AWS adds OpenFold3 preview to SageMaker JumpStart

AWS has added OpenFold3-preview to Amazon SageMaker JumpStart, giving computational biology teams a managed way to deploy the open biomolecular structure model in their own AWS accounts. The AWS-maintained container runs on a managed SageMaker AI endpoint and supports joint predictions for proteins, DNA, RNA, small-molecule ligands and complexes.

The package is still a research preview, not the final OpenFold3 release. The OpenFold3 project says the preview is intended to gather community feedback and let developers start building on the ecosystem. AWS lists OpenFold3 version 0.4.1 with the of3-p2-155k checkpoint. The model returns predicted structures in mmCIF format alongside confidence metrics.

The deployment uses SageMaker Asynchronous Inference because structure predictions can take minutes. A client stores a request in Amazon S3 and calls InvokeEndpointAsync. SageMaker queues the job, returns an identifier and output location, then writes the result or error details to S3. Clients can poll for completion or receive an Amazon SNS notification, and an invocation can run for up to one hour.

For this OpenFold3 package, multi-GPU instances add concurrency; they do not make a single prediction run faster. That differs from GPU partitioning for SageMaker real-time inference. One OpenFold3 prediction uses one GPU, so an ml.g6.12xlarge instance with four GPUs can run as many as four predictions at once when configured for four concurrent invocations. AWS said its tests found larger throughput gains as protein size increased, while the cost per prediction for larger proteins on ml.g6.12xlarge was close to ml.g6.4xlarge. The results were not independently benchmarked, and AWS said outcomes vary by input and Region.

AWS also says the endpoint uses network isolation, does not call external multiple-sequence-alignment servers and accepts precomputed alignments. According to the company, that setup can help keep proprietary sequences within a customer’s AWS environment. This is an AWS privacy claim, not an independent security audit. Requests and predicted structures are stored in S3. AWS recommends access controls, customer-managed KMS encryption and lifecycle rules, and advises users to consult security and legal teams before applying the sample code to proprietary or regulated data.

SageMaker asynchronous endpoints can autoscale to zero when no requests are pending, then queue new work while capacity starts. Endpoint instance charges stop after scale-in, though S3 and other charges continue. The first request from zero capacity must also wait for the scaling alarm and instance startup. AWS warned that G6 capacity and service quotas vary by Region, with ml.g6.12xlarge capacity potentially limited.