chemdraw-bridge

ChemDraw Bridge

A DeepSeek Harness plugin that bridges compound identity and literature structures to native, editable ChemDraw source files on this computer.

It accepts a CAS number, an English/IUPAC compound name, or a compound in a paper/SI PDF.

中文操作说明：README.zh.md

What it produces

• "<title> (CID <cid>).cdxml" — ChemDraw's native XML format. Double-click to open and edit in ChemDraw.
• "<title> (CID <cid>).cdx" — optional binary ChemDraw format, useful when inserting an OLE object into Word. Enable with also_cdx: true.
• "<title> (CID <cid>).meta.json" — provenance: PubChem CID, CAS, formula, InChIKey, SMILES, generation time, ChemDraw version and the validation result.

The .cdxml file is written by ChemDraw itself, not hand-authored: the plugin pastes the PubChem isomeric SMILES into a hidden ChemDraw document and saves that document. ChemDraw therefore supplies the layout, implicit hydrogens, charges and stereo wedges.

How identity is resolved

1. CAS number (e.g. 50-78-2): the check digit is validated locally. PubChem is queried, and the returned CID is accepted only if the exact CAS string appears in that CID's synonym list. A bad check digit or an unconfirmed CAS is rejected rather than guessed.
2. Name (e.g. aspirin, L-ascorbic acid): PubChem autocomplete and ChemicalBook are searched, with prefix probes for common typos. One exact match is used directly; multiple candidates require the operator to select one before generation.
3. CID (CID:2244 or a bare integer): used directly.

Chinese names are resolved through ChemicalBook when possible; CAS remains the most reliable input.

Fuzzy names and operator selection

A name does not have to be exact:

1. PubChem autocomplete supplies name suggestions, and a few shorter prefixes are probed to catch common typos (caffiene → caffeine, asprin → aspirin).
2. ChemicalBook is searched as a secondary, best-effort source; it has no public API and is skipped when unavailable.
3. If one exact match is found, it is drawn directly.
4. If several candidates remain, draw_from_identifier returns the ranked list and asks the caller to use search_compounds; the operator chooses one, then the caller invokes draw_from_identifier with CID:<cid>.

CAS numbers and explicit CIDs never go through fuzzy matching.

SciFinder's search API is OAuth/ELN-oriented and returns a path into its web application rather than the result set. The precise route is the export workflow below; PubChem and ChemicalBook remain the fuzzy-search sources.

SciFinder exports

A logged-in SciFinder user can export the selected substance records and let the plugin read them:

1. In SciFinder, search for the substance, select the target record(s), and export SDF (preferred, carries the authoritative structure) or CSV/TSV with CAS numbers.
2. Call import_scifinder(file_path="...").
3. The tool returns candidates; the operator chooses one.
4. Draw it with draw_from_identifier(identifier="SMILES:<smiles>"), or with CID:<cid> when the CSV row resolved to PubChem.

Supported export formats: .sdf, .csv, .tsv, .txt. SDF is read with RDKit and compared with the CDXML after ChemDraw writes it, so the structure does not depend on a name lookup.

Image recognition and reaction drafts

recognize_image runs OSRA on a tight structure image, then validates every SMILES with RDKit:

recognize_image(image_path="E:\\figs\\compound.png")
recognize_image(pdf_path="E:\\paper.pdf", page=5, box=[100, 200, 360, 430])

This is for compounds with no printed IUPAC name. The output is a candidate list, not a verified identity: OSRA can misread stereochemistry, charges, isotopes and abbreviations. Abbreviated groups such as PhthN, MQ or Boc cannot be reconstructed from pixels; the tool refuses rather than inventing them. Confirm the candidate, then draw it with draw_from_identifier(identifier="SMILES:<smiles>").

draw_reaction drafts a one-step scheme from SMILES:

draw_reaction(
  reactants=["CCO"],
  products=["CC=O"],
  conditions="Pd(OAc)2, 80 °C",
  yield_text="82%",
  also_cdx=true
)

It places structures, arrow, conditions and yield in one ChemDraw document. This is a layout draft: check arrow direction, stoichiometry, conditions and products before using it in a paper.

Publication layout

Structures are laid out with RDKit CoordGen, which avoids ChemDraw's tall clipboard layout for large molecules. ChemDraw then applies the journal-style defaults used by the plugin:

• Arial 9 pt atom labels;
• 0.4 cm (20 pt) bond length;
• 0.6 pt line width;
• 10 pt bond spacing and 1 pt margin;
• compact reaction arrows, 9 pt conditions/yields and joined plus signs.

The settings are written into the CDXML, so the file stays editable rather than being a flattened image.

Literature extraction

draw_from_paper extracts a compound from a paper or its supporting information:

1. It reads the PDF text and locates the requested label (for example 3a) and the IUPAC name printed above it.
2. It resolves that name with the public OPSIN name-to-structure service.
3. RDKit independently computes the formula and InChIKey; a mismatch is a hard failure.
4. If the SI also prints an HRMS formula (for example calc. for C27H19N3O4), the formula is cross-checked, allowing for common Na/K/Li adducts.
5. ChemDraw writes the .cdxml / optional .cdx.

A real example from 10.1039/c7cc02426f (compound 3a):

PDF text  ->  2-((3S,4S)-1-(5-Methoxyquinolin-8-yl)-2-oxo-4-phenylazetidin-3-yl)isoindoline-1,3-dione
OPSIN     ->  COC1=C2C=CC=NC2=C(C=C1)N1C([C@H]([C@@H]1C1=CC=CC=C1)N1C(C2=CC=CC=C2C1=O)=O)=O
RDKit     ->  C27H19N3O4, InChIKey QWWIAFJZFCQLMC-ZEQRLZLVSA-N
HRMS      ->  C27H19N3O4  (match)

source may be a DOI, a PDF URL, or a local PDF path. Publishers often block automated PDF downloads; in that case pass local_pdf with a PDF/SI you downloaded yourself. The DOI is still used for the citation metadata.

The tool refuses to guess when no unambiguous IUPAC name is found. Abbreviated schemes (PhthN, MQ, Boc as labels) cannot be reconstructed by image OCR alone, so a name is the reliable route.

Verification

Before returning, the plugin:

• parses the generated CDXML and compares its element counts with the source SMILES;
• checks that the heavy-atom count matches;
• checks that fragments and bonds exist;
• checks the VjCD0100 header when a .cdx is requested;
• writes the result into the sidecar JSON.

If the atom composition does not match, the generated files are deleted and the call fails.

Install

1. Clone or download this repository to a local folder.

2. Run the Python setup once (used by the CoordGen publication layout, draw_from_paper and import_scifinder):

   powershell -ExecutionPolicy Bypass -File scripts\setup-python.ps1
   

3. In DeepSeek Harness, install that folder as a bundle:

   plugin_manager install_bundle(target="E:\\path\\to\\chemdraw-bridge")
   

4. Enable the chemdraw-bridge plugin row if it is not already active.

For recognize_image, also install the small local OCR binary once:

powershell -ExecutionPolicy Bypass -File scripts\setup-osra.ps1

The vendor/venv and vendor/osra directories are intentionally not committed; the setup scripts create them.

Requirements

• Windows with ChemDraw installed and activated (tested with ChemOffice 2022 / ChemDraw 22.0.0.22).
• Network access to pubchem.ncbi.nlm.nih.gov and www.chemicalbook.com; the literature tool also uses opsin.ch.cam.ac.uk.
• No API key.

Tool

search_compounds

Returns ranked candidates with CID, CAS, formula, source and match score. Show them to the operator, then call draw_from_identifier with the chosen CID:<cid>.

import_scifinder

For an SDF record, call draw_from_identifier with SMILES:<smiles>; for a CSV row that resolved to PubChem, CID:<cid> also works.

recognize_image

Returns OSRA candidates with formula and SMILES. This is not an identity proof; review before drawing.

draw_reaction

The reaction is a one-step layout draft; verify it before publication.

draw_from_identifier

Config keys in cordis.patch.yml are outputDir, alsoCdx, timeoutMs and pythonPath.

draw_from_paper

The publication layout, literature extraction and SciFinder import paths need a Python environment with PyMuPDF and RDKit. On a fresh clone, run this once:

powershell -ExecutionPolicy Bypass -File scripts\setup-python.ps1

It creates vendor/venv and installs requirements.txt. OPSIN is called over HTTPS at https://opsin.ch.cam.ac.uk; no API key is needed.

Command line

The bundled CLI runs the same pipeline:

node tools/cli.mjs "50-78-2" --out "E:\ds\chemdraw"
node tools/cli.mjs "cholesterol" --cdx
node tools/cli.mjs "CID:2244" --name aspirin

Implementation notes

• lib/pubchem.js — PUG REST client and strict identifier resolution.
• lib/literature.js — PDF text, label/name extraction, OPSIN and RDKit checks.
• lib/generate.js — shared SMILES → ChemDraw → CDXML validation step.
• lib/coordgen.js — RDKit CoordGen 2D layout; gives ChemDraw a compact, publication-style starting geometry.
• lib/vision.js — OSRA image recognition with RDKit validation.
• lib/reaction.js — one-step reaction-scheme layout and arrow/caption assembly.
• scripts/pdf_tool.py — PyMuPDF text extraction and RDKit formula/InChIKey validation.
• lib/smiles.js — small atom counter used only for validation.
• lib/cdxml.js — CDXML reader and element-composition comparison.
• lib/chemdraw.js — PowerShell/COM bridge with serialization, timeout and output checks. ChemDraw's SaveAs is unreliable with non-ASCII paths, so it always writes into an ASCII staging directory and Node moves the finished files to the requested path.
• scripts/convert.ps1 — opens a hidden ChemDraw document from a CoordGen MOL (or pastes SMILES as a fallback), applies publication-style settings, and saves CDXML/CDX.

Only one conversion runs at a time because ChemDraw automation and the clipboard are process-global. The tool is marked non-concurrency-safe.

Limitations

• Structures come from PubChem's canonical isomeric SMILES (identifier tool) or from the paper's IUPAC name via OPSIN (literature tool). A record without a SMILES, and a paper without an IUPAC name, cannot be drawn reliably.
• Image-only figures can be sent through recognize_image, but the result is a candidate and must be reviewed. Abbreviated schemes (PhthN, MQ, Boc labels) cannot be reconstructed from pixels alone.
• recognize_image and draw_reaction are drafts: check stereochemistry, charges, arrow direction, conditions and stoichiometry before publication.
• Organometallic coordination bonds are not inferred by ChemDraw's SMILES paste. Ferrocene, for example, arrives as two cyclopentadienyl rings plus a separate Fe atom; add the dative bonds manually if the figure needs them.
• Large structures are laid out with RDKit CoordGen, then styled by ChemDraw. A human may still adjust a crowded substituent or a very long chain.
• If ChemDraw is already open, the plugin attaches to that instance and closes only the document it created. It does not quit the user's session.
• A name that is ambiguous in PubChem is rejected with candidate CIDs instead of silently choosing one.

Changelog

See CHANGELOG.md for the release history, including the 1.1.0 publication-layout, image-recognition and reaction-draft changes.