Boltzgen for de novo protein binder design
A flow-based generative model that samples diverse protein binder scaffolds and binding geometries, expanding the design space beyond diffusion-only approaches
Flow-based generative modeling for proteins
Boltzgen is a generative model that uses normalizing flows to sample from a learned Boltzmann distribution over protein conformations. Unlike diffusion models that reverse a noise process, flow models learn an invertible transformation between a simple prior distribution and the target distribution of valid protein structures.
This architectural difference means Boltzgen explores different regions of protein structure space than RFdiffusion. It can generate backbone topologies and binding geometries that may not appear in diffusion-based outputs, adding structural diversity to the overall candidate pool.
Different models, different designs
Denoising diffusion. High-throughput backbone generation conditioned on target hotspots. Broad topological exploration.
Iterative co-optimization. Scaffold and sequence refined together in a single loop. Fewer but higher-quality candidates per run.
Normalizing flow. Learns a continuous transformation from noise to structure. Different inductive biases produce alternative binding geometries.
When Boltzgen adds the most value
When RFdiffusion scaffolds underperform, Boltzgen explores alternative binding geometries that diffusion models may not sample.
For campaigns where structural diversity in the candidate pool is a priority, Boltzgen adds topologies outside the RFdiffusion distribution.
Boltzgen samples from the Boltzmann distribution, making it useful for understanding conformational ensembles of designed proteins.
Combining Boltzgen and RFdiffusion outputs increases the probability that at least one design approach produces validated binders.
Same validation, different generation
Boltzgen candidates enter the same downstream pipeline as RFdiffusion and BindCraft outputs. ProteinMPNN designs sequences for Boltzgen-generated backbones, and Boltz-2, ESMFold, and ColabFold validate the predicted complex structure.
No separate filtering criteria. The same ipTM, pLDDT, PAE, and solubility thresholds apply to all candidates regardless of which generative model produced them.
Maximize candidate diversity with multiple generative approaches
We run Boltzgen alongside RFdiffusion and BindCraft to give your campaign the widest possible design space. Tell us about your target.