Microbiome gene activity linked to thinking skills in older adults at risk for Alzheimer's

Older individuals (60 y or older) residing independently in Massachusetts were recruited to the Gut-brain Alzheimer's Disease Inflammation and Neurocognitive Study (GAINS). Individuals with AD or MCI (see grouping criteria below) were targeted along with healthy controls, usually a family member of a subject with AD. We excluded participants with antibiotic use in the preceding 60 d and those with gastrointestinal conditions (i.e., bowel shortening surgery or colostomy), which are likely to grossly affect gastrointestinal physiology. We also excluded participants with other neurodegenerative diagnoses, such as Lewy body disease and Huntington's disease. During enrollment, demographics (i.e., age, sex, race, and education) and medical history were collected. At enrollment and each follow-up visit (occurring every 90 d), fecal samples, nutritional status (using the Mini-Nutritional Assessment,40), frailty (measured using the Canadian Study of Health and Aging's 7-point Clinical Frailty Scale (CFS)),41 hospital exposure and medications were collected. We also performed a battery of cognitive assessments at each visit, including the Alzheimer's Disease Assessment Scale-Cognitive Subscale 13 (ADAS-Cog-13),33 the Clinical Dementia Rating Scale (CDR),39 and the Cognitive Battery (CB) of the National Institutes of Health (NIH) Toolbox for the assessment of neurological and behavioral function.42^,43

Subjects were categorized into three groupings post hoc: Alzheimer's disease (AD), mildly cognitively impaired (MCI), and cognitively normal healthy controls (HC). For subjects to be labeled as having AD, they or their caregiver self-reported a diagnosis by a clinician. To confirm the clinical diagnosis, research staff reviewed medical records, including neurological assessments, neuroimaging, and other tests. The current standard of care and clinical guidelines remain largely based on symptomology, supplemented by neuroimaging and CSF analyses where available. Blood-based biomarkers are now available to some institutions, but laboratory value thresholds for diagnosis vary across different institutions. We note that these supplemental tests are often not ordered if the results would not alter the clinical management of patients. For the purposes of this study, AD was a clinical diagnosis given the lack of uniformity of confirmatory neuroimaging in the presence of blood-based markers. The subjects were classified as having MCI if they met all of the following criteria at baseline: 1) a CDR score <1; 2) normal daily functioning; and 3) a score greater than or equal to four for the word list delayed recall task on the ADAS-Cog-13 assessment, which coincides with a score <1 SD below the mean for the general age group.44 Subjects without either a dementia diagnosis or impaired cognitive performance were labeled as HC for healthy control.

The ADAS-Cog-13 subscale was originally developed as an extension of the original 11-item ADAS-Cog, which added tasks to assess attention and concentration, planning and executive function, verbal and nonverbal memory, and praxis.33^,45 Scores ranging from 0 to 85, with higher scores indicative of higher cognitive impairment.

To examine whether performance in the two early-impaired AD-specific cognitive domains of memory and executive function is associated with microbiome composition, we created composite z-scores to ease interpretation and comparison across domains. For our memory z-score, we calculated the sum at each sample point of the word recall score, word recognition score, orientation score, remembering word recognition score, and delayed recall score. The mean and standard deviation of the sum score of all the subjects at each visit time were calculated. A z-score for each visit was calculated by subtracting the mean memory sum score of the visit from the memory sum score of a subject and dividing it by the standard deviation of that visit. An executive function z-score was found by first summing the executive function maze time score, maze number of mistake score, dimensional card sort task (NIH toolbox), and the pattern comparison processing speed test score (NIH toolbox). The mean and standard deviation of each group of tests were calculated for each visit group and used to calculate a visit and sample-specific z-score. For both measures, a higher value is indicative of greater cognitive impairment.

Stool samples were collected at home and mailed to the laboratory. All stool samples were collected as part of the participant's normal bowel habits from defecation using the Fisherbrand™ Commode Specimen Collection System and the sample was placed into the OMNIgene•GUT collection kits (DNAgenotek, catalog no. OMR-200), which stabilizes the sample at room temperature for shipment via mail service.

Samples were stored at −80 °C until DNA extraction and sequencing were performed. DNA was extracted from samples using the QIAGEN DNAeasy PowerSoil kits (QIAGEN, catalog no. 47016). Sequencing libraries were constructed using the Nextera XT DNA Library Prep Kit. Libraries were sequenced on a NextSeq 500 or NextSeq 1000 system with 2 × 150 nucleotide paired-end reads.

Quality control was performed on sequencing reads using the KneadData pipeline (https://github.com/biobakery/kneaddata↗) version v0.12.0, which trims low-quality reads and filters out host contamination sequences (aligned to the human reference version hg37dec_v0.1). Metagenomic reads were profiled for microbial abundance, metabolic pathways, and other functional measures using MetaPhlAn4 and HUMAnN3 databases and tools.46^,47 Sequences were also annotated using the MetaCyc48 and KEGG49 databases for analysis of functional metabolic pathways.

Participants were recruited to the GAINS from community centers and physicians' offices across central Massachusetts. Based on information collected at the first clinic visit, participants were categorized into one of three cognitive status groups: cognitively normal controls (HC, n = 158), those with mild cognitive impairment (MCI, n = 40), and those with Alzheimer's disease (AD, n = 25) (see Methods section for classification details). Over 4 y, these 223 participants provided fecal samples at least once, with an average of 3.8 (SD 2.4) samples collected per participant for a total of 856 samples. The average age of the participants was 71.5 y (SD 7.5), but age differed significantly between the cognitive groups, with MCI being older and HC being younger on average (Table 1). A total of 63.7% of the total GAINS cohort was female, with more females in the HC group and roughly equivalent sex ratios in the MCI and AD groups (Table 1).

Polypharmacy, known to be associated with dementia54 and defined as taking five or more regularly prescribed medications, was prevalent in 34.1% of participants, with the highest percentage (68%) in the AD group and lowest percentage (27.2%) among HCs. The incidence of polypharmacy was statistically significant between the cognitive groups. A total of 17.5% of all participants were exposed to antibiotics in the 6 months preceding sampling (Table 1), with the MCI group having the highest exposure (22.5%). The prevalence of all daily medications is listed in Supplemental Table 1.

Worse clinical frailty and malnutrition indicator scores correlated significantly with cognitive status (Table 1). Both frailty and malnutrition are related to AD and mild cognitive impairment.55-58 Clinical frailty was also positively correlated with age and malnutrition in the HC and MCI groups (Figure 1). Among participants with AD, clinical frailty was positively correlated with malnutrition but negatively correlated with age, indicating that younger patients in this group were frailer. Age did not correlate significantly with malnutrition in any of the three cognitive groups (Figure 1).

A battery of cognitive testing was performed at each visit, providing us with three measures of interest: ADAS-Cog-13 score, memory z-score, and executive function (EF) z-score. The ADAS-Cog-13 is a global measure of cognitive function with a specific emphasis on domains implicated in AD progression,59 including memory, language, executive function, attention, and praxis. As expected, there were statistically significant differences in ADAS-Cog-13 scores, with the AD group scoring highest (worse cognition) and HC group scoring lowest (Table 1). As memory and executive function represent domains profoundly affected in AD during the early clinical stages, we focused on these two domains by calculating composite z-scores for memory and executive function using components of the ADAS-Cog-13 test and the NIH toolbox (see Methods for further explanation of derivation). Like the ADAS-Cog-13 score, higher memory and EF z-score values indicate greater impairment in the given domain or worse performance.

Performance on cognitive tests can be influenced by a myriad of factors, including age and years of education.60 In the HC and MCI groups, age was correlated with higher scores for all three cognitive measures (Figure 1A, B). In participants with AD, age was correlated with lower scores for all tests, suggesting that younger participants are more severely impaired (Figure 1C). Our MCI group will naturally consist of individuals who are experiencing “normal” aging as well as individuals who will progress to AD and other forms of dementia, which we cannot distinguish in the current study. Considering that our MCI group is overall older than the HC and AD groups, our observation of a matching direction of correlation between age and cognitive performance between the HC and MCI groups and an inverse correlation between the MCI and AD groups may indicate that our MCI group consists of a higher proportion of “normal agers.” A negative correlation was found between years of education and cognitive score in the HC cohort but not in the MCI or AD groups (Figure 1). Frailty was positively correlated with all scores in the AD group; in the MCI and HC groups, frailty was positively correlated with ADAS-Cog-13 and EF z-scores (Figure 1). Malnutrition was negatively correlated with ADAS-Cog-13 and EF z-scores in the HC cohort (Figure 1A). Altogether, our analysis suggests that among HC participants, health and education measures are more informative of cognitive scores than are participants with AD. In the AD group, younger age and increased frailty correlated with cognitive scores.

Within the GAINS cohort, alpha diversity, a general metric of microbiome composition within individual samples, was not significantly different across different cognitive statuses (Supplemental Figure 1). However, antibiotic usage within the past 6 months prior to sampling was significantly associated with lower diversity (Supplemental Table 2). To examine associations between microbial abundance and clinical covariates in the GAINS cohort, we ran mixed-effect random forest (MERF) modeling to predict each species abundance as a function of demographics and clinical covariates (Figure 2). A mix of demographic, medication, and environmental factors was significantly predictive of microbial abundance across the 311 microbe-specific models (FDR-adjusted p-value < 0.05; Figure 2). While age was the most frequently selected feature (appearing in 57.2% of models), medication accounted for 54.79% of the significant predictors. Specifically, psychoactive and antihypertensive medications were most prevalently associated with microbiome abundance.

We next sought to identify microbiome features associated with cognitive scores between the patient groups. As above, we built MERF regression models for HC, MCI, and AD patients separately, beginning with species abundance combined with demographics and clinical covariates. Models predicting ADAS-Cog-13 scores in the HC group produced marginally informative results, as evident by the correlation between actual and predicted ADAS-Cog-13 for the species abundance-based model, 0.64 ± 0.05 (Figure 3A, Supplemental Table 3). Conversely, models trained on MCI and AD subsets demonstrated strong predictive performance, with correlation coefficients of 0.87 ± 0.05 and 0.94 ± 0.03, respectively (Figure 3B, C and Supplemental Table 3). These results may reflect the functional design of the ADAS-Cog-13, which is powered to identify differences in MCI and mild AD cases and thus focuses on AD-specific cognitive domains. The test is thus not developed or powered to assess general cognition in healthy individuals.

As the MCI and AD patient's subsets cognitive performance-microbiome model displayed strong predictive power, we sought to determine the most informative microbial species of the predicted signal by analyzing the permutated importance ranking of features used in the model. We found that the leading microbial predictors of ADAS-Cog-13 scores in the MCI group included two Blautia representatives, a known health-associated Clostridium Cluster IV/XIVa SCFA-producing bacterium,61 (ranked 1st and 4th) and Anaerotignum faecicola (ranked 3rd), all of which showed significantly (p-value < 0.05) lower abundance that strongly correlated (|correlation| ≥ 0.4) with worse(higher) ADAS-Cog-13 scores (Figure 3D and Supplemental Figure 2A). Among participants with MCI, the only clinical covariate in the top fifteen predictors correlated with higher ADAS-Cog-13 scores was increased age (ranked 2nd), suggesting that general age-related cognitive decline may be occurring in the MCI group.

A lower abundance of Lachnospiraceae bacterium (also a Clostridium cluster XIVa member), Coriobacteriia bacterium, and Bacteroides caccae were found to be the top predictors of worse ADAS-Cog-13 scores in the AD group (Figure 3E and Supplemental Figure 3A). Corroborating the finding that frailty strongly correlates with ADAS-Cog-13 scores (as in Figure 2C), the model identified higher clinical frailty as a stronger predictor of higher ADAS-Cog-13 scores in AD patients (ranked 3rd, and only the clinical variable in the top fifteen predictors).

We next sought to determine whether we could disentangle the predictive ability of microbial features seen for ADAS-Cog-13 to memory and EF specifically, since these cognitive domains are affected earlier in clinical disease. MERF models trained to predict memory z-scores from species abundance in combination with demographics and clinical features predicted scores that tightly correlated with the actual scores for both the MCI (correlation = 0.815 ± 0.07) and AD (correlation = 0.929 ± 0.01) populations (Figure 4A). The predictions of EF z-scores were also well correlated with the actual scores in both the MCI (0.838 ± 0.04) and AD populations (0.84 ± 0.04) (Figure 4B). Predictions in the cognitively normal HC population were again subpar and thus not further examined (Figure 4A, B).

We again evaluated features that contributed most to the MERF predictions by permutated importance analysis. As with the MERF models for ADAS-Cog-13 scores, nearly all the top predictors were microbial features. In the MERF model for predicting memory z-scores for the MCI group, two Blautia species and an uncultured microbe (GGB2982 SGB3964) were identified as the top predictors with strong negative correlation (p-value < 0.05 and |correlation| ≥ 0.4) (Figure 4C and Supplemental Figure 2B). In the MERF model predicting memory z-scores for the AD group, only one microbial representative (FaecalibacteriumSGB15346↗) was identified as predictive and significant with strong correlation, ranking 11th. All top predicting microbes for memory z-scores in the AD group correlated positively (Figure 4D and Supplemental Figure 3B). Only two of the fifteen top predictors for memory z-scores were common between the MCI and AD groups (GGB2982 SGB3964 and Evtepia gabavorous) but showed opposite correlations (Figure 4C, D).

In the MERF model predicting EF z-score for individuals with MCI, higher abundance of three bacterial species (Clostridium symbiosum, Bacteroides intestinalis, and Desulfovibrio fairfieldensis) associated with inflammation62-64 was identified as significant with strong positive correlations. A lower abundance of one health-associated microbe (Oscillibacter sp. ER4)65 was identified as significantly and strongly correlated (negative correlation) (Figure 4E and Supplemental Figure 2C). In the model predicting EF z-score for the AD group, six bacteria were significant top predictors with strong correlations (Figure 4F and Supplemental Figure 3C). A lower abundance of an Egerthellaceae species, Odoribacter laneus, and Blautia obeum were correlated with worse EF z-scores (negatively correlated), while higher abundances of Blautia wexlerae, Blautia faecis, and Clostridium scindens were predictive of worse EF scores (positively correlated). Between the MCI and AD groups, only three microbial features were common among the top fifteen predictors for EF z-score (Candidatus Cibiobacter qucibialis, Bacteroides uniformis, and B. obeum), two of which showed opposite correlations between the MCI and AD groups (Figure 4E, F). These species have been associated with both pro- and anti-inflammatory activities, making their functional correlations here unclear.66-69

Only two demographic/clinical features were identified among the top fifteen predictors in these models. Age was selected as a top predictor of both the memory z-score and EF z-score in the MCI population (Figure 4C, E), which matched our analysis of clinical covariate correlations (Figure 1B). Like the ADAS-Cog-13 score, increased age was significantly correlated with worse memory and EF z-scores, suggesting that some impairment may result from age-related cognitive impairment in the MCI population. To this end, the MCI group was significantly older (75.1 ± 7.22 y; p-value < 0.001) than both the HC (70.2 ± 7.46 y) and AD groups (73.9 ± 5.75 y). In the MERF model predicting EF z-scores among participants with AD, both age and clinical frailty were identified as among the top fifteen predictors. Age was negatively correlated, though not significant, while clinical frailty was positively correlated and significantly correlated (p-value < 0.05 and |correlation| ≥ 0.4) (Figure 4F). These correlates again matched our earlier analyses (Figure 1C) and suggest that, unlike the MCI group, younger participants with AD seem to have worse cognition.

Examining the top-ranked important features across the MERF models for the three cognitive measures, only a few predictors ranked consistently as predictive and significantly correlated (Figure 5). Among participants with MCI, age was consistently selected as a predictor with a strong and significant correlation (Figure 5A). Two Blautia members (Blautia sp. MSK 21 and Blautia sp. AF19 10LB) were identified as predictors with significant correlations for both ADAS-Cog-13 and memory z-scores. For EF z-scores, none of the species identified as predictive with significant correlation were selected as the top predictors for ADAS-Cog-13 or memory z-scores. Among participants with AD, no microbial or clinical features ranked as a top predictor in all three cognitive measures. Only four species ranked in two of the three tests. We also failed to find consistent patterns for pro- versus anti-inflammatory-associated microbes across the three cognitive measures for the MCI and AD groups. Altogether, these results lead us to hypothesize that it is not the individual species present, but rather the functions of the microbiome that are shared across different microbes or composite microbial communities predict cognitive functioning.

While the results are informative, microbial abundance alone fails to provide insight into the functionality of the microbial community. Analyses of microbial metabolic pathway genes enable a more complete view of the functional microbiome or the function of genes found in fecal samples that produce metabolites in the gut and interact with or influence the host. Analysis of metabolic pathways may also help reconcile incongruous findings in species abundances (i.e., different species or communities with similar functional gene pathways may produce the same metabolites). Accordingly, we profiled the abundances of microbial-encoded metabolic pathways (using the MetaCyc database) from our sequenced microbiome samples. Metabolic gene pathway abundances were determined by aligning our sequencing results against the MetaCyc database.

Starting with the MCI group, we again utilized MERF models to predict ADAS-Cog-13, memory z-scores, and EF z-scores using a combination of basic clinical covariates and metabolic pathway relative abundance, and analyzed pathways contributing to the predictions by permutated importance. The predicted values correlated well with the actual cognitive scores, similar to our microbial abundance analysis above (ADAS-Cog-13 correlation = 0.879 ± 0.044; memory z-score correlation = 0.811 ± 0.062; EF z-score correlation = 0.801 ± 0.059) (Figure 6A). In the model predicting ADAS-Cog-13 scores, four pathways were selected among the top fifteen predictors, with two strongly correlated pathways (Figure 6B, left panel and Supplemental Figure 4A). Putrescine biosynthesis (PWY-6305) showed a strong and significant negative correlation, while L-lysine biosynthesis (PWY-2941) showed a strong and significant positive correlation. The putrescine biosynthesis pathway involves the metabolism of arginine and its metabolite ornithine as well as the production of urea, which can be a nitrogen source for gut bacteria.70^,71 The other two significant pathways (PWY-5505 and GLUDEG-I-PWY) are both related to the metabolism of glutamate, which can act as a potent neurotransmitter and be converted to the neurotransmitter GABA in mammalian cells.

In the MERF model predicting memory z-scores, lower abundance of several pathways was selected as top predictors, though none with strong correlations (Figure 6B, middle panel and Supplemental Figure 4B). These include the same putrescine biosynthesis and glutamate pathways selected in the ADAS-Cog-13 model as well as pathways involved in glycolysis (including pentose phosphate pathways), N-acetylneuraminate metabolism, and formaldehyde assimilation, indicating potential carbon sources for gut bacteria.

In the MERF model predicting EF z-scores, only one pathway was identified as significant with a strong correlation, a CMP-legionaminate biosynthesis pathway (PWY-6749) (Figure 6B, right panel and Supplemental Figure 4C). This pathway involves glutamate as intermediaries to produce sialic acid derivatives, which are used by some pathogens in virulence and host immune evasion. Higher abundance is correlated with worse EF z-scores, possibly indicating a contribution by potential pathobionts. Pathways involved in lysine and heptose biosynthesis were also significantly positively correlated, though not with a strong correlation. Heptose sugars are important components of LPS in most gram-negative bacteria. A lower abundance of pathways in glycolysis, polyamine biosynthesis, N-acetylneuraminate metabolism, and anaerobic sucrose degradation was also significantly correlated.

Across the three cognitive measures, a lower abundance of a glycolysis pathway was selected as common among the top 15 predictors, although it was not strongly correlated (Figure 6C). Nine pathways were common between the two cognitive tests, with correlations in the same direction. Significant pathways include putrescine biosynthesis, glutamate metabolism, lysine biosynthesis, and N-acetylneuraminate degradation.

For the AD group, MERF models predicting ADAS-Cog-13, memory z-scores, and EF z-scores also predicted values that were tightly correlated with actual scores (ADAS-Cog-13 correlation = 0.935 ± 0.02; Memory z-score correlation = 0.944 ± 0.01; EF z-score correlation = 0.88 ± 0.03) (Figure 7A).

In the model predicting ADAS-Cog-13 scores, all fifteen top predictors showed positive correlations. Higher abundances of five pathways were identified as significant with strong correlations (Figure 7B, left panel and Supplemental Figure 5A), including pathways involved in xyloglucan degradation (PWY-6807), vitamin B12 cofactor salvaging (COBALSYN-PWY), chorsimate biosynthesis (PWY-6163), the urea cycle (PWY-4984), and starch degradation (PWY-6731). In the model predicting memory z-scores, all of the top fifteen predictors also showed positive correlations; only one pathway, the urea cycle, was significant (Figure 7B, middle panel and Supplemental Figure 5B). In the model predicting EF z-scores, most of the top fifteen predictors showed positive correlations. Higher abundance of four pathways was identified as significant with strong correlations (Figure 7B, right panel and Supplemental Figure 5C), including starch (PWY-6731) and chitin (PWY-6902) degradation.

Two pathways were selected as important across all three cognitive outcome models. Notably, the urea cycle (PWY-4984) was strongly and significantly correlated with worse performance in each cognitive outcome (Figure 7C), ranking 10th for ADAS-Cog-13, 1st for the memory z-score, and 4th for EF z-score. Few species in our AD samples confidently contributed to the abundance of the urea pathway (Supplemental Figure 8A). Urea is also a significant source of nitrogen for gut bacteria.72^,73 The pathway for vitamin B12 cofactor salvaging (COBALSYN-PWY) was also selected as important for all three outcomes, with higher abundance correlated with worse scores, with strong correlations were strong for ADAS-Cog-13 and EF z-scores. The vitamin B12 cofactor (adenosylcobalamin), the active form of vitamin B12 in the body, is a cofactor for highly conserved enzymes involved in amino acid synthesis and the breakdown of fatty and amino acids (Rowley and Kendall, 2019). A combination of commensals and few pathobionants contributed to the abundance of the vitamin B12 cofactor salvage pathway (Supplemental Figure 8B). Gut bacteria also use B12 as a critical cofactor for many metabolic reactions, including the production of methionine.74

Several pathways were selected as significantly important features contributing to both worse (higher) ADAS-Cog-13 and memory z-scores in MERF models (positive correlations), though none met the strong correlation cutoff of 0.4. These include pathways involved in the biosynthesis of peptidoglycan (PWY-7953, PWY-6387, and PWY-6385), chorismate (PWY-6163), isoleucine (PWY-3001), and isopentenyl pyrophosphate (NONMEVIPP-PWY). Peptidoglycan is an important and major component of bacterial cell walls. Chorismate is an intermediary for the synthesis of aromatic amino acids (i.e., phenylalanine, tyrosine, and tryptophan). Isopentenyl pyrophosphate is a precursor for the synthesis of isoprenoids, a broad class of compounds essential for both eukaryotes and prokaryotes. Some gut pathogens have been reported to use isoprenoids in virulence and host immune evasion.75^,76 Only one pathway, a pathway for starch degradation (PWY-6731), was common among the top predictors for both ADAS-Cog-13 (ranked 15th) and EF z-score (ranked 1st).

The MetaCyc pathways, as used above, are well curated and highly interpretable. However, understanding the interconnections between pathways and their relationships with specific outcomes often requires complex, manual curation. Thus, to supplement our analyses, we also used the Kyoto Encyclopedia of Genes and Genomes (KEGG) ontology database, which offers broader coverage of pathways and functions that can enable a more holistic view of the microbial functional activity.77 KEGG ontology (KO) terms represent functional orthologs that can be linked to a variety of larger pathways and groups. To identify specific gene functions associated with cognitive outcomes in our cohort, we predicted our three cognitive measures using MERF models trained with KO terms encoded by the microbiome in combination with clinical covariates for the MCI and AD groups.

For the MCI group, the predicted scores for our three cognitive measures matched well with the actual scores (Figure 8A) (ADAS-Cog-13 correlation = 0.888 ± 0.036, memory z-score correlation = 0.81 ± 0.048, EF z-score correlation = 0.808 ± 0.055). The permeated variable importance for KO terms was calculated as above. Of the top fifteen predictive terms, notable individual KO terms that significantly and strongly correlated (p ≤ 0.05 and |correlation| ≥ 0.04) with the outcomes of the cognitive tests included virulence factors and transport proteins (Figure 8B and Supplemental Figure 6). For participants with AD, the actual and predicted ADAS-Cog-13 (correlation = 0.943 ± 0.02), memory z-score (correlation = 0.949 ± 0.02), and EF z-score (correlation = 0.87 ± 0.06) were strongly correlated, indicating high performance by the resulting models (Figure 8D). Notably, the top KO terms with significant and strong correlations in this group included the clinical frailty score and various transport proteins (Figure 8E and Supplemental Figure 7).

To discern patterns in overall functional activity, the top fifteen predictors of each cognitive outcome were combined, and enrichment analysis of the functionality of the selected KO terms was performed for both the MCI and AD groups separately. Genes related to folate biosynthesis were enriched (Figure 8C) in participants with MCI. Among participants with AD, genes related to cysteine and methionine metabolism were significantly enriched, along with genes related to the bacterial secretion system and protein export genes (Figure 8F). Folate biosynthesis and metabolism of cysteine and methionine are connected metabolic pathways that share intermediaries and substrates with vitamin B12.74^,78 The convergence of these pathways in our analyses could indicate a functional role for methionine-related microbial metabolites and/or B vitamins, specifically folate and B12, in the pathogenesis of AD.

The strengths of our study include our comprehensive approach to (1) including clinical features in analyses in combination with microbiome data, (2) using multiple objective cognitive assessments, and (3) analyzing metagenomic data using complementary techniques. Together, these aspects should increase confidence in the robustness of our findings, despite limitations to the study, which include a modest size and relatively homogenous sample population due to the geographic location of the study and limited information on patient lifestyle factors (e.g., diet, exercise, social and community supports, etc.), which may improve or worsen dementia symptoms or progression. Related to these aspects is the bi-directionality of these lifestyle contextual factors on the microbiome—e.g., frailer individuals without family, community, or professional support may have poorer diets, which in turn impacts the microbiome. This study is also limited given the absence of host metabolomics, blood‒brain barrier integrity measures, and CSF and plasma markers. Additionally, AD participants were classified based on clinical records as opposed to blood-based biomarkers, which we plan to collect in future studies. The lack of uniform clinical imaging has limited our ability to classify participants as AD based on biological definitions. In this study, we did not confirm a biomarker diagnosis of AD. Due to the longitudinal nature of sample collection and processing, batch effects were unavoidable. However, correction resulted in no substantial changes in the correlations (Supplemental Figure 9). Like all black-box machine learning approaches, MERF models cannot disambiguate the relative contribution of correlated predictors. As age correlates with both microbiome and cognitive outcomes in our cohort, we cannot exclude the possibility that the model is learning an age-mediated relationship. Further, a limitation of MERF models is that the use of a random effect variable based on patients inherently will carry over information relevant to the outcome. Together, these issues represent an inherent model attribution limitation and possible data leakage of the framework used. Further, the stability of features and the models' performance were not evaluated here, as standard feature selection parameters were used. Finally, it is known that medications can shape the microbiome composition, so we cannot clearly state that these pathways do not influence our reported findings. We aim to address these challenges in future studies with expanded cohorts as well as by longitudinal follow-up to replicate and validate our findings.

Microbiome sequencing and related metadata are available on NCBI (BioProject ID PRJNA1446836).