Yinqiao Shihu Decoction (YSD) is derived from the Yinqiao Formula, Liuwei Dihuang Pill, and Dendrobium officinale, as documented in The Clinical Handbook of Chinese Prescriptions and previous literature. The decoction consists of nine herbal ingredients: Rehmannia glutinosa (Shudihuang), Cornus officinalis (Shanzhuyu), Dioscorea opposita (Shanyao), Poria cocos (Fuling), Alisma orientale (Zexie), Paeonia suffruticosa Andrews (Mudanpi), Lonicera japonica (Jinyinhua), Forsythia suspensa (Lianqiao), and Dendrobium nobile (Shihu).

The Traditional Chinese Medicine Systems Pharmacology Database (TCMSP) was used to retrieve the chemical constituents of each herb in the formula by searching the keywords “Lonicera japonica” (Jinyinhua), “Rehmannia glutinosa” (Shudihuang), “Paeonia suffruticosa” (Mudanpi), “Cornus officinalis” (Shanzhuyu), “Alisma orientale” (Zexie), “Poria cocos” (Fuling), “Dioscorea opposita” (Shanyao), and “Forsythia suspensa” (Lianqiao). The active components were screened based on pharmacokinetic parameters of Absorption, Distribution, Metabolism, and Excretion (ADME), with Oral Bioavailability (OB) and Drug-Likeness (DL) serving as key indicators. The constituents of Dendrobium (Shihu) were retrieved from the Traditional Chinese Medicine Integrated Database (TCMID) and cross-validated with the TCMSP database to identify bioactive compounds. The chemical structures of these active ingredients were obtained from the databases and redrawn using ChemDraw Ultra 8.0. The corresponding two-dimensional (2D) structures were converted into three-dimensional (3D) conformations using Chem3D Ultra 8.0 and saved in mol2 format. Potential targets of the active compounds were collected from the TCMSP database. Protein names were standardized and converted into corresponding gene names using Perl scripts and the UniProt database.

The GSE62232 dataset, containing data from 10 normal liver tissue samples and 81 Hepatocellular Carcinoma (HCC) tissue samples, was obtained from the Gene Expression Omnibus (GEO) database. Meanwhile, the TCGA-LIHC dataset, including 50 normal liver tissue samples and 374 primary HCC tissue samples, was downloaded from The Cancer Genome Atlas (TCGA) database. Differential expression analyses were conducted on both datasets using the GEO2R web tool and R packages, and the overlapping Differentially Expressed Genes (DEGs) were identified based on the results. To further determine HCC-related targets, the Therapeutic Target Database (TTD), Online Mendelian Inheritance in Man (OMIM), and GeneCards databases were queried using the keyword “Hepatocellular Carcinoma”, and the identified targets were considered as candidate genes. The DEGs and candidate targets were then integrated using the Venny 2.1.0 online tool and intersected with the targets of YSD. The overlapping genes, visualized in a Venn diagram, were identified as the potential key targets of YSD in the treatment of HCC.

To evaluate the potential protein-protein interactions among the identified key targets, these targets were imported into the database for analysis. The resulting interaction data were saved as a TSV file and subsequently imported into Cytoscape software to construct the protein-protein interaction network. The top five core target proteins were then identified based on their degree of connectivity within the network.

Biological pathways perform specific physiological functions through the interactions among their constituent target proteins, providing the mechanistic basis for understanding the clinical manifestations of diseases [5]. The core target genes were imported into the DAVID database, and official gene symbols were used for identification. Gene Ontology (GO) functional enrichment analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis were then conducted, and the top 15 significantly enriched results were selected for visualization.

Based on the KEGG enrichment analysis results, an “Active Compound-Key Target-Pathway” network was constructed using Cytoscape 3.7.2, reflecting the interactions between active compounds and key targets as well as between key targets and pathways. In this network, nodes represent active compounds, targets, and pathways, while edges indicate the interactions among them. The merge function was used to integrate this network with the Protein-Protein Interaction (PPI) network, and topological analysis was performed using the Network Analyzer plugin. Higher degree values indicate greater importance of the target in the therapeutic process of YSD against HCC. Based on the degree values, the top five core target proteins and corresponding core active compounds were identified.

The selected core target proteins were used as receptors, and their corresponding crystal structures were obtained from the Protein Data Bank (PDB). The active compounds of YSD-quercetin, kaempferol, luteolin, wogonin, and β-sitosterol-were employed as ligands, with the previously saved mol2 files uploaded. Molecular docking of the core target proteins and ligands was performed using AutoDockTools 1.5.6. Grid energy calculations were conducted using AutoGrid, and the Genetic Algorithm was applied for docking. Binding energy was used to assess the affinity between the receptor and ligand.

To further elucidate the potential key genes linking YSD and Hepatocellular Carcinoma (HCC) progression, results from multiple analytical layers were integrated. Genes that were simultaneously ranked among the top nodes in the PPI network and included in the “Active Compound-Key Target-Pathway” network were first shortlisted. Among these, those exhibiting strong binding affinities to major active compounds (binding energy < −5.0 kcal/mol) in molecular docking and previously reported to be associated with tumorigenesis or HCC development were prioritized. Based on these criteria, five representative genes—NCOA2, PTGS1, RELA, CCND1, and RAF1—were selected for subsequent prognostic and single-cell analyses.

The prognostic significance of these targets was evaluated using the GEPIA2 online platform. A 95% confidence interval was applied, and the median expression level (50%) was used as the cut-off value to classify patients into high- and low-expression groups. The correlations between the expression levels of NCOA2, PTGS1, RELA, CCND1, and RAF1 and Overall Survival (OS) outcomes in HCC patients were then analyzed.

We analysed three single-cell RNA sequencing (scRNA-seq) samples from the GEO database (GSM8825750, GSM8825751, GSM8825752) belonging to the series GSE290925 (“Single-Cell Transcriptomic Profiling of the Tumor Microenvironment in Treatment-Naive Hepatocellular Carcinoma Patients”). Each sample represents a surgical puncture tumour specimen obtained before any treatment from a Hepatocellular Carcinoma (HCC) patient. The scRNA libraries were constructed using the 10× Genomics Chromium Single Cell 3' Reagent v3 Kit and sequenced on an Illumina NextSeq 500 platform according to the manufacturer’s protocol. Cells with mitochondrial gene content exceeding 25% were filtered out, followed by data normalization using the LogNormalize method, and 2000 highly variable genes were selected. The Harmony algorithm was subsequently applied to correct batch effects. Cell clustering was performed based on Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE). Cell type annotation was conducted using the SingleR package in conjunction with a human primary cell atlas, and the expression patterns of core genes were visualized on t-SNE plots, providing a systematic characterization of single-cell heterogeneity.

Through database retrieval and screening, 2 compounds of Rehmannia glutinosa (Shudihuang), 20 compounds of Cornus officinalis (Shanzhuyu), 16 compounds of Dioscorea opposita (Shanyao), 10 compounds of Alisma orientale (Zexie), 11 compounds of Paeonia suffruticosa (Mudanpi), 15 compounds of Poria cocos (Fuling), 23 compounds of Lonicera japonica (Jinyinhua), 23 compounds of Forsythia suspensa (Lianqiao), and 3 compounds of Dendrobium (Shihu) were identified (two additional Dendrobium compounds with drug-likeness < 0.18 were included due to limited matches across the two databases; see AppendixTableA1). Targets of these active compounds were collected from the TCMSP database and mapped to gene symbols using the UniProt database. After removing duplicates, a total of 120 unique targets were obtained.

Differential expression analysis was performed using sample data from the GEO and TCGA databases, combined with searches in the TTD, OMIM, and GeneCards databases. A total of 949 overlapping differentially expressed genes and 7100 disease-related targets were identified. In the TCGA dataset, 7821 genes were upregulated and 1493 were downregulated (Figure 1(A)). Based on the intersection with the targets of YSD, 101 key targets for HCC treatment were obtained (Figure 1(B), Figure 1(C)).

Figure 1. Screening of active compounds and identification of key targets related to Hepatocellular Carcinoma (HCC). (A) Differentially expressed genes in HCC vs. normal tissues from TCGA, including 7821 upregulated and 1493 downregulated genes. (B) (C) Venn diagram showing overlap between YSD targets and HCC-related genes; 101 intersecting genes were identified as key targets.

The identified key targets were input into the STRING database to obtain the Protein-Protein Interaction (PPI) network, and the resulting interaction data were imported into Cytoscape for network construction (seeAppendixFigure A1(A)). In the network, larger nodes indicate higher degree values. ALB, IL6, CASP3, EGFR, and VEGFA exhibited the highest degrees and were identified as the core target proteins.

GO functional enrichment analysis using the DAVID online database identified a total of 121 GO terms, including 79 Biological Process (BP) terms, 18 Cellular Component (CC) terms, and 24 Molecular Function (MF) terms. KEGG pathway enrichment analysis revealed 75 significantly enriched pathways. The major biological processes involved positive/negative regulation of RNA polymerase II promoter transcription, DNA transcription, negative regulation of apoptosis, and negative regulation of cell proliferation. Cellular components were primarily associated with the nucleus, cytoplasm, extracellular exosome, and cytosol. Molecular functions included transcription factor activity, sequence-specific DNA binding, zinc ion binding, and transcription activator activity. Significantly enriched pathways included PI3K-Akt, TNF, MAPK, cancer-related pathways, and hepatitis B signaling (see AppendixFigures A1(B)-(E)).

The constructed “Active Compound-Key Target-Pathway” network comprised 62 active compounds, 101 target proteins, and 20 pathways. In the network, red nodes represent active compounds, blue nodes represent key targets, and green nodes represent pathways (see AppendixFigure A2(A)). Based on node degree, the top five core targets were identified as MYC, ESR1, CCND1, ALB, and AR (in addition to the core targets identified from the PPI network), while the top five core active compounds were quercetin, kaempferol, luteolin, wogonin, and β-sitosterol, which were selected for subsequent analyses.

The crystal structures of the 10 selected core targets were downloaded from the Protein Data Bank (PDB), and molecular docking was performed between the core targets and core active compounds (see AppendixFigure A1(B)). The results indicated that the five selected core active compounds exhibited low binding energies with most core targets, with β-sitosterol and wogonin showing the strongest binding affinities. Targets including RAF1, PTGS1, and NCOA2 demonstrated favorable binding with the core compounds; these proteins have been previously reported in HCC-related studies, and their molecular docking results showed relatively low binding energies (seeAppendixFigures A1(C)-(F)).

Analysis revealed that the mRNA expression levels of PTGS1 and RELA were significantly associated with Overall Survival (OS) in patients (P < 0.05), indicating that both serve as adverse prognostic factors in HCC (Figure 2).

Figure 2. Survival analysis of RAF1 and EGFR in HCC patients.

Cells with mitochondrial gene content exceeding 25% were filtered out from the single-cell dataset (Figure 3(A)). After filtering, the number of detected RNAs was highly positively correlated with the number of detected molecules (R = 0.91) (Figure 3(B)). Batch effects were subsequently corrected using the Harmony package, and t-SNE plots were generated before and after batch correction (Figure 3(C), Figure 3(D)). The first 10 Principal Components (PCs) were selected for clustering analysis (Figure 3(E)).

Cell type annotation was performed using the SingleR package, classifying the cell populations into T cells, monocytes, macrophages, fibroblasts, tissue stem cells, and NK cells (Figure 4(A)). The results demonstrated that NCOA2, RELA, and RAF1 exhibited the highest average expression levels in monocytes, PTGS1 showed the highest expression in NK cells, and CCND1 displayed the highest expression in fibroblasts (Figure 4(B)).

Figure 3. Quality control and batch correction of single-cell data. (A) Cells with >25% mitochondrial gene content were removed. (B) A strong positive correlation was observed between the number of detected genes and UMIs (R = 0.91). (C) (D) t-SNE visualization before and after batch correction using Harmony. (E) The top 10 Principal Components (PCs) were selected based on variance. (F) Clustering was performed at a resolution of 1 for downstream analyses.