AI Agent for Virtual Screening with Smina

Overview

This tutorial demonstrates how to build an AI Agent for Virtual Screening using Ollama with local LLMs. The agent acts as a molecular docking expert that can autonomously perform virtual screening tasks using natural language commands.

Tutorial: ADN_T018_AI_Agent_Virtual_Screening_with_Smina.ipynb

What You’ll Learn

1. AI Agent Architecture: Understand how to build an AI agent with tools
2. Ollama Integration: Use Ollama with gpt-oss:latest for local LLM inference
3. Molecular Docking Expert: Create a domain-specific AI persona
4. Tool Integration: Connect computational tools (Smina, OpenBabel) to the agent
5. Function Calling: Implement autonomous tool usage by the AI
6. Web Deployment: Convert to a Streamlit web application

Key Components

🤖 AI Brain

• LLM: Ollama with gpt-oss:latest model
• Framework: Function calling for tool orchestration
• Persona: Molecular docking expert with deep domain knowledge

🛠️ Tools

1. OpenBabel: Protein and ligand preparation (hydrogen addition, format conversion)
2. Smina: Molecular docking engine
3. Pandas: Data analysis and pose ranking
4. py3Dmol: 3D visualization of protein-ligand complexes

💻 Web Interface

• Streamlit: Interactive web app for the AI agent
• Features: File upload, chat interface, 3D visualization, results download

Prerequisites

Software Requirements

1. Ollama - Local LLM runtime

   # Download from https://ollama.ai/download
   # For macOS/Linux, install and pull the model:
   ollama pull gpt-oss:latest
   

2. Conda - For package management (recommended)

Python Packages

# Create environment
conda create -n vs_agent python=3.9
conda activate vs_agent

# Install core packages
conda install -c conda-forge rdkit openbabel smina py3dmol
pip install pandas ollama streamlit stmol streamlit-ketcher

Getting Started

1. Run the Jupyter Notebook

cd notebooks
jupyter notebook ADN_T018_AI_Agent_Virtual_Screening_with_Smina.ipynb

Work through the cells sequentially:

• Section 1: Setup Ollama and verify model
• Section 2: Define molecular docking expert persona
• Section 3: Create tool functions
• Section 4: Build the AI agent
• Section 5: Demo the agent with virtual screening tasks
• Section 6: Deploy as Streamlit web app

2. Example Agent Interactions

The agent can understand natural language commands like:

# Example 1: Protein preparation
agent.chat("""
I have a PDB file called 5R82.pdb. Please:
1. Extract the protein and ligand (residue name RZS) separately
2. Add hydrogens to the protein at pH 7.4
3. Tell me about the structure
""")

# Example 2: Virtual screening
agent.chat("""
Please perform virtual screening:
1. Prepare ligands from '../databases/data-test.sdf'
2. Dock them against 5R82_proteinH.pdb using reference ligand 5R82_ligand.pdb
3. Analyze the top 5 compounds and tell me about their binding affinities
""")

3. Deploy as Web App

After completing the notebook, save the required files:

# Save the tool classes to vs_tools.py
# Save the Streamlit code to streamlit_vs_agent.py

# Run the app
streamlit run streamlit_vs_agent.py

Access at http://localhost:8501

Notebook Structure

ADN_T018_AI_Agent_Virtual_Screening_with_Smina.ipynb
├── 1. Introduction & Overview
├── 2. Setup Ollama and Verification
├── 3. Define Molecular Docking Expert Persona
├── 4. Create Tool Functions
│   ├── prepare_protein()
│   ├── prepare_ligands()
│   ├── perform_docking()
│   ├── analyze_results()
│   └── extract_protein_ligand()
├── 5. Build AI Agent with Function Calling
│   ├── MolecularDockingAgent class
│   ├── Tool specifications
│   └── Chat interface
├── 6. Demo: Virtual Screening Workflow
│   ├── Protein preparation
│   ├── Ligand preparation
│   ├── Molecular docking
│   ├── Results analysis
│   └── 3D visualization
├── 7. Convert to Streamlit Web App
│   ├── Complete app code
│   ├── Deployment instructions
│   └── Usage guide
└── 8. Summary and Extensions

Features

✅ What the AI Agent Can Do

• Autonomous Operation: Break down complex tasks into steps
• Smart Tool Selection: Choose appropriate tools based on context
• Expert Insights: Provide domain knowledge about binding modes and interactions
• Interactive Chat: Natural language interface
• Error Handling: Gracefully handle issues and provide helpful feedback
• Result Visualization: Generate 3D views of docking results

🔄 Example Workflow

1. User: “I want to screen a library of compounds against SARS-CoV-2 main protease”
2. Agent:
   • Extracts protein and reference ligand from PDB
   • Prepares protein (adds hydrogens)
   • Prepares ligand library (protonation, format conversion)
   • Sets up and runs docking with optimal parameters
   • Analyzes results and ranks compounds
   • Visualizes top hits
   • Provides expert commentary on binding modes

Advantages Over Traditional Workflows

Extensions and Next Steps

Additional Tools to Add

• AutoDock Vina integration
• ADMET prediction (SwissADME, pkCSM)
• Protein-ligand interaction fingerprints (PLIF)
• Molecular dynamics preparation
• Multi-target docking

Enhanced Analysis

• Binding site identification
• SAR (Structure-Activity Relationship) analysis
• Pharmacophore modeling
• Fragment-based drug design

Advanced Features

• Batch processing of large libraries (100K+ compounds)
• Cloud deployment (AWS, Google Cloud, Azure)
• Database integration (ChEMBL, PubChem)
• Automated report generation
• Email notifications for long jobs

Model Improvements

• Fine-tune on chemistry/docking data
• Use larger models (GPT-4, Claude)
• Implement RAG (Retrieval-Augmented Generation) with scientific literature

Troubleshooting

Common Issues

1. Ollama not found

   # Verify installation
   which ollama
   ollama --version
   

2. Model not available

   # Pull the model
   ollama pull gpt-oss:latest
   
   # List available models
   ollama list
   

3. Smina not found

   # Install via conda
   conda install -c conda-forge smina
   
   # Or download static binary from sourceforge
   

4. OpenBabel issues

   # Reinstall
   conda install -c conda-forge openbabel --force-reinstall
   

Resources

Documentation

• Ollama: https://ollama.ai/
• Smina: https://sourceforge.net/projects/smina/
• OpenBabel: http://openbabel.org/
• RDKit: https://www.rdkit.org/
• Streamlit: https://streamlit.io/

Related Tutorials

• ADN_T013: Molecular Docking with Smina (basic)
• ADN_T017: Virtual Screening with Smina (intermediate)

Papers and References

• Koes et al. (2013). “Lessons Learned in Empirical Scoring with smina”
• Trott & Olson (2010). “AutoDock Vina: improving the speed and accuracy of docking”

Citation

If you use this tutorial in your research, please cite:

Aouidate, A. (2023-2026). AI Agent for Virtual Screening with Smina and Ollama.
Chemoinformatics Tutorials. Ibn Zohr University, Agadir, Morocco.

Contact

For questions or feedback:

• Open an issue on GitHub
• Contact: [Your contact info]

License

This tutorial is released under the [LICENSE] specified in the repository.

Happy Docking! 🧬🚀