Patent¶

Extract chemical structures from a patent or chemistry PDF with the Deep Origin Patent tool ("DO Patent"). Patent reads the pages of a PDF, detects drawn molecules, and predicts a SMILES string for each one, together with a confidence score and the page it was found on.

Patent is an async tool: submit the job with start(), track it with watch() or wait(), then read the extracted structures with get_results().

Workflow¶

The lifecycle mirrors the other async tools (see Tools on Deep Origin):

1. Create a Patent job from a local .pdf.
2. Estimate the cost with start(quote=True) and inspect estimate.
3. confirm() the quoted execution to run it.
4. Track progress with watch(), or block with wait().
5. Read the extracted structures with get_results().

The PDF is validated on construction (it must exist and end in .pdf) and is uploaded to Deep Origin storage automatically on start().

Extracting structures¶

from deeporigin.drug_discovery import Patent

# Point at a local PDF; the file is validated immediately.
patent = Patent(pdf="path/to/patent.pdf")

# Upload and quote without running.
patent.start(quote=True)
patent.estimate  # estimated cost

# Confirm to run, then wait for completion.
patent.confirm()
patent.wait()

# One row per extracted molecule.
df = patent.get_results()
df.head()

In a notebook, replace patent.wait() with await patent.watch() to render a live progress card. When executing a notebook headlessly (for example with nbconvert), set JOB_WATCH_BLOCK=1 or call await patent.watch(blocking=True) so the cell blocks until the job finishes.

Results¶

get_results() returns a pandas.DataFrame with one row per extracted molecule, or None until the job has succeeded. The main columns are:

Existing executions¶

Every execution has an ID. Reconstruct a Patent object from that ID in a later session and re-fetch its results without re-running the extraction:

patent = Patent.from_id("<executionId>")
df = patent.get_results()

If you already have the execution payload from client.executions.get, use Patent.from_dto(dto) instead. You can also list previous runs and resume the most recent one with Patent.list() and Patent.from_last_run().

See the API reference for the full signature.