Compound-protein Interaction Prediction Based On Molecular Images

The study of compound-protein interactions is of great importance throughout the process of drug discovery.Traditional methods for exploring compound-protein interactions tend to take time-consuming and resource-intensive biological experiments into account,while the proposed models in silico have overcome those to a certain extent due to the rapid development of computer technology recently.Most of these previous prediction methods pay more attention to extracting compound features from graph-based molecular representations.In contrast,the image-based representations of molecules,containing more information on the molecular structure,are rarely reported for compound-protein interaction prediction tasks.Meanwhile,there appears a flurry of superb work in the realm of computer vision,and they lay a solid technical foundation for molecular image feature extraction.Therefore,this paper aims at investigating the feasibility and effectiveness of molecular images in solving compound-protein interaction prediction tasks by combining the techniques of computer vision,multimodal learning,and multi-task unsupervised pre-training.In view of the aforementioned points,the main work and contributions of this thesis are as follows:Firstly,we proposed Image CPI with molecular images.Image CPI is a multimodalbased framework for predicting compound-protein interactions.The model takes molecular images and amino acid sequences as the inputs,incorporating the visual and textual modalities.To bridge the ‘semantic gap’ due to heterogeneous modalities,Image CPI employs a CNN-GRU network to extract features of images and blends them through an attention mechanism.The existing experiments demonstrate the superiority of our model compared with several state-of-the-art precedents.In order to alleviate the “black box” nature of the model and to explore the interpretability of Image CPI,visual experiments are conducted on the prediction results of Image CPI that confirm the ability of capturing compound-protein binding sites.Secondly,to enhance the prediction performance of Image CPI and alleviate the problem of lacking labeled samples in drug discovery,a well-designed unsupervised pre-training model,Mol IMG,is proposed with a framework of multi-task learning,which learns both “molecular similarity” and “molecular legitimacy” tasks simultaneously through hard parameter sharing.The model has been trained intensively and extensively on 2 million unlabeled molecular images,leading to improving the convergence of downstream tasks and alleviating the overfitting problem.Experiments on different downstream tasks show that Mol IMG strengthens the prediction ability of Image CPI as well as improves the performance of several molecular property prediction tasks.In addition,this paper constructs the “Anti-SARS-Co V-2 Drug”dataset which suffers from “lacking labeled samples” and “unbalanced data” problems.As regards to the result,Mol IMG is able to eliminate the drawbacks of the dataset to a certain extent,further,it can predict antiviral drugs beyond the training samples.