Molecule design by reinforcement learning and TOML configs

What it does

REINVENT4 is a command-line generative chemistry toolkit aimed at early-stage drug discovery. It trains models with reinforcement and transfer learning to propose novel compounds that satisfy a multi-component property profile. The software handles de novo design, scaffold hopping, R-group replacement, linker design, and general molecule optimization.

The interesting bit

The scoring layer is built around Python native namespace packages, so you can add custom scoring components by dropping a file on your PYTHONPATH without patching the core repository. Configuration is declarative and TOML-first, with JSON and YAML tolerated as alternatives.

Key highlights

• Unifies de novo design, scaffold hopping, R-group replacement, linker design, and optimization in a single pipeline.
• Extensible scoring via a plugin architecture that uses native namespace packages—no core code changes required.
• Linux is fully validated; Windows and macOS are supported with limited testing and reduced hardware options.
• A GPU is strongly recommended for transfer learning and model training, though pure reinforcement-learning runs lean heavily on CPU-based scoring.
• Public prior models, example configurations, and jupytext notebooks are provided.

Caveats

• Windows and macOS support is explicitly partial and limited; Linux is the only fully validated platform.
• Some optional scoring components and the full test suite require a proprietary OpenEye license.

Verdict

Medicinal chemists and computational drug-discovery teams who want a programmable, extensible generative pipeline will find a solid workhorse here. If you expect a point-and-click GUI or a one-button solution, this is not your shortcut.