What if the Trojan horse had been pulled to pieces, revealing the ruse and fending off the invasion, just as it entered the gates of Troy?

That’s an apt description of a newly characterized bacterial defense system that chops up foreign DNA.

Bacteria and the viruses that infect them, bacteriophages — phages for short — are ceaselessly at odds, with bacteria developing methods to protect themselves against phages that are constantly striving to overcome those safeguards.

New research from the Department of Biology at MIT, recently published in Nature, describes a defense system that is integrated into the protective membrane that encapsulates bacteria. SNIPE, which stands for surface-associated nuclease inhibiting phage entry, contains a nuclease domain that cleaves genetic material, chopping the invading phage genome into harmless fragments before it can appropriate the host’s molecular machinery to make more phages. 

Daniel Saxton, a postdoc in the Laub Lab and the paper’s first author, was initially drawn to studying this bacterial defense system in E. coli, in part because it is highly unusual to have a nuclease that localizes to the membrane, as most nucleases are free-floating in the cytoplasm, the gelatinous fluid that fills the space inside cells.

“The other thing that caught my attention is that this is something we call a direct defense system, meaning that when a phage infects a cell, that cell will actually survive the attack,” Saxton says. “It’s hard to fend off a phage directly in a cell and survive — but this defense system can do it.” 

Light it up

For Saxton, the project came into focus during a fluorescence-based experiment in which viral genetic material would light up if it successfully penetrated the bacteria. 

“SNIPE was obliterating the phage DNA so fast that we couldn’t even see a fluorescent spot,” Saxton recalls. “I don’t think I’ve ever seen such an effective defense system before — you can barrage the bacteria with hundreds of phage per cell, but SNIPE is like god-tier protection.”

When the nuclease domain of SNIPE was mutated so it couldn’t chop up DNA, fluorescent spots appeared as usual, and the bacteria succumbed to the phage infection. 

Bacteria maintain tight control over all their defense systems, lest they be turned against their host. Some systems remain dormant until they flare up, for example, to halt all translation of all proteins in the cell, while others can distinguish between bacterial DNA and foreign, invading phage DNA. There were only two previously characterized mechanisms in the latter category before researchers uncovered SNIPE. 

“Right now, the phage field is at a really interesting spot where people are discovering phage defense systems at a breakneck pace,” Saxton says. 

Problems at the periphery

Saxton says they had to approach the work in a somewhat roundabout way because there are currently no published structures depicting all the steps of phage genome injection. Studying processes at the membrane is challenging: Membranes are dense and chaotic, and phage genome injection is a highly transient process, lasting only a few minutes. 

SNIPE seems to discern viral DNA by interacting with proteins the phage uses to tunnel through the bacteria’s protective membrane. This “subcellular localization,” according to Saxton, may also prevent SNIPE from inadvertently chopping up the bacteria’s own genetic material.

The model outlined in the paper is that one region of SNIPE binds to a bacterial membrane protein called ManYZ, while another region likely binds to the tape measure protein from the phage. 

The tape measure protein got its name because it determines the length of the phage tail — the part of the phage between the small, leglike protrusions and the bulbous head, which contains the phage’s genetic material. The researchers revealed that the phage’s tape measure protein enters the cytoplasm during injection, a phenomenon that had not been physically demonstrated before. 

There may also be other proteins or interactions involved. 

“If you shunt the phage genome injection through an alternate pathway that isn’t ManYZ, suddenly SNIPE doesn’t defend against the phage nearly as well,” Saxton says. “It’s unclear exactly how these proteins interact, but we do know that these two proteins are involved in this genome injection process.” 

Future directions

Saxton hopes that future work will expand our understanding of what occurs during phage genome injection and uncover the structures of the proteins involved, especially the tunnel complex in the membrane through which phages insert their genome.

Members of the Laub Lab are already collaborating with another lab to determine the structure of SNIPE. In the meantime, Saxton has been working on a new defense system in which molecular mimicry — bacterial proteins imitating phage proteins — may play a role. 

Michael T. Laub, the Salvador E. Luria Professor of Biology and a Howard Hughes Medical Institute investigator, notes that one of the breakthrough experiments for demonstrating how SNIPE works came from a brainstorming session at a lab retreat.

“Daniel and I were kind of stuck with how to directly measure the effect of SNIPE during infection, but another postdoc in the lab, Ian Roney, who is a co-author on the paper, came up with a very clever idea that ultimately worked perfectly,” Laub recalls. “It’s a great example of how powerful internal collaborations can be in pushing our science forward.”

Muscles are remarkably effective systems for generating controlled force, and engineers developing hardware for robots or prosthetics have long struggled to create analogs that can approach their unique combination of strength, rapid response, scalability, and control. But now, researchers at the MIT Media Lab and Politecnico di Bari in Italy have developed artificial muscle fibers that come closer to matching many of these qualities.

Like the fibers that bundle together to form biological muscles, these fibers can be arranged in different configurations to meet the demands of a given task. Unlike conventional robotic actuation systems, they are compliant enough to interface comfortably with the human body and operate silently without motors, external pumps, or other bulky supporting hardware.

The new electrofluidic fiber muscles — electrically driven actuators built in fiber format — are described in a recent paper published in Science Robotics. The work is led by Media Lab PhD candidate Ozgun Kilic Afsar; Vito Cacucciolo, a professor at the Politecnico di Bari; and four co-authors.

The new system brings together two technologies, Afsar explains. One is a fluidically driven artificial muscle known as a thin McKibben actuator, and the other is a miniaturized solid-state pump based on electrohydrodynamics (EHD), which can generate pressure inside a sealed fluid compartment without moving parts or an external fluid supply.

Until now, most fluid-driven soft actuators have relied on external “heavy, bulky, oftentimes noisy hydraulic infrastructure,” Afsar says, “which makes them difficult to integrate into systems where mobility or compact, lightweight design is important.” This has created a fundamental bottleneck in the practical use of fluidic actuators in real-world applications.

The key to breaking through that bottleneck was the use of integrated pumps based on electrohydrodynamic principles. These millimeter-scale, electrically driven pumps generate pressure and flow by injecting charge into a dielectric fluid, creating ions that drag the fluid along with them. Weighing just a few grams each and not much thicker than a toothpick, they can be fabricated continuously and scaled easily. “We integrated these fiber pumps into a closed fluidic circuit with the thin McKibben actuators,” Afsar says, noting that this was not a simple task given the different dynamics of the two components.

A key design strategy was to pair these fibers in what are known as antagonistic configurations. Cacucciolo explains that this is where “one muscle contracts while another elongates,” as when you bend your arm and your biceps contract while your triceps stretch. In their system, a millimeter-scale fiber pump sits between two similarly scaled McKibben actuators, driving fluid into one actuator to contract it while simultaneously relaxing the other.

“This is very much reminiscent of how biological muscles are configured and organized,” Afsar says. “We didn’t choose this configuration simply for the sake of biomimicry, but because we needed a way to store the fluid within the muscle design.” The need for an external reservoir open to the atmosphere has been one of the main factors limiting the practical use of EHD pumps in robotic systems outside the lab. By pairing two McKibben fibers in line, with a fiber pump between them to form a closed circuit, the team eliminated that need entirely.

Another key finding was that the muscle fibers needed to be pre-pressurized, rather than simply filled. “There is a minimum internal system pressure that the system can tolerate,” Afsar says, “below which the pump can degrade or temporarily stop working.” This happens because of cavitation, in which vapor bubbles form when the pressure at the pump inlet drops below the vapor pressure of the liquid, eventually leading to dielectric breakdown.

To prevent cavitation, they applied a “bias” pressure from the outset so that the pressure at the fiber pump inlet never falls below the liquid’s vapor pressure. The magnitude of this bias pressure can be adjusted depending on the application. “To achieve the maximum contraction the muscle can generate, we found there is a specific bias pressure range that is optimal,” she says. “If you want to configure the system for faster response, you might increase that bias pressure, though with some reduction in maximum contraction.”

Cacucciolo adds that most of today’s robotic limbs and hands are built around electric servo motors, whose configuration differs fundamentally from that of natural muscles. Servo motors generate rotational motion on a shaft that must be converted into linear movement, whereas muscle fibers naturally contract and extend linearly, as do these electrofluidic fibers. 

“Most robotic arms and humanoid robots are designed around the servo motors that drive them,” he says. “That creates integration constraints, because servo motors are hard to package densely and tend to concentrate mass near the joints they drive. By contrast, artificial muscles in fiber form can be packed tightly inside a robot or exoskeleton and distributed throughout the structure, rather than concentrated near a joint.”

These electrofluidic muscles may be especially useful for wearable applications, such as exoskeletons that help a person lift heavier loads or assistive devices that restore or augment dexterity. But the underlying principles could also apply more broadly. “Our findings extend to fluid-driven robotic systems in general,” Cacucciolo says. “Wherever fluidic actuators are used, or where engineers want to replace external pumps with internal ones, these design principles could apply across a wide range of fluid-driven robotic systems.”

This work “presents a major advancement in fiber-format soft actuation,” which “addresses several long-standing hurdles in the field, particularly regarding portability and power density,” says Herbert Shea, a professor in the Soft Transducers Laboratory at Ecole Polytechnique Federale de Lausanne in Switzerland, who was not associated with this research. “The lack of moving parts in the pump makes these muscles silent, a major advantage for prosthetic devices and assistive clothing,” he says.

Shea adds that “this high-quality and rigorous work bridges the gap between fundamental fluid dynamics and practical robotic applications. The authors provide a complete system-level solution — characterizing the individual components, developing a predictive physical model, and validating it through a range of demonstrators.”

In addition to Afsar and Cacucciolo, the team also included Gabriele Pupillo and Gennaro Vitucci at Politecnico di Bari and Wedyan Babatain and Professor Hiroshi Ishii at the MIT Media Lab. The work was supported by the European Research Council and the Media Lab’s multi-sponsored consortium.

While earning her dual master’s degrees in aeronautics and astronautics and public policy, Carissma McGee SM ’25 learned to navigate between two seemingly distinct worlds, bridging rigorous technical analysis and policy decisions.

As an undergraduate congressional intern and researcher, she saw a persistent gap in space policymaking. Policymakers often lacked technical expertise, while researchers were rarely involved in increasingly complex questions surrounding intellectual property and international collaboration in space.

Her work on intellectual property frameworks for space collaborations directly addresses that gap, combining expertise in gravitational microlensing and space telescope operations with policy analysis to tackle emerging governance challenges.

“I want to bring an expert level in science in the rooms where policy decisions are made,” says McGee, now a doctoral student in aeronautics and astronautics. “That perspective is critical for shaping the future of research and exploration.”

Likewise, she wants to bring her expertise in public policy into the lab.

“I enjoy being able to ask questions about intellectual property, territorial claims, knowledge transfer, or allocation of resources early on in a research project,” adds McGee.

McGee’s fascination with space started during her high school years in Delaware, when she first volunteered at a local observatory and then interned at the NASA Goddard Space Flight Center in Maryland.

Following high school, McGee attended Howard University. She was selected to participate in the Karsh STEM Scholars Program, a full-ride scholarship track for students committed to working continuously toward earning doctoral degrees. Howard, which holds an R1 research classification from the Carnegie Foundation, is in close proximity the Goddard Space Flight Center, as well as the American Astronomical Society and the D.C. Space Grant Consortium.

In 2020, after her first year at Howard, the Covid-19 pandemic sent McGee back to her hometown in Delaware. As it turned out, that gave her an opportunity to work with her local congresswoman, Lisa Blunt Rochester, then a U.S. representative. In addition to supporting the congresswoman’s constituents, she drafted dozens of letters related to STEM education and energy reform.

Working in government gave McGee an opportunity to use her voice to “advocate for astronomy and astrophysics with the American Astronomical Society, advocate for space sciences, and for science representation.”

As an undergraduate, McGee also conducted research linking computational physics and astronomy, working with both NASA’s Jet Propulsion Laboratory and Yale University’s Department of Astronomy. She also continued research begun in 2021 with the Harvard and Smithsonian Center for Astrophysics’ Black Hole Initiative, contributing to work associated with the Event Horizon Telescope.

When she visited MIT in 2023, McGee was struck by the Institute’s openness to interdisciplinary work and support of her interest in combining aeronautics and astronautics with policy.

Once at MIT, she started working in the Space, Telecommunications, Astronomy, and Radiation Laboratory (STAR Lab) with advisor Kerri Cahoy, professor of aeronautics and astronautics. McGee says she experienced a great deal of freedom to craft her own program.

“I was drawn to the lab’s work on satellite missions and CubeSats, and excited to discover that I could pursue exoplanet astrophysics research within this framework and that submitting a dual thesis or focusing on astrophysics applications was possible,” says McGee. “When I expressed interest in participating in the Technology [and] Policy Program for a dual thesis in a framework for space policy, my advisors encouraged me to explore how we could integrate these diverse interests into a path forward.”

In 2024, McGee was awarded a MathWorks Fellowship to pursue research associated with the Nancy Grace Roman Space Telescope and join a NASA mission.

“It was just amazing to join the exoplanet group at NASA,” she says. “I had a front-row seat to see how real researchers and workers navigate complex problems.”

McGee credits MathWorks with helping fellows to “be at the forefront of knowledge and shaping innovation.”

One of her proudest academic accomplishments is PyLIMASS, a software system she developed with collaborators at Louisiana State University, the Ohio State University, and NASA’s Goddard Space Flight Center. The tool enables more accurate mass and distance estimates in gravitational microlensing events, helping the Roman Space Telescope project meet its precision goals for studying exoplanets.

“To build software that didn’t previously exist — and to know it will be used for the Roman mission — is incredibly exciting,” McGee says.

In May 2025, McGee graduated with dual master’s degrees in aeronautics and astronautics and technology and policy. That same month, she presented her research at the American Astronomical Society meeting in Anchorage, Alaska, and at the Technology Management and Policy Conference in Portugal.

McGee remained at MIT to pursue her doctoral degree. Last fall, as an MIT BAMIT Community Advancement Program and Fund Fellow, she hosted a daylong conference for STEM students focused on how intellectual property frameworks shape technical fields.

McGee’s accomplishments and contributions have been celebrated with a number of honors recently. In 2026, she was named Miss Black Massachusetts United States, was recognized among MIT’s Graduate Students of Excellence, and received the MIT MLK Leadership Award in recognition of her service, integrity, and community impact.

Beyond her academic work, McGee is active across campus. She teaches Pilates with MIT Recreation, participates in the Graduate Women in Aerospace Engineering group, and serves as a graduate resident assistant in an undergraduate dorm on East Campus.

She credits the AeroAstro graduate community with keeping her momentum going.

“Even if we’re tired, there’s this powerful camaraderie among AeroAstro graduate students working together. Seeing my peers are pushing through similar research milestones and solve daunting problems motivates you to advance beyond the finish line to further developments in the field.”

Training a large artificial intelligence model is expensive, not just in dollars, but in time, energy, and computational resources. Traditionally, obtaining a smaller, faster model either requires training a massive one first and then trimming it down, or training a small one from scratch and accepting weaker performance. 

Researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), Max Planck Institute for Intelligent Systems, European Laboratory for Learning and Intelligent Systems, ETH, and Liquid AI have now developed a new method that sidesteps this trade-off entirely, compressing models during training, rather than after.

The technique, called CompreSSM, targets a family of AI architectures known as state-space models, which power applications ranging from language processing to audio generation and robotics. By borrowing mathematical tools from control theory, the researchers can identify which parts of a model are pulling their weight and which are dead weight, before surgically removing the unnecessary components early in the training process.

"It's essentially a technique to make models grow smaller and faster as they are training," says Makram Chahine, a PhD student in electrical engineering and computer science, CSAIL affiliate, and lead author of the paper. "During learning, they're also getting rid of parts that are not useful to their development."

The key insight is that the relative importance of different components within these models stabilizes surprisingly early during training. Using a mathematical quantity called Hankel singular values, which measure how much each internal state contributes to the model's overall behavior, the team showed they can reliably rank which dimensions matter and which don't after only about 10 percent of the training process. Once those rankings are established, the less-important components can be safely discarded, and the remaining 90 percent of training proceeds at the speed of a much smaller model.

"What's exciting about this work is that it turns compression from an afterthought into part of the learning process itself,” says senior author Daniela Rus, MIT professor and director of CSAIL. “Instead of training a large model and then figuring out how to make it smaller, CompreSSM lets the model discover its own efficient structure as it learns. That's a fundamentally different way to think about building AI systems.”

The results are striking. On image classification benchmarks, compressed models maintained nearly the same accuracy as their full-sized counterparts while training up to 1.5 times faster. A compressed model reduced to roughly a quarter of its original state dimension achieved 85.7 percent accuracy on the CIFAR-10 benchmark, compared to just 81.8 percent for a model trained at that smaller size from scratch. On Mamba, one of the most widely used state-space architectures, the method achieved approximately 4x training speedups, compressing a 128-dimensional model down to around 12 dimensions while maintaining competitive performance.

"You get the performance of the larger model, because you capture most of the complex dynamics during the warm-up phase, then only keep the most-useful states," Chahine says. "The model is still able to perform at a higher level than training a small model from the start."

What makes CompreSSM distinct from existing approaches is its theoretical grounding. Conventional pruning methods train a full model and then strip away parameters after the fact, meaning you still pay the full computational cost of training the big model. Knowledge distillation, another popular technique, requires training a large "teacher" model to completion and then training a second, smaller "student" model on top of it, essentially doubling the training effort. CompreSSM avoids both of these costs by making informed compression decisions mid-stream.

The team benchmarked CompreSSM head-to-head against both alternatives. Compared to Hankel nuclear norm regularization, a recently proposed spectral technique for encouraging compact state-space models, CompreSSM was more than 40 times faster, while also achieving higher accuracy. The regularization approach slowed training by roughly 16 times because it required expensive eigenvalue computations at every single gradient step, and even then, the resulting models underperformed. Against knowledge distillation on CIFAR-10, CompressSM held a clear advantage for heavily compressed models: At smaller state dimensions, distilled models saw significant accuracy drops, while CompreSSM-compressed models maintained near-full performance. And because distillation requires a forward pass through both the teacher and student at every training step, even its smaller student models trained slower than the full-sized baseline.

The researchers proved mathematically that the importance of individual model states changes smoothly during training, thanks to an application of Weyl's theorem, and showed empirically that the relative rankings of those states remain stable. Together, these findings give practitioners confidence that dimensions identified as negligible early on won't suddenly become critical later.

The method also comes with a pragmatic safety net. If a compression step causes an unexpected performance drop, practitioners can revert to a previously saved checkpoint. "It gives people control over how much they're willing to pay in terms of performance, rather than having to define a less-intuitive energy threshold," Chahine explains.

There are some practical boundaries to the technique. CompreSSM works best on models that exhibit a strong correlation between the internal state dimension and overall performance, a property that varies across tasks and architectures. The method is particularly effective on multi-input, multi-output (MIMO) models, where the relationship between state size and expressivity is strongest. For per-channel, single-input, single-output architectures, the gains are more modest, since those models are less sensitive to state dimension changes in the first place.

The theory applies most cleanly to linear time-invariant systems, although the team has developed extensions for the increasingly popular input-dependent, time-varying architectures. And because the family of state-space models extends to architectures like linear attention, a growing area of interest as an alternative to traditional transformers, the potential scope of application is broad.

Chahine and his collaborators see the work as a stepping stone. The team has already demonstrated an extension to linear time-varying systems like Mamba, and future directions include pushing CompreSSM further into matrix-valued dynamical systems used in linear attention mechanisms, which would bring the technique closer to the transformer architectures that underpin most of today's largest AI systems.

"This had to be the first step, because this is where the theory is neat and the approach can stay principled," Chahine says. "It's the stepping stone to then extend to other architectures that people are using in industry today."

"The work of Chahine and his colleagues provides an intriguing, theoretically grounded perspective on compression for modern state-space models (SSMs)," says Antonio Orvieto, ELLIS Institute Tübingen principal investigator and MPI for Intelligent Systems independent group leader, who wasn't involved in the research. "The method provides evidence that the state dimension of these models can be effectively reduced during training and that a control-theoretic perspective can successfully guide this procedure. The work opens new avenues for future research, and the proposed algorithm has the potential to become a standard approach when pre-training large SSM-based models."

The work, which was accepted as a conference paper at the International Conference on Learning Representations 2026, will be presented later this month. It was supported, in part, by the Max Planck ETH Center for Learning Systems, the Hector Foundation, Boeing, and the U.S. Office of Naval Research.