generative-optim

Molecular optimization using generative models

Installation

Clone this repository:

git clone [email protected]:gmmsb-lncc/generative-optim.git  # ssh
cd generative-optim

When using the HierVAE model, create a virtual environment with Python 3.8 (the latest tested compatible version) and install the dependencies:

conda create --prefix ./venv python=3.8  # using conda for python 3.8
conda activate ./venv
python -m pip install -r requirements-hiervae.txt

Usage

First, initialize a new aim repository for tracking experiments (just once):

aim init

Run the optimization script with the desired arguments:

python optim.py --help  # show help

See the runs-example.sh script for an example of how to run the optimization script.

Optimization algorithm and problems

Choose from the available optimization algorithms and problems (see --help for more details). Objectives are defined in the objectives.conf.json file.

Experiment tracking

To visualize experiments using aim UI, run the following command in the terminal:

aim up

Then, open the browser at http://localhost:43800/ to see the experiments.

By default, a checkpoint of the whole population is saved in a .csv file inside the .aim/meta/chunks/{run_hash}/ folder at the end of each generation. The final population of generated molecules is saved at .aim/meta/chunks/{run_hash}/generated_mols.txt.

Citing

Matheus Müller Pereira da Silva, Jaqueline da Silva Angelo, Isabella Alvim Guedes, and Laurent Emmanuel Dardenne. 2024. A Generative Evolutionary Many-Objective Framework: A Case Study in Antimicrobial Agent Design. In Genetic and Evolutionary Computation Conference (GECCO ’24 Companion), July 14–18, 2024, Melbourne, VIC, Australia. ACM, New York, NY, USA, 8 pages. https://doi.org/10.1145/3638530.3664159