Baseline characteristics of participants and clinical outcomes
One hundred eligible individuals with ND-T2D (54 male participants and 46 female participants) were randomly assigned to the acarbose or vildagliptin arm in a 1:1 ratio. Ninety-two participants completed the 6 month trial, including 48 participants in the acarbose arm and 44 participants in the vildagliptin arm (ESM Fig. 1). In addition, no serious drug-related adverse events were reported. At baseline, no significant differences were found in age and sex distributions between the two arms (p>0.05; Table 1). Baseline levels of clinical variables, including diabetes (HbA1c, fasting plasma glucose [FPG], PPG, fasting insulin [Fins], postprandial insulin [Pins] and HOMA-IR) and obesity (BW, BMI, and VFA and SFA at the L2-L3 and L4-L5 interspaces [L2-L3 VFA, L4-L5 VFA, L2-L3 SFA and L4-L5 SFA, respectively]) variables, blood lipids, gut hormones and adipokines, were also well balanced between groups (ANCOVA, p>0.05; Table 1).
After 6 months of treatment, HbA1c levels in the acarbose (60 vs 46 mmol/mol [7.65% vs 6.40%]) and vildagliptin groups (59 vs 44 mmol/mol [7.55% vs 6.20%]) were lowered to similar levels and both reached the recommended target  (HbA1c < 53 mmol/mol [7%]) (Wilcoxon signed-rank test, adjusted p<0.05; Fig. 1a). Both drugs also significantly improved FPG and PPG (Table 1, Fig. 1b,c), as well as VFAs (L2-L3 VFA and L4-L5 VFA) (Fig. 1i,j). Participants treated with acarbose had significant reductions in Pins, HOMA-IR and BW (adjusted p<0.05; Fig. 1d–g), and the latter two variables were moderately improved by vildagliptin (p<0.05 and adjusted p>0.05; Table 1). The Pins-lowering effect of acarbose was also significantly superior to that of vildagliptin (ANCOVA, adjusted p<0.05; Table 1). Additionally, the changes over time in insulin and glycaemic variables were more highly correlated in the acarbose (p<0.05) than in the vildagliptin group (Spearman’s rank analysis; ESM Fig. 2a, b). Conversely, neither of the two drugs significantly affected Fins levels, blood lipids or blood pressures (adjusted p>0.05; Table 1).
When investigating gut hormones and adipokines, we showed that both drugs significantly increased serum cholecystokinin (CCK) levels (adjusted p<0.05; Fig. 1m), while 6 month vildagliptin specifically increased fasting active GLP-1 levels and acarbose specifically reduced fasting leptin levels (Fig. 1n,o). Neither of the drugs exhibited significant impacts on fasting levels of adiponectin or two gut hormones involved in appetite regulation, ghrelin and peptide YY (PYY) (adjusted p>0.05; Fig. 1p–r) [27, 28]. Despite similar glycaemic efficacy, our results suggest that the two drugs could benefit metabolic variables, gut hormones and adipokines in different ways.
Responses of human gut microbiota to different GLDs
To better understand the impacts of different types of GLDs on human gut microbiota and the relationships between GLDs, microbiota and drug actions, we performed subsequent analyses on faecal samples from the VISA-T2D study (6 month acarbose or vildagliptin) and the previous study (3 month acarbose or glipizide)  using the same pipelines (see ESM Methods).
No significant inter-group differences were observed in alpha diversity, beta diversity and species abundances for baseline samples in the current study (Wilcoxon rank-sum test, p>0.05; Fig. 2a,c) (ESM Table 2). Six months of acarbose but not vildagliptin led to significant decreases in microbial alpha diversity of type 2 diabetes patients (Wilcoxon signed-rank test, p<0.05; Fig. 2a). Acarbose also induced significant changes in the overall gut microbial composition (PERMANOVA for Bray–Curtis distance, p<0.05; Fig. 2b,c, ESM Fig. 3). Conversely, vildagliptin treatment showed no statistically significant impacts on microbial alpha or beta diversity (p>0.05; Fig. 2a–c, ESM Fig. 3). At the taxonomic level, we identified that the RAs of 76 and ten species were altered significantly by 6 month acarbose and vildagliptin monotherapy, respectively (adjusted p<0.05; Fig. 2d,e, ESM Tables 3, 4). At the functional level, acarbose significantly altered the RAs of 115 pathways (adjusted p<0.05) and vildagliptin only had moderate impacts on 51 pathways (p<0.05 and adjusted p>0.05) (ESM Tables 3, 4). In addition, acarbose induced more considerable changes in the structure of species–species co-occurrence networks (pre vs post, correlations for hub scores of species, Spearman’s ρ=0.33) than vildagliptin (Spearman’s ρ=0.75; ESM Fig. 4a, b, and see the ESM Methods). For instance, multiple Streptococcus species (e.g. S. sanguinis and S. salivarius), the most connected gut microbial taxa in the post-acarbose treatment group, exhibited significant positive associations with Bifidobacterium longum and negative associations with Bacteroides spp. (e.g. B. caccae and B. stercoris) (an absolute value of correlation coefficient >0.3; ESM Fig. 4c–e, ESM Table 5). Despite the differences in baseline gut microbial composition between the two acarbose study cohorts (ESM Fig. 5a, b), we demonstrated that 3 or 6 months of acarbose treatment consistently induced significant changes in RAs of 47 species and 39 functional pathways (BH-adjusted p<0.05 in both groups; Fig. 2f, ESM Fig. 5c, d, ESM Tables 3, 6), including the previously reported decreases of diversity and multiple Bacteroides species and the increases of Bifidobacterium and Streptococcus members . Additionally, we also detected many species with differential abundances between acarbose and vildagliptin groups at 6 months, which largely overlapped the acarbose-induced changes (ESM Table 7). These results all supported the greater impacts of acarbose on human gut microbial diversity and ecological structures.
Notably, we found that participants receiving different drugs exhibited consistent changes in RAs of a set of gut microbial species and functional pathways. For instance, there were significant enrichments in RAs of B. adolescentis and reductions in RAs of Bacteroides plebeius, B. caccae, Bacteroides eggerthii, Bacteroides thetaiotaomicron and Paraprevotella distasonis in the gut of participants treated with either single agent (adjusted p<0.05; Fig. 2e, ESM Tables 3, 4). When considering a moderate trend toward significance (p<0.05 in both arms), we showed that 16 species were commonly reduced in the two treatment arms (ESM Fig. 6a) and all belonged to the phylum Bacteroidetes. Furthermore, most of the above species responding to 6 months of acarbose or vildagliptin were altered consistently in participants treated with 3 months of acarbose (adjusted p<0.05; ESM Fig. 6a). Among the commonly increased taxa, two Bifidobacterium members (B. adolescentis and B. longum) [29,30,31,32] and Haemophilus parainfluenzae [33, 34] have been repeatedly shown to have significantly higher RAs in healthy control groups than in type 2 diabetes patients. Both drugs also decreased the RAs of pathways involved in the biosynthesis of queuosine (PWY-6700 and PWY-6703), lipopolysaccharide (PWY-1269) and pyridoxal 5′-phosphate (PLP, PWY0-845 and PYRIDOXSYN-PWY), and increased the pathways of the mixed acid fermentation (FERMENTATION-PWY) and the biosynthesis of seleno amino acid (PWY-6936) (p<0.05 at both arms; ESM Fig. 6b, ESM Tables 3, 4). Correlation analysis on baseline RAs (ESM Fig. 7a) and the RA changes (ESM Fig. 7b) between responding species and pathways consistently revealed positive associations between several Bacteroidetes species and pathways (e.g. PWY0-845 and PYRIDOXSYN-PWY for the biosynthesis of PLP, PWY-7282: 4-amino-2-methyl-5-phosphomethylpyrimidine biosynthesis and ARGININE SYN4-PWY: l-ornithine biosynthesis) which were both reduced after treatment (adjusted p<0.05). Cumulative abundance analysis further supported that the RAs of the above-mentioned highly correlated pathways were mainly contributed by the Bacteroides species (ESM Fig. 7c).
In addition, vildagliptin specifically elevated the RAs of Clostridium bartlettii, a known Firmicutes butyrate producer , and reduced the RAs of Paraprevotella clara and Paraprevotella xylaniphila (adjusted p<0.05 in the vildagliptin group and p>0.05 in 3 month and 6 month acarbose groups; Fig. 2e,f) . We also repeated the previous finding  that glipizide, an effective GLD targeting the SU receptor on pancreatic beta cells , did not significantly alter the RAs of any gut species or pathways (Fig. 2f, ESM Fig. 6a, b, ESM Table 6). Altogether, these results suggested the existence of common and agent-specific gut microbial responses in type 2 diabetes patients receiving acarbose or vildagliptin monotherapy.
Longitudinal associations between microbial abundances and metabolic variables
Given that acarbose and vildagliptin exert their glucose-lowering effects through distinct GI mechanisms, we asked how the responding microbial variables were correlated with metabolic variables during treatments by different agents. To answer this, we performed the GEE analysis and investigated the longitudinal correlations between the responding species/pathways and clinical variables, with adjustment for age and sex (see Methods). We found significant associations between HbA1c and Bifidobacterium species (negative correlations) and a few Bacteroidetes species (positive correlations) in both groups (GEE, adjusted p<0.05; Fig. 3a). Conversely, few microbial features were correlated with obesity variables (e.g. BW, BMI and VFA) in either arm (Fig. 3a,b). There were also drug-dependent longitudinal association patterns including the PPG–microbiome associations in the acarbose group and the GLP-1–microbiome associations in the vildagliptin group (Fig. 3a,b), and most of the correlations remained significant even after adjustment for BMI and L2-L3 VFA (ESM Table 8). We also showed highly consistent association patterns between HbA1c/PPG and 47 acarbose-altered microbial species in the 3 month and 6 month cohorts (GEE, BH-adjusted p<0.05; ESM Fig. 8a, b). By contrast, no significant associations were found in the glipizide-treated type 2 diabetes patients between these species and any clinical variables (ESM Fig. 8c).
Associations between baseline gut microbiota and post-treatment GLP-1 responses
The secretion of GLP-1 could be directly improved by vildagliptin and specific gut bacterial metabolites, such as SCFAs and secondary bile acids [37, 38]. The latter raised the next important question: whether baseline gut microbiota had potential impacts on GLP-1 responses to drug treatments. To answer this, we divided 40 type 2 diabetes patients in the vildagliptin group (who had pre- and post-treatment metagenomes and fasting GLP-1 values) into two subgroups according to their GLP-1 responses (see Methods), namely the high response (HR, n=20, PC%>50.18%) and low response (LR, n=20, PC%≤50.18%) groups (Fig. 4a). Likewise, the HR group also had a greater improvement in Pins levels than the LR group (p<0.05; Fig. 4b). At baseline, the two subgroups had no significant differences in GLP-1, insulin or HOMA-IR (p>0.05; Fig. 4c), but the LR group had worse glycaemic status than the HR group (p<0.05; Fig. 4c, ESM Table 9).
We next performed sPLS-DA to investigate whether baseline microbiota (RAs of species and pathways) could effectively distinguish participants with high and low GLP-1 responses to vildagliptin. We observed a clear separation of individuals between the HR and LR subgroups in the classification model (Fig. 4d). Among the ten selected microbial variables that had the highest contribution to sPLS-DA-1, the baseline RAs of Barnesiella intestinihominis and Clostridium citroniae were enriched in the HR group while those of Veillonella parvula, Prevotella copri and all six selected pathways were enriched in the LR group (Fig. 4e, ESM Fig. 9, ESM Table 10).
Although the fasting GLP-1 did not increase significantly in the acarbose group, we observed similar correlations between the PC%-GLP-1 and baseline RAs of the ten features in the two treatment arms, including the negative associations with PWY-6588: pyruvate fermentation to acetone (p<0.05, Spearman’s rank correlation; Fig. 4f). There were also positive correlations between the PC%-GLP-1 and individual predicted probabilities from classification models for the two groups (vildagliptin: ρ=0.5, p=0.0012; acarbose, ρ=0.27, p=0.09) (Fig. 4g). These results suggested that the baseline gut microbiota, in turn, might impact the heterogeneity of GLP-1 secretory responses to GLDs among different type 2 diabetes patients.