MolCraftDiffusion：面向计算化学的3D分子生成AI框架

MolCraftDiffusion is an open-source framework using diffusion models for 3D molecular generation, providing end-to-end solutions for computational chemistry applications like catalyst discovery and drug design. It bridges deep learning with molecular design to enable efficient generation of molecules with specific properties.

3D molecular structure determines chemical/biological properties. Traditional design relies on intuition and trial-and-error. Diffusion models (successful in image generation via stepwise denoising) are adapted here to let AI generate novel molecules, addressing the need for intelligent molecular design in catalysis and drug development.

The framework offers an end-to-end workflow:

1. Training/Fine-tuning: Supports from-scratch training and domain-specific fine-tuning on pre-trained models.
2. Prediction Models: Trains models to predict physicochemical properties (e.g., excitation energy, dipole moment) for screening.
3. Guided Generation: Built-in strategies for conditional generation of molecules with target properties.

Key features include:

• Curriculum Learning: Progressive training from simple (skeleton) to complex (side chains) structures.
• Attribute-Guided Generation: Generates molecules with specified properties (e.g., HOMO-LUMO gap, solubility).
• Inpainting: Fixes molecular parts (active skeleton) and generates variant side chains/substituents.
• Outpainting: Extends molecules with new fragments (scaffold hopping/fragment linking).
• CLI Interface: Enables command-line execution for easy workflow integration.

Installation:

• GPU: pip install molcraftdiffusion[gpu] (with PyTorch GPU links).
• CPU: pip install molcraftdiffusion[cpu] (with CPU-specific links). Optional packages: Data processing (dscribe SOAP descriptors) and analysis tools (xyz2mol, xtb). Pre-trained Models: Available on Hugging Face or project's models/edm_pretrained/ directory. An interactive demo is provided for zero-installation trial.

Applications include:

• Drug Discovery: Generates target-specific candidates to reduce R&D time/cost.
• Catalyst Design: Creates metal complexes with desired coordination environments.
• Materials Science: Generates molecules with target electronic properties (batteries/photovoltaics).
• Chemical Space Exploration: Discovers novel compounds beyond existing databases.

Limitations: Supports small organic molecules only; generated molecules need DFT validation. Future: Expand to large biomolecules, integrate reinforcement learning, link with experimental automation, and develop a GUI.

MolCraftDiffusion advances AI-computational chemistry integration. It encapsulates diffusion model technology into an accessible tool, letting chemists focus on scientific problems. With improvements, it could become a standard molecular design tool for new materials and drugs.