More than 300 people from academia and industry gathered in the auditorium to participate in the conference BoltzGen Seminar on Thursday, October 30, hosted by Abdul Latif Jameel Machine Learning in Health Clinic (MIT Jameel Clinic). The event's headliner was MIT graduate student and first author of BoltzGen, Hannes Stärk, who had announced BoltzGen just a few days earlier.
Resisting Boltz-2an open-source biomolecular structure prediction model predicting protein binding affinity that gained popularity over the summer BoltzGen (officially unveiled on Sunday, October 26) is the first model of its kind to go a step further by generating new binding proteins that are ready to bring to market the drug discovery pipeline.
Three key innovations make this possible: first, BoltzGen's ability to perform a variety of tasks, standardizing protein design and structure prediction while maintaining state-of-the-art performance. BoltzGen's built-in constraints are then designed based on feedback from wet lab collaborators to ensure that the model will create functional proteins that do not defy the laws of physics and chemistry. Finally, a rigorous evaluation process tests the model on “non-treatable” target diseases, pushing the limits of what is possible to produce the BoltzGen binder.
Most models used in industry or academia can predict the structure or design of proteins. Moreover, they are limited to producing specific types of proteins that bind effectively to easy “targets.” Like students answering a test question that looks like their homework, models often work as long as the training data looks similar to the target values when designing the binder. However, existing methods are almost always evaluated on targets for which structures with binders already exist, and as a result their performance is lower when applied to more demanding targets.
“There have been models that have tried to address binder design, but the problem is that these models are modality dependent,” notes Stärk. “The generic model doesn't just mean we can tackle more tasks. Additionally, we get a better model for a single task because we learn to emulate physics from examples, and with a more general training scheme we provide more such examples containing generalizable physics patterns.”
BoltzGen researchers went to great lengths to test BoltzGen on 26 targets, ranging from therapeutically relevant cases to those explicitly selected for their dissimilarity from the training data.
This comprehensive validation process, which took place in eight wet labs in academia and industry, demonstrates the model's breadth and potential for breakthrough drug development.
Parabilis Medicines, one of the industry collaborators that tested BoltzGen in the wet lab, praised BoltzGen's potential: “We believe that adopting BoltzGen onto our existing Helicon peptide computing platform capabilities promises to accelerate our progress in delivering transformative medicines against major human diseases.”
Although the open source versions of Boltz-1, Boltz-2 and now BoltzGen (which were presented at the conference 7th Molecular Machine Learning Conference October 22) bring new opportunities and transparency to drug development and signal that the biotechnology and pharmaceutical industries may need to re-evaluate their offerings.
Amid the buzz surrounding BoltzGen on social media platform X, Justin Grace, principal machine learning scientist at LabGenius, asked a question. “The performance latency of AI chat systems from private to open is (seven) months and falling,” Grace wrote in post. “It looks like it's even shorter in the protein space. How will Binder-as-a-Service companies be able to (recoup) their investment when we can just wait a few months for the free version?”
For academics, BoltzGen means expanding and accelerating research opportunities. “The question my students often ask me is, 'Where can AI change the therapeutic game?'” says senior co-author and MIT professor Regina Barzilay, head of the department of artificial intelligence at the Jameel Clinic and a fellow at the Computer Science and Artificial Intelligence Laboratory (CSAIL). “Until we identify insurmountable goals and propose a solution, we will not change the game,” he adds. “The emphasis is on unresolved problems, which distinguishes Hannes' work from other work in the field.”
Senior co-author Tommi Jaakkola, the Thomas Siebel Professor of Electrical Engineering and Computer Science affiliated with the Jameel Clinic and CSAIL, notes that “models like BoltzGen, which are made completely open source, enable broader community-wide efforts to accelerate drug design capabilities.”
Looking ahead, Stärk believes that the future of biomolecular design will be turned upside down by artificial intelligence models. “I want to build tools that will help us manipulate biology to solve diseases or perform tasks with molecular machines that we haven't even dreamed of yet,” he says. “I want to provide these tools and enable biologists to imagine things they haven't even thought about before.”
