Liang Qiao

Protein glycosylation, a post‐translational modification involving the attachment of glycans to proteins, plays critical roles in numerous physiological and pathological cellular functions. Characterization of protein glycosylation is one of the most challenging problems due to the high heterogeneity of glycosites and glycan structures. Recently, deep learning has been adopted to predict N ‐glycopeptide tandem mass spectrometry (MS/MS) spectra and exhibited a promising effect in N ‐glycoproteomics analysis. However, current deep learning frameworks struggle to accurately predict O ‐glycopeptide MS/MS spectra due to the complexity of O ‐glycopeptides and the limited availability of training data. In this study, we introduce DeepGPO, a deep learning framework for the prediction of O ‐glycopeptide MS/MS spectra. The DeepGPO incorporates a Transformer module alongside two graph neural network modules designed for handling branched glycans. To address the issue of data scarcity in O ‐glycoproteomics, various training methods are adopted in DeepGPO, such as the introduction of training weights for different MS/MS spectra and the adoption of pre‐training strategies. With the predicted MS/MS, O ‐glycosylation sites can be localized even in the absence of site‐determining ions. Currently, DeepGPO supports both mono‐ and double‐ O ‐glycosylated peptides. It shows promising application in clinical human O ‐glycoproteomics. We anticipate that DeepGPO will inspire future advancements in glycoproteomics research.

In this response, we address key points by commentators on our previous article, "Unsupervised Machine Learning for Differential Analysis in Proteomics" (DOI: 10.1021/acs.analchem.5c03117), concerning the choice and characteristics of statistical testing and machine learning (ML) in differential proteomics. We clarify that while certain ML methods are statistically grounded, many operate on distinct nonparametric principles, offering an alternative approach when data violate standard distributional assumptions or exhibit complex multivariate structures. We also want to clarify our position that ML is proposed not as a replacement for established statistical frameworks but as a valuable expansion of the analytical toolbox, particularly useful in exploratory analysis or with heterogeneous data. We emphasize methodological pluralism, advocating for the combined use of ML and statistical methods across different stages of research, from hypothesis generation to confirmatory testing, to better address the diverse challenges in precision proteomics and to enrich biological discovery.

Existing cellular heterogeneity necessitates continuous analysis in a single cell to capture dynamic information. However, current mass spectrometry-based single-cell metabolomics analysis is a destructive technique and cannot perform multiple samplings from the same cell, which complicates the accurate tracking of metabolite level alterations. In this study, we present a minimally invasive single-cell mass spectrometry technique that enables sampling of less than 10% of the total cell volume per iteration by using nanocapillaries. This minimally invasive process allows for repeated sampling from a single cell without compromising its essential physiological functions. The resulting time-resolved metabolomics analysis provides the alteration in metabolite levels in a single cell in a process of erastin-induced ferroptosis. As compared to the metabolite levels at certain time points, the alteration in metabolite levels exhibits a stronger correlation with the extent of mitochondrial lipid peroxidation. More importantly, by mapping the differentiation trajectory of individual cells during this process, two distinct groups─proliferating and quiescent cells─are identified. These two groups of cells exhibit different sensitivities to ferroptosis, with quiescent cells being more susceptible to ferroptosis, which could only be tracked by using the time-resolved single-cell metabolomics analysis. Ultimately, this technique overcomes the challenge in single-cell dynamic studies posed by cellular heterogeneity, providing more accurate biological insights into the diversity and complexity of cellular metabolic activities.

Endometrial cancer (EC) molecular subtyping is critical for prognosis and treatment but remains hindered by reliance on invasive tissue biopsies and time-consuming genomic sequencing. Here, we present a minimally invasive approach integrating MALDI-TOF mass spectrometry and LC-MS/MS-based peptidomic profiling of plasma extracellular vesicles (EVs) with machine learning for rapid EC screening and subtyping. EVs were isolated from EC patients and controls, and their peptidome fingerprints were analyzed. A machine learning model utilizing 12 discriminative MALDI-TOF MS features, the levels of CA125 and HE4, and clinical features related to cancer risk achieved an AUC of 0.867 in distinguishing EC from the controls. For molecular subtyping (POLE mutant, NSMP, MMRd, P53-abnormal), a multiclassification model demonstrated micro/macro-averaged AUCs of 0.91/0.90. LC-MS/MS identified 7,479 peptides, with fibrinogen α chain (FGA), protease serine 3 (PRSS3), and apolipoprotein A-I (APOA1) emerging as key biomarkers linked to specific subtypes. This study establishes a high-throughput, cost-effective platform for EC management, bridging translational gaps in precision oncology.