Van-The Le , Zi-Jun Zhan , Thi-Thu-Phuong Vu , Muhammad-Shahid Malik , Yu-Yen Ou
{"title":"ProtTrans and multi-window scanning convolutional neural networks for the prediction of protein-peptide interaction sites","authors":"Van-The Le , Zi-Jun Zhan , Thi-Thu-Phuong Vu , Muhammad-Shahid Malik , Yu-Yen Ou","doi":"10.1016/j.jmgm.2024.108777","DOIUrl":null,"url":null,"abstract":"<div><p>This study delves into the prediction of protein-peptide interactions using advanced machine learning techniques, comparing models such as sequence-based, standard CNNs, and traditional classifiers. Leveraging pre-trained language models and multi-view window scanning CNNs, our approach yields significant improvements, with ProtTrans standing out based on 2.1 billion protein sequences and 393 billion amino acids. The integrated model demonstrates remarkable performance, achieving an AUC of 0.856 and 0.823 on the PepBCL Set_1 and Set_2 datasets, respectively. Additionally, it attains a Precision of 0.564 in PepBCL Set 1 and 0.527 in PepBCL Set 2, surpassing the performance of previous methods. Beyond this, we explore the application of this model in cancer therapy, particularly in identifying peptide interactions for selective targeting of cancer cells, and other fields. The findings of this study contribute to bioinformatics, providing valuable insights for drug discovery and therapeutic development.</p></div>","PeriodicalId":16361,"journal":{"name":"Journal of molecular graphics & modelling","volume":"130 ","pages":"Article 108777"},"PeriodicalIF":2.7000,"publicationDate":"2024-04-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of molecular graphics & modelling","FirstCategoryId":"99","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S1093326324000779","RegionNum":4,"RegionCategory":"生物学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"BIOCHEMICAL RESEARCH METHODS","Score":null,"Total":0}
引用次数: 0
Abstract
This study delves into the prediction of protein-peptide interactions using advanced machine learning techniques, comparing models such as sequence-based, standard CNNs, and traditional classifiers. Leveraging pre-trained language models and multi-view window scanning CNNs, our approach yields significant improvements, with ProtTrans standing out based on 2.1 billion protein sequences and 393 billion amino acids. The integrated model demonstrates remarkable performance, achieving an AUC of 0.856 and 0.823 on the PepBCL Set_1 and Set_2 datasets, respectively. Additionally, it attains a Precision of 0.564 in PepBCL Set 1 and 0.527 in PepBCL Set 2, surpassing the performance of previous methods. Beyond this, we explore the application of this model in cancer therapy, particularly in identifying peptide interactions for selective targeting of cancer cells, and other fields. The findings of this study contribute to bioinformatics, providing valuable insights for drug discovery and therapeutic development.
期刊介绍:
The Journal of Molecular Graphics and Modelling is devoted to the publication of papers on the uses of computers in theoretical investigations of molecular structure, function, interaction, and design. The scope of the journal includes all aspects of molecular modeling and computational chemistry, including, for instance, the study of molecular shape and properties, molecular simulations, protein and polymer engineering, drug design, materials design, structure-activity and structure-property relationships, database mining, and compound library design.
As a primary research journal, JMGM seeks to bring new knowledge to the attention of our readers. As such, submissions to the journal need to not only report results, but must draw conclusions and explore implications of the work presented. Authors are strongly encouraged to bear this in mind when preparing manuscripts. Routine applications of standard modelling approaches, providing only very limited new scientific insight, will not meet our criteria for publication. Reproducibility of reported calculations is an important issue. Wherever possible, we urge authors to enhance their papers with Supplementary Data, for example, in QSAR studies machine-readable versions of molecular datasets or in the development of new force-field parameters versions of the topology and force field parameter files. Routine applications of existing methods that do not lead to genuinely new insight will not be considered.