title	Lead Optimization Agent
emoji	🧬
colorFrom	green
colorTo	blue
sdk	streamlit
sdk_version	1.35.0
app_file	app.py
pinned	false
short_description	LangChain agent for iterative drug lead optimization with RDKit

Lead Optimization Agent

Try the live demo on HuggingFace Spaces

---

What it does

LangChain agent that iteratively optimizes drug molecules toward a target ADMET profile. Give it a starting SMILES and a plain-English goal — it proposes one structural change per round, scores each candidate locally via RDKit, and surfaces results in a Streamlit UI.

Swap between Claude, GPT-4o, DeepSeek V4 Flash, Gemini, or Llama by changing one dropdown. No model-specific code.

---

Problem

Drug lead optimization is slow and expensive. A medicinal chemist starts with a promising molecule and must iteratively explore chemical space — proposing analogues, scoring ADMET properties, and deciding whether to continue or pivot. Each feedback loop requires expert chemical intuition and either wet-lab measurements or commercial prediction services.

This project closes that loop: AI reasoning + fast local scoring + scientist oversight, in one interface.

---

How It Works

Agentic loop

Scientist defines goal (SMILES + brief)
         │
         ▼
  Agent proposes structural edit  ←──────────────────┐
         │                                            │
         ▼                                            │
  RDKit scores candidate locally                      │
  (QED, BBB, CNS MPO, solubility, alerts)            │
         │                                            │
         ▼                                            │
  Attempt logged with rationale + property delta      │
         │                                            │
         ▼                                            │
  Scientist reviews → accept / redirect / stop ───────┘
         │
         ▼
  Best candidate surfaced with full audit trail

Stack

Layer	What
LLM routing	LangChain + OpenRouter — any tool-calling model
Agent loop	create_tool_calling_agent + AgentExecutor
Chemistry scoring	RDKit (local, instant, free) via agent_utils.py
UI	Streamlit with live incremental updates

Tools the agent has

Tool	What it does
validate_smiles	RDKit SMILES validation before analysis
analyze_molecule	Full ADMET profile — QED, BBB, CNS MPO, solubility, Lipinski, alerts
compare_candidates	Side-by-side scoring of multiple molecules

Scoring is deterministic and local — no API call for chemistry. The LLM is used only for structural reasoning.

---

Quick Start

git clone https://github.com/mondalsou/lead-optimization-agent.git
cd lead-optimization-agent
pip install -r requirements.txt
# RDKit via pip may fail — use conda if so:
# conda install -c conda-forge rdkit

cp .env.example .env
# add your OpenRouter key to .env:
# OPENROUTER_API_KEY=sk-or-...

streamlit run app.py

Open http://localhost:8501, pick a preset or paste your own SMILES, write the optimization brief, pick a model, and click Run Optimisation.

---

Supported Models (via OpenRouter)

Model	Slug
Claude Sonnet 4.6	anthropic/claude-sonnet-4.6
Claude Haiku 4.5	anthropic/claude-haiku-4.5
Claude Opus 4.6	anthropic/claude-opus-4.6
DeepSeek V4 Flash	deepseek/deepseek-v4-flash
GPT-4o	openai/gpt-4o
Gemini 2.0 Flash	google/gemini-2.0-flash-001
Llama 3.3 70B	meta-llama/llama-3.3-70b-instruct

---

Project Structure

lead-optimization-agent/
├── app.py                  # Streamlit UI + LangChain agent
├── agent_utils.py          # RDKit scoring, SMILES validation
├── requirements.txt
├── .env.example            # Copy to .env and add your key
├── saved_runs/             # Local run persistence (JSON)
├── hermes_tools/           # Hermes agent integration
│   ├── lead_opt.py
│   └── chemistry_skill.md
├── hermes_setup.sh         # One-shot Hermes setup script
└── notebooks/
    ├── 01_admet_tool.ipynb         # Explore the ADMET scorer
    ├── 02_agent_loop.ipynb         # Anthropic SDK agent loop
    ├── 03_visualization.ipynb      # Optimization trajectory plots
    └── 04_langchain_openrouter.ipynb  # LangChain + OpenRouter version

---

Preset Scenarios

• Atenolol → Brain Penetration — reduce polarity, improve BBB / CNS MPO
• Aspirin → CNS Drug Profile — replace carboxylic acid liability
• Ibuprofen → Aqueous Solubility — lower cLogP, add polar groups
• Custom molecule — paste any SMILES

---

Limitations

• Heuristic prototyping tool, not a validated drug-discovery platform
• Property outputs are local approximations, not experimental measurements
• Agent suggestions should be reviewed by a domain expert

---

Author

Sourav Mondal — GitHub · LinkedIn