Skip to content

CLAUDE.md

Latest commit

 

History

History
112 lines (86 loc) · 4.05 KB

CLAUDE.md

File metadata and controls

112 lines (86 loc) · 4.05 KB

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

MolAgent is a multi-agent system for AI-driven molecular property prediction in early-stage drug discovery. It provides MCP (Model Context Protocol) servers that wrap the AutoMol ML framework, enabling agentic AI systems to autonomously train predictive models for molecular properties.

Key Commands

Installation

# Clone with submodules (required for AutoMol backend)
git clone --recurse-submodules https://github.com/openanalytics/MolAgent

# Install (creates .venv with Python 3.12)
./install.sh

# Manual installation
pip install uv
uv venv .venv --python 3.12
source .venv/bin/activate
uv pip install -r requirements.txt
uv pip install pytdc rdkit==2024.3.5
uv pip install AutoMol/automol_resources/ AutoMol/automol/

Starting MCP Servers

# From repository root, in separate terminals:
source .venv/bin/activate && cd MCP && uv run mcp_server/automol_data_server.py     # Port 8000
source .venv/bin/activate && cd MCP && uv run mcp_server/automol_model_server.py    # Port 8001

# Or use startup scripts:
./scripts/server_startup/start_data_server.sh
./scripts/server_startup/start_model_server.sh

Testing

# AutoMol unit tests
cd AutoMol/automol && uv run -m unittest discover -cf

# Plugin tests (from automol-tasks-manager/)
uv run pytest tests/

Claude Code Plugin

# Use as Claude Code plugin with train-pipeline and predict skills
claude --plugin-dir ./automol-tasks-manager/

Architecture

MolAgent/
├── AutoMol/                    # ML backend (git submodule)
│   ├── automol/                # Core AutoMol library
│   └── automol_resources/      # Precomputed molecular features
├── MCP/                        # MCP server layer
│   ├── mcp_server/
│   │   ├── automol_model_server.py  # Port 8001 - regression/classification training
│   │   └── automol_data_server.py   # Port 8000 - TDC data retrieval, 3D processing
│   ├── Tools/training_tools.py      # Training functions called by model server
│   ├── GradioMolAgent.py            # Gradio chatbot interface
│   └── agents.py                    # SmolAgents multi-agent orchestration
├── automol-tasks-manager/      # Claude Code plugin
│   ├── skills/train-pipeline/  # Complete SMILES-to-model workflow
│   ├── skills/predict/         # Inference with trained models
│   └── hooks/setup-automol-env.sh  # Exports AUTOMOL_ROOT, PLUGIN_ROOT
├── MolagentFiles/              # Pipeline outputs (run folders, model_registry.json)
└── Data/                       # Sample datasets

MCP Server Tools

Model Server (port 8001):

  • automol_regression_model - Train regression models for continuous properties
  • automol_classification_model - Train classification models for categorical properties

Data Server (port 8000):

  • retrieve_tdc_data - Download datasets from Therapeutic Data Commons
  • retrieve_3d_data - Process SDF files with 3D structures and PDB files

Pipeline State (automol-tasks-manager)

Pipeline runs create isolated folders: MolagentFiles/{dataset}-{props}-{timestamp}/

State uses "outputs" key for file paths:

outputs = state.get("outputs", state.get("files", {}))

Key conventions:

  • SMILES column standardized to Stand_SMILES after preparation
  • Model files: {property}_stackingregmodel.pt or {property}_stackingclfmodel.pt
  • Refitted models have _refitted suffix in filename
  • Merged models combine per-property files to eliminate encoder duplication

Environment Variables

Variable Purpose
AUTOMOL_ROOT Plugin root directory (set by SessionStart hook)
AUTOMOL_VENV Virtual environment name (default: .venv)

Dependencies

Key packages: rdkit==2024.3.5, molfeat, prolif, scikit-learn, torch, fastmcp, smolagents, pytdc

Requires Python 3.8+ (3.12 recommended), uv package manager, and wkhtmltopdf for PDF generation.