Alzheimer’s disease (AD) is the leading cause of dementia in the elderly. Neuropathological hallmarks of AD are the deposition of extracellular β amyloid (Aβ) plaques, intracellular accumulation of tau-containing neurofibrillary tangles (NFTs), cortical atrophy and neuroinflammation in the brain (Querfurth and Laferla, 2010). The mechanisms leading to this AD pathology are still unknown. A multifactorial genesis seems likely, with inflammatory changes in the brain including altered microglia function playing an important role (McGeer and McGeer, 1995; Wyss-Coray, 2006). The mechanisms triggering these inflammatory changes are also unknown.

The intestine with its mucosal lymphoid tissue is one of the most important components of the immune system, containing about 70–80% of the body’s immune system (Sochocka et al., 2019). The colon contains about 39 × 1,012 bacteria and many more microorganisms (Sender et al., 2016). The totality of these microorganisms is called intestinal microbiome and contains about 100 × more genetic information than the human genome. The gut microbiome plays a pivotal role in supporting human health by regulation of metabolic, endocrine, immune and neurotrophic functions, e.g., by production of short-chain fatty acids (SCFAs) (Silva et al., 2020) and brain-derived neurotrophic factor (BDNF) (Lee et al., 2018; Ranuh et al., 2019). The influence of the gut microbiome is not restricted to the gastrointestinal (GI) tract, it also plays an important role in the bidirectional communication between the GI tract and the brain (microbiota-gut-brain axis; Morais et al., 2021).

Taxonomic and functional changes to the composition of the gut microbiome have been associated with multiple human diseases, e.g., with chronic inflammatory diseases such as Crohn’s disease, ulcerative colitis, and rheumatoid arthritis (Manichanh et al., 2006; Vigsnæs et al., 2012; Zhang et al., 2015). The composition of the gut microbiome can be impacted by host and environmental factors. It also holds different functional features across various life periods (Kundu et al., 2017). While gut microbiome remains mostly stable during the adulthood, it starts at the age of 65 years—interestingly the predilection age for AD—to shift into a less diverse and resilient state, which makes the microbiome more susceptible to environmental factors (Thursby and Juge, 2017).

Changes in the gut microbiome seem to be also linked with AD. Several recent studies have demonstrated changes in intestinal microbiome composition in animal models of AD (Harach et al., 2017; Chen et al., 2020), in human AD patients (Vogt et al., 2017; Li et al., 2019; Liu et al., 2019; Ling et al., 2020) and in MCI patients (Li et al., 2019; Liu et al., 2019). Most of these studies examined taxonomic data, only a few also functional data and none investigated the combination of taxonomic, functional and clinical data using an ensemble learning approach. Composition of intestinal microbiome is altered in animal models of AD even before the presence of amyloid plaques in the brain and thus seems to be an early phenomenon within AD pathogenesis (Chen et al., 2020). In addition, it has been shown that animal models of AD develop in the absence of intestinal microbiome less amyloid pathology in the brain than animal models of AD with existing or replaced intestinal microbiome (Harach et al., 2017; Dodiya et al., 2019, 2020).

The aim of the present study was to examine how well taxonomic and functional intestinal microbiome data and their combination with clinical data can be used to discriminate between amyloid-positive AD patients and cognitively healthy elderly controls.

Characteristics of the study sample are presented in Table 1. HCs revealed significantly higher age (p = 0.032) and MMSE scores (p < 0.0001) compared to amyloid-positive AD patients. Gender was equally distributed between the investigated groups (p = 0.257). As expected, amyloid-positive AD patients took more frequently AChE inhibitors (p < 0.0001) and antidepressants (p < 0.0001). However, we found no statistically significant influence of antidepressant use (p = 0.25) and AChE inhibitor use (p = 0.21) on the relative abundance of genera.

The best model for discrimination of amyloid-positive AD patients from healthy controls with taxonomic data was obtained analyzing 18 genera features, and yielded an area under the receiver operating characteristic curve (AUROC) of 0.76 (95% CI: 0.67–0.84) in the training set and 0.61 (95% CI: 0.42–0.80) in the validation set (Figure 1A). The best models with functional data were obtained analyzing 17 GO features with an AUROC of 0.81 (95% CI: 0.73–0.88) in the training set and 0.75 (95% CI: 0.56–0.94) in the validation set (Figure 1B) and 26 KO features with an AUROC of 0.83 (95% CI: 0.74–0.91) and 0.77 (95% CI: 0.60–0.93), respectively (Figure 1C). Using ensemble learning for these three models including a clinical model with the 4 parameters age, gender, BMI and ApoE yielded an AUROC of 0.92 (95% CI: 0.87–0.97) in the training set and 0.80 (95% CI: 0.65–0.94) in the validation set (Figure 2).

Box plots of the absolute abundances of the 18 genera used in the genera model compared to all genera in our study cohort are shown in Figure 3. It turns out that all model genera are quite evenly distributed around the median of all genera. The differences of logarithmised medians of the features included in the genera, GO and KO model are shown in Tables 2–4. The non-significant associations between genera, GO and KO features and AD clinical and biomarker characteristics are presented in Tables 5–7.

This study investigated the capability of taxonomic and functional intestinal microbiome data and their combination with clinical data in discriminating amyloid-positive AD patients from cognitively healthy elderly controls. To our knowledge, this is the first study addressing this question using taxonomic and functional data of shotgun metagenomics in a training cohort and an independent validation cohort. Within taxonomic data, we identified a genera model and within functional data, we identified a GO and KO model showing best results for discrimination of amyloid-positive AD patients from healthy controls. Functional data showed a higher diagnostic accuracy than taxonomic data. In a next step, we combined taxonomic, functional and clinical models with an ensemble learning approach. This ensemble model performed better than any of its parts in the training data and with respect to genera and clinical model also in the validation data. Our finding of an added value of including clinical variables to gut microbiome data is in line with previous studies, demonstrating an improved accuracy for AD diagnosis adding clinical variables to blood-based biomarkers (O’Bryant et al., 2011) or MRI data (Vemuri et al., 2008).

Several recent studies have demonstrated an altered intestinal microbiome composition in animal models of AD (Harach et al., 2017) and in human AD patients (Vogt et al., 2017; Li et al., 2019; Liu et al., 2019; Ling et al., 2020) and MCI patients (Li et al., 2019; Liu et al., 2019). Three studies investigated the diagnostic accuracy for discrimination of clinically diagnosed AD patients from healthy controls using genera data with AUROC values ranging from 0.78 to 0.94 (Li et al., 2019; Liu et al., 2019; Ling et al., 2020). However, all three studies had some limitations: They used limited bacterial 16S rRNA sequencing technology, they did not validate their results in an independent cohort with the same diagnosis and they did not examine the diagnostic accuracy of functional data. Instead, Li and colleagues (Li et al., 2019) showed that their genera model from AD patients was able to predict 28 of 30 patients with MCI, suggesting a similar pattern of gut microbiome changes in AD and MCI patients. In contrast, Liu and colleagues (Liu et al., 2019) could discriminate AD patients and MCI patients using a genera model with an AUROC of 0.93, suggesting a different pattern of gut microbiome changes in AD and MCI patients. Both results are conflicting and also surprising, because MCI is a heterogeneous syndrome with varying clinical outcomes: up to 60% of MCI patients develop dementia within a 10-year period, however, many people remain cognitively stable or regain normal cognitive function (Manly et al., 2008; Mitchell and Shiri-Feshki, 2009; Korolev et al., 2016). We hope in the future to be able to resolve these conflicting results by providing more information for differentiating microbiome patterns of MCI patients and testing how much overlap there is with more advanced pathology.

Interestingly, 10 of the 18 taxa in our genera model belong to the phylum proteobacteria. Of these, 5 show higher levels in healthy individuals and 5 in AD patients. This means, that 50% of taxa with higher levels in healthy controls and 62.5% of taxa with higher levels in AD patients belong to the phylum proteobacteria. However, on the phylum level we found no significant difference of the relative abundance of proteobacteria between AD patients and healthy controls (p = 0.41). This result indicates that differences of taxa may play rather a role on the genera level than on phylum level. The potential meaning of proteobacteria in AD has been shown in a previous study, describing increased proteobacteria levels in AD patients, which correlated with the severity of cognitive impairment (Liu et al., 2019). From the 17 features in the GO model, aldehyde dehydrogenase showed the highest activity in AD patients. This is an interesting finding, as this enzyme has been associated with neuroprotective effects due to its aldehyde detoxification and with neurodegenerative diseases such as AD (Michel et al., 2010; Grünblatt and Riederer, 2016). From the 26 features in the KO model, putative multiple sugar transport system permease protein showed the highest activity in AD patients. This protein belongs to the ATP-binding cassette (ABC) transporters which have been associated with AD pathogenesis due to their ability to transport Aβ peptides out of the brain (Aβ brain clearance; Behl et al., 2021).

The intestinal microbiome may have a direct impact on amyloid pathology and neuroinflammation in the brain. This hypothesis is supported by recent preclinical studies showing that the absence or presence of intestinal microbiome influenced the degree of cerebral amyloid pathology in animal models of AD (Harach et al., 2017; Dodiya et al., 2019, 2020). Germ-free Alzheimer’s mice had less amyloid plaques in the brain than Alzheimer’s mice with normal intestinal microbiome. The colonization of germ-free Alzheimer’s mice with microbiota from conventionally reared Alzheimer’s mice resulted in a significant increase in cerebral amyloid load in the brain, but not the microbiota of control mice (Harach et al., 2017). In addition, treatment of male Alzheimer’s mice with an antibiotic cocktail resulted in changes of gut microbiome and was associated with reduced amyloid pathology in the brain and transplants of fecal microbiota from male Alzheimer’s mice restored the gut microbiome and partially restored amyloid pathology in the brain (Dodiya et al., 2019). A recent study showed that when the gut microbiota from AD patients were transplanted into Alzheimer’s mice, the recipient mice showed more severe cognitive impairment, activated intestinal NLRP3 inflammasome, increased levels of inflammatory factors in peripheral blood and activated microglia in the hippocampus, and these effects could be reversed by transplantation of healthy human gut microbiota (Shen et al., 2020).

Which mechanisms could mediate the effects of intestinal microbiome on amyloid pathology in AD patients? A recent study has identified the bacterial products lipopolysaccharide (LPS) and the short chain fatty acids (SCFAs) acetic acid and valeric acid to be positively correlated with amyloid load in the brain of AD patients, whereas butyric acid showed an inverse correlation (Marizzoni et al., 2020). In addition, an in vitro study revealed an anti-Aβ aggregation efficacy for the SCFAs valeric acid, butyric acid, and propionic acid (Ho et al., 2018). This anti-amyloidogenic effect of some SCFAs could be explained in part by their influence on maturation and function of microglia in the brain (Erny et al., 2015), which can clean up Aβ fibrils and typically surrounds amyloid plaques in the brains of AD patients (Calsolaro and Edison, 2016). In contrast, LPS is part of the outer membrane of Gram-positive bacteria and has been shown to trigger systemic inflammation and the release of proinflammatory cytokines (Zhao and Lukiw, 2015) and to potentiate amyloid fibrillogenesis (Asti and Gioglio, 2014).

The findings that gut microbiome is changed in AD and directly influences the degree of amyloid pathology and neuroinflammation in the brain indicate that it could represent a promising target for the prevention or treatment of AD. In this context, the use of a probiotic cocktail for 12 weeks has shown beneficial effects on cognitive performance in AD patients (Akbari et al., 2016). The beneficial effect of gut microbiota on cognition could be explained in part by the upregulation of hippocampal BDNF and cAMP-response element-binding protein (CREB), as shown for different lactobacillus species (Lee et al., 2018; Ranuh et al., 2019). However, a meta-analysis examining the effects of probiotic supplementation on cognition in dementia could not find a significant beneficial effect (Krüger et al., 2021).

Despite its significant implications, the present study has some limitations to be discussed. First, this study used a cross-sectional design. To determine whether the characteristic intestinal microbiome in AD is the cause or result of the disease, well designed longitudinal studies with adequate sample size are needed. Second, CSF was not available in our healthy control group. The reason for that is that we got no permission from our ethics committee to perform lumbar puncture in healthy controls without any signs of cognitive impairment. As the prevalence of preclinical AD among cognitively healthy elderly people is estimated to be about 20%, we cannot exclude the same prevalence for our cohort (Parnetti et al., 2019). Therefore, our findings should be replicated in an independent cohort including healthy controls without signs of AD pathology. Third, even though we found no statistically significant influence of antidepressant use (p = 0.25) and AChE inhibitor use (p = 0.21) on the relative abundance of genera, taxonomic and functional differences between the AD patients and healthy controls may have been influenced by drug exposure. Furthermore, we measured no metabolites in the collected gut samples.

In conclusion, we showed that intestinal microbiome can be used to discriminate amyloid-positive AD patients from healthy controls. The diagnostic accuracy increases from taxonomic to functional data and is even better when combining taxonomic, functional and clinical models. Intestinal microbiome represents a promising area for developing novel interventions against AD.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below: https://www.ebi.ac.uk/ena, PRJEB47976.

The studies involving human participants were reviewed and approved by Institut für Ethik und Geschichte der Medizin, Gartenstr. 47, 72074 Tübingen. The patients/participants provided their written informed consent to participate in this study.

CL and MW participated in study concept and design. CL, SM, OP, VR, MM, SP, US, and MW participated in the acquisition, analysis, and interpretation of data, and in the critical revision of the manuscript. CL drafted the manuscript. US did the statistical analysis. All authors read and approved the submitted version of the manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

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