Loss and gain of N-linked glycosylation sequons due to single-nucleotide variation in cancer

Despite availability of sequence site-specific information resulting from years of sequencing and sequence feature curation, there have been few efforts to integrate and annotate this information. In this study, we update the number of human N-linked glycosylation sequons (NLGs), and we investigate cancer-relatedness of glycosylation-impacting somatic nonsynonymous single-nucleotide variation (nsSNV) by mapping human NLGs to cancer variation data and reporting the expected loss or gain of glycosylation sequon. We find 75.8% of all human proteins have at least one NLG for a total of 59,341 unique NLGs (includes predicted and experimentally validated). Only 27.4% of all NLGs are experimentally validated sites on 4,412 glycoproteins. With respect to cancer, 8,895 somatic-only nsSNVs abolish NLGs in 5,204 proteins and 12,939 somatic-only nsSNVs create NLGs in 7,356 proteins in cancer samples. nsSNVs causing loss of 24 NLGs on 23 glycoproteins and nsSNVs creating 41 NLGs on 40 glycoproteins are identified in three or more cancers. Of all identified cancer somatic variants causing potential loss or gain of glycosylation, only 36 have previously known disease associations. Although this work is computational, it builds on existing genomics and glycobiology research to promote identification and rank potential cancer nsSNV biomarkers for experimental validation.


Results
Comprehensive collection of real and predicted NLGs. Overview of human NLGs. Table 1 summarizes the findings from each of the three methods of NLG identification: high-confidence NLGs (those reported in databases with validated evidence or manual assertion), predicted NLGs by NetNGlyc (http://www.cbs.dtu.dk/ services/NetNGlyc/), and string search of all NX(S/T) (X! = P) sequons. Proteins predicted to contain signal peptides or annotated with the UniProt keyword(s) "Secreted" or "Membrane" were used for analysis of disease-relatedness to focus on those entries with the greatest potential of being biologically viable biomarkers. Of the 59,341 identified non-redundant NLGs from 15,318 proteins (See Supplemental Table S1), 7,017 of these proteins either have signal peptides or are annotated with one or both keyword(s) "Secreted" or "Membrane, " prioritizing this subset for consideration as biologically viable markers. Note that of the 15,318 total proteins identified to SCienTifiC RePoRtS | (2018) 8:4322 | DOI: 10.1038/s41598-018-22345-2 contain NLGs and the subset of 7,017 likely viable N-linked glycosylation sites identified by functional annotations, 15,314 and 7,014, respectively, were retrievable by the string search method alone. Additional information about these NLG-containing sequences, including sequence length, existence and positions of signal peptides, "Cellular component" keywords, and more, can be viewed in Supplemental Table S1.
Frequency of NLGs and experimentally validated sites in human proteins. Within the UniProtKB/Swiss-Prot dataset, there are 15,318 of 20,199 human proteins, or 75.8%, of all human proteins with at least one NLG. Of the 59,341 NLGs on these human proteins, for an average of 3.9 per protein, 16,253 of 21,956, or 74.03% of total NLGs from 4,412 glycoproteins are well-characterized N-linked glycosylation sites. In this context, the term glycoprotein is applied based on the observation of at least one occupied NLG. Among the high-confidence sites, 2,511 NLGs are experimentally verified with evidence from publications and 13,500 are manually asserted by UniProtKB/Swiss-Prot curators. 232 additional NLGs were retrieved from other databases and are included in the table of potential NLGs (see Materials and Methods section). Figure 1 shows the relative contribution of sources to the identification of NLGs. Furthermore, 61 NLGs belonging to these high-confidence N-linked glycosylation sites are atypical in that the corresponding UniProtKB/Swiss-Prot FT line annotations do not follow the consensus NX(S/T) (X! = P) configuration. Therefore, although the current rate of experimentally validated human NLGs is 27.39%, the rate of experimentally validated NLGs in the subset of known glycoproteins is 74.03%, aligning with Apweiler's previous hypothesis that three quarters of glycoproteins should be N-linked.
Distribution and spacing of NLGs and experimentally validated N-linked glycosylation sites. The distribution of NLGs per protein is shown in Fig. 2 (for more information, see Supplemental Table S2). Most proteins have a small number of NLGs, but some proteins have large numbers of NLGs. The most extreme example of this case is Mucin-16 (MUC16, UniProtKB accession: Q8WXI7), which has 265 NLGs. As the largest cell-associated mucin, the length of MUC16 at 22,152 residues 56 increases the likelihood for multiple NLGs in the sequence. Mucins are a major component of mucus with a normal protective physiological role, but it has been shown that glycan  of sequons and proteins identified from three methods 3,452 out of 20,199 proteins in the  human proteome have signal peptides; 14,921 proteins in the human proteome have at least one "Cellular component" keyword ("Secreted, " "Membrane, " "Cytoplasm, " or "Nucleus"); the same protein can belong to multiple "Cellular component" categories if it is observed in more than one cellular location. 8,772 proteins have either signal peptides or are annotated with keyword(s) "Secreted" or "Membrane. " String search results include almost all the NLGs from the other two methods except for 121 atypical cases which do not follow consensus NX(S/T) (X! = P) according to high-confidence criterion, 61 reported from UniProt FT lines. There are 59,341 non-redundant NLGs from 15,318 proteins in total from these three methods. 7,017 of them either have signal peptides or are annotated with keyword(s) "Secreted" or "Membrane". a Annotated NLGs from UniProtKB/ Swiss-Prot, HPRD 9.0, dbPTM 3.0, neXtProt and NCBI-CDD were treated as high-confidence results. attachment to MUC16 is altered in response to oxidative stress in pancreatic cancer 57 and is a major carrier of altered sialylation characteristic of malignant conditions in serous ovarian tumors 58 . When normalized by length, MUC16 still shows a greater than average number of NLGs per unit length ( Table 2), but Sialomucin core protein 24 (CD164, Q04900) has the greatest density of NLGs at approximately one per every 22 residues, compared to the average of approximately one per every 118 residues. CD164 belongs to a class of heavily glycosylated proteins, the sialomucins, involved in regulation of cellular adhesion, proliferation, differentiation, and migration of hematopoietic stem cells 59 . Although its expression, not glycosylation, has been implicated in various cancers 60,61 , absence of proper terminal N-glycan attachment on CD164 protein has been observed to prevent functional interactions, disrupting post-endocytic sorting of the protein 59 . Table 2 demonstrates similar trends in the spacing of both total NLGs and verified N-linked glycosylation sites when averaged for all proteins. Identification of LOG and GOG resulting from variation. Mapping of nsSNVs to NLGs led to identification of variations that may lead to loss of glycosylation (LOG) and gain of glycosylation (GOG). In some cases, different variations can lead to abolition or creation of the same NLG, and proteins may contain multiple NLGs. Thus, we expect the number of variations to be greater than or equal to the number of affected sequons, which is expected to be greater than or equal to the number of affected proteins. Our observations follow this trend, as can be seen in Table 3 (additional information about LOG and GOG variations and corresponding sequons can be found in Supplemental Tables S3-S8).

Summary of collected variation data.
In this study, we report 16,253 high-confidence (experimentally validated or manually asserted) sites with 2,898 and 7,508 sequons abolished by somatic and germline variations, respectively. Although experimental validation of GOG is outside the scope of this study, we report 15,504 and 38,711 potentially induced sequons by somatic and germline variations, respectively. Furthermore, in this update we report that 3,930 of 4,412 high-confidence glycoproteins lose or gain NLGs due to germline or cancer somatic nsSNV compared to 1,091 human proteins previously reported to undergo LOG or GOG due to polymorphisms 63 . For all datasets, the second position in the sequon is the least subject to variations that would induce or abolish a sequon, as is expected due to the flexibility of residues at this position (can be any residue except P). Conversely, the first position requires an N, and is therefore expected to be the most subject to NLG-altering mutations. The third position, which can be occupied by either an S or a T, is neither as limited as the first position nor as flexible as the second, so we expect the frequency of sequon-altering variations to be between the corresponding frequencies for the first and second positions. When we average across all datasets, these trends do hold true; however, when we look at just the subset of somatic LOG mutations, we actually find that the third position is subject to a greater frequency of mutations than the first position (Fig. 3). The majority of proteins are annotated with smaller numbers of NLGs, and therefore the average density is less than 1 NLG per protein, when normalized by unit length. The distribution of NLGs per protein is plotted as the count of proteins with a given density of NLGs for (A) all NLGs in the human proteome, (B) LOG-causing NLGs in the somatic subset, and C) GOG-causing NLGs in the somatic subset.

Sequons
Real N-glycosylation sites  Tables S7-S10). If we only include high-confidence NLGs reported in databases, we find 13 LOG-causing somatic-only variations in 13 sequons of 12 proteins. Figure 4 shows a schematic of the localization of these proteins. Interestingly, considering only the subset of high-confidence NLGs, there are 13 somatic-only LOG variations associated with three or more cancer types each ( Table 4), two of which are associated with four or more cancer types each. Under these criteria, the most significant variations are Mast/stem cell growth factor receptor Kit (KIT, P10721) N486D and Thrombospondin type-1 domain-containing protein 7A (THSD7A, Q9UPZ6) T236M.
KIT is a transmembrane receptor tyrosine kinase commonly expressed in hematopoietic progenitor cells 64 . Inhibition of N-linked glycosylation of KIT has been reported to affect cellular signaling and cell-surface expression of KIT, inducing apoptosis in acute myeloid leukemia (AML) 65 , and glucose metabolism mediated by KIT in response to imatinib has been used to predict tumor sensitivity to the drug in gastrointestinal stromal tumors (GIST) 66 . Evidence suggests that stem cell factor, the activator of KIT's autophosphorylation domain, only  Table 3. Numbers of variations, affected sequons, and affected proteins from germline and cancer somatic LOG and GOG variation. a Variations from cancer genomics databases (somatic origin), but not in dbSNP (germline origin) b Sequons for GOG sets are reported by unique positions-Note that the number of unique motifs per position identified is equal to the total number of nsSNVs for that set. c GOG predicted by string search alone d Overlap between string search and NetNGlyc is 100% of NetNGlyc results.  . Schematic of biologically relevant LOG and GOG variations in at least three cancers. Red circles are proteins in the LOG dataset, green circles are proteins in the GOG dataset. Protein completely within the lipid bilayer are tagged with cellular localization term "Membrane" while proteins spanning both the membrane and the adjacent cytoplasm or extracellular environment are tagged both with cellular localization term "Membrane" and "Cytoplasm" or "Secreted, " respectively. Note that position within membrane is not indicative of status as integral or peripheral proteins. Dotted lines represent proteins that also have cellular localization term "Nucleus. " Black stars on proteins signify the presence of a signal peptide on that protein. Also, note that Mast stem cell growth factor receptor Kit, (KIT, P10721) is the only protein to appear with mutations that could cause loss or gain of glycosylation in more than three cancers each.  69 . While the lack of literature support for cancer involvement could indicate a lack of importance of THSD7A variation in cancer, this variant's presence in four different cancers reported across four cancer databases (as listed above), in conjunction with its link to CFS, warrants additional examination. Thus, THSD7A represents a prime candidate for downstream validation of cancer involvement.
Because newly created NLGs are not expected to be experimentally verified glycosylation sites described in publications, we limit the scope of potential GOGs by relevant cellular compartment keyword and signal peptide annotation. Under these criteria, there are 41 somatic-only GOG-causing variants related to three or more cancers each ( Table 5), six of which are related to four or more cancers each, and two related to five or more cancers each. These top two variations are Neurogenic locus notch homolog protein 1 (NOTCH1, P46531) A465T and Mucin-2 (MUC2, Q02817) T1750N.
NOTCH1 is a member of the highly conserved Notch signaling pathway, responsible for cell fate during development and homeostasis 70,71 . Extensive N-and O-linked glycosylation has been observed in the Notch extracellular domain 71 , where ligand binding induces conformational change and functional receptor activation 72 . A variety of roles have been described for Notch with respect to carcinogenesis and cancer stem cells 73  Mucin-2 is known to be densely O-glycosylated 74,75 , and has been observed with downregulated expression in colorectal cancer 76 . In fact, expression of MUC2 has been associated with less aggressive forms of urothelial bladder cancer 28 . While there are many studies regarding MUC2 expression in cancer 77  However, this protein is annotated to be localized mostly to the cytoplasm or nucleus and does not have a signal peptide, implying that glycosylation is unlikely. Low-density lipoprotein receptor-related protein 1B (LRP1B, Q9NZR2) is a membrane protein with a signal peptide and has been observed to have recurrent mutations in urachal cancer 79 and ovarian clear cell carcinoma 80 . LRP1B is commonly deleted across cancers 81 and it has been proposed to be a tumor suppressor via modulation of the extracellular tumor environment in thyroid cancer cells 82 . In our dataset, LRP1B has 24 predicted somatic-only GOGs affecting 23 positions. Although no literature exists regarding glycosylation of LRP1B in cancer, its known participation in multiple cancers suggests that altered glycosylation of the protein could affect the course of the disease.  Significant proteins. If we divide the number of affected sequons per protein by the length of the corresponding protein, we can identify the proteins with the highest density of nsSNV-affected NLGs per unit length. Considering only the proteins that are annotated with keywords "secreted" or "membrane, " in the LOG set, the top-ranked protein by density is Carcinoembryonic antigen-related cell adhesion molecule 7 (CEACAM7, Q14002) with five mutations abolishing sequons at five distinct positions. CEACAM7 belongs to the immunoglobulin superfamily, specifically to the carcinoembryonic antigen (CEA) gene family, and normally functions in regulation of cellular differentiation 83 .
UniProtKB reports CEACAM7 to be strongly downregulated in colonic adenocarcinomas, and it has been reported to be a predictive marker for rectal cancer recurrence 84 . Among the five somatic-only variations leading to LOG on CEACAM7 three are observed in skin cancer, the other two in prostate and pharynx cancers, independently.
Although this protein only has five LOG variants, it is a relatively small protein (265 residues in length) resulting in a density of one LOG per every 53 residues. Another CEA family gene, Carcinoembryonic antigen-related cell adhesion molecule 5 (CEACAM 5, P06731), also has one of the highest densities of somatic-only LOG variants. CEACAM5 has 12 variants occurring at 11 positions, for a density of one LOG per every 64 residues. Six of the variants are observed in skin cancer, with others reported in colon, kidney, uterine, ovarian, and hematologic cancers.
In the somatic-only GOG set, the top-ranked protein by density is Cellular tumor antigen p53 (TP53, P04637), with 15 GOG variants for an average of one per every 27 residues. While mutation of TP53 is well-characterized to interrupt its normal transcription factor function and lead to the development of many different cancers 85 , its cytoplasmic and nuclear localizations, as well as lack of signal peptide, make it an unlikely candidate for biologically viable N-glycosylation. Epididymal-specific lipocalin-8 (LCN8, Q6JVE9) is a secreted protein with a signal peptide with a density of one GOG variant for every 59 residues. Its three GOG variants are reported in liver and skin cancer datasets, but there is no current literature linking this gene to cancer.
Pathway and biomarker analysis summary. LOG/GOG-cancer associations. Of 8,894 distinct LOG somatic-only variants, only 20 were indicated in publications reported from HGMD (Human Gene Mutation Database, June 2016 release) 86 , only one of which is directly cancer-related (CREBBP Q92793 N1978D; ovarian cancer) and another nine with cancer-related syndromes. Similarly, of 12,939 distinct GOG somatic-only variants, only 16 overlap with HGMD disease annotations, none of which are directly cancer-related, and two of which confer increased risk of certain cancers. With respect to germline variants mapping to the LOG and GOG subsets, we see 349 LOG-associated germline variants and 436 GOG-associated germline variants mapping to HGMD mutations, out of a total of 40,365 HGMD mutations found in the entire nonsynonymous germline dataset retrieved from dbSNP. We hypothesize that mutations already associated with any disease (not restricted to cancer alone) identified in our somatic-only datasets could become high-value biomarkers for different cancer types based on existing literature evidence for functional disruption. At the same time, the minimal overlap between the HGMD dataset and our somatic-only findings suggests enormous potential for validation of as yet unpublished cancer-associated LOG/ GOG-causing variants. (See Supplemental Tables S7 and S8 for more information.) Enrichment analysis. The top five canonical MetaCore ™ pathways enriched across the set of all cancer-associated NLG variations (including both LOG-and GOG-causing variants) include muscle contraction GPCRs in the regulation of smooth muscle tone (P = 2.120e-09), signal transduction mTORC2 downstream signaling (P = 1.058e-8), development regulation of epithelial-to-mesenchymal transition (EMT) (P = 1.477e-8), breast cancer (general schema) (P = 9.996e-8), and nociception nociceptin receptor signaling (P = 1.259e-7). Interestingly, protein folding and maturation POMC processing (P = 2.373e-10) is strongly enriched in the GOG dataset, including 19/30 associated pathway genes, but having only two genes in common with the corresponding LOG dataset and no additional genes unique to the LOG dataset. This implies a strong bias for this pathway in potential pathogenesis of cancer associated with GOG mutations. The top five pathways for the GOG somatic variants (considered independently of any potential LOG overlap, ranked by statistical significance) are cytoskeleton remodeling_ TGF, WNT and cytoskeletal remodeling (P = 4.857e-12), cytoskeleton remodeling_cytoskeleton remodeling (P = 4.778e-11), ovarian cancer (main signaling cascades) (P = 1.668e-10), cell adhesion_chemokines and adhesion (P = 9.465e-10), and development_regulation of cytoskeleton proteins in oligodendrocyte differentiation and myelination (P = 1.166e-9). Similarly, the top five LOG somatic variants include signal transduction_mTORC2  downstream signaling (P = 4.862e-11), development_regulation of epithelial-to-mesenchymal transition (EMT) (P = 1.985e-8), cell adhesion_histamine H1 receptor signaling in the interruption of cell barrier integrity (P = 2.192e-8), cell adhesion_cadherin-mediated cell adhesion (P = 4.214e-8), and some pathways of EMT in cancer cells (P = 1.002e-7).

Discussion
The primary aims of this study were to: (1) provide a comprehensive view of protein N-linked glycosylation sequons within the human genome; and (2) compare cancer-centric nsSNVs and germline SNPs with respect to their relative impact on protein N-linked glycosylation sequons. Despite the in silico nature of this work, the consideration of existing experimental literature and functional annotations as evidence supporting the likelihood of possible glycosylation at various sites will allow downstream users of these findings to prioritize sites for additional study and to provide experimental validation that is currently lacking. With this goal in mind, we aimed to delineate which sites were verified sites of glycosylation and which were not, further classifying un-verified sites as possible and probable based on the evidence available for each site. To this end, we provided a comprehensive survey of human NLGs including up-to-date values for: (1) the number of experimentally verified human N-linked glycosylation sites (16,253, or 27.4% of all NLGs); (2) the proportion of human proteins containing at least one NLG (75.8%, or 15,318 of 20,199 total proteins); and (3) the average number of NLGs per human protein (3.9). We further aimed to associate possible loss or gain of glycosylation with the development of cancer by looking at the subset of variants expected to cause loss or gain of an NLG in somatic-only cancer variant calls. While loss or creation of an NLG does not necessarily impact N-glycosylation, we used additional functional information (including presence of signal peptides and subcellular locations), existing literature regarding role in N-glycosylation, and structural modeling predictions to devise a high-confidence criterion for likelihood of biologically plausible glycosylation. Even if glycosylation has been observed at a given site, effects on so-called "normal" glycosylation at a single site does not guarantee a functional consequence. However, combining information regarding the pervasiveness of independent variants across multiple cancers with evidence supporting plausible glycosylation allows us to rank variants. We have provided lists of variants (both LOG and GOG) meeting at least one high-confidence requirement for glycosylation and appearing in at least three cancers (Supplemental Tables S9 and S10). We suggest prioritizing these variants for additional study including possible experimental validation of N-glycosylation. While the change of the glycosylation status itself may not be pathogenic, the complexity of N-glycosylation and its potential impact on a number of factors may warrant multi-faceted exploration to determine whether a possible association with cancer could be directly explained by simple variation, loss or gain of glycosylation, or downstream effects of altered glycosylation. We expect that these cancer-associated, high-confidence, LOG/GOG-causing variants may lead to biomarker development and help to better elucidate the role of glycosylation in carcinogenesis.

Materials and Methods
Data retrieval and integration. Somatic nsSNVs were collected from COSMIC (CosmicCompleteExport_ v73), IntOGen (Release-2014-12), ICGC (Data Release v0.10a), TCGA (Release-2015-01-27), ClinVar (ClinVarFullRelease_2015-02-05), and literature mining methods 87 ; germline SNPs were collected from dbSNP (Human Build 149). Cancer-related somatic nsSNVs were integrated using the previously described BioMuta workflow 23 for pan-cancer analysis such that all cancer types were mapped to DO terms by corresponding disease ontology IDs (DOIDs) 62 . Annotated sequence functional site data for N-linked glycosylation retrieved from UniProtKB/Swiss-Prot (release-2015_01), HPRD 9.0 88 , dbPTM 3.0 89 , neXtProt release 2017-01-21 90 and NCBI-CDD (2015_Jan) 91 were treated as high-confidence results. For the purpose of this paper, high-confidence NLGs are those which are annotated as a glycosylation site in a UniProt FT line associated with a PMID or manual assertion, or reported directly from the other databases listed above.
Identification of real and predicted human NLGs. The identification of all possible human NX(S/T) (X! = P) sequons was performed by three methods. The reference UniProtKB/Swiss-Prot human proteome was obtained in January 2015 from the UniProt ftp site at ftp://ftp.uniprot.org/pub/databases/uniprot/previous_ releases/release-2015_01/knowledgebase/knowledgebase2015_01.tar.gz, and annotated N-linked glycosylation sites were retrieved from this set of protein entries. Additionally, the NetNGlyc-1.0 [http://www.cbs.dtu.dk/services/NetNGlyc/] tool was used to predict possible N-linked glycosylation sites. Finally, all NX(S/T) (X! = P) sequons in the human proteome were identified by searching all human protein sequences using custom Python scripting.
Calculating frequency of human NLG occurrence and occupancy rate. The proportion of NLG-containing human proteins was reported as the number of human proteins in UniProtKB/Swiss-Prot with at least one NLG divided by the total number of human proteins in UniProtKB/Swiss-Prot. Frequency of NLGs was calculated as the total number of human NLGs divided by the total number of human proteins in UniProtKB/ Swiss-Prot. Occupancy was calculated as the number of real N-linked glycosylation sites divided by the number of human NLGs in UniProtKB/Swiss-Prot.
Mapping NLGs to variations and reporting of LOG/GOG. Each potential NLG was mapped to variation datasets and reported as LOG or GOG. An altered NLG was considered a LOG if the altered version had an N abolished at the first position, an S or T abolished at the third position, or a newly generated P at the second position. Following the rationale of our previous study 55 , variants that alter T to S (or S to T) at the third position were not considered to functionally affect NLGs. An altered NLG was considered a GOG if the altered version contained a newly generated N at the first position, a newly generated S or T at the third position, or an abolished SCienTifiC RePoRtS | (2018) 8:4322 | DOI:10.1038/s41598-018-22345-2 P within a normal NPS/T subsequence. Variants were also mapped to the human proteome and analyzed using NetNGlyc to predict additional GOGs. Figure 5 shows the workflow for the LOG/GOG identification and variation mapping processes. It should be noted that the NetNGlyc tool may predict a GOG by modification of a residue outside the sequon and may also predict GOG including P at the middle sequon position: the number of GOGs suggested as candidates for further analysis was based on overlap between simple mapping results and NetNGlyc predictions.

LOG and GOG identification and pan-cancer analysis. Significant variations, sequons, and proteins
were identified according to the mapping results for LOG and GOG, especially for variations observed in multiple types of cancer. To assess the potential enrichment of a given feature and discrepancy of occurrence of that feature between datasets, a binomial test was used as described in an earlier study 92 . For example, to determine the P-value for significance of abolished sequons, the total number of abolished sequons across all methods was used to determine the expected rate of abolished sequons; this rate was then compared to the observed rate of abolished sequons for a given method. To compare somatic and germline variations, we used the somatic-only nsSNVs and germline SNPs and excluded any overlapping variations between the two datasets. Numbers of associated cancer types were counted by summing distinct DOIDs annotated at the variant level. Note that DOID mapping in BioMuta was done only for the subset of cancer-related DOIDs (CDO slim) 62 . Since some samples are of a particular cellular subtype, they may automatically map to multiple terms designating tissue-level and cell-level specificity separately. HGMD comparison. Our set of cancer-related, NLG-impacting variants was compared to variants in the Human Gene Mutation Database (HGMD, HGMD_Professional_2016.2) 93 . Because HGMD includes only published gene lesions, any identified nsSNV-affected LOG/GOG not cross-referenced by HGMD represented possible novel findings. Enrichment analysis. Enrichment of pathways was analyzed with MetaCore ™ (https://portal.genego.com).
Data availability. All data generated or analyzed during this study are available as supplemental tables accompanying this publication and can be browsed and downloaded from https://hive.biochemistry.gwu.edu/ kbdata/view/loss_and_gain_of_n_linked_glycosylation_sequons_in_cancer by selecting tables with prefix "NLGPaper" from the dropdown menu. Figure 5. Flowchart of the identification of LOG and GOG. The complete human proteome was retrieved from UniProtKB/Swiss-Prot, and sequences of included proteins were analyzed by string search and by NetNGlyc to identify all potential NLGs. High-confidence annotations of NLGs were also retrieved from the specified databases and incorporated into the comprehensive NLG dataset. NLGs were then mapped to somatic nsSNVs reported by cancer genomics databases and germline variations reported by dbSNP. The impact of variation on NLGs was analyzed, and for the subset resulting in loss or gain of NLG (LOG and GOG, respectively), presence in cancer samples was reported.