Skip to content
AI.info

The Pulse

AWS details OpenFold3 research preview deployment on SageMaker

AWS has published a deployment guide for OpenFold3-preview on SageMaker JumpStart. The model predicts biomolecular structures using an asynchronous endpoint that can scale to zero when idle.

AWS details OpenFold3 research preview deployment on SageMaker

AWS customers can deploy OpenFold3-preview through SageMaker JumpStart to predict the structures of proteins, DNA, RNA and other molecules in their own accounts. A November 13 guide from Amazon Web Services describes how to run the research-preview model on a managed endpoint, submit prediction jobs and retrieve their results.

AWS had already listed OpenFold3 as available on JumpStart on September 14. The new guide details the deployment configuration, including asynchronous processing, multi-GPU predictions and an option to scale the endpoint to zero when it is idle.

What the model predicts

The OpenFold Consortium and the AlQuraishi Lab at Columbia University developed OpenFold3 as an open-source, all-atom biomolecular structure prediction model. It predicts the joint three-dimensional structure of molecules from their sequences, including multi-chain protein complexes, protein–nucleic acid assemblies and protein–ligand complexes.

The JumpStart version is OpenFold3 0.4.1 with the of3-p2-155k checkpoint. Its code and weights are released under the Apache 2.0 license, and its developers publish the training data and methods. The developers characterize OpenFold3-preview as a research preview.

AWS lists possible uses including predicting target and complex structures, evaluating designed antibody or protein variants, and generating binding-pose hypotheses by folding a protein with a small molecule. Small molecules can be supplied as SMILES strings; protein, DNA and RNA chains are supplied as sequences.

Jobs run asynchronously

The deployment uses SageMaker Asynchronous Inference rather than a real-time endpoint. AWS says structure predictions can take several minutes and that folding time increases with the number and length of chains in a query.

Users upload a request to Amazon S3 and invoke the endpoint asynchronously. The endpoint queues the job, processes it with an invocation timeout of up to one hour, and writes either the result or error details to a chosen S3 location. Users can check for the output or receive a notification through Amazon Simple Notification Service.

The endpoint can queue batches and can be configured to scale to zero instances when the queue is empty. AWS says SageMaker endpoints are billed per instance-hour while running, including when idle.

Deployment and output

OpenFold3 runs in an AWS-maintained container on a managed SageMaker AI endpoint. AWS says the endpoint uses network isolation, which can help keep sequences within the customer’s AWS account. Requests and results are stored in S3.

The supported instance types are ml.g6.2xlarge, ml.g6.4xlarge and ml.g6.12xlarge; ml.g6.4xlarge is the default. The walkthrough requires an AWS account, an execution role, the SageMaker Python SDK v3 and sufficient service quota. Users must also accept the model’s license terms to deploy it.

Predictions are returned in mmCIF format with confidence metrics, including pLDDT, PAE and PDE. The guide also describes supplying custom multiple sequence alignments, which AWS says can improve prediction accuracy while helping keep sequences private.

Sources

Companies in this story

Related stories

Explore

More articles