This repository contains the PyTorch implementation of the molecule generation models. Our goal is to create a Neural Networks that generates a molecules that meet multiple property conditions. We implemented a transformer-based model and a contrastive learning-based model for molecular design. We tested each different models using the ChEMBL dataset.

Generated molecular visualization results

Latent space visualization of contrastive learning-based VAE

The result of visualizing the latent space distribution of the last layer of the encoder using 200,000 validation datasets

Comparison of validity, uniquenss, and novelty evaluation metrics

@K means the result of each indicator when K is generated

Model	Validity(↑) @1000	Uniqueness(↑) @1000	Novelty(↑) @1000
Char-RNN
Tranformer-VAE	0.844	0.944	0.902