A comprehensive landscape of human organ N-glycoproteome Authors: Xiaoyu Hu, Tao Wang, Chao Qin, Yanggang Yuan, Suideng Qin, Hongwei Liang, Zhixin Tian Published: 2026 ABSTRACT AbstractN-glycosylation, as a common post-translational modification with complex structures, plays key roles in protein folding, cellular recognition and signaling pathways. State-of-the-art mass spectrometry-based N-glycoproteomics has enabled deep N-glycoproteome characterization of various human organs. However, a comprehensive N-glycoproteome landscape of human organs remains lacking. Here we present a systematic human N-glycoproteome atlas spanning 74 subjects across 18 organs/tissues with identification of 57,884 N-glycan structure-level and 23,863 monosaccharide composition-level intact N-glycopeptides on 7543 N-glycosites of 5062 N-glycoproteins. Tissue-specific N-glycosylation patterns and crosstalk between sialylation and fucosylation are observed. An organ classifier using multiple machine learning models on our N-glycopeptide dataset is trained. This atlas, complemented by a unified multi-software analysis framework, provides insights into organ-specific glycobiology and establishes a fundamental reference for understanding physiological N-glycosylation characteristics of human tissues.Subject terms: Proteomics, Translational research, Glycobiology, Mass spectrometry Introduction The human body is governed by three interconnected macromolecular systems: the genome, transcriptome, and proteome. Their relationship is defined by the central dogma of molecular biology. A landmark achievement was the sequencing of the human genome in 2001, which identified an estimated 19,000 to 20,000 protein-coding genes1. Subsequent research focused on characterizing the human proteome. In 2014, mass spectrometry-based analysis of 30 human tissues detected proteins from 17,294 genes, covering approximately 84% of the predicted proteome2. By 2025, the Human Proteome Organization had updated this landscape by integrating data from neXtProt and UniProtKB, validating 18,397 proteins and achieving 93% coverage of the human proteome3 (https://hupo.org/). The functional diversity of proteins is greatly enhanced by post-translational modifications (PTMs), with over 400 types reported4,5. The roles of phosphorylation6, acetylation7, ubiquitination8, and methylation9 are comprehensively characterized, and it is now clear that PTMs often act cooperatively to fine-tune protein structure and function10–12. N-Glycosylation is a prevalent PTM that critically influences protein folding13, cellular recognition14, signaling pathways15 and immune system regulation16. Unlike small-molecule PTMs, such as methylation, glycosylation exhibits a unique level of structural complexity. The complexity of glycosylation is reflected in its two forms of heterogeneity. Macro-heterogeneity arises from the occupancy of multiple glycosites within a single protein. Micro-heterogeneity, conversely, refers to the presence of multiple linkages at a single glycosite. To systematically catalog this complexity, the Human Glycome Atlas Project was launched in 202317. Mass spectrometry-based N-glycoproteomics has emerged as the primary technique for the comprehensive site- and structure-specific characterization of the N-glycoproteome. The latter provides simultaneous information on the peptide backbone (which identifies the protein), the precise glycosite and the detailed linkage of the N-glycan. Aiming precision and personalized medicine, N-glycoproteomics has been widely adopted for the discovery of putative diagnosis and prognosis biomarkers, drug targets, and subtyping of various human pathological conditions (especially cancers) using human tissues (paired tumor and adjacent), including pancreas18,19, brain20–25, lung26,27, ovary28,29, kidney30, and liver31–34. However, a comprehensive draft of the human organ N-glycoproteome is still lacking. In this study, we present a systematic atlas of the human N-glycoproteome, profiling 74 subjects across 18 normal tissues through 219 mass spectrometry runs. In total, 6,897,919 MS/MS spectra were acquired, where 4,227,689 oxonium ion-containing spectra led to the identification of 57,884 unique N-glycopeptides corresponding to 7543 N-glycosites on 5062 distinct N-glycoproteins. This represents a relatively comprehensive dataset of intact N-glycopeptides of humans. An organ classifier using multiple machine learning models on our intact N-glycopeptide dataset was trained. The favorable classification result indicated that the fundamental N-glycosylation signature common to each organ is more substantial than inter-individual heterogeneity. Our site- and structure-specific analysis unveiled broad tissue-spanning patterns and the precise distribution of N-glycosylation macro- and micro-heterogeneity. This dataset also established a foundational resource for understanding tissue-specific N-glycosylation and its biological implications. Results A systemic atlas of N-glycosylation across 18 human organs A label-free N-glycoproteomics workflow was employed to systematically characterize the N-glycoproteomes of 74 subjects (2–6 per organ) from 18 healthy human organs (Supplementary Data 1), revealing 57,884 plausible N-glycan structure-level and 23,863 monosaccharide composition-level intact N-glycopeptides on 7543 N-glycosites of 5062 N-glycoproteins (Fig. 1a and Supplementary Data 2). For false discovery rate (FDR) control at the intact N-glycopeptide spectrum match (GPSM) level, longer peptide length, charge states of 4–5, and lower glycan masses are more reliable (Supplementary Fig. 1a–c). In all the 219 raw LC-MS/MS files, the number of MS2 scans is 10,980,678, and the number of scans containing oxonium ions is 6,924,949 (63.1%). The adrenal gland exhibits the highest proportion of oxonium ion-containing scans (88.4%) (Supplementary Fig. 1d, e). In addition, N-glycosite-determining fragment ions are observed for most (ranging from 66.0 to 85.5%) of the IDs, where N-glycosites are unambiguously localized; structure-diagnostic fragment ions are observed for more than 50% (ranging from 49.9 to 80.1%) of IDs, where monosaccharide sequence isomers (with the same molecular formula including the same monosaccharide compositions) are distinctly differentiated (Supplementary Fig. 1f). In comparison with the known N-glycan DBs of GlyConnect (637 compositions) and GlyGen (350 compositions), 141 unreported compositions are observed in this study (Fig. 1b). The N-glycosylation motif N-x-S/T/C is highly conserved, with longer conserved sequences in fat and urinary bladder (Supplementary Fig. 2). The frequencies of the N-x-S, N-x-T, and N-x-C motifs are 43.8%, 50.5%, and 5.8%, respectively (Fig. 1c). Overall profiles of the intact N-glycopeptides identified from the 18 human organs/tissues. a Total numbers of IDs; b N-glycan compositions validated against the public databases; c motif analysis of the identified intact N-glycopeptides; d N-glycan type, e core type, f the number of N-glycosites per protein and g the number of linkages per N-glycosite distribution of IDs; h, i distribution of N-glycan branch patterns across 18 organs from adrenal gland to urinary bladder. Regarding the number of N-glycosites per protein, the urinary bladder exhibits the greatest heterogeneity, 26.7% glycoproteins contain 2 or more glycosites (Fig. 1d). And in the aspect of N-glycan linkages per N-glycosite, more than 50% glycopeptides exist isomer glycans in multiple organs (Fig. 1e). High-mannose N-glycans and the core structure HexNAc(2)Hex(3) are predominant across organs (Fig. 1f, g). Notably, hybrid N-glycans accounted for nearly one-third in the lung, making it the most abundant. Similar to the previous findings18–20, the bisect structure was rich in brain (10.8%) and kidney (6.8%). The mannose terminal is remarkable in liver (66.6%) (Fig. 1h). The distribution patterns of HexNAc, LacNAc (HexNAc-Hex), and sialyl LacNAc, along with their fucosylated forms, show consistency across organs (Fig. 1i). Analysis of organ-specific IDs revealed 13,591 unique intact N-glycopeptides are restricted to specific organs, while the remaining IDs are shared across multiple organs. The N-glycans defined by intersections 39 and 40 in the Upset plot are exclusively of the high-mannose type (Supplementary Fig. 3a). A general trend was observed wherein glycans identified in more organs exhibited higher abundance. The 18 organs are clustered into three groups, in partially consistent with the established human anatomical systems (Supplementary Fig. 3b); notably, the intact N-glycopeptides detected in no less than 14 organs are exclusively high-mannose, suggesting a conserved biological feature. Negligible batch effect is observed in quality assessment (Supplementary Fig. 4). Analysis of the normalized data from all 219 raw files reveals no batch effects and shows concordant expression profiles within each organ after removing IDs only detected in a single organ (Supplementary Fig. 5). The cross-validation of this study and the previous study also shows the reliability of this data. Substantial overlap (73.8%) at the monosaccharide composition level with the study of mouse35 is found; it is worth noting that 17 Neu5Gc-containing glycans are not identified in this study of human (Supplementary Fig. 6a). Further, this study, identification of 234 glycans of brain, heart, kidney, liver and lung, covers 51.2% glycans of Pan Fang et al.36, among the 98 nonoverlapping glycans, 38 contain Neu5Gc (Supplementary Fig. 6b). The glycan patterns of the five organs are also approximately 50% (50.0–59.7%) overlapped with each other (Supplementary Fig. 6c). The uniform manifold approximation and projection results are also similar, and the five organs are well distinguished (Supplementary Fig. 6d, e). Consistent with the published data, the correlation between the five organs is very low, but the same organs are grouped together (Supplementary Fig. 6f). Cross-validation has verified the credibility of the data in this study. Organ-specific profiles of co-identified glycopeptides (>60%) are observed. N-glycopeptides of the intestine, kidney, muscle, pancreas, and spleen show minimal sialylation (Supplementary Figs. 7–24). High-mannose Six thousand twenty-three high-mannose intact N-glycopeptides are identified in this study. Analysis of high-mannose glycan distribution and branching patterns across organs reveals that tri-antennary is predominated with a percentage of 58.6–81.0%, where liver exhibits the highest proportion of 81.0% (Fig. 2a). Among the 20 identified distinct high-mannose linkages, 11 have been detected in all the organs and are relatively abundant; whereas HexNAc(2)Hex(11) is only detected in liver (Supplementary Fig. 25). Mannose 5 (M5) exhibit high abundance in muscle, urinary bladder and brain. Mannose 6 (M6) exhibits high abundance in the kidney (Fig. 2b). The cumulative abundance trends of M3 to M9 in specific organs are categorized into three distinct patterns (Fig. 2c). In the adrenal gland, brain and liver, expression levels increase initially and then decrease. Fat, urinary bladder, muscle, and testis show a gentle trend. Other organs show an increase-decrease-increase trend. As for the depth of N-glycosite identification, prolow-density lipoprotein receptor-related protein 1 (Q07954, also called CD91), being found in 15 different organs, contains 17 N-glycosites and 13 high-mannose linkages (Fig. 2d). In addition, alpha-2-macroglobulin-like protein 1 (A8K2U0) is identified in all oesophagus samples with 3 potential glycosites among all the 5 glycosites (Fig. 2e and Supplementary Data 3). N-glycosite N104 of VSGVGNNISFEEK is recognized with y9* and N-x-T motif (Supplementary Fig. 26a). N-glycosite N120 of QGNGTFVQTDKPLYTPGQQVYFR is confirmed with site-determining fragment ions of b3*, b5*, b6*, b10*, b11*, b14*, and b15* (Supplementary Fig. 26b). N-glycosite N867 of LGHINFTISTK is confirmed by the site-determining fragment ion pairs of any of the seven b ions (b5*/b6*/b7*/b8*/b9*/b10*) with any of the four y ions (y7*/y8*/y9*/y10*) (Supplementary Fig. 27). High-mannose N-glycosylation across the 18 human organs. a The percentage distribution of the number of antennas; b heatmap of the 18 distinct sequence structures; c distribution of cumulative abundance from HexNAc(2)Hex(3) (M3) to HexNAc(2)Hex(9) (M9) of the 18 organs; d the N-glycosites and linkages of prolow-density lipoprotein receptor-related protein 1; e the N-glycosites and linkages of alpha-2-macroglobulin-like protein 1. Fucosylation Fucosylation has two types of position isomers of core (CF) and branch (BF). In this study, 790 fucosylated N-glycans from 7679 intact N-glycopeptides are identified. BF is the predominant form (65.0%, Fig. 3a). As for branch numbers, the bi-antennary is most prevalent (60.0%, Fig. 3b). A total of 20 N-glycan compositions (mostly mono-fucosylated) from 99 N-glycopeptides are identified in ten or more organs. Most of the CF N-glycans are bi-antennary featuring a (sialyl) LacNAc structure, which is consistent with the previous study37. Moreover, all CF-only glycans identified across the organs have similar trends except for the brain and liver. HexNAc terminal is predominant in brain and liver with manual verification (Supplementary Fig. 28a). In these glycans, most of the plausible BF-only N-linkages are bi- or tri-antennary with (sialyl) Lewis X structures (Fig. 3c). In BF-only glycans with fucose linked to HexNAc, verification of the local structure was carried out. In these glycans, 68.2% contain (sialyl) Lewis X structures, where heart, muscle and the urinary bladder have the highest percentages of 81.4%, 79.4% and 79.2%, respectively (Supplementary Fig. 28b). For the cross-validation of gPeptide results, a non-targeted N-glycome profiling strategy is implemented via an oxonium ion matching tool38. Distribution of the spectral overlap between the 350.1, 512.2, 657.2 amu ions and BF-only IDs from gPeptide across 18 organs shows the widespread presence of diagnostic glycan fragments. 70.7, 78.8, and 68.9% of these IDs identified in gPeptide from prostate are verified with the 350.1, 512.2, and 657.2 amu ions, separately (Supplementary Fig. 29a–c). With adrenal gland 1 (technical replicate 1) as an example, these IDs are mostly verified with 350.1, 512.2, and 657.2 amu ions (Supplementary Fig. 29d). The graphical fragmentation maps display a strong match of Lewis X (Supplementary Fig. 29e–g) and sialyl Lewis X (Supplementary Fig. 29h–j). Two glycopeptides from lumican (P51884) are identified across in several samples from ten organs in a dual group. The graphical fragmentation maps of the P51884_LHINHNNLTESVGPLPK_7_HexNAc(4)Hex(5)Fuc(3) from P51884 are shown in Supplementary Figs. 30 and 31. In fat 2 (technical replicate 3), the core fucose is determined by a series of core fucose-containing Y ions, including YI1, YI2, YI3, and YI4, of which the isotopic envelope fingerprint matches are displayed. The CI2 ion explained the possible presence of the fucose attached to Hex. And the observed mass difference (Δm) between 3,4AI3 and 2,5AI3 accounts for the potential presence of a fucose attached to HexNAc. Manual validation is performed on the matches of these two glycopeptides in representative samples from ten organs (Supplementary Data 4). Fucosylation across the 18 human organs. a Distribution of core fucosylation (CF) and branch fucosylation (BF); b the percentage distribution of the number of antennas; c 20 representative fucosylated N-glycan compositions identified in at least 2 samples per organ and in more than half of the organs; d organ-specific and fucose-number-specific distribution. The cumulative abundance of fucosylated intact N-glycopeptides varies significantly across organs (Fig. 3d). Mono-fucosylation (F1) has the highest abundance in fat, prostate, urinary bladder, testis, and pancreas. Among glycopeptides with a single fucose type, bi-antennary glycans are dominant. Glycopeptides containing both core and branch fucose exhibit pronounced organ specificity. A general trend of an increase in tri-antennary and tetra-antennary N-glycopeptides is observed with dual fucose types. The urinary bladder, lung, and heart are dominated by tri-antennary glycoforms, and the kidney shows the most abundant expression of tetra-antennary glycoforms. In contrast, the ovary displays a more diverse and abundant expression profile compared to other organs (Supplementary Fig. 32). The four-fucose (F4) glycan is absent in the lung, stomach, heart, liver, and spleen. Sialylation Among the 519 sialylated N-glycans from 4242 sialylated intact N-glycopeptides, the mono- (S1), di- (S2), tri- (S3) and tetra- (S4) sialylated ones are 73.1%, 23.9%, 2.8%, and 0.2%, respectively (Fig. 4). Most sialylated intact N-glycopeptides are either bi-antennary (45.1%) or tri-antennary (41.3%). Except for the absence of detection in the stomach, heart, and pancreas, the S1 glycoform is ubiquitously detected across the remaining 15 organs; whereas, only in the colon and spleen, the cumulative abundance of S2 is significantly greater than that of S1. As for the cumulative abundance, urinary bladder and testis always rank among the top five organs, irrespective of sialic acid number; at the same time, they also have abundant sialylated N-glycopeptides; furthermore, highly sialylated N-glycans (S4) are exclusively found in these three organs as well as in prostate, ovary, fat, and adrenal gland. Sialylation across the 18 human organs. The distribution of mono-, bi-, tri-, and tetra-antenna and cumulative abundance of intact N-glycopeptides with different numbers of sialic acids. In order to observe the functional differences of N-glycoproteins with different sialylation, the results of the bioinformatics analysis based on the common IDs (in no less than 30% of the original files) are statistically processed and compared. In molecular function, both S1 and S2 groups are significantly enriched in core extracellular matrix (ECM) structural functions, including ECM structural constituents and collagen binding. S2 group exhibits a broader functional scope by additionally engaging in glycosaminoglycan binding and chaperone binding (Supplementary Fig. 33a). In pathways, the S1 group is primarily associated with processes related to clot formation and cell adhesion, such as blood coagulation and platelet aggregation. In contrast, the S2 group is strongly linked to inflammatory and defense responses, including the acute-phase response and defense response. Both S1 and S2 groups share foundational ECM-related molecular functions, but they diverge in their biological processes; S1 is primarily implicated in hemostasis and adhesion, whereas S2 is predominantly involved in inflammation and immune defense (Supplementary Fig. 33b). Co-occurring fucosylation and sialylation Pearson correlation coefficient and Cramér’s V values are adopted for the analysis of the crosstalk between fucosylation and sialylation. The median expression levels across all samples within the heart, ovary, and spleen exhibit moderate correlations, with Cramér’s V values exceeding 0.3 and Pearson correlation coefficients greater than 0.2 (Fig. 5a). Among these three organs, 2694 form 4590 IDs contain sialic acid, fucose, or both. To identify the specific structural forms of these co-occurring glycans, the types of fucose and the number of branches per N-glycan are counted. A clear preference for HexNAc-Fucose-containing BF glycans of 72.0% is observed in the fucosylation and sialylation (FxSy) group compared to the fucosylation only (FxS0) group (49.9%, Fig. 5b). Concurrently, profiling of branches shows that in the sialylation only (F0Sy) group, sialyl LacNAc constitutes for nearly half of the total branches (46.7%, Fig. 5c), while in FxSy group, the branch forms are mainly the fucosylated form of sialyl LacNAc (30.3%) and LacNAc (12.1%, Fig. 5d). Eight hundred five N-glycopeptides are found to contain the sialyl Lewis X structure in the FxSy group (1083 IDs, Supplementary Data 5). It is proposed from the aforementioned observation that the co-occurrence form of fucose and sialic acid in these organs is mainly the sialyl Lewis X structure. Cross-talk between fucosylation and sialylation. a Cramer’s V coefficient and Pearson correlation coefficient, sample sizes (n) and numbers of subjects are as follows: adrenal gland, n = 12 from 4 subjects; brain, n = 12 from 4 subjects; colon, n = 12 from 4 subjects; fat, n = 14 from 5 subjects; heart, n = 10 from 4 subjects (all with 3 technical replicates except for heart 1, which had 1 replicate); intestine, n = 14 from 5 subjects (all with 3 technical replicates except for intestine 5, which had 2 replicates); kidney, n = 12 from 4 subjects; liver, n = 15 from 5 subjects; lung, n = 12 from 4 subjects; muscle, n = 15 from 6 subjects (all with 3 technical replicates except for muscle 4 with 2 and muscle 5 with 1); oesophagus, n = 18 from 6 subjects; ovary, n = 18 from 6 subjects; pancreas, n = 9 from 3 subjects; prostate, n = 9 from 3 subjects; spleen, n = 7 from 3 subjects (all with 3 technical replicates except for spleen 2 and 3); stomach, n = 9 from 3 subjects; testis, n = 6 from 2 subjects; urinary bladder, n = 6 from 2 subjects. In the box plots, the center line represents the mean, the bounds of the box represent the 25th and 75th percentiles (interquartile range, IQR), the whiskers extend to the minima and maxima values within 1.5 × IQR from the box bounds, and individual data points are overlaid; b distributions of co-modified (FxSy) and fucosylation-only (FxS0) N-glycopeptides in core fucose (CF) glycans, HexNAc-Fucose-containing BF glycans and Hex-Fucose-containing BF glycans; c, d branching patterns of intact N-glycopeptides with FxSy and sialylation-only (F0Sy); e KEGG pathways of N-glycoproteins with different FS compositions (FxSy, FxS0, and F0Sy). Statistical significance was calculated using a one-sided hypergeometric test via the STRING database, p-values were adjusted for multiple comparisons using the Benjamini-Hochberg method to generate the false discovery rate (FDR); f heatmap of common intact N-glycopeptides identified from 6 samples of the ovary. For the three FxSy groups, KEGG (Fig. 5e) and GO analysis (Supplementary Fig. 34) of the corresponding N-glycoproteins reveal different pathways. The FxSy group shows several unique pathways associated with pathological processes. Among the identified N-glycopeptides in at least 13 ovary samples, different FS compositions share similar expression trends within individual samples. Moreover, the six glycans in the FxSy group consistently exhibit sialyl Lewis X structure (Fig. 5f). Common identification In addition to CD91 and lumican, laminin subunit gamma-1 (LAMC1, P11047) emerged as one of the most extensively glycosylated proteins in our dataset. We identified a total of 101 site-specific glycans across 8 N-glycosylation sites from 14 different organs (Fig. 6a). Notably, the N1241 exhibited remarkable microheterogeneity, carrying 25 distinct glycan compositions (Fig. 6b). High-mannose glycans are only found in the ovary samples (Fig. 6c). Based on the Human Protein Atlas (https://www.proteinatlas.org/), we hypothesize that this truncation of glycan maturation may be attributed to the relatively low expression of α-mannosidase (e.g., MAN1A1) and GlcNAc-transferase I (MGAT1) in ovarian tissues compared to other organs. Besides, fucose-containing glycans were highly prevalent in the adrenal gland, intestine, and especially fat, which displayed the highest abundance of fucosylation. This spatial heterogeneity aligns with the elevated expression levels of FUT8, which is the enzyme responsible for core fucosylation. These findings provide a systematic atlas of how tissue-specific enzymatic shapes the glycome of essential structural proteins like laminin. Site-specific glycan distribution of laminin subunit gamma-1 (P11047). a Site-specific glycan distribution across all the organs; b number of glycans of each glycosite; c number of unique site-specific glycans of each organ. In the aspect of intact of N-glycopeptides, common N-glycosylation refers to IDs found in no less than 85 raw files across the samples. K-nearest neighbors (KNN)-based data imputation and normalization generate an expression matrix of 85 intact N-glycopeptides across the 18 organs for hierarchical clustering analysis. The organ clustering is largely matched to human anatomical systems, except for brain tissue (Fig. 7). Organ cluster 1 is composed of organs from the endocrine system (including adrenal gland, testis, ovary, and fat), the digestive system (oesophagus, intestine, and colon), with the spleen representing the immune system. Organ cluster 3 is composed of digestive system organs, including the liver, pancreas, and stomach. Organ cluster 4 is comprised of prostate and urinary bladder from the urinary system. Our analysis reveals that the brain exhibits pronounced differences from all other organs, extending and strengthening previous findings of its distinctiveness from the heart, kidney, liver, and lung36. Common N-glycosylation analysis. Hierarchical clustering of intact N-glycopeptides identified in 18 human organs/tissues and bioinformatics analysis corresponding to four glycopeptide clusters. Clustering of intact N-glycopeptides at the linkage structure level reveals distinct patterns. Sialylation and fucosylation are mainly found in NGP cluster 1. This also reflects our previous assertion that the expression patterns of the same FS composition are likely to be similar. Notably, NGP cluster 1 shows high expression in most organs but low expression in organ cluster 3 and the brain. In contrast, NGP clusters 2 and 3 primarily consist of high mannose glycans. These two NGP clusters exhibit low expression in organ cluster 4, but are highly expressed in other organs. KEGG and GO analysis of the N-glycoproteins from the four clusters reveal distinct functional profiles. The functional enrichment analysis revealed four distinct protein clusters with specific biological roles. NGP cluster 1 is significantly associated with coagulation and fibrinolysis, encompassing zymogen activation and integrin interactions. NGP cluster 2 is characterized by glycoproteins involved in co-translational modification and processing within the endoplasmic reticulum, particularly related to N-glycan biosynthesis. NGP cluster 3 comprises components of the endomembrane system, including the lysosome and endoplasmic reticulum membrane. Finally, NGP cluster 4 is defined by its involvement in ECM organization, featuring glycoproteins with glycosaminoglycan-binding capacity. N-glycoproteome multi-model classifier To further evaluate the discriminative power of N-glycosylation patterns, we trained a multi-class classifier based on the common N-glycosylation matrix (Supplementary Data 6), using four distinct machine learning algorithms of decision tree, neural network, random forest and support vector machine (SVM). Before classification, T-distributed stochastic neighbor embedding (tSNE) analysis is executed to check the similarity among samples, and samples from the same organ can be grouped together (Supplementary Fig. 35). The performance is assessed by F1 scores and ROC curves, demonstrating that organs are divisible based on the co-identified IDs (Fig. 8a, b). Among the algorithms, random forest achieved the highest average F1 score (0.959 ± 0.030), closely followed by SVM (0.939 ± 0.032) and neural network (0.874 ± 0.054). These trends are consistent with the 2022 study35. Decision tree performs relatively poorly, with an average F1 score of comparatively 0.772 ± 0.060. While neural networks, random forest, and SVM excel in organ classification, the decision tree algorithm struggles specifically to differentiate the testis and urinary bladder. This limitation likely stems from the inherently smaller training set sizes for these two organs, as decision tree is more prone to overfitting and reduced generalization in learning from limited data compared to ensemble or neural network methods (Fig. 8c). The ROC curves show that neural network, random forest and SVM excel in most of the organs, except for brain, which might be due to its significantly unique N-glycosylation heterogeneity (Fig. 8d). It is important to note that our classifier is trained on an expression matrix derived from intact N-glycopeptides. Consequently, our approach demonstrated discriminative power in organ classification, attributable to the biological specificity offered by N-glycopeptide-level information. Thus, tissue-specific N-glycosylation patterns are distinct to serve as unique molecular fingerprints. Model performance of four supervised machine learning methods. a F1 scores of decision tree, neural network, random forest, and support vector machines (SVM), data are presented as mean values ± SD, statistics are derived from n = 50 independent iterations of model training and evaluation; b AUC curves for the four models and the most frequent class model; c F1 scores for each organ in four models; d AUC curves of each organ in four models. Multi-software N-glycoproteomics analysis Besides gPeptide, three other intact N-glycopeptide software tools are adopted to search the raw LC-MS/MS datasets of this study to provide complementary and comprehensive identification (Fig. 9a). Fragpipe is characterized by a mass offset-based approach, while the other three tools employ Y ion-based identification strategies. Glyco-Decipher is notable for its glycan database-independent search algorithm. gPeptide is recognized for its structural diagnostic capabilities, and N-glycan sequence structures and isomers are identified and confirmed with structure-diagnostic fragment ions. Evaluation of four common N-glycoproteome identification tools of Glyco-Decipher, pGlyco, FragPipe, and gPeptide. a Common and unique features; b overlap of intact N-glycopeptide identification at the N-glycan monosaccharide composition level. All search results are rigorously filtered to ensure methodological consistency (Supplementary Data 7). First, adhering to a one identification from one spectrum, only the highest-scoring ID is retained for each mass spectrum. Second, all identifications featuring post-glycosylation modifications (PGMs) are excluded. With the aforementioned criteria, gPeptide, Glyco-Decipher, Fragpipe, and pGlyco identified 16,563, 13,515, 15,985, and 14,880 intact N-glycopeptides, with 7935, 6592, 6377 and 3601 unique IDs, respectively. A total of 3496 intact N-glycopeptides are consistently identified by all four software tools, which is the fifth intersection of all the overlaps (Fig. 9b). The multi-software approach provides a comprehensive landscape of potential glycoforms, capturing a broader spectrum of the N-glycoproteome. Discussion Here, we publicly present an intact N-glycopeptide dataset covering 219 runs of 74 subjects from 18 different human organ tissues, with site- and structural-specific identification. This is of certain significance in promoting the advancement of human post-translational modification proteomics. In-depth data analysis is performed using HPLC-MS/MS. The identification results from gPeptide are cross-validated with a multi-software search workflow and a non-targeted N-glycome profiling tool. Our systematic comparative analysis framework represents an advance in understanding tissue N-glycosylation heterogeneity. By examining three N-glycan types, four core structures, 12 branching patterns, and three fucosylation types, we uncover distinct clustering patterns among different tissues. Besides, the co-occurrence of sialylation and fucosylation modifications exists in some specific organs, suggesting coordinated regulation of these modifications in certain biological processes. These findings provide insights into tissue-specific N-glycosylation regulation mechanisms that were previously unclear. Supported by four machine learning models, our data reveals that the N-glycoproteome maintains intra-organ homogeneity despite significant inter-individual heterogeneity. This lays the groundwork for future comparative PTM-omics, where classification performance can be benchmarked across species to explore the evolutionary conservation of glycosylation machineries. Besides, we performed a comprehensive multi-software evaluation to capture a broader landscape of the human N-glycoproteome. By implementing standardized database searching and ID mapping methods, we enabled systematic comparison across different analytical platforms. Approximately 21.1% of intact N-glycopeptides being consistently identified across all platforms. This intersection highlights the highly divergent scoring algorithms and spectral interpretation strategies currently employed by different glycoproteomics search engines. While gPeptide achieved superior performance in spectral interpretation, there is a lack of tools like peptideshaker39 capable of global FDR rescoring and result integration at present. While our study provides insights into tissue-specific N-glycosylation patterns, several limitations should be noted. First, the output of 3′ and 6′ glycans is based on the matching conditions in the database, which remain challenging to differentiate using the current LC-MS/MS-based pipeline. Besides, since the antenna fucose migration during HCD fragmentation, the position of fucoses requires more cross-validation40. The unique IDs from each software should be interpreted with caution, as not all possibilities have been taken into account, such as excessive acetylation of the N-terminal/methionine, amide adducts, and so on41. At last, the current analytical workflow may not capture all possible N-glycans and PGMs, particularly those present at very low abundances. Looking forward, this comprehensive human organ N-glycoproteome atlas serves as a basic resource for the glycobiology community. Our findings lay the groundwork for understanding how different tissues maintain their unique glycobiological identities and how these differences might contribute to organ-specific functions. The analytical framework we developed can be readily applied to investigate N-glycosylation changes in pathological conditions, potentially leading to the identification of disease-specific N-glycosylation signatures. Methods Ethical statement The experiments were performed in accordance with the Code of Ethics of the World Medical Association (Declaration of Helsinki) and the guidelines of Nanjing University and approved by the ethics committee of Nanjing University Medical School (2202-SR-598). Sex and gender were considered based on SAGER guidelines. The study included a total of 74 subjects consisted of 45 males and 29 females. The participants’ ages ranged from 32 to 66 years old (median age: 48). The requirement for specific informed consent for this study was waived. This waiver was granted because (i) all participants had previously signed a broad informed consent form upon admission or during the diagnostic procedure, explicitly agreeing that their residual clinical specimens could be utilized for future medical research; and (ii) the research utilized fully de-identified and anonymized residual samples post-pathological examination, posing no more than minimal risk to the subjects and involving no sensitive personal or genetic information. The study procedures strictly adhere to the current ethical regulations and guidelines in China. Preparation of human organ samples All samples were frozen in liquid nitrogen and stored at −80 °C after collection to preserve their current state. The colon, intestine, esophagus, stomach, and urinary bladder were thoroughly washed with PBS to eliminate any fecal and food residues. Other tissues were also washed repeatedly with 1×PBS until all blood were removed. Sample preparation of intact N-glycopeptides About a 50–100 mg sample was thawed on ice. After cutting the tissue into pieces while keeping it on ice, the sample was washed by pre-cold 1×PBS several times to avoid blood residue. Lysis buffer (0.1 M Tris/HCl, 4% SDS, pH 8.0) and protease inhibitor (1% v/v) were added. The sample was then homogenized for 1 min at 8000 × g and incubated on ice for 30 min to obtain lysates. Lysates were centrifuged at 14,000 × g for 30 min at 4 °C, and the supernatant was then mixed with cold acetone (1:5, v/v) at −20 °C for 3 h in order to obtain protein precipitation. After redissolving with 8 M urea, protein concentrations were determined by BCA assay. One milligram proteins were reduced in 10 mM DTT for 20 min at 55 °C, alkylated in 20 mM IAA for 30 min at RT in the dark; DTT was then added to a final concentration of 10 mM to terminate the reaction. Samples were diluted with 50 mM ABC and digested with trypsin at a 1:50 enzyme-to-substrate ratio overnight at 37 °C. Tryptic peptides were desalted by C18 (Phenomenex Jupiter, 15 μm, 300 Å) solid phase extraction (SPE) columns and dried in a SpeedVac (Thermo Scientific). The ZIC-HILIC (Merck, 5 μm, 200 Å) SPE column was equilibrated three times with 300 μL of 80% ACN, 1% TFA (v/v). Tryptic peptides were redissolved in 300 μL 80% ACN, 0.1% TFA (v/v) and loaded on the ZIC-HILIC column for five times; after wash with 80%ACN, 0.1% TFA (v/v), intact glycopeptides were eluted in 300 μL 0.1% TFA and 200 μL 50 mM ABC. The eluants were combined and dried in the SpeedVac. C18-RPLC-MS/MS (HCD) analysis of intact N-glycopeptides For LC-MS/MS analysis, a Q-Exactive mass spectrometer (Thermo Fisher Scientific) equipped with an Ultimate 3000 nano-LC system (Thermo Fisher Scientific) was used. For LC separation, a homemade 70-cm-long C18 analytical column (360 μm o.d. × 75 μm i.d.) was protected by a homemade 5-cm-long trap column (360 μm o.d. × 200 μm i.d.), which were both packed with Phenomenex Jupiter C18 resin (particle size 5 μm, pore size 300 Å). The mobile phase flow rate was 300 nL/min, consisting of 0.2% formic acid in water (Buffer A) and 0.2% formic acid in 95% acetonitrile (Buffer B). The sample was trapped in 2% Buffer B for 25 min and separated at Buffer B from 2 to 40% for 135 min. Buffer B was then changed to 95% over 10 min for flushing posts and held on it for 5 min. At last Buffer B was changed to 2% over 10 min and re-equilibrated for 10 min. The eluted N-glycopeptides were ionized under 2.8 KV, ion transfer capillary temperature was 250 °C. The mass range of MS1 was 700–2000 with a resolution of 70,000. The automatic gain control (AGC) target was 2e5, and the maximum injection time was 50 ms. The top 20 ions were selected for higher-energy collisional dissociation (HCD) with stepped normalized collision energy of 20, 30, and 30%. The resolution of MS2 was 17,500, and the isolation window was 3 m/z. The AGC target and maximum injection time were 5e5 and 250 ms. Three technical replicates were performed for each sample. Identification and quantification of intact N-glycopeptides with gPeptide and skyline A comprehensive glycoproteomic analysis was performed utilizing a custom-built glycan database comprising 6709 linkages derived from 603 distinct compositions in gPeptide format. MS raw data from three technical replicates per sample were searched against the UniProt Homo sapiens reference proteome (version 2023-07-28, containing 20,423 protein sequences) and the glycan database. Searches were performed using trypsin as the proteolytic enzyme, allowing up to two missed cleavages. Minimum peptide length was 6. Mass tolerance was set to 20 ppm for both precursor and fragment ions. Other parameters included isotopic peak abundance cutoff ≥40%, isotopic peak abundance deviation ≥50%, percentage of matched peptide backbone product ions (p-PMPs) ≥10%, matched peptide backbone product ions (MPs) ≥5, isotopic peaks (IPs coverage) ≥10%, abundance coverage ≥10%, matched fragment ions of the N-glycan moiety ≥1, TopN hits N = 2 (top1 hits have the lowest p score), Y1 ions Top4. Spectrum-level FDR was controlled at 1% using the target-decoy approach. The quantitative extracted ion chromatogram of each ID was extracted by Skyline42,43. The conversion of glycan databases and the search of Glyco-Decipher, pGlyco, and FragPipe The glycan database preparation followed a comprehensive approach. A linkage-based and custom-built glycan database of gPeptide was converted to pGlyco 3.1 format through a custom Python script. Additionally, a composition-based glycan database was converted to accommodate both FragPipe and Glyco-Decipher analysis platforms using Microsoft Excel. To ensure consistency across all glycoproteomics analyses, the same protein database and common parameters were utilized throughout all software platforms. Glyco-Decipher-v1.0.444, pGlyco3.145, and FragPipe-22.046–49 were used. It is worth noting that the search expansion of Glyco-Decipher does not depend on the glycan database. In FragPipe, a workflow based on “glyco-N-HCD” is used. PTM Shepherd was enabled, using the “Glyco search” parameter. Glycan FDR and peptide FDR were both controlled at 1%. All search results were filtered into one ID per spectrum. Python was used to edit the script and unify all IDs into the format of “Access_Peptide_Glycosite_Composition,” such as “P43308_SAPASNVSHTVVLR_6_HexNAc(2)Hex(8)” for subsequent analyses. Data process and model training The data processing flow of this study has been presented in Supplementary Fig. 36. An expression matrix was consolidated from all 219 raw files. IDs identified in only a single organ were filtered out from the matrix. The data then underwent sample total abundance normalization. For some analysis, the total abundance-normalized matrix after log2-transformed was used. To handle missing values, IDs with identifications in ≥85 raw files were subjected to KNN imputation, with the same organ and the same individual carrying greater weight and missing values for which insufficient neighbors were available were filled using the minimum value. Batch effect correction was subsequently performed on the imputed data using ComBat. The resulting batch-corrected matrix was used for model training. Data analysis and data visualization are implemented through Python 3.9.13 or R 4.4.1. softwares. Packages “tidyverse,” “factoextra,” “tidyr,” and “dplyr” were used for data analysis. “RColorBrewer,” “ComplexUpset,” “ggplot2,” “pheatmap,” “ComplexHeatmap,” “svglite,” and “motifStack” were used for data visualization. The grid-based parameter optimization has been applied to decision trees and neural networks. To evaluate model performance, the dataset was randomly split into training (70%) and testing (30%) sets using stratified random sampling based on organ class labels, ensuring that the class distribution remained consistent in both subsets. The operation time of this model can be completed within several minutes. The model was trained repeatedly, 50 times, to ensure its stability. Baseline comparisons to trivial models of 1-nearest neighbor, and the most frequent class are provided in the code. Ablation experiments were conducted on four different models respectively. Data analysis and machine learning modeling were implemented through Python. Package “scikit-learn” (v1.3.0) was used for machine learning pipelines, feature selection, and model evaluation. “pandas” and “numpy” were utilized for data manipulation and numerical computations. “matplotlib” was employed for data visualization and ROC curve generation. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1–7 Reporting Summary Transparent Peer Review File Source data Source Data 1 Source Data 2 Author contributions Z.X.T. conceived the study and revised the manuscript; X.Y.H. did the experiments and wrote the draft; X.Y.H. and S.D.Q. did the data analysis; H.W.L., T.W., C.Q., and Y.G.Y. collected the clinical samples. Peer review Peer review information Nature Communications thanks Shin-Ichiro Nishimura, Juan Rojas, Philip Toukach, and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Funding X.Y.H., T.W., C.Q., Y.G.Y., S.D.Q., H.W.L., and Z.X.T. disclose support for the research of this work from the National Science Foundation of China (22074105) and Shanghai Science and Technology Commission (14DZ2261100). Data availability The raw LC-MS/MS datasets from analysis of intact N-glycopeptides enriched from the 18 human organs/tissues have been uploaded to the Pride (https://www.ebi.ac.uk/pride/) under accession code PXD070547. Source data are provided with this paper. Unless otherwise stated, all data supporting the results of this study can be found in the article, supplementary, and Source data files. Source data are provided with this paper. Code availability Data analysis and data visualization were used with Python 3.9.13 or R 4.4.1. softwares. This code can be obtained as needed and has been stored in the GitHub repository (https://github.com/HuXiaoyu-TJU/A-comprehensive-landscape-of-human-organ-N-glycoproteome). We have obtained a (10.5281/zenodo.20181185) for the GitHub repository on Zenodo (https://zenodo.org/records/20181185). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Hongwei Liang, Email: hwliang@nju.edu.cn. Zhixin Tian, Email: zhixintian@tongji.edu.cn. Supplementary information The online version contains Supplementary material available at 10.1038/s41467-026-74631-7. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Data 1–7 Reporting Summary Transparent Peer Review File Source Data 1 Source Data 2 Data Availability Statement The raw LC-MS/MS datasets from analysis of intact N-glycopeptides enriched from the 18 human organs/tissues have been uploaded to the Pride (https://www.ebi.ac.uk/pride/) under accession code PXD070547. Source data are provided with this paper. Unless otherwise stated, all data supporting the results of this study can be found in the article, supplementary, and Source data files. Source data are provided with this paper. Data analysis and data visualization were used with Python 3.9.13 or R 4.4.1. softwares. This code can be obtained as needed and has been stored in the GitHub repository (https://github.com/HuXiaoyu-TJU/A-comprehensive-landscape-of-human-organ-N-glycoproteome). We have obtained a (10.5281/zenodo.20181185) for the GitHub repository on Zenodo (https://zenodo.org/records/20181185).