OpenFold3 on SageMaker JumpStart: Async Protein Folding in AWS

AWS added OpenFold3-preview to SageMaker JumpStart on November 13, 2026, with a full walkthrough of asynchronous inference, bring-your-own multiple sequence alignments (MSAs), multi-GPU throughput, and scale-to-zero idle savings. The AWS Machine Learning Blog post targets computational drug discovery teams who already live in AWS but do not want to babysit CUDA stacks for structure prediction.

OpenFold3 is an all-atom model from the OpenFold Consortium and Columbia’s AlQuraishi Lab — Apache 2.0 (see the OpenFold3 GitHub repository), version 0.4.1 with the of3-p2-155k checkpoint. It is the same family NVIDIA cites inside BioNeMo protein design workflows and Claude Science integrations.

TL;DR

Why async inference matters

Real-time SageMaker endpoints must answer within 60 seconds. Protein and protein–ligand jobs routinely exceed that. AWS routes OpenFold3 through Asynchronous Inference:

1. Upload JSON payload to S3.
2. Call InvokeEndpointAsync — returns immediately.
3. Endpoint queues work (timeout up to one hour).
4. Results land in your output S3 prefix (failures in a separate failure path).

That pattern suits batch screening: queue hundreds of sequences, autoscale on queue depth, drain results from S3 into your LIMS. It parallels how teams already orchestrate Bedrock AgentCore long jobs, but for structural biology.

OpenFold3 on SageMaker JumpStart predicts biomolecular 3D structures from protein and ligand sequences

Prerequisites (from AWS)

AWS lists:

• SageMaker execution role with least-privilege S3 on your request/result prefixes.
• Caller permissions to create, invoke, and delete endpoints plus iam:PassRole.
• SageMaker Python SDK v3 and optional py3Dmol for notebook viewing.
• Service quota for at least one ml.g6.4xlarge (more if you use ml.g6.12xlarge).

AWS also warns: endpoints accrue instance-hour charges while running, idle included. The post’s cleanup section and scale-to-zero instructions are not optional niceties — they are cost controls.

Deploy pattern (SDK v3 sketch)

The blog’s canonical deploy uses ModelBuilder.from_jumpstart_config:

from sagemaker.core.inference_config import AsyncInferenceConfig
from sagemaker.core.jumpstart.configs import JumpStartConfig
from sagemaker.core.training.configs import Compute
from sagemaker.serve.model_builder import ModelBuilder

model_builder = ModelBuilder.from_jumpstart_config(
    jumpstart_config=JumpStartConfig(
        model_id="huggingface-proteinfolding-openfold3",
        accept_eula=True,
    ),
    compute=Compute(instance_type="ml.g6.4xlarge"),
)
model = model_builder.build()

async_config = AsyncInferenceConfig(
    output_path=f"s3://{bucket}/{prefix}/output",
    failure_path=f"s3://{bucket}/{prefix}/failure",
)
endpoint = model_builder.deploy(inference_config=async_config)

accept_eula=True is mandatory — you are accepting model license terms on deploy.

Submit a ubiquitin example

AWS’s human ubiquitin demo packs a protein chain in JSON, stages to S3, and polls output locations until JSON returns structures or an error object with "status": "error".

Confidence summaries include avg_plddt, pTM, ipTM, and sample_ranking_score — use ranking score to pick among five diffusion samples per query (default).

Multi-GPU: throughput, not single-job speed

A single fold uses one GPU. ml.g6.12xlarge helps when you run several jobs concurrently — set max_concurrent_invocations_per_instance=4 to match four GPUs. AWS notes ml.g6.12xlarge capacity can be scarce; scaling out on multiple ml.g6.4xlarge nodes may be easier.

For antibody or screening campaigns, think queue depth + autoscaling, not “one protein faster.”

Bring your own MSA (privacy and accuracy)

JumpStart endpoints run network isolated and do not hit public MSA servers. AWS recommends MSAs for most real targets; ubiquitin may work sequence-only, but complexes like carbonic anhydrase + zinc + acetazolamide depend on alignment quality.

Each protein chain can include:

"precomputed_msa": [
  {"name": "colabfold_main", "content": "<A3M or Stockholm text>"}
]

Generate alignments on infrastructure you control (ColabFold pipeline, internal HPC, etc.), then pass text into the request. Combine with KMS on async output paths when handling proprietary sequences.

Bring-your-own MSA workflows keep proprietary protein sequences inside your AWS account during OpenFold3 prediction

Scale to zero

Async endpoints can scale in to zero instances when the queue is empty — critical for academic groups and biotech startups that fold in bursts. Wire CloudWatch alarms on queue depth if you need wake latency SLAs.

Compare cost thinking to tracking Bedrock spend per principal: structure endpoints are GPU-heavy line items.

How this fits explainx.ai’s existing map

None replace experimental validation — pLDDT is confidence, not binding proof. AWS repeats that ligand poses are hypotheses for medicinal chemistry, not clinic-ready assets.

Confidence metrics without fooling yourself

AWS documents multiple outputs per sample:

• pLDDT — per-atom local confidence; values above ~70 often indicate trustworthy local geometry.
• pTM / ipTM — global and interface confidence for complexes; for ligands, ipTM is the number medicinal chemists watch.
• PAE / PDE — pair errors for domain placement; read these when mixing protein and DNA chains.

A strong pLDDT on a ligand pose does not prove binding affinity. Combine with orthogonal checks: molecular dynamics, assay data, or retrospective benchmarks on known PDB complexes like AWS’s carbonic anhydrase + acetazolamide demo (PDB 3HS4 cited in the post).

For teams comparing to AlphaFold-era explainers, OpenFold3’s pitch is all-atom complexes in one forward pass, not a pipeline of separate folding and docking tools.

Ligand co-folding JSON shape

AWS’s complex example packs three chains: protein sequence, zinc ion as SMILES [Zn+2], and acetazolamide as a SMILES string on separate ligand chain IDs. That pattern generalizes to other covalent and ion cofactors if your chemistry is encoded correctly.

Common failure modes from the post:

• Invalid SMILES → "status": "error" object with message instead of structures.
• Missing MSA on hard targets → low ipTM despite pretty pLDDT visuals.
• Oversized complexes → timeout unless you raise async timeout below the one-hour cap.

Use py3Dmol in SageMaker Studio to color by B-factor (pLDDT proxy) exactly as AWS demonstrates — a fast visual sanity check before paying for wet-lab synthesis.

IAM, encryption, and compliance hooks

AWS recommends:

• Dedicated S3 bucket or prefix with SSE-KMS for proprietary sequences.
• Bucket policies restricting roles to the SageMaker execution role and your pipeline workers.
• Lifecycle rules to expire request/result objects after N days for GDPR or trade-secret retention policies.

If you already govern LLM spend with Bedrock KB ACL interceptors, reuse the same data-classification labels on structure inputs — many biotech teams treat sequences as confidential even when models are open weights.

Cleanup and cost traps

The blog’s Clean up section is emphatic for a reason. Async endpoints with autoscaling can still leave minimum instances above zero if misconfigured. Scale-to-zero helps, but cold starts add latency for the first fold after idle periods — plan queue polling accordingly.

Document spend in FinOps reviews next to four inference choices on Bedrock: structure GPUs are a separate SKU from token billing, but finance still rolls them into the same AI budget envelope.

What readers should do this week

1. Quota-check ml.g6 in your region before demo day.
2. Prototype async with ubiquitin, then one real complex with MSA.
3. Set lifecycle rules on S3 prefixes holding sequences.
4. Tear down endpoints — schedule a calendar reminder.
5. If you also run LLM agents on structures, read AWS agent DLC production patterns for hygiene, not this post’s scope.

Notebook users should bookmark the JumpStart example notebook AWS links from Studio — it is the maintained source of truth when SDK constructor names drift between SageMaker versions. Pin SDK versions in your repo the same way you pin inference containers, or async polling scripts will break silently when response JSON shapes change.

Batch orchestration pattern: Many teams wrap InvokeEndpointAsync behind Step Functions or Apache Airflow tasks that (1) shard FASTA or JSON inputs, (2) write manifests to S3, (3) poll failure prefixes for malformed sequences before requeueing, and (4) merge mmCIF outputs into structure databases. That glue code is not in the AWS blog, but it is where engineering time actually goes after the demo ubiquitin fold works.

When not to use JumpStart: If you need interactive single-structure latency under a minute in a browser UI, a dedicated GPU notebook with open weights may feel better. JumpStart wins when compliance, IAM, and async batch matter more than sub-minute feedback — typical for discovery IT groups already standardized on SageMaker for SGLang Trainium DLC experiments.

Partner handoffs: CRO teams often need mmCIF in both internal Vault and partner-facing portals. Automate virus scanning and metadata tagging on S3 output prefixes the same way you gate model weights downloads — structure files still contain sequence intellectual property even when the model weights are Apache licensed.

Disaster recovery: Snapshot your endpoint configuration, async S3 paths, and autoscaling policies in Infrastructure-as-Code. Re-deploying OpenFold from scratch after a region outage is painful if the only copy of AsyncInferenceConfig lived in a lost notebook kernel.

Support expectations: OpenFold3-preview is a research preview per AWS’s table footnotes. Production drug programs should plan validation gates with structural biologists before advancing poses to synthesis, regardless of how polished the JumpStart container looks. Keep a human sign-off step even when async queues are fully automated.

Related reading

• NVIDIA BioNeMo inference runtime
• BioNeMo + Claude Science toolkit
• RFD4-Proteina and OpenFold3 rewards
• Bedrock AgentCore next-gen runtime
• SageMaker SGLang on Trainium DLC
• Cap Bedrock spend guide
• AWS source: OpenFold3 JumpStart post

AWS steps and instance names reflect the November 13, 2026 blog post.