DEV Community

News
News

Posted on

AWS details OpenFold3-preview protein structure prediction on SageMaker JumpStart

Infographic cover

๐Ÿ“… November 13, 2026 โ€ข AWS machine learning blog

A structure model you can deploy from a catalog. An AWS machine learning blog post dated November 13, 2026 details deploying OpenFold3-preview from Amazon SageMaker JumpStart, per the post. OpenFold3 is an open-source biomolecular structure prediction model developed by the OpenFold Consortium together with the AlQuraishi Lab at Columbia University, per AWS. The model predicts 3D structures of proteins, DNA, RNA, small-molecule ligands and the complexes they form, working at the all-atom level, per the same post. The open source code lives in the aqlaboratory/openfold-3 repository on GitHub, so the code behind the preview is publicly inspectable.

Deployment stays inside your account. The model deploys into a customer AWS account with a few lines of code via SageMaker JumpStart and runs in an AWS-maintained container inside that account, per the post. Asynchronous inference endpoints handle long-running predictions, and results are written to Amazon S3, per the same source. On compute, multi-GPU instances such as ml.g6.12xlarge run four concurrent predictions, per the post. Endpoints scale to zero instances when idle and scale back up automatically on new requests, per AWS, so capacity follows the experiment schedule rather than a fixed footprint, which matters for workloads that arrive in bursts.

Sequence data stays put. Network isolation keeps sequences inside the account, so the endpoint cannot call external servers, and customers may supply precomputed multiple sequence alignments, or MSAs, per the blog post. Results are delivered in mmCIF format with confidence metrics pLDDT, PAE, pTM and ipTM, per the same post. AWS classifies OpenFold3-preview as a research preview and cautions that its demonstrated prediction of a well-studied complex is a workflow example, not an accuracy benchmark for novel targets, per the post, so teams should treat demonstration results accordingly.

What this means in practice. A research team gets a short path from idea to structure: a few lines of code stand up an endpoint inside its own account, a batch of sequences goes in, and mmCIF files with confidence metrics land in S3, with multi-GPU instances such as ml.g6.12xlarge running four concurrent predictions and idle endpoints scaling to zero between runs. Because the endpoint is network isolated, sequence data does not leave the account, and teams holding precomputed MSAs can plug them in instead of recomputing them. The research preview label means results on novel targets should be validated before use, per the caution in the post.

Sources: aws.amazon.com - github.com

Key figures

IMPACT checklist


Continue the conversation: https://t.me/AIAutomationArchitecture ยท Bluesky

AI Automation

Top comments (0)