The term “untreatable” has always said more about the limits of our tools than the limits of biology itself. That assumption is now being directly challenged by BoltzGen, a generative AI model from MIT that doesn’t just analyze disease targets, it actively designs brand-new molecules to reach them. If it lives up to its promise, entire categories of conditions may soon lose their status as therapeutically out of bounds.
Developed by a research team at MIT, BoltzGen aims to break through this long-standing barrier and rethink how new medicines are conceived, built, and evaluated.
The model stepped into the spotlight during a BoltzGen seminar at the Abdul Latif Jameel Clinic for Machine Learning in Health, where more than 300 researchers from academia and industry filled an auditorium to hear its debut. Leading the presentation was MIT PhD student and first author Hannes Stärk, who had only days earlier introduced the scientific community to the system.
BoltzGen officially launched on October 26, building directly on the team’s prior open-source work, Boltz-2: a model designed to predict how proteins bind at the structural level. But where Boltz-2 stopped at analysis, BoltzGen makes the leap into creation. Instead of merely evaluating molecular interactions, it generates entirely new protein binders that can move directly into the drug development pipeline. In doing so, the model closes a long-standing gap between prediction and invention that has slowed therapeutic discovery for years.
Three design innovations make this shift possible.
First, BoltzGen operates as a unified system, combining both protein structure prediction and protein design within a single general model. This integration enables it to achieve state-of-the-art performance while addressing multiple tasks at once.
Second, strict physical and chemical constraints were built into the model through close collaboration with laboratory scientists. These safeguards ensure that generated proteins remain biologically feasible, grounded in real-world molecular behavior rather than conceptual constructs that defy chemistry or physics.
Third, the team reengineered the validation process itself. Instead of relying on familiar benchmarks, they deliberately tested the model against disease targets historically labeled “undruggable,” pushing its capabilities into uncharted territory.
This strategy directly tackles a core weakness of today’s protein-modeling landscape. Most AI tools specialize in either structure prediction or binder design, but rarely both. Even when they perform well, their success tends to depend on how closely a target resembles patterns already present in training data. As soon as the model encounters unfamiliar biological challenges, performance often collapses, much like a student excelling only when exam questions mirror their homework.
Stärk believes such modality-restricted approaches inherently cap progress. “A general model does not only mean we can address more tasks,” he explains. “It also produces better results on individual challenges because physical principles are learned through broad exposure, patterns that translate across problems rather than locking into narrow examples.”
To demonstrate BoltzGen’s ability to operate far outside predictable scenarios, the team tested it on 26 disease targets, spanning both therapeutically familiar cases and others chosen specifically for their distance from the training data. Validation took place across eight wet labs throughout academia and industry, an unusually broad real-world assessment for a computational drug design model. The results highlighted both the system’s flexibility and its potential to advance therapies where previous approaches have fallen short.
One of the participating industry collaborators, Parabilis Medicines, integrated BoltzGen into its Helicon peptide computational platform and described the move as one that promises to accelerate the development of “transformational drugs against major human diseases.”
Beyond the laboratory, BoltzGen’s open-source release raises questions for the business of biotech itself. With the availability of Boltz-1, Boltz-2, and now BoltzGen, the latest previewed publicly at the 7th Molecular Machine Learning Conference on October 22, the pace at which private tools are overtaken by open alternatives continues to shrink. On X, Justin Grace, a principal machine learning scientist at LabGenius, noted that the performance gap between proprietary and open chat AI models has already fallen to around seven months and continues to close. In protein modeling, he suggested, the cycle appears even shorter, posing challenges for companies offering commercial “binder-as-a-service” models seeking to recoup their investments.
For academic researchers, however, the story reads very differently. BoltzGen represents not disruption, but expansion, an invitation to ask tougher questions and pursue more ambitious goals. Regina Barzilay, MIT professor, senior co-author on the study, and AI faculty lead at the Jameel Clinic, sees this focus as essential. Students often ask her where AI can truly change the therapeutic landscape. Her response remains consistent: impact demands confronting targets that have previously resisted solutions. Without tackling such unsolved problems, she argues, the field cannot fundamentally transform and that focus distinguishes the BoltzGen effort.
Tommi Jaakkola, Thomas Siebel Professor of Electrical Engineering and Computer Science at MIT and another senior co-author, emphasizes the role of openness itself. Fully open-source releases like BoltzGen, he notes, empower the broader scientific community, aligning global efforts behind the shared acceleration of drug discovery.
Looking ahead, Stärk envisions AI not as an optimization layer but as a foundational engine for biomolecular engineering. “I want to build tools that help us manipulate biology to solve disease,” he says, “or use molecular machines to perform tasks we haven’t yet imagined. The goal is to give biologists instruments that unlock entirely new ways of thinking.”
