Inversion-based Latent Bayesian Optimization (InvBO)

Official PyTorch implementation of the "Inversion-based Latent Bayesian Optimization". (NeurIPS 2024)

Jaewon Chu*, Jinyoung Park*, Seunghun Lee, Hyunwoo J. Kim†.

Setup

• Clone repository

git clone https://github.com/mlvlab/InvBO.git
cd InvBO

• Install Environment

conda env create -f invbo.yml
conda activate invbo
pip install molsets==0.3.1 --no-deps

Run Experiments

This repository provides InvBO applied to CoBO [Lee et al., NeurIPS 2023] for small budget setting.

python exec.py --cuda 0 --task_id [TASK]

Since we predefined the coefficients for VAE loss terms in exec.py provided by CoBO, the available tasks for [TASK] are:

task_id	Task Name
med2	Median molecules 2
pdop	Perindopril MPO
osmb	Osimertinib MPO
adip	Amlodipine MPO
zale	Zaleplon MPO
valt	Valsartan SMARTS
rano	Ranolazine MPO

However, we can also run on the remaining Guacamol tasks when we define coefficients for VAE loss terms in exec.py:

task_id	Task Name
med1	Median molecules 1
siga	Sitagliptin MPO
dhop	Deco Hop
shop	Scaffold Hop
fexo	Fexofenadine MPO

Weights and Biases (wandb) tracking

You can track the optimization process using the wandb library.

You can use wandb tracking by simply setting '--track_with_wandb', 'True' and '--wandb_entity', 'YOUR ENTITRY' in exec.py.

Acknowledgements

This repository is based on CoBO.