ระบบปัญญาประดิษฐ์เพื่อทำนายและอธิบายความสามารถในการจับตัวระหว่างสารและโปรตีน

ชื่อนักเรียนผู้จัดทำโครงงานวิทยาศาสตร์

กรวิชญ์ ปอยสูงเนิน, พงค์ภักดิ์ มะโนเรศ

อาจารย์ที่ปรึกษาโครงงานวิทยาศาสตร์

พิษณุ จันทรเสวต, กอบชัย ดวงรัตนเลิศ

โรงเรียนที่กำกับดูแลโครงงานวิทยาศาสตร์

โรงเรียนเตรียมอุดมศึกษา

ปีที่จัดทำโครงงานวิทยาศาสตร์

พ.ศ. 2564

บทคัดย่อโครงงานวิทยาศาสตร์

As the result of recent pandemics, many people have lost their lives, and most have suffered greatly from losing their loved ones, experiencing physical debilitation, and striving against socio-economic disruption. This challenges the capability of the pharmaceutical industry in designing new drugs to remedy and prevent the situation from further escalation. However, conventional drug-discovery protocols are costly, labor-intensive, and time-consuming. Because of these, artificial intelligence (A.I.) technologies have been increasingly used to facilitate the process, especially in screening new drug candidates to target disease-relevant proteins. In this research, we proposed a new pipeline for predicting the binding activity between protein and compound using their amino acid and SMILES strings. While sequence representation does not have intrinsic information of 3D conformation or binding configuration, it is not difficult to acquire experimentally and is therefore much more available than 3D representation data. We built a neural network model with the additive attention mechanism to learn the significance of each unit in the sequences of protein and compound that contribute to the binding process. The result suggested that the proposed deep learning model is a practical approach for candidate screening. The model achieved the remarkable AUC score of 0.86 in the BindingDB benchmark datasets. Extensive analysis on multiple groups of drugs including, antivirals, antibacterials, anti-tumor, and painkillers had also confirmed the capability of the model to indicate the crucial residue/token interaction of existing drugs and their disease targets. In addition, the application programming interface for binding prediction was deployed on the cloud servers to service​ interested​ external​ users such​ as pharmaceutical companies, cosmetics industries, and any related scientists. At this stage of the project, we will continue exploring new architecture and various combinations of hyperparameters to improve the model performance. Future collaboration with molecular docking and De Novo drug design should enhance the opportunity to discover new promising drugs for any given protein of interest.