Saturn: Sample-efficient Generative Molecular Design using Memory Manipulation

<img src="saturn.jpeg" alt="Saturn Logo" width="300"/>

Saturn is a language model based molecular generative design framework that is focused on sample-efficient de novo small molecule design.

In the experimental_reproduction sub-folder, prepared files and checkpoint models are provided to reproduce the experiments. There is also a Jupyter notebook to construct your own configuration files to run Saturn.

Git Hash Code Versions

• Saturn Pre-print: fee0179
• TANGO Constrained Synthesizability Pre-print: de5cd7f
• Steerable and Granular Synthesizability Control Pre-print: 468b1f4

Installation

1. Install Conda

2. Clone this Git repository

3. Open terminal and install the saturn environment:

    $ source setup.sh
   

Potential Installation Issues

• GLIBCXX_3.4.29 version not found - thank you to @PatWalters for flagging this and solving via:

    $ conda uninstall openbabel 
    $ conda install gcc_linux-64
    $ conda install gxx_linux-64
    $ conda install -c conda-forge openbabel
  

• causal-conv1d and mamba-ssm installation error - see Issue 1 - thank you to @surendraphd for sharing their solution.

System Requirements

• Python 3.10
• Cuda-enabled GPU (CPU-only works but runs times will be much slower)
• Tested on Linux

Acknowledgements

The Mamba architecture code was adapted from the following sources:

• Official Repository
• Mamba Protein Language Model
• Mamba CPU

References

1. Saturn Pre-print
2. Generating Synthesizable Molecules - Coupling Saturn with Retrosynthesis Models
3. TANGO Constrained Synthesizability Pre-print
4. Steerable and Granular Synthesizability Control Pre-print
5. Augmented Memory
6. Beam Enumeration
7. GraphGA