MolDigger is a molecular structure search and clustering tool available as both a PyQt5 desktop app and a browser-based web app. It uses FPSim2 for fingerprint-based similarity search (screening millions of compounds in milliseconds) and RDKit for substructure matching, Butina clustering, 2D depiction, and property calculation.
- Similarity search — Tanimoto, Dice, and Tversky metrics with adjustable min/max threshold range
- Substructure search — SMILES or SMARTS queries with multi-threaded RDKit matching
- Clustering — Butina clustering of search results with adjustable similarity cutoff; cluster ID column added to results table; always uses Morgan ECFP4 for best chemical groupings
- Atom highlighting — MCS highlighted in similarity hits; matched atoms highlighted in substructure hits (optional, uses RDKit default highlight colour)
- GPU acceleration — NVIDIA CUDA via FPSim2's CudaEngine (Tanimoto only)
- Multiple fingerprint types — Morgan/ECFP4, ECFP6, FCFP4, RDKit Topological, MACCS Keys, Atom Pairs, Topological Torsion
- Auto-detects fingerprint type from the loaded database file
- Structure editor — Ketcher launched in browser; drawn structures sent back to the app automatically
- Results table — sortable, with 2D thumbnails, MW, ClogP; right-click to copy SMILES or use hit as new query
- Stop button — cancel any running search mid-way
- Database builder — create FPSim2
.h5databases from SDF or SMILES files within the app - Export — save results to CSV
Full-featured PyQt5 GUI. Run with:
conda activate moldigger
python moldigger.pyBrowser-based interface built with FastAPI and uvicorn. Useful on headless servers or when a native GUI is not available.
conda activate moldigger
python moldigger_web.py
# opens http://localhost:8000 in your browserBoth apps share the same database format and support the same search, clustering, and highlighting features.
- Python 3.10+
- Conda (recommended) or pip
conda create -n moldigger python=3.11
conda activate moldigger
conda install -c conda-forge rdkit
pip install fpsim2 PyQt5 numpy tablespip install fastapi uvicornRequires an NVIDIA GPU with CUDA installed. Match the cupy version to your CUDA installation:
pip install cupy-cuda12x # for CUDA 12.x
# or
pip install cupy-cuda11x # for CUDA 11.xGPU availability is detected automatically at startup. If available, a Use GPU checkbox appears in the search parameters panel.
- Database tab → load an existing
.h5file, or build one from an SDF/SMILES file - Structure Search tab → type a SMILES or SMARTS query (live 2D preview updates as you type)
- Choose Search type (Similarity or Substructure), fingerprint, metric, and threshold range
- Click Search — results appear sorted by score with 2D thumbnails
- Optionally click Apply Clustering above the results table to group hits by structural similarity
- Click Search again while running to stop it
- Go to the Database tab
- Select an input file (SDF or SMILES)
- Choose fingerprint type and output path
- Click Create Database
The app writes a .h5 FPSim2 database and a companion .h5.smiles.json file that stores SMILES strings and molecule names for display in the results table.
Finds molecules with similar fingerprints to the query using a chosen metric:
| Metric | Description |
|---|---|
| Tanimoto | Standard Jaccard similarity — most common in cheminformatics |
| Dice | 2·|A∩B| / (|A|+|B|) — gives higher scores than Tanimoto |
| Tversky | Asymmetric; α=1, β=0 finds larger molecules containing your scaffold |
Set the Min and Max threshold sliders to control the score range returned. Results are colour-coded green (score = 1.00) → yellow → orange → red (low similarity). The MCS (maximum common substructure) between the query and each hit is highlighted in the 2D thumbnail.
Finds all molecules containing the query as a substructure. Accepts:
- SMILES — exact substructure match
- SMARTS — flexible pattern matching, e.g.:
c1ccccc1— any benzene ring[#6]-C(=O)-[#7]— amide bond[F,Cl,Br,I]— any halogen[n;H1]— NH in an aromatic ring
Matched atoms are highlighted in the hit thumbnails. Runs on CPU using all configured worker threads.
After a search, a Cluster toolbar appears above the results table. Clustering is decoupled from search — you can adjust the cutoff and re-cluster without repeating the database search:
- Set Min similarity (0.10–0.90, default 0.40) — molecules with ≥ this similarity will tend to be grouped together
- Click Apply Clustering — a Cluster column appears and results are sorted by cluster ID
- Click Clear to remove clustering and restore score order
Clustering always uses Morgan ECFP4 fingerprints (radius=2, 2048 bits) regardless of the fingerprint type used for the similarity search, as Morgan ECFP4 gives the best chemical groupings for diverse compound sets.
The underlying algorithm is Butina clustering (rdkit.ML.Cluster.Butina), which is a standard single-pass, sphere-exclusion method widely used in cheminformatics.
| Name | FPSim2 type | Notes |
|---|---|---|
| Morgan / ECFP4 | Morgan, radius=2 | Most common for drug-like molecules |
| Morgan / ECFP6 | Morgan, radius=3 | Larger neighbourhood |
| Morgan / FCFP4 | Morgan, radius=2 | Feature-based (pharmacophore-aware) |
| RDKit Topological | RDKit | Path-based |
| MACCS Keys | MACCSKeys | 166-bit, interpretable |
| Atom Pairs | AtomPair | Counts atom-pair types |
| Topological Torsion | TopologicalTorsion | Encodes torsion angles |
The fingerprint type is automatically detected from the loaded .h5 file.
- FPSim2 screens millions of molecules in < 1 second on CPU (multi-threaded)
- GPU mode (CUDA) provides an additional 5–50× speedup for large databases
- Substructure search is parallelised across all CPU workers using
ThreadPoolExecutor - The
.h5database is memory-mapped — loading is near-instantaneous - The results label reports both the search time (FPSim2/RDKit computation) and total time (including 2D rendering)
MolDigger integrates Ketcher (MIT) as a structure editor. Click Draw Structure to open Ketcher in your browser. Draw or paste a structure, then click Use this structure — the SMILES is sent back to MolDigger automatically. You can modify and resubmit without restarting.
Note for WSL2 users: Qt WebEngine (embedded browser) does not work reliably under WSL2 due to OpenGL/GLX limitations. MolDigger automatically detects this and falls back to launching Ketcher in your system browser on a local port (18920). On native Linux or Windows this limitation does not apply.
| Package | Purpose | License |
|---|---|---|
| FPSim2 | Fingerprint similarity search | MIT |
| RDKit | Cheminformatics, substructure search, clustering, depiction | BSD-3-Clause |
| PyQt5 | Desktop GUI framework | GPL v3 / commercial |
| FastAPI | Web app framework | MIT |
| uvicorn | ASGI server for web app | BSD |
| PyTables | HDF5 I/O | BSD |
| NumPy | Array operations | BSD |
| CuPy (optional) | GPU array library for CUDA | MIT |
| Ketcher (optional) | Structure editor | MIT |
MIT