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

Lead Optimization Agent

Try the live demo on HuggingFace Spaces

---

Problem

Drug lead optimization is one of the most expensive and slow phases of drug discovery. A medicinal chemist starts with a promising molecule and must iteratively explore chemical space — proposing analogues, synthesizing them, measuring ADMET properties (absorption, distribution, metabolism, excretion, toxicity), and deciding whether to continue or pivot.

This process is:

• Manual and slow — each analogue requires expert chemical intuition to design
• Disconnected — scoring tools, structure editors, and decision notes live in separate places
• Opaque — it's hard to audit why a particular direction was pursued across dozens of iterations
• Expensive to iterate — without fast local scoring, every feedback loop requires wet-lab or commercial prediction services

There is no lightweight tool that closes the loop between AI-assisted structural reasoning, fast local property scoring, and scientist oversight — all in one interface.

---

Solution

Lead Optimization Agent is an AI-native iterative sandbox for medicinal chemistry exploration. Given a starting molecule and a plain-English optimization goal, the agent proposes one structural change at a time, scores each candidate locally using RDKit, and surfaces results for the scientist to review, accept, or redirect.

The scientist stays in control. The agent accelerates the ideation and scoring cycle.

---

How It Works

Agentic loop

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

Two distinct components

Agent (Claude via Anthropic API) — handles the chemistry reasoning layer:

• reads the current molecule and property scores
• proposes the next structural edit and explains the logic
• does not perform property calculations

Local scoring (RDKit in agent_utils.py) — deterministic, fast, no API call:

• QED (drug-likeness)
• Lipinski rule-of-five summary
• BBB permeability heuristic
• CNS MPO score
• Aqueous solubility estimate
• GI absorption heuristic
• Structural alerts
• Synthetic accessibility (SA) heuristic

This separation means chemistry scoring is reproducible and free; the LLM is used only for reasoning about what change to try next.

UI

The Streamlit app presents each iteration as a candidate card with:

• 2D structure with the changed region highlighted
• plain-English change summary
• per-property deltas vs. the starting molecule
• Candidate Journey tab for attempt-by-attempt review
• Performance Overview tab with trajectory plots and a start-vs-best radar chart

---

Quick Start

git clone https://github.com/mondalsou/lead_optimization_agent.git
cd lead_optimization_agent
pip install -r requirements.txt
# if RDKit fails via pip: conda install -c conda-forge rdkit
export ANTHROPIC_API_KEY=sk-ant-...
streamlit run app.py

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

---

Project Structure

lead_optimization_agent/
├── app.py              # Streamlit UI + agent orchestration
├── agent_utils.py      # RDKit scoring, SMILES validation, helpers
├── requirements.txt
├── candidates.json
├── saved_runs/         # Local run persistence (JSON)
└── notebooks/
    ├── 01_admet_tool.ipynb
    ├── 02_agent_loop.ipynb
    └── 03_visualization.ipynb

---

Preset Scenarios

• Atenolol → Brain Penetration
• Aspirin → CNS Drug Profile
• Ibuprofen → Aqueous Solubility
• Custom molecule

---

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 before serious decisions

---

Author

Sourav Mondal — GitHub · LinkedIn