MolecularAI/REINVENT4

Leverages reinforcement learning with multi-component scoring to optimize molecules toward user-defined property profiles, with transfer learning support for biasing generation toward reference compounds. Built on PyTorch with a modular plugin architecture for custom scoring components, enabling seamless integration of domain-specific metrics without modifying core code. Supports multiple GPU backends (NVIDIA CUDA, AMD ROCm, Intel XPU) and CPU execution, configured via TOML/JSON/YAML files with pre-trained priors available on Zenodo.

707 stars. Actively maintained with 7 commits in the last 30 days.