OpenFold3 arrives on SageMaker for protein, DNA and drug-molecule predictions

AWS has added the OpenFold3-preview biomolecular structure-prediction model to SageMaker JumpStart, letting research teams run predictions through a managed endpoint in their own AWS account. The useful shift is practical: complex folding jobs can be queued and run asynchronously, rather than forcing researchers to build the serving infrastructure themselves.

AWS AI Watch analysis

What happened

OpenFold3 predicts 3D structures for proteins, DNA, RNA and small-molecule ligands, including complexes made from multiple kinds of molecule. AWS says the model is available under the Apache 2.0 licence and runs on SageMaker Asynchronous Inference. Requests go through Amazon S3; the endpoint can take up to an hour to process one, then writes the result or error details back to S3.

The deployment supports batches of jobs, multi-GPU instances and user-supplied multiple sequence alignments. AWS also describes scaling the endpoint to zero when its queue is empty. Its walkthrough uses version 0.4.1 and lists three supported instance sizes, from ml.g6.2xlarge to ml.g6.12xlarge. The service is a research preview, and AWS warns that running instances are billed by the hour, including while idle.

Why it matters

Biomolecular prediction can demand GPUs, large model files and a serving setup built for jobs that take minutes, not milliseconds. A managed endpoint and built-in job queue lower some of that operational friction. Scaling to zero could also help avoid paying for idle endpoint capacity between experiments, though the compute used while it is running is still billable.

For teams working with proprietary sequences, AWS says the requests and results are stored in S3. Its guidance points to restricting bucket access, using encryption and setting lifecycle rules to expire data that is no longer needed. That is useful operational detail, not a claim that sensitive data handling happens automatically.

Our read

This is a meaningful way to make an open model easier to try and operate, especially for research teams that would rather spend time on biology than assembling a CUDA serving stack. But a smoother deployment is not a scientific validation: OpenFold3 is labelled a preview, and the announcement does not establish how well its predictions perform across real research workloads. Start with a controlled test, and keep an eye on instance costs and sequence storage.

What to watch

• Whether AWS moves OpenFold3 beyond preview and expands its supported instance options.
• How the model performs on researchers’ own prediction tasks and supplied alignments.
• Whether AWS publishes clearer cost guidance for typical jobs and endpoint configurations.

Discussion spark: Would managed deployment make you more willing to test an open biomolecular model, or would preview status and cloud costs keep the work on your own infrastructure?

not affiliated with or endorsed by Amazon Web Services (AWS)