Drybench Molecule.

What the app shows, how it works and how it was built, in the words of its own write-up, followed by the parameter defaults and the data sources with their licences.

What it shows

Give it a molecule by drawing it, pasting a SMILES string or typing a name, and it reports the standard medicinal-chemistry descriptors (molecular weight, Crippen logP, polar surface area, hydrogen-bond donors and acceptors, rotatable bonds, rings, fraction sp3), Lipinski's rule of 5 and Veber's rules with the reason for each pass or fail, the QED drug-likeness score with the eight inputs that set it, and any PAINS substructure alerts.

It then builds a 3D shape, finds the ten most structurally similar approved drugs, and places the molecule on a map of 2,197 approved small-molecule drugs from ChEMBL 37, so you can see whether its property profile looks like that of approved drugs (ChEMBL max_phase 4: approved in at least one country, including some later withdrawn). The map colours each drug by the routes ChEMBL records for it: oral, parenteral (injected or infused) but not oral, topical only, or other or unknown; a drug given both by mouth and by injection counts as oral. Lipinski's rule of 5 is about absorption and permeation and Veber's rules about oral bioavailability, so the colours show where orally given drugs sit among the rest. They are records for approved drugs, not a forecast for your molecule. Links of the form ?smiles= or ?name= open a molecule directly, and the address updates as you work, so any view can be shared. If a link has both, the SMILES decides the structure and the name is only a label.

How it works

All chemistry runs in your browser, apart from the opening example (imatinib), whose results the pipeline works out in advance with the same code, so opening the page does not download the chemistry engine; it loads when you choose, type or draw a molecule, or open a link to one. RDKit.js (the RDKit C++ toolkit compiled to WebAssembly) runs in a Web Worker and computes the descriptors, the 2D drawing and Morgan fingerprints (radius 2, 2,048 bits). RDKit.js has no QED, no PAINS catalogue and no similarity function, so those are ported: QED follows RDKit's QED.py with the same patterns, parameters and weights, PAINS uses the 480 filters from RDKit's FilterCatalog, and Tanimoto similarity is computed on the fingerprint bits. A salt or other multi-part input is described by its largest organic fragment (imatinib mesylate as imatinib), unless another part holds a metal other than a simple counter-ion such as sodium or calcium: PubChem writes carboplatin as its ligands plus a separate platinum ion, so such an input is kept whole and treated as a metal complex. A structure that matches a library drug only after redrawing, such as another tautomer, is recognised by its standard InChIKey.

The ports are checked against Python RDKit of the same release (2026.03.6). On all 2,197 library drugs, the browser gives the same QED to within 3 × 10⁻¹⁶, the same descriptors, formulae and canonical SMILES, the same Morgan bits and the same PAINS matches; 124 of those drugs match at least one PAINS filter, so the positive case is covered too.

The 3D shape comes from OpenChemLib in a second worker: its conformer generator starts from the most likely torsion angles in its built-in torsion tables and avoids clashes, then the MMFF94s+ force field minimises the result. 3Dmol.js draws it. The map is a principal component analysis of nine standardised descriptors. Your molecule is projected with the stored means, standard deviations and loadings, so it lands in the same fixed space as the library, and a drug already in the library lands exactly on its own point.

How it was built

A Node pipeline downloads every approved molecule that ChEMBL types as a small molecule (max_phase 4) from the ChEMBL web services with timeouts and retries, checks that the release is ChEMBL 37, keeps the 2,197 parent molecules that have a structure, describes each with the same RDKit.js code the page uses, fits the PCA and writes one 250 KB JSON file (about 80 KB compressed) with ChEMBL ids, names, SMILES, ChEMBL approval year (its first_approval field), route flags and map coordinates. For the route flags it also downloads every ChEMBL record flagged oral, parenteral or topical, so a flag that ChEMBL sets only on a salt or other form of a drug still counts for the drug. The pipeline also runs the page's own analysis and similarity code on the opening example and saves the result (about 27 KB compressed); a test checks that it matches what the page computes. A Python script with RDKit produces the reference values the tests compare against and converts RDKit's PAINS definitions.

Fingerprints are not shipped. Storing them would add about 110 KB compressed, while computing them from the SMILES takes a second or two in the background worker once RDKit has loaded (the page shows the time it took on your device), and guarantees they match the query fingerprint bit for bit. Names are resolved against the library first and then with PubChem's PUG REST service, using its current SMILES property (requests for the older IsomericSMILES and CanonicalSMILES names now return SMILES and ConnectivitySMILES).

Parameter defaults

Engine

RDKit.js 2026.03.6 (WebAssembly) in a Web Worker; OpenChemLib in a second worker for the 3D conformer.

Fingerprints

Morgan, radius 2, 2,048 bits; similarity is the Tanimoto coefficient on those bits.

Similar drugs

The ten most similar approved drugs in the shipped library.

Rules of thumb

Lipinski's rule of 5 allows one violation; Veber's rules allow none.

QED and PAINS

QED as RDKit's QED.py (same patterns, parameters and weights); PAINS from the 480 filters of RDKit's FilterCatalog.

Map

A principal component analysis of nine standardised descriptors over the 2,197 approved small-molecule drugs of ChEMBL 37.

Input limits

SMILES up to 2,000 characters and structures up to 500 heavy atoms; the editor draws up to 150 atoms.

Data sources, licences and versions

Times on screen are local time with the UTC offset; times in exports are UTC. Dates with no time are shown at the precision the source gave them.

Today the app has no file export. What you can take away is the address, whose ?smiles= or ?name= link reopens the same molecule, and the canonical SMILES, which the Copy SMILES button puts on your clipboard. The library's licence (ChEMBL, CC BY-SA 3.0) and release are shown beside the results.

Release gate: before the app leaves the preview, its exports and printed reports will carry the data versions, parameters, limits and a timestamp needed to reproduce a result, as the regulatory position says. This section will describe each format as it ships.

Each module below is a set of automated tests in the app's source code, with the expected results written into the tests. They check the app's logic against recorded inputs and stored reference records; none of them calls a live service.

src/components/tools/molecule-workbench/fixtures/rdkit-reference.json

Expected outputs from Python RDKit 2026.03.6 (descriptors, QED, PAINS matches and Morgan bits) for reference molecules, written by pipeline/molecule-workbench/rdkit_reference.py. The browser's results are compared with it.

src/components/tools/molecule-workbench/chem.test.ts

The fixture and RDKit.js are the same RDKit release; descriptors, QED, PAINS and Morgan bits match Python RDKit; invalid and oversized input is refused; salts, metal complexes and unspecified stereocentres are handled; a tautomer of a library drug has the same standard InChIKey.

src/components/tools/molecule-workbench/logic.test.ts

Tanimoto on packed bits equals set Tanimoto in Python RDKit, nearest-neighbour ranking, the PCA and the stored map projection, Lipinski and Veber, deep links, PubChem property names, the library file's licence and release, route flags and formula ordering.

src/components/tools/molecule-workbench/default-example.test.ts

The precomputed opening example (imatinib) equals what the worker computes from the current library.

Release gate: before the app leaves the preview, each release will publish a verification dataset with expected outputs on this page, so you can run your own checks. None has been published yet.

The app is a set of static files that runs in your browser, with no account and no server of ours in the data path. The privacy statement, the hosts it contacts and the headers its workspace is served with are generated from its registry entry, the same record that sets the browser's policy.

What leaves your browser

Hosts the app contacts

• pubchem.ncbi.nlm.nih.gov

  Operator

  NCBI, U.S. National Library of Medicine

  What for

  Looks up a structure by name when the name is not in the app's library of approved drugs.

  What it receives

  the name you typed

  When

  on lookup

Headers the workspace is served with