df_fabales = df_species[df_species.Order == 'Fabales'].reset_index(drop = True)Overview
Glycans are a fundamental biological sequence, similar to DNA, RNA, or proteins. Glycans are complex carbohydrates that can form branched structures comprising monosaccharides and linkages as constituents. Despite being conspicuously absent from most research, glycans are ubiquitous in biology. They decorate most proteins and lipids and direct the stability and functions of biomolecules, cells, and organisms. This also makes glycans relevant to every human disease.
The analysis of glycans is made difficult by their nonlinearity and their astounding diversity, given the large number of monosaccharides and types of linkages. Glycowork is a Python package designed to process and analyze glycan sequences, with a special emphasis on glycan-focused machine learning. Next to various functions to work with glycans, glycowork also contains some of the largest curated glycan datasets in the field, that can be used for model training, motif comparisons, composition matching, etc.
Where possible, glycowork uses the IUPAC-condensed nomenclature for glycans. However, since v1.0, glycowork broadly supports IUPAC-extended, LinearCode, Oxford, GlycoCT, and WURCS as input formats as well. No conversions or special settings are necessary.
If you do want to use IUPAC-condensed (still recommended), here are some best practices: An example for this would be Man(a1-3)[Man(a1-6)]Man(b1-4)GlcNAc(b1-4)GlcNAc, in which parentheses enclose the linkage description and brackets indicate branches. Ideally, the longest branch should constitute the main chain and side branches should be ordered ascendingly based on the last linkage of the branch (yet all functions in glycowork can deal with violations to these guidelines). Fragments with unclear attachment points are added within curly brackets. Where possible/known, monosaccharide modifications are indicated by the ring position to which they are attached (e.g., Glc6Me), while the type of connection is given if the exact position is unknown (e.g., GlcOMe). Linkage uncertainty is expressed with a ‘?’ placeholder (e.g., a1-6, ?1-6, a1-?). As suggested in IUPAC-condensed, we omit the D-/L-prefix for the “usual” case and only add it for the rarer option (e.g., Glc vs. L-Glc).
However, it should be noted that nearly all functions in glycowork operate on pure graphs (facilitated by glycowork.motif.graph.glycan_to_nxGraph), avoiding all ambiguity of the IUPAC nomenclature. We use the IUPAC nomenclature here as a human-readable glycan nomenclature that can be readily used for motif interpretation and user-friendly usage. The conversion to graphs in all end user-functions and operations, such as motif annotation via highly optimized subgraph isomorphism tests, are handled behind the scenes, allowing you to work with glycans in a human-readable format.
All modules in glycowork, except for ml, can be run on any machine. For most parts of ml, however, a GPU is needed to load and run torch_geometric.
If you want to run glycowork.ml in a Jupyter Notebook (in the cloud or wherever you have a GPU), just install glycowork with its optional dependencies for machine learning:
!pip install "glycowork[ml]"Okay, how about an example! (For more example workflows, head over to the examples section)
Let’s say you’re interested in glycans of the plant order Fabales that includes the legumes. Luckily, we’ve got you covered and df_species has >700 glycan sequences that stem from Fabales. So the first step is easy: import df_species from glycowork.glycan_data.loader and filter for Fabales glycans.
| Species | Genus | Family | Order | Class | Phylum | Kingdom | Domain | ref | |
|---|---|---|---|---|---|---|---|---|---|
| glycan | |||||||||
| Man(a1-3)Man(a1-6)[Xyl(b1-2)][Man(a1-3)]Man(a1-6)Man(b1-4)GlcNAc(b1-4)GlcNAc | Abrus_precatorius | Abrus | Fabaceae | Fabales | Dicotyledons | Angiosperms | Plantae | Eukarya | nan |
| Man(a1-3)[Man(a1-3)Man(a1-6)][Xyl(b1-2)]Man(b1-4)GlcNAc(b1-4)GlcNAc | Abrus_precatorius | Abrus | Fabaceae | Fabales | Dicotyledons | Angiosperms | Plantae | Eukarya | nan |
| Man(a1-6)[Man(a1-3)]Man(a1-6)[Man(a1-3)]Man(b1-4)GlcNAc(b1-4)GlcNAc | Abrus_precatorius | Abrus | Fabaceae | Fabales | Dicotyledons | Angiosperms | Plantae | Eukarya | nan |
| Man(a1-6)[Xyl(b1-2)][Man(a1-3)]Man(a1-6)Man(b1-4)GlcNAc(b1-4)[Fuc(a1-3)]GlcNAc | Abrus_precatorius | Abrus | Fabaceae | Fabales | Dicotyledons | Angiosperms | Plantae | Eukarya | nan |
| Rha(a1-2)Gal(b1-2)GlcA6Me | Abrus_precatorius | Abrus | Fabaceae | Fabales | Dicotyledons | Angiosperms | Plantae | Eukarya | nan |
Now, if you’re not a plant glycobiologist, that may seem a bit overwhelming to you. But fear not, we can just plot glycans. How? Easy: Glycan representations!
Representations or embeddings are really just a bunch of numbers learned by a machine learning model that express similarity. Meaning that if two glycans are similar (in sequence/properties/etc.), the difference between their representations should be small. Or, in other words, they should cluster in a plot.
Now, you might not be in possession of a GPU and/or the wherewithal to train a glycan-focused deep learning model. That’s why glycowork has both trained models and already calculated glycan representations waiting for you! And the best thing: you don’t even need to worry about them. You just pass glycan sequences to our functions and glycowork will handle glycan representations behind the scenes.
So, for our Fabales example, we pass our glycan sequences to glycowork.motif.analysis.plot_embeddings, which will retrieve their learned representations, transform them into two dimensions via t-distributed stochastic neighbor embedding (t-SNE), and plot them! If you want, it can even color the resulting scatterplots by a label, for instance the taxonomic family within Fabales.
plot_embeddings(df_fabales.glycan.values.tolist(), label_list = df_fabales.Family.values.tolist())
Hmm, interesting. We definitely see clear glycan clusters (remember, every point is one glycan sequence). What’s even cooler is that the families Polygalaceae and Quillajaceae cluster in narrow regions of the whole plot. I wonder what differentiates their glycans from Fabaceae?
If you’re also curious, proceed to the next function. First, we need to convert our data into a count table, with counts of every glycan in every family of Fabales. For this, you can use glycowork.motif.processing.presence_to_matrix, which was designed to work with tables such as df_species. Changing label_col_name to ‘Family’ will for instance generate a count table for all families in Fabales.
df_map = presence_to_matrix(df_fabales, label_col_name = 'Family').T| Family | Fabaceae | Fagaceae | Polygalaceae | Quillajaceae |
|---|---|---|---|---|
| glycan | ||||
| Apif(a1-2)Xyl(b1-2)[Glc6Ac(b1-4)]Glc | 1 | 0 | 0 | 0 |
| Ara(a1-2)Ara(a1-6)GlcNAc | 4 | 0 | 0 | 0 |
| Ara(a1-2)Glc(b1-2)Ara | 1 | 0 | 0 | 0 |
| Ara(a1-2)GlcA | 3 | 0 | 0 | 0 |
| Ara(a1-2)[Glc(b1-6)]Glc | 1 | 0 | 0 | 0 |
Okay, we’re ready for some heatmaps! In glycowork.motif.analysis.get_heatmap, we have a function which…makes heatmaps.
Specifically, it will generate clustermaps, that are aiming to cluster samples based on their glycan profile. Calling get_heatmap with df_map without any other arguments will try to cluster the samples based on their full glycan sequences. But we also have the motif mode, which extracts motifs from these glycans and clusters samples based on those. For instance, you can use the known (curated glycan motifs) and the exhaustive (all observed monosaccharides and disaccharides) options. Check out the documentation for more options!
We can immediately see that we have a lot more N-linked glycans in Fabaceae, for instance based on the prominence of Trimannosylcore and Fuc(a1-3)GlcNAc. Glycans from Polygalaceae and Quillajaceae in df_species seem to mainly contain small molecule- and lipid-linked glycans, which of course are quite different from N-linked glycans and are more similar to each other than to glycans from Fabaceae in df_species.
get_heatmap(df_map, motifs = True, feature_set = ['exhaustive'], datatype = 'presence',
yticklabels = 1)
Let’s say you’re interested in a specific glycan from Fabales, GlcNAc(b1-2)Man(a1-3)[GlcNAc(b1-2)Man(a1-6)][Xyl(b1-2)]Man(b1-4)GlcNAc(b1-4)[Fuc(a1-3)]GlcNAc. Finding out more about it could be annoying (try to Google anything more complex than a short motif and you run into problems). That’s why we have glycowork.motif.query.get_insight, a function that returns some metadata about that glycan. The nice thing is that you don’t need to bother with database IDs or notational ambiguity, as glycowork will convert that glycan into a graph behind the scenes and search our databases.
Let’s say you develop an acute interest in the distribution of xylose in plant glycans. In such a situation, one option for further inquiry would be glycowork.motif.analysis.characterize_monosaccharide. This neat little function allows you to get a sense of what a particular monosaccharide is typically directly connected to and which modifications it can acquire. Further, you can filter the results to any taxonomic level you’re interested in! (as long as it’s in df_species or your dataset of interest of course.)
So, for xylose in plants, we immediately see Man (in M3X/M3FX structures) and Glc/Xyl (Xyloglucans) as prominent neighbors. Additionally, we note that xylose in plant glycans can be methylated, which is a rather typical modification in plant glycans in general.
There is quite a bit more that you can do with characterize_monosaccharide, so have a look at its documentation for all the options!
characterize_monosaccharide('Xyl', rank = 'Kingdom', focus = 'Plantae', modifications = True)
To show you the true power of characterize_monosaccharide we’ll switch gears to bacteria and have a look at how they handle their glucose connections. This is a lot more diverse than with xylose! Note how we put thresh to 50 (from its default value of 10), as otherwise the plot would become illegible with all the rare modifications and pairings.
characterize_monosaccharide('Glc', rank = 'Kingdom', mode = 'sugarbond',
focus = 'Bacteria', modifications = True, thresh = 50)
And that’s just a super-brief glimpse at the really cool functions contained in glycowork! The rest of the documentation describes all functions and what they’re good for. Be sure to check out more, suggest improvements/fixes, and maybe even contribute something yourself. We are looking forward to your discoveries!