Meet MedGAN: A Deep Learning Model based on Wasserstein Generative Adversarial Networks and Graph Convolutional Networks for Novel Molecule Design

The growing urgency for innovative drugs in various medical fields, such as antibiotics, cancer treatments, autoimmune disorders, and antiviral therapies, underscores the need for increased research and development efforts. Drug discovery, a complex process involving exploring a vast chemical space, can benefit from computational methods and, more recently, deep learning. Deep learning, particularly generative AI, proves promising in efficiently exploring extensive chemical libraries, predicting new bioactive molecules, and enhancing drug candidate development by learning and recognizing patterns over time.

Researchers from Faculty of Medicine, University of Porto, Porto, Portugal, Department of Community Medicine, Information and Decision in Health, Faculty of Medicine, University of Porto, Porto, Portugal, Center for Health Technology and Services Research (CINTESIS), Porto, Portugal, Faculty of Health Sciences, University Fernando Pessoa, Porto, Portugal, SIGIL Scientific Enterprises, Dubai, UAE, and MedFacts Lda., Lisbon, Portugal has created MedGAN. This deep learning model utilizes Wasserstein Generative Adversarial Networks and Graph Convolutional Networks. It aims to generate novel quinoline scaffold molecules by working with intricate molecular graphs. The development process involved fine-tuning hyperparameters and assessing drug-like qualities such as pharmacokinetics, toxicity, and synthetic accessibility.

The study discusses the urgent need for new and effective drugs in various classes, such as antibiotics, cancer treatments, autoimmune disorders, and antiviral treatments, due to emerging challenges in drug delivery, disease mechanisms, and rapid mutation rates. It highlights the potential of generative AI in drug discovery, including drug repurposing, drug optimization, and de novo design, using techniques like recursive neural networks, autoencoders, generative adversarial networks, and reinforcement learning. The study emphasizes the importance of exploring the vast chemical space for drug discovery and the role of computational methods in guiding the process toward optimal goals.

The study utilized the WGAN architecture to develop a new GAN model for creating quinoline-like molecules. The objective was to improve and optimize the model’s output by emphasizing the learning of particular key patterns, such as the molecular scaffold inherent to the quinoline structure. The model was fine-tuned using an optimized GAN approach, where three different models (models 1, 2, and 3) were trained and evaluated based on their ability to generate valid chemical structures. Models 2 and 3 showed marked improvement over the base model, achieving higher scores for developing valid chemical structures. These models were selected for further fine-tuning using a larger dataset of quinoline molecules.

The study also divided the ZINC15 dataset into three subsets based on complexity, which were used sequentially for fine-tuning training. The subsets included quinoline molecules of different sizes and constitutions, allowing for a more tailored approach to generating molecules with superior chemical properties.

The MedGAN model has been optimized to create quinoline scaffold molecules for drug discovery and has achieved impressive results. The best model developed 25% valid molecules and 62% fully connected, of which 92% were quinolines, and 93% were unique. It preserved important properties such as chirality, atom charge, and favorable drug-like attributes. It successfully generated 4831 fully connected and unique quinoline molecules not present in the original training dataset. These generated molecules adhere to Lipinski’s rule of 5, which indicates their potential bioavailability and synthetic accessibility. 

In conclusion, The study presents MedGAN, an optimized GAN with GCN for molecule design. The generated molecules preserved important drug-like properties, including chirality, atom charge, and favorable pharmacokinetics. The model demonstrated the potential to create new molecular structures and enhance deep learning applications in computational drug design. The study highlights the impact of various factors, such as activation functions, optimizers, learning rates, molecule size, and scaffold structure, on the performance of generative models. MedGAN offers a promising approach to rapidly access and explore chemical libraries, uncovering new patterns and interconnections for drug discovery.

Check out the Paper and Github. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. Join our 36k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel