SweetOrigins: Extracting Evolutionary Information from Glycans

Abstract

Glycans, the most diverse biopolymer and crucial for many biological processes, are shaped by evolutionary pressures stemming in particular from host-pathogen interactions. While this positions glycans as being essential for understanding and targeting host-pathogen interactions, their considerable diversity and a lack of methods has hitherto stymied progress in leveraging their predictive potential. Here, we utilize a curated dataset of 12,674 glycans from 1,726 species to develop and apply machine learning methods to extract evolutionary information from glycans. Our deep learning-based language model SweetOrigins provides evolution-informed glycan representations that we utilize to discover and investigate motifs used for molecular mimicrymediated immune evasion by commensals and pathogens. Novel glycan alignment methods enable us to identify and contextualize virulence-determining motifs in the capsular polysaccharide of Staphylococcus aureus and Acinetobacter baumannii. Further, we show that glycan-based phylogenetic trees contain most of the information present in traditional 16S rRNA-based phylogenies and improve on the differentiation of genetically closely related but phenotypically divergent species, such as Bacillus cereus and Bacillus anthracis. Leveraging the evolutionary information inherent in glycans with machine learning methodology is poised to provide furthercritically needed -insights into host-pathogen interactions, sequence-to-function relationships, and the major influence of glycans on phenotypic plasticity.In contrast to RNA and proteins, whose sequences can be elucidated from their DNA sequence, glycans are the only biological polymer that falls outside the rules of the central dogma of molecular biology. Glycans are present as modifications on all other biopolymers 1 , exerting varying effects on biomolecules, including stabilization and modulation of their functionality 2,3 . Apart from influencing the function of individual proteins, glycans are also crucial for cell-cell contact 4 and mediate essential developmental processes 5 . Although glycans are synthesized by specific, DNA-encoded enzymes 6 , an individual glycan sequence is dependent on an intricate interplay between multiple enzymes and cellular conditions, providing glycans with important roles in phenotypic plasticity. Different glycosylation sites, even on the same protein, can exhibit considerable glycoform heterogeneity, depending on accessibility and glycosyltransferase kinetics, as has been shown for protein disulfide isomerase 7 . The glycan alphabet exceeds 1,000 monomers, allowing for an astronomical number of potential oligosaccharides built with different monosaccharides, lengths, connectivity, and branching.

Using graph convolutional neural networks to learn a representation for glycans

“…The nonlinear branching structure of glycans, together with their diversity, has hitherto presented an obstacle to the development of machine learning models for glycobiology that fully capitalized on the rich information in glycan sequences. The use of a glycoword-based language model overcame some of these limitations, allowing for the prediction of glycan immunogenicity, pathogenicity, or taxonomic class (Bojar et al, 2020a(Bojar et al, , 2020b(Bojar et al, , 2021; data augmentation inspired by graph isomorphism further improved predictions (Bojar et al, 2021). This led us to consider whether the structure of glycans as graphs or trees could be better captured by neural network architectures specifically developed for modeling graphs.…”

“…Computational approaches to analyzing glycans are mostly limited to counting the occurrence of curated sequence motifs and using this information as input for models predicting glycan properties (Bao et al, 2019;Coff et al, 2020). Recently, deep learning has been applied to the analysis of glycan sequences, creating glycan language models based on recurrent neural networks (Bojar et al, 2020a(Bojar et al, , 2020b(Bojar et al, , 2021. The glycan language model SweetTalk views glycans as a sequence of ''glycowords'' (subsequences that describe structural contexts of a glycan) and was used to predict the taxonomic class of glycans as well as their properties, such as immunogenicity or contribution to pathogenicity.…”

GlyNet: A Multi-Task Neural Network for Predicting Protein-Glycan Interactions

“…The ideal way to compare different systems is to run each on the same input and obtain their respective predictions, then compare their outputs with scores such as AUC, R-squares and mean square errors. Here we compare GlyNet with CCARL by Coff et al 30 and SweetTalk by Bojar et al [31][32][33][34] .…”

“…Previous machine learning (ML) approaches to protein-glycan interaction employed techniques like support vector machines 17,18 , graph kernels 18 , modularity optimization methods 19 , and Markov models [20][21][22] , to identify glycan motifs, substructures that specific proteins recognize-for reviews see Mamitsuka 23 , Haab 24 , and Sese 25 . Several publications from Aoki-Kinoshita and co-workers 20,22,[26][27][28][29] , Coff and co-workers 30 , Cummings and co-workers 14 , and recently Bojar and co-workers [31][32][33][34] focus on ML classification models that predict qualitative features (think "strong vs. weak" interactions) for each glycan-protein pair. Woods and co-workers combined molecular mechanics, automated 3D glycan structure generation and docking techniques to produce computational carbohydrate grafting that can qualitatively predict binding between a carbohydrate fragment and a known 3D protein structure 35 .…”

“…Dahl 39 , Ma 40 , and Jensen et al 41 pioneered the use of deep NN to predict physicochemical properties of organic molecules. Neural networks have also been used to predict patterns of enzymatic processing on glycopeptides 38 , taxonomic classifications of organisms synthesizing glycans 33 , and biological properties like immunogenicity of glycans 32,34 . At the time of the preparation of this manuscript, Bojar and coworkers released a bioRxiv manuscript describing the first graph neural network (GNN) that predicts properties of glycan graphs such as evolutionary origin, immunogenicity and recognition by viral proteins 34 .…”