like how similar the binding site has to be to the known motif or how regulatory activity changes with the distance of the binding sites from the splice sites,” McGurk says. 

Starting simple

Although predictive models can be very powerful at presenting possible hypotheses, many are considered “black boxes,” meaning the rationale that gives rise to their conclusions is unclear. KATMAP, on the other hand, is an interpretable model that enables researchers to quickly generate hypotheses and interpret splicing patterns in terms of regulatory factors while also understanding how the predictions were made. 

“I don’t just want to predict things, I want to explain and understand,” McGurk says. “We set up the model to learn from existing information about splicing and binding, which gives us biologically interpretable parameters.” 

The researchers did have to make some simplifying assumptions in order to develop the model. KATMAP considers only one splicing factor at a time, although it is possible for splicing factors to work in concert with one another. The RNA target sequence could also be folded in such a way that the factor wouldn’t be able to access a predicted binding site, so the site is present but not utilized.

“When you try to build up complete pictures of complex phenomena, it’s usually best to start simple,” McGurk says. “A model that only considers one splicing factor at a time is a good starting point.” 

David McWaters, another postdoc in the Burge Lab and a co-author on the paper, conducted key experiments to test and validate that aspect of the KATMAP model.

Future directions

The Burge lab is collaborating with researchers at Dana-Farber Cancer Institute to apply KATMAP to the question of how splicing factors are altered in disease contexts, as well as with other researchers at MIT as part of an MIT HEALS grant to model splicing factor changes in stress responses. McGurk also hopes to extend the model to incorporate cooperative regulation for splicing factors that work together. 

“We’re still in a very exploratory phase, but I would like to be able to apply these models to try to understand splicing regulation in disease or development. In terms of variation of splicing factors, they are related, and we need to understand both,” McGurk says.

Burge, the Uncas (1923) and Helen Whitaker Professor and senior author of the paper, will continue to work on generalizing this approach to build interpretable models for other aspects of gene regulation.

“We now have a tool that can learn the pattern of activity of a splicing factor from types of data that can be readily generated for any factor of interest,” says Burge, who is also an extra-mural member of the Koch Institute for Integrative Cancer Research and an associate member of the Broad Institute of MIT and Harvard. “As we build up more of these models, we’ll be better able to infer which splicing factors have altered activity in a disease state from transcriptomic data, to help understand which splicing factors are driving pathology.”

MIT engineers have developed a flexible drug-delivery patch that can be placed on the heart after a heart attack to help promote healing and regeneration of cardiac tissue.

The new patch is designed to carry several different drugs that can be released at different times, on a pre-programmed schedule. In a study of rats, the researchers showed that this treatment reduced the amount of damaged heart tissue by 50 percent and significantly improved cardiac function.

If approved for use in humans, this type of patch could help heart attack victims recover more of their cardiac function than is now possible, the researchers say.

“When someone suffers a major heart attack, the damaged cardiac tissue doesn’t regenerate effectively, leading to a permanent loss of heart function. The tissue that was damaged doesn’t recover,” says Ana Jaklenec, a principal investigator at MIT’s Koch Institute for Integrative Cancer Research. “Our goal is to restore that function and help people regain a stronger, more resilient heart after a myocardial infarction.”

Jaklenec and Robert Langer, the David H. Koch Institute Professor at MIT and a member of the Koch Institute, are the senior authors of the new study, which appears today in Cell Biomaterials. Former MIT postdoc Erika Wangis the lead author of the paper.

Programmed drug delivery

After a heart attack, many patients end up having bypass surgery, which improves blood flow to the heart but doesn’t repair the cardiac tissue that was damaged. In the new study, the MIT team wanted to create a patch that could be applied to the heart at the same time that the surgery is performed.

This patch, they hoped, could deliver drugs over an extended time period to promote tissue healing. Many diseases, including heart conditions, require phase-specific treatment, but most systems release drugs all at once. Timed delivery better synchronizes therapy with recovery.

“We wanted to see if it’s possible to deliver a precisely orchestrated therapeutic intervention to help heal the heart, right at the site of damage, while the surgeon is already performing open-heart surgery,” Jaklenec says.

To achieve this, the researchers set out to adapt drug-delivery microparticles they had previously developed, which consist of capsules similar to tiny coffee cups with lids. These capsules are made from a polymer called PLGA and can be sealed with a drug inside.

By changing the molecular weight of the polymers used to form the lids, the researchers can control how quickly they degrade, which enables them to program the particles to release their contents at specific times. For this application, the researchers designed particles that break down during days 1-3, days 7-9, and days 12-14 after implantation.

This allowed them to devise a regimen of three drugs that promote heart healing in different ways. The first set of particles release neuregulin-1, a growth factor that helps to prevent cell death. At the next time point, particles release VEGF, a growth factor that promotes formation of blood vessels surrounding the heart. The last batch of particles releases a small molecule drug called GW788388, which inhibits the formation of scar tissue that can occur following a heart attack.

“When tissue regenerates, it follows a carefully timed series of steps,” Jaklenec says. “Dr. Wang created a system that delivers key components at just the right time, in the sequence that the body naturally uses to heal.”

The researchers embedded rows of these particles into thin sheets of a tough but flexible hydrogel, similar to a contact lens. This hydrogel is made from alginate and PEGDA, two biocompatible polymers that eventually break down in the body. For this study, the researchers created compact, miniature patches only a few millimeters across.

“We encapsulate arrays of these particles in a hydrogel patch, and then we can surgically implant this patch into the heart. In this way, we’re really programming the treatment into this material,” Wang says.

Better heart function

Once they created these patches, the researchers tested them on spheres of heart tissue that included cardiomyocytes generated from induced pluripotent stem cells. These spheres also included endothelial cells and human ventricular cardiac fibroblasts, which are also important components of the heart.

The researchers exposed those spheres to low-oxygen conditions, mimicking the effects of a heart attack, then placed the patches over them. They found that the patches promoted blood vessel growth, helped more cells to survive, and reduced the amount of fibrosis that developed.

In tests in a rat model of heart attack, the researchers also saw significant improvements following treatment with the patch. Compared to no treatment or IV injection of the same drugs, animals treated with the patch showed 33 percent higher survival rates, a 50 percent reduction in the amount of damaged tissue, and significantly increased cardiac output.

The researchers showed that the patches would eventually dissolve over time, becoming a very thin layer over the course of a year without disrupting the heart’s mechanical function.

“This is an important way to combine drug delivery and biomaterials to potentially new treatments for patients,” Langer says.

Of the drugs tested in this study, neuregulin-1 and VEGF have been tested in clinical trials to treat heart conditions, but GW788388 has only been explored in animal models. The researchers now hope to test their patches in additional animal models in hopes of running a clinical trial in the future.

The current version of the patch needs to be implanted surgically, but the researchers are exploring the possibility of incorporating these microparticles into stents that could be inserted into arteries to deliver drugs on a programmed schedule.

Other authors of the paper include Elizabeth Calle, Binbin Ying, Behnaz Eshaghi, Linzixuan Zhang, Xin Yang, Stacey Qiaohui Lin, Jooli Han, Alanna Backx, Yuting Huang, Sevinj Mursalova, Chuhan Joyce Qi, and Yi Liu.

The researchers were supported by the Natural Sciences and Engineering Research Council of Canada and the U.S. National Heart, Lung, and Blood Institute.

More than 70 aircraft are struck by lightning every day. If you happen to be flying when a strike occurs, chances are you won’t feel a thing, thanks to lightning protection measures that are embedded in key zones throughout the aircraft.

Lightning protection systems work well, largely because they are designed for planes with a “tube-and-wing” structure, a simple geometry common to most aircraft today. But future airplanes may not look and fly the same way. The aviation industry is exploring new designs, including blended-wing bodies and truss-braced wings, partly to reduce fuel and weight costs. But researchers don’t yet know how these unconventional designs might respond to lightning strikes.

MIT aerospace engineers are hoping to change that with a new physics-based approach that predicts how lightning would sweep across a plane with any design. The tool then generates a zoning map highlighting sections of an aircraft that would require various degrees of lightning protection, given how they are likely to experience a strike.

“People are starting to conceive aircraft that look very different from what we’re used to, and we can’t apply exactly what we know from historical data to these new configurations because they’re just too different,” says Carmen Guerra-Garcia, associate professor of aeronautics and astronautics (AeroAstro) at MIT. “Physics-based methods are universal. They’re agnostic to the type of geometry or vehicle. This is the path forward to be able to do this lightning zoning and protect future aircraft.”

She and her colleagues report their results in a study appearing this week in IEEE Access. The study’s first author is AeroAstro graduate student Nathanael Jenkins. Other co-authors include Louisa Michael and Benjamin Westin of Boeing Research and Technology.

First strike

When lightning strikes, it first attaches to a part of a plane — typically a sharp edge or extremity — and hangs on for up to a second. During this brief flash, the plane continues speeding through the air, causing the lightning current to “sweep” over parts of its surface, potentially changing in intensity and re-attaching at certain points where the intense current flow could damage vulnerable sections of an aircraft.

In previous work, Guerra-Garcia’s group developed a model to predict the parts of a plane where lightning is most likely to first connect. That work, led by graduate student Sam Austin, established a starting point for the team’s new work, which aims to predict how and where the lightning will then sweep over the plane’s surface. The team next converted their lightning sweep predictions into zoning maps to identify vulnerable regions requiring certain levels of protection.

A typical tube-and-wing plane is divided into three main zones, as classified by the aviation industry. Each zone has a clear description of the level of current it must withstand in order to be certified for flight. Parts of a plane that are more likely to be hit by lightning are generally classified as zone 1 and require more protection, which can include embedded metal foil in the skin of the airplane that conducts away a lightning current.

To date, an airplane’s lightning zones have been determined over many years of flight inspections after lightning strikes and fine-tuning of protection measures. Guerra-Garcia and her colleagues looked to develop a zoning approach based on physics, rather than historical flight data. Such a physics-based mapping could be applied to any shape of aircraft, such as unconventional and largely untested designs, to identify regions that really require reinforcement.

“Protecting aircraft from lightning is heavy,” Jenkins says. “Embedding copper mesh or foil throughout an aircraft is an added weight penalty. And if we had the greatest level of protection for every part of the plane’s surface, the plane would weigh far too much. So zoning is about trying to optimize the weight of the system while also having it be as safe as possible.”

In the zone

For their new approach, the team developed a model to predict the pattern of lightning sweep and the corresponding lightning protection zones, for a given airplane geometry. Starting with a specific airplane shape — in their case, a typical tube-and-wing structure — the researchers simulated the fluid dynamics, or how air would flow around a plane, given a certain speed, altitude, and pitch angle. They also incorporated their previous model that predicts the places where lightning is more likely to initially attach.

For each initial attachment point, the team simulated tens of thousands of potential lightning arcs, or angles from which the current strikes the plane. They then ran the model forward to predict how the tens of thousands of potential strikes would follow the air flow across the plane’s surface. These runs produced a statistical representation of where lightning, striking a specific point on a plane, is likely to flow and potentially cause damage. The team converted this statistical representation into a map of zones of varying vulnerability.

They validated the method on a conventional tube-and-wing structure, showing that the zoning maps generated by the physics-based approach were consistent with what the aviation industry has determined over decades of fine-tuning.

“We now have a physics-based tool that provides some metrics like the probability of lightning attachment and dwell time, which is how long an arc will linger at a specific point,” Guerra-Garcia explains. “We convert those physics metrics into zoning maps to show, if I’m in this red region, the lightning arc will stay for a long time, so that region needs to be heavily protected.”

The team is starting to apply the approach to new geometries, such as blended-wing designs and truss-braced structures. The researchers envision that the tool can help designers incorporate safe and efficient lightning-protection systems early on in the design process.

“Lightning is incredible and terrifying at the same time, and I have full confidence in flying on planes at the moment,” Jenkins says. “I want to have that same confidence in 20 years’ time. So, we need a new way to zone aircraft.”

“With physics-based methods like the ones developed with professor Guerra-Garcia’s group we have the opportunity to shape industry standards and as an industry rely on the underlying physics to develop guidelines for aircraft certification through simulation,” says co-author Louisa Michael of Boeing Technology Innovation. Currently, we are engaging with industrial committees to propose these methods to be included in Aerospace Recommended Practices.”

“Zoning unconventional aircraft is not an easy task,” adds co-author Ben Westin of Boeing Technology Innovation. “But these methods will allow us to confidently identify which threat levels each part of the aircraft needs to be protected against and certified for, and they give our design engineers a platform to do their best work to optimize aircraft design.”

Beyond airplanes, Guerra-Garcia is looking at ways to adapt the lightning protection model to other technologies, including wind turbines.

“About 60 percent of blade losses are due to lightning and will become worse as we move offshore because wind turbines will be even bigger and more susceptible to upward lightning,” she says. “They have many of the same challenges of a flowing gas environment. It’s more complex, and we will apply this same sort of methodology to this space.”

This research was funded, in part, by the Boeing Company.

Quantum computers have the potential to model new molecules and weather patterns better than any computer today. They may also one day accelerate artificial intelligence algorithms at a much lower energy footprint. But anyone interested in using quantum computers faces a steep learning curve that starts with getting access to quantum devices and then figuring out one of the many quantum software programs on the market.

Now qBraid, founded by Kanav Setia and Jason Necaise ’20, is providing a gateway to quantum computing with a platform that gives users access to the leading quantum devices and software. Users can log on to qBraid’s cloud-based interface and connect with quantum devices and other computing resources from leading companies like Nvidia, Microsoft, and IBM. In a few clicks, they can start coding or deploy cutting-edge software that works across devices.

“The mission is to take you from not knowing anything about quantum computing to running your first program on these amazing machines in less than 10 minutes,” Setia says. “We’re a one-stop platform that gives access to everything the quantum ecosystem has to offer. Our goal is to enable anyone — whether they’re enterprise customers, academics, or individual users — to build and ultimately deploy applications.”

Since its founding in June of 2020, qBraid has helped more than 20,000 people in more than 120 countries deploy code on quantum devices. That traction is ultimately helping to drive innovation in a nascent industry that’s expected to play a key role in our future.

“This lowers the barrier to entry for a lot of newcomers,” Setia says. “They can be up and running in a few minutes instead of a few weeks. That’s why we’ve gotten so much adoption around the world. We’re one of the most popular platforms for accessing quantum software and hardware.”

A quantum “software sandbox”

Setia met Necaise while the two interned at IBM. At the time, Necaise was an undergraduate at MIT majoring in physics, while Setia was at Dartmouth College. The two enjoyed working together, and Necaise said if Setia ever started a company, he’d be interested in joining.

A few months later, Setia decided to take him up on the offer. At Dartmouth, Setia had taken one of the first applied quantum computing classes, but students spent weeks struggling to install all the necessary software programs before they could even start coding.

“We hadn’t even gotten close to developing any useful algorithms,” Seita said. “The idea for qBraid was, ‘Why don’t we build a software sandbox in the cloud and give people an easy programming setup out of the box?’ Connection with the hardware would already be done.”

The founders received early support from the MIT Sandbox Innovation Fund and took part in the delta v summer startup accelerator run by the Martin Trust Center for MIT Entrepreneurship.

“Both programs provided us with very strong mentorship,” Setia says. “They give you frameworks on what a startup should look like, and they bring in some of the smartest people in the world to mentor you — people you’d never have access to otherwise.”

Necaise left the company in 2021. Setia, meanwhile, continued to find problems with quantum software outside of the classroom.

“This is a massive bottleneck,” Setia says. “I’d worked on several quantum software programs that pushed out updates or changes, and suddenly all hell broke loose on my codebase. I’d spend two to four weeks jostling with these updates that had almost nothing to do with the quantum algorithms I was working on.”

QBraid started as a platform with pre-installed software that let developers start writing code immediately. The company also added support for version-controlled quantum software so developers could build applications on top without worrying about changes. Over time, qBraid added connections to quantum computers and tools that lets quantum programs run across different devices.

“The pitch was you don’t need to manage a bunch of software or a whole bunch of cloud accounts,” Setia says. “We’re a single platform: the quantum cloud.”

QBraid also launched qBook, a learning platform that offers interactive courses in quantum computing.

“If you see a piece of code you like, you just click play and the code runs,” Setia says. “You can run a whole bunch of code, modify it on the fly, and you can understand how it works. It runs on laptops, iPads, and phones. A significant portion of our users are from developing countries, and they’re developing applications from their phones.”

Democratizing quantum computing

Today qBraid’s 20,000 users come from over 400 universities and 100 companies around the world. As qBraid’s user base has grown, the company went from integrating quantum computers onto their platform from the outside to creating a quantum operating system, qBraid-OS, that is currently being used by four leading quantum companies.

“We are productizing these quantum computers,” Setia explains. “Many quantum companies are realizing they want to focus their energy completely on the hardware, with us productizing their infrastructure. We’re like the operating system for quantum computers.”

People are using qBraid to build quantum applications in AI and machine learning, to discover new molecules or develop new drugs, and to develop applications in finance and cybersecurity. With every new use case, Setia says qBraid is democratizing quantum computing to create the quantum workforce that will continue to advance the field.

“[In 2018], an article in The New York Times said there were possibly less than 1,000 people in the world that could be called experts in quantum programming,” Setia says. “A lot of people want to access these cutting-edge machines, but they don’t have the right software backgrounds. They are just getting started and want to play with algorithms. QBraid gives those people an easy programming setup out of the box.”

With the rapid advancement of generative artificial intelligence, teachers and school leaders are looking for answers to complicated questions about successfully integrating technology into lessons, while also ensuring students actually learn what they’re trying to teach. 

Justin Reich, an associate professor in MIT’s Comparative Media Studies/Writing program, hopes a new guidebook published by the MIT Teaching Systems Lab can support K-12 educators as they determine what AI policies or guidelines to craft.

“Throughout my career, I’ve tried to be a person who researches education and technology and translates findings for people who work in the field,” says Reich. “When tricky things come along I try to jump in and be helpful.” 

“A Guide to AI in Schools: Perspectives for the Perplexed,” published this fall, was developed with the support of an expert advisory panel and other researchers. The project includes input from more than 100 students and teachers from around the United States, sharing their experiences teaching and learning with new generative AI tools. 

“We’re trying to advocate for an ethos of humility as we examine AI in schools,” Reich says. “We’re sharing some examples from educators about how they’re using AI in interesting ways, some of which might prove sturdy and some of which might prove faulty. And we won’t know which is which for a long time.”

Finding answers to AI and education questions

The guidebook attempts to help K-12 educators, students, school leaders, policymakers, and others collect and share information, experiences, and resources. AI’s arrival has left schools scrambling to respond to multiple challenges, like how to ensure academic integrity and maintain data privacy. 

Reich cautions that the guidebook is not meant to be prescriptive or definitive, but something that will help spark thought and discussion. 

“Writing a guidebook on generative AI in schools in 2025 is a little bit like writing a guidebook of aviation in 1905,” the guidebook’s authors note. “No one in 2025 can say how best to manage AI in schools.”

Schools are also struggling to measure how student learning loss looks in the age of AI. “How does bypassing productive thinking with AI look in practice?” Reich asks. “If we think teachers provide content and context to support learning and students no longer perform the exercises housing the content and providing the context, that’s a serious problem.”

Reich invites people directly impacted by AI to help develop solutions to the challenges its ubiquity presents. “It’s like observing a conversation in the teacher’s lounge and inviting students, parents, and other people to participate about how teachers think about AI,” he says, “what they are seeing in their classrooms, and what they’ve tried and how it went.”

The guidebook, in Reich’s view, is ultimately a collection of hypotheses expressed in interviews with teachers: well-informed, initial guesses about the paths that schools could follow in the years ahead. 

Producing educator resources in a podcast

In addition to the guidebook, the Teaching Systems Lab also recently produced “The Homework Machine,” a seven-part series from the Teachlab podcast that explores how AI is reshaping K-12 education. 

Reich produced the podcast in collaboration with journalist Jesse Dukes. Each episode tackles a specific area, asking important questions about challenges related to issues like AI adoption, poetry as a tool for student engagement, post-Covid learning loss, pedagogy, and book bans. The podcast allows Reich to share timely information about education-related updates and collaborate with people interested in helping further the work.

“The academic publishing cycle doesn’t lend itself to helping people with near-term challenges like those AI presents,” Reich says. “Peer review takes a long time, and the research produced isn’t always in a form that’s helpful to educators.” Schools and districts are grappling with AI in real time, bypassing time-tested quality control measures. 

The podcast can help reduce the time it takes to share, test, and evaluate AI-related solutions to new challenges, which could prove useful in creating training and resources.  

“We hope the podcast will spark thought and discussion, allowing people to draw from others’ experiences,” Reich says.

The podcast was also produced into an hour-long radio special, which was broadcast by public radio stations across the country.

“We’re fumbling around in the dark”

Reich is direct in his assessment of where we are with understanding AI and its impacts on education. “We’re fumbling around in the dark,” he says, recalling past attempts to quickly integrate new tech into classrooms. These failures, Reich suggests, highlight the importance of patience and humility as AI research continues. “AI bypassed normal procurement processes in education; it just showed up on kids’ phones,” he notes. 

“We’ve been really wrong about tech in the past,” Reich says. Despite districts’ spending on tools like smartboards, for example, research indicates there’s no evidence that they improve learning or outcomes. In a new article for article for The Conversation, he argues that early teacher guidance in areas like web literacy has produced bad advice that still exists in our educational system. “We taught students and educators not to trust Wikipedia,” he recalls, “and to search for website credibility markers, both of which turned out to be incorrect.” Reich wants to avoid a similar rush to judgment on AI, recommending that we avoid guessing at AI-enabled instructional strategies.

These challenges, coupled with potential and observed student impacts, significantly raise the stakes for schools and students’ families in the AI race. “Education technology always provokes teacher anxiety,” Reich notes, “but the breadth of AI-related concerns is much greater than in other tech-related areas.” 

The dawn of the AI age is different from how we’ve previously received tech into our classrooms, Reich says. AI wasn’t adopted like other tech. It simply arrived. It’s now upending educational models and, in some cases, complicating efforts to improve student outcomes.

Reich is quick to point out that there are no clear, definitive answers on effective AI implementation and use in classrooms; those answers don’t currently exist. Each of the resources Reich helped develop invite engagement from the audiences they target, aggregating valuable responses others might find useful.

“We can develop long-term solutions to schools’ AI challenges, but it will take time and work,” he says. “AI isn’t like learning to tie knots; we don’t know what AI is, or is going to be, yet.” 

Reich also recommends learning more about AI implementation from a variety of sources. “Decentralized pockets of learning can help us test ideas, search for themes, and collect evidence on what works,” he says. “We need to know if learning is actually better with AI.” 

While teachers don’t get to choose regarding AI’s existence, Reich believes it’s important that we solicit their input and involve students and other stakeholders to help develop solutions that improve learning and outcomes. 

“Let’s race to answers that are right, not first,” Reich says.

A recent study from Oregon State University estimated that more than 3,500 animal species are at risk of extinction because of factors including habitat alterations, natural resources being overexploited, and climate change.

To better understand these changes and protect vulnerable wildlife, conservationists like MIT PhD student and Computer Science and Artificial Intelligence Laboratory (CSAIL) researcher Justin Kay are developing computer vision algorithms that carefully monitor animal populations. A member of the lab of MIT Department of Electrical Engineering and Computer Science assistant professor and CSAIL principal investigator Sara Beery, Kay is currently working on tracking salmon in the Pacific Northwest, where they provide crucial nutrients to predators like birds and bears, while managing the population of prey, like bugs.

With all that wildlife data, though, researchers have lots of information to sort through and many AI models to choose from to analyze it all. Kay and his colleagues at CSAIL and the University of Massachusetts Amherst are developing AI methods that make this data-crunching process much more efficient, including a new approach called “consensus-driven active model selection” (or “CODA”) that helps conservationists choose which AI model to use. Their work was named a Highlight Paper at the International Conference on Computer Vision (ICCV) in October.

That research was supported, in part, by the National Science Foundation, Natural Sciences and Engineering Research Council of Canada, and Abdul Latif Jameel Water and Food Systems Lab (J-WAFS). Here, Kay discusses this project, among other conservation efforts.

Q: In your paper, you pose the question of which AI models will perform the best on a particular dataset. With as many as 1.9 million pre-trained models available in the HuggingFace Models repository alone, how does CODA help us address that challenge?

A: Until recently, using AI for data analysis has typically meant training your own model. This requires significant effort to collect and annotate a representative training dataset, as well as iteratively train and validate models. You also need a certain technical skill set to run and modify AI training code. The way people interact with AI is changing, though — in particular, there are now millions of publicly available pre-trained models that can perform a variety of predictive tasks very well. This potentially enables people to use AI to analyze their data without developing their own model, simply by downloading an existing model with the capabilities they need. But this poses a new challenge: Which model, of the millions available, should they use to analyze their data? 

Typically, answering this model selection question also requires you to spend a lot of time collecting and annotating a large dataset, albeit for testing models rather than training them. This is especially true for real applications where user needs are specific, data distributions are imbalanced and constantly changing, and model performance may be inconsistent across samples. Our goal with CODA was to substantially reduce this effort. We do this by making the data annotation process “active.” Instead of requiring users to bulk-annotate a large test dataset all at once, in active model selection we make the process interactive, guiding users to annotate the most informative data points in their raw data. This is remarkably effective, often requiring users to annotate as few as 25 examples to identify the best model from their set of candidates. 

We’re very excited about CODA offering a new perspective on how to best utilize human effort in the development and deployment of machine-learning (ML) systems. As AI models become more commonplace, our work emphasizes the value of focusing effort on robust evaluation pipelines, rather than solely on training.

Q: You applied the CODA method to classifying wildlife in images. Why did it perform so well, and what role can systems like this have in monitoring ecosystems in the future?

A: One key insight was that when considering a collection of candidate AI models, the consensus of all of their predictions is more informative than any individual model’s predictions. This can be seen as a sort of “wisdom of the crowd:” On average, pooling the votes of all models gives you a decent prior over what the labels of individual data points in your raw dataset should be. Our approach with CODA is based on estimating a “confusion matrix” for each AI model — given the true label for some data point is class X, what is the probability that an individual model predicts class X, Y, or Z? This creates informative dependencies between all of the candidate models, the categories you want to label, and the unlabeled points in your dataset.

Consider an example application where you are a wildlife ecologist who has just collected a dataset containing potentially hundreds of thousands of images from cameras deployed in the wild. You want to know what species are in these images, a time-consuming task that computer vision classifiers can help automate. You are trying to decide which species classification model to run on your data. If you have labeled 50 images of tigers so far, and some model has performed well on those 50 images, you can be pretty confident it will perform well on the remainder of the (currently unlabeled) images of tigers in your raw dataset as well. You also know that when that model predicts some image contains a tiger, it is likely to be correct, and therefore that any model that predicts a different label for that image is more likely to be wrong. You can use all these interdependencies to construct probabilistic estimates of each model’s confusion matrix, as well as a probability distribution over which model has the highest accuracy on the overall dataset. These design choices allow us to make more informed choices over which data points to label and ultimately are the reason why CODA performs model selection much more efficiently than past work.

There are also a lot of exciting possibilities for building on top of our work. We think there may be even better ways of constructing informative priors for model selection based on domain expertise — for instance, if it is already known that one model performs exceptionally well on some subset of classes or poorly on others. There are also opportunities to extend the framework to support more complex machine-learning tasks and more sophisticated probabilistic models of performance. We hope our work can provide inspiration and a starting point for other researchers to keep pushing the state of the art.

Q: You work in the Beerylab, led by Sara Beery, where researchers are combining the pattern-recognition capabilities of machine-learning algorithms with computer vision technology to monitor wildlife. What are some other ways your team is tracking and analyzing the natural world, beyond CODA?

A: The lab is a really exciting place to work, and new projects are emerging all the time. We have ongoing projects monitoring coral reefs with drones, re-identifying individual elephants over time, and fusing multi-modal Earth observation data from satellites and in-situ cameras, just to name a few. Broadly, we look at emerging technologies for biodiversity monitoring and try to understand where the data analysis bottlenecks are, and develop new computer vision and machine-learning approaches that address those problems in a widely applicable way. It’s an exciting way of approaching problems that sort of targets the “meta-questions” underlying particular data challenges we face. 

The computer vision algorithms I’ve worked on that count migrating salmon in underwater sonar video are examples of that work. We often deal with shifting data distributions, even as we try to construct the most diverse training datasets we can. We always encounter something new when we deploy a new camera, and this tends to degrade the performance of computer vision algorithms. This is one instance of a general problem in machine learning called domain adaptation, but when we tried to apply existing domain adaptation algorithms to our fisheries data we realized there were serious limitations in how existing algorithms were trained and evaluated. We were able to develop a new domain adaptation framework, published earlier this year in Transactions on Machine Learning Research, that addressed these limitations and led to advancements in fish counting, and even self-driving and spacecraft analysis.

One line of work that I’m particularly excited about is understanding how to better develop and analyze the performance of predictive ML algorithms in the context of what they are actually used for. Usually, the outputs from some computer vision algorithm — say, bounding boxes around animals in images — are not actually the thing that people care about, but rather a means to an end to answer a larger problem — say, what species live here, and how is that changing over time? We have been working on methods to analyze predictive performance in this context and reconsider the ways that we input human expertise into ML systems with this in mind. CODA was one example of this, where we showed that we could actually consider the ML models themselves as fixed and build a statistical framework to understand their performance very efficiently. We have been working recently on similar integrated analyses combining ML predictions with multi-stage prediction pipelines, as well as ecological statistical models. 

The natural world is changing at unprecedented rates and scales, and being able to quickly move from scientific hypotheses or management questions to data-driven answers is more important than ever for protecting ecosystems and the communities that depend on them. Advancements in AI can play an important role, but we need to think critically about the ways that we design, train, and evaluate algorithms in the context of these very real challenges.

By stimulating cancer cells to produce a molecule that activates a signaling pathway in nearby immune cells, MIT researchers have found a way to force tumors to trigger their own destruction.

Activating this signaling pathway, known as the cGAS-STING pathway, worked even better when combined with existing immunotherapy drugs known as checkpoint blockade inhibitors, in a study of mice. That dual treatment was successfully able to control tumor growth.

The researchers turned on the cGAS-STING pathway in immune cells using messenger RNA delivered to cancer cells. This approach may avoid the side effects of delivering large doses of a STING activator, and takes advantage of a natural process in the body. This could make it easier to develop a treatment for use in patients, the researchers say.

“Our approach harnesses the tumor’s own machinery to produce immune-stimulating molecules, creating a powerful antitumor response,” says Natalie Artzi, a principal research scientist at MIT’s Institute for Medical Engineering and Science, an associate professor of medicine at Harvard Medical School, a core faculty member at the Wyss Institute for Biologically Inspired Engineering at Harvard, and the senior author of the study.

“By increasing cGAS levels inside cancer cells, we can enhance delivery efficiency — compared to targeting the more scarce immune cells in the tumor microenvironment — and stimulate the natural production of cGAMP, which then activates immune cells locally,” she says. “This strategy not only strengthens antitumor immunity but also reduces the toxicity associated with direct STING agonist delivery, bringing us closer to safer and more effective cancer immunotherapies.”

Alexander Cryer, a visiting scholar at IMES, is the lead author of the paper, which appears this week in the Proceedings of the National Academy of Sciences.

Immune activation

STING (short for stimulator of interferon genes) is a protein that helps to trigger immune responses. When STING is activated, it turns on a pathway that initiates production of type one interferons, which are cytokines that stimulate immune cells.

Many research groups, including Artzi’s, have explored the possibility of artificially stimulating this pathway with molecules called STING agonists, which could help immune cells to recognize and attack tumor cells. This approach has worked well in animal models, but it has had limited success in clinical trials, in part because the required doses can cause harmful side effects.

While working on a project exploring new ways to deliver STING agonists, Cryer became intrigued when he learned from previous work that cancer cells can produce a STING activator known as cGAMP. The cells then secrete cGAMP, which can activate nearby immune cells.

“Part of my philosophy of science is that I really enjoy using endogenous processes that the body already has, and trying to utilize them in a slightly different context. Evolution has done all the hard work. We just need to figure out how push it in a different direction,” Cryer says. “Once I saw that cancer cells produce this molecule, I thought: Maybe there’s a way to take this process and supercharge it.”

Within cells, the production of cGAMP is catalyzed by an enzyme called cGAS. To get tumor cells to activate STING in immune cells, the researchers devised a way to deliver messenger RNA that encodes cGAS. When this enzyme detects double-stranded DNA in the cell body, which can be a sign of either infection or cancer-induced damage, it begins producing cGAMP.

“It just so happens that cancer cells, because they’re dividing so fast and not particularly accurately, tend to have more double-stranded DNA fragments than healthy cells,” Cryer says.

The tumor cells then release cGAMP into tumor microenvironment, where it can be taken up by neighboring immune cells and activate their STING pathway.

Targeting tumors

Using a mouse model of melanoma, the researchers evaluated their new strategy’s potential to kill cancer cells. They injected mRNA encoding cGAS, encapsulated in lipid nanoparticles, into tumors. One group of mice received this treatment alone, while another received a checkpoint blockade inhibitor, and a third received both treatments.

Given on their own, cGAS and the checkpoint inhibitor each significantly slowed tumor growth. However, the best results were seen in the mice that received both treatments. In that group, tumors were completely eradicated in 30 percent of the mice, while none of the tumors were fully eliminated in the groups that received just one treatment.

An analysis of the immune response showed that the mRNA treatment stimulated production of interferon as well as many other immune signaling molecules. A variety of immune cells, including macrophages and dendritic cells, were activated. These cells help to stimulate T cells, which can then destroy cancer cells.

The researchers were able to elicit these responses with just a small dose of cancer-cell-produced cGAMP, which could help to overcome one of the potential obstacles to using cGAMP on its own as therapy: Large doses are required to stimulate an immune response, and these doses can lead to widespread inflammation, tissue damage, and autoimmune reactions. When injected on its own, cGAMP tends to spread through the body and is rapidly cleared from the tumor, while in this study, the mRNA nanoparticles and cGAMP remained at the tumor site.

“The side effects of this class of molecule can be pretty severe, and one of the potential advantages of our approach is that you’re able to potentially subvert some toxicity that you might see if you’re giving the free molecules,” Cryer says.

The researchers now hope to work on adapting the delivery system so that it could be given as a systemic injection, rather than injecting it into the tumor. They also plan to test the mRNA therapy in combination with chemotherapy drugs or radiotherapy that damage DNA, which could make the therapy even more effective because there could be even more double-stranded DNA available to help activate the synthesis of cGAMP.

Managing a power grid is like trying to solve an enormous puzzle.

Grid operators must ensure the proper amount of power is flowing to the right areas at the exact time when it is needed, and they must do this in a way that minimizes costs without overloading physical infrastructure. Even more, they must solve this complicated problem repeatedly, as rapidly as possible, to meet constantly changing demand.

To help crack this consistent conundrum, MIT researchers developed a problem-solving tool that finds the optimal solution much faster than traditional approaches while ensuring the solution doesn’t violate any of the system’s constraints. In a power grid, constraints could be things like generator and line capacity.

This new tool incorporates a feasibility-seeking step into a powerful machine-learning model trained to solve the problem. The feasibility-seeking step uses the model’s prediction as a starting point, iteratively refining the solution until it finds the best achievable answer.

The MIT system can unravel complex problems several times faster than traditional solvers, while providing strong guarantees of success. For some extremely complex problems, it could find better solutions than tried-and-true tools. The technique also outperformed pure machine learning approaches, which are fast but can’t always find feasible solutions.

In addition to helping schedule power production in an electric grid, this new tool could be applied to many types of complicated problems, such as designing new products, managing investment portfolios, or planning production to meet consumer demand.

“Solving these especially thorny problems well requires us to combine tools from machine learning, optimization, and electrical engineering to develop methods that hit the right tradeoffs in terms of providing value to the domain, while also meeting its requirements. You have to look at the needs of the application and design methods in a way that actually fulfills those needs,” says Priya Donti, the Silverman Family Career Development Professor in the Department of Electrical Engineering and Computer Science (EECS) and a principal investigator at the Laboratory for Information and Decision Systems (LIDS).

Donti, senior author of an open-access paper on this new tool, called FSNet, is joined by lead author Hoang Nguyen, an EECS graduate student. The paper will be presented at the Conference on Neural Information Processing Systems.

Combining approaches

Ensuring optimal power flow in an electric grid is an extremely hard problem that is becoming more difficult for operators to solve quickly.

“As we try to integrate more renewables into the grid, operators must deal with the fact that the amount of power generation is going to vary moment to moment. At the same time, there are many more distributed devices to coordinate,” Donti explains.

Grid operators often rely on traditional solvers, which provide mathematical guarantees that the optimal solution doesn’t violate any problem constraints. But these tools can take hours or even days to arrive at that solution if the problem is especially convoluted.

On the other hand, deep-learning models can solve even very hard problems in a fraction of the time, but the solution might ignore some important constraints. For a power grid operator, this could result in issues like unsafe voltage levels or even grid outages.

“Machine-learning models struggle to satisfy all the constraints due to the many errors that occur during the training process,” Nguyen explains.

For FSNet, the researchers combined the best of both approaches into a two-step problem-solving framework.

Focusing on feasibility

In the first step, a neural network predicts a solution to the optimization problem. Very loosely inspired by neurons in the human brain, neural networks are deep learning models that excel at recognizing patterns in data.

Next, a traditional solver that has been incorporated into FSNet performs a feasibility-seeking step. This optimization algorithm iteratively refines the initial prediction while ensuring the solution does not violate any constraints.

Because the feasibility-seeking step is based on a mathematical model of the problem, it can guarantee the solution is deployable.

“This step is very important. In FSNet, we can have the rigorous guarantees that we need in practice,” Hoang says.

The researchers designed FSNet to address both main types of constraints (equality and inequality) at the same time. This makes it easier to use than other approaches that may require customizing the neural network or solving for each type of constraint separately.

“Here, you can just plug and play with different optimization solvers,” Donti says.

By thinking differently about how the neural network solves complex optimization problems, the researchers were able to unlock a new technique that works better, she adds.

They compared FSNet to traditional solvers and pure machine-learning approaches on a range of challenging problems, including power grid optimization. Their system cut solving times by orders of magnitude compared to the baseline approaches, while respecting all problem constraints.

FSNet also found better solutions to some of the trickiest problems.

“While this was surprising to us, it does make sense. Our neural network can figure out by itself some additional structure in the data that the original optimization solver was not designed to exploit,” Donti explains.

In the future, the researchers want to make FSNet less memory-intensive, incorporate more efficient optimization algorithms, and scale it up to tackle more realistic problems.

“Finding solutions to challenging optimization problems that are feasible is paramount to finding ones that are close to optimal. Especially for physical systems like power grids, close to optimal means nothing without feasibility. This work provides an important step toward ensuring that deep-learning models can produce predictions that satisfy constraints, with explicit guarantees on constraint enforcement,” says Kyri Baker, an associate professor at the University of Colorado Boulder, who was not involved with this work.

"A persistent challenge for machine learning-based optimization is feasibility. This work elegantly couples end-to-end learning with an unrolled feasibility-seeking procedure that minimizes equality and inequality violations. The results are very promising and I look forward to see where this research will head," adds Ferdinando Fioretto, an assistant professor at the University of Virginia, who was not involved with this work.

Good management of aid projects in developing countries reduces violence in those areas — but poorly managed projects increase the chances of local violence, according to a new study by an MIT economist.

The research, examining World Bank projects in Africa, illuminates a major question surrounding international aid. Observers have long wondered if aid projects, by bringing new resources into developing countries, lead to conflict over those goods as an unintended consequence. Previously, some scholars have identified an increase in violence attached to aid, while others have found a decrease.

The new study shows those prior results are not necessarily wrong, but not entirely right, either. Instead, aid oversight matters. World Bank programs earning the highest evaluation scores for their implementation reduce the likelihood of conflict by up to 12 percent, compared to the worst-managed programs.

“I find that the management quality of these projects has a really strong effect on whether that project leads to conflict or not,” says MIT economist Jacob Moscona, who conducted the research. “Well-managed aid projects can actually reduce conflict, and poorly managed projects increase conflict, relative to no project. So, the way aid programs are organized is very important.”

The findings also suggest aid projects can work well almost anywhere. At times, observers have suggested the political conditions in some countries prevent aid from being effective. But the new study finds otherwise.

“There are ways these programs can have their positive effects without the negative consequences,” Moscona says. “And it’s not the result of what politics looks like on the receiving end; it’s about the organization itself.”

Moscona’s paper detailing the study, “The Management of Aid and Conflict in Africa,” is published in the November issue of the American Economic Journal: Economic Policy. Moscona, the paper’s sole author, is the 3M Career Development Assistant Professor in MIT’s Department of Economics.

Decisions on the ground

To conduct the study, Moscona examined World Bank data from the 1997-2014 time period, using the information compiled by AidData, a nonprofit group that also studies World Bank programs. Importantly, the World Bank conducts extensive evaluations of its projects and includes the identities of project leaders as part of those reviews.

“There are a lot of decisions on the ground made by managers of aid, and aid organizations themselves, that can have a huge impact on whether or not aid leads to conflict, and how aid resources are used and whether they are misappropriated or captured and get into the wrong hands,” Moscona says.

For instance, diligent daily checks about food distribution programs can and have substantially reduced the amount of food that is stolen or “leaks” out of the program. Other projects have created innovative ways of tagging small devices to ensure those objects are used by program participants, reducing appropriation by others.

Moscona combined the World Bank data with statistics from the Armed Conflict Location and Event Data Project (ACLED), a nonprofit that monitors political violence. That enabled him to evaluate how the quality of aid project implementation — and even the quality of the project leadership — influenced local outcomes.

For instance, by looking at the ratings of World Bank project leaders, Moscona found that shifting from a project leader at the 25th percentile, in terms of how frequently projects are linked with conflict, to one at the 75th percentile, increases the chances of local conflict by 15 percent.

“The magnitudes are pretty large, in terms of the probability that a conflict starts in the vicinity of a project,” Moscona observes.

Moscona’s research identified several other aspects of the interaction between aid and conflict that hold up over the region and time period. The establishment of aid programs does not seem to lead to long-term strategic activity by non-government forces, such as land acquisition or the establishment of rebel bases. The effects are also larger in areas that have had recent political violence. And armed conflict is greater when the resources at stake can be expropriated — such as food or medical devices.

“It matters most if you have more divertable resources, like food and medical devices that can be captured, as opposed to infrastructure projects,” Moscona says.

Reconciling the previous results

Moscona also found a clear trend in the data about the timing of violence in relation to aid. Government and other armed groups do not engage in much armed conflict when aid programs are being established; it is the appearance of desired goods themselves that sets off violent activity.

“You don’t see much conflict when the projects are getting off the ground,” Moscona says.” You really see the conflict start when the money is coming in or when the resources start to flow. Which is consistent with the idea of the relevant mechanism being about aid resources and their misappropriation, rather than groups trying to deligitimize a project.”

All told, Moscona’s study finds a logical mechanism explaining the varying results other scholars have found with regard to aid and conflict. If aid programs are not equally well-administered, it stands to reason that their outcomes will not be identical, either.

“There wasn’t much work trying to make those two sets of results speak to each other,” says Moscona. “I see it less as overturning existing results than providing a way to reconcile different results and experiences.”

Moscona’s findings may also speak to the value of aid in general — and provide actionable ideas for institutions such as the World Bank. If better management makes such a difference, then the potential effectiveness of aid programs may increase.

“One goal is to change the conversation about aid,” Moscona says. The data, he suggests, shows that the public discourse about aid can be “less defeatist about the potential negative consequences of aid, and the idea that it’s out of the control of the people who administer it.”