To systematically investigate potential mitochondria-related candidate biomarkers for LLF in the treatment of DN, we designed a four-phase analytical workflow (Figure 1). In Phase I, we integrated transcriptomic data from the GSE142025 dataset (training set, whole kidney, n=36) and GSE96804 (validation set, glomerulus, n = 61) with 1,136 mitochondria-related genes from the MitoCarta 3.0 database and 517 predicted targets of 9 active ingredients from the TCMSP database. Overlapping these three gene sets yielded 9 candidate genes. In Phase II, four machine learning models (RF, KNN, PLS, and SVM) were applied to prioritize feature genes using RMSE < 0.281 as the threshold. Cross-dataset validation using ROC analysis (AUC > 0.7 in both datasets) identified four candidate biomarkers: CAT, FABP1, MAOB, and MAOA. In Phase III, we performed GSEA to identify enriched KEGG pathways, immune infiltration analysis using CIBERSORT, m6A modification prediction, and constructed lncRNA-miRNA-mRNA, active ingredient–biomarker, and active ingredient–biomarker–pathway networks, followed by molecular docking. In Phase IV, the pharmacodynamic effects of LLF and changes in mRNA expression of the four candidate biomarkers were evaluated in a db/db mouse model of DN.
Screening of candidate genes for LLF treating DN
In the GSE142025 dataset, 3,810 DEGs were identified between the DN and control groups, including 1,904 upregulated and 1,906 downregulated DEGs (Figure 2A,B). Thirteen active ingredients from LLF were predicted using the TCMSP database, namely beta-sitosterol, kaempferol, taxifolin, Lucidumoside D, Lucidumoside D_qt, (20S)-24-ene-3,20-diol-3-acetate, eriodictyol, syringaresinol diglucoside_qt, Lucidusculine, Olitoriside, Olitoriside_qt, luteolin, and quercetin (Table 2). Four active ingredients—Lucidumoside D_qt, (20S)-24-ene-3,20-diol-3-acetate, syringaresinol diglucoside_qt, and Olitoriside_qt—did not predict any potential target genes, while the remaining nine ingredients predicted 517 potential target genes. By overlapping the 3,810 DEGs, 1,136 MRGs, and 517 potential target genes, nine candidate genes were identified: GPX1, BAX, CASP8, MAOA, MAOB, CAT, AKR1B10, ALDH2, and FABP1 (Figure 2C). An active ingredient-candidate gene network was subsequently constructed (Figure 2D). These nine candidate genes were enriched in 341 GO terms, including response to toxic substances, organic hydroxy compound catabolic process, and cellular detoxification (Figure 2E). Additionally, they were associated with 52 KEGG pathways, such as tryptophan metabolism, neurodegenerative pathways, and histidine metabolism (Figure 2F).
Screening of candidate biomarkers for DN treatment in LLF
The PPI network revealed seven nodes and eight edges, with MAOA, ALDH2, MAOB, and AKR1B10 interacting (Figure 3A). Genes with RMSE values less than 0.281 across four machine learning models were identified as feature genes: CAT, MAOB, MAOA, BAX, and FABP1 (Figure 3B-E). Expression analysis showed that CAT, FABP1, MAOB, and MAOA were significantly different between the DN and control groups and were consistent in both the GSE142025 and GSE96804 datasets (Figure 3F,G). Furthermore, their AUC values in ROC curve analysis exceeded 0.7 in both datasets, indicating that these genes could effectively differentiate DN samples from control samples and serve as candidate biomarkers for DN treatment in LLF (Figure 4A-H).
Significant enrichment of candidate biomarkers in inflammatory and immue-related pathways
GSEA identified four candidate biomarkers prominently enriched in the chemokine signaling pathway and cytokine-cytokine receptor interactions (Figure 5A-D). Among these, the peroxidase signaling pathway showed a significant association with CAT, MAOA, and MAOB.
Correlation of candidate biomarkers with immune cells
Notable differences in the expression of nine immune cell types-naive B cells, M0 Macrophage, M1 Macrophage, M2 Macrophage, activated Mast cell, activated NK cell, resting memory CD4+ T cell, naive CD4+ T cell, and CD8+ T cell—were observed between the DN and control samples (P < 0.05) (Figure 6A,B). A significant positive correlation (cor = 0.6) was found between naive B cells and activated NK cells, while a significant negative correlation (cor = -0.69) was detected between naive B cells and activated Mast cells (Figure 6C). All candidate biomarkers exhibited strong negative correlations with CD8+ T cells and activated Mast cells and positive correlations with activated NK cells and naive B cells (Figure 6D).
Interaction of key modified m6A proteins with candidate biomarkers
The m6A RNA methylation modification profoundly affects RNA synthesis and metabolism and is implicated in the pathogenesis of various diseases29. The locations of the m6A modification sites in the candidate biomarkers and their high-confidence positions in the secondary structures are illustrated in Figure 7A-H. Further analysis revealed that key m6A-modified proteins interacting with CAT included AQR and RBM22, while FABP1 interacted with both SF3A3 and AQR. MAOA was found to interact with IGF2BP3 and IGF2BP2, and MAOB with TIA1 (Table 3).
Favorable in silico binding predictions for taxifolin, beta-sitosterol, and eriodictyol in LLF treating DN
In miRNet, CAT was predicted to interact with 24 miRNAs, while FABP1 was associated with five miRNAs. Additionally, MAOB and MAOA were linked to 29 and 26 miRNAs, respectively. Among these, 23 lncRNAs were identified in both the TarBase and Starbase databases. A lncRNA-miRNA-mRNA regulatory network was then constructed, incorporating four candidate biomarkers, 74 miRNAs, and 23 lncRNAs (Figure 8A). Potential active ingredients targeting the candidate biomarkers included luteolin, beta-sitosterol, eriodictyol, kaempferol, quercetin, and taxifolin (Figure 8B). Moreover, an active ingredient-biomarker-pathway network was established based on the active ingredients, candidate biomarkers, and the top five pathways identified in GSEA (Figure 8C). For instance, taxifolin targeted CAT in the peroxisome pathway. The binding energies between CAT and taxifolin (-8.8 kcal/mol), FABP1 and beta-sitosterol (-8.1 kcal/mol), and MAOB and eriodictyol (-9.8 kcal/mol) were all below -5 kcal/mol, suggesting strong affinities between these candidate biomarkers and their respective active ingredients27. Taxifolin, beta-sitosterol, and eriodictyol were identified as potential active ingredients with favorable in silico binding predictions in LLF treating DN (Figure 8D-F). However, they are presented as database-predicted constituents rather than confirmed bioactive intermediates of the observed in vivo effects.
Validate candidate biomarkers in the DN mouse model
Pharmacodynamic evaluation of LLF in treating DN mice
During the administration period, blood glucose and urinary microalbumin levels in mice were monitored (Figure 9A-D). Compared with the control group, blood glucose and urinary microalbumin in DN model group was significantly increased (P < 0.01); compared with the DN model group, blood glucose of the mice in the treatment group were significantly decreased after 4 weeks of administration (P < 0.01) and urinary microalbumin of the mice in the treatment group were significantly decreased after 8 weeks of administration (P < 0.05). The results suggest that LLF could be beneficial in treating DN.
Pathological evaluation of LLF in treating DN mice
Following HE staining, the control group exhibited clear glomerular structures in kidney tissue. In contrast, the DN model group showed glomerular nuclear pyknosis and hyperchromasia, along with inflammatory cell infiltration around the glomeruli, compared to the normal group. Treatment with LLF ameliorated pathological damage in the kidneys of db/db mice (Figure 9E).
RT-PCR analysis of the expression of candidate biomarkers in DN mice
Following the successful establishment of a DN mouse model and the observation of significant improvement in symptoms with LLF treatment, RT-qPCR was further used to analyze the changes of candidate biomarkers. Compared to the control group, the DN group exhibited significantly reduced expression of CAT and MAOA (P < 0.05 or P < 0.001). Conversely, the treatment group showed significantly higher CAT and MAOA expression than the DN group (P < 0.05). However, no statistically significant differences were observed in the expression of MAOB and FABP1 among the groups (Figure 9F-I).
Data availability
Gene expression datasets analyzed in this study are publicly available from the Gene Expression Omnibus (GEO) under accession numbers GSE142025 and GSE96804. The R scripts used for the bioinformatics analyses, together with the experimental source data (blood glucose, urinary microalbumin, and RT-qPCR data), are provided in Supplemental File 1. All other databases, software, and web resources used in this study are listed in the Table of Materials.

Figure 1: Workflow of the study. Transcriptomic datasets, mitochondria-related genes, and predicted targets of Ligustri Lucidi Fructus were integrated to identify candidate genes. Four machine-learning algorithms were then used to prioritize feature genes, followed by cross-dataset validation, functional characterization, and experimental validation in db/db mice. Abbreviations: DN = diabetic nephropathy; DEGs = differentially expressed genes; MRGs = mitochondria-related genes; LLF = Ligustri Lucidi Fructus; RF, random forest; KNN = k-nearest neighbor; PLS = partial least squares; SVM = support vector machine; RMSE = root mean square error; GSEA = gene set enrichment analysis; RT-qPCR = reverse transcription quantitative polymerase chain reaction. Please click here to view a larger version of this figure.

Figure 2: Screening and functional characterization of candidate genes for LLF treatment of DN. (A) Volcano plot showing differentially expressed genes between DN and control samples in GSE142025. (B) Heatmap of the top 10 upregulated and top 10 downregulated genes ranked by |log2FC|. (C) Venn diagram showing the intersection of DEGs, MRGs, and predicted LLF target genes. (D) Active ingredient–candidate gene network. (E) Gene Ontology enrichment analysis of candidate genes. Bar height represents enrichment significance, and the z-score indicates the predicted direction of functional regulation. (F) Kyoto Encyclopedia of Genes and Genomes pathway enrichment analysis of candidate genes. Abbreviations: DN = diabetic nephropathy; LLF = Ligustri Lucidi Fructus; DEGs = differentially expressed genes; MRGs = mitochondria-related genes; GO = Gene Ontology; KEGG = Kyoto Encyclopedia of Genes and Genomes. Please click here to view a larger version of this figure.

Figure 3: Machine-learning-based identification of candidate biomarkers. (A) Protein–protein interaction network of proteins encoded by the candidate genes. (B) Reverse cumulative distribution of residuals for the RF, KNN, PLS, and SVM models. (C) Boxplots showing residual distributions of the four models; the red point indicates the root mean square error. (D) RMSE-based importance of candidate genes across the four machine-learning models. (E) Intersection of feature genes meeting the RMSE < 0.281 criterion across all four models. (F,G) Expression of the selected feature genes in GSE142025 and GSE96804, respectively. Abbreviations: RF = random forest; KNN = k-nearest neighbor; PLS = partial least squares; SVM = support vector machine; RMSE = root mean square error. Please click here to view a larger version of this figure.

Figure 4: Receiver operating characteristic curves of the four candidate biomarkers. ROC curves for CAT, FABP1, MAOB, and MAOA in the (A-D) GSE142025 training dataset and (E-H) GSE96804 validation dataset. The AUC represents the area under the receiver operating characteristic curve. Abbreviations: ROC = receiver operating characteristic; AUC = area under the curve. Please click here to view a larger version of this figure.

Figure 5: Gene set enrichment analysis of candidate biomarkers. GSEA showing significantly enriched KEGG pathways associated with (A) CAT, (B) FABP1, (C) MAOA, and (D) MAOB in the GSE142025 dataset. Abbreviations: GSEA = gene set enrichment analysis; KEGG = Kyoto Encyclopedia of Genes and Genomes. Please click here to view a larger version of this figure.

Figure 6: Immune-cell infiltration and its association with candidate biomarkers in DN. (A) Relative proportions of 22 immune cell types estimated by CIBERSORT in DN and control samples. (B) Comparison of significantly different immune-cell fractions between DN and control groups. (C) Correlation matrix among the differentially abundant immune cell types. (D) Spearman correlations between the expression of CAT, FABP1, MAOA, and MAOB and the differentially abundant immune cell types. Abbreviations: DN = diabetic nephropathy. Please click here to view a larger version of this figure.

Figure 7: Predicted m6A modification sites and RNA secondary structures of candidate biomarker transcripts. Predicted m6A modification sites in (A) CAT, (B) FABP1, (C) MAOA, and (D) MAOB. Predicted RNA secondary structures showing high-confidence m6A-associated regions of (E) CAT, (F) FABP1, (G) MAOA, and (H) MAOB. Yellow-highlighted regions indicate the predicted sequence regions containing m6A modification sites. Abbreviation: m6A = N6-methyladenosine. Please click here to view a larger version of this figure.

Figure 8: Regulatory networks and molecular docking of potential active ingredients of LLF. (A) Predicted lncRNA–miRNA–mRNA regulatory network involving the candidate biomarkers. (B) Network of potential LLF active ingredients and candidate biomarkers. (C) Active ingredient–biomarker–pathway network based on the GSEA results. (D-F) Predicted molecular docking conformations of (D) CAT with taxifolin, (E) FABP1 with beta-sitosterol, and (F) MAOB with eriodictyol. Abbreviations: LLF = Ligustri Lucidi Fructus; lncRNA = long noncoding RNA; miRNA = microRNA; GSEA = gene set enrichment analysis. Please click here to view a larger version of this figure.

Figure 9: Effects of LLF treatment on biochemical indicators, renal histopathology, and candidate biomarker expression in db/db mice. (A,B) Blood glucose levels at baseline and week 8, respectively. (C,D) Urinary microalbumin levels at baseline and week 8, respectively. (E) Representative hematoxylin and eosin-stained kidney sections from the Control, DN, and Treatment groups (magnification, ×40; scale bar = 25 µm). (F-I) Relative renal mRNA expression levels of Cat, Maoa, Maob, and Fabp1, respectively, measured by RT-qPCR. #P < 0.05, ##P < 0.01, and ###P < 0.001 versus the Control group; *P < 0.05, **P < 0.01, and ***P < 0.001 versus the DN group. Abbreviations: LLF = Ligustri Lucidi Fructus; DN = diabetic nephropathy; RT-qPCR = reverse transcription quantitative polymerase chain reaction. Please click here to view a larger version of this figure.
| primer | sequences |
| CAT F | TCACTGACGAGATGGCACAC |
| CAT R | ATCGAACGGCAATAGGGGTC |
| FABP1 F | CAATAGGTCTGCCCGAGGAC |
| FABP1 R | GTCATGGTCTCCAGTTCGCA |
| MAOB F | GCACTGAAACAGCCTCACAC |
| MAOB R | TCGTGCAGGGACATCCAAAG |
| MAOA F | ACTTACCCATTCCGTGGTGC |
| MAOA R | ACCACAGGGCAGATACCTCA |
| M-GAPDH F | CCTTCCGTGTTCCTACCCC |
| M-GAPDH R | GCCCAAGATGCCCTTCAGT |
Table 1: Primer sequences used for RT-qPCR analysis of mouse kidney tissues. Abbreviations: F = forward primer; R = reverse primer; RT-qPCR = reverse transcription quantitative polymerase chain reaction.
| MOL ID | Molecule name | OB (%) | DL | Target number |
| MOL000358 | beta-sitosterol | 36.91 | 0.75 | 100 |
| MOL000422 | kaempferol | 41.88 | 0.24 | 103 |
| MOL004576 | taxifolin | 57.84 | 0.27 | 92 |
| MOL005146 | Lucidumoside D | 48.87 | 0.71 | 104 |
| MOL005147 | Lucidumoside D_qt | 54.41 | 0.47 | 0 |
| MOL005169 | (20S)-24-ene-3,20-diol-3-acetate | 40.23 | 0.82 | 0 |
| MOL005190 | eriodictyol | 71.79 | 0.24 | 101 |
| MOL005195 | syringaresinol diglucoside_qt | 83.12 | 0.8 | 0 |
| MOL005209 | Lucidusculine | 30.11 | 0.75 | 105 |
| MOL005211 | Olitoriside | 65.45 | 0.23 | 100 |
| MOL005212 | Olitoriside_qt | 103.23 | 0.78 | 0 |
| MOL000006 | luteolin | 36.16 | 0.25 | 102 |
| MOL000098 | quercetin | 46.43 | 0.28 | 103 |
Table 2: Thirteen active ingredients of Ligustri Lucidi Fructus identified using the TCMSP database. Abbreviations: OB = oral bioavailability; DL = drug-likeness.
| mRNA | Protein | RF | SVM |
| CAT | AQR | 0.7 | 0.98 |
| CAT | RBM22 | 0.8 | 0.97 |
| FABP1 | AQR | 0.65 | 0.94 |
| FABP1 | SF3A3 | 0.7 | 0.8 |
| MAOA | IGF2BP2 | 0.75 | 0.97 |
| MAOA | IGF2BP3 | 0.75 | 0.97 |
| MAOB | TIA1 | 0.85 | 0.89 |
Table 3: Predicted interactions between four mitochondrial biomarker mRNAs and m6A-related RNA-binding proteins. CAT, FABP1, MAOA, and MAOB denote human biomarker mRNAs; AQR, RBM22, SF3A3, IGF2BP2, IGF2BP3, and TIA1 denote RNA-binding proteins. RF and SVM scores > 0.5 indicate predicted RNA–protein interactions. Abbreviations: RF = random forest; SVM = support vector machine.
Supplemental File 1. Bioinformatics scripts and experimental source data. This archive contains the R scripts used for data processing, differential expression analysis, functional enrichment, machine learning, receiver operating characteristic analysis, gene set enrichment analysis, Spearman correlation analysis, and CIBERSORT immune-cell infiltration analysis, together with the source data for blood glucose, urinary microalbumin, and RT-qPCR experiments. Please click here to download this file.