Scholarlink

<p>Protein fucosylation is a biologically significant post-translational modification that modulates protein stability, receptor signaling, and immune function. However, the structural heterogeneity of <i>N</i>-glycopeptides complicates reliable discrimination of fucosylation topology, particularly between core and outer modifications. In this study, we developed and systematically evaluated a machine learning–based framework for automated classification of <i>N</i>-glycopeptide fucosylation into four categories (none, core, outer, and dual) directly from tandem mass spectrometry (MS/MS) data. A total of 1320 glycopeptide spectra derived from immunoglobulin G (IgG) and alpha-1-acid glycoprotein (AGP) were used for supervised training and testing. Thirteen diagnostic fragment ions were extracted as quantitative features and applied to deep neural network (DNN) and support vector machine (SVM) classifiers. Model robustness was enhanced through architectural refinement and imbalance-aware optimization strategies, including focal loss and synthetic minority over-sampling. The optimized DNN model using collision-induced dissociation (CID) spectra achieved 96.5% accuracy on the standard glycoproteins test set and 86.6% overall accuracy on an independent human plasma validation cohort of 5960 glycopeptide spectral matches. Per-class evaluation under this expanded cohort revealed strong recall fornone (97.1%), core (78.4%), and outer (67.3%) topologies while identifying dual fucosylation as the principal performance bottleneck (recall 22.3%, precision 13.5%, AUC 0.98). This asymmetry, which is not detectable from overall-accuracy metrics alone, is mechanistically attributable to the low MS/MS detectability of core-fucosylation marker ions (Y1F–Y4F, 16–32% across CID and HCD), which produces missing-not-at-random feature vectors that bias prediction away from the dual class while leaving probability ranking informative. Notably, CID-only models consistently outperformed combined CID/HCD datasets, contradicting the common assumption that multimodal fragmentation enhances classification and indicating that higher-energy fragmentation degrades preservation of topology-defining outer-fucosylation ions. These findings provide a per-class, mechanistic re-evaluation of four-class fucosylation classification and identify acquisition strategy and missing-data structure as the principal levers for future improvement in dual-class recovery.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A machine learning model for site-specific classification of N-glycoprotein fucosylation using tandem mass spectrometry and deep neural network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月29日

查看全文
  • Mina Park,
  • Catia Mota,
  • Heeyoun Hwang

摘要

Protein fucosylation is a biologically significant post-translational modification that modulates protein stability, receptor signaling, and immune function. However, the structural heterogeneity of N-glycopeptides complicates reliable discrimination of fucosylation topology, particularly between core and outer modifications. In this study, we developed and systematically evaluated a machine learning–based framework for automated classification of N-glycopeptide fucosylation into four categories (none, core, outer, and dual) directly from tandem mass spectrometry (MS/MS) data. A total of 1320 glycopeptide spectra derived from immunoglobulin G (IgG) and alpha-1-acid glycoprotein (AGP) were used for supervised training and testing. Thirteen diagnostic fragment ions were extracted as quantitative features and applied to deep neural network (DNN) and support vector machine (SVM) classifiers. Model robustness was enhanced through architectural refinement and imbalance-aware optimization strategies, including focal loss and synthetic minority over-sampling. The optimized DNN model using collision-induced dissociation (CID) spectra achieved 96.5% accuracy on the standard glycoproteins test set and 86.6% overall accuracy on an independent human plasma validation cohort of 5960 glycopeptide spectral matches. Per-class evaluation under this expanded cohort revealed strong recall fornone (97.1%), core (78.4%), and outer (67.3%) topologies while identifying dual fucosylation as the principal performance bottleneck (recall 22.3%, precision 13.5%, AUC 0.98). This asymmetry, which is not detectable from overall-accuracy metrics alone, is mechanistically attributable to the low MS/MS detectability of core-fucosylation marker ions (Y1F–Y4F, 16–32% across CID and HCD), which produces missing-not-at-random feature vectors that bias prediction away from the dual class while leaving probability ranking informative. Notably, CID-only models consistently outperformed combined CID/HCD datasets, contradicting the common assumption that multimodal fragmentation enhances classification and indicating that higher-energy fragmentation degrades preservation of topology-defining outer-fucosylation ions. These findings provide a per-class, mechanistic re-evaluation of four-class fucosylation classification and identify acquisition strategy and missing-data structure as the principal levers for future improvement in dual-class recovery.