Powerpoint

Resource

Deep-Learning Resources for Studying Glycan-

Mediated Host-Microbe Interactions

Graphical Abstract

Highlights

d Glycan-focused language models can be used for sequence-

to-function models

d Information in glycans predicts immunogenicity,

pathogenicity, and taxonomic origin

d Glycan alignments shed light into bacterial virulence

Bojar et al., 2021, Cell Host & Microbe 29, 132–144 January 13, 2021 ª 2020 The Author(s). Published by Elsevier In https://doi.org/10.1016/j.chom.2020.10.004

Authors

Daniel Bojar, Rani K. Powers,

Diogo M. Camacho, James J. Collins

Correspondence diogo.camacho@wyss.harvard.edu (D.M.C.), jimjc@mit.edu (J.J.C.)

In Brief

Bojar et al. present a workflow that

combines machine learning and

bioinformatics techniques to analyze the

prominent role of glycans in host-microbe

interactions. The herein developed

glycan-focused language models and

alignments allow for the prediction and

analysis of glycan immunogenicity,

association with pathogenicity, and

taxonomic classification.

c. ll

OPEN ACCESS

ll

Resource

Deep-Learning Resources for Studying Glycan-Mediated Host-Microbe Interactions Daniel Bojar,1,2 Rani K. Powers,1,2 Diogo M. Camacho,1,4,* and James J. Collins1,2,3,4,5,* 1Wyss Institute for Biologically Inspired Engineering, Harvard University, Boston, MA 02115, USA 2Department of Biological Engineering and Institute for Medical Engineering & Science, Massachusetts Institute of Technology, Cambridge,

MA 02139, USA 3Broad Institute of MIT and Harvard, Cambridge, MA 02142, USA 4These authors contributed equally 5Lead Contact

*Correspondence: diogo.camacho@wyss.harvard.edu (D.M.C.), jimjc@mit.edu (J.J.C.)

https://doi.org/10.1016/j.chom.2020.10.004

SUMMARY

Glycans, the most diverse biopolymer, are shaped by evolutionary pressures stemming from host-microbe interactions. Here, we present machine learning and bioinformatics methods to leverage the evolutionary in- formation present in glycans to gain insights into how pathogens and commensals interact with hosts. By using techniques from natural language processing, we develop deep-learning models for glycans that are trained on a curated dataset of 19,299 unique glycans and can be used to study and predict glycan functions. We show that these models can be utilized to predict glycan immunogenicity and the pathogenicity of bac- terial strains, as well as investigate glycan-mediated immune evasion via molecular mimicry. We also develop glycan-alignment methods and use these to analyze virulence-determining glycan motifs in the capsular polysaccharides of bacterial pathogens. These resources enable one to identify and study glycan motifs involved in immunogenicity, pathogenicity, molecular mimicry, and immune evasion, expanding our under- standing of host-microbe interactions.

INTRODUCTION

In contrast to RNA and proteins, whose sequences can be eluci-

dated from their associated DNA sequence, glycans are the only

biopolymer outside the rules of the central dogma of molecular

biology. Although glycans are synthesized by DNA-encoded en-

zymes (Lairson et al., 2008), an individual glycan sequence is

dependent on the interplay between multiple enzymes and

cellular conditions. Additionally, the expansive glycan alphabet

of hundreds of different monosaccharides allows for a large

number of potential oligosaccharides, built with different mono-

saccharides, lengths, connectivity, and branching. Glycans are

present as modifications on all other biopolymers (Varki, 2017),

exerting varying effects on biomolecules, including stabilization

and modulation of their functionality (Dekkers et al., 2017; Solá

and Griebenow, 2009). Apart from influencing the function of in-

dividual proteins, glycans are also crucial for cell-cell contact in

the case of glycan-glycan interactions during the attachment of

pathogenic bacteria to host cells (Day et al., 2015), and they

mediate essential developmental processes such as nervous

system development (Haltiwanger and Lowe, 2004). Recently,

Lauc et al. hypothesized that the plethora of available glycoforms

and their plasticity facilitated the evolution of complex multicel-

lular lifeforms (Lauc et al., 2014), reasoning that is supported

by the essential roles of glycans in developmental processes

132 Cell Host & Microbe 29, 132–144, January 13, 2021 ª 2020 The This is an open access article under the CC BY-NC-ND license (http://

and cell-cell communication and emphasizes the evolutionary

information in glycans.

Because glycans make up the outermost layer of both eukary-

otic and prokaryotic cells, cross-kingdom interactions will

necessarily involve these molecules (Day et al., 2015). The prom-

inent role of glycans in host-pathogen interactions (Varki, 2017)

has resulted in evolutionary pressures and opportunities on

both sides of the interaction—natural selection can modify

host glycan receptors used by pathogens without losing their

functionalities, whereas pathogens and commensals need to

alter their glycans to evade the host immune system. These inter-

actions provide a window into understanding glycan-mediated

host-microbe relationships. Glycans display great phenotypic

variability: sequences can be changed depending on environ-

mental conditions, such as the level of extracellular metabolites

(Park et al., 2017), without the need for genetic mutations, poten-

tially facilitating rapid responses to changes in host-microbe

relationships.

Given the aforementioned glycan-mediated host-microbe in-

teractions, glycans could provide insights into pathogenicity

and commensalism determinants, as, for instance, molecular

mimicry of host glycans by both pathogens and commensals fa-

cilitates their immune evasion (Carlin et al., 2009; Varki and Gag-

neux, 2015). Additional therapeutic potential is enabled by the

widespread usage of glycans by viruses for cell adhesion and

Author(s). Published by Elsevier Inc. creativecommons.org/licenses/by-nc-nd/4.0/).

C

D Bonds Made by NeuNAc

α2-8

α2-3

α2-6

N um

b e

r o f G

ly c

a ns

0

200

400

α2-6

α2-3

α2-8

Bonds Made by NeuNGc

0

40

80

NeuNAc

NeuNGc Kdo

Monosaccharides with Bond α2-3

0

200

400

Monosaccharides Paired with Fuc

N um

b e

r o f G

ly c

a ns

Branching

O c

c ur

e nc

e s

Position

Main

Side

G a

l G

lc

G a

lN A

c M

a n

Fu c

G lc

N A

c N

G lc

N A

c O

S

G lc

A Rh

a

G lc

N A

c G

ro G

a lO

A c

G a

lA G

lc N

1200

800

400

0

3000

2000

1000

0

G lc

N A

c

G a

lO S

A 12,674 Species-Specific Glycans

6,969 eukaryotic 6,119 prokaryotic 152 viral

19,299 Unique Glycans

1,027 Glycoletters 19,866 Glycowords 9,152 Glycans with at least one label

1600

6000

Domain

Order

Kingdom

Family

Phylum

Genus

Class

Species

Number of Glycans

Number of Glycans 0 2000 4000 6000 0 2000 4000 0 20001000 3000

0 800 0 400 800 0 400 800

Virus

Archaea

Primates Pseudomonadales Fabales

Rhizobiales

Saccharomycetales

Lactobacillales Artiodactyla

Burkholderiales

Actinomycetales

0 1000 2000

Plantae Animalia

Fungi Excavata

Virus Euryarchaeota Riboviria Chromista Proteoarchaeota

0 2000 4000

Hominidae

Pseudomonadaceae Fabaceae

Saccharomycetaceae Rhizobiaceae

Pasteurellaceae

Burkholderiaceae

Solanaceae Muridae

Angiosperms Chordata

Ascomycota Firmicutes

Basidiomycota Actinobacteria

Euglenozoa

Virus Arthropoda

Homo

Salmonella

Burkholderia Shigella

Bos Sus Streptococcus

Lactobacillus

Dicotyledons Mammalia

Bacilli Alphaproteobacteria Monocotyledons

Saccharomycetes

Betaproteobacteria

Sordariomycetes Actinobacteria

Sus scrofa Mus musculus

Rattus norvegicus

Shigella dysenteriae Gallus gallus

Pseudomonas sp.

Saccharomyces cerevisiae

B Eukarya

Bacteria

Bacteria Proteobacteria Gammaproteobacteria

Enterobacterales Enterobacteriaceae Escherichia Homo

Pseudomonas Homo sapiens

Bos taurus

Escherichia coli

Figure 1. Using a Curated Glycan Dataset as a Resource for Glycobiology and Analyzing Host-Microbe Interactions (A) Building curated datasets of species-specific and unique glycan sequences. Glycans stemming from proteins, lipids, small molecules, or cellular surfaces

were gathered from UniCarbKB, CSDB, GlyTouCan, and the academic literature. We deposited these datasets in our database SugarBase, containing additional

associated metadata, such as linkage and immunogenicity information.

(legend continued on next page)

ll OPEN ACCESSResource

Cell Host & Microbe 29, 132–144, January 13, 2021 133

ll OPEN ACCESS Resource

entry (Thompson et al., 2019) and pathogenic bacteria (Poole

et al., 2018).

In addition to previous work developing computational ap-

proaches to glycan analysis (McDonald et al., 2016; Spahn

et al., 2016), identifying relevant glycan motifs and their roles in

host-microbe interactions at scale would benefit from pattern-

learning algorithms, such as machine learning, that can uncover

statistical dependencies in biological sequences (Camacho

et al., 2018). Research on other biopolymers has shown that lan-

guage models, originally developed for the analysis of human

languages, perform best in this task (Alley et al., 2019; Almagro

Armenteros et al., 2020; Strodthoff et al., 2020), because they

can leverage evolutionarily conserved regularities and lan-

guage-like properties in such sequences. Language models,

with their memory-like features, are well suited for leveraging

patterns and implicit structure in biopolymers such as those un-

derlying nucleic acids (Valeri et al., 2020) and proteins (Alley

et al., 2019), because information in these sequences is order

dependent, and non-neighboring residues can have meaningful

interactions. Applying a natural language-processing approach

to biological sequences also enables learning a representation

of a molecule that can be used to analyze sequence motifs

and predict functional properties. These types of models are

therefore a suitable starting point for the analysis of glycan

sequences.

Here, we present a resource toolkit comprising machine

learning and bioinformatics methods as well as a large glycan

database to leverage the evolutionary information present in gly-

cans for predictive purposes in the context of host-microbe in-

teractions, e.g., by understanding pathogenicity-associated

glycan motifs. This toolkit can be used as a complete workflow

for investigating host-microbe interactions, from a glycan data-

set to glycan motifs identified by machine learning and further

investigated by glycan alignments, or as separate modules. Un-

derlying all of this is our language model for glycans, SweetTalk,

trained on a dataset of 19,299 unique glycan sequences. With

this, we demonstrate that similarities between glycans can be

visualized and used to predict glycan properties such as human

immunogenicity. Another part of our platform is SweetOrigins, a

language-model-based classifier predicting the taxonomic

origin of glycans that we use to obtain evolution-informed repre-

sentations of glycans. To achieve this in the context of glycan-

mediated host-microbe interactions, we manually curated a

comprehensive dataset comprising 12,674 glycans with species

annotations. These datasets were combined into a database,

SugarBase, that is amenable to programmatic access and inte-

gration into deep-learning pipelines, thus providing resources for

analyses involving host-microbe interactions.

In this work, we demonstrate the potential and generaliz-

ability of using SugarBase, SweetTalk, SweetOrigins, and a

glycan-alignment methodology for studying glycan-mediated

host-microbe interactions. We show that a language-model-

based classifier trained on glycan sequences can accurately

(B) Glycan species distribution in the species-specific glycan dataset. For all gly

taxonomic level are shown with their number of glycans.

(C and D) Analyzing the local structural context of glycoletters. We identified the m

local structural context together with its likely position in the glycan structure (main

sialic acids (D).

134 Cell Host & Microbe 29, 132–144, January 13, 2021

predict glycan immunogenicity and the pathogenicity of

E. coli strains, revealing predictive glycan motifs. We also

leverage the evolutionary information gained by SweetOrigins

to analyze glycan motifs that could be used for molecular-

mimicry-mediated immune evasion by commensals and

pathogens. Applying our glycan-alignment methodology to

the example of the capsular polysaccharides of Staphylo-

coccus aureus and Acinetobacter baumannii, we uncover a

potential connection to the enterobacterial common antigen

and hypothesize a mechanism for the increased virulence

mediated by these glycan motifs. Taken together, these

resources offer a powerful and generalizable platform for

studying and understanding the role of glycans in host-

microbe interactions.

RESULTS

Curating Glycan Datasets for Glycobiology and Glycan- Mediated Host-Microbe Interactions To investigate the role of glycans in host-microbe interactions,

we constructed a dataset of species-specific glycan sequences

that could be used to train machine-learning models. For this, we

gathered and curated a dataset with glycans from GlyTouCan

(Tiemeyer et al., 2017), UniCarbKB (Campbell et al., 2014), the

Carbohydrate Structure Database (CSDB) (Toukach and Egor-

ova, 2016), and targeted literature searches (see STAR

Methods). To facilitate training deep-learning models on glycan

sequences, we only included glycans with fully elucidated se-

quences, including the determination of linkages between

monosaccharides. Our dataset contained 12,674 highly diverse

glycans with a deposited species association (Figure 1A; Table

S1) and included glycans from 1,726 species (corresponding to

39 taxonomic phyla; Figure 1B). Specifically, our dataset con-

tained 6,969 eukaryotic, 6,119 prokaryotic, and 152 viral gly-

cans. Because we included all species for which we could find

glycans, this dataset constituted a comprehensive snapshot of

currently known species-specific glycans, with glycans from

numerous bacteria, facilitating the study of glycan-mediated

host-microbe interactions.

We further reasoned that the inclusion of glycan sequences

without a deposited species label would strengthen the lan-

guage models we describe below. This approach is supported

by the success of transfer learning in the field of machine learning

(Howard and Ruder, 2018), in which models are initially trained

on large datasets without labels and then finetuned on smaller

datasets with labels. This makes more data available to learn

general patterns, such as sequence motifs, that can be lever-

aged to predict glycan properties. Accordingly, we curated a

separate dataset in which we used the databases mentioned

above to gather 19,299 unique glycan sequences, irrespective

of whether species information was available (Figure 1A;

STAR Methods; Table S2). To gain a comprehensive view of

glycobiology, we included all glycan categories, encompassing

cans with species information, up to the 10 most abundant classes for each

ost frequent monosaccharides following fucose in glycans (C), highlighting its

versus side branch). Additionally, we compared the binding behavior of several

Featurize Input

/ // /

Language Model (Glycoletters)

Classifiers (Glycowords)

SweetTalk

Xt-1 Xt

LSTMRv LSTMRv LSTMRv

Xt+1

Yt-1 Yt Yt+1

2-3x

Embedding

A

Language Model Output

C Glycowords With Existing Alphabet

Possible Realized

Data Processing

GlyTouCan Literature

D

Datasets

α2

β4 α3 β3 β3

Ser/Thr

α2

β4 α3 β3 β3

Ser/Thr

α2

β4 α3 β3 β3

Ser/Thr

α2

β4 α3 β3 β3

Ser/Thr

β3 β3

α3 β3

β4 α3

α2 β4

Fuc

Gal GalNAc GlcNAc

tSNE Dim 1

Glycoletter Embeddings

Fuc

Glc Neu

Man

B Glycowords

tS N

E D

im 2

Bonds

Gal

NeuNAc

Glc

α3 β3

10 -

30 -

-10 -

-30 -

-30

-

-10

-

30

-

10

-

N um

b e

r o f G

ly c

o w

o rd

s

Possible Realized

U M

A P

D im

2

UMAP Dim 1

8 -

4 -

0 -

-4 -

-10

-

-6

-

-2

-

2

-

6

-

LSTMFw LSTMFw LSTMFw

19,299 Glycans

E F G

2-2-6 UMAP Dim 1

6

-8

-4

0

4

U M

A P

D im

2

Non -immuno

-genic

Immuno -genic

N um

b e

r o f U

nm a

sk e

d G

ly c

o w

o rd

s

Probability Immunogenic Probability Immunogenic

α3 α6 α6

α3

β2

β4 β4 α6

α2 α3 α3 α2

α6

α2 β4 β4 α3 β4 β4 α6

WT α2 α3

α2 α6

1 2

3

3

1 2

2

α6 α3

α3 α6

β4 β21

1

2

1

1

1

α6 β4 β3 β4

WT

1 2

β4 β31

1

β3 β6

α6 α3

Fuc Gal GalNAcGlcNAc ManNAcGlc ManRhaNeuNAcNeuNGc Xyl

αGal

0.2

-

0.6

-

1.0

-

0.0

-

0.4

-

0.8

-

0.2

-

0.6

-

1.0

-

0.0

-

0.4

-

0.8

-

1 -

2 - 3 -

4 - 5 -

6 - 7 -

0.2

-

0.6

-

1.0

-

0.2

-

0.6

-

1.0

-

1 -

2 -

3 -

4 -

0.2

-

0.6

-

1.0

-

0.2

-

0.6

-

1.0

-

Homo sapiens

Non-Reducing Reducing

Homo sapiens

Non-Reducing Reducing

Ruminococcus gnavus

Homo sapiens

High MMannose

ghHig haRh

NN-Glycans

nsO-Glycan

1012

108

104

100

(legend on next page)

ll OPEN ACCESSResource

Cell Host & Microbe 29, 132–144, January 13, 2021 135

ll OPEN ACCESS Resource

protein-, lipid-, and small molecule-associated glycans, as well

as capsular and extracellular polysaccharides.

In our dataset, we observed 1,027 unique monosaccharides or

bonds that were present in glycan sequences and comprised the

smallest units of an alphabet for a glycan language. Analogous to

natural language processing, we termed these entities ‘‘glycolet-

ters’’ and constructed ‘‘glycowords’’ by considering trisaccha-

rides (i.e., three monosaccharides and two connecting bonds,

or five glycoletters), yielding 19,866 unique glycowords in our da-

taset. With this, we sought to incorporate local structural infor-

mation into our models and enable the discovery of relevant mo-

tifs, which usually contain subsequences larger than a single

monosaccharide. Even larger substructures would preclude

the analysis of shorter glycans and lead to an exponential in-

crease in the size of the resulting vocabulary. We would also

like to note that although we chose trisaccharides as building

blocks, glycan substructures of any length can be used to build

a vocabulary for our models without considerable changes.

To make these data and analysis resources readily accessible

and facilitate further advances in glycobiology, we created Sug-

arBase, a comprehensive glycan database with metadata and

analytical tools based on this work (Figure S1A; Table S2;

https://webapps.wyss.harvard.edu/sugarbase). SugarBase of-

fers accessible glycan data, explorable glycan representations

learned by our language models, and many of the methods

developed here as tools, such as the local structural context

of any glycoletter (Figure S1B) and glycan alignments, described

below.

Reasoning that our glycan datasets constitute broad re-

sources for glycobiology and host-microbe interactions, we set

out to investigate host glycan substructures that could be

emulated by microbes for molecular mimicry. Analyzing the envi-

ronment of the monosaccharide fucose as an example, we

observed N-acetylglucosamine (GlcNAc) and galactose (Gal)

as typical connected monosaccharides (Figure 1C), which is

consistent with the fucosyltransferase substrate specificities an-

notated in glycosyltransferase family 10 (Lombard et al., 2014).

Thus, microbial glycans containing fucose could potentially

include either GlcNAc or Gal in direct proximity to maximize sim-

ilarity with host glycans. This insight aids in formulating hypoth-

Figure 2. Learning the Language of Glycans Revealed Regularities in S

(A) Building a language model for glycobiology. We used glycowords, overlapping

based bidirectional RNN, SweetTalk, that was trained by predicting the next glyc

symbol nomenclature for glycans (SNFG).

(B) Learned representation of glycoletters by SweetTalk. We visualized the embe

SNE). Areas enriched for modified monosaccharides of one type are colored.

(C) Comparing the abundance of possible and observed glycowords. Possible

exhaustive combination (36 bonds and 991 monosaccharides).

(D) Comparing the distribution of possible and observed glycowords. We gene

monosaccharides and bonds and formed their embedding by averaging their co

jection (UMAP) of these generated glycowords (blue) and all observed glycoword

(E) Glycan embeddings learned by the immunogenicity classifier. Embeddings fo

according to whether they were immunogenic (blue) or non-immunogenic (orang

(F) Glycoword masking to probe the immunogenicity classifier. Glycowords were

Reducing’’/‘‘Reducing’’) and used as input for the trained immunogenicity classifi

glycan is for prediction, with the bar representing the full-length glycan at the bo

(G) Glycan in silico alterations to probe immunogenicity classifier. For 4,000

monosaccharide or bond. If the resulting glycowords were observed, we used th

probability is plotted together with the altered glycan sequences, with the wildtyp

monosaccharide was modified. The addition of an ‘‘S’’ implies a sulfurylated mo

136 Cell Host & Microbe 29, 132–144, January 13, 2021

eses and identifying glycan motifs relevant for molecular mim-

icry, as we describe below. We also differentiated binding

orientation preferences for different sialic acids, a crucial mono-

saccharide type in host-pathogen interactions (Figure 1D;

Haines-menges et al., 2015), revealing a preference for the char-

acteristic human monosaccharide NeuNAc to be (a2-3)-linked,

relative to other sialic acids such as NeuNGc. These types of an-

alyses can directly lead to hypotheses of glycan motifs that can

be investigated by using the methods presented in this work.

Using Natural Language Processing to Learn the Grammar of Glycans Next, we used our curated dataset of 19,299 glycan sequences

(Table S2) to develop a deep-learning-based language model,

SweetTalk. For this, we chose a bidirectional recurrent neural

network (RNN; Figure 2A; Sherstinsky, 2020), because this

type of model has delivered state-of-the-art results for other bio-

polymers, such as protein sequences (Alley et al., 2019; Almagro

Armenteros et al., 2020; Strodthoff et al., 2020). Originally devel-

oped for human languages, RNNs exhibit memory-like elements

by predicting the next word given the preceding words (Sherstin-

sky, 2020); this enables RNNs to learn complex, order-depen-

dent interactions in proteins by viewing amino acids as letters

and predicting the next amino acid given the preceding

sequence (Alley et al., 2019). Two of the main usages for a trained

language model are as follows: (1) extracting a learned represen-

tation for each word and (2) finetuning the model for predicting

structural or functional properties of a sequence. For the former,

a representation or embedding that characterizes a word in

terms of context, usage, and meaning is constructed in the pa-

rameters of the trained model for each word in the vocabulary.

This learned representation can be used to quantify the similarity

of two glycan sequences or analyze language properties, which

we demonstrate with the analysis of molecular mimicry in host-

microbe interactions. The latter—finetuning a general language

model on a predictive task such as predicting pathogenicity—

is also known as transfer learning (Howard and Ruder, 2018;

Tan et al., 2018), and in our case it involves general glycan fea-

tures that are learned by the language model to predict func-

tional properties.

ubstructures and Can Be Used to Predict Glycan Immunogenicity

units consisting of three monosaccharides and two bonds, for our glycoletter-

oletter given previous glycoletters. Glycans are drawn in accordance with the

dding for every glycoletter by t-distributed stochastic neighbor embedding (t-

glycowords were calculated from the pool of observed glycoletters and their

rated 250,000 glycowords by randomly sampling from the observed pool of

nstituent glycoletter embeddings. A uniform manifold approximation and pro-

s (orange) is shown.

r glycans from our immunogenicity dataset are shown via UMAP and colored

e).

progressively exchanged with padding (‘‘masking’’) from both termini (‘‘Non-

er. Inferred immunogenicity probability indicates how crucial each region of a

ttom.

iterations, single monosaccharides or bonds were replaced with a random

em as input for the trained immunogenicity classifier. Inferred immunogenicity

e glycan found at the bottom. In case of ambiguity, a number indicates which

nosaccharide, whereas ‘‘Me’’ implies a methylated monosaccharide.

ll OPEN ACCESSResource

Glycans are the only nonlinear biopolymer, with up to multiple

branches per sequence. To enable a language model despite

this branching, we extracted partially overlapping ‘‘glyco-

words’’ from the non-reducing end to the reducing end of gly-

cans in the bracket notation (Figure 2A), comprising three

monosaccharides and two bonds. These glycowords repre-

sented snapshots of structural contexts that characterize a

glycan sequence. By using monosaccharides and bonds as

‘‘glycoletters,’’ we then trained a glycoletter-based language

model, SweetTalk, predicting the next most probable glycolet-

ter given the preceding glycoletters in the context of these gly-

cowords (Table S3). This operation, instead of directly training

on full sequences, avoids learning specious relationships be-

tween glycoletters that are close in the bracket notation but

far apart in the actual glycan structure due to branching. We

then demonstrated the necessity of accounting for the order-

dependent information in glycans by training SweetTalk on

scrambled glycan sequences, randomizing the order but keep-

ing the composition of a sequence—this resulted in severely

degraded model performance, emphasizing the language-like

elements inherent in glycan sequences (Table S3). Analyzing

the learned embeddings of glycoletters after training SweetTalk

revealed similar positions in embedding space for monosac-

charides and their modified counterparts (e.g., sulfurylated

galactose, GalOS, and sulfurylated N-acetylgalactosamine,

GalNAcOS; Figure 2B), implying similarity in their language

characteristics and context. This finding is reminiscent of ob-

servations made on the popular word2vec embeddings that

also learn a representation of words in a human language by

considering their neighboring words/context, in which seman-

tically similar words form clusters (Mikolov et al., 2013).

We then constructed glycoword embeddings by averaging the

embeddings of their constituent glycoletters. Our first observa-

tion was that from the close to 1.2 trillion possible glycowords

(given our observed glycoletters), only 19,866 distinct glyco-

words (�0.0000016%) were observed here (Figure 2C). More- over, these 19,866 glycowords were not evenly distributed in

the learned embedding space, as existing glycowords formed

clusters compared to in silico-generated, possible glycowords

(Figure 2D). The observation that the glycoword space (and,

thus, glycan space) is sparsely populated is potentially a conse-

quence of having to evolve dedicated enzymes for constructing

specific glycan substructures from a species-specific set of

monosaccharide building blocks, making most combinations

inaccessible.

Predicting Glycan Immunogenicity with a Glycan-Based Language Model Given the important role glycans play in human immunity (Kap-

pler and Hennet, 2020; Reusch and Tejada, 2015), we curated

known immunogenic glycans from the literature (Table S2) to fi-

netune a SweetTalk-based classifier with glycan sequences as

input to predict their immunogenicity to humans. On an indepen-

dent validation dataset, our model achieved an accuracy of

�92% (F1 score or balanced F score: 0.915), in comparison with an accuracy of �51% for a model trained on scrambled glycan sequences (Figures 2E–2G; Table S4). Alternative ma-

chine-learning models that did not treat glycan sequences as a

language, such as random forest classifiers, only achieved accu-

racies ranging from �80%–88% for this task (Table S4), empha- sizing the importance of order and patterns for elucidating

glycan properties.

Rhamnose-rich glycans, a common monosaccharide in bac-

teria but not in mammals, were unambiguously assigned to an

immunogenic cluster by our RNN-based model and presented

the most striking motif for glycan immunogenicity (Figure 2E).

The cluster containing high-mannose glycans provided addi-

tional ambiguity, because it included both immature human gly-

cans and immunogenic fungal glycans, potentially suggesting

the immunogenicity of unintentionally exposed immature human

glycans. Indeed, the presence of immature high-mannose gly-

cans on viral surfaces has been noted to influence immunoge-

nicity, with many broadly neutralizing antibodies targeting the

high-mannose glycans on HIV glycoproteins (Lavine et al.,

2012). We also found that human mucosal O-glycans, character-

ized by their interactions with bacteria, were interspersed with

bacterial immunogenic glycans in the embedding space, in

contrast to N-linked glycans. This adds to the notion of an immu-

nological compromise of recognizing these bacterial glycans at

the expense of targeting human O-glycans with shared motifs,

such as the ABH blood group antigens (Kappler and Hennet,

2020). These analyses indicate that embeddings from glycan-

focused language models could be used to study characteristics

of glycans on a large scale and with many potential applications,

such as the exploration of glycan-immune system interactions.

Using Deep Learning to Provide Evolution-Informed Glycan Representations We next hypothesized that the evolutionary pressures on gly-

cans stemming from host-pathogen interactions could be ex-

tracted by a deep-learning model. For this, we constructed a lan-

guage-model-based classifier, SweetOrigins, to predict the

taxonomic origin of a glycan (Figure 3A). In distinguishing taxo-

nomic classes, SweetOrigins could learn species-specific fea-

tures of glycans that are indicative of their evolutionary history.

Based on a bidirectional RNN, we first pre-trained SweetOrigins

with a SweetTalk model as described above. We then used the

language-like properties learned in this process to finetune the

model on a different task—predicting the taxonomic group of

glycans. By doing this for every taxonomic level, from the spe-

cies level up to the domain level, we obtained eight SweetOrigins

models with the same basic model architecture except for

different final layers. These final layers could learn how to

combine the extracted information from glycans for predicting

their taxonomic group, and they differed in terms of their number

of output nodes, as the number of classes varied for each taxo-

nomic level. This strategy was successful in extracting evolu-

tionary information from glycans, as SweetOrigins models clas-

sified the taxonomic group of a glycan with high accuracy

(Table 1).

In contrast to other biological sequences such as DNA or pro-

teins, the number of available sequences for glycans is still

limited, which is compounded by their high diversity. This is

especially visible in prediction tasks in which only few glycans

per class are available, such as for the species-level SweetOri-

gins model, resulting in lower model performance for rare clas-

ses and less useful glycan representations for downstream ana-

lyses. As knowledge of host-microbe interactions at the species

Cell Host & Microbe 29, 132–144, January 13, 2021 137

A

Glycan

G ly

c o

w o

rd s

SweetOrigins

b iL

ST M

b iL

ST M

b iL

ST M

Domain: Bacteria

Kingdom: Bacteria

Phylum: Proteobacteria

Class: Gammaproteobacteria

Order: Enterobacterales

Family: Enteriobacteriaceae

Genus: Escherichia

Classification ResultFully Connected Layer

Species: Escherichia coli

Glycoword Embeddings

α6β2 β2

α3

β4 β4 α3

α6β2

β2

α3 β4 β4

α3 α6

β2

β2

α3

β4 β4 α3 α6

β2

β2 α3 β4 β4

α3

α6β2 β2

α3

β4 β4 α3

α6

β2

β2

α3 β4 β4 α3 α6

β2

β2 α3 β4 β4

α3

GlcNAc(b1-2)Man(a1-6)[Xyl(b1-2)][Man(a1-3)]Man(b1-4)GlcNAc(b1-4)[Fuc(a1-3)]GlcNAc

GlcNAc(b1-2)Man(a1-6) [Xyl(b1-2)][Man(a1-3)]...

GlcNAc(b1-2)Man(a1-6) [Man(a1-3)][Xyl(b1-2)]...

Xyl(b1-2)[GlcNAc(b1-2) Man(a1-6)][Man(a1-3)]..

Xyl(b1-2)[Man(a1-3)] [GlcNAc(b1-2)Man(a1-6)]..

Man(a1-3)[GlcNAc(b1-2) Man(a1-6)][Xyl(b1-2)]..

Man(a1-3)[Xyl(b1-2)] [GlcNAc(b1-2)Man(a1-6)]..

B

400

t-SNE Dim 1 -40

t- SN

E D

im 2

20

40

0

-20

αGal-engineered

O8/O9

K-12

O86/O127/O128

O13/O148/O150 C

F470O6

O174

O4/O25

O6

J5

C D

t-SNE Dim 1 -20 0 20 60

40

0

-40

Yes Unknown No

Pathogenic

O157:H7

K-12

O111:B4

α3 β3 β3

α6 α4

α6 α4

β3α4

α3β3

α3 β3

Figure 3. Deep-Learning-Based Classifiers Use Glycans to Predict Taxonomic Origin and Pathogenicity

(A) Exemplary schematic of SweetOrigins to predict taxonomic origin from glycans. Lists of glycowords are used as input for a SweetOrigins model to predict the

taxonomic class ranging from the domain level down to the species level.

(B) Glycan data augmentation strategy. Different bracket notations describing the same glycan can be generated by alternating double branches as well as

replacing side branches with main branches to increase model robustness.

(C) Glycans of E. coli in embedding space distinguish strains. The embedding for all 1,010 E. coli-derived glycans with strain information from the trained species-

level SweetOrigins model is plotted via t-SNE and colored for areas enriched for annotated E. coli strains.

(D) E. coli glycans predict pathogenicity. For all E. coli-derived glycans, representations learned by a model predicting pathogenicity are plotted via t-SNE and

colored as to whether they stem from pathogenic, non-pathogenic, or unlabeled E. coli. Example strains for all cases are annotated.

ll OPEN ACCESS Resource

138 Cell Host & Microbe 29, 132–144, January 13, 2021

Table 1. Metrics of Trained SweetOrigins Models

Taxonomic Level Classes

Baseline

Accuracy Cross-Entropy Loss Accuracy MCC

Random Max Base Aug Base Aug Base Aug

Domain 4 (4) 0.2500 0.99 0.2841 0.1906 0.9128 0.9313 0.8134 0.8693

Kingdom 9 (11) 0.1111 0.98 0.3844 0.3249 0.8733 0.8953 0.8001 0.8390

Phylum 33 (39) 0.0303 0.98 0.8685 0.7543 0.7779 0.8008 0.7018 0.7341

Class 71 (101) 0.0141 0.96 1.3283 1.1729 0.6803 0.7149 0.6218 0.6638

Order 145 (207) 0.0069 0.92 2.2498 2.1132 0.4937 0.5333 0.4602 0.5066

Family 258 (411) 0.0039 0.90 2.9834 2.7068 0.4134 0.4660 0.3873 0.4428

Genus 405 (919) 0.0025 0.86 3.6588 3.4081 0.3658 0.3849 0.3505 0.3682

Species 581 (1,726) 0.0017 0.86 4.3704 3.9550 0.3052 0.3651 0.2870 0.3496

Taxonomic groups with fewer than five unique glycans were not used for model training or validation. Number of classes indicates the number of

included taxonomic groups, whereas the full number of taxonomic groups in our dataset is given in parentheses. Models were trained with the standard

set of glycans (Base) or after data augmentation (Aug). As an accuracy baseline, a random prediction of classes was used for each model. Max in-

dicates the maximum theoretically possible accuracy given shared glycan sequences across taxonomic groups. Cross-entropy loss, accuracy,

and Matthew’s correlation coefficient (MCC) of the trained model on a separate validation set are given for each taxonomic level. For each metric

and taxonomic level, the superior value is bolded.

ll OPEN ACCESSResource

level could offer insights, we developed methods that enable

training glycan-focused machine-learning models on small data-

sets. This goal motivated our transfer-learning approach of pre-

training a language model on all glycan sequences and then fine-

tuning the model on a smaller dataset, because this approach in

natural language processing has in some cases reduced the

necessary dataset size by a factor of 100 (Howard and Ruder,

2018). In other domains of deep learning, such as image classi-

fication, data augmentation routinely results in improved model

quality and robustness by providing the model with slightly modi-

fied versions of the data (Perez and Wang, 2017), such as

rotating images or changing their brightness. We reasoned that

the same could be achieved for biomolecules such as glycans;

we thus designed a data-augmentation method, specifically for

glycans, by conceptualizing glycans as graphs and forming a

set of isomorphic graphs comprising slightly different lists of gly-

cowords that we used as inputs for SweetOrigins (Figure 3B;

STAR Methods). Capitalizing on the ambiguity of the bracket no-

tation (Tanaka et al., 2014), we generated bracket notations that

differed in their ordering of branches but still described the same

glycan. This led to model performance improvements at every

classification level, with absolute accuracy increases of up to

6%, by effectively increasing the amount of available data. As

we envisioned, classifications with less data per class, such as

the species level, benefited most from data augmentation (Table

1), paving the way for using glycan-based deep-learning models

with smaller datasets.

In general, our predictions were robust, and we could, for

example, accurately predict glycans from the kingdoms Animalia

(91.1%) and Bacteria (97.2%), as well as glycans from the phyla

Chordata (91.9%) and Firmicutes (90.4%) in our validation data-

set (Figures S2A–S2C). This demonstrates that SweetOrigins

can learn glycan representations from both hosts and microbes,

enabling the analyses presented below. Any misclassifications

occurred among closely related groups, such as viral glycans

misclassified as those of their hosts (Figures S2A–S2C). Glycan

embeddings from our trained SweetOrigins model illustrated

clusters reminiscent of taxonomic groups (Figure S2D). We

next used our trained SweetOrigins models to infer the taxo-

nomic origin of the 10,333 glycans without a species label in

our dataset (Table S2). For several randomly selected glycans,

we performed literature searches to validate the predictions

made by SweetOrigins (Figure S2E; Table S5), indicating that

our trained SweetOrigins models had accurately learned spe-

cies- or group-specific glycan motifs.

We next used SweetOrigins models to investigate host-path-

ogen interactions, specifically in the context of the well-studied

bacterium E. coli. Although SweetOrigins classifiers were only

trained up to the species level, we hypothesized that subspe-

cies-level information could be extracted from the rich glycan

representation learned by the species-level SweetOrigins model.

To test this, we gathered 1,010 glycan sequences from E. coli

with strain-level annotation from CSDB and used these as inputs

to our trained model, yielding learned representations that we

used to differentiate serotypes. We could readily identify clusters

enriched for several strains in the representations, such as the

serotypes O8/O9, characterized by a special polymannose O-

antigen (Greenfield et al., 2012), and the K-12 strain popular in

molecular biology research (Figure 3C), demonstrating the diver-

sity and characteristic features of glycans for different E. coli

strains.

We next reasoned, given the prominent role of glycans in host-

microbe interactions, that these glycan differences could be

used to predict E. coli pathogenicity, because E. coli strains

can range from being non-colonizing to commensal or patho-

genic (Lim et al., 2010). Accordingly, we trained a deep-

learning-based classifier with the same language-model archi-

tecture as SweetOrigins on glycan sequences to elucidate

whether information in glycans can predict pathogenicity. With

a threshold of 0.5 in the predicted probability of pathogenicity,

we found that we were able to predict E. coli strain pathogenicity

with an accuracy of �89% on a separate validation dataset (Fig- ure 3D; F1 score: �0.906). This positioned E. coli strains along a continuum of predicted pathogenicity and supported the role of

glycans in mediating pathogenicity. Interestingly, E. coli strains

such as O111:B4, which were labeled as ‘‘unknown’’ in the

Cell Host & Microbe 29, 132–144, January 13, 2021 139

ll OPEN ACCESS Resource

dataset and therefore not available during model training, were

predicted to be among the pathogenic strains and confirmed

to cause gastric disease (Viljanen et al., 1990). Our trained model

placed the majority of E. coli glycans from unknown pathoge-

nicity strains between pathogenic and non-pathogenic strains,

adding to the notion of a continuum of pathogenicity (Casade-

vall, 2017).

Because glycans appear to be predictive of pathogenicity, we

reasoned that certain glycan motifs in E. coli strains on the path-

ogenic end of the spectrum might provide further insight into

pathogenesis. To address this notion, we identified glycan motifs

that are enriched in regions populated by predominantly patho-

genic E. coli strains in the representation learned by our model

(Figure 3D). Motifs in these pathogenicity-associated glycans

exhibited a striking resemblance to host mucosal glycans, with

an enrichment for a1-2-linked fucose and the core 1 O-glycan

structure (also known as T antigen; Gal(b1-3)GalNAc) prevalent

in mucins (Figures S3A and S3B). Consistent with our local struc-

tural context analysis (Figure 1B), the majority of a1-2-linked

fucose residues in pathogenic E. coli strains were linked to

galactose (Figure S3C), forming part of the human blood group

H antigen. Indeed, when analyzing the glycan motifs most pre-

dictive of E. coli strain pathogenicity, both Gal(b1-3)GalNAc

and Fuc(a1-2)Gal disaccharides were among the top 20 motifs

(Figure S3D). On the other hand, the presence of typical bacterial

glycan components, such as rhamnose or L-Glycero-D-Manno-

Heptose (LDManHep), was associated with lower predicted

pathogenicity (Figure S3D).

Using Glycan Alignments to Study Virulence Determinants in Bacterial Pathogens To better understand the function of glycans in host-microbe in-

teractions, we developed a sequence-alignment method. For

DNA and protein sequences, alignments use sequence changes

due to mutations and insertions to enable, for example, the iden-

tification of conserved motifs in protein families (Do�gan and Kar-

açalı, 2013). To facilitate analogous analyses for glycans and capitalize on the evolutionary influence of host-pathogen inter-

actions on glycans, we developed methods for gapped, pairwise

alignments of glycan sequences based on the Needleman-

Wunsch alignment algorithm (Needleman and Wunsch, 1970).

For this, we constructed a substitution matrix (which we termed

GLYSUM; Table S6), analogous to the BLOSUM matrices used

in protein alignments, that utilizes the likelihood of substituting

two monosaccharides to calculate alignment scores. To assess

whether our glycan alignments performed as envisioned, we

analyzed viral glycans that are predominantly derived from their

host organisms and thus should align to host glycans. As ex-

pected, the optimal alignment for the viral glycans was indeed

from their host organisms (Figures 4A and 4B), supporting the

validity of our glycan-alignment method.

We reasoned that functionally relevant glycan motifs for host-

pathogen interactions are likely conserved to some extent and

could be analyzed with glycan alignments. As an example, we

used our glycan-alignment method to align the serotype 5

capsular polysaccharide of the clinically relevant pathogen

S. aureus, which is known to increase bacterial virulence (Tziana-

bos et al., 2001), against our dataset. Because the capsular poly-

saccharides of S. aureus mediate its evasion of the immune sys-

140 Cell Host & Microbe 29, 132–144, January 13, 2021

tem (Weidenmaier and Lee, 2015), we hypothesized that

comparing these to similar sequences might offer insights to un-

derstand their pathogenicity. Notably, the best alignment results

were achieved with the enterobacterial common antigen, ECA

(Figure 4C), conserved in the Enterobacteriaceae family, which

has been shown to be important for virulence (Gilbreath et al.,

2012) and outer membrane permeability (Mitchell et al., 2018).

These findings are supported by experiments demonstrating

that ECA deficiency in E. coli can be rescued by the expression

of enzymes from serotype 5 S. aureus (Kiser and Lee, 1998).

Such a phenotype complementation could suggest that this

ECA-like glycan motif fulfills a similar role in S. aureus as the ca-

nonical ECA in E. coli.

To further probe the connection of ECA-like glycans and

increased virulence, we aligned the canonical ECA motif against

our dataset to compile a list of ECA-like sequences and their

alignment distances; we used these distances to construct a

dendrogram detailing the relationships between ECA-like glycan

sequences (Figure 4D). Although most of the S. aureus-derived

ECA-like sequences formed a separate cluster, the type 5

capsular polysaccharide was located in a different cluster with

the canonical ECA sequences. Of note, we observed an ECA-

like motif in the capsular polysaccharide of A. baumannii (Fig-

ure 4D, bold), one of the most problematic hospital-acquired

pathogens, in the same cluster dominated by canonical ECA se-

quences. The capsular polysaccharide of A. baumannii has been

implicated with antibiotic resistance and virulence (Geisinger

and Isberg, 2015), providing an intriguing potential link to the

functions of the canonical ECA. For other pathogens, such as

Haemophilus ducreyi, the expression of a gene cluster synthe-

sizing a putative ECA-like glycan has also been linked to

increased virulence (Banks et al., 2008), further suggesting a

connection of this motif with virulence. Notably, the genera

Staphylococcus, Acinetobacter, and Haemophilus are not part

of the Enterobacteriaceae family that is typically associated

with the ECA, highlighting the importance of our glycan align-

ments for screening thousands of glycans to aid in understand-

ing motifs important for pathogenicity, such as the ECA-like gly-

cans from S. aureus and A. baumannii.

DISCUSSION

Here, we presented a set of resources—a collection of deep-

learning and bioinformatics methods, together with large,

curated datasets of glycan sequences—that can be used to

gain insights into many facets of glycan-mediated host-microbe

interactions. The aggregation of many glycan sequences in our

datasets leads to robust machine-learning models that are

largely unaffected by data-entry errors, thereby adjusting for

database errors. By training a language model to understand

the hidden grammar of glycan sequences, we demonstrated

that the information in glycans can be used to predict a range

of glycan properties, such as immunogenicity or pathogenicity.

We also showed that sequences can be compared and clustered

by learning a representation for each glycan via our trained

models. For applications involving glycoproteins, the distribution

of variant glycans on a protein (Wu et al., 2018) could be ac-

counted for by averaging their representations, potentially even

weighted by their relative abundance. By developing both

Human Immunodeficiency Virus

Gal β1-4 GlcNAc β1-2 Gal β1-4 GlcNAc β1-4 Man a1-3 Gal β1-4 GlcNAc β1-2 Man a1-6 Man β1-4 GlcNAc β1-4 Fuc a1-6 GlcNAc

Gal β1-4 GlcNAc β1-2 Gal β1-4 GlcNAc β1-4 Man a1-3 Gal β1-4 GlcNAc β1-2 Man a1-6 Man β1-4 GlcNAc β1-4 Fuc a1-6 GlcNAc

1 23

Alignment Score: 115 Percent Identity: 100.0 Percent Coverage: 100.0 Species: Homo sapiens

A

B SARS-CoV-2

Man a1-3 Man a1-6 Man a1-6 Man a1-3 Man β1-4 GlcNAc β1-4 GlcNAc

Man a1-3 Man a1-6 Man a1-6 Man a1-3 Man β1-4 GlcNAc β1-4 GlcNAc

1 13

Alignment Score: 65 Percent Identity: 100.0 Percent Coverage: 100.0 Species: Homo sapiens

C D Staphylococcus aureus

ManNAcA β1-4 FucNAcOAc a1-3 D-FucNAc β1-4 ManNAcA 1 7

Alignment Score: 26 Percent Identity: 71.4 Percent Coverage: 100.0 Species: Escherichia coli

ManNAcA β1-4 GlcNAc a1-3 D-FucNAc a1-4 ManNAcA

ManNAcA β1-4 FucNAcOAc a1-3 D-FucNAc β1-4 ManNAcA 1 7

Alignment Score: 22 Percent Identity: 57.1 Percent Coverage: 100.0 Species: Escherichia coli

ManNAcA β1-4 GlcNAc a1-3 FucNAc a1-4 ManNAcA

ManNAcA β1-4 FucNAcOAc a1-3 D-FucNAc β1-4 ManNAcA 1 7

Alignment Score: 21 Percent Identity: 71.4 Percent Coverage: 100.0 Species: Yersinia pestis

ManNAcA β1-4 GlcNAcOAc a1-3 D-FucNAc a1-4 ManNAcA

ManNAcβ1-4Glcα1-4ManNAc S.marcescens

ManNAcAβ1-4FucNAcOAcα1-3D-FucNAcβ1-4ManNAcA S.aureus

Manα1-2Fucα1-2GlcOAcAβ1-3GalNAc P.alcalifaciens

Manα1-3FucNAcα1-3GlcNAcβ1-3FucNAc C.universalis

ManNAcAβ1-4GlcNAcα1-3D-FucNAcα1-4ManNAcA S.sonnei

ManNAcAβ1-4GlcNAcα1-3D-FucNAcα1-4ManNAcA E.coli

ManNAcOAcAβ1-3FucNAcα1-3D-FucNAcβ1-3ManNAcOAcA S.aureus

Manα1-3D-Fucα1-3GlcNAcβ1-3Rha P.agglomerans

ManNAcα1-3Rhaβ1-4GlcNAcα1-2Man.1 S.dysenteriae

ManNAcAβ1-4GlcNAcα1-3D-FucNAcα1-4ManNAcA Y.enterocolitica

ManNAcβ1-4Glcβ1-3ManNAc.1 C.werkmanii

GalNAcβ1-4GlcAβ1-3D-FucNAcNβ1-3GalNAc P.temperata

ManNAcβ1-4Glcβ1-3ManNAc C.braakii

ManNAcAβ1-4GlcNAcα1-3FucNAcα1-4ManNAcA E.coli

ManNAcAβ1-4GlcNAcα1-3D-FucNAcα1-4ManNAcA S.enterica

Manα1-3Fucα1-3GlcNAcβ1-4GalNAc E.tarda

ManNAcAβ1-4GlcNAcα1-3D-FucNAcα1-4ManNAcA Y.pestis

Manα1-2Fucα1-2Glcβ1-3GlcNAc P.rustigianii

Manβ1-4Glcβ1-3D-FucNAcOAcα1-4GalNAc C.gillenii

ManNAcβ1-4GlcNAcα1-3D-FucNAcα1-4ManNAc E.coli

ManNAcOAcAβ1-4FucNAcα1-3D-FucNAcβ1-4ManNAcOAcA S.aureus

Manα1-3Fucα1-3GlcNAcα1-2Man.1 Y.entomophaga

ManNAcAβ1-4GlcNAcα1-3D-FucNAcα1-4ManNAcA P.mirabilis

ManNAcAβ1-4GlcNAcOAcβ1-3D-FucNAcα1-4ManNAcA Y.pestis

ManNAcOAcAβ1-3FucNAcα1-3FucNAcα1-3ManNAcOAcA S.aureus

ManNAcOAcAβ1-3FucNAcα1-3D-FucNAcα1-3ManNAcOAcA S.aureus

ManNAcβ1-4GlcNAcα1-4ManNAc H.alvei

ManNAcAβ1-4ManNAcAβ1-3D-FucNAcα1-4ManNAcA A.baumannii

Manα1-3Fucα1-3GlcNAcα1-2Man Y.pseudotuberculosis

ManNAcAα1-4FucNAcα1-3D-FucNAcβ1-4ManNAcA S.aureus

ManNAcβ1-4GlcNAcβ1-6GlcNAcα1-4ManNAc B.anthracis

ManNAcAβ1-4GlcNAcα1-4ManNAcA P.putida ManNAcAβ1-4GlcNAcα1-4ManNAcAβ1-3D-FucNAcα1-4ManNAcA A.globiformis

ManNAcAβ1-4L-GulNAcOAcAα1-3QuiNAcNButα1-4ManNAcA A.haemolyticus ManNAcAβ1-4GlcNAcAβ1-6Glca1-4ManNAcA A.cyaneus

ManNAcα1-3Rhaβ1-4GlcNAcα1-2Man R.terrigena

ManNAcAβ1-4GlcNAcα1-3D-FucNAcα1-4ManNAcA P.shigelloides

ManNAcAβ1-4GlcNAcNAmAβ1-3GlcNAcα1-4ManNAcA E.albertii

ManNAcAβ1-4GlcNAcOAcα1-3D-FucNAcα1-4ManNAcA Y.pestis

C an

on ic

al S

. a ur

eu s

Enterobacterial common antigen

Figure 4. Glycan Alignments Identify Pathogenicity-Associated Glycan Motifs

(A and B) Viral glycans aligned to host glycans. We aligned viral glycans to all glycans and depicted the highest scoring alignment.

(C) Glycan alignments using serotype 5 capsular polysaccharide of S. aureus. The repeating unit of the glycan was aligned against our database, and the best

three alignments are shown.

(D) ECA and ECA-like glycans. We aligned the canonical ECA sequence against our entire dataset, curated ECA-like sequences from the best 50 alignments, and

constructed a dendrogram from alignment distances.

ll OPEN ACCESSResource

transfer-learning and data-augmentation methods for glycan-

focused machine learning, we also addressed the pressing issue

of the limited availability of glycan sequences due to experi-

mental difficulties, enabling machine learning for many applica-

tions in glycobiology.

Our deep-learning strategies enabled us to introduce lan-

guage models for glycans, while our curated datasets offer a

state-of-the-art coverage for glycan sequences across a multi-

tude of organisms. In contrast to word2vec-type models (Miko-

lov et al., 2013), our language-model-based approach captured

sequential information beyond mere co-occurrences in glycan

sequences and thus achieved better predictive results than

alternative machine-learning techniques. This also enabled us

to analyze glycan motifs, such as those important for immuno-

genicity and pathogenicity, that are dependent on sequential

information and their relative position in glycans. Additionally,

starting from a glycoletter-based model allowed for the con-

struction of embeddings for close to 1.2 trillion glycowords,

making SweetTalk easily extendable to the full diversity of gly-

cobiology. SweetTalk can also incorporate position-specific

modifications, illustrating its flexibility and potential for the anal-

ysis of information-rich glycosaminoglycans to predict, for

instance, viral binding such as required for severe acute respi-

ratory syndrome coronavirus 2 (SARS-CoV-2) cell entry (Liu

et al., 2020).

Our resources can be utilized as a complete workflow, from a

glycan dataset to motifs obtained by machine learning and

further analyzed by glycan alignment, or as separate modules.

The accuracy exhibited by our SweetOrigins models demon-

strated that glycans can be used to distinguish closely related

taxonomic groups and provided the means to leverage the

evolutionary information in glycans for predictive purposes.

Our observation that E. coli glycans are predictive of pathoge-

nicity adds to the role of glycans as mediators of host-microbe

relationships (Poole et al., 2018). The continuum of pathogenicity

of E. coli strains, suggested by our deep-learning model, further

adds to the redefinition of the notion of pathogenicity from a bi-

nary concept to a gradual, environmentally controlled process

(Casadevall, 2017), mediated and influenced by glycans.

Both glycan alignments and glycan classification can connect

glycan functions with sequence patterns, which we have used to

derive insight from glycan motifs by analyzing glycans that could

potentially be used for molecular-mimicry-mediated immune

evasion by pathogenic E. coli strains. We further hypothesized

Cell Host & Microbe 29, 132–144, January 13, 2021 141

ll OPEN ACCESS Resource

that glycan-based molecular mimicry, in addition to mimicking

host glycans, could also extend to approximating glycans from

other bacteria for increased virulence, e.g., as in the case of

the capsular polysaccharides of S. aureus and A. baumannii, in

which we hypothesized that they potentially mimicked the ECA

of other bacteria. Our glycan-alignment method readily facili-

tated a hypothesis of the ECA mimicry performed by glycans

of these pathogens, with a potentially broader relevance of this

phenomenon in other pathogens, such as H. ducreyi, that are

predicted to engage in ECA mimicry as well. In general, the re-

sources developed here enable rapid discovery, understanding,

and utilization of functionally relevant glycan motifs from glycan

datasets, especially in the context of host-pathogen interac-

tions. Another important feature of trained machine-learning

models is the prediction of properties for newly acquired sam-

ples, such as predicting the pathogenic potential of newly iden-

tified E. coli strains based on their glycans. As glycobiology pro-

gresses, SugarBase and our deep-learning models could be

readily expanded and updated, enabling an even more compre-

hensive investigation of glycan-mediated host-microbe interac-

tions. This will eventually allow for precise classification at the

subspecies level using language-model-based approaches,

facilitating the glycan-based study of host-microbe interactions

at unprecedented resolution.

STAR+METHODS

Detailed methods are provided in the online version of this paper

and include the following:

d KEY RESOURCES TABLE

d RESOURCE AVAILABILITY

B Lead Contact

B Materials Availability

B Data and Code Availability

d METHOD DETAILS

B Dataset

B Data Processing

B Analyzing Links in Glycan Sequences

B Glycan In Silico Modification

B Glycan Alignment

B Model Training

d QUANTIFICATION AND STATISTICAL ANALYSIS

SUPPLEMENTAL INFORMATION

Supplemental Information can be found online at https://doi.org/10.1016/j.

chom.2020.10.004.

ACKNOWLEDGMENTS

The authors would like to thank Jacqueline Valeri and Mathieu Groussin for

helpful discussions. This work was supported by the Predictive BioAnalytics

Initiative at the Wyss Institute for Biologically Inspired Engineering.

AUTHOR CONTRIBUTIONS

D.B. conceived the method. D.B., D.M.C., and J.J.C. designed the experi-

ments. D.B. performed the experiments and implemented the method.

R.K.P. developed the SugarBase web tool. D.M.C. and J.J.C. supervised the

work. D.B., R.K.P., D.M.C., and J.J.C. wrote and edited the manuscript.

142 Cell Host & Microbe 29, 132–144, January 13, 2021

DECLARATION OF INTERESTS

The authors declare no competing interests.

Received: June 29, 2020

Revised: September 9, 2020

Accepted: October 8, 2020

Published: October 28, 2020

REFERENCES

Alley, E.C., Khimulya, G., Biswas, S., AlQuraishi, M., and Church, G.M. (2019).

Unified rational protein engineering with sequence-based deep representation

learning. Nat. Methods 16, 1315–1322.

Almagro Armenteros, J.J., Johansen, A.R., Winther, O., and Nielsen, H. (2020).

Language modelling for biological sequences – curated datasets and base-

lines. bioRxiv. https://doi.org/10.1101/2020.03.09.983585.

Banks, K.E., Fortney, K.R., Baker, B., Billings, S.D., Katz, B.P., Munson, R.S.,

Jr., and Spinola, S.M. (2008). The enterobacterial common antigen-like gene

cluster of Haemophilus ducreyi contributes to virulence in humans. J. Infect.

Dis. 197, 1531–1536.

Bardor, M., Faveeuw, C., Fitchette, A.-C., Gilbert, D., Galas, L., Trottein, F.,

Faye, L., and Lerouge, P. (2003). Immunoreactivity in mammals of two typical

plant glyco-epitopes, core alpha(1,3)-fucose and core xylose. Glycobiology

13, 427–434.

Bashir, S., Leviatan Ben Arye, S., Reuven, E.M., Yu, H., Costa, C., Galiñanes,

M., Bottio, T., Chen, X., and Padler-Karavani, V. (2019). Presentation Mode of

Glycans Affect Recognition of Human Serum anti-Neu5Gc IgG Antibodies.

Bioconjug. Chem. 30, 161–168.

Bovin, N., Obukhova, P., Shilova, N., Rapoport, E., Popova, I., Navakouski, M.,

Unverzagt, C., Vuskovic, M., and Huflejt, M. (2012). Repertoire of human nat-

ural anti-glycan immunoglobulins. Do we have auto-antibodies? Biochim.

Biophys. Acta 1820, 1373–1382.

Camacho, D.M., Collins, K.M., Powers, R.K., Costello, J.C., and Collins, J.J.

(2018). Next-Generation Machine Learning for Biological Networks. Cell 173,

1581–1592.

Campbell, M.P., Peterson, R., Mariethoz, J., Gasteiger, E., Akune, Y., Aoki-

Kinoshita, K.F., Lisacek, F., and Packer, N.H. (2014). UniCarbKB: building a

knowledge platform for glycoproteomics. Nucleic Acids Res. 42, D215–D221.

Carlin, A.F., Uchiyama, S., Chang, Y.-C., Lewis, A.L., Nizet, V., and Varki, A.

(2009). Molecular mimicry of host sialylated glycans allows a bacterial path-

ogen to engage neutrophil Siglec-9 and dampen the innate immune response.

Blood 113, 3333–3336.

Casadevall, A. (2017). The Pathogenic Potential of a Microbe. MSphere 2,

e00015–e00017.

Day, C.J., Tran, E.N., Semchenko, E.A., Tram, G., Hartley-Tassell, L.E., Ng,

P.S.K., King, R.M., Ulanovsky, R., McAtamney, S., Apicella, M.A., et al.

(2015). Glycan:glycan interactions: High affinity biomolecular interactions

that can mediate binding of pathogenic bacteria to host cells. Proc. Natl.

Acad. Sci. USA 112, E7266–E7275.

Dekkers, G., Treffers, L., Plomp, R., Bentlage, A.E.H., de Boer, M., Koeleman,

C.A.M., Lissenberg-Thunnissen, S.N., Visser, R., Brouwer, M., Mok, J.Y., et al.

(2017). Decoding the Human Immunoglobulin G-Glycan Repertoire Reveals a

Spectrum of Fc-Receptor- and Complement-Mediated-Effector Activities.

Front. Immunol. 8, 877.

Do�gan, T., and Karaçalı, B. (2013). Automatic identification of highly conserved family regions and relationships in genome wide datasets including remote

protein sequences. PLoS One 8, e75458.

Dotan, N., Altstock, R.T., Schwarz, M., and Dukler, A. (2006). Anti-glycan an-

tibodies as biomarkers for diagnosis and prognosis. Lupus 15, 442–450.

Geisinger, E., and Isberg, R.R. (2015). Antibiotic modulation of capsular exo-

polysaccharide and virulence in Acinetobacter baumannii. PLoS Pathog. 11,

e1004691.

Gilbreath, J.J., Colvocoresses Dodds, J., Rick, P.D., Soloski, M.J., Merrell,

D.S., and Metcalf, E.S. (2012). Enterobacterial common antigen mutants of

ll OPEN ACCESSResource

Salmonella enterica serovar Typhimurium establish a persistent infection and

provide protection against subsequent lethal challenge. Infect. Immun. 80,

441–450.

Glorot, X., and Bengio, Y. (2010). Understanding the difficulty of training deep

feedforward neural networks, in: Proceedings of the Thirteenth International

Conference on Artificial Intelligence and Statistics. Presented at the

Proceedings of the Thirteenth International Conference on Artificial

Intelligence and Statistics, pp. 249–256.

Greenfield, L.K., Richards, M.R., Li, J., Wakarchuk, W.W., Lowary, T.L., and

Whitfield, C. (2012). Biosynthesis of the polymannose lipopolysaccharide O-

antigens from Escherichia coli serotypes O8 and O9a requires a unique com-

bination of single- and multiple-active site mannosyltransferases. J. Biol.

Chem. 287, 35078–35091.

Haines-menges, B.L., Whitaker, W.B., Lubin, J.B., and Boyd, E.F. (2015). Host

Sialic Acids: A Delicacy for the Pathogen with Discerning Taste. In Metabolism

and Bacterial Pathogenesis, C. Conway, ed. (American Society of

Microbiology), pp. 321–342.

Haltiwanger, R.S., and Lowe, J.B. (2004). Role of glycosylation in develop-

ment. Annu. Rev. Biochem. 73, 491–537.

Hochreiter, S., and Schmidhuber, J. (1997). Long short-term memory. Neural

Comput. 9, 1735–1780.

Hong, Y., and Reeves, P.R. (2014). Diversity of o-antigen repeat unit structures

can account for the substantial sequence variation of wzx translocases.

J. Bacteriol. 196, 1713–1722.

Howard, J., and Ruder, S. (2018). Universal Language Model Fine-tuning for

Text Classification. arXiv.

Kappler, K., and Hennet, T. (2020). Emergence and significance of carbohy-

drate-specific antibodies. Genes Immun. 21, 224–239.

Khasbiullina, N.R., Shilova, N.V., Navakouski, M.J., Nokel, A.Yu., Blixt, O.,

Kononov, L.O., Knirel, Y.A., and Bovin, N.V. (2019). The Repertoire of

Human Antiglycan Antibodies and Its Dynamics in the First Year of Life.

Biochemistry (Mosc.) 84, 608–616.

Kiser, K.B., and Lee, J.C. (1998). Staphylococcus aureus cap5O and cap5P

genes functionally complement mutations affecting enterobacterial com-

mon-antigen biosynthesis in Escherichia coli. J. Bacteriol. 180, 403–406.

Knirel, Y.A. (2011). Structure of O-Antigens. In Bacterial Lipopolysaccharides,

Y.A. Knirel and M.A. Valvano, eds. (Springer Vienna), pp. 41–115.

Lairson, L.L., Henrissat, B., Davies, G.J., and Withers, S.G. (2008).

Glycosyltransferases: structures, functions, and mechanisms. Annu. Rev.

Biochem. 77, 521–555.

Lauc, G., Kri�sti�c, J., and Zoldo�s, V. (2014). Glycans - the third revolution in evo-

lution. Front. Genet. 5, 145.

Lavine, C.L., Lao, S., Montefiori, D.C., Haynes, B.F., Sodroski, J.G., and Yang,

X.; NIAID Center for HIV/AIDS Vaccine Immunology (CHAVI) (2012). High-

mannose glycan-dependent epitopes are frequently targeted in broad neutral-

izing antibody responses during human immunodeficiency virus type 1 infec-

tion. J. Virol. 86, 2153–2164.

Lim, J.Y., Yoon, J., and Hovde, C.J. (2010). A brief overview of Escherichia coli

O157:H7 and its plasmid O157. J. Microbiol. Biotechnol. 20, 5–14.

Liu, L., Chopra, P., Li, X., Wolfert, M.A., Tompkins, S.M., and Boons, G.-J.

(2020). SARS-CoV-2 spike protein binds heparan sulfate in a length- and

sequence-dependent manner. bioRxiv. 2020.05.10.087288. https://doi.org/

10.1101/2020.05.10.087288.

Lombard, V., Golaconda Ramulu, H., Drula, E., Coutinho, P.M., and Henrissat,

B. (2014). The carbohydrate-active enzymes database (CAZy) in 2013. Nucleic

Acids Res. 42, D490–D495.

Lundberg, S.M., and Lee, S.-I. (2017). A Unified Approach to Interpreting

Model Predictions. In Advances in Neural Information Processing Systems,

Volume 30, I. Guyon, U.V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S.

Vishwanathan, and R. Garnett, eds. (Curran Associates, Inc), pp. 4765–4774.

McDonald, A.G., Tipton, K.F., and Davey, G.P. (2016). A Knowledge-Based

System for Display and Prediction of O-Glycosylation Network Behaviour in

Response to Enzyme Knockouts. PLoS Comput. Biol. 12, e1004844.

Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013). Efficient Estimation of

Word Representations in Vector Space. arXiv.

Mitchell, A.M., Srikumar, T., and Silhavy, T.J. (2018). Cyclic Enterobacterial

Common Antigen Maintains the Outer Membrane Permeability Barrier of

Escherichia coli in a Manner Controlled by YhdP. mBio 9, e01321-18.

Needleman, S.B., and Wunsch, C.D. (1970). A general method applicable to

the search for similarities in the amino acid sequence of two proteins.

J. Mol. Biol. 48, 443–453.

Park, D., Xu, G., Barboza, M., Shah, I.M., Wong, M., Raybould, H., Mills, D.A.,

and Lebrilla, C.B. (2017). Enterocyte glycosylation is responsive to changes in

extracellular conditions: implications for membrane functions. Glycobiology

27, 847–860.

Paschinger, K., Fabini, G., Schuster, D., Rendi�c, D., and Wilson, I.B.H. (2005).

Definition of immunogenic carbohydrate epitopes. Acta Biochim. Pol. 52,

629–632.

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen,

T., Lin, Z., Gimelshein, N., Antiga, L., et al. (2019). PyTorch: An Imperative

Style, High-Performance Deep Learning Library. arXiv.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,

Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., et al. (2011). Scikit-learn:

Machine learning in Python. J. Mach. Learn. Res. 12, 2825–2830.

Perez, L., and Wang, J. (2017). The Effectiveness of Data Augmentation in

Image Classification using Deep Learning. arXiv.

Pochechueva, T., Jacob, F., Fedier, A., and Heinzelmann-Schwarz, V. (2012).

Tumor-associated glycans and their role in gynecological cancers: acceler-

ating translational research by novel high-throughput approaches.

Metabolites 2, 913–939.

Poole, J., Day, C.J., von Itzstein, M., Paton, J.C., and Jennings, M.P. (2018).

Glycointeractions in bacterial pathogenesis. Nat. Rev. Microbiol. 16, 440–452.

Reusch, D., and Tejada, M.L. (2015). Fc glycans of therapeutic antibodies as

critical quality attributes. Glycobiology 25, 1325–1334.

Samraj, A.N., Bertrand, K.A., Luben, R., Khedri, Z., Yu, H., Nguyen, D., Gregg,

C.J., Diaz, S.L., Sawyer, S., Chen, X., et al. (2018). Polyclonal human anti-

bodies against glycans bearing red meat-derived non-human sialic acid N-gly-

colylneuraminic acid are stable, reproducible, complex and vary between indi-

viduals: Total antibody levels are associated with colorectal cancer risk. PLoS

One 13, e0197464.

Sherstinsky, A. (2020). Fundamentals of Recurrent Neural Network (RNN) and

Long Short-Term Memory (LSTM) network. Phys. Nonlinear Phenom. 404,

132306.

Silipo, A., and Molinaro, A. (2010). The Diversity of the Core Oligosaccharide in

Lipopolysaccharides. In Endotoxins: Structure, Function and Recognition, X.

Wang and P.J. Quinn, eds. (Springer Netherlands), pp. 69–99.

Solá, R.J., and Griebenow, K. (2009). Effects of glycosylation on the stability of

protein pharmaceuticals. J. Pharm. Sci. 98, 1223–1245.

Spahn, P.N., Hansen, A.H., Hansen, H.G., Arnsdorf, J., Kildegaard, H.F., and

Lewis, N.E. (2016). A Markov chain model for N-linked protein glycosylation–

towards a low-parameter tool for model-driven glycoengineering. Metab.

Eng. 33, 52–66.

Strodthoff, N., Wagner, P., Wenzel, M., and Samek, W. (2020). UDSMProt: uni-

versal deep sequence models for protein classification. Bioinformatics 36,

2401–2409.

Tan, C., Sun, F., Kong, T., Zhang, W., Yang, C., and Liu, C. (2018). A Survey on

Deep Transfer Learning. arXiv.

Tanaka, K., Aoki-Kinoshita, K.F., Kotera, M., Sawaki, H., Tsuchiya, S., Fujita,

N., Shikanai, T., Kato, M., Kawano, S., Yamada, I., and Narimatsu, H. (2014).

WURCS: the Web3 unique representation of carbohydrate structures.

J. Chem. Inf. Model. 54, 1558–1566.

Thompson, A.J., de Vries, R.P., and Paulson, J.C. (2019). Virus recognition of

glycan receptors. Curr. Opin. Virol. 34, 117–129.

Tiemeyer, M., Aoki, K., Paulson, J., Cummings, R.D., York, W.S., Karlsson,

N.G., Lisacek, F., Packer, N.H., Campbell, M.P., Aoki, N.P., et al. (2017).

GlyTouCan: an accessible glycan structure repository. Glycobiology 27,

915–919.

Cell Host & Microbe 29, 132–144, January 13, 2021 143

ll OPEN ACCESS Resource

Toukach, P.V., and Egorova, K.S. (2016). Carbohydrate structure database

merged from bacterial, archaeal, plant and fungal parts. Nucleic Acids Res.

44 (D1), D1229–D1236.

Tsuchiya, S., Yamada, I., and Aoki-Kinoshita, K.F. (2019).

GlycanFormatConverter: a conversion tool for translating the complexities of

glycans. Bioinformatics 35, 2434–2440.

Tzianabos, A.O., Wang, J.Y., and Lee, J.C. (2001). Structural rationale for the

modulation of abscess formation by Staphylococcus aureus capsular poly-

saccharides. Proc. Natl. Acad. Sci. USA 98, 9365–9370.

Valeri, J.A., Collins, K.M., Ramesh, P., Alcantar, M.A., Lepe, B.A., Lu, T.K., and

Camacho, D.M. (2020). Sequence-to-function deep learning frameworks for

engineered riboregulators. Nat Commun 11, 5058, https://doi.org/10.1038/

s41467-020-18676-2.

Varki, A. (2017). Biological roles of glycans. Glycobiology 27, 3–49.

144 Cell Host & Microbe 29, 132–144, January 13, 2021

Varki, A., and Gagneux, P. (2015). Biological Functions of Glycans. In

Essentials of Glycobiology, A. Varki, R.D. Cummings, J.D. Esko, P. Stanley,

G.W. Hart, M. Aebi, A.G. Darvill, T. Kinoshita, N.H. Packer, and J.H.

Prestegard, et al., eds. (Cold Spring Harbor Laboratory Press).

Viljanen, M.K., Peltola, T., Junnila, S.Y., Olkkonen, L., J€arvinen, H., Kuistila, M.,

and Huovinen, P. (1990). Outbreak of diarrhoea due to Escherichia coli

O111:B4 in schoolchildren and adults: association of Vi antigen-like reactivity.

Lancet 336, 831–834.

Weidenmaier, C., and Lee, J.C. (2015). Structure and Function of Surface

Polysaccharides of Staphylococcus aureus. In Staphylococcus Aureus, F.

Bagnoli, R. Rappuoli, and G. Grandi, eds. (Springer International Publishing),

pp. 57–93.

Wu, D., Struwe, W.B., Harvey, D.J., Ferguson, M.A.J., and Robinson, C.V.

(2018). N-glycan microheterogeneity regulates interactions of plasma pro-

teins. Proc. Natl. Acad. Sci. USA 115, 8763–8768.

ll OPEN ACCESSResource

STAR+METHODS

KEY RESOURCES TABLE

REAGENT or RESOURCE SOURCE IDENTIFIER

Software and Algorithms

PyTorch Paszke et al., 2019 https://github.com/pytorch/pytorch

Scikit-learn Pedregosa et al., 2011 https://github.com/scikit-learn/scikit-learn

Apex N/A https://github.com/NVIDIA/apex

Python-alignment N/A https://github.com/eseraygun/python-alignment

SHAP Lundberg and Lee, 2017 https://github.com/slundberg/shap

SweetTalk This paper https://github.com/midas-wyss/sweettalk

SweetOrigins This paper https://github.com/midas-wyss/sweetorigins

SugarBase This paper https://webapps.wyss.harvard.edu/sugarbase

RESOURCE AVAILABILITY

Lead Contact Communication should be directed to the lead contact, James J. Collins (jimjc@mit.edu).

Materials Availability This study did not generate new unique reagents.

Data and Code Availability Data used for all analyses can be found in the supplementary tables. All code and trained models can be found at https://github.com/

midas-wyss/sweettalk and https://github.com/midas-wyss/sweetorigins.

METHOD DETAILS

Dataset To create a comprehensive glycan dataset annotated with species labels, we manually curated 12,674 glycan sequences from three

sources: UniCarbKB (Campbell et al., 2014), the Carbohydrate Structure Database (CSDB) (Toukach and Egorova, 2016), and the

peer-reviewed scientific literature. From UniCarbKB, we compiled all glycans with species information, a length of at least three

monosaccharides to facilitate usage with machine learning models, and a working link to PubChem to retrieve their sequences.

We further complemented and extended this list by gathering glycans deposited in the Carbohydrate Structure Database (CSDB)

up to December 2019 with a length of at least three monosaccharides. For species with more than 15 strains available on CSDB,

only glycans from the first 15 strains were recorded to prevent taxonomic bias. For the model organism E. coli, all available glycan

sequences were recorded to facilitate a strain-based analysis. Labels for E. coli strain pathogenicity were assigned, if possible, via

the peer-reviewed academic literature. Finally, we performed additional literature searches, predominantly adding viral and archaeal

glycans, which are underrepresented in the other databases. We revised and completed the annotations for all species’ taxonomic

characterization (species, genus, family, order, class, phylum, kingdom, domain) based on the NCBI Taxonomy Browser. In total, the

dataset contained sequences from 1,726 different species from a range of 39 taxonomic phyla. To the best of our knowledge, this

database represents the most comprehensive and current resource of glycans and their species information to date (Table S1).

To enable transfer learning by first pre-training a language model, we also added glycan sequences that lacked species informa-

tion, by extracting the Web3 Unique Representation of Carbohydrate Structures (WURCS) representation (Tanaka et al., 2014) of the

set of all glycans with at least three monosaccharides deposited on GlyTouCan (Tiemeyer et al., 2017) that were also available on

PubChem (n = 18,926) and the databases mentioned above; this resulted in an augmented database containing 19,299 unique glycan

sequences (Table S2). For all glycans, we relied on the quality control of the respective database. All glycans in WURCS represen-

tation were reformatted into the IUPAC condensed representation, using the GlycanFormatConverter software (Tsuchiya et al.,

2019). For the immunogenicity classifier, all GlycoEpitope (https://www.glycoepitope.jp) entries with a minimum length of at least

three monosaccharides were extracted. This list was further complemented by targeted literature searches (Bardor et al., 2003; Ba-

shir et al., 2019; Bovin et al., 2012; Dotan et al., 2006; Hong and Reeves, 2014; Khasbiullina et al., 2019; Knirel, 2011; Paschinger et al.,

2005; Pochechueva et al., 2012; Samraj et al., 2018; Silipo and Molinaro, 2010) resulting in the final set of immunogenic glycans

(n = 685, Table S2). We included protein-, lipid-, and small molecule-associated glycans as well as capsular and extracellular

Cell Host & Microbe 29, 132–144.e1–e3, January 13, 2021 e1

ll OPEN ACCESS Resource

polysaccharides in our dataset of 19,299 glycans. All these glycans were paired with an ID to allow for our relational database Sugar-

Base, linking all available information (linkage type, species information, human immunogenicity, etc.) to a glycan sequence (Table

S2). Additionally, we included representations learned by our language model for all observed glycoletters (monosaccharides or

bonds) as well as glycowords (trisaccharides).

Data Processing Glycan sequences were processed by removing dangling bonds (e.g., ‘(a1-’). Analogous to word stemming in natural language pro-

cessing, unifying different inflections of the same word, we removed position-specific information of monosaccharide modifications

to reduce vocabulary size. Then, we harmonized capitalization and, in the case of glycan repeat structures, appended the first mono-

saccharide to their end to capture more sequence context. Additional steps to exclude duplicated glycans included strict ordering of

multiple branches with equal lengths by ascending connection to the main branch (e.g., branch ending in ‘a1-2’ before branch ending

in ‘b1-4’). For branches closest to the non-reducing end, the longest branch was defined as the main chain. Observed monosaccha-

ride modifications necessitated a hierarchy of order (in case of multiple modifications on the same monosaccharide) to avoid dupli-

cates or mislabeling: NAc > OAc > NGc > OGc > NS > OS > NP > OP > NAm > OAm > NBut > OBut > NProp > OProp > NMe > OMe >

CMe > NFo > OFo > OPPEtn > OPEtn > OEtn > A > N > SH > OPCho > OPyr > OVac > OPam > OEtg > OFer > OSin > OAep > OCoum >

ODco > OLau > OSte > OOle > OBz > OCin > OAch > OMal > OMar > OOrn > rest.

Data processing for model training included featurization of glycan sequences into glycoletters (e.g., ‘Gal’), as well as glycowords

(three monosaccharides connected by two bonds). The conversion of a glycan sequence into glycowords, from the non-reducing to

the reducing end, resulted in a list of partially overlapping glycowords, with maximum overlap so that two subsequent glycowords

only differed in one monosaccharide and one bond. The aim of these glycowords is to capture representative characteristics and

local structural contexts of a given glycan. The dataset comprising all glycowords (n = 113,112) was then used to train a context-spe-

cific, glycoletter-based language model. For scrambled glycan sequences, the order of glycoletters in any given glycan was randomly

shuffled to maintain composition but erase patterns. All abbreviations for glycan nomenclature in this work can be found in Table S7.

Analyzing Links in Glycan Sequences To determine typical local structural contexts of monosaccharides and bonds, we quantified the frequency of a given monosaccha-

ride co-occurring with any other monosaccharide in our extensive database of unique glycans. Additionally, we also compared the

relative frequencies of a particular monosaccharide being observed in the glycan main branch versus a side branch in our database.

Glycan In Silico Modification We performed in silico modification of glycans by replacing monosaccharides and/or bonds with other observed monosaccharides/

bonds. We used exhaustive modification, replacing glycoletters with all possible glycoletters, while only retaining modified glycans

comprising previously observed glycowords. This ensured physiological relevance, given the extreme sparsity of observed glycan

sequences compared to the theoretical number of possibilities.

Glycan Alignment Global sequence alignment of glycans was implemented according to the Needleman-Wunsch algorithm (Needleman and Wunsch,

1970) by adapting the Python Alignment library (https://github.com/eseraygun/python-alignment). For our GLYcan SUbstitution Ma-

trix (GLYSUM; Table S6), the exhaustive list of in silico modifications resulting in glycans with observed glycowords was generated

(n = 1,238,879). All thereby observed monosaccharide and/or bond substitutions were recorded in a symmetric matrix and converted

into substitution frequencies by dividing them by the total number of retained modifications. The substitution score Sij for each

possible substitution was then calculated with the following formula:

Sij = l log

� pij

qi � qj

�

The substitution frequency is hereby denoted as pij, while qi and qj describe the observed base frequencies of the respective gly-

coletters. Additionally, we used l as a scaling factor (a value of four in this work) to arrive at suitable integer values by rounding all

values up or down. Substitutions never observed during this procedure received a final value of �5, lower than any of the observed substitution scores, while the diagonal values of the substitution matrix were set at 5, higher than any of the observed substitution

scores. The penalty for gaps for alignments in this work was set at �5, to match the minimal substitution score.

Model Training All models were trained on an NVIDIA� Tesla� K80 GPU using PyTorch (Paszke et al., 2019). For all models, architecture and hyper- parameters were optimized by minimizing the respective loss function. For the language models, we used mixed precision training

utilizing the Apex library (https://github.com/nvidia/apex). For language models and classifiers, we randomly split the respective da-

taset into 80% for training and 20% for validation. A modified stratified shuffle split was used to randomly split glycans into training

and validation sets for the species classifier so that, for every class, 80% of the glycans were present in the training set and 20% in the

validation set. Further, only classes comprising at least five glycans were used for training and testing the SweetOrigins models. We

e2 Cell Host & Microbe 29, 132–144.e1–e3, January 13, 2021

ll OPEN ACCESSResource

employed data augmentation by forming a generalizable subset of all possible isomorphic glycans if a glycan sequence had isomor-

phic glycans. Specifically, we swapped the order of double branches and exhaustively exchanged the main branch with the side

branches closest to the non-reducing end in the bracket notation (Figure 3B). The resulting sequence in the bracket notation still

described the same glycan in a slightly different way, increasing model robustness during training. Glycans were converted into lists

of glycowords describing the glycans, brought to equal lengths using a padding token facilitating model training, and used in batches

of 32 glycans for training and testing.

SweetTalk and the SweetOrigins models for each taxonomic level consisted of a three-layered, bidirectional recurrent neural

network using long short-term memory (LSTM) units (Hochreiter and Schmidhuber, 1997) with 128 nodes per layer, including an

embedding layer for the glycowords. The concatenated hidden representation learned by the bidirectional LSTMs was then pro-

jected to a fully connected layer at the end for the final prediction. The language model SweetTalk was trained by predicting the

next glycoletters, given preceding glycoletters, in the context of glycowords, thereby learning the local structural context of glyco-

letters. The embedding layer for classifiers was derived by first training a glycoletter-based language model and then extracting the

learned glycoletters embedding and calculating initial glycoword embeddings for SweetOrigins. The last, fully connected layer in all

models was initialized by Xavier initialization (Glorot and Bengio, 2010) and the number of nodes was determined by the number of

classes for each classifier. We used a cross-entropy loss function and the ADAM optimizer with a starting learning rate of 0.0001

(decaying it with a cosine function over 100 epochs during training) and a weight decay of 0.005. Additionally, we employed an early

stopping criterion after 10 epochs without improvement in validation loss for regularization.

The model for predicting E. coli strain pathogenicity followed the same architecture except for using 150 nodes per layer, a binary

cross-entropy loss function, and a learning rate of 0.0005. Machine learning models used for comparison comprised random forest

classifiers and support vector machines for classification. For the implementation of these models, we used the scikit-learn imple-

mentation (Pedregosa et al., 2011). Feature importances were extracted using SHAP (SHapley Additive exPlanations) values (Lund-

berg and Lee, 2017). Hyperparameters for all methods were optimized by maximization of accuracy via 5-fold cross-validation.

QUANTIFICATION AND STATISTICAL ANALYSIS

This study did not use statistical analysis. All experimental details can be found in the STAR Methods section.

Cell Host & Microbe 29, 132–144.e1–e3, January 13, 2021 e3