AI-Powered Biased Ligand Design

点击翻译按钮获取中文摘要

AI-Powered Biased Ligand Design A biased ligand which elicits a certain cellular signal but does not affect other pathways is an attractive drug candidate as it can minimize unwanted or adverse effects. Unfortunately, very few current computer-aided drug design methods can enable biased ligand design. Moreover, there is an urgent need to expand the druggable chemical space for those very promising drug targets which have plenty of potent ligands developed, but unfortunately, no approved drugs. We plan to apply the artificial intelligence (AI) techniques to address the two challenges by developing interaction profile scoring function models to enable biased ligand design, and Drug-GAN models to achieve de novo chemical structure design. The central hypothesis of this application states that the function as well as the signaling pathways elicited by a ligand is encoded in the ligand- residue interaction profile (IP), and machine learning algorithms can learn the key attributes of the IP and generate scoring functions, coined IPSFs, to recognize similar ligands in a screening library. The second hypothesis of this application states that generative adversarial networks (GAN) can learn chemical patterns from input and de novo design novel chemical structures. Thus, the AI-powered algorithms and Drug-GAN models will be able to tackle the two challenges, and likely revolutionize future drug discovery. Cannabinoid receptors, CB1R and CB2R, are an ideal model target system for experimental evaluation. The proposal has four aims. In Aim 1, we will develop IPSFs to specifically design agonists or antagonists of CB1R or CB2R, and agonists which can activate a certain signaling pathway. Those target-specific, function-specific and signaling pathway-specific IPSFs will enable biased ligand design. In Aim 2, we will develop Drug- GAN models to rationally design novel chemical structures as potential agonists or antagonists of CB1R and CB2R. In Aim 3, we will acquire top hits of screening compounds and Drug-GAN designed compounds, and conduct binding and functional assays to validate the predictions. In Aim 4, we will develop an expandable computational platform called PBLD to integrate the developed IPSF models and Drug-GAN-generated druglike chemicals, and launch webtools and APIs to conduct biased ligand design using the developed IPSF models and de novo design using the developed Drug-GAN models. We estimate that IPSFs and Drug-GAN models can be generated for about 300 drug targets based on a recent statistics analysis on the ChEMBL database. PBLD has the potential to become a national resource for biased ligand design with more and more IPSF and Drug-GAN models implemented to PBLD.