Amazon SageMaker AI now supports BoltzGen, a diffusion-based generative model for designing proteins and peptides that can bind to specific biomolecular targets. The integration streamlines the complex process of protein design by managing GPU compute infrastructure end to end, allowing researchers to focus on design iteration rather than infrastructure operations. The platform handles instance provisioning, data movement, and resource cleanup, making it easier to scale 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 capability is particularly useful for academic research labs, biotech startups, and pharmaceutical R&D groups working in protein binder design or therapeutic protein engineering. 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. The same configuration that validates a design specification with 10 candidates scales to larger campaigns — only the parameter values change. Instance types ranging from ml.g4dn (lowest cost T4 GPUs) to ml.g6e (NVIDIA L40S GPUs) give you flexibility for matching throughput to budget.
The repository includes setup scripts, both execution modes, multi-GPU and multi-instance scaling support, and a quick start guide. The implementation is available in the Boltzgen on SageMaker repository on GitHub. The repository includes setup scripts, both execution modes, multi-GPU and multi-instance scaling support, and a quick start guide. Two execution modes are provided, optimized for different stages of research, so you can start with quick experiments and graduate to production workflows as needs grow. SageMaker AI processing jobs provide direct batch execution for quick experiments. After submitting a job, SageMaker AI provisions the GPU instance, runs the container, and shuts down the instance when complete. A single Python script triggers the entire workflow, without multi-step orchestration. For early-stage work where the goal is to test design specifications before scaling up, processing jobs minimize setup time. SageMaker AI Pipelines provide a 5-step orchestrated workflow with step-level caching, automatic scaling, and visual monitoring in SageMaker AI Studio.
Production workloads benefit from parameter iteration without re-running expensive compute steps. Each step (design, inverse folding, folding, analysis, filtering) can be scaled independently, and cached results are reused when inputs have not changed. Both modes follow the same workflow. The container from Amazon ECR runs on a provisioned GPU instance, BoltzGen processes the design specification, and SageMaker AI stores results in Amazon S3 before shutting down compute resources. Getting started requires an AWS account, a few tools installed locally, and an Amazon S3 bucket for your data. The following sections walk you through each prerequisite and then guide you through your first protein design experiment. Prerequisites include an active AWS account with billing enabled, the AWS Command Line Interface (AWS CLI) installed and configured, sufficient service quotas for GPU instances, an IAM execution role for SageMaker AI, and local installation of Python 3.11+, Boto3, SageMaker SDK, and Docker. Finally, creating an Amazon S3 bucket in your target Region is necessary. Running your first protein design experiment is streamlined with Amazon SageMaker Studio or a notebook instance, which come preconfigured with AWS credentials, the SageMaker SDK, and Docker support.
Launching SageMaker AI Studio involves opening the SageMaker AI console, creating a new domain, enabling Docker Access through local mode using AWS CLI, and launching a JupyterLab space. Quota limits for ml.g4dn.xlarge and ml.g5.xlarge processing job usage must be raised to greater than 0. The repository includes setup scripts, both execution modes, and example design specifications. The container image packages BoltzGen with its dependencies and GPU drivers, and can be built once and reused across experiments. Setting your region and running the build script, along with configuring AWS credentials, is required for SageMaker AI to provision compute resources and store results in Amazon S3. The example .env file includes AWS_REGION and AWS_S3_BUCKET settings.
Source: awsml
