GPU-Accelerated Molecular Docking for macOS
A native macOS application for structure-based drug discovery.
Built from the ground up for Apple Silicon with Metal compute shaders.
Requirements — macOS 26.0 or later, Apple Silicon (M1 or newer). Click the button above to download the
.dmg, open it, and drag Druse into Applications.
Note — Druse is currently distributed unsigned. On first launch, macOS Gatekeeper will block the app. Right-click (or Control-click)
Druse.appand select Open, then click Open in the confirmation dialog. Alternatively, runxattr -cr /path/to/Druse.appin Terminal.
Build from Source
Prerequisites
- macOS 26.0+, Apple Silicon (M1 or newer)
- Xcode 26.0+ (with Command Line Tools)
- Homebrew
- XcodeGen 2.35+
1. Install dependencies
brew install rdkit boost eigen nanoflann tbb
brew install xcodegenOptional (for GFN2-xTB quantum refinement):
brew install xtb2. Clone the repository
git clone https://github.com/Vitruves/Druse.git
cd Druse3. Build the C++ core
cd CppCore
mkdir -p build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j$(sysctl -n hw.ncpu)
cd ../..OpenMM is fetched and built automatically by CMake. The first run can take several minutes.
4. Generate the Xcode project and build the app
xcodegen generate
xcodebuild -project Druse.xcodeproj -scheme Druse -configuration Release buildThe built Druse.app lands in Xcode's DerivedData directory; the build log prints the exact path. To launch it directly:
open "$(xcodebuild -project Druse.xcodeproj -scheme Druse -configuration Release -showBuildSettings 2>/dev/null | awk -F' = ' '/ BUILT_PRODUCTS_DIR /{print $2}')/Druse.app"Running the benchmark
The CASF-2016 benchmark suite requires the PDBbind refined-set + CASF-2016 core set data (not redistributed with the repo) and a manifest generated from them.
- Obtain the data. Place the extracted archives under
Benchmark/data/so the layout is:
Benchmark/data/
├── CASF-2016/ # CASF-2016 core set (CoreSet.dat + complexes)
└── refined-set/ # PDBbind refined-set (one folder per PDB ID)
- Install the Python dependencies (Python 3.10+ recommended):
pip install -r Benchmark/requirements.txt- Build the manifest from the data above:
python Benchmark/prepare.py
# writes Benchmark/manifests/casf_manifest.json- Run a benchmark. Each invocation drives
xcodebuild testagainst theDruseTeststarget — a GUI window will open during the run.
# Quick smoke test: redock 10 complexes with Vina scoring
python Benchmark/run_benchmark.py --casf --scoring vina --preset standard --quick 10
# Compare all scoring functions on the full core set
python Benchmark/run_benchmark.py --casf --scoring vina,drusina,druseaf,pignet2
# All options
python Benchmark/run_benchmark.py --helpResults land in Benchmark/results/ as timestamped JSON files. If the run reports PASSED in seconds with no per-complex output, the manifest is missing — re-run step 3.
Most docking software was designed for Linux clusters and command-line workflows. Druse takes a different approach — a fully native macOS application that puts the entire docking pipeline in a single window, accelerated by Metal on the GPU you already have.
| Native & Fast | Built in Swift and Metal. No Python runtimes, no Docker, no X11. Just a .app that launches instantly. |
| GPU Everything | 20+ Metal compute kernels — from genetic algorithm search to neural network scoring to surface rendering. All on your M-series chip. |
| Unified Memory | Apple Silicon's shared CPU/GPU memory means zero-copy data transfer between the docking engine, ML models, and 3D renderer. |
| Interactive | Real-time 3D visualization with impostor rendering, Connolly surfaces, and ribbon diagrams — all at retina resolution. |
Druse covers the full structure-based drug discovery workflow in a single application.
- Fetch from RCSB — Enter a PDB ID and Druse downloads the structure directly
- Format support — PDB, mmCIF, SDF (V2000/V3000), MOL2, SMILES
- Protein preparation — Automated pipeline: water removal, alternate conformer selection, missing atom reconstruction, polar hydrogen addition, H-bond network optimization (Asn/Gln flips, His tautomers), sidechain packing (FASPR with Dunbrack rotamers), and energy minimization
- Protonation — pH-dependent protonation with table lookup or optional GFN2-xTB quantum pKa prediction
- Partial charges — Four methods: EEM (GPU-accelerated), QEq, Gasteiger, and RDKit template matching
- Alpha-sphere probing — Geometric pocket detection with DBSCAN clustering and 26-direction buriedness scoring
- ML-enhanced detection — Optional neural network pocket predictor combined with geometric scoring
- Ligand-guided — Auto-center from a co-crystallized ligand, or define a custom box manually
- Pocket metrics — Volume (ų), buriedness score, druggability rating, and resident residue identification
- Genetic Algorithm — Population-based search with adaptive mutation, crossover, and iterated local search — all on GPU
- Fragment-Based Docking — Hierarchical fragment growth with anchor placement and beam search pruning
- Parallel Tempering — Replica exchange Monte Carlo for enhanced sampling of flexible ligands
- Diffusion-Guided Docking — Reverse diffusion process with DruseAF attention gradient guidance
- Flexible Docking — Simultaneous optimization of ligand pose and sidechain rotamers (3–6 residues)
- Pharmacophore Constraints — H-bond, hydrophobic, aromatic, salt bridge, and metal coordination constraints with configurable strength
- Auto-Tuning — Automatically adapts search parameters to protein size, pocket shape, and ligand flexibility
- Analytical Gradients — SIMD-cooperative local search, faster than numerical finite differences
Four scoring functions, each with different strengths:
| Scoring Function | Type | Description |
|---|---|---|
| Vina | Empirical | Gaussian + repulsion + hydrophobic + H-bond + torsion penalty |
| Drusina | Extended empirical | Vina baseline plus electrostatics, salt bridges, π-π stacking, π-cation, halogen bonds, metal coordination, CH-π, amide-π, chalcogen bonds, and torsion strain |
| DruseAF v4 | Neural network | SE(3)-equivariant pairwise geometric network for pKd prediction |
| PIGNet2 | Physics-informed GNN | Graph neural network with physics-based interaction constraints |
- Pose clustering — GPU-accelerated pairwise RMSD with hierarchical clustering and consensus pose extraction
- Energy decomposition — Per-term breakdown across all scoring components
- Interaction detection — 10+ interaction types with color-coded 3D visualization and 2D interaction diagrams
- Explicit atom rescoring — Top clusters re-scored against explicit receptor atoms with basin hopping refinement
- Strain filtering — MMFF94 torsion strain penalty to flag unrealistic conformations
- GFN2-xTB refinement — Optional semi-empirical QM post-docking optimization with D4 dispersion and ALPB solvation
Batch-dock up to 100,000 molecules with shared grid reuse and parallel 3D generation.
- Pre-filtering — Rotatable bond cutoff, rapid pre-scoring for fast rejection
- ADMET filtering — Lipinski Rule of Five, Veber rules
- Multiple scoring functions — Vina, Drusina, DruseAF v4, or PIGNet2 as the primary scorer
- Ranked export — CSV with full metrics and SDF with posed geometries
Generate and evaluate analogs directly inside Druse.
- Analog generation — 18+ curated substitution rules: halogen swaps, alkyl extensions, heteroatom substitutions, aromatic replacements, functional group interchanges, ring size modifications
- Property sliders — Dial in polarity, rigidity, lipophilicity, and size to bias generation toward desired property space
- Mini-docking — Sub-second per-analog docking with RMSD tracking against the parent compound
- ADMET gating — Real-time Lipinski and Veber checks on every analog
- Trade-off analysis — Binding affinity vs. ADMET property landscape with Pareto frontier identification
Real-time Metal-rendered 3D molecular graphics, triple-buffered for smooth interaction.
- Rendering modes — Ball & stick, space filling, wireframe, ribbon (helix/sheet/coil)
- Surfaces — Connolly (solvent-accessible) and Gaussian blob isosurfaces with electrostatic potential, hydrophobicity, or pharmacophore coloring
- Interactions — Color-coded lines for hydrogen bonds, salt bridges, π-stacking, halogen bonds, metal coordination, and more
- GPU picking — Click any atom or residue directly in the 3D view for instant selection and inspection
- Ghost ligands — Translucent overlay of docking poses for comparison
- Z-slab clipping — Slice through the structure to focus on the binding site interior
A built-in molecular database for organizing your compounds.
- Multi-format import — SDF, MOL2, PDB files, co-crystallized ligands from loaded structures, or direct SMILES input
- Chemical form enumeration — Tautomers, protomers, and cross-enumeration with Boltzmann population weighting
- Conformer generation — RDKit 3D embedding with energy-ranked conformer selection
- Batch preparation — Hydrogen addition, charge calculation, and minimization across the library
- Search & filter — By name, SMILES, molecular weight, rotatable bonds, H-bond donors/acceptors
New to molecular docking? Druse includes a fully interactive guided walkthrough that docks Nafamostat into Trypsin (PDB: 3PTB) — from protein fetch to scored poses — with step-by-step narration. Just click Demo on the welcome screen.
Druse runs 20+ specialized Metal compute kernels on Apple Silicon:
- Genetic algorithm evolution and selection
- Analytical gradient local search (SIMD-cooperative)
- Grid map generation (steric, hydrophobic, H-bond, affinity)
- Neural network inference (DruseAF v4, PIGNet2)
- EEM partial charge calculation
- Connolly surface generation
- Pairwise RMSD computation
- Fragment growth and diffusion denoising
- Flexible sidechain scoring
- Impostor sphere/cylinder rendering with depth-correct post-processing
Half-precision grid storage, shared memory tiling, and on-demand rendering keep memory usage low and battery life long.
Druse — Molecular docking, native on your Mac.
Druse is released under the Apache License 2.0. See NOTICE for third-party attributions and academic citation details.
Copyright © 2026 Johan H.G. Natter