Use your own model in Talindrew, including AlphaFold 3
Maintained by the Talindrew team · Last updated
A custom model in Talindrew is an HTTPS endpoint you run that fills one stage of a FoldRx drug-discovery pipeline: fold, pockets, dock, ADMET, pose checks, selectivity or design. Talindrew sends it exactly the JSON the built-in model for that stage receives and expects the same shape back, either at once or by answering 202 with an nvcf-reqid header that Talindrew then polls. It appears in that stage's model list for your account only. This is how a lab that holds an AlphaFold 3 licence folds its screens with AlphaFold 3, which Talindrew may not run for anyone itself.
Watch · 90 sec
Drug Discovery in 90 Seconds: Build and Run a Screen in Talindrew's AI Workspace
Chapters
What does the fold stage send and expect back?
The fold stage posts {"sequence": "MKT...", "params": {...}} with the protein's one-letter sequence. Answer with JSON whose "pdb" field is a PDB-format structure with ATOM records. Optional fields are used when present: "confidence" as the mean pLDDT on a 0 to 1 scale (divide AlphaFold's 0 to 100 values by 100), "per_residue_confidence" as a list on the same scale, and "ptm". The other stages work the same way: your endpoint receives what Talindrew's own model for that stage receives.
How do long-running models answer?
A model that takes minutes should answer the POST with HTTP 202 and an nvcf-reqid header holding a job id. Talindrew then polls GET <status base>/<job id>, by default the endpoint's origin plus /status, every few seconds: 202 means still running, 200 carries the result, and any 4xx or 5xx with a {"detail": "..."} body fails the stage with that message. This is the protocol NVIDIA's hosted NIMs use, so a NIM you run can be registered as it is.
Serve your own AlphaFold 3, step by step
The adapter is one Python file with no dependencies beyond Biopython. It runs your AlphaFold 3 for each request and answers in the fold stage's shape.
- Install AlphaFold 3 under your licence from Google DeepMind and confirm run_alphafold.py works on one input.
- Download the adapter from www.talindrew.com/downloads/alphafold3_adapter.py and install Biopython (pip install biopython).
- Set AF3_COMMAND to the command you already use to run AlphaFold 3, with {json} where the input file goes and {out} where the output folder goes; the adapter's header has a complete example with your model and database folders.
- Set ADAPTER_TOKEN to a long random string and start the adapter: python alphafold3_adapter.py (it listens on port 8787).
- Put it behind HTTPS, with your reverse proxy or a tunnel; Talindrew only calls https addresses.
- In the AI workspace, select the Fold stage, choose Use your own model, pick AlphaFold 3 (my licence), paste the URL ending in /fold and the token, and press Add and test.
What does Talindrew check, and what does it never do?
Every saved URL must be https and must not resolve to a private, loopback, link-local or cloud metadata address; it is checked when you save it and again before every request, and redirects are not followed. Your token is stored encrypted, sent only as a bearer token to your endpoint, and never shown again. A custom model is visible and usable only by the account that registered it. Talindrew does not copy, cache or run your model; a moved endpoint counts as a new model, so earlier results are not reused for it.
Custom model contract by stage
| Stage | Talindrew sends | Your endpoint returns |
|---|---|---|
| Fold | sequence, params | pdb (required); confidence 0 to 1, per_residue_confidence, ptm |
| Pockets | pdb, params | pockets: centre, score, volume per pocket |
| Dock | receptor, pocket, compounds, params | poses: id, smiles, score (kcal/mol) per compound |
| ADMET, pose checks, selectivity, design | the same inputs as the built-in model | the same shape as the built-in model |
Frequently asked questions
Why can Talindrew not run AlphaFold 3 for me?
Google DeepMind's terms for the AlphaFold 3 model parameters allow use only by or on behalf of non-commercial organisations and prohibit use in connection with commercial activities. Talindrew is a commercial service, so it cannot run the parameters for you. A licensee running AlphaFold 3 on their own hardware and calling it from Talindrew is a different arrangement; check that your own use fits the terms.
Can I use ESMFold2 or a fine-tuned model the same way?
Yes. Any model you can put behind an HTTPS endpoint that accepts the stage's request and answers in its shape works, including ESMFold2 (MIT licensed), a fine-tuned structure model, or an in-house docking engine.
Do custom models cost runs on my plan?
A FoldRx run is counted the same way whichever models fill its stages, because the other stages still run on Talindrew's compute. Your own endpoint's costs are yours.
References
- Google DeepMind. AlphaFold 3 model parameters terms of use. https://github.com/google-deepmind/alphafold3/blob/main/WEIGHTS_TERMS_OF_USE.md
- Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630: 493–500 (2024). https://doi.org/10.1038/s41586-024-07487-w
- NVIDIA. NIM for BioNeMo: asynchronous requests and the nvcf-reqid status protocol. https://docs.api.nvidia.com/nim/reference/mit-boltz2
Related
Design your Use your own model in Talindrew, including AlphaFold 3 construct in Talindrew.
Free to use. No download required. Works in any modern browser.