REINVENT4 Documentation

Command-line tool for generative molecular design using RNNs and Transformers.

Overview

REINVENT4 generates SMILES strings and optimizes them against user-defined scoring functions. All behavior is controlled via a TOML configuration file.

Overview Taken from Loeffler et al., J. Cheminformatics (2024) under a Creative Commons Attribution 4.0 International License.

Generators

Four generators are available: Reinvent (de novo), LibInvent (scaffold decoration), LinkInvent (fragment linking), and Mol2Mol (molecule optimization). See Core Concepts for details.

Model Adaptation

Pre-trained prior models for all generators are available on Zenodo. These can be adapted to a specific task in two ways, which can be combined:

  • Transfer Learning (TL): fine-tune the prior on a focused set of SMILES (e.g. known actives for a target).

  • Reinforcement / Curriculum Learning (RL/CL): iteratively bias the agent toward molecules that score well on a user-defined scoring function. Multiple RL stages can be chained (curriculum learning).

Configuration

All run modes (sampling, TL, RL/CL, scoring) are configured through a single TOML file. JSON is also accepted.

Installation

  1. Clone the repository:

    git clone git@github.com:MolecularAI/REINVENT4.git --depth 1

  2. Create a environment (Python 3.10+):

    conda create --name reinvent4 python=3.10
conda activate reinvent4

  3. Install dependencies: Run the installation script with your processor type (e.g., cu126, rocm6.4, xpu, cpu, or mac):

    python install.py cpu  # Replace 'cpu' with your target platform

  4. Verify:

    reinvent --help

Concepts

Tutorials