OpenFold3 Lands on SageMaker JumpStart for Asynchronous Biomolecular Structure Prediction
Summary:AWS has made OpenFold3-preview available through SageMaker JumpStart, allowing research groups and pharmaceutical teams to deploy the Apache 2.0-licensed all-atom structure prediction model inside their own AWS accounts. The managed endpoint supports proteins, DNA, RNA, ligands, and multi-chain complexes through asynchronous inference backed by Amazon S3. Its deployment path addresses several practical obstacles in biomolecular inference: GPU and CUDA/PyTorch setup, minute-scale serving jobs, m
OpenFold3 Enters a Managed Inference Stack
AWS has added OpenFold3-preview to SageMaker JumpStart. Developed by the OpenFold Consortium and Columbia University's AlQuraishi Lab, the model is distributed under the Apache 2.0 license and performs joint all-atom structure prediction for proteins, DNA, RNA, small-molecule ligands, and their complexes.
The current JumpStart offering uses OpenFold3 0.4.1 with the of3-p2-155k checkpoint and supports ml.g6.2xlarge, ml.g6.4xlarge, and ml.g6.12xlarge. Predictions are returned as mmCIF structures together with pLDDT, pTM, ipTM, PAE, PDE, and other confidence metrics.
The deployment's main value is not another model announcement. It turns a GPU-intensive scientific model that normally requires CUDA, PyTorch, checkpoint management, and a custom serving layer into a production endpoint managed through the SageMaker Python SDK.
Why Asynchronous Inference Is Necessary
Structure prediction time grows with the number and length of chains and with complex composition. Real biological queries often take several minutes, exceeding the 60-second ceiling typical of real-time inference. OpenFold3 therefore uses SageMaker Asynchronous Inference:
- The client writes the request to Amazon S3 and calls
InvokeEndpointAsync. - The endpoint queues and processes the job, with an invocation timeout of up to one hour.
- Results or failure details are written to S3, with optional SNS notifications.
This pattern absorbs bursts of sequence submissions without requiring a separate job queue. Its cost is weaker request semantics: callers must handle queueing, polling, failure objects, and timeouts. Interactive workflows should retain a synchronous job layer rather than treating every prediction as an asynchronous API call.
Multi-GPU Scaling Improves Throughput, Not Single-Job Latency
One OpenFold3 prediction runs on one GPU. The four GPUs in an ml.g6.12xlarge instance therefore do not accelerate an individual fold. Instead, max_concurrent_invocations_per_instance=4 assigns four concurrent requests to four separate GPUs, increasing the number of completed predictions per unit time.
AWS reports that the gain grows for larger proteins and that, using on-demand pricing, the measured cost per prediction on ml.g6.12xlarge approaches that on ml.g6.4xlarge. The result depends on protein size, queue utilization, regional pricing, and instance availability; it should not be generalized to every complex without workload-specific measurements.
Two operational issues remain: multi-GPU utilization does not eliminate the cost of idle instances, and G6 capacity can be constrained in some regions. Cross-instance autoscaling or another region can address availability, but it adds operational complexity.
MSAs Are the Central Accuracy and Privacy Lever
OpenFold3 can predict from a single sequence or accept precomputed multiple sequence alignments. Covariation across aligned homologs provides evolutionary constraints on residue contacts and complex interfaces, making MSAs particularly useful for targets without experimental structures, protein–nucleic-acid assemblies, and protein–ligand co-folding.
The JumpStart endpoint runs with network isolation and does not call external MSA servers. Users generate alignments with ColabFold/MMseqs2, HHblits, or JackHMMER in infrastructure they control, then submit A3M or Stockholm-formatted data. This avoids sending proprietary sequences to a public search service.
AWS compared the carbonic anhydrase II–zinc–acetazolamide complex with and without an MSA. With the MSA, average pLDDT rose from 30 to 98, pTM from 0.22 to 0.98, and ipTM from 0.49 to 0.98. Backbone Cα RMSD improved from 22 Å to 0.2 Å, while both runs took approximately 1.5 minutes.
This is not a strict novel-target benchmark. The complex is extensively studied in the Protein Data Bank, and related structures may be represented in the training data. The experiment demonstrates the operational value of an MSA-based workflow, not universal accuracy on unseen targets. A rigorous evaluation should use temporally split or training-deduplicated targets and compare MSA, single-sequence, and experimental references consistently.
Confidence Scores Are Not Experimental Validation
pLDDT estimates local structural confidence, with values above roughly 70 often treated as more reliable. It does not answer every question. For complexes, ipTM and chain-pair ipTM are critical for assessing inter-chain placement, while PAE and PDE expose uncertain domain or chain arrangements.
By default, OpenFold3 generates five diffusion samples per query. The sample_ranking_score can rank candidates, but it remains a model-derived estimate rather than external validation. Important targets still require geometry and clash checks, ligand chemistry and valence validation, energy assessment, and biochemical or crystallographic confirmation.
Scale-to-Zero Saves Cost at the Expense of Cold Starts
The asynchronous endpoint can scale using ApproximateBacklogSizePerInstance and use HasBacklogWithoutCapacity to restart from zero instances. AWS observed scale-in roughly 15 to 20 minutes after the queue emptied.
This suits bursty drug-discovery workloads: GPUs are absent between experiments and return when new jobs arrive. Interactive use should retain MinCapacity=1, otherwise the first request must wait for alarm evaluation, instance startup, and GPU warm-up in addition to inference time.
Scaling to zero reduces endpoint compute charges but does not remove S3 storage or other service costs. An idle endpoint still incurs instance-hour charges, so experimental deployments should delete endpoints, configurations, models, and unnecessary request and result objects when finished.
Deployment Guidance and Security Boundaries
Start with ml.g6.4xlarge to validate accuracy, then select ml.g6.12xlarge according to concurrent queue depth rather than the number of proteins being evaluated. Batch pipelines should place requests and outputs in dedicated S3 prefixes, encrypt them with customer-managed KMS keys, and grant the execution role access only to the relevant paths.
Regulated workloads also require audits of container images, model weights, retention periods, regional boundaries, and experiment records. Apache 2.0 permits inspection, fine-tuning, and reuse, but it does not automatically resolve training-data rights, regulatory approval, or biosafety requirements.
Overall, the JumpStart integration turns OpenFold3 into an operable batch service: asynchronous queues accommodate minute-scale inference, multi-GPU instances increase throughput, private MSAs support both accuracy and data isolation, and scale-to-zero reduces idle compute costs. The model's accuracy on genuinely novel targets still requires independent evaluation.
— END —Source: AWS Machine Learning Blog (2026-11-13)
Tags: #OpenFold3 #Biomolecular Structure Prediction #SageMaker JumpStart #Asynchronous Inference #Multiple Sequence Alignment
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments