Gut microbiome-based prediction of autoimmune neuroinflammation

Gut commensals are linked to neurodegenerative diseases, yet little is known about causal and functional roles of microbial risk factors in the gut–brain axis. Here, we employed a pre-clinical model of multiple sclerosis in mice harboring distinct complex microbiotas and six defined strain combinations of a functionally-characterized synthetic human microbiota. Discrete microbiota compositions resulted in different probabilities for development of severe autoimmune neuroinflammation. Nevertheless, assessing presence or the relative abundances of a suspected microbial risk factor failed to predict disease courses across different microbiota compositions. Importantly, we found considerable inter-individual disease course variations between mice harboring the same microbiota. Evaluation of multiple microbiome-associated functional characteristics and host immune responses demonstrated that the immunoglobulin A-coating index of Bacteroides ovatus before disease onset is a robust individual predictor for disease development. Our study highlights that the “microbial risk factor” concept needs to be seen in the context of a given microbial community network, and host-specific responses to that community must be considered when aiming for predicting disease risk based on microbiota characteristics.


27
Gut commensals are linked to neurodegenerative diseases, yet little is known about causal and functional roles 28 of microbial risk factors in the gut-brain axis. Here, we employed a pre-clinical model of multiple sclerosis in 29 mice harboring distinct complex microbiotas and six defined strain combinations of a functionally-30 characterized synthetic human microbiota. Discrete microbiota compositions resulted in different probabilities 31 for development of severe autoimmune neuroinflammation. Nevertheless, assessing presence or the relative 32 abundances of a suspected microbial risk factor failed to predict disease courses across different microbiota 33 compositions. Importantly, we found considerable inter-individual disease course variations between mice 34 harboring the same microbiota. Evaluation of multiple microbiome-associated functional characteristics and 35 host immune responses demonstrated that the immunoglobulin A-coating index of Bacteroides ovatus before 36 disease onset is a robust individual predictor for disease development. Our study highlights that the "microbial 37 risk factor" concept needs to be seen in the context of a given microbial community network, and host-specific 38 responses to that community must be considered when aiming for predicting disease risk based on microbiota 39 characteristics.

INTRODUCTION 41
Compared to healthy controls, autoimmune disease patients exhibit distinct microbiota compositions (1), 42 especially in the context of multiple sclerosis (MS) (2). Thus, determining whether susceptibility or 43 progression of MS can be predicted by the microbiota composition is a necessary precondition to develop 44 patient-targeted microbiota modulations (Fig. 1a). A common approach to elucidate microbiota-related, MS-45 promoting predictors compares relative abundances of bacterial taxa -often determined by 16S rRNA gene-46 based sequencing -between MS-affected and healthy individuals (2-9). Although certain differentially 47 abundant taxa identified across different human cohort studies tend to be concordant, e.g., increased 48 abundances of Akkermansia (2,4,5,(7)(8)(9) or decreased abundances of Prevotella (3,(5)(6)(7)(8) in MS patients 49 compared to healthy controls, such cohort-level observations do not explain inter-individual differences (10) 50 in disease course or susceptibility. Therefore, it remains challenging to reliably link taxa abundances across 51 individuals to microbiota characteristics that impact MS disease course.

52
Given the limitations of correlation-focused human cohort studies to uncover reliable microbial risk factors 53 for MS susceptibility, the experimental autoimmune encephalomyelitis (EAE) mouse model is commonly used 54 to verify presumed causality between presence of suspected microbial risk factors and development of 55 autoimmune neuroinflammation (11)(12)(13). However, it is unclear whether the causality of a singular species 56 alone or within only one specific background microbiota, i.e. in mice harboring a relatively consistent, specific 57 pathogen-free (SPF) microbiota composition, can be translateable to the plethora of individual microbiota 58 compositions found across a given population (1). Although certain inter-microbial interactions, that promote 59 EAE development, have previously been revealed (11,14), the mutual impact between the background 60 microbiota and potential commensal risk factors on disease-promoting properties of the microbiota is poorly 61 understood.

62
Here, we investigated whether the EAE disease course can be predicted before disease onset by microbiota-63 associated readouts. For reliable EAE disease prediction, we weighed in microbial taxonomic composition 64 analyses against microbiota-associated, functional analyses, and we investigated the host immune responses.

65
Our comprehensive approach allowed us to examine how individual host−microbe interactions interfere with 66 disease predictability.

RESULTS 68
Muc2-deficiency in mice is associated with less severe experimental autoimmune encephalomyelitis 69 During experimental autoimmune encephalomyelitis (EAE), the microbiota composition impacts how the host 70 immune system is shaped (15), impacting the degree of neuroinfammation. Investigating how different disease 71 phenotypes could be predicted based on the microbiota composition could be a key toward microbiota-based 72 intervention (Fig. 1a). As a first step toward testing our hypothesis, we set out to identify the EAE 73 development-associated commensal genera within a complex microbiota. To do so, we induced EAE in 74 specific pathogen free (SPF) mice of different origins and genotypes. We fed these mice diets with different 75 fiber contents, given the impact of dietary fiber quality and quantity on relative abundances of indigenous 76 commensals (16,17). First, we addressed the question of whether changing relative abundances of taxa within 77 a given microbiota might affect outcome of EAE. Toward this goal, we used wildtype C57BL/6J mice 78 purchased from Charles River (CR), which were fed either a standard laboratory chow (fiber-rich; FR) or a 79 fiber-free diet (FF) diet for 20 days, followed by induction of EAE (Fig. 1b). Feeding these diets did not result 80 in different disease outcomes (Fig. 1c-e), indicating that dietary fiber quantity and quality, and the associated 81 effects on relative abundances of taxa within this particular indigenous microbiota, are not determining factors 82 in mediating EAE (Fig. 1f).

83
Next, we sought to elucidate whether we could observe distinct EAE outcomes between mice whose native 84 microbiota differentiated considerably by the presence of certain taxa, rather than by relative abundances of a 85 shared core microbiota. To do this, we induced EAE in mice deficient for the Muc2 protein (Muc2 −/− ), as this 86 genetic modification results in an impaired mucus barrier (18). We expected a significantly different 87 indigenous microbiota composition due to anticipated reduction in commensals relying on an intact mucus 88 layer as a functional or nutritional niche (19). As controls, we used littermate mice homozygous for the 89 presence of Muc2 gene (Muc2 +/+ ). While Muc2 +/+ mice were fed both FR or FF diet, Muc2 −/− mice were only 90 fed a FR diet (Fig. 1g). We observed a significant difference in disease progression between the genotypes, 91 with Muc2 −/− mice being significantly less susceptible to EAE induction compared to Muc2 +/+ mice, regardless 92 of diet ( Fig. 1h-j). As observed for CR mice (Fig. 1f), diet-mediated influences on disease development were 93 negligeble (Fig. 1k).
Higher abundances of Akkermansia muciniphila are associated with less severe EAE in mice with a 95 complex microbiota 96 To evaluate a potential contribution of the microbiota to the observed differences, we performed 16S rRNA 97 gene-based sequencing analyses on DNA isolated from fecal samples taken before EAE induction (Fig. 1l-m; 98 Extended Data Fig. 1a-f) and during the EAE course (Extended Data Fig. 1a-e). Intriguingly, the overall 99 microbiota β-diversity (Extended Data Fig. 1a−d) and α-diversity (Extended Data Fig. 1e) were 100 disconnected from the EAE disease course. Since all four groups of mice expressing the Muc2 protein (CR 101 mice and Muc2 +/+ mice) provided a comparable EAE disease course (Fig. 1c, h), which was significantly 102 different from the one observed in Muc2 knockout (KO) mice (Fig. 1h), we assessed potential EAE-relevant 103 microbiota differences by comparing Muc2 −/− mice (KO) with all Muc2-expressing mice combined (WT), 104 irrespective of origin or diet. We identified 11 differentially abundant genera that explained more than 70% of 105 the variance detected in the Bray−Curtis distance matrix between WT and KO mice before induction of EAE 106 (pre-EAE). Pre-EAE relative abundance of the genus Akkermansia alone explained 14.4% of said variance 107 ( Fig. 1l, right panel), correlated negatively with various EAE readouts upon induction of disease ( Fig. 1l left   108 panel), and was significantly higher in Muc2 −/− mice compared to WT counterparts (Fig. 1m), suggesting 109 possible disease-preventing properties of Akkermansia.

110
However, our observations so far are inadequate to attribute distinct EAE phenoytpes exclusively to changes 111 in Akkermansia abundance, or micobiota changes in general, because potential Muc2 knockout-associated 112 changes in host responses were not specifically addressed. Nonetheless, given that Akkermansia is consistently 113 reported as a potential risk factor for MS (2,4,5,(7)(8)(9) and due to its controversial role in EAE development 114 (2,20), we asked whether the observed potential disease-preventing properties in our experiments might be 115 rooted in distinct background microbiota compositions, as Akkermansia was embedded in diverse microbiotas 116 with different taxa present or absent (Extended Data Fig. 1f).

118
To better understand a potential causal role of Akkermansia in EAE development and to evaluate its potential 119 as a disease-risk predictor, we colonized germ-free (GF) C57BL/6 mice with a functionally-characterized 14-120 member human synthetic microbiota (SM14) (16, 21) (Fig. 2a). This approach allowed us to drop out specific 121 species-of-interest from this community to investigate the contribution of a single microbe on EAE 122 development in a genetically homogenous host. Akkermansia muciniphila, the type species for the 123 Akkermansia genus, is a member of this SM14 community. Thus, we colonized GF C57BL/6 mice with either 124 the complete SM14 community or a SM13 community, lacking A. muciniphila, followed by induction of EAE 125 (Fig. 2b). SM13-colonized mice exhibited a significantly less severe EAE phenotype compared to SM14-126 colonized counterparts (Fig. 2c, left panel; Fig. 2d−f), highlighting the general contribution of the microbiota 127 to EAE development and the disease-driving role of A. muciniphila in the SM14 microbiota-based mouse 128 model when this species is combined with the 13 strains listed in Fig. 2a. As controls, we induced EAE in A.

131
To evaluate whether changes in relative abundances of SM14-constituent strains might affect EAE disease 132 course, we fed SM14-and SM13-colonized mice an FF diet, followed by EAE induction (Fig. 2b). GF mice 133 were also fed a FF diet to exclude microbiota-independent but diet-mediated effects on EAE. Feeding SM14-134 colonized mice the FF diet resulted in significantly increased A. muciniphila relative abundances compared to 135 equally colonized FR-fed mice (Fig. 2g). However, we did not detect any significant differences in any EAE-136 associated readout (Fig. 2c, right panel; Fig. 2f) between FR-and FF-fed mice harboring the same microbiota.

137
Removal of A. muciniphila from the SM14 community explained between 22% and 28% of the variance for 138 different EAE-associated readouts (Fig. 2e,

145
To evaluate how A. muciniphila might alter microbiota function (Fig. 1a) within the SM14 reference 146 microbiota, we performed metabolomic and metatranscriptomic analyses. EAE is associated with changes in 147 either plasma metabolite profiles (22,23) or changes in metabolic pathways of the intestinal microbiota (24). Furthermore, A. muciniphila mediates other pathologies, at least in part, via secretion of certain metabolites 149 (25). Thus, we asked whether different levels of neuroinflammation between SM14-and SM13-colonized mice 150 could be explained by A. muciniphila-associated metabolite patterns in the cecum or serum. In addition to 151 collecting cecal and serum samples from EAE-induced GF, SM01-, SM13-and SM14-colonized mice, we 152 also collected the same samples from the same groups of non-EAE induced mice. The cecal metabolite profiles 153 were similar between EAE-induced and non-EAE induced groups harboring the same microbiota, as well as 154 between EAE-induced SM13-colonized and SM14-colonized mice (Fig. 3a, b). As broader metabolic profiles 155 were disconnected from the EAE disease course (Extended Data Fig. 3a), we reasoned that only a few cecal 156 metabolites, if any, might causally influence the EAE disease course.

157
To identify such potential EAE-impacting metabolites, we developed a metabolite-of-interest screening 158 pipeline including 20 independent analyses (Extended Data Fig. 3b−e). We proposed that a potential A. 159 muciniphila-associated and disease-mediating metabolite should fulfill five different criteria. The rationale for 160 these criteria and the analytical approach is specified in the Materials and Methods section. Among the 18 161 metabolites that significantly correlated with at least one EAE-associated readout, only γ-amino butyric acid 162 (GABA) emerged as a metabolite-of-interest in cecal samples (Fig. 3c). Of note, its concentration was 163 significantly elevated in non-EAE-induced mice harboring an SM combination which resulted in severe EAE 164 upon disease induction. (Fig. 3d). Given that GABA concentrations were higher in disease-prone, A. 165 muciniphila-harboring mice, these results suggested that the cecal concentration of GABA before EAE 166 induction already defined disease development upon disease induction and that its concentration was linked to 167 disease-influencing properties of the tested microbial communities harboring A. muciniphila. Additionally, we 168 did not identify any metabolite-of-interest in serum samples (analyses not shown), indicating that potential 169 metabolite-driven impacts on EAE disease course occur locally in the intestine.

172
Given these community-dependent alterations in microbial metabolite profiles, we rationalized that the 173 presence or absence of A. muciniphila had a significant impact on gene expression profiles of the overall 174 microbiota, possibly contributing to distinct EAE phenotypes. Thus, we performed metatranscriptomic analysis of cecal contents obtained from non EAE-induced SM14-and SM13-colonized mice ( Fig. 3e−g).

176
When comparing transcript profiles of both groups, we found 117 genes expressed only in SM14-colonized 177 mice (Fig. 3f). Although we expected that these transcripts would be mostly from A. muciniphila, in fact, most 178 of these genes were exclusively expressed by either Roseburia intestinalis or Marvinbryantia formatexigens 179 (Fig. 3g). Of the 30 genes expressed only in SM13-colonized mice, the majority were expressed by 180 Eubacterium rectale (Fig. 3g). These findings highlight the crucial impact of the presence of a single 181 commensal on the gene expression pattern of other microbial community members, likely impacting their 182 "function" (Fig. 1a) within a given community. These indirect influences on community function might also 183 contribute to microbiota-mediated effects on EAE development and thus impact disease-mediating properties 184 of potential risk-predicting species.

185
Mucin-degrading capacity of the microbiota is not linked to EAE severity 186 So far, our results suggest that predicting EAE development, based on the presence or absence of A. 187 muciniphila, a mucin-specialist bacterium (16), was only possible in mice harboring a variation of the SM14 188 community (Figs. 2, 3) and not in mice harboring a complex community (Fig. 1). To further address potential 189 reasons for these discrepancies in general, and the apparent crucial impact of the presence of A. muciniphila 190 in SM14-colonized mice in particular, we next hypothesized that changes in relative abundances of other 191 strains in response to dropping out A. muciniphila from the SM14 community might causally impact EAE 192 development. We observed that four strains were significantly higher in abundance in SM13-colonized mice 193 compared to SM14-colonized mice (Fig. 4a). To address their potential contribution to EAE development, we 194 colonized mice with three additional SM combinations (Fig. 4b, Extended Data Fig. 4a). In the first of these 195 combinations, we colonized GF mice with an SM12 community (Fig. 4b), lacking A. muciniphila and 196 Faecalibacterium prausnitzii. This experiment was performed to elucidate whether the >1000-fold increase in 197 relative abundance of F. prausnitzii (Fig. 4a, extreme right panel) -a species known for gut health-promoting 198 properties (26) and decreased abundances in MS patients (10) -in mice lacking A. muciniphila (Fig. 4a) was 199 responsible for EAE-preventing properties of the SM13 community. Intriguingly, SM12-colonized mice ( Fig.   200 4c−e) provided a comparable disease course as SM13-colonized mice (Fig. 2), suggesting that F. prausnitzii 201 expansion in SM13-colonized mice is not responsible for decreased EAE in SM13-colonized mice. At the 202 same time, these data point out the A. muciniphila-mediated inhibitory effects on the expansion of an anti-203 inflammatory bacterium, F. prausnitzii.

204
Removal of A. muciniphila from the SM14 community further resulted in expansion of three mucin glycan-205 degrading (16) Bacteroidetes species (Fig. 4a) in the SM13 community. Thus, we investigated whether 206 colonization with the three mucin glycan-degrading strains alone (SM03) resulted in decreased EAE compared 207 to SM14-colonized mice and whether addition of A. muciniphila (SM04) might counteract a potential 208 beneficial effect. While SM03-and SM04-colonized mice showed comparable EAE disease courses ( Fig.   209 4c−e) to SM14-colonized mice (Fig. 2), they differed significantly from SM13-colonized mice. Additionally, 210 the three mucin glycan-degrading Bacteroidetes strains appeared to not provide disease-reducing properties 211 but, on the contrary, disease-promoting properties in the absence of the remaining 10 strains within the SM13 212 community. To evaluate whether dysregulated mucin turnover might contribute to the observed results in these 213 mice, we assessed various indirect measures for intestinal barrier integrity. We did not detect any correlations 214 between EAE outcome and glycan-degrading enzymatic activities (Extended Data Fig. 4b−

218
were not an individual predictor for EAE disease development.

219
Microbiota composition can be used to estimate the probability of severe EAE incidence 220 Thus far, groupwise comparisons of EAE-associated readouts and microbiota compositions failed to identify 221 reliable predictors for disease development in EAE-induced mice. Therefore, we next aimed to elucidate 222 common denominators on a group-based and individual level to help uncover more reliable potential predictors 223 for microbiota-mediated impacts on disease course. First, we conducted group-based comparison of EAE 224 outcomes between all 10 tested diet−colonization combinations ("groups") ( Fig. 5a, b). Performing 225 hierarchical clustering (Fig. 5c) based on group means of key EAE-associated readouts (Fig. 5b) revealed 226 three distinct group phenotypes: "moderate", "intermediate" and "severe". While diet explained less than 8% 227 of the variance observed for EAE-associated readouts, microbiota composition (SM) explained between 11% 228 and 27% (Fig. 5d). Given these low values, rooted in considerable intra-group variances (Fig. 5b), we 229 performed individual EAE phenotype clustering, treating all mice across all groups individually (Fig. 5e). T-230 distributed stochastic neighbor embedding (t-SNE) analysis of all EAE-induced individuals resulted in two 231 disease clusters: "Cluster 1", comprising mice showing strong EAE symptoms, and "Cluster 2", comprising 232 mice showing minor EAE symptoms (Fig. 5e). Besides SM03-and SM04-colonized mice, every group 233 included mice of both phenotypes (Fig. 5f), however with varying proportions. These proportions broadly, but 234 not completely, corresponded to the group-based phenotype classification (Fig. 5c). In summary, these results 235 ( Fig. 5a−f) indicate that knowing the composition of the microbiota, in combination with information on 236 dietary conditions, enables estimation of the probability for either moderate or severe disease, but is unsuitable 237 to predict individual EAE outcomes.

238
Given that IL-17-and IFNγ-producing CD4 + cells (11), CD8 + cells (27), and IgA + IL-10 + plasma cells (28) Fig. 5e) showed significant differences for seven populations, with IFNγ-expressing Th17 252 cells in the spinal cords significantly increased in Cluster 1-mice (Fig. 5h). Mouse-specific T cell polarization 253 profiles aligned better with disease outcome (Fig. 5h) than with SM−diet combinations (Fig. 5h). Thus, we 254 concluded that host-specific differences must occur before T cell activation in EAE-induced mice, most probably due to individual microbiota-mediated signals that appeared to be distinct even in mice harboring the 256 same set of strains.

257
When analyzing T cell subsets in non-EAE induced mice, we found that the microbiota composition primed 258 CD4 + T cells towards a pro-inflammatory Th17 response before EAE-induction (Extended Data Fig. 5b).

259
Although we found more significant correlations of these populations with EAE-associated readouts in the 260 ileum, overall T cell population distribution in the colon aligned best with emerging EAE group phenotypes 261 upon EAE induction (Extended Data Fig. 5c,d), suggesting a crucial contribution of T cell priming in the 262 colon by the microbiota to disease development upon EAE induction.

263
In summary, it was impossible to predict individual disease development based on microbiota−diet 264 combinations alone despite apparent microbiota-mediated priming of the local adaptive immune system before 265 EAE induction. This observation suggests that microbiota-mediated signals that influence the adaptive immune 266 system to either promote or decelerate EAE development are relatively constant before disease induction, but 267 are prone to individual changes upon disease onset.  Fig. 6b, c). To do so, we only included mice harboring at least 12 different strains, thus excluding mice with 275 low-diversity microbiotas (SM04, SM03, SM01). Correlations for each strain were only assessed for those 276 mice, which were gavaged with the respective strain and calculations were performed by either including mice 277 harboring SM14, SM13 and SM12 communities, or specific combinations thereof, into the analysis. We found 278 statistically significant correlations between pre-EAE bacterial relative abundances with EAE-associated 279 readouts for some strains (Fig. 6b). However, the few statistically significant correlations we determined were 280 generally weak (R<0.4) and Pearson correlation values for a given strain were highly dependent on the 281 background microbiota (Fig. 6b). In line with this, relative abundances of strains only explained very low 282 proportions of the variances across all groups for all assessed EAE-associated readouts (Fig. 6c).

283
Next, we asked whether presence or absence of a given strain might be a better predictor for individual EAE 284 development. Thus, we performed a linear mixed model regression for three EAE-associated readouts with 285 presence of the strain as an independent variable and colonization as a random intercept effect (Fig. 6d,

286
Extended Data Fig. 6b). Given the setup of our tested SM-combinations, we could only assess A. muciniphila 287 and F. prausnitzii separately and had to analyze the remaining 12 strains in groups of two combinations.

288
Presence or absence of a specific strain or strain combination was insufficient to predict the individual outcome 289 of any of the tested EAE-associated readouts (

295
but were strongly connected to the microbiota composition (Extended Data Fig. 6d). Interestingly, we found 296 a significant correlation between group means of sIgA concentrations and corresponding EAE susceptibility 297 incidence (Extended Data Fig. 6e). Owing to these observations and given that the "IgA-coating index" (ICI) 298 was previously suggested to be a measure of autoimmunity-promoting potential of a given commensal species 299 (32), we determined ICIs for each strain within each high-diversity SM combination (SM12, SM13 and SM14)

304
Additionally, ICIs of these strains varied not only between distinct groups, but also between individuals within 305 groups. Thus, we reasoned that the individual ICI of these strains might reflect individual EAE-promoting 306 properties of the microbiota in a certain host. Correlation analysis of strain-specific ICI, as determined from 307 fecal samples obtained before EAE induction, with EAE outcome in the same individual, revealed significant 308 correlations with some EAE-associated readouts for four strains (Fig. 6f). However, the only strain whose 309 individual ICI provided significant correlations with the two most important EAE-associated readouts (AUC 310 and maximum achieved EAE score) was B. ovatus ( Fig. 6f, g), thus allowing for individual prediction of EAE 311 disease course across all B. ovatus-encompassing SM combinations (Fig. 6g).

DISCUSSION 313
Given the association between the intestinal microbiota and extra-intestinal autoimmune diseases (1), 314 microbiota manipulation is a feasible approach to boost existing therapy options for MS patients. Potential 315 strategies for microbiota modulation include administration of antibiotics or probiotics (33), dietary 316 interventions (34, 35) or fecal microbiota transplantation (36). However, such untargeted strategies could lead 317 to broad-scale changes in the microbiota with potentially unpredictable outcomes. On the contrary, given the 318 unique microbiota composition in a given individual, personalized approaches might be more promising.

319
A better understanding of what exactly makes a specific microbiota composition in a given individual disease 320 prone is a precondition for a targeted personalized approach. Such knowledge is expected to result in analytical 321 procedures to evaluate the average risk of disease or even predict individual outcomes. Previously suggested 322 measures to evaluate the MS-mediating risk of a microbial community, such as the microbiota α-diversity (37) 323 or the Firmicutes-to-Bacteroidetes ratio (38), emerged as unsuitable tools (10) and more focus is currently 324 being put on individual taxa or combinations of taxa (10). A recent mouse study suggested that up to 50 325 different microbial taxa could be associated with disease (39). Disease-associated bacterial taxa are often 326 referred to as "microbial risk factors", which are mostly identified by differences in presence or relative 327 abundances from cross-sectional human cohort studies. Akkermansia muciniphila represents such a potential 328 microbial risk factor as it was reportedly increased in MS patients across various human cohorts (2,5,7,40).

329
Other studies, however, report on positive effects of A. muciniphila on maintaining general gut homeostasis 330 (41,42) or on progression of autoimmune neuroinflammation in mice (20).

331
At first sight, these observations might appear contradictory. They are, however, corroborated by our findings.

332
By comparing development of experimental autoimmune encephalomyelitis (EAE) in mice of different genetic 333 backgrounds and with distinct complex microbiotas, we found the genus Akkermansia to be the most 334 negatively associated with EAE disease development, thus representing a potential hallmark genus for less 335 severe EAE, when considering the microbiota composition as the only variable and ignoring host genetics 336 (Extended Data Fig. 7). Next, we evaluated whether this finding could be reproduced in gnotobiotic, 337 genetically homogenous mice harboring different combinations of a reduced reference microbiota, with or 338 without A. muciniphila. Interestingly, we found A. muciniphila to be positively associated with EAE severity 339 in certain mice harboring specific reduced communities. This was associated with increased cecal levels of γ-340 amino butyric acid (GABA). However, we cannot reliably conclude whether increased GABA concentrations 341 are a general risk factor for EAE induction or whether this only applies to a certain microbiota composition.

342
Since elevated GABA levels were previously reported to be associated with less neuroinflammation (43, 44), 343 we deemed assessment of intestinal GABA concentration, without corresponding information on microbiota 344 composition, unsuitable for disease course prediction. It is unclear whether this neurotransmitter directly 345 mediates EAE-influencing host responses via interaction with local host receptors (45) or whether it might act 346 as a signaling molecule or energy source (46) for other species, which finally mediate disease promotion.

347
Our study suggests that focusing on microbiota composition only, as we did in our experiments using mice of 348 different genotypes, can result in misleading conclusions. Co-variates, such as diet, sex, medication use, 349 geographic location, disease subtypes and genetic heterogeneity of study participants make interpretation of 350 microbiota data from human cohort studies complicated, although certain biostatistic approaches help to 351 reduce the risk of misinterpretation, as elegantly shown in a recent publication from the iMSMS Consortium 352 (10). In addition, we found that 16S rRNA gene sequencing-based determination of relative taxa abundances 353 is unsuitable to make meaningful assumptions on disease-mediating properties of a given microbiome.

354
Assessing the presence or absence of taxa allowed us to determine the probability of severe disease ( Fig. 6h:

355
"Risk of disease"). However, this observation was unrelated to presence or absence of a single taxon, 356 suggesting that focusing on combinations of taxa and/or environmental factors, rather than single taxa alone, 357 may be required to form reliable conclusions. Thus, our results suggest that mutual influences between a 358 suspected risk factor and the microbial environment crucially shape the overall microbiota's disease-impacting 359 potential (Extended Data Figure 7).

360
Our metatranscriptomics analyses revealed that even minor changes in microbiota composition, i.e. by 361 removing A. muciniphila from a reduced community, resulted in profound changes in gene expression patterns 362 of some, but not all, intestinal microbes. Such influences may also affect disease-mediating properties of the 363 microbiota. Therefore, we suggest to put more focus on microbial network analysis to disentangle specific 364 inter-microbial interactions.
Although not yet a technically and analytically refined approach (47), 365 metagenomic-based microbiota network analyses are currently being explored as an analysis tool (48) and 366 might be superior to statistical analysis of species−species co-abundances (10). A key study, evaluating the 367 effects of multiple defined microbiota compositions on fitness of Drosophila melanogaster, already pointed 368 out that microbial network interactions are more important than relative abundances of a given species alone 369 (49). Our study documents similar innovative findings, based on comprehensive datasets, in a controlled 370 vertebrate gnotobiotic disease model..

371
In addition to these microbiota-specific effects, host-specific effects appeared to be a decisive factor for 372 individual EAE development in our experiments, further complicating the quest for reliable disease predictors.

373
Even in genetically homogenous mice of the same sex and age, harboring the exact same set of commensal 374 bacteria and living under the same standardized conditions, we found considerable individual differences in 375 EAE disease course, suggesting that the individual disease development is mediated by either 376 microbe−microbe or microbe−host interactions (50, 51) (Fig. 6h). After extensive evaluation of multiple 377 microbiota-associated readouts, we found the IgA-coating index (ICI) of the fiber-degrader Bacteroides ovatus 378 (16) to be capable of sensing the individual EAE-influencing properties of the microbiota, irrespective of its 379 definite composition. Determining the ICI of B. ovatus before disease induction correctly predicted EAE 380 outcome in every individual (Fig 6h: "Prediction of disease"). Thus, we propose that B. ovatus acts as a 381 "reporter species", reflecting the individual microbiota-and host-mediated dual influences on EAE progression 382 while taking into account distinct microbiota functions across different hosts ( Fig. 6h: "host-mediated effects 383 on Function"). In the current study, although such a property of B. ovatus was only evaluated in reduced 384 microbial communities, the concept of "reporter species" might also apply to other strains and more complex 385 communities, including MS patient microbiotas. A recent study showed that different strains of B. ovatus are 386 capable of driving variable host IgA secretion (52), which might also impact the IgA coating of B. ovatus.

387
Although we did not evaluate in our study whether different strains of B. ovatus evolve in different mice, future 388 studies need to consider such aspects.

389
In summary, we demonstrate that making disease-course predictions based on microbiota characteristics is 390 generally possible, but is not as black-and-white as it might appear. We therefore strongly argue for a 391 reconsideration of how microbiota-related data are analyzed and interpreted. In particular, we advocate for 392 higher analytical standards, with more sophisticated data integration to better account for discrepancies in host-393 specific microbiota function.   Given that EAE phenotypes were disconnected from the overall metabolome pattern, we looked for single 581 metabolites which might explain observed differences in EAE disease course. Thus, we implemented a 582 screening pipeline, comprising 20 independent analyses, to identify potential metabolites-of-interest that might 583 explain differences in EAE outcomes. These analyses included evaluation of the contribution of each 584 metabolite to the variance of the PC1 and PC2 axes in a mulidimensional reduction PCA plot (Extended Data 585 Fig. 3b), correlation analyses (Extended Data Fig. 3c), as well as group-wise comparisons of metabolite 586 concentrations (Extended Data Fig. 3d). By combining information obtained from these analyses, our goal 587 was to shortlist microbiota-induced cecal metabolites that enable prediction of either the overall disease course 588 or the relapse occurrence in EAE-induced mice. We concluded that a potential metabolite-of-interest should composition, we concluded that these shifts occured independent from the EAE disease phenotype given the 604 significant differences in EAE outcomes in mice harboring different microbiota compositions. In summary, 605 we concluded that these shifts on the PC2 axis were either a direct result of EAE induction (independent from 606 the disease phenotype) or a consequence of different microbiota colonization times since EAE-induced mice 607 harbored the respective microbiota for 3 more weeks compared to non-EAE-induced mice of the same 608 microbiota composition. Thus, we concluded that metabolites, which significantly contributed to the PC2 axis shift, were not causal to different EAE phenotypes. Instead, their concentrations in EAE-induced mice 610 appeared to be a feedback effect from either longer colonization or EAE induction itself. In consequence, these 611 metabolites were removed from the list of metabolites-of-interest. (3) a significant correlation with either the 612 overall disease course (Fig. 3c, "AUC"; Extended Data Fig. 3e) or the mean score during the relapse phase 613 (REL) in EAE-induced mice. (4) a significant correlation with the presence of A. muciniphila (Fig. 3c, " for 5 min using a mixer mill and the aqueous phase was recovered after centrifugation at 4°C for 3 min at 18 633 000 × g. 500 µL of phenol:chloroform (125:24:1) pH 4.3 was added to each sample and centrifuged as 634 previously described. Again, the aqueous phase was recovered and 1/10 volume of 3M sodium acetate (pH ~ 635 5.5) and 1 volume of ice-cold 100% ethanol was added and gently mixed by inversion. Samples were incubated 636 for 20 min on wet ice and then washed twice with 500 µL of ice-cold 70% ethanol and centrifuged at 4°C for 637 20 min at 18 000 × g. Pellets were air-dried for 10 min and resuspended in 50 µL nuclease-free water. DNase 638 treatment was performed by adding 10 µL 10X buffer, 40 µL nuclease-free water (to reach 100 µL final 639 volume) and 2 µL DNase I (Thermo Scientific, DNase I, RNase-free kit, #EN0521) followed by 30 min

Bacterial IgA coating index
Fecal samples stored at -20 °C were resuspended in 500 µL ice-cold sterile PBS per fecal pellet and 664 mechanically homogenized using a plastic incolulation loop. Pellets were then thoroughly shaken for 20 min 665 at 1100 rpm and 4 °C. After adding 2 × volume of ice-cold PBS, samples were centrifuged for 3 min at 100 × 666 g and 4°C to sediment undissolved debris. Clear supernatant was recovered and passed through a 70 µm sieve 667 (Imtec, #U3_10070_70) followed by centrifugation for 5 min at 10 000 × g and 4 °C to sediment bacteria.

668
Supernatant was removed and pellet resuspended in 1 mL ice-cold PBS. Next, optical density of this 669 suspension at 600 nm (OD600) was detected and approximate concentration of bacteria was estimated based on 670 the assumption that OD600 = 1 equals 5 × 10 8 bacteria per mL. The respective volume corresponding to 10 9 671 bacteria was centrifuged for 5 min at 10 000 × g and 4 °C. Pellet was then resuspended in 400 µL 5 % goat 672 serum (Gibco, #11540526) in PBS and incubated for 20 min on ice. After incubation, pellet was washed once 673 in ice-cold PBS and centrifuged for 5 min at 10 000 × g and 4 °C. Pellet was then resuspended in 100 µL ice-674 cold PBS containing 4 µg of FITC-coupled goat anti-mouse IgA antibody (SouthernBiotech, Imtec Diagnostic, 675 #1040-02). The ratio of 4 µg of the respective antibody to stain 10 9 bacteria was previously evaluated to be the 676 maximum amount of antibody that can be used without resulting in unspecific staining of non-IgA coated 677 bacteria by using fecal samples from Rag1 −/− mice as non IgA-coated negative controls. Samples were then 678 incubated for 30 min at 4 °C on a shaker at 800 rpm. After incubation, 1 mL ice-cold PBS was added followed 679 by centrifugation for 5 min at 10 000 × g and 4 °C. Samples were then washed once in ice-cold PBS, either

697
10% of the suspension volume was used for bacterial DNA staining using SYTO™ 60 Red Fluorescent Nucleic 698 Acid Stain as described above, whereas purity was generally >90% for both fractions; (2) To purify bacterial 699 DNA for subsequent 16S rRNA gene sequencing of bacteria within the different fractions, 90% of the 700 suspension volume was centrifuged for 10 min at 10 000 × g and 4 °C, supernatant was discarded and the dry 701 pellet was stored at -20 °C. DNA isolation and 16S rRNA gene sequencing was then performed as described 702 above. The IgA-coating index (ICI) for a given species x (ICIx) was calculated by the following equation, with 703 Ax + representing the strain-specific relative abundance in the IgA + fraction and Ax − representing the strain-704 specific relative abundance in the IgA − negative fraction: = . Strains were classified as "highly 705 coated", when > 2 × (ICIx > 0.301) and as "low coated", when ICIx < −0.301.

709
Plates were then washed four times in wash buffer (10 mM Trizma Base, 154 mM NaCl, 1% (v/v) Tween-10).          964 h) Relative abundances of T cell subsets, which were found to be statistically significant by EAE phenotype 965 cluster affiliation, as determined by unpaired t-tests. *, 0.01 < p < 0.05; **, p < 0.01.   h) Graphical summary: Taxonomic microbiota information can be used to assess the "Risk" of a given 993 individual to develop disease (as defined by a chance < 100 %). For "Predictions", as defined by a 100 % 994 chance to develop disease, host-mediated influences on the microbiota function within a given individual must 995 have to be taken into account.