torch-molecule provides a deep learning framework for molecular discovery, targeting researchers and practitioners in chemistry, biology, and materials science. It simplifies the implementation and deployment of molecular AI models for property prediction, inverse design, and representation learning with an intuitive scikit-learn-style API.

How It Works

The package offers a unified interface for various molecular tasks, abstracting away complex model architectures and training procedures. It supports a wide array of predictive, generative, and representation models, including graph neural networks (GNNs) and transformer-based approaches. This design allows users to easily swap models and datasets, facilitating rapid experimentation and benchmarking.

Quick Start & Requirements

• Installation: pip install torch-molecule (v0.1.3) or from source.
• Prerequisites: Python 3.11.7. Some models require torch-scatter (install with CUDA-specific wheels) or transformers.
• Datasets: Includes loaders for QM9, ChEMBL2k, ToxCast, ADMET, and Gasperm.
• Usage: Examples and documentation are available in the examples and tests folders.
• Model Checkpoints: Integration with Hugging Face for saving and loading pretrained models.

Highlighted Details

• Supports a broad range of molecular AI models, including GRIN, BFGNN, SGIR, GREA, Graph DiT, DiGress, GDSS, MolGPT, JTVAE, and various GNNs and Transformers.
• Features an autofit method for automatic hyperparameter tuning.
• Enables seamless saving and loading of trained models to/from Hugging Face repositories.
• Provides access to multiple pre-trained representation models like ChemBERTa and ChemGPT.

Maintenance & Community

The project template was adapted from lwaekfjlk/python-project-template. Contributions and dataset suggestions are welcomed.

Licensing & Compatibility

The repository does not explicitly state a license in the provided README. Users should verify licensing for commercial or closed-source use.

Limitations & Caveats

The README does not specify a license, which may impact commercial adoption. Some advanced models require specific PyTorch extensions like torch-scatter, necessitating careful installation based on CUDA version.