yangshun2005/openscience-cn/molecular-rag

Molecular RAG (Retrieval-Augmented Generation)

Overview

LLMs hallucinate molecular properties. This skill grounds predictions by retrieving structurally similar compounds with experimentally measured properties from ChEMBL and ZINC. When the agent says "this compound should have good hERG safety," it can now check what happened with similar compounds in real assays.

Based on:

• MolRAG (Xian et al., 2025, ACL): RAG for molecular property prediction — retrieves similar compounds to ground LLM predictions

When to Use This Skill

• Before property prediction: Retrieve analogs with known properties for context
• Lead optimization: Find what modifications worked for similar scaffolds
• Novelty assessment: Check if your generated molecule is truly novel or already known
• SAR grounding: Ground structure-activity reasoning in experimental data

Do NOT use this skill for:

• Bulk database queries (use chembl-database or pubchem-database directly)
• De novo generation (use denovo-design)

Related Skills

• chembl-database: Direct ChEMBL API access
• pubchem-database: PubChem compound lookup
• zinc-database: ZINC compound search
• admet-reasoning: Interpret properties of retrieved analogs

Installation

pip install rdkit-pypi requests pandas

Core Workflows

1. Find Similar Compounds with Known Properties

python scripts/retrieve_analogs.py \
    --smiles "c1ccc(NC(=O)c2ccccc2Cl)cc1" \
    --similarity-threshold 0.6 \
    --max-results 20 \
    --output analogs.json

2. Target-Specific Analog Search

python scripts/retrieve_analogs.py \
    --smiles "c1ccc(NC(=O)c2ccccc2Cl)cc1" \
    --target CHEMBL25 \
    --output target_analogs.json

3. SAR Context for Optimization

python scripts/retrieve_analogs.py \
    --smiles "c1ccc(NC(=O)c2ccccc2Cl)cc1" \
    --include-activities \
    --output sar_context.json

Script Reference