VibeGen: AI That Designs Molecular Motion in Proteins

The Marvel of Molecular Machines

Proteins are not just dietary nutrients – they are living molecular machines constantly at work in our cells. Unlike the static ribbon diagrams we see in textbooks, proteins twist, flex, and even walk as part of their function. For example, tiny motor proteins literally stride along cellular fibers hauling cargo, and spring-like proteins in our muscles stretch and rebound to absorb shock. Nature has evolved proteins like elastin, silk, and collagen that exhibit remarkable mechanical properties, outperforming many man-made materials. These dynamic movements and mechanical feats are central to life, allowing proteins to act as engines, clamps, switches, and sensors in the cell. It’s no wonder scientists view proteins as “dynamic molecular machines” – and now, we are on the cusp of designing these machines with customized motions and properties.

Physics-Aware AI: A New Paradigm for Intelligent Design

For AI to revolutionize scientific discovery, it must go beyond static representations and incorporate a deeper, first-principles-based understanding of the physical world—its forces, constraints, and governing principles such as conserved quantities. The essence of life and material performance lies not just in structure but in movement. Everything from protein folding to the deformation of materials under stress follows the fundamental laws of physics. Dynamics—how systems evolve over time—governs not only biological function but also engineering, medicine, and materials science. This is why we need physics-aware AI, capable of reasoning about motion, deformation, and energy landscapes. Such AI must go beyond analyzing static forms to understanding how structure and motion are fundamentally intertwined.

Now, we are beginning to bridge this gap. We have AI models that can not only describe how objects move but also tailor their dynamic behaviors with unprecedented accuracy. Nowhere is this more critical than in designing proteins with specific dynamical properties—a challenge that once seemed nearly insurmountable. Only recently have breakthroughs in computational science and AI started to unravel the secrets. A new approach called VibeGen is pushing the frontier even further—using generative AI to create proteins with tailor-made dynamics. Before diving into this latest innovation, we'll take a quick look at how we got here.

Historical Context: From Static Designs to Dynamic Dreams

Designing proteins de novo (from scratch) has long been a “holy grail” of bioengineering. Early pioneers like Nobel Laureate David Baker showed in 2003 that it was possible to build entirely new proteins not found in nature. One of the first successes was a small protein named Top7 – a brand-new fold that proved it was possible to custom-build a protein’s 3D shape.

This breakthrough opened the imagination of scientists: if we can design new protein structures, we might also craft novel enzymes, therapeutics, and nanomaterials. In fact, even back then researchers suggested that designing artificial proteins “opens the way for… enzymes for use as medicines or industrial catalysts.” However, early protein design was like picking a lock in the dark. Scientists had to make educated guesses at amino acid sequences and then experimentally test which ones folded correctly. It could take tens of thousands of tries to find one sequence that worked as intended.

The focus was mostly on getting stable structures – ensuring the protein folded into the desired shape (the classic "structure = function" paradigm). Dynamics – how that protein might bend or flex – were often an afterthought in these early designs. Yet, biologists were learning that a protein’s function often depends on its flexibility and motion. By the late 1990s and 2000s, new experimental techniques allowed researchers to tug on proteins like titin (a giant muscle protein) with atomic force microscopes, revealing they behave like molecular springs that unfold domain by domain under force (Rief, Fernandez, Gaub, et al., Science, 1997). It became clear that mechanics and motion matter: a protein’s conformational dance is just as important as its static form. This is especially true if we want to understand how a protein binds to receptors or how it behaves when assembled into materials like silk, skin, or blood vessels.

The origins of analyses of movements in proteins trace back to early developments in theoretical physics and structural biology. Inspired by the success of normal mode theory in solid-state physics, where it was used to describe collective vibrations in crystals, researchers began applying similar principles to macromolecules in the 1970s. The first application of normal mode analysis to proteins is attributed to Nobel Laureate Martin Karplus and colleagues, who in 1983 demonstrated that low-frequency normal modes could capture large-scale collective motions relevant to biological function. This landmark study showed that normal mode approximations could predict functionally relevant conformational changes, such as hinge motions in enzymes. Since then, normal mode analysis has become a fundamental tool in structural biology, aiding in protein flexibility analysis and molecular docking.

2010 onwards: Convergence of biology and AI

The 2010s and 2020s brought a convergence of biology and AI that dramatically accelerated protein design. Advances like DeepMind’s AlphaFold developed by Nobel Laureates John Jumper and Demis Hassabis solved protein structure prediction, and other generative design tools emerged to create new proteins by computational design. A key development was treating protein sequences like a language that a computer can learn. Just as AI language models learn grammar from sentences, protein language models trained on thousands of natural sequences learned the “grammar” of what makes a foldable, functional protein. By coupling these models with powerful generative algorithms, scientists began to design proteins with specific goals. These methods focused primarily on static properties—a snapshot of a protein, like a single frame in a movie.

Recognizing that such approaches are insufficient to understand the complex dynamical features that govern living materials, algorithms emerged that began to understand physical properties from a more holistic perspective integrating both spatial and temporal features across scales. In 2023, for instance, an MIT-led team used AI diffusion models (inspired by image generators like DALL-E) to design proteins with particular structural features linked to dynamical mechanical traits like stiffness and elasticity. Essentially, they guided an AI to “dream up” proteins that could form materials as resilient as spider silk or as rubbery as elastin. This generative approach called ForceGen was incredibly efficient – whereas older methods were trial-and-error, these AIs could produce millions of novel sequences in days, each meeting the design criteria and tailored mechanical properties that implicitly modeled temporal-spatial multiscale processes. The field was shifting from painstaking manual design to AI-driven creation, moving beyond what evolution has given us, and allowing for the direct design of proteins with targeted properties that are the result of complex multi-scale processes.

Despite these advances, one big challenge remained: dynamics. Most computational designs still dealt with static targets or treated them only implicitly as in ForceGen – a desired shape or binding capability. But what if we want to design a protein’s movement directly? Its flexibility, its unfolding pathway, its internal vibrations, to specify exactly how it moves? These aspects are crucial for truly treating proteins as engineered machines.

Meet VibeGen – Generative AI for Tailored Protein Dynamics

Until now, there was no reliable way to specify and achieve a protein’s dynamic properties. This has now been accomplished with a new AI model called VibeGen, marking a new era where we not only design how a protein looks, but how it behaves dynamically. This model has learned how its sequence relate to the protein's movements. Notably, these relationships are modeled in both directions, such that we can predict a molecule's movement based on sequence and to identify a sequence to meet a certain movement.

VibeGen is an agentic AI-driven platform for end-to-end protein design that addresses the problem of protein dynamics, seamlessly integrating earlier theories about protein motions and protein design. The name “VibeGen” evokes vibrations and generation, reflecting how this system can “tune” a protein’s internal motions like a musical instrument. Developed by a team of researchers, VibeGen builds on the idea of diffusion models and protein language models to tackle a previously impossible feat: designing proteins that meet specific mechanical and dynamic objectives.

How does it work? In simple terms, VibeGen is like a master chef following a recipe to achieve certain dynamical properties. Scientists provide a target “recipe” – for example, “a stretchy protein that moves in this particular way” – and VibeGen cooks up amino acid sequences that should have those properties. Under the hood, it uses a protein language diffusion model, a kind of generative AI that blends two cutting-edge techniques. First, a pretrained protein language model provides deep knowledge of protein “grammar” learned from nature – it knows what sequence patterns tend to fold and function.

Second, a diffusion model component iteratively refines protein designs, much like an image AI refining a blurry picture into clarity. During this refinement that can be pictured as multiple AIs talking to each other to negotiate the best solution, the model is steered toward the desired dynamic outcome (e.g., a certain movement pattern, stiffness or unfolding energy). The AI behaves agentically – it developed sophisticated reasoning abilities that allows it to act as an autonomous designer, exploring countless possibilities and homing in on those that satisfy the mechanical criteria set by the researchers. This agentic end-to-end process means minimal human tweaking; the AI itself develops an awareness about its own predictions, and learns how to achieve the goal by linking sequence features to physical performance.

VibeGen was able to generate de novo proteins - sequences never seen in nature - that exhibit unique, targeted dynamic behaviors. For example, they designed proteins with a specific movement profile – and when these were tested in detailed simulations of the nano-scale physics, they indeed behaved in that way and moved as specified in the original design objective. In technical terms, the proteins fulfilled the targeted vibrational properties.

What’s more, VibeGen operates in silico (on the computer) and end-to-end, meaning it goes straight from a desired property to a protein sequence, which can then be synthesized and tested. This dramatically speeds up innovation. This approach opens rapid pathways to explore the enormous protein sequence space” without being limited to what biology has already made. In other words, VibeGen can venture into uncharted territory, proposing protein designs beyond the repertoire of evolution, tailored purely to our specifications. It’s as if we’ve invented a new creative engine that designs molecular machines on demand.

Together with other methods, like ForceGen, these tools offer powerful new ways to generate mechanically optimized protein structures for dynamical properties. For instance, while VibeGen targets dynamical motions, ForceGen achieves an exact force-extension behavior is like designing a mini shock absorber at the molecular scale – a feat unimaginable a few years ago.

Implications and Real-World Applications

The ability to tailor protein dynamics has far-reaching implications across multiple fields. Here are a few exciting possibilities that VibeGen and similar AI-driven design tools could unlock:

Medicine: Smarter Enzymes and Shape-Shifting Therapeutics

In medicine, proteins are both targets and tools – think of antibodies, enzymes, and vaccine components. VibeGen could enable the design of flexible enzymes and therapeutics that adapt to their job in real time. For instance, an enzyme drug might be designed to be highly dynamic, so it can mold itself around a tricky disease-causing molecule and break it down effectively. Traditional protein design already envisioned creating novel enzymes for therapies, but now we can imagine enzymes that not only perform a new reaction but do so with optimized flexibility and stability, making them more efficient in the body.

Consider drugs for conditions where the cellular environment is variable – a therapeutic protein that can switch conformation in response to pH or cellular signals could maintain effectiveness where a static protein would fail. Another example is allosteric medicines: we could design a protein that changes shape when it binds a small molecule drug, acting as a molecular switch to turn a pathway on or off only when needed. This kind of dynamic control is key for minimizing side effects – the protein drug could remain inactive (harmless) until a trigger molecule activates its change in shape to do the therapeutic work. With AI-guided dynamic design, researchers might also craft better antibodies or vaccine proteins that have the right amount of flexibility to bind tightly to viruses or cancer cells but remain stable enough to circulate in the body. In short, medicine could see a new class of protein therapeutics that are programmable in motion – adapting their structure to fit the task, which could improve efficacy and reduce unintended interactions.

Materials Science: Proteins as Living Materials with Tuned Mechanics

Materials scientists have long looked at proteins like silk and collagen with envy – these natural materials are incredibly strong, resilient, and versatile. Now imagine being able to create new protein-based materials on demand, with precisely tuned mechanical responses. The ability to design for properties like stiffness, elasticity, or energy absorption means we could specify “make a protein gel that is soft under gentle movement but becomes firm under sudden impact” and actually get a molecule that does that. This could lead to smart biomaterials – for example, a shock-absorbing coating for helmets or joints that uses networks of designed proteins to dissipate force (acting like a nano-scale airbag). Conversely, one could design ultra-stiff protein fibers for use in construction or clothing, all biodegradable and produced without petroleum. In fact, researchers suggest such biologically inspired materials could replace plastics or ceramics one day, offering a much smaller carbon footprint while still achieving high performance.

Another exciting application is in tissue engineering and medical implants. Today’s synthetic implants (like heart valve replacements or cartilage scaffolds) often use static materials that don’t adapt to the body. With tailor-made protein materials, we could create implants that match the mechanical behavior of natural tissues – for instance, a designed protein-based scaffold for cartilage that has the same springiness as real cartilage, promoting better integration and healing. Because we can dial in properties, we might also program such materials to respond to changes: a wound dressing material that stiffens if a wound swells (providing support) and relaxes when the swelling goes down, all thanks to embedded protein fibers that react to force or chemical signals. These ideas fall under “dynamic biomaterials”, and they’re increasingly plausible as we gain tools like VibeGen. Indeed, the creators of VibeGen note that exploring the huge design space of proteins with mechanical targets can yield “protein materials with superior mechanical properties.”

Future skyscrapers or vehicles might even incorporate protein-based components that heal themselves after stress or adjust their rigidity on the fly – a form of bio-inspired adaptive material science is eager to develop.

Synthetic Biology: Building New Life Components with Novel Dynamics

Synthetic biology seeks to engineer living systems by reprogramming biology’s building blocks. Proteins are crucial parts of this, serving as sensors, actuators, and scaffolds in synthetic circuits. With the power to design dynamic properties, we could expand the toolkit of synthetic biology to include proteins that act in ways evolution never attempted. Imagine engineering a molecular motor from scratch – a protein machine that converts chemical energy into mechanical motion for a specific task, like moving cargo inside a synthetic cell or powering a tiny pump. By tailoring the protein’s dynamics, one could ensure this artificial motor moves at the desired speed or force. Similarly, we could design regulatory proteins that behave like logical switches or oscillators, flipping between states in a controlled rhythm. For example, a protein could be made to alternate between two shapes every few seconds, effectively working as a timing device for cellular processes – something like a nanoscale clock.

Another area is biosensors. Synthetic biologists often create sensor proteins that detect environmental signals (like toxins, light, or metabolic changes) and then trigger a response. With dynamic design, one could create a sensor protein that not only binds a target molecule but undergoes a big conformational change when it does so. This change could be harnessed to, say, bring together other molecules or activate an enzyme, making the sensor extremely sensitive and switch-like. Because we aren’t limited to what exists in nature, these sensor proteins could detect novel chemicals or conditions. Additionally, designing proteins with novel allostery (where binding at one site affects activity at another site through a change in shape) could allow us to build complex control circuits inside cells, akin to electronic circuits but composed of interacting proteins. The bottom line is that synthetic biology would get a wealth of new parts – not static Lego blocks, but dynamic components that behave in predictable ways by design. This could lead to engineered cells that perform new functions, from bio-remediation systems that adapt to toxins, to cellular micro-factories that modulate their output based on environmental feedback, all powered by designer proteins.

The Future: Proteins with On-Demand Dynamics

It’s hard to overstate how revolutionary these developments could be. By mastering not just protein structures but their motions, we are beginning to learn to engineer life’s movers and shakers at will. VibeGen and its ilk hint at a future where if you can dream up a molecular machine, you can get an AI to build it. Need a protein that acts as a hinge, moving only within a certain angle range? Or a molecular spring that operates at a specific frequency? These could become mere design specifications in a software tool. The design space of possible proteins is astronomically large – vastly more sequences than there are atoms in the universe – yet AI is giving us a map and compass to navigate it. Each new algorithm (from language models to diffusion models, to perhaps future quantum models) serves as an assistant, or indeed an agent, helping us explore options and discover solutions far beyond what human intuition could find.

One exciting prospect is the synergy of AI-designed proteins with experimental evolution. Scientists could use AI to create a batch of proteins with desired dynamics, then put them into a lab setting to see which ones perform best and even let them evolve further. This combination of rational design and evolution could accelerate the invention of molecular machines that are highly optimized and robust. We might also see the rise of “protobots” – protein-based robots at the nanoscale. These would be ensembles of designed proteins that work together like parts of a machine (gears, levers, etc. at molecular size) to achieve a task, whether that’s repairing tissue at a microscopic site or performing computations inside a cell. With precise control over dynamics, each part can be tuned to sync with the others.

In materials science, the future could bring bulk materials made of designed proteins that have capabilities we only see in science fiction. Envision clothing that adapts to temperature by stiffening or relaxing its fiber weave, or buildings that use dynamic protein lattices to dampen vibrations during an earthquake. Because proteins are biodegradable, this could usher in a new era of green materials – high-tech functionality without environmental baggage.

In human health, having AI-customized proteins might enable personalized medicine at an unheard-of level. For instance, if someone has a particular genetic disorder that causes a protein to misfold, we could design a custom chaperone protein that dynamically binds and stabilizes the misfolding protein only when it’s in the wrong shape, thereby preventing disease – a highly specific fix that adapts to the protein’s behavior. Cancer treatments could employ designer proteins that exploit the subtle dynamic differences between healthy and cancerous cell proteins, targeting only the malignant ones.

This vision is bold, but it’s grounded in the real progress we’re already seeing. Biomolecular machines lie at the heart of cellular function, and as our ability to design these machines grows, so do their potential applications. We are witnessing the dawn of a new era where biology becomes an engineering discipline. Proteins, the workhorses of life, can be refashioned with the precision of mechanical parts, yet retaining the beauty and complexity of living systems. From medicines to materials to entire synthetic organisms, the ability to engineer proteins with precise dynamic properties could transform innovation across the board.

The journey is just beginning. Not long ago, designing even a static protein was a monumental task – now, AI-driven tools like VibeGen hint that designing dynamic, functional proteins might one day be as straightforward as coding an app. The implications are profound: we could end up cracking the code of life’s motion, allowing us to create solutions to problems in health, industry, and the environment that were previously unsolvable. It’s an extraordinary time where the once impossible feat of building tailor-made molecular machines is becoming a reality, one amino acid at a time. As these technologies mature, we might soon find ourselves collaborating with AI “colleagues” to invent proteins that propel us into a new age of bio-innovation – an age where the only limit to what proteins can do is our imagination.

References and Notes

• Kuhlman, B., Dantas, G., Ireton, G. C., Varani, G., Stoddard, B. L., & Baker, D. (2003). Design of a novel globular protein fold with atomic-level accuracy. Science, 302(5649), 1364-1368.DOI: 10.1126/science.1089427
• Brooks, B. & Karplus, M. (1983). Normal modes for specific motions of macromolecules: Application to the hinge-bending mode of lysozyme. Proc. Natl. Acad. Sci. USA, 80(21), 6571-6575
• Isralewitz, B., Gao, M., & Schulten, K. (2001). Steered molecular dynamics and mechanical functions of proteins. Current Opinion in Structural Biology, 11(2), 224-230. DOI: 10.1016/S0959-440X(00)00194-9
• Rief, M., Gautel, M., Oesterhelt, F., Fernandez, J. M., & Gaub, H. E. (1997). Reversible unfolding of individual titin immunoglobulin domains by AFM. Science, 276(5315), 1109-1112.
• B. Ni, D. Kaplan, M.J. Buehler, ForceGen (2023): End-to-end de novo protein generation based on nonlinear mechanical unfolding responses using a language diffusion model: https://www.science.org/doi/10.1126/sciadv.adl4000
• Nature Machine Intelligence. (2024). AI protein shake-up. Nature Machine Intelligence, 6, 121.
• B. Ni and M.J. Buehler, VibeGen, an Agentic End-to-End De Novo Protein Design for Tailored Dynamics Using a Language Diffusion Model (2025): https://arxiv.org/abs/2502.10173
• VibeGen Code: https://github.com/lamm-mit/ModeShapeDiffusionDesign
• VibeGen Model weights: https://huggingface.co/lamm-mit/VibeGen