RAD
Retrieval Augmented Docking
RAD uses Hierarchical Navigable Small World (HNSW) graphs to accelerate virtual screening workflows. This public service allows researchers to screen a chemical space of 2 billion ZINC-22 lead-like molecules with their own scoring function.
Easily Screen 2 Billion Molecules
Connect to our public HNSW service with your own scoring function and begin virtual screening of 2 billion molecules in minutes
# Define your custom scoring function
def score_molecule(smiles: str) -> float:
return score
# Connect to the public RAD server
from rad.hnsw_service import create_remote_hnsw_service
hnsw_service = create_remote_hnsw_service("https://rad.docking.org")
# Start Screening
from rad.traverser import RADTraverser
traverser = RADTraverser(hnsw_service=hnsw_service, scoring_fn=score_molecule)
traverser.prime()
traverser.traverse(n_workers=8, n_to_score=100000)
The RAD Method
Ultra-large chemical libraries of billions of molecules are pre-organized into Hierarchical Navigable Small World (HNSW) graphs with molecules distributed across multiple layers and connected based on similarity.
User-defined scoring functions guide a traversal through this graph, allowing researchers to quickly find top-scoring molecules while avoiding expensive calculations on a majority of the library.
Hierarchical Organization
Sparse hierarchy allows rapid identification of promising areas of chemical space
Small-World Graphs
Graph edges are based on Tanimoto similarity and chosen to maintain the small-world property, enabling efficient exploration
Best First Traversal
User-defined scoring functions guide graph traversal to quickly find good-scoring molecules while avoiding unpromising areas of chemical space
Distributed Scoring
Lightweight scoring workers coordinate with a single central process to enable massively parallel scoring
Why Choose RAD?
Efficient Screening
RAD can find over 50% of the top-scoring molecules in a library, while doing the expensive scoring calculation for only 1%.
Billion-Scale Validated
Proven performance on billion-scale chemical libraries with traditional molecular docking and modern ML models.
Distributed Computing
Scale from single machines to HPC clusters with Redis-coordinated workers that can run anywhere.
Flexible Infrastructure
Deploy locally, use our hosted public service to avoid setup complexity, or run your own infrastructure.
Get Started & Learn More
Installation & Documentation
Install
git clone --recursive https://github.com/keiserlab/rad.git
cd rad
pip install .
Publications
Original RAD Paper
Read the original paper describing the method and showcasing the performance on the DUDE-Z dataset.
Read PaperBillion-Scale Application
See RAD applied to billion-scale chemical libraries with UCSF DOCK and the ML model Chemprop.
Read Paper