Analysis of each taxon between controls and DLB, and controls and PD
We obtained fecal samples in 224 PD patients, 26 iRBD patients, 28 DLB patients, and 147 controls. The numbers of PD patients at Hoehn & Yahr stages 1 to 5 with or without dementia are indicated in Supplementary Table 1. The collation of the demographic and clinical features between (i) controls and DLB, (ii) controls and PD, and (iii) controls and iRBD is indicated in Table 1. Five to six features out of the seven collated features were statistically different in either DLB, PD, or iRBD compared to controls. Next, we examined taxonomic differences between controls and DLB using Analysis of Compositions of Microbiomes with Bias Correction (ANCOM-BC), which examines taxonomic differences in two groups41, and Wilcoxon rank sum test (Supplementary Table 2a at the genus level and 2b at the family level). In ANCOM-BC, five confounding factors (age, sex, BMI, constipation, and PPI) were included in the analysis. In DLB, at the genus level, three genera were increased (Collinsella, Eggerthella, and Ruminococcus torques) and seven genera were decreased (Agathobacter, Lachnospiraceae ND3007 group, Butyricicoccus, Coprococcus, Faecalibacterium, Fusicatenibacter, and Haemophilus) after adjusting for the confounding factors (Fig. 1 and Supplementary Table 2a). In DLB, at the family level, four families were increased (Eggerthellaceae, Desulfovibrionaceae, Coriobacteriaceae, and Anaerovoracaceae) and one family was decreased (Ruminococcaceae) after adjusting for the confounding factors (Fig. 1 and Supplementary Table 2b). Nested cross-validation of random forest models to differentiate controls and DLB gave rise to the area under the receiver operating characteristic curve (AUROC) of 0.816 (95% confidence interval: 0.714–0.917) (Supplementary Fig. 1), indicating that gut bacteria were able to differentiate controls and DLB efficiently. Fifteen genera made the maximum AUROC by leave-one-out cross-validation in recursive feature elimination (Supplementary Fig. 1 and Supplementary Table 4). The three genera (Collinsella, Eggerthella, and Ruminococcus torques) that were significantly increased in ANCOM-BC and Wilcoxon rank sum test were also essential determinants in random forest models.
Table 1 Clinical and demographic features of controls, DLB, PD, and iRBD patients. Full size table
Fig. 1: Plots of ten genera and five families that were significantly changed between controls (C) and DLB (D). Medians are indicted by red bars. P-values are calculated by Wilcoxon rank sum test. Q-values by the Benjamini-Hochberg method are indicated in Supplementary Table 2a at the genus level and 2b at the family level. Full size image
We previously analyzed almost identical fecal samples in controls and PD using ANCOM42 and Wilcoxon rank sum test20. We then analyzed confounding factors in 18 genera and 5 families with generalized linear modeling (GLM)20. Here, we compared controls and PD using ANCOM-BC by simultaneously adjusting for the five confounding factors. We previously showed that eight genera (Christensenellaceae R-7 group, Ruminococcaceae_anonymous, UBA1819, Oscillibacter, Family XIII_anonymous, Alistipes, Akkermansia, and Family XIII AD3011 group) were increased in PD (see Supporting Information Fig. S2 in our previous report20), whereas only two genera, Akkermansia and Oscillibacter, which were a subset of the eight previous genera, were increased in our current analysis (Supplementary Table 3a). Similarly, we previously showed that seven genera (Fusicatenibacter, Butyricicoccus, Lachnospiraceae ND3007 group, Faecalibacteriumb, Roseburia, Blautia, and Ruminococcaceae UCG-013) were decreased in PD (see Supporting Information Fig. S2 in our previous report20), while four previous genera (Butyricicoccus, Blautia, Fusicatenibacter, and Lachnospiraceae ND3007 group) and three new genera (Coprococcus, Monoglobus, and Agathobacter) were decreased in our current analysis (Supplementary Table 3a). We previously concluded by additionally performing meta-analysis of gut microbiota in PD in five countries that PD patients had increased Akkermansia and decreased SCFA-producing genera20. The changes in these genera were indeed shared between our previous and current analyses. Nested cross-validation of random forest models to differentiate controls and PD, which were not generated in our previous report20, yielded the AUROC of 0.762 (0.714–0.810) (Supplementary Fig. 2). Twenty-five genera made the maximum AUROC by leave-one-out cross validation in recursive feature elimination (Supplementary Fig. 2 and Supplementary Table 5).
When DLB and PD were compared, five out of the seven decreased genera in DLB (Agathobacter, Lachnospiraceae ND3007 group, Butyricicoccus, Coprococcus, and Fusicatenibacter) were also decreased in PD, whereas none of the three increased genera in DLB were increased in PD. Random forest modeling showed that gut bacteria differentiated controls and PD less efficiently than controls and DLB, which was likely due to a broad spectrum of disease severities in PD compared to those in DLB.
Analysis of the overall composition of gut microbiota
We performed PERMANOVA to examine the overall composition of gut microbiota in controls and DLB (Table 2). The overall composition of gut microbiota between controls and DLB was statistically different by all three distance metrics (Table 2a). We also found that age, sex, and PPI affected the overall composition of gut microbiota (Table 2b). Donepezil and memantine, both of which were used to treat dementia, did not affect the overall composition of gut microbiota in DLB patients (Table 2c). PERMANOVA analyses between controls and PD20 and between controls and iRBD21 were performed previously using almost the same samples, and were not repeated in this communication.
Table 2 PERMANOVA to examine the effect of each factor on the overall bacterial composition. Full size table
PCoA analysis, as well as integrated topological analysis with tmap for simultaneous mapping of the overall gut microbiota, disease states, and clinical features
PCoA to examine the difference in the overall composition of gut microbiota revealed that the centers of gravity were shifted from the lower right to the upper left with the disease progression in PD, and that the center of gravity in DLB was close to those in Hoehn & Yahr stages 3 and 4 in PD (HY3&4) and PD with Mini-Mental State Examination (MMSE) < 26 (PD with cognitive decline, PDD+) (Fig. 2a). Next, we performed tmap43 to examine the relationship between taxonomic abundances, disease states, and clinical features in the same dimensions. The tmap analysis revealed that controls were closely located to SCFA-producing genera (Faecalibacterium, Coprococcus, Anaerostipes, Lachnospiraceae ND 3007 group, and Fusicatenibacter), indicating that controls were rich in SCFA-producing genera (Fig. 2b). In addition, DLB was closely located to PDD+ and HY3&4 (Fig. 2b), which was in accordance with the PCoA analysis (Fig. 2a).
Fig. 2: PCoA and tmap plots. a PCoA plot showing the centers of gravity and the standard errors of the overall compositions of gut microbiota in nine disease states. Bray-Curtis distance was used as a distance metric. b An integrated topological map, tmap, showing how close genera, disease states, and clinical features are to each other. Blue, red, and green circles indicate genera, disease states, and clinical features, respectively. The size of circles indicates the SAFE score, which represents the network-level association of a target feature and is used as an effect size. Full size image
Random forest models to differentiate DLB and HY3&4, as well as DLB and PDD+
According to PCoA and tmap, the overall composition of gut microbiota in DLB was similar to those of HY3&4 and PDD+ . In order to identify bacteria that were uniquely changed in DLB, we made random forest models to differentiate DLB (n = 28) and HY3&4 (n = 91) (including both PDD− and PDD+), as well as DLB (n = 28) and PDD+ (n = 31) (including all HY stages). The AUROC to differentiate DLB and HY3&4 was 0.756 (95% confidence interval: 0.649–0.864) (Fig. 3a) by nested cross-validation. Three genera (Ruminococcus torques, Bifidobacterium, and Collinsella) made the maximum AUROC by leave-one-out cross-validation in recursive feature elimination (Fig. 3b). The top ten genera remained in recursive feature elimination are indicated in Supplementary Table 6. We analyzed taxonomic differences between DLB and HY3&4 by ANCOM-BC and Wilcoxon rank sum test (Supplementary Table 7a). Wilcoxon rank sum test showed that Ruminococcus torques, Bifidobacterium, and Collinsella were ranked first, third, and seventh, respectively. None of the 94 analyzed genera, however, were significantly changed after being corrected for multiple comparisons.
Fig. 3: Random forest models and essential intestinal genera to differentiate DLB and Hoehn & Yahr stages 3 and 4 (HY3&4) including both PDD− and PDD+ , and plots of fecal bile acids. a ROC curves of nested cross-validation of random forest models to differentiate DLB and HY3&4 (both PDD− and PDD+). The optimal point by Youden index is indicated by a dot with the specificity and sensitivity in parentheses. b AUROCs by leave-one-out cross-validation of random forest models while genera were recursively eliminated. An arrow points to the maximum AUROC with the number of genera. The top ten genera that differentiated DLB and HY3&4 (both PDD− and PDD+), as well as the exact AUROC values, are indicated in Supplementary Table 6. c ROC curves of leave-one-out cross-validation of random forest models generated with three genera indicated by an arrow in b. The optimal point by Youden index is indicated by a dot with the specificity and sensitivity in parentheses. d, e, f Relative abundances of three genera indicated by an arrow in b in controls (n = 147), HY3&4 with MMSE ≥ 26116 (PDD−; n = 71), HY3&4 with MMSE < 26116 (PDD+ ; n = 20), and DLB (n = 28). P-values by Kruskal-Wallis test were all less than 0.05. P-values by Dunn’s post hoc test are indicated with an asterisk for p < 0.05. g Relative abundance of Ruminococcus gnavus, which also produces ursodeoxycholic acid (UDCA) from 7-ketolithocholic acid (7K-LCA), in the four categories. Although p-value by Kruskal-Wallis test was 0.40, p-values by Dunn’s post hoc test are indicated. Note that relative abundance is plotted on a logarithmic scale to clearly indicate medians and interquartile range. h Fecal UDCA/7K-LCA ratios in the four categories. P-value by Kruskal-Wallis test was 0.044. P-values by Dunn’s post hoc test are indicated with an asterisk for p < 0.05. UDCA and 7K-LCA were randomly measured in available fecal samples. (d, e, f, g, h) Median and interquartile range are indicated in red. Full size image
In contrast to a model to differentiate DLB and HY3&4, the AUROC to differentiate DLB and PDD+ was 0.603 (0.451–0.754) by nested cross-validation, which indicated that gut microbiota could not efficiently differentiate DLB and PDD+. Taxonomic differences between DLB and PDD+ by ANCOM-BC and Wilcoxon rank sum test were indicated in Supplementary Table 7b. Collinsella was the only genera that was significantly increased in DLB compared to PDD+ by ANCOM-BC.
Analysis of three genera in patients with or without cognitive decline
As shown above (Fig. 3b), three genera, Ruminococcus torques, Bifidobacterium, and Collinsella, were essential determinants to differentiate DLB and HY3&4. When relative abundances of the three genera were compared in controls (n = 147), PDD− at HY3&4 (n = 71), PDD+ at HY3&4 (n = 20), and DLB (n = 28), (i) Ruminococcus torques was increased in DLB compared to controls, (ii) Bifidobacterium was decreased in DLB compared to PDD−, and (iii) Collinsella was increased in DLB compared to controls (Fig. 3d–f). Thus, increased Ruminococcus torques, decreased Bifidobacterium, and increased Collinsella were unique to DLB.
Correlation between five clinical features and bacterial abundances in DLB
We calculated Spearman’s rank correlation coefficients between five clinical features [age, disease duration, MMSE, total Movement Disorder Society’s (MDS) version of the Unified Parkinson’s Disease Rating Scale (UPDRS), MDS-UPDRS III] and the abundances of ten genera that were significantly changed in DLB compared to controls (Supplementary Table 8). Ruminococcus torques was negatively correlated with MMSE. Eggerthella and Coprococcus were positively and negatively correlated with total MDS-UPDRS, respectively. Thus, Ruminococcus torques was likely to be increased in dementia, whereas Eggerthella was likely to be increased and Coprococcus was likely to be decreased with the progression of parkinsonism in DLB. In contrast to DLB, neither of the three genera was significantly changed in PD in our meta-analysis of five countries20.
Quantification of fecal bile acids
Three genera (Ruminococcus torques, Collinsella, and Ruminococcus gnavus), which had relative abundances of more than 0.5% in our cohort, carry 7β-hydroxysteroid dehydrogenase (7BHD) [EC 1.1.1.201] to catalyze bidirectional reactions between 7-ketolithocholic acid (7K-LCA) and ursodeoxycholic acid (UDCA) according to KEGG and UniRef90. We showed above that both Ruminococcus torques and Collinsella were high in DLB (Fig. 3d, f). Ruminococcus gnavus tended to be high in DLB and PDD+ at HY3&4 (Fig. 3g), which was similar to Ruminococcus torques. We quantified fecal UDCA and 7K-LCA concentrations, and calculated the ratio of UDCA/7K-LCA to estimate the activity of 7BHD. The UDCA/7K-LCA ratio was significantly increased in DLB compared to controls (Fig. 3h). The median of the UDCA/7K-LCA ratios was high in PDD− and PDD+ at HY3&4 compared to controls, but p-values were both greater than 0.999 (Fig. 3h). Spearman’s rank correlation coefficients between the UDCA/7K-LCA ratios and Ruminococcus torques, Collinsella, and Ruminococcus gnavus were −0.009 (p = 0.922), −0.189 (p = 0.036), and 0.396 (p < 0.0001), respectively.
Comparison of four genera (Ruminococcus torques, Bifidobacterium, Collinsella, and Ruminococcus gnavus) between controls, iRBD, PD, and DLB
We additionally plotted the four genera (Ruminococcus torques, Bifidobacterium, Collinsella, and Ruminococcus gnavus) indicated in Fig. 3 in controls, iRBD, PD, and DLB (Supplementary Fig. 3). As we observed in the comparisons between DLB and HY3&4 (PDD− and PDD+) (Fig. 3d–g), Ruminococcus torques, Collinsella, and Ruminococcus gnavus were increased in DLB, and Bifidobacterium was increased in PD, although statistical significance was not always observed. In addition, the abundances of the four genera in iRBD were similar to those in controls.