Longitudinal Analysis of the Microbiome and Metabolome in the 5xfAD Mouse Model of Alzheimer’s Disease

Shotgun metagenomics were performed on 66 cecal samples from 4- and 18-month-old mice and on 170 fecal samples from 4-, 8-, 12-, and 18-month-old mice. The samples had a slightly higher representation of WT animals than 5xfAD animals (128 WT versus 108 5xfAD) and slightly more females than males (124 females versus 114 males). The average number of reads per sample for each cohort ranged from 1.1 to 1.9 million, with 18 to 28% of those reads aligning to a genome in the GenBank NT database.

Clustered heat maps of the 100 most abundant microbial species in cecal (Fig. 2A) and fecal (Fig. 2B) samples showed many abundance variations with respect to both age and genotype. For example, the first three bacteria in the heat map for the cecal samples (Limosilactobacillus reuteri, Turicibacter sp. H121, and Faecalibaculum rodentium) had lower abundances in the 18-month 5xfAD samples, relative to both the 18-month WT samples and the 4-month samples of both genotypes. Those microbes in the seventh cecal cluster (comprised of 17 species from the genera Burkholderiales, Akkermansia, Ligilactobacillus, and Bacteroides, among others) were generally more abundant at 18 months of age than at 4 months of age. In the fecal samples (Fig. 2B), T. sp. H121 and L. reuteri were found together in the first cluster with Staphylococcus nepalensis, while F. rodentium was found in the third cluster. Organisms in the seventh cluster of the fecal samples had increased abundances in 18 month old mice. Among others, this cluster included multiple species from the genera Ligilactobacillus, Bacteroides, Alistipes, and Duncaniella. Microbial abundance differences with respect to sex were not readily apparent in this representation of the data.

The richness (Fig. 3A) and the Shannon diversity (Fig. 3B) of the cecal microbiomes did not significantly differ with respect to age or genotype (P > 0.05 via a Kruskal-Wallis one-way ANOVA). For the fecal samples, both the richness and the Shannon diversity showed significant overall differences with respect to age but not with respect to genotype (via a two-way ANOVA). Both of the fecal alpha diversity metrics increased between 4 and 8 months of age and then successively decreased thereafter. Separating the cecal and fecal data based on sex (i.e., evaluating males and females separately) (Fig. S1A) also failed to uncover significant Shannon diversity differences with respect to genotype.

Beta diversity, as evaluated through both nonmetric multidimensional scaling (NMDS) (Fig. 3C and D) and a principal coordinates analysis (PCoA) (Fig. 3E and F) of the Bray-Curtis dissimilarity matrix, showed general separation by age but not by genotype. The NMDS results exhibited high stress values (cecal = 0.179; fecal = 0.193) and a substantial overlap of the 95% confidence intervals for genotype, indicating weak associations. Inspection of the PCoA loadings (gray vectors in Fig. 3E and F) revealed separation arising from contributions of many species (as opposed to a few species driving differences between groups). The PCoA of the males and females separately (Fig. S1B–E) showed more distinct separation by age but not by genotype.

Permutational multivariate analysis of variance (PERMANOVA) was performed in multiple ways to explore significant differences in beta diversity, including for: cecal samples only (dashed bars in Fig. 3G), fecal samples only (spotted bars in Fig. 3G), cecal and fecal samples together (“both” in Fig. 3G), and each sample type and age in isolation (Fig. 3H). Significant differences (P < 0.05 via PERMANOVA; data tables in Fig. 3G and H) were observed for all variables examined (genotype, sex, sample type, cohort [age], and cage assignment), with the exception of genotype for the 4-month-old cecal and fecal samples. In all cases examined, the cage resulted in the largest proportion of variance explained (PERMANOVA R² > 50%). The proportion of variance explained by genotype showed a remarkable increase as the animals aged, indicating a gradual divergence in microbiome composition between the WT and 5xfAD animals (Fig. 3H, inset). For both the 18-month-old cecal and fecal samples, nearly 10% of the variance was explained by genotype. Age also accounted for a substantial portion of the variance (R² > 16% in all cases). Sex and genotype accounted for more of the variance in the cecal samples than in the fecal samples. The full PERAMANOVA statistics are provided in Table S1. An important point to note about this analysis is that many of the variables used in this comparison are inextricably conflated as: (i) littermates of the same sex were often (but not always) housed together; (ii) males and females were always housed separately; and (iii) mice of different age groups were housed separately. Because littermates were housed together, no effort was made to disentangle cage effects from maternal effects.

Random forest was used to elucidate specific species-level differences in the microbiome between genotypes (Fig. 4A–F). AD pathology became more pronounced as the mice age. Therefore, the 18-month-old mice were the primary focus of this analysis. Clear separation was observed between genotypes, both when data from the cecal and fecal sample types were combined (Fig. 4A) and when they were considered separately (Fig. 4B and C). In all three cases, overlapping ellipses (which represent the 95% confidence intervals) showed that the groups share many features in common. Turicibacter sp. H121 was the species most responsible for the random forest separation (>10% mean decreased accuracy in all separations), with Romboutsia ilealis and Lactobacillus johnsonii also playing substantial roles (Fig. 4D–F).

Random forest of the 18-month-old samples with grouping by sample type and sex (ignoring genotype differences) also showed interesting trends (Fig. S2A–D). Grouping by sample type showed an out-of-bag accuracy of 90%, with Oscillibacter sp. PEA192, Flavonifractor plautii, and Desulfovibrio desulfuricans contributing heavily to the separation (Fig. SA and B2). Grouping by sex was 97% accurate, with Limosilactobacillus reuteri and Lactobacillus johnsonii playing the largest role in the separation (Fig. S2C and D). Differences with respect to age were also examined (Fig. S2E and F), with large overlap between each age cohort. The 4-month samples showed the most variation and the least overlap relative to the other age cohorts. Staphylococcus nepalensis was the species most responsible for the separation by age, with a mean decreased accuracy (MDA) of almost 12%.

Random forest is generally applied as a classification method and can lead to false positives when used to identify differentially abundant species. Therefore, we also applied a linear mixed-effects (LME) model to identify bacterial species whose abundances significantly differed between the WT and 5xfAD genotypes (Fig. 4G–I). Using housing ID as the random effect, LME modeling was performed on all samples in aggregate (i.e., 5xfAD [all age groups & sample types] versus WT [all age groups & sample types]) as well as to each age group and sample type individually. The comparisons made for the 18-month samples were: 5xfAD fecal and cecal versus WT cecal and fecal (Fig. 4G), 5xfAD cecal versus WT cecal (Fig. 4H), and 5xfAD fecal versus WT fecal (Fig. 4I). After correcting for the false discovery rate, significantly different species were only found in the 18-month samples. The 18-month cecal samples had 15 differentially abundant species (P adj < 0.05), whereas the other two groupings each had six differentially abundant species. Five species were significant across all three groupings: Turicibacter sp. H121, Romboutsia ilealis, Lactobacillus sp. P38, Ligilactobacillus murinus, and Ligilactobacillus animalis. Among others, six species from the genus Alistipes were differentially abundant in the 18-month cecal samples only. A large proportion of the species (11/17) were also among the top 20 species contributing to the random forest separations.

Relative abundance plots for all 17 differentially abundant species can be found in Fig. S3, with faceting by sample type and sex to allow for the inspection of more subtle abundance differences. The abundances of 8 of the 17 significant species were lower in the 18-month-old 5xfAD mice, and 9 of the 17 species were present at a higher abundance. With the exceptions of Ligilactobacillus murinus (relative abundances as high as 50%), Lactobacillus johnsonii (relative abundance as high as 30%), and Muribaculum intestinale (relative abundance as high as 16%), the differentially abundant species were minor contributors to the overall microbiome composition, displaying relative abundances below 5%.

Turicibacter was recently identified as a potential mediator of the gut-brain axis in AD through its ability to consume and regulate the production of, 5-hydroxytryptamine (5-HT, serotonin) (33). Further studies have highlighted the potential role of serotonin in AD (recently reviewed by Aaldijk et al. in 2022) (34). These established findings, as well as the importance of Turicibacter sp. H121 in our random forest and LME analysis, motivated us to pursue a more detailed investigation. A Bowtie2 sequence alignment was performed against reference genomes for three species of Turicibacter (T. sp. H121, T. sp. HGF1, and T. sanguinis) that encode a neurotransmitter sodium symporter-related protein with sequence and structural homology to the mammalian serotonin transporter (SERT; see Materials and Methods for more details). All three species were present in both the fecal and cecal microbiomes (Fig. 5A and B), and, in agreement with the NCBI BLAST data for T. sp. H121 shown in Fig. S3, they were significantly less abundant in the 5xfAD mice, relative to the WT animals at 18 months of age (P < 0.05). Any apparent differences at other ages were not statistically significant.

Quantitative metabolomics was performed on the blood plasma of the 12- and 18-month-old mice in our study. For these data, we utilized the Biocrates Absolute-IDQ P180 Kit (see Materials and Methods), which was previously applied to develop blood-based biomarker assays for preclinical AD (35–37). From this list of 188 potential analytes, we received concentrations for 137 metabolites within the assay’s quantitation limits for more than 50% of our samples, which was the quality threshold used for this report.

The metabolites showed similar aggregate concentrations in the 12 and 18 month samples (P > 0.05 via a Kruskal-Wallis test) (Fig. S4A). The Shannon alpha diversity was also similar across all age and genotype groupings (P > 0.05 via a Kruskal-Wallis test) (Fig. S4B). PCoA did not reveal clear separation with respect to age, sex, or genotype (Fig. S4C). PERMANOVA of the Bray-Curtis dissimilarity matrix (Fig. S4D) showed sex to be the only significant distinguishing factor (P = 0.037; R² = 0.069). Housing ID accounted for almost 40% of the variance in the data set, but it was not significant (P = 0.357; R² = 0.39).

A clustered heat map of all 137 quantified plasma metabolites (Fig. 6A) shows metabolite specific differences with respect to age, genotype, and sex. Relative to their 12-month-old counterparts, the 18-month-old mice appeared to have lower concentrations of α-aminoadipic acid, serotonin, glutamic acid, aspartic acid, spermidine, and spermine. Most of the detected phosphatidylcholines (PCs) were depleted in the 18-month-old 5xfAD mice (both diacyl [aa] and acyl-alkyl [ae] PCs), while many amino acids (shown primarily in the sixth cluster) had an increased relative concentration. Sex-specific differences were pronounced for most of the detected PCs, which clustered together in the fourth cluster.

Random forest (Fig. S5) showed separation by age (75% out-of-bag accuracy), sex (78% accurate), and genotype (81% accurate for the 18-month samples only). Serotonin was the metabolite most responsible for age-related separation (MDA = 8.5%), closely followed by spermidine (MDA = 7%) and spermine (MDA = 7%). Separation with respect to sex was dominated by sphingomyelin sphingolipids (SMs) and phosphatidylcholines (PCs) from both the diacyl (aa) and acyl-alkyl (ae) subclasses. SM C20:2, PC aa C40:3, PC aa C42:6, PC ae C42:0, PC aa C34:3, PC ae C34:2, and SM (OH) C22:2 topped this list, with each accounting for more than 5% of the MDA. Random forest separation with respect to genotype was negligible when both of the age groups were considered together (61% accurate). The most predictive molecules for genotype at 18 months of age were glycine (MDA = 7%), carnosine (MDA = 4.5%), serine (MDA = 4.5), SM C24:1 (MDA = 4.2%), and serotonin (MDA = 4%).

LME was also performed, using the housing ID as the random effect, and several significantly different metabolites (P adj < 0.05) were observed with respect to age, sex, and genotype (Fig. S6A). The concentrations of the metabolites that significantly differed between genotypes are shown in Fig. 6B and C, between ages in Fig. 6D–G, and between sexes in Fig. S6B. Carnosine and lysoPC a C18:1 had significantly lower plasma concentrations in the 5xfAD mice, relative to those of the WT mice (P adj < 0.05). Serotonin, spermine, spermidine, and PC aa C36:0 each had significantly lower concentrations in the 18-month-old mice, relative to the younger cohort. The concentrations of two SM lipids, two PC ae lipids, six PC aa lipids, and lysine significantly differed between sexes. Separately, a targeted analysis using the Wilcoxon signed rank test revealed serotonin to be more concentrated in the 18-month-old 5xfAD animals, relative to their age-matched counterparts (P = 0.038) (Fig. S7A).

Spearman correlations were generated to better evaluate the relationships between the significant microbes at 18 months of age and the plasma metabolome. The Spearman correlations between all 137 quantified plasma metabolites and the 17 significant microbes for the 18-month-old mice are provided in Fig. 7. The analogous correlations for all 137 plasma metabolites and the 100 most abundant microbes are provided in Fig. S8A and S8B.

For the cecal samples (Fig. 7A), strong correlations were observed between the first cluster of bacteria (Ligilactobacillus murinus, Ligilactobacillus animalis, and Lactobacillus sp. P38) and PC ae C30:0, ornithine, acetylornithine, tyrosine, histidine, methionine, arginine, and proline (among others). These bacteria were anticorrelated with most lipids. Similarly, the six Alistipes species (which were found together in the second microbial cluster) weakly co-occurred with several amino acids in the last two metabolite groupings. The remaining 8 microbes (which were all at reduced relative abundances in the 18-month-old 5xfAD samples) showed correlation trends similar to one another. They were weakly anticorrelated with most of the metabolites, aside from three acylcarnitines (propionylcarnitine [C3], butyrylcarnitine [C4], exadecenoylcarnitine [C16-OH]), PC aa C34:2, PC aa C36:2, citrulline, and acetylornithine.

All 17 microbes in the fecal correlation (Fig. 7B) were either anticorrelated or weakly correlated with the top group of 102 metabolites. Ligilactobacillus murinus, Ligilactobacillus animalis, and Lactobacillus sp. P38 were correlated with most of the amino acids and many of the acylcarnitines, and they were strongly anticorrelated with most of the lipids. The six Alistipes species, Muribaculum intestinale, Duncaniella sp. B8 and Duncaniella sp. C9 were weakly correlated with or anticorrelated with all 137 metabolites. Turicibacter sp. H121 and Romboutsia ilealis were weakly correlated with many amino acids, acylcarnitines, and biogenic amines, and they differed from Burkholderiales bacterium YL45, based primarily on the last cluster of 8 amino acids.

Untargeted metabolomics (see Materials and Methods for more details) were also performed on a subset of fecal samples with adequate material, namely, 40 samples from 4-, 8-, and 12-month-old mice. Unfortunately, there was insufficient material to perform both metabolomics and genomics for the 18-month-old mice and therefore there are no fecal metabolomics data for this age cohort. Overall trends did not reveal any major differences with respect to genotype. Fig. S9 contains a more thorough review of this data set, including the normalized abundances of the six fecal metabolites that were found to be significantly different in the 12-month-old females by LME (Fig. S9A–F) and the Spearman correlations between the 100 most abundant fecal metabolites and 17 significant species at 18 months of age (Fig. S9G).

Overall, the richness and Shannon diversity of the fecal microbiomes increased between 4 and 8 months, and they declined thereafter (Fig. 3A and B). Among other factors, changes in gut microbial diversity have been associated with age (38), sex (38), obesity (39), and general health (40). PERMANOVA (Fig. 3G and H) revealed a marked age-associated increase in microbiome variation attributable to genotype (Fig. 3H), rising from 0.9 to 9% for the cecal samples, and from 0.3 to 8% for the fecal samples. This trend indicates that the 5xfAD and WT microbiomes increasingly diverge as the animals age and that the AD pathophysiology becomes more pronounced and chronic. Most of the variance in the fecal and cecal microbiomes was attributable to the specific cages in which the animals resided, a finding that is consistent with the results from other studies of fecal microbiomes in 5xfAD mice (21, 30). Such “cage effects” are routinely observed in microbiome studies, and they serve to increase variability and complicate the associations between trends in the microbial population and the experimental factors of interest (41–44). The consistency of these findings emphasizes the need for the thorough documentation of housing protocols and the accounting for cage effects in the experimental design. Cage effects have been covered extensively in the microbiome literature, and these and other experimental pitfalls were expertly discussed by Kim et al. (45).

At first approximation, housing animals from the control and experimental cohorts together (as we did here) helps the researcher disentangle cage effects from genotype differences; however, this may obscure the true extent of microbiome dysbiosis. For example, in their study of the effects of sodium oligomannate on 5xfAD mice (some of which were housed together with WT controls and others separate), Wang and coworkers found that the WT microbiomes more closely resembled those from cohoused 5xfAD animals than those from WT controls housed separately. Remarkably, Alzheimer’s-associated disease hallmarks, including those both behavioral (e.g., decline in discrimination learning) and physical (e.g., decreased development of M1-type glial cells), also converged with cohousing. This convergence was attributed to coprophagy-induced microbiome dysbiosis of the WT animals (21). Another option is to house each animal in its own cage (suggested by Kim et al. and others) (45); however, this may be prohibitively expensive or even lead to the unintended distortion of results due to social isolation.

Comparing the microbiome composition of the 5xfAD mice to that of the WT B6J controls revealed differences later in life (Fig. 4). The abundances of 17 species of bacteria were found via LME to be significantly different between genotypes at 18 months of age (15 species in the cecal samples and 6 in the fecal samples) (Fig. 4G–I). Of the 17 significant species, 8 were found at reduced abundances in the 5xfAD mice, relative to the WT controls (Turicibacter sp. H121, Romboutsia ilealis, Burkholderiales bacterium YL45, Adlercreutzia equolifaciens, Duncaniella sp. PC9, Duncaniella sp. B8, Muribaculum intestinale, and Lactobacillus johnsonii), and 9 were found at increased abundances (Lactobacillus sp. P38, Ligilactobacillus murinus, Ligilactobacillus animalis, and 6 Alistipes species). A closer look at the random forest analysis (Fig. 4A–F; Fig. S2) shows some concordance between the top variables of importance that are responsible for the separation by genotype and sex, including Turicibacter sp. H121 and Lactobacillus johnsonii. This agreement may be the reflection of a sex-specific difference in AD-pathology manifesting as differential microbial dysbiosis.

Our findings are largely consistent with those of previous studies of the fecal microbiome of people with AD. Using 16S rRNA sequencing, Vogt et al. reported significantly decreased levels of 7 genera of bacteria in people with AD, including Turicibacter and Adlercreutzia. They also reported 7 genera with increased abundances in people with AD, including Alistipes (25). Considering the vast differences in physiology, diet, and environment, the concordance of our findings in 5xfAD mice with those in people with AD is remarkable, and may suggest AD related differences in the gut microbiome that are independent of the host.

The finding of reduced Turicibacter spp. is also noteworthy because members of this genus can both metabolize and induce the production of serotonin in the gut. The association between serotonin and T. sanguinis has been shown to be both causative and bidirectional (46). Along with depleted Turicibacter (25), both blood-derived and central nervous system-derived serotonin are depleted in humans with AD pathology (34, 47, 48).

Here, we found three Turicibacter spp. (i.e., T. sp. H121, T. sp. HGF1, and T. sanguinis) to be significantly less abundant in 5xfAD mice (relative to WT mice) at 18 months of age (Fig. 5).

Our measurement of serotonin in the blood plasma of 12- and 18-month-old mice showed the neurotransmitter to be at a significantly higher concentration in 18-month-old 5xfAD animals relative to age-matched WT controls (Fig. 6D; Fig. S7A). Although this finding is surprising due to the difference between our serotonin results and corresponding measurements in human plasma, it has been reported by others for 5xfAD mice (21). Because of the potential importance of serotonin to AD, it must be investigated whether this discrepancy between patients with AD and 5xfAD mice extends to other animal models. The other remarkable finding with regard to serotonin was the dramatic drop in concentration between 12 and 18 months of age, which may have larger implications for the health and experimental validity of the older mice. Alternatively, the drop in serotonin levels may be related to dynamic interactions between chronic, established AD pathology and associated nonlinear physiological responses.

Serotonin was not detected in our fecal metabolomics data; however, we did detect tryptophan (which is the primary biochemical precursor of serotonin synthesis) as well as one of its bioactive metabolites (5-hydroxyindoleacetic acid; 5-HIAA) (Fig. S7B and C). 5-HIAA was significantly more abundant in the fecal material of 5xfAD mice, relative to their WT littermates at both 8 and 12 months of age, whereas the differences in tryptophan abundance were not significant. This finding may indicate the increased synthesis and metabolism of serotonin in the gastrointestinal tracts of 5xfAD mice, which in turn results in higher levels of its primary metabolite 5-HIAA.

It is important to note that the relationship between gut, blood, and brain serotonin is complicated and that causative associations are difficult to establish. For example, although more than 90% of the body’s serotonin is made in the gut, it does not cross the blood-brain barrier (48).

The two plasma metabolites found via LME to be significantly lower in concentration in 18-month-old 5xfAD animals (carnosine and lysoPC a C18:1) (Fig. 6B and C) have both been associated with AD. Among other positive effects (49), carnosine (a dipeptide comprised of beta-alanine and histidine) and related compounds have been found to suppress beta-amyloid toxicity (50), suggesting that lower concentrations in the bloodstream may exacerbate AD symptoms. Interestingly, lysoPC a C18:2 (differing from lysoPC a C18:1 only by the position of the double bond) was previously found by Fiandaca et al. to be lower in people who were at risk for developing AD (36). Several other P180 biomarkers highlighted by both Fiandaca et al. (36) and Mapstone et al. (35) showed trends in our data that were similar to those in people with AD. However, following quality control, they lacked adequate statistical power with which to provide meaningful context to this report.

The dominance of cage effects undoubtedly obscures other interesting trends in the microbiome and metabolome, including differences due to sex, age, and genotype. The cage effects were also inexorably linked to other factors, including maternal identity, as same-sex littermates were housed together. This linkage may cause even more pronounced cage effects, contributing to the large proportion of variance attributed to the housing ID. Interestingly, the proportion of variance attributable to housing was approximately equivalent for both the fecal and the cecal samples (Fig. 3G and H; Table S1). In our opinion, the housing of AD model animals needs careful and systematic study, not only from the perspective of the microbiome and the metabolome but also with the consideration of microbiome-induced changes in behavior and neuropathology. Further innovation in animal husbandry strategies may also be warranted (45, 51).

Our results from plasma metabolomics, although powerful, were limited by many measurements falling outside the quantitative range. This resulted in the necessary exclusion of many metabolites and the deterioration of statistical power. Future experiments should strive to overcome this issue by using a larger sample volume or perhaps a more sensitive metabolomics method. Depending on the question at hand, earlier time points (e.g., 4 months) could also be informative.

Unfortunately, we had inadequate material to perform both genomics and metabolomics on all fecal samples, forcing us to choose between the two techniques in some cases, including at the critical time point of 18 months. We prioritized genomics, and, as a result, we were only able to collect metabolomics for a subset of the samples (namely, the 4-, 8-, and 12-month fecal samples). Due to the advanced stage of AD pathology, it is likely that an analysis of the 18-month metabolomes would have produced more meaningful results than the analysis with the younger animals.

Finally, in this work, we largely confined our analysis of the microbiome to the perspective of taxonomy. Evaluating this same data from the standpoint of functional genomics, that is, how functionally predicted genes identified through metagenomics correlate with detected metabolites in the blood or feces, may yield additional insights.

All of the animal experiments were approved by the UC Irvine Institutional Animal Care and Use Committee, and they were conducted in compliance with all of the relevant ethical regulations for animal testing and research. The animals were bred and aged together in the Transgenic Mouse Facility at UCI, and they were maintained in a 12/12-h light/dark cycle. 5xFAD hemizygous (B6.Cg Tg(APPSwFlLon,PSEN1*M146L*L286V)6799Vas/Mmjax, Stock number 34848-JAX, MMRRC) and its wild-type littermates were produced via crossing or IVF procedures with C57BL/6 J (Jackson Laboratory, ME) females. After weaning, they were housed together with same-sex littermates until harvest. Littermate female and male mice from 4, 8, 12, and 18 months of age were used in this study. All of the mice were fed a LabDiet Irr6f (7% fat) diet and were sustained on pH 2.5 to 3.0 autoclaved water. Samples were collected at 4, 8, 12, and 18 months of age, based on multiple factors, including the known progression of AD pathology, interest in dynamic accumulation, long-term exposure to AD pathology, especially amyloid plaques, and synergy with other ongoing studies of the same animals.

Immediately prior to sampling or sacrifice, the animals were isolated into individual cages for fecal collection. All working areas and tools were sterilized with 70% ethanol. Fecal pellets and cecum samples were deposited directly into 1.5 mL Eppendorf tubes. All sample tubes were placed on dry ice following collection. All samples were stored at −80°C until analysis.

DNA was extracted from both fecal and cecal samples by Zymo Research Corp. using the ZymoBIOMICS-96 MagBead DNA Kit (Zymo Research, Cat. number D4302). The quantity of DNA in each sample was measured with a Quant-iT PicoGreen dsDNA Assay Kit (ThermoFisher, Cat. number P11496) and read with a Synergy H1 Microplate Reader (BioTek, Cat. number BTH1M). The sequence libraries were prepared from the extracted DNA samples using a Nextera DNA Flex Library Prep Kit (Illumina, Cat. number 20018705), and a low volume variation of the standard protocol was followed (52).

The average input DNA for each sample was 391.6 ng (std dev = 54.8 ng). Samples were prepared for PCR with Kapa HiFi HotStart ReadyMix (Roche, Cat. number 07958935001) using the following primers: KAPA-PCR-F: 5′–AATGATACGGCGACCACCG*A−3′ and KAPA-PCR-R: 5′–CAAGCAGAAGACGGCATACG*A–3′.

PCR was performed with an Eppendorf Mastercycler Nexus Gradient (Eppendorf, Cat. number 2231000665), using the standard thermal cycles for the Nextera Flex Kit. The resulting sequence fragments were analyzed on an Agilent Bioanalyzer to determine the fragment length distribution (Agilent, Cat. number G2939BA). The sequence libraries were pooled based on DNA concentration, as determined by the Bioanalyzer. After pooling, the libraries were sent to Novogene Co., Ltd. for sequencing on an Illumina HiSeq4000.

The sequence libraries were downloaded from the Novogene FTP website. The libraries consisted of 456 samples pooled into a single dual-index, paired-end library and run across two lanes. The two paired-end aggregate FASTQ files were parsed using BBDuk to trim adapters and remove artifact mouse DNA (GCA_000001635.8), and rat DNA (GCA_000001895.4). The cleaned files were demultiplexed using demuxbyname.sh inside the BBmap suite. To assign microbial taxonomy, GenBank’s NT database (version 1/29/20) was retrieved using the update_blastdb PERL script. Each pair of metagenome files was queried against NT using GenBank’s MEGABLAST algorithm from the BLAST+ package (version 2.10.0), using all default parameters. For the phylogenic assignment, the NCBI Taxonomy database dump files were downloaded from the NCBI FTP site. The BLAST outputs were parsed to retrieve the accession numbers of the hits, and then GenBank’s blastdbcmd program, along with the taxonomy dump files, were used to obtain the full taxonomic rank of each hit.

When a sequence read maps to multiple taxa, it is general practice to either ignore all hits for that read or divide the hit evenly across the taxa. The first option excludes taxa with homologous sequences, whereas the second option introduces false positives into the taxonomical assignments and alters abundance estimates. We chose the second option and mitigated false positives by excluding taxa with abundance counts of less than one.

During the analysis, a major update was made to the Lactobacillus taxonomical assignments, which included the parsing of Lactobacillus into 25 new genera (53). To accommodate this change, we updated the names of our final species-level taxonomical assignments using the NCBI Taxonomy Browser (https://www.ncbi.nlm.nih.gov/Taxonomy/TaxIdentifier/tax_identifier.cgi).

Approximately 0.5 to 1.0 mL of blood was drawn after the cessation of respiration via a closed-chest cardiac blood draw, using a tuberculin syringe (1.0 mL) and a 25-gauge needle. The aspirated blood was immediately transferred to a 2.0 mL EDTA tube at a rate of approximately 1 mL/30 s to mitigate hemolysis. The EDTA tube was inverted 10 times and was placed in wet ice. The EDTA blood tubes were processed within 1 h of placement into ice. Blood components were separated via centrifugation at 2,600 rpm (1,500 × g) for 10 min at 20°C. Plasma was carefully aspirated from the EDTA tube using a micropipette and aliquoted as 100 μL volumes into siliconized cryovials and stored at −80°C until metabolomic analysis. The small plasma volumes allowed for a single freeze-thaw cycle prior to the metabolomic analysis.

We performed targeted metabolomic measurements using stable isotope dilution multiple reaction monitoring mass spectrometry (SID-MRM-MS). We used the Biocrates AbsoluteIDQ P180 (BIOCRATES, Life Science AG, Innsbruck, Austria) Kit, as in our previous work (35–37). For this targeted analysis, frozen plasma samples were thawed on ice, vortexed, and processed as per the manufacturer’s instructions. The metabolite abundances were measured in a 96-well format, including seven calibration standards and three quality control samples that were integrated as part of the kit. The samples were analyzed with a triple quadrupole mass spectrometer (Xevo TQ-S, Waters Corporation, USA), operating in the multiple reaction monitoring (MRM) mode. The kit is capable of providing absolute quantitation for 21 amino acids, hexoses and carnitines, 39 acylcarnitines, 15 sphingomyelins, 90 phosphatidylcholines, and 19 biogenic amines. The flow injection analysis (FIA) tandem mass spectrometry (MS/MS) method was used to simultaneously quantify the 144 lipids via multiple reaction monitoring. The other metabolites were resolved via ultra-performance liquid chromatography (UPLC) and quantified using scheduled MRMs. The abundance of each metabolite was calculated from the area under the curve by normalizing to the respective isotope-labeled internal standard provided with the kit. The concentration of each metabolite is expressed in nmol/L. We used EDTA plasma samples spiked with standard metabolites as quality control samples to assess the reproducibility of the assay. The quantitated data were uploaded to RStudio, and values less than the limit of detection (LOD) were set to NA or 0, as appropriate for the specific analysis.

Fecal material was shipped on dry ice to the West Coast Metabolomics Center (WCMC) at UC Davis for extraction and analysis via hydrophilic interaction chromatography (HILIC) time-of-flight (TOF) mass spectrometry (MS). Standardized WCMC extraction and analysis protocols for “Biogenic amines by HILIC-QTOF MS/MS” were used throughout (analogous to those described by Ding [2021]) (54). Approximately 4 mg of fecal material were used from each sample. The final extract was dissolved in 100 μL acetonitrile/water (4:1, vol/vol) that contained internal control standards, and 3 μL of each extract was analyzed on a SCIEX 6600 TTOF, using both positive and negative ionization. We received the resulting data from the WCMC as a table containing ion identifications (where possible) and peak heights normalized to the sum of the internal standards.

The processed sequence data were analyzed using RStudio Version 1.4.1106 (55) without rarefaction. Most of the graphics were created using the R package ggplot2 (56). The heat maps in Fig. 2 were generated from the standard normalized abundance data via the pheatmap package with “average” clustering (57). The Shannon diversity, nonmetric multidimensional scaling (NMDS), and permutational multivariate analysis of variance (PERMANOVA) were computed using the Vegan package (58). Both NMDS and PERMANOVA were applied to the Bray-Curtis dissimilarity matrix. PERMANOVA was performed using the Adonis function with 999 permutations and the following formula (SampleType and Cohort omitted for some analyses): data ~ Genotype + Sex + SampleType + Cohort + HousingID. Principal coordinates analysis (PCoA) was performed on the Bray-Curtis dissimilarities of the relative abundances using the cmdscale function from the stats package, and the resulting data were visualized with ggplot2 (56). Random forest was performed using the rfPermute package (v2.1.81) with a relative abundance threshold of 0.001. Linear mixed-effects modeling (LME) was performed on the relative abundance data using the lmer function in the lme4 package (46) with a relative abundance threshold of 0.001 and housing ID as the random effect.

To determine the abundance of Turicibacter spp. in our samples, three reference genomes were downloaded from NCBI: Turicibacter sp. H121 (GenBank assembly accession: GCA_001543345.1), Turicibacter sp. HGF1 (GenBank assembly accession: GCA_000191865.2), and Turicibacter sanguinis (GenBank assembly accession: GCA_013046825.1). Bowtie2 v2.4.4 was used with the default parameters to align the reads from our samples to each Turicibacter reference genome (59, 60). The Turicibacter reads per sample were obtained and converted to relative abundances by using the total number of reads per sample. A Wilcoxon test was used for the significance testing in R v3.6.

The analysis of both the fecal and plasma metabolomics data was performed using methods adapted from the microbiome analysis described above. The Spearman correlations between the metabolites and the species were generated using the cor function (method = spearman) from the stats package, and the resulting data were plotted using the pheatmap function from the pheatmap package (57).

Raw sequence reads are available through the Sequence Read Archive (PRJNA902000). Both metabolomics and processed sequence data has been archived on Dryad (https://doi.org/10.7280/D1KH5N). The R scripts used to analyze all data are available on GitHub (https://github.com/sjbd1/5xfAD_mBio2022).