Correcting for sparsity and non-independence in glycomic data through a systems biology framework

Abstract

23Glycans are fundamental cellular building blocks, involved in many organismal functions. 24Advances in glycomics are elucidating the roles of glycans, but it remains challenging to 25 properly analyze large glycomics datasets, since the data are sparse (each sample often has only a 26 few measured glycans) and detected glycans are non-independent (sharing many intermediate 27 biosynthetic steps). We address these challenges with GlyCompare, a glycomic data analysis 28 approach that leverages shared biosynthetic pathway intermediates to correct for sparsity and 29 non-independence in glycomics. Specifically, quantities of measured glycans are propagated to 30 intermediate glycan substructures, which enables direct comparison of different glycoprofiles 31 and increases statistical power. Using GlyCompare, we studied diverse N-glycan profiles from 32 glycoengineered erythropoietin. We obtained biologically meaningful clustering of mutant cell 33 glycoprofiles and identified knockout-specific effects of fucosyltransferase mutants on tetra-34 antennary structures. We further analyzed human milk oligosaccharide profiles and identified 35 novel impacts that the mother's secretor-status on fucosylation and sialylation. Our substructure-36 oriented approach will enable researchers to take full advantage of the growing power and size of 37 glycomics data. 38 39 40 the rapid generation of many glycoprofiles with detailed glycan composition 7-10 , exposing the 54 complex and heterogeneous glycosylation patterns on lipids and proteins 11,12 . Large glycoprofile 55 datasets and supporting databases are also emerging, including GlyTouCan 13 , UnicarbDB 14 , 56GlyGen and UniCarbKB 15 . 57These new technologies and databases provide opportunities to examine global trends in 58 glycan function and their association with disease. However, the rapid and accurate comparison 59 of glycoprofiles can be challenging with the size, sparsity and heterogeneity of such datasets. 60Indeed, in any one glycoprofile, only a few glycans may be detected among the thousands of 61 possible glycans 16 . Thus, if there is a major perturbation to glycosylation in a dataset, few 62 glycans, if any, may overlap between samples. However, these non-overlapping glycans may 63 4 only differ in their synthesis by as few as one enzymatic step. Thus, it can be difficult to know 64 which glycans to compare. Furthermore, since glycans often share substantial portions of their 65 biosynthetic pathways with each other, statistical methods that assume independence (e.g., t-66 tests, ANOVA, etc) are inappropriate for glycomics. Here we address these challenges by 67 proposing glycan substructures, or intermediates, as the appropriate functional units for 68 meaningful glycoprofile comparisons, since each substructure can capture one step in the 69 complex process of glycan synthesis. Thus, using substructures for comparison, we account for 70 the shared dependencies across glycans. 71Previous work has investigated the similarity across glycans using glycan motifs, ...

Elucidating Human Milk Oligosaccharide biosynthetic genes through network-based multiomics integration

“…SP1, EGR1, ETS1, ETV4 and ERG are all predicted to positively influence expression associated with the biosynthetically related HMOs: 3'SL, 3FL, LSTb and DSLNT; 3'SL and 3FL share a common substrate (lactose) while LSTb is a likely precursor to DSLNT. The motif-level analysis showed opposing regulation between IKZF1: upregulating gene expression signatures associated with the 3'SL and LSTb substructure abundance 17 (X34 and X62 respectively, see Figure S 19) and downregulating gene expression associated with GlcNAC-lactose, LNT and LNFPI substructure abundance (X18, X40 and X65 respectively, see Figure S 19), while EGR1, ERG and ETS1 have the opposite predicted impact (Figure S 17). The motif-level predictions are consistent with the HMO-level predictions of upregulation on 3'SL and LSTb while adding an additional point of contrast.…”

“…Absolute and relative concentrations of the 16 most abundant HMOs was measured. Starting from a scaffold of all possible reactions [13][14][15][16][17][18] , we used constraint-based modeling 19,20 to reduce the network to a set of relevant reactions and most plausible HMO structures when not known to form the basis for a mechanistic model 13,14,22 . This resulted in a ranked ensemble of candidate biosynthetic pathway topologies.…”

“…IPA discovered 57 TFs significantly (|z|≥3; p < 0.001) associated with the 16 HMO-specific gene expression signatures. We performed differential expression on HMO substructure abundance and substructure abundance ratios 17 ; IPA found 66 and 49 TFs significantly (|z|≥3; p < 0.001)) associated with HMO substructure and substructure ratio specific gene expression signatures. Using MEME, we identified three putative TF regulatory sites (TF motifs I, II and III) for 6 selected glycosyltransferases responsible for the HMO biosynthesis (Table 2 and Figure S 15).…”

“…Technicians were blinded to sample metadata. HMO composition and abundance measurement for cohort 1 were fully described in 17 . Measurements for cohort 2 are previously unpublished and used the same methodology.…”

“…The differential expression analysis was conducted on three different datasets: 1) 16 different HMOs (2'FL, 3'SL, 3FL, FLNH, LNT, LNnT, LSTb, LNFP-III, LNFP-II, LNFP-I, DFLNT, LSTc, DSLNT, FDSLNH, DSLNH, DFLNH), 2) 19 glycan motifs (X18, X32, X34, X35, X37, X40, X62, X63, X64, X65, X66, X94, X106, X113, X120, X127, X141, X142, X143, see Figure S 19), and 3) 4 differential motifs for the difference ("conversion rate") between related motifs (X65-X40, X106-X62, X63-X37, X62-X40, see Figure S 19). Substructure abundance for glycan motifs and conversion ratios were computed using Glycompare v1 17 . The gene expression data were downloaded from the Gene Expression Omnibus 113 (GSE36936).…”