The human large intestine supports an extremely dense and diverse microbial community—up to 100 trillion individuals—known as the gut microbiota (Bäckhed et al., 2005). Over the last decade, the microbiome has been shown to play an important role in human health, and numerous studies have documented the link between microbiota composition and metabolic diseases, such as type 2 diabetes (T2D) (Qin et al., 2012), obesity, colorectal cancer (Thomas et al., 2019), immune response to treatment (Tanoue et al., 2019), and inflammatory bowel diseases such as Crohn’s disease (Henke et al., 2019), and also athletic performance (Scheiman et al., 2019). Microbial dysbiosis has been correlated to modern lifestyle, environmental parameters, medication, and western diet (Mosca et al., 2016). Indeed, one of the main drivers of microbiota composition has been shown to be diet, with long-term differences and fast responses to drastic diet changes, both in the metagenome and in the transcriptome (Filippo et al., 2010; David et al., 2013; Tap et al., 2015). One of the parameters playing a part in the booming number of individuals affected by metabolic disorders is the reduction of polysaccharide diversity in the day-to-day diet. As an example, migration from a non-Western country to the United States has been associated with immediate loss of gut microbiome diversity and function (Vangay et al., 2018). The mechanism of action being the depletion of dietary fibers—a nutrient category that includes a broad array of polysaccharides that are not digestible by human enzymes—in industrialized countries’ diet (Burkitt et al., 1972; Faith et al., 2011; Sonnenburg and Sonnenburg, 2014; Zmora et al., 2018).

However, the responses to a given diet are characterized by a large and not-yet-understood individual variability (Leshem et al., 2020) that complicates the design of specific diets or targeted foods and the understanding of glycan breakdown. The human microbiota produces complementary enzymes enabling the depolymerization and hydrolysis of dietary polysaccharides into sugars that can further be fermented into short-chain fatty acids (SCFAs). The enzymes completing this task, named carbohydrate-active enzymes (CAZymes), are involved in complex metabolic networks for the synthesis [glycosyltransferases (GTs)], degradation [glycoside hydrolases (GHs), polysaccharide lyases (PLs), carbohydrate esterases (CEs), and enzymes for the auxiliary activities (AAs)], and recognition [carbohydrate-binding module (CBM)] of all the carbohydrates on Earth. CAZymes are found in all living organisms (typically 1–3% of the gene content) and are particularly abundant (more than 3% of the gene content) in plants and microbes, e.g., Bacteroides thetaiotaomicron encodes for 391 CAZymes, which represent 8.2% of its genomic/gene content. In humans eating more fiber for 5 days, the expression of GH related to dietary fiber increased while mucus degrading GH were downregulated (Tap et al., 2015). Another argument supporting the link between CAZyme gene presence and strains’ metabolic capacities is the trend between CAZyme gene count and diversity, being almost a classifier of the bacterial genomes’ phylogeny (El Kaoutari et al., 2013). CAZyme gene expression was demonstrated in vitro using RNA-seq when strains were grown on specific polysaccharides (Rey et al., 2010; Martens et al., 2011; Scott et al., 2011). However, except for the Bacteroidetes phylum, there are still very few enzymatic systems being described and characterized. Annotating the glycosyl hydrolase genes and their loci organization/synteny is mandatory to characterize the human gut microbial capacity to breakdown glycans. Today, the CAZyme database of sequences and subfamilies is a reference for such genes. The database is built by an academic laboratory using manual expert editing of the annotation (Cantarel et al., 2008), while an open-source annotation tool called dbCAN is also available (Yin et al., 2012; Zhang et al., 2018).

Annotating CAZyme genes is not an easy task due to the modularity of the gene structure (Cantarel et al., 2008). Beyond difficulties to annotate CAZyme genes, the practical question to further investigate the link between the metabolic disorders (phenotype-level) and the genomic content of the microbiome is still an open question. Several publications reported that the CAZyme genes’ abundance in genomes could account for a richer metabolic network able to degrade different fibers or carbohydrates. Despite the link between specific GH and specific fibers documented by several authors (Zhao et al., 2018; Kovatcheva-Datchary et al., 2019), a consistent and complete list of CAZyme genes associated to a specific fiber is currently missing to any newcomer willing to understand the metabolic capacities of a strain from its genome sequence (as a starting point).

In addition to a complicated underlying link between carbohydrate breakdown and specific CAZy genes, taking into account the polysaccharide utilization loci (PUL) appears critical to understand the capacity of strains to hydrolase carbohydrates, in particular for the Bacteroidetes (Lapébie et al., 2019). Specifically, within the Bacteroidetes phylum, PULs are reported to be the genomic area that encodes the capacity to attach, degrade the fiber, and import oligomers. The term PUL was first coined by Bjursell et al. (2006) to describe clusters of colocalized, coregulated genes that contain functions such as detection, sequestration, enzymatic digestion, and transport of complex carbohydrates (Martens et al., 2009; Grondin et al., 2017). Indeed, the PULs encode a complement of cell surface glycan-binding proteins (SGBPs), TonB-dependent transporters (TBDTs), CAZymes (most frequently GHs and also PLs and CEs), and carbohydrate sensors/transcriptional regulators. Less reported in the literature, a similar structure has also been mentioned for Gram-positives (gpPULs) for butyrate-producing species belonging to the Firmicutes (Sheridan et al., 2015). Therefore, glycosyl hydrolases are encoded within operons that are not taken into account when simply annotating CAZymes. Currently, obtaining PUL annotation from a genome is not straightforward, and we found two available published resources. The PULDB database (Terrapon et al., 2017) of experimentally and non-experimentally proven PULs in Bacteroidetes is built as an extension to the CAZy database. The other one is a prediction tool called PULpy, identifying CAZymes that are co-localized with susCD gene pairs. The authors present their tool as a public version of the PULDB algorithm (Stewart et al., 2018). These current resources have drawbacks for non-experts: the first database has to be queried with either fibers or known species. Therefore, it is hardly usable for any new isolated strain. In addition, some interesting polysaccharides with prebiotic properties, such as inulin, are missing. The other tool could help since it provides an algorithm searching for hits similar to PULDB. However, the output is a prediction of a potential PUL with a number referring to the PULDB number and not specifically to a carbohydrate.

Taking this context into account, the goal of this manuscript is to attempt to bridge the gap between the microbiome metabolic capacities and strains’ genomes using a predictive model. Our approach is based on a simple microbiological standpoint where we consider the strains’ ability to grow on a specific carbohydrate as a measure of their metabolic capacities. This simple proxy allows us to have a simple experimental measure for which we suppose the specific genomic content could be predictive of such capacities. Hence, from a genome sequence, microbiologists could obtain a prediction of the metabolic capacities of a strain without having it grown on the fiber. The proof of principle is based on a benchmark data set that documented the growth differences of strains on different carbohydrates. We show that in taking PUL structures into account, our FiberGrowth tool improves growth prediction in comparison to only relying on CAZyme annotations of single genes.

In this manuscript, we designed a pipeline to predict the growth of Bacteroidetes species from the human gut on seven different polysaccharides using a combination of in silico modeling and validation with microbiology data. To our knowledge, our work provides the first integrated pipeline to use PUL to investigate the growth of human gut strains on specific polysaccharides.

PUL models have already been described by Terrapon et al. (2015), but the approach is different, centered on specific genomes to provide a unique model based on previous biochemical characterizations of the enzymes and proteins involved. For taxonomic assignment and phylogenetic placement of existing GH or new GH, the SACCHARIS pipeline automatically annotates GH and provides accurate phylogenetic functional trees (Jones et al., 2018).

Our hypothesis was that, by using growth and transcriptomics data from the literature, new fiber-specific PUL models could be built and assessed on a distinct growth data set. Hence, the integrated pipeline, including a new automated annotation of PULs, could provide microbiologists, in a minute, with growth predictions from genome sequences. The pipeline showed very different accuracies, depending on the fiber, from excellent, up to 96%, to 38% for mucin, close to a random association.

Despite building a PUL model that followed a good agreement within the scientific community for the GH and associated genes involved in starch hydrolysis, and using available transcriptomics data to build the model, it performed poorly compared to the other models. Anderson and Salyers (1989) first reported in the late 80s that the breakdown of starch by B. thetaiotaomicron involved outer membrane bound-attached starch-binding sites and periplasmic starch-degrading enzymes, rather than only extracellular enzymes. Since then, starch catabolism has been largely characterized, and Foley et al. (2016) described the Sus operon as the model system for starch uptake in Bacteroidetes. The complexity and substrates’ diversity of starch-related polysaccharides hardly fit the CAZyme database. GH13 displays a wide phylogenetic diversity (as described by Stam et al., 2007) that is now classified under GH13 subfamilies in CAZyme. The visualization of the genes’ organization of our PUL models for few Bacteroides genomes is consistent with the large diversity and lack of synteny between close genomes. In the in vitro experiments we considered, starch was from potato starch (Eric Martens’ data), and the results would be different from a different origin or amylose/amylopectin ratio. Furthermore, starch can be found under distinct biochemical structures, in in vitro experiments (RS2 or RS3) and in food: not only starch as found in fruits but also the different resistant starch structures, formed after cooking and cooling down starchy food, or chemically processed resistant starch (RS4). In humans and in pigs, the microbial community composition was found to be linked to the starch structure, emphasizing that variability can be explored and understood only through the use of starches with highly characterized structures (Warren et al., 2018). It is then possible to speculate that other uncharacterized operons for starch breakdown exist in Bacteroidetes genomes, and, as a matter of fact, we detected incomplete PUL, where annotations are missing. Furthermore, starch is a carbohydrate storage in plants, hence, relative to the global evolution, has not been consumed as cooked by humans until recently. Whether or not this large diversity of amylase systems reflects recent evolution remains to be determined. It is worth noting that in other bacterial phyla, genomic systems described for the breakdown of starch differ from the common set of genes typically reported and used in our model. It has been recently demonstrated that in Ruminococcus bromii L2-63 a cell surface amylosome and sporulation capacity exist for starch breakdown in strains from the human and rumen microbiota (Mukhopadhya et al., 2017).

Laminarin is also a storage glucan and the growth prediction was low. The performance was similar to the one for starch, despite a very different situation: our PUL model was inadequate since, based on the methodology and annotation, using GH3 and GH16, we might not have found an appropriate marker GH. However, a seven-gene PUL has been described for Bacteroides uniformis, which includes CBM6/GH3, GH158, and GH16.

Devillé et al. (2007) reported that laminarin modulated the microbiome and increased propionate and butyrate in fermenters, pointing toward an effect on not only Bacteroidetes but also Firmicutes. Furthermore, it modulated mucus composition in rats’ gastrointestinal tract (Devillé et al., 2007). It was recently shown by another research group (Strain et al., 2019) that an addition of laminarin into digesters increased Lachnospiraceae, compared to fructooligosaccharide (FOS) and cellulose. The mucin model led to the lowest prediction performance. Interestingly seven out of the eight non-Bacteroides genomes led to a false prediction. It was built using PULDB and included 11 genes, but several key GH involved in mucin breakdown were not accounted for. The PUL model was based on transcriptomics data, and because of the complexity of mucin, the GH genes over-expressed at this time point were limited. Another issue is cDNA annotation. For instance, one of the main causes of misannotation of the fuc genes is their similarity to the genes for rhamnose utilization. Both the FucK and RhaB proteins belong to the FGGY family of carbohydrate kinases (Pfam: PF02782).

Mucin used in experiments is very frequently pig gastric mucin type III. However, mucin composition depends upon the individual and the GI tract segment, and its complexity has only recently been acknowledged. Mucin is at a crossroad between dietary substrates and human secretion. Indeed, using droplet microfluidics, we recently demonstrated and characterized a new GH enzyme, active on human gangliosides, with similar structure to human mucin and milk oligosides and overrepresented in IBD patients’ metagenomes (Tauzin et al., 2020). The mucin composition and degradation pathways still remain to be elucidated.

Interestingly, the link between mucin and heparin has been shown in vivo: the expression of heparin PUL in mice colonized with B. thetaiotaomicron (B. theta) could only be observed in bacteria occupying the mucosal layer of the gastrointestinal tract, suggesting that in vivo mucus could be a source of heparin in vivo (Li et al., 2015). Furthermore, GH88 included in our heparin PUL model plays a key role in bacteria–mucus interaction: when mice were co-colonized with six other Bacteroidetes strains in addition to the B. theta mutant, the B. theta GH88–mutant was much lower in abundance than in mono-colonized mice, indicating that the ability to degrade heparin is under increased selection pressure for B. theta in the presence of other Bacteroidetes. Heparin is an interesting glycan: it is an anticoagulant drug, with a structure of the glycosaminoglycan family of carbohydrates, bearing similarities with mucin, since the most common disaccharide unit is composed of a 2-O-sulfated iduronic acid and 6-O-sulfated, N-sulfated glucosamine. As a matter of fact, heparins from most commercial preparations originate from beef lung or porcine intestinal mucosa. Our model led to a good growth prediction, and interestingly, a posteriori, we had obtained a PUL structure very similar to the one built by Joglekar et al. (2018). The authors documented the complex synteny for heparin PULs within the Bacteroides genus: for example, in Bacteroides eggerthii and Bacteroides galinarium genomes, another heparin PUL exists, with PL15 but not PL13. The prediction from our model could be expected to be low for strains that do not encode for a PL13 or PL15, such as B. eggerthii DSM20697, B. galinarium DSM 18171, B. stercoris ATCC 43183, and B. YIT 12058. Because the human epithelium contains high amounts of heparin sulfate, the biologically relevant glycan PUL is heparin sulfate, hence with PL13 as a priority gene to target. We did not include GH95, involved in mucin degradation, because its α-L-fucosidase activity is only found in very few heparin PULs, in some B. thetaiotaomicron, B. ovatus, and B. finegoldii genomes.

Growth prediction for inulin shows good performance. FOS are short-chain oligosaccharides that are generated by hydrolysis of the polysaccharide inulin, which is composed of 2–60 fructose monomers. We presented the results of inulin and levan next to each other because they represent two distinct glycosidic linkages (2–6 in levan and 2–1 in inulin) that are present in the fructan homopolymers and that are available to the gut microbiota. Inulin is found in different nutrients such as wheat, onion, garlic, and banana and is the most common used fiber in prebiotics that, when used in combination with other probiotics, is able to promote the growth of specific beneficial gut bacteria such as bifidobacteria (Gibson et al., 1995). GH91, an inulin lyase, has been described as involved in the hydrolysis of inulin. Interestingly, while GH32 appears to be always necessary, our results showed that GH91 is not. A close examination of GH91 indicates that the enzyme activity releases difructofuranosyl 1,2 23 diamyhide that seems kept within the cytoplasmic compartment, consistent with the absence of a signal peptide on the gene sequence (Henrissat, personal communication). The absence of release of fructose and the location of the enzymatic activity seem to indicate that GH91 is not mandatory and that its role might not be on catabolism but potentially for intra-cytoplasm metabolism or storage.

Recent work by the group of Joglekar et al. (2018) on Bt strains VPI-5482 (same strain as used in our study) and Bt-8736 contrasted levan and inulin or fructan operon with GH phylogenetic trees. They demonstrated that related genetic loci can encode diversified biochemical pathways in strains from the same B. thetaiotaomicron species. The presence of GH32, SGBP, and SusD and SusC-like domain, corresponding to outer membrane binding proteins, explained the capacity to grow. Accordingly, in our growth prediction model, B. finegoldii, in the same phylogenetic cluster as Bt VPI-5482 for the SusC and SusD, does not grow on inulin. Furthermore, it has been demonstrated that the presence of the divergent susC/susD gene alone enabled the hybrid Bt(8736-2) strain to outcompete the wild-type strain in vivo in mice fed an inulin diet. This pathway does match our model, which includes enzymes and carbohydrate-binding and import proteins with distinct substrate specificities, which could not have been predicted previously based on sequence data alone.

The discrepancy between databases regarding GH annotation can be highlighted in the case of levan PUL, looking at the domain level. Our results showed that taking into account three GH32 distinct gene copies and a levanase within a PUL model was sufficient to predict growth on levan, but the model over-predicted growth in Bacteroides genomes. However, the PUL was built using transcriptomics data, which raises questions on cross-annotation of cDNA and genomes. Indeed, Pfam annotation from the transcriptome provided an N-terminal domain for the GH32 used in this model, instead of a catalytic domain. One can predict that the catalytic domain from Pfam would provide a better performance. Recent work brought some insight onto the PUL related to inulin or levan metabolism. A closer look at the inulin/levan- or fructan-associated operon recently described by Joglekar et al. (2018) confirmed that the specificity for the 2-6 linkage found in levan is from a GH32 cell surface endo-levanase and an ortholog of BT1761, a surface glycan-binding protein. The presence of the cell surface levanase of B. thetaiotaomicron VPI-5482 was critical for the ability of this strain to use the levan. The authors very elegantly demonstrated how structural differences present in dietary polysaccharides such as fructans can result in distinct molecular mechanisms for utilization of these polymers.

Despite obtaining low performance for some fibers, the overall goal of this analysis is a metabolic functional assessment of the different strains. We also tested whether an approach involving the use of GH genes only would be sufficient to obtain a growth prediction. Several teams applied the following method: gathering enzymes into functional groups, for example GH23, GH25, and GH73 being dedicated to peptidoglycan breakdown and GH13 is dedicated for starch breakdown. A first drawback in implementing this method into a pipeline is that the attribution of a function or the link between GH and a substrate or fiber varies substantially across the literature. For instance, we gathered such associations from four different publications (El Kaoutari et al., 2013; Park et al., 2018; Zhao et al., 2018; Kovatcheva-Datchary et al., 2019) and found that discrepancies can be observed. For example, GH95 being either associated with mucin degradation or with cellulose degradation depending on the publication. Supporting our PUL approach, using GH only led to growth predictions that were not meaningful.

Several parameters play a part in the model prediction performance. Both annotation quality and discrepancy between datasets and gene and enzyme terminology differences have been a hurdle in designing new PUL models. As shown by the comparison of GH annotation tools, an accurate annotation of CAZymes is key to improve prediction performance. Most automated annotation pipelines for transcriptomic data do not accurately annotate for GH. Then, the Pfam domain used for functional domain characterization may easily provide the right annotation of only one domain of the GH, as we observed for several PUL where we did not capture the GH catalytic domain. Another aspect is that our PUL models do not have a size limit, as long as the distance between neighboring genes annotated as the same PUL family is less than 5 kb. This is consistent with large syntenies observed in PULDB. For example, for our PUL amylase model, B. uniformis PUL is slightly over 6 kb while B. salyersae PUL reached 10 kb.

False prediction may originate from several reasons including in vitro conditions or how the model processes growth data. The PUL models were obviously sensitive to the growth/OD threshold. For arabinoxylan, the three strains predicted “no growth” by the model and grew were B. massiliensis, B. oleiciplinus, and B. merdae. Interestingly, in the four cases where the model predicted “no growth” and growth was counted positive based on the model threshold, the OD_600–nm recorded were below 0.1 (B. intestinalis, 0.05; B. johnsonii and B. clarius, 0.06; and B. finegoldii, 0.14). A similar phenomenon happened for heparin, and a reset growth threshold improved performance. The growth threshold could be adjusted accordingly for the different fibers, when more experimental data are available. However, the experimental growth medium itself influences growth yield in a strain-to-strain manner that is difficult to predict. The results would need to be reassessed or taken with caution if strains were grown in a culture medium that is different from the reference. The experimental data we used to measure the pipeline performance were obtained in PY medium. However, the transcriptomics data used to build the PULs models were obtained on CM medium, which could account for differential genes being over-expressed, at the time of sampling, compared to growth validation data on PY medium. Indeed, the growth medium may change which genes are prioritized in the model. It remains to be determined whether this bias would be more impactful on complex PUL models with large numbers of genes or complex operon structure where genes, not co-transcribed, might not be captured in a single transcriptome time point. The extreme modularity of some PULs indicates that the conclusions about growth results from one species to another have to be taken with caution since, like others, we detected strain-to-strain variations within the same species for inulin and levan.

The complexity of fiber structures makes the links between CAZyme genes and functional interpretation very uncertain. In order to determine the strain capacity to break down a specific fiber, we linked CAZy genes to a given fiber or prebiotic. However, this representation has limitations because (i) the complexity and substrate diversity hardly fit the CAZyme databases and several families are displaying a wide phylogenetic diversity, such as amylase GH13 (as described by Stam et al., 2007) and (ii) certain families can be dedicated to/involved in different substrates (GH32 targeting inulin and levan as showed by Sheridan et al., 2015). Different authors reported different gene and substrate associations. As the gene content is then associated with functional capacities, it can have far-reaching consequences on the conclusions and on further applications or recommendations.

Which experimentations could help refine the pipeline? A major challenge toward quantifying the degree of redundancy of CAZymes will be to obtain more in vitro information on growth and CAZyme expression patterns. Up to now, there are still very few measurements of purified enzymes to pinpoint the specific substrate specificity of the CAZymes involved in the hydrolysis and fermentation of fiber or of the human gut mucus. We also modeled Firmicutes PULs (Supplementary Figure S5), but despite encouraging results, the scarcity of transcriptomics data available led us to focus on Bacteroidetes.

One way to circumvent these drawbacks or limits is to obtain massive large datasets performed in the same conditions, with strains for which the genomes are available. To improve FiberGrowth pipeline performance, a larger number of strains should be grown in the exact same conditions. This could be done using robots under anaerobic chamber, similar to the work by Zou et al. (2019), leading to > 6,000 isolated strains and 1,520 new genomes sequenced. Note that if more experimental data become available, it would also be possible to adopt a machine-learning approach to infer potential PUL rather than having a deterministic approach such as FiberGrowth.

Another way to increase our capacity to obtain large growth dataset arises from the recent advances in droplet microfluidic culture, which can increase the capacity and lower the cost of purified substrates, to screen for growth on fiber. Villa et al. (2020) recently demonstrated the potential of microfluidic droplet assays for comparing the growth rates and functions of individual bacterial strains isolated from gut microbial communities. This would be very useful in order to improve FiberGrowth pipeline on a large number of strains grown in the exact same conditions.

Interestingly, the authors also investigated how screening for GH with microfluidics could lead toward the differentiation of subjects, based on the “fiber profile” being metabolized by their microbiota. A limit of the characterization of a fiber-metabolizing potential per microbiota remains the complexity and modularity of the GH and PUL operons. Predicting the overall microbiome response to a specific fiber requires to account for the variability of GH protein structures. However, an accurate annotation of PUL seems difficult in metagenomics datasets. We already showed, using 40-kb E. coli fosmid libraries built from fecal samples or distal ileum mucosa, that glycosyl hydrolases were modular and subject to recent horizontal gene transfer, not just between phylogenetically close genomes but above the genus level (Tasse et al., 2010; Tauzin et al., 2020). A challenge remains to accurately annotate glycosyl hydrolase loci in metagenomic data, only feasible with long assembly with enough coverage.

Predicting individual response to prebiotic or fiber intake, we previously had demonstrated that the diversity of fiber-rich food items correlated with microbiome 16S rDNA diversity in young adults (Tap et al., 2015). Others have demonstrated that a 5-day regimen with a plant-based diet compared to animal product-based diet (David et al., 2013) increased microbial diversity. Mining existing large datasets such as iHMP2 (Proctor et al., 2019) can be of importance to find correlations between the breakdown of polymers from human origin, which have been linked to inflammatory bowel diseases. Those include not only human mucus and related compounds, such as heparin, but also human gangliosides. As a growing number of metagenomics datasets are being generated worldwide, correlations can be inferred between dietary intake and microbiome composition. However, most food questionnaires are not designed to provide indication on specific fiber. Interestingly, new AI-assisted tools for computing food images on smartphones may provide a more accurate picture of the ingested fiber content. There is still a lack of knowledge between a food element/items and the set of glycosyl hydrolases necessary for its breakdown and the subsequent production of beneficial metabolites through fermentation.

The diversity of the CAZy gene families involved in the breakdown of glycans and the extreme modularity of the operons that encode them, sometimes at the strain level, offer a challenge to building PUL models that are generic for a given fiber. However, we demonstrated, for the first time, that a pipeline, combining automated genome screening and annotation of PULs, allowed us to build growth prediction models and measure their accuracies on Bacteroidetes available growth data. Some PUL models require optimization, while for levan, heparin, inulin, and arabinoxylan, FiberGrowth pipeline can compute growth estimations with TPR > 0.8 and FPR < 0.33 in 1 min on an unannotated genome.

This work also demonstrated that, despite advances in PUL bioinformatics screening and computing modeling, the lack of biochemical characterization of glycosyl hydrolases and PUL systems remains a main issue.

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author/s.

BC and CS designed and built the pipeline, performed all computing, and participated in the manuscript writing. FP supervised computing and assisted in the manuscript editing. ML was involved in the pipeline design and supervised the manuscript writing. All authors contributed to the article and approved the submitted version.

FP and CS work at Pendulum Therapeutics Inc. ML was member of the scientific advisory board for Pendulum Therapeutics. The remaining author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.