OpenFold3 Now Available on Amazon SageMaker JumpStart for Biomolecular Structure Prediction

OpenFold3-preview is now available on Amazon SageMaker JumpStart, allowing users to deploy the model into their own AWS accounts and predict the 3D structures of proteins, DNA, RNA, small-molecule ligands, and the complexes they form. The model was developed by the OpenFold Consortium and the AlQuraishi Lab at Columbia University.

Model details and availability

OpenFold3 is an all-atom structure prediction model that predicts the joint 3D structure of a set of molecules from their sequences. It can fold multi-chain protein complexes, protein–nucleic acid assemblies, and protein–ligand complexes in one prediction. The developers release its code and model weights under the Apache 2.0 license and publish the training data and methods. The developers describe OpenFold3-preview as a research preview.

The JumpStart model ID is huggingface-proteinfolding-openfold3. Supported instances are ml.g6.2xlarge, ml.g6.4xlarge (default), and ml.g6.12xlarge. The model version is OpenFold3 0.4.1, of3-p2-155k checkpoint.

How deployment works

OpenFold3 on JumpStart runs on SageMaker Asynchronous Inference. Users upload a request to Amazon S3 and call InvokeEndpointAsync, which returns immediately. The endpoint queues the request and processes it, with an invocation timeout of up to 1 hour. Results or error details are written to an S3 location chosen by the user.

Asynchronous endpoints queue bursts of requests internally and can scale to zero instances when the queue is empty, helping users avoid paying for idle instances between experiments. SageMaker endpoints are billed per instance-hour while they are running, including while they are idle.

Deployment typically completes within minutes and includes a warmup step in which the container runs a short prediction on each GPU.

Prediction inputs and outputs

A request is a JSON document with one or more named queries, and each query lists its chains. Each chain has fields for molecule_type, chain_ids, sequence, and smiles. Molecule types include protein, dna, rna, and ligand. Small-molecule ligands are supplied as SMILES strings.

The response contains a status, a list of predicted structures in mmCIF format, and their matching confidences. By default, OpenFold3 generates five diffusion samples per query, so each query returns five candidate structures. Each sample comes with two confidence files: a detailed one and a summary. The summary includes a sample_ranking_score.

Confidence metrics include pLDDT (per-atom confidence on a scale of 0–100), pTM and ipTM (overall confidence on a scale of 0–1), and PAE and PDE (per-residue-pair error estimates).

Multi-GPU and MSA options

A single OpenFold3 prediction runs on one GPU, so a multi-GPU instance does not make an individual prediction faster. Instead, users can run several predictions at once. On an ml.g6.12xlarge instance, which has four GPUs, the container assigns each concurrent request its own GPU.

OpenFold3 can take a multiple sequence alignment (MSA) as an additional input for each protein chain. The JumpStart endpoint runs with network isolation and does not call external MSA servers. Instead, it accepts precomputed MSAs in the request. If no MSA is supplied, the endpoint predicts from the sequence alone.

Why it matters

• OpenFold3 is an open source biomolecular structure prediction model developed by the OpenFold Consortium and the AlQuraishi Lab at Columbia University.
• The model is available on Amazon SageMaker JumpStart under the model ID huggingface-proteinfolding-openfold3.
• It predicts 3D structures of proteins, DNA, RNA, small-molecule ligands, and their complexes.
• Deployment runs on SageMaker Asynchronous Inference with support for scale-to-zero and multi-GPU parallel predictions.

---

Sources

• aws.amazon.com: OpenFold3 on Amazon SageMaker JumpStart: Predict biomolecular structures | Artificial Intelligence

This article was drafted with AI assistance and checked against the sources above. Company claims are reported as claims. Cover image is AI-generated.