2.1.1. Data Preprocessing

In the first step, the raw genus or species abundance counts obtained from 16S rRNA gene amplicon or shotgun metagenomics sequences are converted into relative abundance counts. The counts are then filtered to retain only those with a frequency above 0.001%. Subsequently, the data is transformed into relative abundances based on the total sum per sample. Concurrently, a factor is created from the metadata labels column (e.g., disease or control) using R, enabling the storage of metadata as levels.

2.1.2. DAA (Association Testing)

In the second step, the dependencies between the preprocessed 16S rRNA gene (or shotgun) sequence features and the metadata factors are analyzed through association testing. For this, we utilize the SIAMCAT (statistical inference of associations between microbial communities and host phenotypes) R-toolbox developed by Wirbel et al. [27], building on the earlier work by Zeller et al. [28].

SIAMCAT performs DAA across two classes using the nonparametric Wilcoxon test with multiple hypothesis correction to assess the significance of associations. This analysis yields model metrics such as the p value and the FDR adjusted p value (q value), which can be used to evaluate the importance of selected taxa. Additionally, SIAMCAT calculates a nonparametric effect size measure for features using area under the receiver operating characteristic (AU-ROC) curve scores. For our analyses, we used the check.associations function of the SIAMCAT packages with default parameters.

While numerous differential abundance testing methods exist, we selected SIAMCAT due to its robust statistical approach and its extensive adoption in the field (see [29–31]).

2.1.3. ML Methods

In parallel to the DAA, we applied three state-of-the-art ML methods to predict biomarkers: LASSO, Boruta, and LEfSe. Each method adopts a different approach for feature selection, as detailed below.

2.1.4. Biomarker Selection

2.3.1. Datasets Used for Evaluation

To evaluate our KMB consensus approach and to compare its performance with individual methods, four datasets were used. An overview of these datasets, their sources, and the results from the original studies are summarized in Table 1. Further details about each dataset are provided below.

2.3.2. Performance Assessment

The individual models and the consensus-based KMB model were evaluated based on their ability to accurately predict biomarkers (or biomarker signatures) identified in the original studies and corroborated by other literature. Additionally, we evaluated the statistical significance association with the predictions using the metrics employed by the individual methods.

For the DAA with SIAMCAT, the significance of the log fold changes was assessed using p values and q values (or adjusted p values). These values were derived from two-tailed Z-tests, as described by Cardoso et al. [63] and Wirbel et al. [27]. A significance threshold of p ≤ 0.05 and q ≤ 0.05 was applied. For the LR method (which includes LASSO and ridge regression), we used the median relative feature weights as a performance metric. For the RF model (included in Boruta), we used Gini impurity scores to assess the feature performance. Note that there is no fixed threshold or acceptance criteria for Gini impurity scores; hence, features were ranked in decreasing order of Gini impurity scores. For the LEfSe method, we used LDA scores to assign the importance of taxa (feature) in discriminating between groups, and p values to assess the significance. Note that we used the default settings for the p value cut-off (≤ 0.05) and log LDA score (≥ 2.0).

The overall performance characteristics of the LR (LASSO/elastic net) and RF (Boruta) models were further assessed using the AU-ROC and the area under the precision recall curve (AU-PRC) plots. ROC curves were obtained by plotting the true positive rate (TPR—also known as sensitivity) against the false positive rate (FPR—also known as specificity). While ROC is typically used for plotting the diagnostic ability of a binary classifier, it can also be applied to a multiclass classification problem as found in biological datasets with the help of pairwise comparison analysis. As there is no generic optimal AU-ROC threshold value, we used 0.75 as a cut-off, which is in between 0.5 (random performance of the model) and 1 (perfect model performance). Precision–recall (PR) curves were obtained by plotting the precision (calculated by dividing the number of true positive predictions by the total number of positive predictions) and recall (calculated by dividing the number of true positive predictions by the total number of true positives). Similarly, for PR curves there is no generic optimal threshold value; however, the closer the curve is to the upper right corner, the better the model’s performance.

2.3.3. Confounding Effects

All datasets included in this study were analyzed for potential confounding factors (e.g., patient and other metadata) by the authors of the original studies. To summarize, Kostic et al. [46] observed a minor correlation between patient geographic location and microbial diversity, likely due to the difference in sample collection methods. An association between higher tumor grades and an increased microbial diversity was also noted. Zeller et al. [28] indicated that patient age significantly differed between control and CRC cases. Ling et al. [44] analyzed different clinical indicators and found that the mini mental state examination (MMSE), Weschler adult intelligence scale (WAIS), and Barthel scores were significantly lower in AD patients compared to controls. Finally, Qian et al. [48] found that only PD status influenced the PD Index in their cohort.

Since the potential effect of confounding factors was thoroughly examined by the original authors, no additional confounder analysis was performed in this study.

3.1.1. Abundance and Diversity Analysis

In line with the findings of Kostic et al. [46], the phyla Bacteroidetes and Firmicutes were significantly more abundant in healthy control samples, whereas the relative abundance of the phylum Fusobacteria was enriched in tumor samples (Figure 2a). Next, we assessed microbial richness in healthy controls and CRC patients using three alpha diversity indices: observed OTUs, Shannon, and Simpson. While Kostic et al. did not find significant differences using cumulative OTU counts and the Chao1 diversity index, our evaluation consistently indicated a significant reduction (p value ≤ 0.05) in microbial diversity in CRC patient samples across all three indices (Figure 2b). Importantly, two other major studies that investigated the association between CRC and gut microbiome [64, 65] also reported decreased bacterial diversity in tumor samples, using the Shannon diversity and evenness indices.

3.1.2. Performance of the KMB Framework

We further performed biomarker predictions on the Kostic et al. dataset using our KMB framework. In total, 13 genera were detected by KMB, 10 of which were also identified by Kostic et al. using LEfSe for their analysis (LDA significance threshold of 4.2) (Table 2). Notably, our KMB framework confirmed the enriched presence of Fusobacterium in CRC-associated metagenomes, a finding validated by quantitative PCR and metagenomic shotgun analyses and corroborated by subsequent studies (particularly involving Fusobacterium nucleatum). This genus ranked among the top three biomarkers in our analysis (LEfSe p = 5.57e − 05; DAA q = 0.00248). Other genera, identified by both KMB and Kostic et al. and frequently associated with CRC in the literature, include Faecalibacterium, Bacteroides, and Alistipes. Our analysis also predicted a significant enrichment of Eubacterium and Campylobacter, which were not reported by Kostic et al. Both genera had significant predictions via LEfSe (p < 0.01) and DAA (q ≤ 0.05) and at least one additional method (LASSO and Boruta). The role of Eubacterium in the proliferation of CRC has drawn considerable attention due to its production of butyrate, a short-chain fatty acid (SCFA) shown to lower tumor progression and with novel therapeutic effects [66]. In line with our results, Zhang et al. [67] reported decreased levels of Eubacterium (p = 0.009) in gut microbiota of healthy subjects. Specific species such as E. hallii [65] and E. rectale (p ≤ 0.05) [68] have been identified as potential CRC “driver” bacteria. Recent in vitro and in vivo studies by Ryu et al. [69] demonstrated anti-CRC effects of E. callanderi KGMB02377, which also contains pathway genes for γ-aminobutyric acid (GABA) synthesis, thus potentially explaining the inhibition of CRC progression. Similarly, studies linked an increased abundance of Campylobacter spp. to CRC proliferation, particularly the oral pathogenic strain C. jejuni in a mouse study [70]. Importantly, C. jejuni produces cytolethal distending toxin (CDT), which has been shown to promote intestinal inflammation.

3.1.3. Performance of the Individual Biomarker Prediction Methods

At the method level, DAA and LEfSe demonstrated a strong and consistent performance agreeing on the top nine biomarkers, albeit with varying order of significance. By contrast, LASSO and Boruta predicted fewer biomarkers (four and six, respectively). Notably, LASSO’s features fell below the AU-ROC threshold (0.68 < 0.75), leading to potential incorrect predictions. For instance, LASSO detected an increased abundance of unclassified Clostridiales genus in tumor samples, whereas both LEfSe and DAA detected its enrichment in healthy controls. Boruta achieved higher accuracy (AU − ROC = 0.89 ≥ 0.75), but its top biomarkers (based on Gini index values) were less significant by LEfSe and DAA (based on p and q values, respectively).

3.2.1. Abundance and Diversity Analysis

In line with the original study by Zeller et al. [28], our KMB analysis confirmed an enrichment of phyla Fusobacteriota (synonym Fusobacteria), Pseudomonodota (synonym Proteobacteria), and Bacteroidota (synonym Bacteroidetes), alongside a depletion of phyla Actinomycetota (synonym Actinobacteria) and Bacillota (synonym Firmicutes) in CRC samples (Supporting Information 1: Figure S1A). These results partially align with Kostic et al., who reported an enrichment of Fusobacterium and depletion of Firmicutes in CRC samples but observed a depletion of the phylum Bacteroidetes, which is in coherence with Zeller et al. and own findings. Alpha diversity analysis (Supporting Information 1: Figure S1B), that is Shannon and Simpson indices, revealed a slight but nonsignificant decrease (p > 0.05) decrease in community diversity in tumor samples, which corresponds with our previous analysis of 16S rRNA taxonomic profiles. However, the observed OTUs index showed a small but significant (p ≤ 0.05) increased richness in tumor samples.

3.2.2. Performance of the KMB Framework

Next, we evaluated the performance of the KMB framework and its individual methods against the original biomarker predictions by Zeller et al. Our approach predicted a total of 26 bacterial species as differentially abundant between healthy and tumor groups (see Table 3). Of these, 15 species were also part of the consensus signature proposed by Zeller et al., which consisted of 22 marker species collectively associated with CRC. Notably, the KMB framework accurately predicted the enriched presence of Fusobacterium nucleatum in tumor samples. The role of this periodontal pathogen in CRC progression has been extensively studied in recent years leading to improved insights into its functional mechanisms. For instance, Zhu et al. [71] recently performed experiments in mice and human subjects, demonstrating the increased ability of F. nucleatum to invade tumor cells and to bind to a specific tumor-expressed protein, DHX15. These findings further support the hypothesis of a distinct interaction between host genotype and gut microbiome in CRC. Furthermore, several bacterial species consistently associated with CRC tumors were confirmed by our KMB framework and are supported by epidemiological evidence as well as mouse model experiments (see [72] for an overview). Amongst these species were Porphyromonas asaccharolytica [73], Peptostreptococcus stomatis [74], and Bacteroides fragilis [75], all of which were predicted by KMB and were included in Zeller et al.’s consensus signature. Conversely Parvimonas micra, known to promote CRC tumorigenesis by upregulating cell-associated cytokines [76], was predicted by the KMB among the biomarkers but was not part of Zeller et al.’s consensus signature. However, Zeller et al. did report a significantly increased abundance of P. micra in CRC patients compared to controls (p = 2.06e − 02)).

3.2.3. Contributions of the Individual Biomarker Prediction Methods to the KMB Model

Regarding the contribution of individual biomarker prediction methods to the consensus model, all 26 predictions were supported by the DAA method. The LASSO method, however, demonstrated the highest overlap with Zeller et al.’s findings, with 12 out of 18 predictions aligning with their results. While this consistency with Zeller et al. is partly expected, given that they used LASSO as well (albeit with a different procedure), it is worth noting that the LASSO model in the KMB framework exceeded the set AU-ROC threshold value (AU-ROC = 0.81 ≥ 0.75). The model’s high performance is also reflected by the high AU-PRC score (Figure 3). These results highlight LASSO’s significant contribution to our consensus approach on the shotgun-based taxonomic profiles. Notably, LASSO and DAA were the only methods to predict the enriched presence of F. nucleatum and B. fragilis in CRC samples. Additionally, both methods identified significant shifts in four other species (Porphyromonas uenonis, Campylobacter rectus, Campylobacter gracilis, and Leptotrichia hofstadii) that were neither reported by the Boruta and LEfSe methods nor were part of the consensus signature of Zeller et al. However, Zeller et al. reported significant difference in abundances (p < 0.01) of L. hofstadii and C. rectus between CRC and healthy individuals.

Interestingly, P. uenonis was found to be significantly decreased in CRC in the recent study by Zhang et al. [77], which analyzed 705 fecal samples across six metagenomic sequencing cohorts with diverse geographical and ethnic backgrounds. Additionally, a meta-transcriptome study by Warren et al. [78] reported coenrichment of Fusobacterium, Lepotrichia, and Campylobacter spp. in CRC tissues. At the species level, the largest number of uniquely mapping sequences corresponded to F. nucleatum, L. hofstadii and C. showae (phenotypically similar to C. rectus). The authors did not find unique sequence matches to C. jejuni, which was also absent in KMB’s predictions.

The LEfSe method demonstrated lower consistency with DAA on shotgun-based profiles compared to 16S rRNA gene-based profiles of the Kostic et al. study. LEfSe detected only 12 biomarkers, 6 of which were found by Zeller et al.’s study. However, 5 of the top 6 LEfSe’s predicted biomarkers (LDA value > 3) corresponded to Zeller et al.’s biomarkers. Similarly, the Boruta method contributed significantly with 8 out of the Top 10 OTUs included in the final KMB selection, 5 of which were also found in Zeller et al.’s study. Unlike LASSO, Boruta accurately predicted the depletion of E. rectale and the enrichment of P. micra in CRC samples. Importantly, the accuracy of Boruta model exceeded the set AU-ROC threshold value (AU-ROC = 0.91 ≥ 0.75) and the high AU-PRC indicates a strong model performance (Figure 3b). These findings highlight the robustness of the method ensemble implemented in the KMB framework to accurately predict microbial signatures associated with disease.

3.3.1. Abundance and Diversity Analysis

At the phylum level, our taxonomic classifications (abundances) mirrored those reported by Ling et al. [44]. Of the 10 phyla classified by the authors (using a minimum sequence occurrence threshold of > 0.005%), 8 were consistent with the Top 10 most abundant phyla classified in our analysis (Supporting Information 2: Figure S2A). However, direct comparison of exact percentages was not possible due to absence of numerical data in the Ling et al. study. The most abundant phyla were Firmicutes, Proteobacteria, Actinobacteria, Bacteroidetes and Verrucomicrobiae, which aligns with earlier findings by Zhuang et al. [79].

Our analysis additionally detected a significant presence of the phylum Euryarchaeota, which was not found in the Ling et al. study due to the removal of sequences identified as archaea. Similarly, the phylum Streptophyta was absent from the top 10 list of most of Ling et al. study, although our analysis reported its presence at very low abundance. Other phyla such as Candidates, Saccharibacteria and Lentisphaerae were classified as the 11th and 12th most abundant phylum in our study, whereas they were ranked 8th and 10th in the Ling et al. study. At the genus level, our classifications also largely corresponded with those reported by Ling et al., confirming the high inter-individual variation observed by the authors (Figure 4a). While the relative proportions of bacterial genera could not be directly compared (see above), 9 out of the Top 10 most abundant genera in our analysis were included in the top 19 most abundant genera reported by Ling et al. Additionally, our analysis found a high relative abundance of the genus Gemmiger. This is likely attributed to the difference in the reference databases used: Ling et al. relied on the Greengenes database [80], while our classifications relied on the RDP database. The Greengenes database has not been updated since 2013, and recent research by Campos et al. [81] on chicken cecal microbiota composition highlighted potential outdatedness issue with the Greengenes classification of Faecalibacterium, specifically of F. prausnitzii which was reclassified as distinct from the closely related Gemmiger/Subdoligranulum cluster [82]. Consequently, Campos et al. reported that sequences classified as Faecalibacterium using Greengenes were instead classified as Gemmiger with the RDP database. Interestingly, Ling et al. classified Subdoligranulum amongst their Top 19 abundant genera but also identified genus Gemmiger as a biomarker using LEfSe. This finding is unsurprising though, as LEfSe relies on the RDP taxonomy. Overall, these results underscore the significant impact a database can have on analysis outcomes.

We also assessed species richness using three alpha diversity indices—observed OTU, Shannon, and Simpson—which were also employed by Ling et al. In our analysis, all three indices indicated a reduction in bacterial diversity among AD patients (Supporting Information 2: Figure S2B), corroborating the study results of Ling et al. who also reported higher diversity in control samples. However, none of the indices in our analysis was statistically significantly (p > 0.05), whereas Ling et al. reported significant differences for all three indices (p ≤ 0.05). This discrepancy aligns with a recent review by Zhu et al. [83], which evaluated 14 AD case-control studies, and found that microbial diversity changes were often not significant. To illustrate, eight studies using the Shannon index reported no significant differences between groups, while four studies reported a significant decrease in diversity among AD patients. The beta diversity, in our analysis and in that of Ling et al., identified a significant variability in community composition between controls and AD patients (Figure 4b). This finding aligns with the broader consensus; 12 out of 14 studies in the Zhu et al. review also reported significant differences in beta diversity between these groups.

3.3.2. Performance of the KMB Framework: Enriched Microbial Genera in Healthy Controls

Subsequently, we compared the biomarker predictions of the KMB analysis framework against those of the original study. Consistent with the findings of Ling et al., a substantial number of biomarkers were associated with microbiota dysbiosis in AD (Table 4). At the genus level, KMB predicted 33 biomarkers. Of these, 11 were supported by LEfSe with LDA scores > 3, of which 9 overlapped with the predictions of Ling et al. who applied the same LDA threshold in their LEfSe analysis. Among the overlapping genera enriched in healthy controls were several butyrate-producing bacteria, such as Faecalibacterium, Roseburia, and Coprococcus, all members of the family Lachnospiraceae. Butyrate is a SCFA, produced during the fermentation of fibers and resistant starch, and garnered increasing attention for its role in gut-brain axis communication [84, 85]. Specifically, SCFAs influence two major signaling pathways: binding to G protein-coupled receptors on enteroendocrine cells and inhibition of histone deacetylases. These pathways can directly or indirectly impact effects on the central nervous system functioning. Numerous animal and human studies have demonstrated a reduced abundance of SCFA-producing bacteria in AD patients [79, 86]. Butyrate in particular has been shown to promote gastrointestinal health and to reduce inflammation by inhibiting proinflammatory cytokines [87]. Ling et al. observed a negative correlation between these three butyrate-producing genera and the levels of proinflammatory cytokine TNF-α and chemokine IP-10. The KMP framework identified Fusicatenibacter and an unclassified Lachnospiraceae as biomarkers, supported by LEfSe (LDA scores > 3), LASSO and DAA, though these were not detected in the original study. Notably, Fusicatenbacteria and other members of Lachnospiraceae are also butyrate producers, with evidence of their depleted levels in AD patients reported across multiple studies. For instance, a meta-analysis by Hung et al. [88] encompassing 11 studies (N = 805; with 427 AD patients and 378 controls from United States and Chinese cohorts) and a study by Yildirim et al. [89] (N = 127; with 47 AD patients, 27 mild cognitive impairment patients and 51 nondemended controls) observed a negative association between Fusicatenibacter and AD in a Turkish cohort. A Thai population by Wanapaisan et al. [90] reported enriched presence of Lachnospiracea (p = 0.001) and Fusicatenibacter (p = 0.0007) in healthy controls compared to AD and MCI patients though the cohort size was relatively small (N = 52).

Furthermore, KMB identified additional butyrate-producing bacteria enriched in healthy controls, supported by LEfSe, albeit lower LDA scores between 2 and 3. These include Lachnospira, Butyrivibrio, Pseudobutyvibrio, and Anaerobium (all Lachnospiraceae) and Butyriccicoccus (family Oscillospiraceae). While these genera were identified as significant biomarkers by DAA and LASSO, only Butyriccicoccus was reported in the Ling et al. study. KMB also confirmed a significant depletion of Eubacterium in AD patients (supported by DAA, LASSO, and LEfSe with LDA > 3). Notably, a study by Cattaneo et al. [91] reported a lower abundance of the anti-inflammatory species Eubacterium rectale (p < 0.001) and a reduction of anti-inflammatory cytokine IL-10 in patients with cognitive impairment and brain amyloidosis. A more recent study by Haran et al. [92], with shotgun metagenomics analysis on a cohort of N = 76 (25 AD patients and 51 controls), similarly confirmed the decrease of key butyrate-producing bacteria (two members of Butyrivibrio and three members of Eubacterium, including E. rectale). Interestingly, in patients without dementia, the study observed an increase in butyrate-encoding enzyme genes, correlating with the induction of the anti-inflammatory P-Glycoprotein pathway. However, the positive correlation observed between the pro-inflammatory taxon Escherichia/Shigella observed by Cattaneo et al. could not be confirmed by the KMB analysis, nor by the studies of Ling et al. and Haran et al.

3.3.3. Performance of the KMB Framework: Enriched Microbial Genera in AD Patients

As regards the identification of enriched microbial genera in AD patients, the KMB framework confirmed four genera that were found identified as such by Ling et al. Two of these genera, Akkermansia and Bifidobacterium, are typically regarded as health-promoting microbes and are known butyrate-producers. The authors, however, associated the increased abundance of these genera in AD patients with their capacity to produce lactate and the SCFA propionate. The association of the genus Bifidobacterium to the AD spectrum displays conflicting results in the literature. For instance, Vogt et al. [10] in a study of N = 119 subjects (94 controls and 25 AD patients) identified Bifidobacterium as a driver of gut dysbiosis and found it significantly depleted in AD patients from the United States. Conversely, the meta-analysis by Hung et al., encompassing a combined, larger cohort, found a significantly increased levels of Bifidobacterium in AD patients (p < 0.001). As Ling et al. suggested, it is plausible that different Bifidobacterium species may be involved with different effects on AD pathology. Supporting this, Haran et al., in a shotgun metagenomics sequencing study, observed a decrease in two Bifidobacterium species (B. bifidum and B. longum) and an increase in Akkermansia muciniphila in AD patients. While these findings corroborate Ling et al.’s hypothesis, a substantially larger number of shotgun metagenomics samples would be required for definitive conclusions.

The KMB framework also confirmed the two other genera identified by Ling et al.: Enterococcus and Eggerthella. Both have been implicated by several studies as potential biomarkers for AD diagnosis. For example, Underly et al. [93] demonstrated in an in vitro experiment with rat cell cultures that Enterococcus faecalis could generate early neurofibrillary epitopes, leading to abnormal tau phosphorylation. Tau pathology, along with β-amyloid deposit, is the two major hallmarks of AD. After infecting rat cortical neuron cell cultures with E. faecalis, the authors observed a strong increase in the reactivity of monoclonal antibodies Alz-50 and CP13, which specifically target tau phosphorylation. These findings align with earlier research linking oral health and AD, and the potential role of E. faecalis in the development of chronical periodontal disease [94] and its capacity to translocate to the brain, potentially causing abscess [95]. On the contrary, a recent study by Hou et al. [96] involving a cohort of N = 77 (30 AD patients and 47 healthy controls) indicated higher abundance of Enterococcus in controls (p ≤ 0.05), attributing this to this genus’s ability to produce SCFAs. While most Enterococcus species produce acetate as the main SCFA, butyrate production is species and strain specific. For instance, strain Enterococcus durans M4-5 can produce butyrate [97], whereas Enterococcus faecalis species cannot. However, some E. faecalis strains can produce propionate in addition to acetate. Haran et al. did not find report significant variations in Enterococcus species, but reported three SCFA producing bacteria (Odoribacter splanchnicus, Eubacterium eligens, and Eubacterium rectale) as the most discriminative markers for AD. Importantly, O. splanchnicus primarily produces acetate and propionate, whereas the two Eubacterium species are primarily acetate producers and generally lack the ability to produce butyrate. The KMB framework also confirmed increased levels of Eggerthella lenta in AD patients, a finding supported by Balakrishnan et al. [47], who demonstrated that gavaging mice with E. lenta led to reduced fecal butyrate levels. These observations support Ling et al.’s hypothesis that gut dysbiosis in AD patients is characterized by a shift from butyrate producing to lactate and propionate-producing bacteria, potentially extending to acetate-producing bacteria. However, since the studies of Ling et al. and Hou et al., as well as most other AD gut microbiome studies, were based on 16S rRNA gene V3-V4 sequencing data, species level resolution remained unattainable. Finally, the KMB analysis also identified a significant increase in Erysipelatoclostridium in AD patients, a genus not detected by Ling et al. This genus has recently been proposed as a biomarker for AD by Xi et al. [98], though with contrasting findings (i.e., a significant increase in healthy controls). The genus is hypothesized to influence cytokine IFN-γ levels, potentially impacting AD pathogenesis. While there is limited evidence linking Erysipelatoclostridium dysbiosis to AD, some strains of species E. ramosum (previously Clostridium ramosum) are known to cause invasive infections, particularly in immunocompromised patients [99].

3.3.4. Summarizing Results on the AD Dataset

In summary, the KMB framework confirmed several biomarkers identified by Ling et al., demonstrating an analogous shift from butyrate producers to lactate, propionate as well as acetate producers in AD samples. The KMB framework also identified some interesting novel genera as potential biomarkers, many of which have been reported in other studies, albeit with contrasting findings. These discrepancies are likely to stem from species- and strain-specific characteristics in metabolite-production, underscoring the need to further investigate using shotgun metagenomics sequencing data and larger cohorts.

3.4.1. Abundance and Diversity Analysis

At the phylum level, our abundance analysis confirmed the findings of Qian et al. [48] who identified Bacteroidetes, Firmicutes, Proteobacteria, and Actinobacteria as the most abundant bacterial phyla in both the control and disease groups (Supporting Information 3: Figure S3A). Moreover, healthy controls were characterized by a higher abundance of the phylum Bacteroides and lower abundances of phyla Firmicutes, Proteobacteria, and Actinobacteria compared to the PD group, in line with the findings of Qian et al. Furthermore, we observed a high abundance of the viral phylum Uroviricota, with an increased level in the PD group. The original study by Qian et al. also identified an enrichment of viruses in PD patients, although the authors could not assign a more specific taxonomy and only reported an unclassified/no name phylum for viruses. This could be related to the fact that the authors used MetaPhlan ([58]) to assign taxonomies based on a specific set of marker genes identified from whole genomes. Until recently, these genomes included only a limited number of viruses (approximately 3500 viral genomes for MetaPhlan2.0 and MetaPhlan3.0). In contrast, our analysis used Kraken2-Bracken software for taxonomic classification, combined with a more recent export of whole genome sequences from the NCBI RefSeq database including 11,562 viral genomes.

We then measured alpha diversity using the observed OTU richness, Shannon, and Simpson indices. All three indices indicated increased species diversity in PD samples, with significant p values for the observed OTUs and Shannon indices (p = 0.00586 and p = 0.00447, respectively, both ≤ 0.05; Supporting Information 3: Figure S3B). Additionally, Qian et al. found a significantly increased Shannon index in PD patients (p = 0.0084) based on a profile of 1,118,355 gut microbial genes. These findings are consistent with the current consensus in the field. In a recent review by Tan et al. [49], which compared the results of 30 independent case-control studies, 14 studies reported no significant difference, while nine studies observed an increase in the diversity in PD patients. Six other studies did not report on diversity measures, and one study indicated a decreased diversity in PD patients.

3.4.2. Performance of the KMB Framework: Comparison With the Original Study

Next, we predicted microbial markers using our KMB framework and compared the results to the original findings by Qian et al. In total, 22 species were identified by KMB, all of which were enriched in PD patients (Figure 4). Of these, 10 were supported by LASSO, although the model quality was just below the set AU-ROC threshold value (AU-ROC = 0.74 < 0.75). In contrast, the accuracy of the Boruta model was above the threshold value (AU-ROC = 0.89 ≥ 0.75), yet none of the 10 most contributing species predicted by Boruta was found to be significant by another method. As such, the Boruta markers were not included in the KMB list. Furthermore, 15 biomarkers were supported by significant LEfSe predictions, of which 4 had an LDA score > 3. Surprisingly, only one species identified by KMB (Gordonibacter pamelaeae) matched with the 12 biomarkers identified by Qian et al. Additionally, KMB identified a second species from the same genus, G. urolithinfaciens, as a significant biomarker. Both species were supported by DAA and LASSO, although the features of the latter model were just below the significance level. While the specific role of Gordonibacter in the development of PD is not well understood, some negative correlations with PD have been suggested based on the capacity of G. pamelaeae (strain DSM 19378^T) and G. urolithinfaciens (strain DSM 27213^T) to produce intermediary urolithins. Urolithins are anti-inflammatory molecules produced by specific gut microbes upon the intake of dietary polyphenols [100] and may possibly enhance gut-barrier integrity. Consequently, urolithin-producing metabotypes are generally considered health-promoting bacteria. Moreover, there is some evidence for a neuroprotective role of urolithins, particularly Urolithin A, though the mechanisms of action have mostly been shown in rodent-based research [101]. Nonetheless, Romo-Vaquero [102], who recently investigated the association between urolithin metabotypes, gut dysbiosis, and disease severity in PD patients (N = 169, 52 PD and 117 HC), did not find a significant difference in Gordonibacter abundance between HC and PD. The genus was found to be decreased in Severe PD compared to Mild PD, although at low significance (Gordonibacter had the lowest LDA score of the genera enriched in PD). In contrast, Lubomski et al. [103] found decreased levels of Gordonibacter in PD patients (N = 128, 74 PD and 74 HC) 12 months after initiation of levodopa–carbidopa intestinal gel (LCIG) therapy. Although the number of LCIG patients was small (N = 10), this finding aligns with earlier studies that detected associations between gut dysbiosis in PD and the use of levodopa–carbidopa, the primary medication to treat PD. Importantly, Maini Rekdal et al. [104] and Kessel et al. [105] demonstrated that bacteria harboring tyrosine decarboxylase genes, particularly Lactobacillus and Enterococcus faecalis, decreased the efficacy of levodopa–carbidopa. It may, therefore, be hypothesized that species of the genus Gordonibacter could also interact with levodopa metabolism. To support this hypothesis, Maini Rekdal and collaborators found that a specific strain from the family Eggerthellaceae, Eggerthella lenta A2, is involved in the reduced efficacy of levodopa–carbidopa. The authors showed that E. lenta A2 harbors a molybdenum-dependent dehydroxylase (Dadh) which catalyzes the conversion of dopamine to m-tyramine. Notably, Gordonibacter sp. is part of the same family, and, from a BLASTP search of close family relatives, Maini Rekdal et al. found a homologous dopamine dehydroxylase in two unclassified Gordonibacter strains, An232A and An230, with amino acid identities of 94% and 93%, respectively. However, the same study indicated that two other strains (Gordonibacter sp. 28C and Gordonibacter pamelaeae 3A) were unable to convert dopamine to m-tyramine. Clearly, deeper investigations are required to understand the association between decreased levels of Gordonibacter, PD, and levodopa–carbidopa administration. Nonetheless, current evidence seems to favor a possible role of molybdenum-dependent dehydroxylases, rather than a beneficial effect due to the capacity of Gordonibacter sp. to produce urolithin. Additionally, the KMB framework identified another member of the family Eggerthellaceae, Slackia isoflavoniconvertens, as significantly enriched in PD patients. This strain encodes several molybdopterin-oxidoreductases (UniProt accessions A0A3N0IGL9, A0A3N0IK48, and A0A3N0I7E1), which belong to the same family as Dadh (molybdopterin-containing oxidoreductases). Finally, Qian et al. found a negative correlation between Streptococcus salivarius and levodopa equivalent dose, along with an overall decreased presence in PD patients. This association could, however, not be reproduced by the KMB framework, and this species was not referred in the studies of Maini Rekdal et al. [104] and Kessel et al. [105].

3.4.3. Performance of the KMB Framework: Comparison With Other Shotgun Metagenomics Studies

Next, we extended our comparison to eight other studies included in the review by Tan et al. [49], which investigated dysbiosis at the species level. Of these, only the study by Bedarf et al. [50] used shotgun metagenomics sequencing, while the other seven studies [51, 52, 106–110] based their analysis on 16S rRNA gene data, albeit at a higher resolution. As shown in Table 5, the species Akkermansia muciniphila and Megasphaera elsdenii were confirmed by one or more other studies reviewed by Tan et al. [49]. Notably, these species exhibit low fold changes and are supported by both DAA and LEfSe within the KMB framework (with LDA scores > 3). The detected enrichment of Akkermansia muciniphila in PD is consistent with findings from four other biomarker studies [50–52, 110]. At the genus level, Akkermansia is also largely associated with PD, as illustrated by Tan et al. [49], who noted that 14 studies reported same result. Although Qian et al. did not identify Akkermansia spp. (nor A. muciniphila) as a significant microbial biomarker based on MetaPhlAn taxonomic analysis, they did find two MGS clusters annotated to the genus Akkermansia and one specifically to the strain A. muciniphila ATCC BAA-835. In summary, Akkermansia spp. and particularly A. muciniphila are considered among the most important biomarkers for PD diagnosis. However, as previously noted, this finding is quite remarkable given that Akkermansia is generally considered a health-promoting bacterium. Particularly, A. muciniphila promotes the integrity of the gut barrier (mucin layer) and is involved in immune response modulation [111, 112]. The beneficial effect of A. muciniphila has been also demonstrated in metabolic disorders such as obesity [113] and diabetes Type 2 [114]. At the same time, increased abundance of Akkermansia spp. has been reported in several neurological diseases, including multiple sclerosis [115] and AD (see above). These findings suggest the potential existence of a common mechanism underlying the progression of various neurological diseases. To support this hypothesis, Duvallet et al. [13] showed that many genera are typically associated with multiple diseases, indicating that involvement of bacteria is often not disease specific. Although several studies link the increased presence of Akkermansia to a modified immune response (Zhai et al. [116]) and constipation (Vandeputte et al. [117]), more investigations are required to understand the exact role of the different Akkermansia (sub)species in the progression of PD and other neurological disorders. Moreover, caution must be taken when interpreting these results, as confounding factors such as drug treatment and stool frequency can introduce specific microbial signatures [118].

Megasphaera elsdenii, which was identified by KMB and Tan et al. [110] as a potential biomarker, is also a commensal human gut microbe, although typically found in low abundance. This species is known to ferment lactate into SFCAs, including butyrate (Shetty et al. [119]), and with this capability, the strain has been used as a probiotic to prevent ruminal acidosis [120]. Another important characteristic of M. elsdenii is its ability to produce high levels of hexanoic acid [121]. Notably, a recent study by Abdik and Çakır [122] predicted hexanoic acid as a candidate metabolite biomarker for PD. Specifically, the authors aligned 13 postmortem PD transcriptome datasets from the substantia nigra (the brain region most affected by PD) against the Human-GEM metabolism model using the popular TIMBR (transcriptionally inferred metabolic biomarker response) algorithm, originally developed by Blais et al. [123]. Based on a cohort of N = 263 (141 PD and 112 HC), the authors found increased production of hexanoic acid (hexanoylcarnitine) in more than 75% of the comparisons performed. In addition to M. elsdenii, the genus Megasphaera spp. also includes other hexanoic-acid producing species, such as M. hexanoica, which was also predicted as a biomarker by our KMB framework. The stronger association of M. elsdenii with PD compared to M. hexanoica, as indicated by the lower p and q values in LEfSe and DAA analyses, may be explained by the fact M. elsdenii produces higher amounts of hexanoic acid and can utilize a broader variety of carbon sources (e.g., sucrose, glucose, maltose, and fructose) than M. hexanoica, which primarily ferments fructose. In contrast, a recent study by Ren et al. [108], included in the review by Tan et al. [49], identified increased levels of the species Megasphaera micronuciformis in healthy controls. However, the LDA score with LEfSe was notably low (close to 2). Moreover, M. micronuciformis is incapable of producing acids from carbon sources, nor can it ferment the nonconventional carbon gluconate [124], a typical characteristic of all Megasphaera species, with the production of gas.

Of the 17 other species that were identified by the KMB framework, six belonged to the family Oscillospiraceae (Vescimonas coprocola, Bittarella massiliensis, and Angelakisella massiliensis and three Pseudoflavonifractor species: P. gallinarum, P. phocaeensis, and P. capillosus). All of these biomarkers were supported by significant DAA and LEfSe predictions. Notably, a meta-analysis study conducted by Romano et al. [125] reanalyzed 10 available 16S rRNA gene-based gut microbiome datasets and found significant increase of various Oscillospiraceae genera and species in PD patients. While relatively little is known about the role of these species in the human gut microbiome, partly due to difficulty to cultivate these bacteria, substantial evidence suggests a link between low BMI and constipation [126], both key indicators in PD. The biomarker identified with the highest LEfSe LDA score was Vescimonas coprocola (LDA = 3.742), a species recently isolated from the human gut [127]. Unsurprisingly, the literature provides limited information about its possible associations with PD, although one possible link might be an increased plasmid abundance in PD patients [128]. The Pseudoflavonifractor genus was found to be elevated in PD patients undergoing LCIG therapy [125] based on a small cohort of N = 31 (21 PD and 10 HC). Despite the small sample size, this study finding suggests a possible link between Pseudoflavonifractor spp. and levodopa metabolism, as previously observed within the members of Eggerthellaceae family. For example, the genome of P. capillosus ARCC 29799 encodes a molybdopterin oxidoreductase enzyme (UniProt accession A6NTH0), which shows a strong sequence similarity to Dadh and is classified as part of the molybdopterin-containing oxidoreductase family. Regarding the other species identified by the KMB framework, while the studies reviewed by Tan et al. [49] did not report specific evidence of their involvement in PD-related gut microbial dysbiosis, the families they belong to (Streptococcaceae, Clostridiaceae, Lactobacillaceae and Desulfovibrionaceae) were found significantly increased in PD patients. An exception was Pusillimonas faecalis, for which an opposite trend was observed by Vascellari et al. [109], who reported depleted levels of the family Alcaligenaceae in PD patients. Of note, P. faecalis was very recently isolated from human feces [127]. For the remaining families, Pseudonocardiaceae, Paenibacillaceae, Acidominococcaceae, Lactobacillaceae, and Eubacteriales incertae sedis, Tan et al. [49] found no literature evidence supporting decrease or increase abundances in PD patients compared to controls. However, as remarked earlier, gut dysbiosis is often characterized by changes of very specific species and/or strains. To facilitate better understanding on the contribution of individual species and strains and the functional mechanisms underlying disease progression, analysis at higher taxonomic resolution datasets (e.g., with shotgun metagenomics) and larger cohorts for statistical rigor are, therefore, recommended.