Accelerating Drug Discovery through Artificial Intelligence with Enhanced Interpretability and Reliability

Abstract

Drug discovery and development have been known for requiring lengthy timelines and astronomical costs. Unfortunately, both time and costs have continued to rise, significantly reducing the sustainability of the pharmaceutical industries. To efficiently reduce the costs and time of drug discovery and development, the use of artificial intelligence (AI) is gaining remarkable attention. There have been many attempts to utilize AI throughout the drug development process recently; however, most lack practicality due to low interpretability and reliability. In this dissertation, I address these problems by developing advanced AI models based on enhanced interpretability and reliability, especially in the hit identification and lead optimization stages. The first study involves developing a generative model that creates hit-like candidates for specific cancers based on the genotype of cancer cell lines. Inspired by text-to-image generative models, the proposed model, named G2D-Diff, processes the input of genotype and response class and generates the molecular structures that meet the given conditions. Through various evaluation settings, the model showed outstanding performance in generating hit-like candidates by enhancing reliability and interpretability. In particular, the model proved its high practicality by successfully generating hit-like candidates through zero-shot case studies. The second study focuses on developing a predictive model that predicts hERG toxicity with high accuracy based on the molecular structure. Incorporating transfer learning, Bayesian deep learning technique, and an attention mechanism, the proposed model, named BayeshERG, achieved robustness, reliability, and interpretability. The proposed model not only demonstrated robust predictive performance on various benchmark datasets but also offered highly calibrated predictions, enhancing reliability. Furthermore, the model proved its interpretability by suggesting hERG-related molecular substructures using the attention mechanism. In conclusion, the AI models and the silver-bullet approach that enhance interpretability and reliability discussed in this dissertation are expected to significantly contribute to accelerating drug discovery and development.