BREAKING: BoltzGen Now Live on Rowan
a new foray into generative protein-binder design; what makes BoltzGen different; experimental validation; democratizing tools; running BoltzGen on Rowan
Yesterday, Hannes Stärk and co-workers from MIT (and other institutions) released BoltzGen (paper, repository), an open-source all-atom generative model for designing protein and peptide binders to a diverse array of targets. BoltzGen launched with extensive experimental validation from laboratories in academia and industry—while this tool is less than 24 hours old, it’s clear that this is going to be an important and impactful addition to the computational life-sciences toolbox.
We’re excited to be launching support for BoltzGen on Rowan today. In this post, we’ll (1) briefly review how the field got here, (2) explain what makes BoltzGen special, and (3) discuss how to use BoltzGen through Rowan.
Protein Binder Design
Protein binder design is something that’s new for Rowan—but it’s not a new field. Scientists like David Baker (last year’s Nobel Laureate in Chemistry) have been working on protein design for decades now, with myriad successes along the way. While this process used to be a highly manual and iterative process, the rise of deep learning has led to massive optimizations and accelerations in protein design.
For newcomers, it’s hard to understand just how fast this field has progressed. Just three years ago, a Nature paper from Baker and co-workers reported that miniproteins that bound to target proteins could be designed solely based on the structure of the target protein. While this result was groundbreaking (hence the Nature publication), the hit rates were abysmal by today’s standards: screening 100,000 binders against a given target returned hundreds of binders in the best case and fewer than ten in the worst case, for an optimistic hit rate of ~0.5%.
Followup work from Baker and co-workers in 2023 showed that deep learning methods like AlphaFold2 or RoseTTAFold could be used to triage designs in silico, leading to c. 10x improvements in success rates. And another 2023 paper from Baker and co-workers showed that backbone generation via equivariant diffusion conditioned on the target (RFdiffusion) also lead to significant enhancements in generality and hit rate.
The current state-of-the-art method for protein binder design is fast-moving, but the BoltzGen paper specifically calls out three methods as key previous techniques:
RFDiffusion. The aforementioned 2023 model from Baker and co-workers, which uses diffusion to design protein backbones.
BindCraft. This open-source pipeline for protein design was published about a year ago by Martin Pacesa and co-workers. It uses AlphaFold2 and ProteinMPNN and, to our knowledge, is state-of-the-art for miniprotein binders (10–100% reported success rate).
Germinal. This antibody-design pipeline combines ColabDesign, AbMPNN, and AlphaFold3 to generate antibodies against targeted epitopes (with a 4–22% reported success rate). The use of AlphaFold3 makes this challenging for non-academic users.
This seems to represent general sentiment in the field; in recent protein-design competitions like the 2024 Adaptyv EGFR binder competition, BindCraft and RFDiffusion were some of the most used and most successful models.
BoltzGen
How does BoltzGen fit into this landscape? The key advantages of BoltzGen, according to the authors, are:
Generality. BoltzGen isn’t narrowly tailored to a specific class of biomolecules, but instead uses a single model and pipeline for nanobody design, protein-binder design, peptide design, and protein–small molecule design. The authors anticipate that this flexibility will push models to learn the correct underlying physics: “as models learn to emulate physics primarily through examples provided, we believe expanding the generality of the method further improves its design capabilities for specific classes as well.”
One all-atom model. I’ll quote from the paper directly (emphases added):
At its core, the BoltzGen pipeline uses a single, all-atom generative model that unifies design and structure prediction. A purely geometry-based representation of designed residue types enables scalable training on both tasks simultaneously. As a result, unlike any previous design model, BoltzGen matches the performance of state-of-the-art folding models.
This isn’t trivial to accomplish. The BoltzGen model uses a 14-atom, geometry-based amino-acid representation for designed residues. This makes it possible to represent all structures at atomic resolution rather than needing to combine discrete residue tokens with structure coordinates and allows the model to operate in a smooth, continuous space (enabling the joint training of structure prediction and design).
Tunability. BoltzGen comes with a diverse design specification language which allows virtually any feature of the structure-generation process to be tuned: residue types, bonds, secondary structure, binding site, and so on and so forth. The authors see this tunability as necessary and vital (emphasis added):
Lastly, we comment on how there is a tendency in the field to claim that binder design models are “zero-shot” and “plug-and-play” solutions without a chance for failure. We do not make this claim and encourage users to use BoltzGen thoughtfully, carefully inspect the generated structures, and potentially rerun the pipeline multiple times, first at smaller and then larger scales. BoltzGen’s rich design specification language provides a large degree of control that should be experimented with for optimal results.
It’s worth talking about how BoltzGen works in slightly greater detail. At a high level, the BoltzGen workflow has seven steps:
Design. Design initial binder structures according to the design specification using a diffusion model (using the core BoltzGen model).
Inverse folding. Predict sequences of amino acids that will fold into those structures (”inverse folding” using the BoltzIF model).
Design folding. Refold the newly-predicted amino-acid sequences with their targets using Boltz-2 to validate that they’re actually predicted to bind the target.
Folding. Refold standalone structures of newly-predicted amino-acid sequences (skipped for peptides or nanobodies) to validate that the designed proteins will be stable on their own.
Affinity. Predict protein–ligand binding affinities (if designing a small-molecule binder).
Analysis. Analyze to predict design quality.
Filtering. Filter and rank designs to select best candidate binders.
(In an accompanying piece, we’ve outlined the key features of BoltzGen in much more detail.)
The ultimate test for any protein-binder-design workflow, however, is how well it works in real life. The BoltzGen team included lots of experiments in conjunction with 26 external labs showing that BoltzGen-designed binders can be expressed, are soluble, and in many cases actually bind the target that they’re supposed to. These validation experiments were designed to be quite challenging!
The team used BoltzGen to design nanobodies and miniprotein binders against 9 novel protein targets with high dissimilarity to the PDB, ran real-world experiments with Adaptyv Bio, and in both cases found 6 out of 9 binders had nM binding affinity. BoltzGen-designed peptides were able to bind disordered regions of proteins and existing targets without known peptide binders (like RagC GTPase), while BoltzGen could also design binders for arbitrary small molecules like rucaparib and a rhodamine derivative. (See the BoltzGen paper for additional results and full details.)
While many more validation experiments will doubtless be run in the future, these results are powerful evidence that BoltzGen can actually work against real and challenging targets.
Running BoltzGen Through Rowan
In an interview with Benchling, BoltzGen’s first author Hannes Stärk envisioned a world where tools like BoltzGen can be “immediately at the fingertips of biologists” who “don’t need to be ML experts to start designing and testing” their binders. In Stärk’s words, “that accessibility is what will make AI a real part of everyday science.”
We couldn’t agree more—as longtime readers know, one of our central goals at Rowan is to make it possible for scientists to use the best ML models to accelerate their research. As of today, all Rowan users can run BoltzGen calculations through our new protein-binder-design workflow.
Users can input the desired sequences, SMILES strings, or structure files into Rowan. (It’s easy to load existing crystallographic files and remove existing binders using Rowan’s protein editor.) The below workflow will design cyclic peptide binders against KRas G12D using the peptide-anything protocol starting from an existing PDB structure (8JJS) with the binder removed.
Running Boltz produces a list of output binders and quality scores, which can be viewed and downloaded through Rowan’s web interface. Additional analytics will roll out over the next few days as we add support for more of BoltzGen’s complex configurable options.
It’s worth noting that running a proper binder-design campaign with BoltzGen can get pretty expensive—the authors recommend designing around 50,000 binders for each target. Since each input requires a BoltzGen inference call, a BoltzIF inference call, and 1–2 Boltz-2 inference calls, the GPU requirements quickly become substantial (several GPU-days of computer time, which translates into thousands of real-world dollars).
To make sure that users’ first BoltzGen calculations finish successfully, we’re capping the number of designs that Rowan users can run to 100 by default (10 for free-tier users). This aligns with advice from Stärk (emphasis added):
Treat BoltzGen as an iterative design partner. Start small, inspect your results, and adjust parameters. Don’t just go with the pre-sets. Explore different binding sites or constraints, rerun your designs, and compare. We’ve intentionally exposed many control options — binding-site flexibility, sequence length, exclusion zones — so users can see how changes affect outcomes.
This isn’t intended to limit your science! If you want to run larger numbers of designs or are interested in running BoltzGen-powered protein-design campaigns through Rowan, please reach out. We’d be delighted to work with you and find a cost-effective way to run BoltzGen against your targets.