Loading…
Protein complex structure prediction from amino acid sequence, deployed as a self-hosted SageMaker endpoint that returns mmCIF and ipTM confidence.
Anthrofold is a protein structure prediction model from Anthrogen that takes amino acid sequences for one or more polypeptide chains and returns 3D atomic coordinates. It is distributed exclusively as an Amazon SageMaker model package on AWS Marketplace, where the current release is v1.1.0, published in July 2026 and available in sixteen AWS regions. Rather than an API or a downloadable checkpoint, it ships as a container the customer deploys into their own account, and the listing states that no data leaves that account and the container makes no external API calls.
Anthrofold is a separate product line from Anthrogen's other published model, Odyssey, a 102-billion-parameter multimodal system that generates and edits protein sequence and structure through discrete diffusion. Where Odyssey is a design engine, Anthrofold is a fixed-checkpoint predictor — the read-out side of the same workflow, used to fold a given sequence and score how confidently it assembles. Anthrogen's own website documents Odyssey and does not mention Anthrofold; the marketplace listing and a companion deployment repository are the only documentation that exists for it.
That absence shapes how the model should be read. There is no paper, preprint, or technical report, no model source code or released weights, and no published benchmark results — the listing reports no accuracy figures of any kind, and architecture, parameter count, and training data are undisclosed. This sets it apart from openly distributed complex predictors such as Boltz-2, Chai-1, and AlphaFold 3, whose performance can be measured independently.
What the listing documents is the behavior of the deployed container rather than the model that sits inside it. On first deployment the container synchronizes roughly 1 TB of ColabFold MSA databases and a 224 GB mmCIF template corpus onto instance NVMe, a one-time sync of 60-80 minutes served from a Requester Pays S3 bucket, so the egress is billed to the customer. Inference is asynchronous only: predictions take one to three minutes for complexes up to 250 residues and three to seven minutes for larger antibody-antigen complexes, exceeding SageMaker's 60-second real-time invocation timeout, and each instance is restricted to a single concurrent invocation. The model samples across random seeds and uses its ranking score to select which sample is returned. The v1.1.0 release notes describe a new architecture giving more predictable inference-time scaling and more accurate confidence prediction than the prior version. Glycans, ligands, and post-translational modifications are not supported: the request schema reserves fields for chain modifications and covalent bonds but requires them to be empty.
The listing positions Anthrofold for antibody engineering and enterprise drug discovery, and its output schema reflects that — ipTM is singled out as the metric that matters for antibody-antigen interfaces, with a documented band above 0.8 for high-confidence predictions and 0.6-0.8 as useful for triage. The other intended uses are deep mutational scanning, target identification, and structural checking of engineered or de novo designed sequences, including candidates coming out of generative design models. The delivery model suits organizations already building on AWS that want a folding step inside an existing pipeline without routing proprietary sequences through an external service.
Anthrofold is most notable for how it is delivered: a commercial folding engine packaged for in-account, in-VPC deployment with no software fee, where comparable capability is usually obtained either as an open checkpoint or a hosted API. Its confidence vocabulary matches what downstream tooling already expects from structure predictors, which makes substitution into an existing pipeline straightforward. Its standing in the field is nonetheless hard to establish: without a paper, released weights, or any published benchmark, the model cannot be compared against open predictors on common targets, and adopters must validate it on their own systems before trusting it — a real cost for a step that most structural biology pipelines expect to be able to audit.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.