MedGAN: optimized generative adversarial network with graph convolutional networks for novel molecule design

Bruno Macedo,Inês Ribeiro Vaz,Tiago Taveira Gomes

DOI: https://doi.org/10.1038/s41598-023-50834-6

IF: 4.6

2024-01-13

Scientific Reports

Abstract:Generative Artificial Intelligence can be an important asset in the drug discovery process to meet the demand for novel medicines. This work outlines the optimization and fine-tuning steps of MedGAN, a deep learning model based on Wasserstein Generative Adversarial Networks and Graph Convolutional Networks, developed to generate new quinoline-scaffold molecules from complex molecular graphs, including hyperparameter adjustments and evaluations of drug-likeness attributes such as pharmacokinetics, toxicity, and synthetic accessibility. The best model was capable of generating 25% valid molecules, 62% fully connected, from which 92% were quinolines, 93% were novel, and 95% unique, preserving chirality, atom charge, and favorable drug-like properties while generating 4831 novel quinolines. These results provide valuable insights into how activation functions, optimizers, learning rates, neuron units, molecule size and constitution, and scaffold structure affect the performance of generative models and their potential to create new molecular structures, enhancing deep learning applications in computational drug design.

multidisciplinary sciences

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve

This paper aims to design a new model named MedGAN by optimizing the combination of Generative Adversarial Networks (GAN) and Graph Convolutional Networks (GCN) to generate new quinoline scaffold molecules. Specifically, the research objectives include:

1. **Optimizing Model Parameters**: Optimize the MedGAN model to generate effective quinoline molecules by adjusting hyperparameters, activation functions, optimizers, learning rates, etc.
2. **Generating Novel Molecules**: Utilize the MedGAN model to generate new quinoline molecules with drug potential, characterized by good pharmacokinetic properties, low toxicity, and ease of synthesis.
3. **Improving Generation Quality**: Maintain the chirality and charge properties of molecules during the generation process to ensure that the generated molecules have high uniqueness and novelty.

Through these optimization steps, MedGAN can generate new quinoline scaffold molecules based on complex molecular graphs and evaluate their key attributes such as pharmacokinetics, toxicity, and synthetic accessibility. Experimental results show that the optimal model can generate 25% effective molecules, of which 62% are fully connected molecules, 92% are quinoline molecules, 93% are newly discovered, and 95% are unique, retaining chirality, atomic charge, and favorable drug-like properties. These results provide valuable insights, demonstrating how activation functions, optimizers, learning rates, the number of neuron units, molecular size, and structure affect the performance of the generative model and its potential applications in computational drug design.