Prediction of Ligand Selectivity and Efficacy Using Artificial Intelligence Algorithms

dc.creatorYeung, Tatianaen_US
dc.creatorMathew, Ezeken_US
dc.creatorLiu, Kevinen_US
dc.creatorMadugula, Sirishaen_US
dc.creatorNguyen, Trongen_US
dc.creatorPham, Tyleren_US
dc.creatorLiu, Jinen_US
dc.creator.orcid0009-0003-4083-1896 (Yeung, Tatiana)
dc.description.abstractIntroduction: Bringing new pharmaceuticals to market is a time-intensive and expensive process. The purpose of this project is to combine computational structure based approaches such as docking and machine learning methodologies to yield ideal pharmaceutical candidates for future exploration. The ligands of interest were those that bind to the Dopamine 4 (D4) and Sigma 1 (S1) receptors, serving as prime candidates for treatment of neurological ailments. Design of more efficacious and selective ligands could allow researchers and clinicians to improve treatment of patients with such conditions. The primary objective is to leverage advancements in computational chemistry to approach the problem of identifying ideal drug candidates using both ligand based and structure-based approaches. Methods: Receptor structures were identified for both the D4 receptor using the PDB 5WIU, and the S1 receptor using the PDB 5HK1. A list of possible ligands was obtained from a DrugBank database, collaborators at the University of Nebraska Medical Center, and a similarity search. The DrugBank database of FDA approved drugs was scanned for ligands to both the D4 and S1 receptors, and a list of 1415 drug ligands was compiled. Using Autodock software, we docked each ligand with 10 poses. After docking, the ligands were ranked by binding affinity. Using Autodock software, the 83 ligands we received from our collaborators were also docked and ranked by binding affinity to the D4 and S1 receptors. To further expand our pool of potential candidates for further study, a similarity search was conducted by screening through a drug database (ZINC) to identify the FDA approved drugs that were most structurally similar to the 83 ligands. Ligands from both the DrugBank and the similarity search were integrated into a machine learning pipeline using graph neural networks to predict the Ki values, thereby identifying compounds with high binding affinity. Once the ligands of highest affinity are identified by the machine learning model, they will be sent to our collaborators for in vitro testing. Results: Of the top 50/1450 FDA approved drugs with the lowest Ki values for both D4, 18 overlapped with the top 50/1450 lowest Ki values for S1. Docking the 83 ligands to the D4 and S1 receptors showed that the ligands were generally more strongly bonded to S1 than D4. We will deliver the top ligand candidates belonging to both D4 and S1 ligand classes as identified by the machine learning model to our collaborators so they can perform further in vitro testing. This will allow us to validate our computational efforts with real world testing. Conclusion: We will leverage the trained machine learning model to search through more databases and identify other prime candidates for future exploration.en_US
dc.titlePrediction of Ligand Selectivity and Efficacy Using Artificial Intelligence Algorithmsen_US