Xinxin Yu, Yuanting Chen, Long Chen, Weihua Li, Yuhao Wang, Yun Tang, Guixia Liu
{"title":"GCLmf: A Novel Molecular Graph Contrastive Learning Framework Based on Hard Negatives and Application in Toxicity Prediction.","authors":"Xinxin Yu, Yuanting Chen, Long Chen, Weihua Li, Yuhao Wang, Yun Tang, Guixia Liu","doi":"10.1002/minf.202400169","DOIUrl":null,"url":null,"abstract":"<p><p>In silico methods for prediction of chemical toxicity can decrease the cost and increase the efficiency in the early stage of drug discovery. However, due to low accessibility of sufficient and reliable toxicity data, constructing robust and accurate prediction models is challenging. Contrastive learning, a type of self-supervised learning, leverages large unlabeled data to obtain more expressive molecular representations, which can boost the prediction performance on downstream tasks. While molecular graph contrastive learning has gathered growing attentions, current models neglect the quality of negative data set. Here, we proposed a self-supervised pretraining deep learning framework named GCLmf. We first utilized molecular fragments that meet specific conditions as hard negative samples to boost the quality of the negative set and thus increase the difficulty of the proxy tasks during pre-training to learn informative representations. GCLmf has shown excellent predictive power on various molecular property benchmarks and demonstrates high performance in 33 toxicity tasks in comparison with multiple baselines. In addition, we further investigated the necessity of introducing hard negatives in model building and the impact of the proportion of hard negatives on the model.</p>","PeriodicalId":18853,"journal":{"name":"Molecular Informatics","volume":" ","pages":"e202400169"},"PeriodicalIF":2.8000,"publicationDate":"2024-10-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Molecular Informatics","FirstCategoryId":"3","ListUrlMain":"https://doi.org/10.1002/minf.202400169","RegionNum":4,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q3","JCRName":"CHEMISTRY, MEDICINAL","Score":null,"Total":0}
引用次数: 0
Abstract
In silico methods for prediction of chemical toxicity can decrease the cost and increase the efficiency in the early stage of drug discovery. However, due to low accessibility of sufficient and reliable toxicity data, constructing robust and accurate prediction models is challenging. Contrastive learning, a type of self-supervised learning, leverages large unlabeled data to obtain more expressive molecular representations, which can boost the prediction performance on downstream tasks. While molecular graph contrastive learning has gathered growing attentions, current models neglect the quality of negative data set. Here, we proposed a self-supervised pretraining deep learning framework named GCLmf. We first utilized molecular fragments that meet specific conditions as hard negative samples to boost the quality of the negative set and thus increase the difficulty of the proxy tasks during pre-training to learn informative representations. GCLmf has shown excellent predictive power on various molecular property benchmarks and demonstrates high performance in 33 toxicity tasks in comparison with multiple baselines. In addition, we further investigated the necessity of introducing hard negatives in model building and the impact of the proportion of hard negatives on the model.
期刊介绍:
Molecular Informatics is a peer-reviewed, international forum for publication of high-quality, interdisciplinary research on all molecular aspects of bio/cheminformatics and computer-assisted molecular design. Molecular Informatics succeeded QSAR & Combinatorial Science in 2010.
Molecular Informatics presents methodological innovations that will lead to a deeper understanding of ligand-receptor interactions, macromolecular complexes, molecular networks, design concepts and processes that demonstrate how ideas and design concepts lead to molecules with a desired structure or function, preferably including experimental validation.
The journal''s scope includes but is not limited to the fields of drug discovery and chemical biology, protein and nucleic acid engineering and design, the design of nanomolecular structures, strategies for modeling of macromolecular assemblies, molecular networks and systems, pharmaco- and chemogenomics, computer-assisted screening strategies, as well as novel technologies for the de novo design of biologically active molecules. As a unique feature Molecular Informatics publishes so-called "Methods Corner" review-type articles which feature important technological concepts and advances within the scope of the journal.