SpotGAN: A Reverse-Transformer GAN Generates Scaffold-Constrained Molecules with Property Optimization

Abstract

Generating molecules with a given scaffold is a challenging task in drug-discovery. Scaffolds impose strict constraints on the generation of molecules. Moreover, the order of the simplified molecular-input line-entry system (SMILES) strings changes substantially during sequence expansion. This study presents a scaffold-constrained, property-optimized transformer GAN (SpotGAN) to solve these issues. SpotGAN employs a decoration generator that fills decorations into a given scaffold using a transformer-decoder variant. The discriminator is a transformer-encoder variant with a global receptive field that improves the realism of the generated molecules. The chemical properties are optimized through reinforcement learning (RL), affording molecules with high property scores. Additionally, an extension of SpotGAN, called SpotWGAN, is proposed to optimize and stabilize the training process leveraging the Wasserstein distance and mini-batch discrimination. Experimental results show the usefulness of the proposed model on scaffold-constrained molecular-generation tasks in terms of the drug-likeness, solubility, synthesizability, and bioactivity of the generated molecules( $^1$ 1 Our code is available at: https://github.com/naruto7283/SpotGAN ).