Amazon SageMaker AI accelerates BoltzGen protein design with 90% compute savings

BoltzGen is an all-atom diffusion-based generative model for designing proteins and peptides that bind to specific biomolecular targets. A typical design campaign involves multiple GPU-intensive steps: backbone generation, inverse folding (using the BoltzIF model), structural validation with Boltz2 folding predictions, and candidate ranking. Running these steps across hundreds to millions of candidates traditionally required managing GPU instance provisioning, data pipelines, failure recovery, and cost tracking. Amazon SageMaker AI directly addresses these bottlenecks by provisioning GPU instances, executing BoltzGen inside a container, writing results to Amazon S3, and releasing instances when done—billing per second with no idle costs. A 1,000-sample campaign on a 4-GPU ml.g5.12xlarge instance takes about 375 hours, while a 2-hour test run on ml.g4dn.xlarge costs roughly $1.50.

SageMaker AI offers two execution modes: processing jobs for quick experiments (single Python script, direct batch execution) and pipeline mode with step-level caching. In pipeline mode, each step’s output is cached in S3 with a 7-day expiry. When iterating on filtering parameters, the design generation step (which accounts for ~90% of compute cost) does not re-run, drastically reducing expenses. The implementation supports multi-GPU parallelization within a single instance and multi-instance scaling across a pipeline. Setup scripts in the GitHub repository build the container, push it to Amazon ECR, and include working examples to submit a first design job within minutes. Instance types range from low-cost ml.g4dn (T4 GPUs) to high-throughput ml.g6e (NVIDIA L40S), giving flexibility across budget and performance needs.