Deep Generative Methods For Target Specific Drug Design

 Efficiently shortening the early drug design phase can be achieved by directly generating potential binding chemical compounds based on the properties of target receptors. Traditionally, drug design heavily relies on high throughput screening (HTS), which lacks prior information for selecting compounds to test. In this dissertation, we integrate receptors' properties into a deep generative model framework to directly and efficiently generate high-binding chemical compounds. Chapters 1 to 4 provide background information on drug design and deep learning methods. The subsequent chapters formally introduce our work. The first part introduces a design comprising a graph neural network and a general adversarial network for shape-constrained small-molecule drugs specific to receptors. This method generates 3D conformation-ready molecules for a given receptor, performing both scaffold-hopping and de novo ligand design tasks. The shape-constrained molecule generator proves more efficient in producing high-binding molecules for a receptor compared to standard HTS datasets such as Enamine REAL. The second part presents a Monte Carlo sampling method operating in a latent space to generate protein-binding specific peptide drugs. This method, incorporating limited iterations of feedback from molecular dynamic simulations, computationally and experimentally identifies effective binding peptide drugs in two protein systems. A subsequent improvement involves redesigning the peptide sampler's optimization loop, allowing more feedback iterations and producing peptides with superior binding qualities in less time. In summary, our work demonstrates how incorporating receptors' properties into deep learning models enhances the efficiency of the early drug design process.