Distinct multi-omics signatures of clinical subgroups of type 2 diabetes define heterogeneous responses to an insulin sensitizer

Abstract

Type 2 diabetes (T2D) subgroups defined by clinical variables differ in disease progression and treatment response. To uncover potential molecular drivers of this heterogeneity, we performed a multi-omics analysis of 826 drug-naïve T2D patients from two phase 3 trials of the insulin sensitizer chiglitazar. Here we show that severe insulin-resistant diabetes (SIRD) is characterized by distinct miRNA profiles (e.g., miR-122-5p) correlated with liver injury, and metabolic shifts in amino acids and primary bile acids. Mild obesity-related diabetes (MOD) showed the lowest level of phenylacetylglutamine, a metabolite known to promote cardiovascular disease. Severe insulin-deficient diabetes (SIDD) exhibited high pancreas-specific miR-7-5p, while mild age-related diabetes (MARD) presented the mildest abnormalities. Finally, integrating these multi-omics signatures into machine learning models enhanced prediction of insulin sensitizer efficacy over clinical data alone. Our findings define the distinct molecular signatures of T2D subgroups, facilitating the prediction of heterogeneous treatment responses and supporting personalized clinical management.

The authors identified molecular signatures for type 2 diabetes subgroups in Chinese patients using multi-omics data. These signatures enhance drug response predictions, supporting personalized diabetes management.

Introduction

High heterogeneity of type 2 diabetes (T2D) greatly hampers its effective management^1–3. Researchers have tried to stratify patients with T2D into subgroups with different characteristics, thereby guiding subgroup-specific precision medicine¹. The most frequently replicated subgroups were established by Ahlqvist et al.⁴, which have been replicated in populations of diverse ancestries^5–10. The data-driven subgrouping strategy applied cluster analysis with six clinical variables widely used in diabetes diagnosis: glycated hemoglobin A1c (HbA1c), age at diagnosis, body mass index (BMI), homeostasis model assessment 2 estimated insulin resistance (HOMA2-IR) and beta cell function (HOMA2-B) and glutamic acid decarboxylase antibody. Four non-autoimmune T2D subgroups were identified, labeled as severe insulin-deficient diabetes (SIDD), mild age-related diabetes (MARD), mild obesity-related diabetes (MOD), and severe insulin-resistant diabetes (SIRD)⁴. The four subgroups significantly differed in clinical outcomes, including disease progression^4,11, complication risk^{4,7,8,10,12–14}, and treatment response^7,11,15. For example, the SIRD subgroup shows the highest risk of T2D complications including kidney disease^4,7,8,10 and metabolic dysfunction-associated steatotic liver disease (MASLD, previously known as nonalcoholic fatty liver disease NAFLD)¹⁰, and better outcomes of metabolic surgery¹⁵ and thiazolidinediones therapy¹¹. The MOD subgroup shows the lowest risk of cardiovascular disease (CVD)^12–14. The SIDD subgroup shows the best benefit from receiving glargine⁷, and the MARD subgroup shows the best efficacy of sulfonylureas¹¹.

The subgroup differences in clinical outcomes motivate further investigation of the underlying mechanisms. Advances in multi-omics technologies have greatly facilitated the characterization of complex diseases, providing opportunities to gain mechanistic insights into disease etiology and effective treatment^16–18. In the context of T2D subgroups, several studies, including the Innovative Medicines Initiative (IMI)-Risk Assessment and ProgreSsiOn of Diabetes (RHAPSODY) study¹⁹, the Qatar Biobank (QBB) study⁹, the Outcome Reduction With Initial Glargine Intervention (ORIGIN) study²⁰ and the German Diabetes Study (GDS)²¹, have investigated subgroup differences in circulating proteins and metabolites. Among the subgroups, SIRD exhibited the most distinct molecular signatures, characterized by aberrant circulating levels of metabolites (e.g., branched-chain amino acids BCAAs, diacylglycerol, and triacylglycerol [TG]¹⁹) and proteins (e.g., interleukin 6 [IL6]²¹) involved in insulin resistance and inflammation. The MOD subgroup showed elevated levels of leptin and enzymes involved in lipid metabolism (e.g., fatty acid binding proteins)^9,20, and the SIDD subgroup showed signatures associated with glycemic control (e.g., 1,5-anhydroglucitol⁹). In contrast, the MARD subgroup displayed the healthiest molecular profiles and was the most similar to the non-diabetic controls⁹. Unfortunately, although these signatures have been identified to distinguish T2D subgroups, the relationships between molecular signatures and clinical outcomes were lacking, limiting their application in T2D treatment.

In this study, we aimed to improve T2D treatment by systematically characterizing the molecular heterogeneity among T2D subgroups. We stratified 826 drug-naïve T2D patients undergoing phase 3 clinical trials into four T2D subgroups using the cluster approach described by Ahlqvist et al.⁴, and comprehensively investigated their miRNAome, proteome, and metabolome. We revealed distinct molecular signatures underlying the subgroups, providing mechanistic insights into the observed differences in clinical outcomes. We further demonstrated that the multi-omic signatures can help stratify patients with heterogeneous responses to antidiabetic therapies. In particular, we focused on chiglitazar, an oral PPAR pan-agonist already approved for T2D treatment^22,23 and shown to improve MASLD outcomes in phase 2 trials²⁴. Our study provides the most comprehensive multi-omics atlas of T2D subgroups to date (http://fudan-pgx.org/chihope), offering a valuable resource to guide precision treatment strategies for T2D in the future.

Results

Cohort characteristics and subgroup stratification

We studied drug-naïve patients with T2D from the Chiglitazar perturbed Human multi-Omics ProfilE (ChiHOPE) cohort, which was derived from two randomized, double-blind, phase 3 trials of the insulin sensitizer Chiglitazar (the CMAP trial, controlled by the placebo²²; and the CMAS trial, controlled by the dipeptidyl peptidase 4 [DPP4] inhibitor sitagliptin²³). The cohort was characterized male (n = 524 [63.4%]), middle-age (age: 50.8 ± 9.8 years) and obesity (BMI: 26.0 ± 3.1 kg/m²), with insufficient glycemic control (HbA1c: 70.4 ± 7.9 mmol/mol) despite adherence to strict diet and exercise regimens. A total of 826 patients were included for cluster analysis based on HbA1c, age at diagnosis, BMI, HOMA2-IR and HOMA2-B⁴. Four robust T2D subgroups were identified: SIDD (n = 268, 32%), MARD (n = 310, 38%), MOD (n = 163, 20%), and SIRD (n = 85, 10%) (Fig. 1a, b). No significant differences in sex proportion were observed across the four subgroups (Supplementary Data 1).

Fig. 1. Study design and cohort characteristics.

a Study design. b The t-distributed stochastic neighbor embedding (t-SNE) visualization of five cluster variables (i.e., HbA1c, BMI, age at diagnosis, HOMA2-B and HOMA2-IR) in a two-dimensional space. For the t-SNE plot, each dot represented one patient, and the color represented the subgroup to which the patient belongs. c–g Characteristics of the five cluster variables in patients from different T2D subgroups. In box plots, the center line indicates the median, box bounds indicate the first and third quartiles, whiskers extend to 1.5 × IQR, and points outside the whiskers are shown as outliers. The notch represents an approximate 95% confidence interval around the median, calculated as median ± 1.58 × IQR/sqrt(n). Differences among subgroups in clinical variables were identified by the KW test, and the differed subgroup pairs were determined by post hoc Dunn’s multiple comparison test. All statistical tests were two-sided unless otherwise specified. *P < 0.05, **P < 0.01, ***P < 0.001. n represents the number of patients. SIDD severe insulin-deficient diabetes (n = 268), MARD mild age-related diabetes (n = 310), MOD mild obesity-related diabetes (n = 163), SIRD severe insulin-resistant diabetes (n = 85). Source data are provided as a Source data file. a was created in BioRender. Zheng, Y. (2026) https://BioRender.com/sqhlyu7.

The four T2D subgroups showed significantly distinct clinical characteristics (Fig. 1c–g, Supplementary Data 1). Specifically, patients in the SIDD subgroup were characterized by the worst glycemic control (the highest HbA1c, fasting plasma glucose [FPG] and 2-h postprandial blood glucose [PBG]) and poor insulin secretion (low HOMA2-B index). Patients in the MARD subgroup showed the oldest age at diagnosis and the mildest metabolic derangement in glycemic control (relatively low HbA1c, FPG and PBG). Patients in the MOD subgroup showed the youngest age at diagnosis, obesity (the highest BMI and waist circumference [WC]), and elevated insulin resistance (high HOMA2-IR index). Patients in the SIRD subgroup showed severer insulin resistance (the highest HOMA2-IR index), retained insulin secretion (the highest HOMA2-B index), greater obesity (high BMI and WC), worse lipid profiles (low high-density lipoprotein-cholesterol [HDL-C] and high TG), higher liver disease biomarkers (high alanine aminotransferase [ALT] and aspartate aminotransferase [AST]), kidney disease biomarkers (high urine microalbumin creatinine ratio [UACR]) and inflammatory biomarkers (high hypersensitive C-reactive protein [hsCRP]) than other subgroups.

We constructed the genetic risk scores (GRSs) to assess subgroup differences in the genetic risk for T2D etiology. The GRSs were calculated based on previously reported genome-wide significant SNPs, summing dosages weighted by effect sizes, reflecting inherited genetic susceptibility for T2D and related pathways (beta-cell, pro-insulin, obesity, lipodystrophy and liver)^25,26. The GRS for T2D and partial GRS (pGRS) for beta-cell function and pro-insulin were the lowest in the SIRD subgroup after adjusting for sex, trial and the first five principal components (Supplementary Fig. 1a, Supplementary Data 2), with a significantly lower frequency of risk alleles of impaired beta-cell function at genes including CDC123/CAMK1D, KCNJ11, HNF1B and CDKAL1 (Supplementary Data 3). No significant differences in pGRS for other mechanistic pathways including obesity, lipodystrophy and liver functions were observed among subgroups (Supplementary Fig. 1a, Supplementary Data 2).

We also constructed the blood-derived methylation risk score (MRS)²⁷ to assess epigenetic risk of T2D among subgroups. The MRS was calculated from five reproducible CpG sites, summing standardized methylation values weighted by site-specific effect sizes. Because DNA methylation profiles derived from whole blood are strongly influenced by the relative proportions of constituent cell types, we adjusted for estimated cell composition when examining the MRS differences among subgroups to reduce potential confounding. The MRS for T2D was the lowest in the MARD subgroup after adjusting for sex, trial, estimated proportions of cell components, and the first five principal components, which was in concordance with previous research that MARD is the healthiest T2D subgroup with modest metabolic derangements^9,28. The MARD subgroup showed healthier levels of methylation sites associated with T2D risk, including hypermethylation of cg19693031 at TXNIP, hypermethylation of cg02650017 at PHOSPHO1, and hypomethylation of cg06500161 at ABCG1. No significant differences in methylation sites at SREBF1 and SOCS3 were observed among subgroups (Supplementary Fig. 1b, Supplementary Data 4).

SIRD miRNA signatures were significantly correlated with liver disease biomarkers

We compared circulating miRNA profiles among T2D subgroups by miRNA sequencing. Distinct circulating miRNA profiles were shown among diabetes subgroups, with circulating levels of 22 miRNAs significantly differed among subgroups (Kruskal–Wallis [KW] Test P_adj < 0.05). The SIRD subgroup was significantly distinguished from other subgroups in circulating levels of certain miRNA signatures (Fig. 2a), and the largest differences were observed between the SIRD and SIDD subgroups (Fig. 2b, Supplementary Data 5).

Fig. 2. SIRD miRNA signatures were strongly correlated with liver disease biomarkers.

a Differences in miRNAs between a given subgroup and all other subgroups combined. Effect sizes were estimated with Cliff’s delta. Subgroup-differed miRNAs were shown in purple, identified by P_adj < 0.05 from the KW test among four subgroups. b Pairwise miRNA differences between subgroups. Effect sizes were estimated using Cliff’s delta; blue and red indicate significantly lower and higher levels, respectively, in one subgroup versus the other. Pairwise significance was assessed using post hoc Dunn’s multiple-comparison test. c Z-scaled mean circulating levels of subgroup-differed miRNAs across four subgroups (left) and Cliff’s delta values for subgroup pairs (right). Stars show the statistical significance determined by post hoc Dunn’s test. *P_adj < 0.05, **P_adj < 0.01, ***P_adj < 0.001. Tissue-enriched miRNAs are labeled. d Sex-adjusted partial Spearman correlations (pSc) between subgroup-differed miRNAs shown in the left heatmap and clinical measurements. Significant correlations (P_adj < 0.05) were shown. Marginal histograms show the number of significant correlations for each omics signature and clinical variable, colored by correlation direction. e Violin plots show circulating levels of tissue-enriched subgroup-differed miRNAs (counts per million [Log2]) in four subgroups (SIDD: n = 220, MARD: n = 248, MOD: n = 127, SIRD: n = 78; n represents the number of patients). Embedded box plots show the median as the center line, first and third quartiles as box bounds, whiskers extending to 1.5 × IQR, and points beyond whiskers as outliers. The notch represents an approximate 95% confidence interval around the median, calculated as median ± 1.58 × IQR/sqrt(n). Stars show the statistical significance determined by post hoc Dunn’s test. *P_adj < 0.05, **P_adj < 0.01, ***P_adj < 0.001. f Barplot shows the most significantly enriched Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways (P_adj < 0.05, excluding cancer-related ones) with target genes of subgroup-differed miRNAs. g Networks show the miRNAs and their targets that are involved in the fatty acid metabolism pathway. Significance was regarded as a two-sided P value < 0.05, and type I error was further controlled by the FDR method. Source data are provided as a Source data file.

The SIRD subgroup showed distinct circulating levels of certain miRNAs, including the lowest levels of let-7b, miR-185-5p, miR-16-2-3p, miR-7-5p and miR-20a-5p, and the highest levels of miR-99/100, miR-122, miR-193a/193b, miR-885, miR-125b-2-3p, miR-505 and miR-27b-3p (Fig. 2c). The subgroup-differed miRNAs were significantly correlated with liver disease biomarkers (i.e., ALT and AST; Fig. 2d, Supplementary Data 6). Specifically, the miRNAs that were expressed at higher levels in the SIRD subgroup were positively correlated with liver disease biomarkers, while those expressed at lower levels in the SIRD subgroup were negatively correlated with liver disease biomarkers (Fig. 2d). Thus, the dysregulation of circulating miRNA profiles in the SIRD subgroup could reflect liver dysfunction.

We evaluated the tissue specificity of subgroup-enriched miRNAs using the human miRNA tissue atlas (miRNATissueAtlas2)²⁹, which includes human tissue samples across multiple organ types. Tissue specificity was quantified using the tissue specificity index (TSI), and miRNAs with a TSI greater than 0.8 were considered tissue-specific. Several subgroup-differed miRNAs were specifically expressed in metabolic tissue. For example, miR-122-5p, miR-192-5p, miR-194-5p and miR-885 were specifically expressed in the human liver. MiR-193a/193b were specifically expressed in the human adipose tissue, and in the brown adipose tissue of the mouse³⁰ (Fig. 2c). The levels of these liver- and adipose- specific miRNAs were the highest in the SIRD subgroup, indicating the dysregulation of insulin resistance in these tissues (Fig. 2e). In addition, miR-7-5p was specifically expressed in the pancreas (Fig. 2c), and β-cell-specific overexpression of miR-7a in mice has been shown to reduce insulin secretion³¹. Thus, the highest level of miR-7 observed in the SIDD subgroup could underlie dysregulated insulin secretion in the pancreas (Fig. 2e).

To investigate the biological pathways regulated by subgroup-differed miRNAs, we mapped their target genes and carried out Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses. A total of 47 non-cancer pathways were enriched with the target genes (P_adj < 0.05, Supplementary Data 7), and the most significant pathways were fatty acid biosynthesis and metabolism pathways (Fig. 2f). Several genes involved in fatty acid metabolism, including ACSL3, ACSL4, ELOVL5, ELOVL6, and SCD, were targeted by multiple subgroup-differed miRNAs, highlighting their potential regulatory differences across subgroups (Fig. 2g). These genes also showed tissue specificity according to the Human Protein Atlas (HPA) (https://www.proteinatlas.org/). For example, ELOVL6 is specifically expressed in liver, and SCD is specifically expressed in adipose tissue, brain, and liver, indicating that subgroup-specific regulation may converge on lipid metabolic pathways in a tissue-dependent manner.

SIRD and MOD proteomic signatures were involved in complement and coagulation

We compared circulating proteomic profiles among diabetes subgroups quantified by mass spectrometry (MS) technology, and measured more than 600 proteins (Supplementary Data 8). Both the SIRD and MOD subgroups were distinguished from other subgroups in circulating levels of various proteomic signatures regarding the number of differed signatures and the quantitative differences (Fig. 3a). Similar to the miRNA profiles, modest differences in protein profiles were observed between the MARD and SIDD subgroups (Fig. 3b).

Fig. 3. SIRD and MOD proteomic signatures were involved in complement and coagulation.

a Differences in proteins between a given subgroup and all other subgroups combined. Effect sizes were estimated with Cliff’s delta. Subgroup-differed proteins were shown in purple, identified by P_adj < 0.05 from the KW test among four subgroups. b Pairwise protein differences between subgroups. Effect sizes were estimated using Cliff’s delta; blue and red indicate significantly lower and higher levels, respectively, in one subgroup versus the other. Pairwise significance was assessed using post hoc Dunn’s multiple-comparison test. c Z-scaled mean values of circulating levels of subgroup-differed proteins in four subgroups (left) and Cliff’s delta values of subgroup-differed proteins between subgroup pairs (right). Stars show the statistical significance determined by post hoc Dunn’s test. *P_adj < 0.05, **P_adj < 0.01, ***P_adj < 0.001. The tissue-enriched proteins were labeled. d The sex-adjusted pSc between subgroup-differed proteins shown in the left heatmap and clinical measurements. Significant correlations (P_adj < 0.05) were shown. The histogram represents the number of significant correlations between each omic signature (rows) and each clinical variable (columns), colored by the correlation direction. e Enriched Gene Ontology (GO) terms of subgroup-differed proteins. Seven subgroup-differed proteins were enriched in six GO terms, which were divided into three groups. Each node represents a term, with the color indicating its group and the size indicating the enrichment significance of the gene set in the term. Each edge represents the intersection between two gene sets, and the thickness indicates the percentage of the intersection. f Protein–protein interaction (PPI) network of subgroup-differed proteins. Each node represents a protein. The outer ring of the node represents the GO term group to which the protein belongs. An edge between two nodes indicates an interaction between the proteins, and the thickness of the edge reflects the confidence score of the interaction. g Violin plots show circulating levels of subgroup-differed proteins (intensity [Log2]) in four subgroups (SIDD: n = 242, MARD: n = 268, MOD: n = 139, SIRD: n = 80; n represents the number of patients). Embedded box plots show the median as the center line, first and third quartiles as box bounds, whiskers extending to 1.5 × IQR, and points beyond whiskers as outliers. The notch represents an approximate 95% confidence interval around the median, calculated as median ± 1.58 × IQR/sqrt(n). Stars show the statistical significance determined by post hoc Dunn’s test. *P_adj < 0.05, **P_adj < 0.01, ***P_adj < 0.001. Significance was regarded as a two-sided P value < 0.05, and type I error was further controlled by the FDR method. Source data are provided as a Source data file.

Circulating levels of 22 proteins were significantly different among subgroups (KW Test P_adj < 0.05), including previously reported apolipoprotein D (APOD), milk fat globule EGF and factor V/VIII domain containing (MFGE8), 14-3-3 protein theta (YWHAQ), a2-macroglobulin (A2M), complement factor H (CFH), sex hormone binding globulin (SHBG) and serpin family F member 2 (SERPINF2)¹⁹. In the MOD and SIRD subgroups, APOD levels were lower¹⁹, whereas MFGE8 and CFH levels were higher¹⁹. YWHAQ exhibited the highest levels in the SIDD subgroup and the lowest in the SIRD subgroup¹⁹. SERPINF2, A2M, and SHBG levels significantly differed between SIRD and MOD subgroups, that the MOD subgroup showed the highest SERPINF2 levels and the lowest A2M and SHBG levels¹⁹.

In addition to the previously reported proteins, we identified fifteen proteins that differed significantly among the subgroups, highlighting previously unrecognized molecular heterogeneity. For instance, the SIRD and MOD subgroups showed lower levels of circulating hemoglobin subunits (HBA1 and HBZ) and APOF, while the SIDD and MARD subgroups showed higher levels of these proteins. Meanwhile, circulating levels of vitronectin (VTN), tubulin alpha 4a (TUBA4A), nucleosome assembly protein 1 like 1 (NAP1L1), transgelin (TAGLN), proteoglycan 4 (PRG4) and complement C3 were higher in the SIRD and MOD subgroups, while lower in the SIDD and MARD subgroups (Fig. 3c). Of these, eight were enriched or enhanced in the liver (Fig. 3c), indicating the hepatic dysregulation in the two subgroups. The subgroup-differed proteins were widely correlated with T2D phenotypes (Fig. 3d). Most correlations were found between PRG4 levels and T2D phenotypes. The circulating levels of PRG4 were positively correlated with blood glucose, pressure, lipid, and obesity, along with the liver and kidney injury phenotypes, indicating its critical role in T2D and its complications. A previous study has shown that Prg4 KO mice are protected from high-fat diet-induced glucose intolerance and fatty liver disease³². Thus, PRG4 appears to play a deleterious role in metabolic regulation, potentially linking extracellular matrix remodeling to insulin resistance, lipid accumulation, and inflammation.

The proteins with higher levels in the MOD and SIRD subgroups were enriched in the complement and coagulation and the blood microparticle pathways with extensive interactions, indicating dysregulation of inflammatory processes (Fig. 3e, f). For example, A2M, APOE, C3, CFH, HBA1, SERPINF2, and VTN were involved in the blood microparticle pathway. A2M, C3, and CFH were also implicated in the regulation of complement activation and humoral immune response. APOE, SERPINF2, and VTN were involved in the negative regulation of blood coagulation and hemostasis (Fig. 3f, g). These findings are consistent with prior reports that T2D, particularly insulin-resistant and obesity-related subtypes, exhibit heightened complement activity, coagulation abnormalities, and microparticle-mediated vascular inflammation^33–35.

Diabetes subgroups significantly differed in amino acid metabolism

We quantified 203 circulating metabolites with targeted metabolomic technology, mainly consisting of fatty acids (n = 50 [25%]), amino acids (n = 44 [22%]), and organic acids (n = 30 [15%]). Circulating levels of 61 metabolites were significantly and widely different among subgroups (KW Test P_adj < 0.05, Fig. 4a, b). Amino acids showed the largest differences regarding the number of differentially expressed signatures (n = 22 [36%], Fig. 4c) and quantified differences (Supplementary Fig. 2a). Concordantly, the subgroup-differed metabolites were enriched in multiple amino acids metabolism pathways, including valine, leucine and isoleucine biosynthesis, alanine, aspartate and glutamate metabolism, arginine biosynthesis, glycine, serine and threonine metabolism and tyrosine metabolism (Fig. 4d).

Fig. 4. Diabetes subgroups significantly differed in amino acid metabolism.

a Differences in metabolites between a given subgroup and all other subgroups combined. Effect sizes were estimated with Cliff’s delta. Subgroup-differed metabolites were shown in purple, identified by P_adj < 0.05 from KW test among four subgroups. b Pairwise differences in metabolites between two subgroups. Effect sizes were estimated with Cliff’s delta. Subgroup-differed metabolites were shown in blue (significantly lower in one subgroup versus the other) and red (significantly higher in one subgroup versus the other). Significant subgroup differences were determined by post hoc Dunn’s test (P_adj < 0.05). c Percentages of metabolite classes of all detected metabolites and subgroup-differed metabolites. d The significantly enriched KEGG pathways (P_adj < 0.05) with subgroup-differed metabolites. e Subgroup differences in key metabolic pathways. The heatmaps show Z-scaled mean values of circulating levels of subgroup-differed metabolites in four subgroups. Metabolites not detected were shown in gray. f Violin plots show circulating levels of subgroup-differed metabolites (concentration [μΜ, Log2]) in four subgroups of different enriched pathways (SIDD: n = 242, MARD: n = 269, MOD: n = 141, SIRD: n = 80; n represents the number of patients). Embedded box plots show the median as the center line, first and third quartiles as box bounds, whiskers extending to 1.5 × IQR, and points beyond whiskers as outliers. The notch represents an approximate 95% confidence interval around the median, calculated as median ± 1.58 × IQR/sqrt(n). Stars show the statistical significance determined by post hoc Dunn’s test. *P_adj < 0.05, **P_adj < 0.01, ***P_adj < 0.001. Significance was regarded as a two-sided P value < 0.05, and type I error was further controlled by the FDR method. Source data are provided as a Source data file.

Consistent with findings from the IMI-RHAPSODY study¹⁹, both the SIRD and MOD subgroups showed high circulating levels of three BCAAs (isoleucine, valine and leucine) and their cognate α-ketoacids (Fig. 4e, f, Supplementary Fig. 2b). Circulating levels of two aromatic amino acids (AAAs; tyrosine and tryptophan) and the tyrosine-related metabolites homovanillic acid (HVA) and hydroxyphenyllactic acid (HPLA) were also the highest in the SIRD subgroup (Supplementary Fig. 2b). Besides, the SIRD subgroup showed a subgroup-specific activation of tricarboxylic acid (TCA) cycle as shown by the increased levels of isocitrate, α-ketoglutarate (α-KG) and malate (Fig. 4e, f, Supplementary Fig. 2b), and the levels of two relevant metabolites (i.e., pyruvate and lactate) were also higher in the SIRD subgroup (Fig. 4e, f, Supplementary Fig. 2b). In addition, the SIRD subgroup showed the highest circulating levels of primary bile acids (BAs; cholic acid [CA], chenodeoxycholic acid [CDCA]) and their conjugated forms (glycocholic acid [G-CA], taurocholic acid [T-CA] and taurochenodeoxycholic acid [T-CDCA]) (Fig. 4e, f, Supplementary Fig. 2a, b).

Circulating levels of several microbiota-dependent metabolites including phenylacetic acid (PAA), α-N-phenylacetyl-L-glutamine (PAGln), 2-phenylproionate, hydrocinnamic acid and hippuric acid were significantly lower in the MOD subgroup than in the other subgroups (Supplementary Fig. 2b). Among them, PAA and PAGln were phenylalanine metabolites in the oxidative pathway mediated by gut microbiota, while circulating levels of their precursors (i.e., phenylpyruvic acid) and the metabolites in the reductive pathway (i.e., phenyllactic acid) showed no differences between the MOD subgroup and other subgroups (Fig. 4e, f, Supplementary Fig. 2a, b). PAGln is a CVD-promoting metabolite, which accelerates the formation of platelet clots and enhances the potential for thrombosis³⁶. Thus, the lowest levels of PAGln could explain the lowest CVD risk of the MOD subgroup reported by previous research^{12–14,37,38}. The MOD subgroup also showed the lowest circulating level of citrulline, a metabolite involved in the urea cycle, and low levels of asparagine and N-acetyl-L-aspartate (NAA), which are derived from aspartate (Fig. 4e, f, Supplementary Fig. 2a, b). Meanwhile, the MOD subgroup showed stronger elevation in the levels of fatty acids including 2-hydroxybutyric acid, octanoic acid, adrenic acid, 8,11,14-Eicosatrienoic acid, 2-Hydroxy-3-methylbutyric acid, and 3-Hydroxybutyrylcarnitine (Supplementary Fig. 2a, b), and the SIDD subgroup showed elevated levels of carbohydrates (Supplementary Fig. 2a, b).

Subgroup-differed multi-omics signatures predicted efficacy of the insulin sensitizer

As we have identified subgroup-differed multi-omics signatures to capture the molecular heterogeneity among diabetes subgroups, we intended to investigate whether these signatures can help predict the efficacy of antidiabetic drugs. After treatment for 24 weeks, 476 (90%) of 529 participants in the insulin sensitizer groups completed the trial. The overall efficacy (defined as HbA1c at week 24 < 53 mmol/mol) markedly differed among the four subgroups, which was more frequent in patients in the SIRD subgroup (42 [74%] of 57 participants) compared with patients in the SIDD (59 [38%] of 154 participants, P < 0.001), MARD (95 [55%] of 173 participants, P = 0.009) and MOD (50 [54%] of 92 participants, P = 0.028) subgroups after adjusting for sex, trial and baseline HbA_1c (Fig. 5a, Supplementary Data 14). Accordingly, the reduction in HbA1c after 24 weeks was the greatest in patients in the SIRD subgroup (22.4 ± 1.4 mmol/mol) compared with patients in the SIDD (15.4 ± 1.1 mmol/mol, P_adj < 0.001), MARD (17.2 ± 1.0 mmol/mol, P_adj = 0.012), and MOD (17.1 ± 1.1 mmol/mol, P_adj = 0.015) subgroups after adjusting for sex, trial and baseline HbA_1c (Fig. 5b, Supplementary Data 15). The SIRD and MARD subgroups showed a trend toward better efficacy with the DPP4 inhibitor sitagliptin, and the MARD subgroup showed a trend toward better efficacy after lifestyle intervention, although neither reached statistical significance (Supplementary Fig. 3, Supplementary Data 14 and 15).

Fig. 5. Subgroup-differed multi-omics signatures predicted efficacy of the insulin sensitizer.

a Percentage of participants that achieved efficacy in each subgroup after treatment with the insulin sensitizer for 24 weeks. Categorical data were analyzed by a log-binomial regression model adjusted by sex, trial, and baseline HbA_1c. b Reduction of HbA_1c adjusted for sex, trial, and baseline HbA_1c in each subgroup after treatment with the insulin sensitizer for 24 weeks. Continuous data were analyzed using the ANCOVA model adjusted by sex, trial, and baseline HbA_1c, followed by post hoc comparison (the ANCOVA P value was shown in the left corner). *P_adj < 0.05, **P_adj < 0.01, ***P_adj < 0.001. c Receiver operating characteristic (ROC) curves for the model constructed by subgroup (dashed) and multi-omics signatures (continuous) models applied on the external validation cohort. The dotted line indicates random performance. The AUC values and 95% CI were labeled. Statistical analysis was performed using DeLong’s test. d The percentage of patients whose predicted efficacy was concordant with actual efficacy. e SHAP summary plot shows the importance of the top 30 important signatures in the combined subgroup and signature model. The points within the graph correspond to individual study participants. The y-axis ranks the predictors in descending order of importance, while the x-axis indicates the impact on the model output. Larger values denote a more significant positive influence on the prediction. The legend employs a gradient color scheme to represent different values of each variable. Significance was regarded as a two-sided P value < 0.05, and type I error was further controlled by the FDR method. Source data are provided as a Source data file.

Then, we constructed machine learning models to predict the efficacy of insulin sensitizer treatment, defined as achieving HbA1c < 53 mmol/mol after 24 weeks. The training dataset was constructed from patients who received the insulin sensitizer treatment and were among the first 50% of patients enrolled in one of the 53 clinical centers conducting one of the two clinical trials (n = 183). Three models, including different features, were derived using: (1) the four T2D subgroups defined in the previous section (SIDD, MARD, MOD, and SIRD by one-hot encoding), (2) subgroup-differed multi-omics signatures, and (3) integrated T2D subgroups and subgroup-differed multi-omics signatures. Within the pipeline, features were first imputed and scaled, and then fed into an unweighted ensemble classifier. Each ensemble consisted of three parallel algorithms: logistic regression with elastic net regularization, support vector machine (SVM), and eXtreme Gradient Boosting (XGBoost). The three algorithm scores were then averaged to form a predictor. A five-fold cross-validation scheme was used to optimize model hyperparameters (Supplementary Data 16 and 17).

The fully trained models were tested for validation on an external cohort of 178 patients who received the insulin sensitizer therapy and were later enrolled in the trials. In the external cohort, the area under the curve (AUC) was significantly higher in the model constructed by subgroup-differed multi-omics signatures (0.782) than in the model constructed by subgroup identity (0.580, DeLong’s test P < 0.001, Fig. 5c, Supplementary Data 18). The concordance rate, which is the percentage of participants with predicted efficacy concordant to actual efficacy, was also higher in the integrated (70.2%) or multi-omics model (68.0%) than in the subgroup model (54.5%, Fig. 5d, Supplementary Data 18). The fully integrated model increased the proportion of true negatives (TN) and reduced the proportion of false positives (FP) compared to the model with only subgroup identity (Supplementary Data 18). In other words, the integrated model outperformed other models in identifying T2D patients who do not respond effectively to the insulin sensitizer treatment, thereby avoiding insulin sensitizer treatment for this group of patients.

We explored the importance of the features used in the integrated training model combining subgroup-differed multi-omics signatures with subgroup identity. The contributions of each variable to the model were estimated using the absolute median SHapley Additive exPlanations (SHAP) values among all individuals in the validation cohort. The subgroup-differed signatures showed higher importance than the subgroup identity, with the largest contribution by metabolomic features (Fig. 5e, Supplementary Data 19). The most dominant features were glucose, 2-hydroxy-3-methylbutyric acid, let-7b-5p, and TAGLN (transgelin), and patients with lower levels of these features showed better efficacy after the treatment (Fig. 5e). From the perspective of utilization, we tested the model with four selected features of the highest importance together with the subgroup identity, and achieved a moderate performance in the external validation cohort (AUC = 0.728). This was significantly better than using the subgroup alone (DeLong’s test P = 0.0008), although lower than the fully integrated model with no statistical difference (DeLong’s test P = 0.061) (Supplementary Fig. 4).

Discussion

We validated T2D subgroups in our cohort consisting of 826 drug-naïve T2D patients based on cluster analyses with five clinical variables⁴. The percentages of patients in the subgroups were similar to those reported in other Asian populations, with MARD as the predominant subgroup⁵ and a higher percentage of SIDD than in European populations^9,39. After identifying subgroup differences in multi-omics signatures and therapeutic outcomes, we demonstrated that the multi-omics signatures could help enhance prediction of treatment response.

SIRD was the most distinguishable subgroup in our study, exhibiting the most distinct multi-omics signatures^4,19,20,28 and the best efficacy to insulin sensitizer treatment (Table 1). SIRD is clinically important for its high risk for diabetes-related comorbidities, such as hepatic fibrosis and nephropathy¹⁰. Consistently, we observed elevated ALT and AST levels in SIRD, and the AST/ALT ratio was lower in SIRD than SIDD and MARD (Supplementary Data 1). These routinely available liver function markers (AST and ALT) may serve as low-cost and easily accessible indicators for identifying SIRD. Notably, we identified SIRD-specific miRNA profiles that correlated with liver injury markers and were specifically expressed in the liver. Several of these miRNAs, such as miR-122, miR-192, and miR-34a-5p, have been previously implicated in MASLD and metabolic dysfunction-associated steatohepatitis (MASH, previously known as nonalcoholic steatohepatitis [NASH])^{4,7,8,10,40–45}. Meanwhile, SIRD showed higher levels of A2M and APOE, both associated with MASLD progression^40,46. The higher levels of A2M and APOE in SIRD compared with MOD are independent of BMI, indicating subgroup-specific differences beyond obesity (Supplementary Data 22).

Table 1.

Subgroup differences in genetics, molecular phenotypes and clinical outcome

Type	SIDD	MARD	MOD	SIRD
Clinical	1. The worst glycemic control (the highest HbA1c, fasting plasma glucose [FPG] and 2-h postprandial blood glucose [PBG])	1. The oldest age at diagnosis	1. The youngest age at diagnosis	1. Severe insulin resistance (the highest HOMA2-IR index)
	2. Poor insulin secretion (low HOMA2-B index)	2. The mildest metabolic derangement in glycemic control (relatively low HbA1c, FPG and PBG)	2. Obesity (the highest BMI and waist circumference [WC])	2. Retained insulin secretion (the highest HOMA2-B index)
			3. Elevated insulin resistance (high HOMA2-IR index)	3. Obesity (high BMI and WC)
				4. Worse lipid profiles (low high-density lipoprotein-cholesterol [HDL-C] and high triacylglycerol [TG])
				5. Elevated liver disease biomarkers (high alanine aminotransferase [ALT] and aspartate aminotransferase [AST]), kidney disease biomarkers (high urine microalbumin creatinine ratio [UACR]), and inflammatory biomarker (high hypersensitive C-reactive protein [hsCRP])
Genomics				1. The lowest genetic risk for beta cell function and pro-insulin
				2. Lower frequency of risk alleles of impaired beta cell function at genes including CDC123/CAMK1D, KCNJ11, HNF1B and CDKAL1
Methylomics		1. The lowest methylation risk score (MRS) for T2D
		2. Healthier levels of methylation sites associated with T2D risk, including hypermethylation of cg19693031 at TXNIP, hypermethylation of cg02650017 at PHOSPHO1, and hypomethylation of cg06500161 at ABCG1
MiRNAomics	1. The highest level of pancreas-specific miR-7-5p			1. The lowest levels of let-7b, miR-185-5p, miR-16-2-3p, miR-7-5p and miR-20a-5p
				2. The highest levels of miR-99/100, miR-122, miR-193a/193b, miR-885, miR-125b-2-3p, miR-505, and miR-27b-3p
Proteomics	1. The highest levels of 14-3-3 proteins (YWHAQ and YWHAH)		1. The lowest levels of A2M and SHBG	1. The lowest levels of SERPINF2, APOF, APOD
			2. The highest level of SERPINF2 and VTN	2. The highest levels of TUBA4A, NAP1L1, CFH, TAGLN, METTL18, APOE and A2M
Metabolomics			1. The lowest levels of phenylalanine metabolism metabolites including phenylacetic acid (PAA) and phenylacetylglutamine (PAGln)	1. The highest levels of branched-chain amino acids (BCAAs), aromatic amino acids (AAAs), tricarboxylic acid (TCA) cycle metabolites, and primary bile acids
Drug response		1. Better efficacy of the DPP4 inhibitor and lifestyle intervention		1. Best efficacy of the insulin sensitizer
				2. Better efficacy of the DPP4 inhibitor

SIRD showed high circulating levels of the TCA cycle intermediates, BCAAs, and AAAs. The TCA cycle plays a central role in mitochondrial metabolism. Excessive hepatic mitochondrial TCA cycle was observed in patients with insulin resistance and MASLD^47,48, indicating an adaptation to metabolic challenges. BCAAs and AAAs were predictors for future development of T2D⁴⁹ and contribute to the development of obesity-associated insulin resistance⁵⁰. SIRD also showed high circulating levels of primary BAs and their conjugated forms, indicating SIRD-specific dysfunction in the classic pathway of BA metabolism from cholesterol via cytochrome P450 7α hydroxylase (CYP7A1) enzyme. Elevated primary bile acids (CA and CDCA) and their conjugated forms (G-CA, T-CA, and T-CDCA) were dysregulated in patients with MASLD⁵¹. Mechanistically, it has been indicated that CYP7A1 expression is upregulated in patients with MASLD⁵².

Our data highlight the possibility of using the data-driven diabetes subgroups to guide clinical decisions on insulin sensitizer use for glucose-lowering effects. We found that patients in the SIRD subgroup had better efficacy after insulin sensitizer treatment. Notably, although SIRD and MOD subgroups showed similar molecular signatures, their efficacy is significantly different, indicating pathological differences between the two subgroups. Previous research showed that patients in the SIRD subgroup also exhibited better outcomes for metabolic surgery in terms of T2D remission and renal function¹⁵. Our results support the notion that data-driven classification can help refine the indications for T2D management. Meanwhile, we found the SIRD also showed a trend toward better efficacy with the DPP4 inhibitor, although without statistical significance. One possible explanation is that the SIRD patients have preserved β-cell function despite severe insulin resistance, which may allow them to benefit from the glucose-dependent insulin secretion enhanced by DPP4 inhibition. Therefore, this observation should be interpreted with caution and requires confirmation in larger, independent cohorts.

Importantly, we found that the predictive ability for the insulin sensitizer efficacy could be improved by adding subgroup-differed signatures, and thus the signatures could become biomarkers for future utilization. The most dominant features were glucose, 2-hydroxy-3-methylbutyric acid, let-7b-5p, and TAGLN. Patients with lower baseline levels of let-7b-5p and TAGLN exhibited better responses to insulin sensitizer treatment. Let-7b-5p, a member of the let-7 microRNA family that is highly expressed in the pancreas, showed the highest expression in the SIDD subgroup and the lowest in the SIRD subgroup (Fig. 2e). In vivo studies have demonstrated that global and pancreas-specific overexpression of let-7 in mice impairs glucose tolerance and reduces glucose-stimulated insulin secretion⁵³, supporting its regulatory role in insulin secretion. Thus, lower baseline let-7b-5p levels may indicate more preserved β-cell function and greater capacity for pharmacological improvement with insulin sensitizers. TAGLN, an actin-binding protein, regulates cytoskeletal dynamics and is involved in smooth muscle differentiation, extracellular matrix remodeling, and cellular stress responses. Cytoskeletal organization and actin remodeling have been reported to modulate insulin receptor trafficking and downstream signaling⁵⁴. However, direct mechanistic evidence linking reduced TAGLN expression to T2D pathophysiology and treatment response is currently limited, and further mechanistic studies will be necessary to clarify whether low TAGLN expression directly enhances responsiveness to insulin sensitizers. We also examined the potential influence of sex on the identified omics signatures. Sex-stratified analyses indicated that the key signatures of interest (e.g., let-7b-5p, TAGLN) remained significant, suggesting that these findings are largely independent of sex (Supplementary Data 23).

Subgroup differences in several biomarkers identified by other omics studies (IMI-RHAPSODY¹⁹, QBB⁹ and ORIGIN²⁰) were validated by our study, including higher levels of BCAAs¹⁹, tyrosine¹⁹, MFGE8¹⁹ and lower level of YWHAQ¹⁹ in the SIRD subgroup, lower levels of A2M¹⁹ and SHBG¹⁹ in the MOD subgroup, and higher levels of threonine⁹, asparagine⁹, glutamine⁹ and SHBG²⁰ in the MARD subgroup (Supplementary Data 8 and 11). However, comparability between studies was limited by the notable heterogeneity in population structure, biomarker assays and statistical methods.

We acknowledge several limitations of our current study. First, the analyses were based on Chinese patients enrolled in trials with inclusion criteria of insufficient glycemic control and exclusion criteria of certain medications or complications, who also adhered to a strict diet and exercise regimen. Thus, the generalizability to patients with milder or more severe diabetes, or to other ethnicities, may be limited. For example, the obesity- and lipodystrophy-related pGRS did not differ between MOD and SIRD, likely due to reduced transferability of European-derived scores to Chinese cohorts. Nevertheless, several omics signatures identified in our study were validated in European cohorts, supporting their broader relevance. In addition, our analyses were based on a relatively short treatment duration, which limited the assessment of subgroup differences in drug resistance. Future studies including more diverse populations and longer follow-up will be essential to confirm and extend these findings.

Secondly, our measurements failed to capture important biomarkers reported by others (e.g., leptin^9,20,55), which prevented comparisons with previous studies. This is primarily due to the technical sensitivity limits of the mass spectrometry-based proteomic platform used in our study. In future studies, these low-abundance biomarkers could be examined using targeted proteomics or immunoassay-based methods.

Third, we acknowledge that translating the molecular signatures into routine clinical practice may be limited by resource availability in many healthcare systems. The cost and accessibility of high-throughput omics technologies remain a barrier for large-scale implementation. Despite these challenges, identifying smaller sets of key biomarkers could facilitate the development of more practical, clinically applicable assays for patient stratification. We found the model with four selected features of the highest importance together with the subgroup identity, and achieved a moderate performance in the external validation cohort (Supplementary Fig. 4). Thus, a reduced panel, such as the top four predictors combined with subgroup identity, will be essential for future clinical translation. With larger cohorts and more comprehensive treatment data, the performance of this panel could be evaluated across multiple drug classes simultaneously.

Finally, we did not establish causality between subgroup-specific molecular signatures and T2D outcomes, and future work will incorporate causal inference approaches such as Mendelian randomization, alongside targeted experimental validation, to clarify the mechanistic roles of key signatures.

In summary, we systematically characterized molecular signatures of diabetes subgroups associated with disease etiology and therapeutic outcomes in drug-naïve T2D patients. These signatures can be accessed at our website (http://fudan-pgx.org/chihope) and may be further developed to realize precision medicine to improve treatment for diabetes subgroups.

Methods

Ethical compliance

We confirm that this research complies with all relevant ethical regulations. The study protocols for the Phase 3 clinical trials of chiglitazar were approved by the Ethics Committee of Peking University People’s Hospital (for NCT02121717) and the Ethics Committee of Shanghai 6th People’s Hospital (for NCT02173457), which served as the lead centers, as well as by the institutional review boards at each of the participating clinical sites. The study was conducted in accordance with the principles of the Declaration of Helsinki and Good Clinical Practice (GCP) guidelines. Written informed consent was obtained from all study participants prior to enrollment, specifically covering the collection of blood samples and their subsequent use for multi-omics research purposes.

Study populations

Drug-naïve Chinese patients with T2D were collected from two multi-center, randomized, double-blind phase 3 trials of chiglitazar (CMAP, placebo-controlled, n = 350, ClinTrials.gov registration no. NCT02121717; CMAS, sitagliptin-controlled, n = 485, ClinTrials.gov registration no. NCT02173457). Briefly, patients with previously untreated recent-onset T2D were eligible for inclusion if they were aged 18–70 years with a BMI range of 18.5–35.0 kg/m² and insufficient glycemic control (HbA_1c ≥ 79.6 mmol/mol and ≤85.8 mmol/mol) despite a strict diet and exercise regimen^22,23. The patients were randomly assigned to receive chiglitazar 32 mg/day, chiglitazar 48 mg/day, sitagliptin 100 mg/day or placebo. Ethical approvals were obtained from the Ethical Committees of the study centers. The study cohort comprised 524 (63.4%) males and 302 (36.6%) females. Specifically, the distribution of sex across the four subgroups was as follows: SIDD (179 males, 66.8%), MARD (184 males, 59.4%), MOD (111 males, 68.1%), and SIRD (50 males, 58.8%), with no significant difference observed in sex distribution between groups (P = 0.120, Fisher’s exact test). Participant sex was determined based on self-report at the time of enrollment. To account for potential biological variability, sex was included as a covariate in all relevant statistical analyses, including partial Spearman correlation and machine learning models.

Data generation

Genetic data

Genetic data of whole blood samples from 835 patients at baseline were generated by Affymetrix Axiom^® Precision Medicine Research (PMR) Arrays under instructions from the manufacturer. Samples were excluded for ambiguous gender, call rate <98%, extra heterozygosity (>3 SDs from the mean), and any duplicate or related individuals (pi_hat ≥0.2). SNPs were excluded for those with a minor allele frequency (MAF) < 0.05, missingness rate >2%, or Hardy-Weinberg equilibrium deviation (P < 10⁻⁶). The ChinaMAP reference panel⁵⁶ v1.0 (www.mbiobank.com) was used to impute after transforming the reference genome from hg19 to hg38 with CrossMap⁵⁷. Poorly imputed variants, defined as imputation quality R² < 0.9, were excluded from downstream analyses.

DNA methylation data

DNA methylation data of whole blood samples from 835 patients at baseline were generated by Infinium MethylationEPIC BeadChips (Illumina) under instructions from the manufacturer. Genomic DNA from each whole blood sample was bisulfite converted using the EZ DNA Methylation Kit (Zymo Research). BeadChips were processed following the manufacturer’s guide and protocol for the Infinium MethylationEPIC array. DNA was hybridized to BeadChips and single base extension was performed using a Freedom EVO robot (Tecan). BeadChips were subsequently imaged using the iScan Microarray Scanner (Illumina). The idat files were processed with the R package minfi⁵⁸, without background correction. Probes with SNPs were removed using the function dropLociWithSnps from minfi, and probes on sex chromosomes were also removed. The β values with P values of detection greater than 0.05 were set to NA, and only probes with less than 5% missing values in all samples were retained. The missing β values were imputed with the function impute.knn (k = 5). Type-2 probe bias was corrected using beta-mixture quantile normalization (BMIQ)⁵⁹. We corrected the slide effects by ComBat on M values (logit of β values) and then transformed the M values back to β values. The proportion of cell type components were estimated by the R package EpiDISH⁶⁰.

MiRNAomic data

MiRNAomic data of 1990 plasma samples from 746 patients with T2D at three time points were sequenced by the NovaSeq6000 platform (Illumina). Highly detected miRNAs were included (n = 396), which were defined as those with more than 10 counts in at least 30% of samples. The count data were transformed to log2-transformed counts per million (CPM) for further analyses^61–63. The TSI in each organ (or tissue) of miRNAs was annotated by miRNATissueAtlas2 (https://www.ccb.uni-saarland.de/tissueatlas2)²⁹. The miRNAs with a TSI greater than 0.8 were considered tissue-specific. The expression of each miRNA in a given organ was calculated as the median value of log2 reads per million in all samples, and only organs with data from at least four samples were displayed (Supplementary Fig. 2a). The mRNAs targeted by miRNAs were identified by microT-CDS, and gene-set enrichment analyses to investigate the enrichment of target mRNAs in the KEGG pathways were conducted using DIANA-miRPath v.3.0 (http://www.microrna.gr/miRPathv3)⁶⁴.

Proteomic data

Peptides of 1419 plasma samples from 723 patients with T2D at two time points were measured using LC-MS instrumentation consisting of an EASY-nLC 1200 ultra-high-pressure system (Thermo Fisher Scientific) coupled via a nano-electrospray ion source (Thermo Fisher Scientific) to a Q Exactive HF-x (Thermo Fisher Scientific). Peptide-query level intensities were assembled into protein groups (identified gene symbols) by iBAQ protein quantification. A quantile normalization approach was utilized after log2 transformation to remove the overall inter-sample bias. Each protein quantity was corrected by a locally weighted linear regression (LOESS) fitting in each batch group separately⁶⁵. The median centering method was applied to remove the residual bias existing in these discrete batch groups. Proteins with no missing values in at least 70% of samples were included for further analysis (n = 858). The tissue specificity and transcripts per million (TPM) in each organ (or tissue) of proteins were annotated by the HPA (https://www.proteinatlas.org/)⁶⁶. Pathway enrichment analyses of subgroup-differed proteins were conducted by MetaScape (http://metascape.org)⁶⁷.

Metabolomic data

Metabolites of 2148 plasma samples from 724 patients with T2D at three time points were measured by targeted UHLPC-MS/MS. Metabolites with no missing values in at least 70% of samples were included for further analysis (n = 203). Missing values were imputed by half of the minimum of non-missing values. The original data were log2 transformed for analyses.

Outcomes

The categorical efficacy was defined as large efficacy (HbA_1c < 48 mmol/mol) and modest efficacy (HbA_1c < 53 mmol/mol) at week 24, and the continuous efficacy was defined as the change of HbA_1c from baseline to week 24. Other outcomes, including changes in other phenotypes from baseline to week 24, were also assessed.

Cluster analysis

We applied the k-means algorithm as proposed by Ahlqvist et al.⁴ to stratify drug-naïve patients with T2D into four subgroups. Age at diagnosis, BMI, HbA_1c, HOMA2-B and HOMA2-IR based on serum insulin were used as clustering variables. Cluster analysis was done on values centered to a mean value of 0 and an SD value of 1. Patients without measurements of five clinical variables (n = 2) or of extreme outliers (>5 SDs from the mean; n = 7) were excluded. K-means clustering was done with a k value of 4 using the kmeansruns function (runs = 100) of the R package fpc. Cluster-wise stability was assessed through resampling the dataset 2000 times and computing the Jaccard similarities to the original cluster. Men and women were clustered together to achieve an overall higher Jaccard similarity. Cluster labels were assigned by examining cluster variable means. Four subgroups that recapitulated the basic phenotypes of the clusters observed by other studies were obtained in our cohort (mean Jaccard similarities = 0.84, 0.83, 0.62 and 0.75 for SIDD, MARD, MOD and SIRD, respectively). To provide a better visualization of the clustering results, we applied t-SNE to project the original data into a 2-dimensional space using the Rtsne function with default parameters in the R package Rtsne.

Machine learning predictive models

Training and validation datasets

The study subjects consisted of 534 T2D patients from the ChiHOPE cohort who received insulin sensitizer treatment. The training cohort was constructed from the first 270 (50.6%) T2D patients enrolled at each trial center, while the external validation cohort was constructed from the remaining 264 (49.4%) patients. Specifically, 183 and 178 T2D patients with baseline data of all included features and 24-week HbA1c data were used as the training and validation cohorts, respectively. The predictive endpoint was defined as whether HbA1c levels dropped below 53 mmol/mol after 24 weeks of insulin sensitizer treatment. Predictive features included the baseline values of multi-omics subgroup-differed features.

Model training

In the training dataset of 183 T2D patients, a five-fold cross-validation strategy was adopted. All patients were randomly divided into five equal parts (folds). In each iteration of cross-validation, four of these folds were used as training data to train the model, while the remaining fold was used as validation data to evaluate the model’s performance. This process was repeated five times, with each fold serving as the validation set once and as part of the training set four times. After all five iterations were completed, the AUC-ROC values were averaged across the five validation folds. During the model training, hyperparameters were automatically optimized to maximize the averaged AUC-ROC by a grid search strategy. The remaining parameters were determined by retraining the model on the entire training dataset using the optimal hyperparameters. To ensure robustness, the optimization process was repeated 10 times with different cross-validation seeds, resulting in 10 alternative predictors. These 10 predictors together formed the “model,” with the final predictor score taken as the average of the 10 individual predictor scores. All retrained predictors were frozen and then independently validated on an external dataset of 178 patients.

Predictor architecture

The machine learning framework was implemented using Python v3.9.0, leveraging libraries such as scikit-learn v1.3.2, numpy v1.26.3, scipy v1.11.4, and pandas v2.1.4.⁹The predictor was developed using three separate pipelines, each composed of four key steps: imputation, scaling, feature selection, and classification. Initially, missing data were imputed using the median value for each feature across all patients. Subsequent to imputation, z-score normalization was applied to standardize each feature across the patient cohort. For feature selection, a univariable approach was employed, retaining only those features ranked within the top k based on their ANOVA F-value concerning the binary response variable. The final step involved classification, utilizing logistic regression in the first pipeline, an SVM classifier in the second, and a random forest classifier in the third. The three pipeline scores were finally averaged to form an ensemble predictor. Hyperparameter optimization for each pipeline was conducted through a randomized grid search spanning 1000 iterations, coupled with five-fold cross-validation, aiming to maximize the averaged AUC-ROC. The logistic regression model was implemented with elastic net regularization and the SAGA solver, exploring C parameters in the range of 10⁻³ to 10³ and L1 ratios from 0.1 to 1. For the SVM classifier, a linear kernel was employed, with C parameters ranging from 10⁻³ to 10³. The random forest classifier was configured with estimators numbering between 5 and 100, or up to the maximum allowed, with maximum features set between 5 and 70% of the total, and the minimum number of samples required per split ranging from 2 to 15. Details of the final optimized hyperparameters are available in Supplementary Data 17.

Model testing

All frozen predictors were applied on the external validation cohort, and their respective AUC-ROC, true positive (TP), true negative (TN), false positive (FP), false negative (FN), precision, sensitivity/recall, specificity, accuracy, F1 score, and Matthews correlation coefficient (MCC) values were evaluated with the functions roc, auc and confusionMatrix from the R package caret (6.0.94). The 95% CI of each test AUC was obtained with the DeLong’s test by the function roc.test of R package pROC (1.18.5).

Feature importance

Feature importance was determined for each algorithm (logistic regression, SVM and random forest) after refitting on the full training cohorts, which was estimated by the SHAP values calculated by the Python library shap (v0.44.0). The logistic regression and the SVM models were explained with the linear SHAP explainer, while the random forest model was explained with the tree SHAP explainer. To ensure comparability of SHAP values across different algorithms, we removed univariable selection from the predictor architecture and retrained three pipelines over the same set of input features. SHAP values from the three algorithms were then aggregated and averaged to produce SHAP values for ensemble predictors. To ensure the robustness of feature ranking, the optimization process was repeated 10 times with different cross-validation seeds, effectively training 10 alternative predictors. The ensembled SHAP values from these 10 predictors were further aggregated and averaged. For each feature, the importance was calculated by the averaged absolute SHAP values of all patients.

Statistical analysis

Genetic risk score

We constructed GRS for T2D and pGRSs for mechanistic pathways of T2D, including beta-cell, pro-insulin, obesity, lipodystrophy and liver, based on summary statistics of previously reported genome-wide significant SNPs described by Spracklen et al.²⁵ and Udler et al.²⁶, respectively (Supplementary Data 20). The GRS was calculated as the sum of the dosage weighted by the effect size of each SNP by the Eq. (1).

$$GRS=\frac{{\sum }_{i=1}^n\left({SNP}_{i\left(LocalCohort\right)}\times Effect{Size}_{i\left(DiscoveryCohort\right)}\right)}{{\sum }_{i=1}^n{EffectSize}_{i\left(DiscoveryCohort\right)}}$$ 1

We conducted linear regression to examine the associations of GRSs and included SNPs with subgroups (the SIRD subgroup as reference), adjusted for sex and the first five principal components. Regression models were not adjusted for age at onset, BMI, HbA1c, HOMA2-B or HOMA2-IR to avoid multicollinearity.

Methylation risk score

We constructed a MRS for T2D based on summary statistics of five reproducible CpG sites described by Chambers et al.²⁷ (Supplementary Data 21). The MRS was calculated as the sum of the standardized methylation values at each CpG site, weighted by site-specific effect size by the Eq. (2).

$$MRS=\sum _{i=1}^5\left(\frac{methylation\beta {value}_{i\left(LocalCohort\right)}-mean{methylation}_{i\left(LOLIPOPCohort\right)}}{SD{methylation}_{i\left(LOLIPOPCohort\right)}}\right)\times {\beta coefficient}_{i\left(LOLIPOPCohort\right)}$$ 2

We conducted linear regression to examine the associations of the MRS and included CpG sites with subgroups (the MARD subgroup as reference), adjusted for sex, estimated proportions of cell components and the first five principal components. Regression models were not adjusted for age at onset, BMI, HbA1c, HOMA2-B or HOMA2-IR to avoid multicollinearity.

Subgroup differed and specific signatures

Subgroup differed signatures were identified by the KW test among four subgroups (P_adj < 0.05), and the differed subgroup pairs were determined by post hoc Dunn’s multiple comparison test (P_adj < 0.05). Cliff’s delta values were calculated as estimates of effect sizes to quantify the differences between subgroups. For a given signature, if its level in one subgroup was significantly different from all of the other subgroups, it was defined as a “subgroup-specific” signature for that subgroup.

Subgroup efficacy differences

The categorical efficacy (proportion of individuals achieving HbA_1c < 53 mmol/mol) was analyzed using a log-binomial regression model (the SIRD subgroup as reference) adjusted by sex, trials, and baseline HbA_1c. The continuous efficacy was analyzed using analysis of covariance (ANCOVA) model with sex, trial, and baseline HbA_1c as covariates to assess the least-squares (LS) mean differences and the associated two-sided 95% confidence interval (CI) among different subgroups under each treatment. Post hoc comparisons were performed if the overall P value was significant, and type one error was further controlled by the false discovery rate (FDR) method if applicable.

All statistical analyses were conducted in R version 4.2.2 (R Foundation for Statistical Computing, Vienna, Austria). Significance was regarded as a two-sided P value < 0.05 unless otherwise specified, and type I error was further controlled by the FDR method if applicable.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Y.T.Z., L.S. and Y.Z. conceived and oversaw the study. Y.T.Z. and L.S. coordinated multi-omics data generation and data analysis. D.P., L.J., W.J. and X.L. coordinated sample collection during clinical trials. Z.C. conducted most data analysis and generated most figures; R.Z. conducted data analysis and generated figures; and H.C. conducted predictive modeling analysis and website development. Y.Z., N.Z., Q.C.C., Y.M., X.Y., Q.W.C., J.L., D.P., L.J., W.J, X.L. and A.V.P. participated in data analysis, validation, or interpretation. Z.C., Y.T.Z. and L.S. wrote and revised the manuscript. All authors reviewed and approved the manuscript.

Peer review

Peer review information

Nature Communications thanks Constantinos Christodoulides, Rajna Golubic, and Yan Li for their contribution to the peer review of this work. A peer review file is available.

Funding

This study was supported in part by the National Natural Science Foundation of China (T2425013, 32470692, 32370701 and 32170657), Shanghai Municipal Science and Technology Major Project (2023SHZDZX02), the National Key R&D Project of China (2023YFC3402501), and the 111 Project (B13016).

Data availability

The multi-omics dataset of the ChiHOPE cohort (genomics, methylomics, miRNAomic) has been deposited at Genome Sequence Archive (GSA) under accession code PRJCA026766 and is publicly available as of the date of publication. The proteomics dataset of the ChiHOPE cohort has been deposited at the ProteomeXchange partner database under accession code PXD068273. The multi-omics dataset of the ChiHOPE cohort has also been deposited in the National Omics Data Encyclopedia (NODE, https://www.biosino.org/node) and is publicly available as of the date of publication at the accession number OEP00002924 or through the URL. The raw clinical data are protected and are not available due to data privacy laws and ethical restrictions. The de-identified clinical data are available under restricted access for reanalyzing the results reported in this paper; access can be obtained by contacting the corresponding author with a formal research proposal and a signed Data Transfer Agreement. The source data generated in this study are provided in the Source Data file. Source data are provided with this paper.

Code availability

All original code has been deposited on Zenodo and is publicly available at https://doi.org/10.5281/zenodo.14923390.

Competing interests

D.P. and X.L. are employees and stockholders of Shenzhen Chipscreen Biosciences Co. Ltd. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77187-8.