BoltzGen on Amazon SageMaker AI accelerates protein binder design by managing GPU compute infrastructure end to end. BoltzGen is a diffusion-based generative model that designs proteins and peptides capable of binding to specific biomolecular targets. A typical design campaign involves multiple GPU-intensive steps: backbone generation, inverse folding, structural validation, and candidate ranking. Running these steps across hundreds, thousands, or even millions of design candidates introduces operational overhead in provisioning instances, moving data between steps, and tracking costs. SageMaker AI manages this compute lifecycle from instance provisioning through result delivery and resource cleanup, so you can focus on design iteration rather than infrastructure operations.

In this post, we demonstrate how to deploy BoltzGen on SageMaker AI and run an end-to-end protein design experiment. By the end of the walkthrough, you have a working setup that scales from quick validation runs to production batch processing. The setup offers two execution modes for different stages of research and uses step-level caching to reduce compute expenses during iterative workflows.

This walkthrough applies to academic research labs, biotech startups, pharmaceutical R&D groups, and educational programs, whether you work in protein binder design, therapeutic protein engineering, or de novo protein architecture.

Each step in a BoltzGen campaign runs on GPU hardware and processes one design specification at a time. On a 4-GPU instance (ml.g5.12xlarge), a campaign of 1,000 samples takes approximately 375 hours to complete, based on the repository’s benchmark data. Operating this infrastructure involves building CUDA environments (e.g. install CUDA driver and setup toolkit), coordinating GPU instance lifecycles, constructing data pipelines between steps, and recovering from failures in long-running jobs.

SageMaker AI addresses each of these bottlenecks directly. After you submit a job, SageMaker AI provisions GPU instances and executes BoltzGen inside the container. It writes results to Amazon Simple Storage Service (Amazon S3) and releases the instances when processing completes. Billing is per-second, so there are no idle GPU costs. A 2-hour design run on ml.g4dn.xlarge costs approximately $1.50 based on on-demand pricing.

The implementation supports multi-GPU parallelization within a single instance and multi-instance scaling across a pipeline. In pipeline mode, each step’s output is cached in Amazon S3 with a 7-day expiry, so when you iterate on filtering parameters, the design generation step that accounts for approximately 90 percent of compute cost does not re-run.