ESMFold2 API Guide

A practical ESMFold2 API flow has four parts: choose the current official client, create a structure prediction input, set folding configuration, and save the resulting mmCIF plus optional confidence outputs.

The exact class names and package paths must be checked against Biohub's current model page, GitHub repository, or model card before publication.

The free tools on this site do not ask for a private Biohub token. Code examples use environment variables or placeholders, and generated code is created locally in the browser.

If hosted prediction is added later, token handling will be replaced by a first-party paid prediction layer with its own compliance review.

The MVP supports starter templates for a single protein, a protein plus RNA/DNA example, and a protein plus CCD ligand example. It does not promise arbitrary ligand docking.

Colab mode is useful for learning because it keeps the environment setup visible and easier to reset. Local Python mode is useful when the user already manages dependencies, tokens, and output directories in a controlled environment.

The generated snippets should not hide the fact that official examples can change. Any published code should include a "Last verified" date and should point readers back to Biohub sources for the current install and import path.

A responsible API guide should describe inputs at the workflow level without asking users to paste private sequences into the site. The browser-local generator can show safe placeholders, demo sequences, CCD-style ligand examples, and configuration fields.

For outputs, the guide should emphasize mmCIF or PDB handling, optional PAE where supported, and confidence interpretation. It should also say plainly that predicted structures are not experimental validation.

Proxying a free platform API creates cost, trust, and terms-of-use risk. It would require receiving user sequences and tokens, storing or forwarding sensitive scientific inputs, and taking responsibility for abuse controls.

The Track A MVP avoids that problem by staying local in the browser. Phase 2 hosted prediction, if launched, should run real ESMFold2 inference through a compliant paid layer rather than reselling another platform endpoint.

Before running generated ESMFold2 code, confirm the official model ID, package install command, input object shape, output save path, and whether PAE or confidence outputs are enabled by the current example.

Then confirm where your token lives. A safe local or Colab workflow should read a token from an environment variable or secret manager. It should not paste the token into a third-party website, a public notebook, or a shared script.

Finally, decide what you will do with the output. A generated mmCIF file, PDB file, pLDDT values, or PAE matrix should be treated as computational output that needs domain review before it is used in a report or downstream experiment.

This launch page is written as a third-party guide, not as an official Biohub page. It links to Biohub, GitHub, Colab, model-card, paper, and reporting sources so readers can verify current details before copying code or making technical decisions.

The free Track A site does not collect protein sequences, private Biohub tokens, prediction outputs, or uploaded structure files. Browser-local tools can prepare starter code and examples, but hosted prediction, stored results, queues, and credits remain Phase 2 features.

Predicted structures should not be described as experimental validation, clinical evidence, or drug-development proof. When this page mentions model scope or comparisons, it keeps claims tied to official or clearly attributed sources.

If a source changes after this page is published, the source should win. The page should be updated with a new verification date instead of preserving stale client names, model IDs, pricing assumptions, or unsupported benchmark language.