CASCADE brings together molecular sketching, 3D conformation, ADMET profiling, graph neural network inference, and protein-ligand docking into a single browser-based environment.
Capabilities
Each module is independently useful and collectively forms a coherent drug-discovery research environment.
A fully custom HTML5 canvas-based sketcher with draw, erase, move, zoom, pan, undo/redo, configurable bond orders, and 23 ring and scaffold templates — no third-party sketcher dependency.
ETKDGv3 with MMFF94/MMFF94s/UFF force-field minimization, interactive 3Dmol.js viewer, and SDF, XYZ, MOL2 export.
PyTorch MPNN producing per-atom and molecule-level property estimates from atom types and pairwise 3D distances. SHA-256 weight validation at load time.
Lipinski Ro5 and Veber filters, live molecular descriptors, ADMET-AI integration, and a heuristic design score with RDKit fallback.
Local SQLite/FTS5 drug search with Morgan-fingerprint Tanimoto similarity comparison — instant lookup without external API calls.
PDB fetch or upload, ligand selection, geometry-based interaction profiling, and optional GNINA CNN-based docking with Ollama NLP commands for the 3D viewer.
Per-user and per-job directory isolation with UUID path canonicalization, SHA-256 GNN weight validation with fail-closed loading, session security with rate-limited login, and isolated Docker service ports. Health probes at /health/live and /health/ready.
Workflow
Every downstream analysis is anchored to a single canonical SMILES string, ensuring reproducible results across all modules.
Draw on the canvas, paste a SMILES, or type a drug name. Resolved to an RDKit molecule and canonical SMILES string — locally or via PubChem fallback.
RDKit validates the structure and generates clean 2D coordinates with proper layout and bond representation.
Explicit hydrogens are added, ETKDGv3 generates a starting 3D conformer, and MMFF94/MMFF94s/UFF relaxes the geometry.
Molecular descriptors, Lipinski Ro5, Veber filters, ADMET-AI predictions, and a heuristic design score are computed in sequence.
The MPNN processes atom types and pairwise 3D distances, producing per-atom and molecule-level property estimates.
Tanimoto similarity search over 2.89M compounds; optional PDB fetch and geometry-based protein-ligand contact profiling.
CASCADE is a research tool. GNN predictions, docking poses, ADMET profiles, and heuristic scores are computational approximations and do not substitute for in vitro or clinical validation.
Technology
Get Started
Desktop installers bundle Python, RDKit, PyTorch, and Flask. Docker is supported on all platforms and is recommended for custom AI configuration.
Native installer with Program Files setup, Start Menu shortcuts, automatic WebView2 configuration, in-place upgrades, and an uninstaller.
Apple Silicon DMG with standard drag-to-Applications layout. Intel Mac users should use the Docker installation path.
One command installs and starts everything. Works on Linux, macOS (Intel and Apple Silicon), and Windows. Recommended for Ollama and custom AI configuration.
git clone https://github.com/prawinin/CASCADE.git
cd CASCADE
python compose_up.pycompose_up.pyAudit & Performance
A comprehensive engineering and security audit covers the full implementation stack. Key findings are documented inline below.
The model loader validates a SHA-256 checksum against the packaged manifest before any inference runs. A mismatched or missing checkpoint aborts startup rather than proceeding with untrusted weights.
Every computation runs inside a UUID-keyed job workspace under jobs/<user-id>/<job-id>/. All paths are canonicalized before use to prevent directory traversal attacks.
Login endpoints are rate-limited server-side. Sessions carry secure flags and task ownership is enforced at every Celery status and cancel route — users can only read or cancel their own jobs.
The 2.89M compound database, fingerprint index, and GNN weights are each verified against a SHA-256 manifest in packaging/assets.json before use. Downloads resume on interruption.
MODEL_DEVICE=cpu is the safe default. GPU execution requires an explicit opt-in via the environment file. This eliminates accidental CUDA exposure in constrained deployments.
/health/live reports liveness; /health/ready verifies Redis connectivity, SQLite availability, writable job storage, and the model checkpoint before accepting traffic.
Provenance
CASCADE was developed by Prawin under the supervision of Dr. Rajiv K. Kar (Assistant Professor, Jyoti and Bhupat Mehta School of Health Sciences & Technology, IIT Guwahati).