Ptarmigan-1 documentation

Ptarmigan-1 predicts which compounds engage a protein target from the protein's sequence alone, with no structure needed. Pick a human protein and a compound library, run a screen, and every compound in the library comes back ranked by predicted engagement. The method is described in the Ptarmigan-1 preprint.

To try it now, the quickstart runs a first screen in the web app for a few cents.

Ways to use it

You want to Use Guide
Run screens and explore results in your browser The web app at https://ptarmigan.app Using the web app
Let an AI agent screen for you MCP, with an API key Using an agent
Script screens from a shell The REST API with curl REST with curl
Analyse results in Python or a notebook The REST API with httpx Python, which links a getting-started notebook to download

All four use the same account, libraries and balance. A screen started in one shows up in the others.

Concepts

Target. The protein you screen against. Search by gene, name or UniProt accession. You can screen only the human proteins Ptarmigan-1 provides; adding your own protein is not supported yet.

Library. A set of compounds, each a SMILES string with an optional id and name. Use a provided library, or upload your own; see Library files.

Screen. One target against one library. A screen is queued, runs, and ranks every compound in the library. It is charged only when it completes; a screen that fails or is cancelled costs nothing.

Predicted engagement. A number between 0 and 1: how likely the compound is to engage the target. A screen lists compounds from highest to lowest predicted engagement; see Reading the results.

Tokens and cost. A screen costs one token per compound. The price per token is on the pricing page, with what is free, and every estimate shows a screen's cost before anything is spent.

Balance. Screens are paid from a prepaid balance in dollars. A new account starts with $250 of promotional funds, which are spent first. You can add funds by card from the Billing page. Funds expire 6 months after they are added. Funds are dated to the start of the hour (UTC) in which they were added, so a grant's granted_at and expires_at in /api/v1/balance fall on the hour: funds bought at 21:29 show 21:00.

Results. A completed screen's full ranking stays available in the app for 6 months. Download it to keep it for longer.

Reference