The team, led by researchers at the First Affiliated Hospital of Shihezi University and Shihezi University School of Medicine in China, began by predicting which human proteins artemisinin might interact with. They gathered known pharmacological targets from the HERB database and used the compound’s chemical structure, retrieved from PubChem, to run predictions through the Swiss Target Prediction and SuperPred platforms. After converting identifiers through UniProt and removing duplicates, the search yielded 550 artemisinin-related targets.

To narrow this list to genes relevant to gastric cancer, the authors analyzed RNA sequencing data from the UCSC Xena database, drawing on 624 samples: 414 gastric tumor tissues and 210 normal tissues from GTEx and adjacent normal TCGA samples. Differential expression analysis with the limma package identified 2,892 differentially expressed genes, 1,371 upregulated and 1,521 downregulated, using an adjusted p-value threshold of 0.05 and a log fold-change cutoff of 1. Weighted gene co-expression network analysis, performed on the top 5,000 genes ranked by median absolute deviation, produced 11 co-expression modules. The blue module, containing 1,201 genes, correlated most strongly with gastric cancer (r = 0.91, p < 0.001) and was selected as the key module. Intersecting its genes with the differential expression list produced 507 gastric cancer-related genes, and intersecting those with the 550 artemisinin targets left 27 candidate targets.

The researchers then applied a two-layer screening strategy. First, they built a protein-protein interaction network of the 27 candidates using the STRING database and the CytoHubba plugin in Cytoscape, running three topological algorithms—MCC, MNC, and Degree—and taking the top 10 genes from each. Their intersection yielded seven candidate hub genes: CCNA2, CCNB1, CCNB2, CCNE1, CDC25A, CDK1, and TOP2A. Second, the team used three machine learning algorithms with different mathematical foundations: random forest, which captures nonlinear relationships; LASSO regression, which favors sparse feature selection; and SVM-RFE, which performs robustly on high-dimensional data. Each algorithm classified genes by their ability to discriminate tumor from normal tissue. The intersection of all three methods converged on TOP2A, CCNB2, and CCNA2 as the hub targets. The authors note that this machine learning step prioritizes tumor phenotype drivers and does not by itself establish a drug-target relationship; that link was tested separately by docking and by measuring gene expression after drug treatment.

Enrichment analyses pointed toward cell cycle regulation as the common thread. Gene Ontology analysis showed the candidate targets enriched in biological processes including regulation of cyclin-dependent protein serine/threonine kinase activity and cell cycle G2/M phase transition, while KEGG pathway analysis highlighted the cell cycle, progesterone-mediated oocyte maturation, cellular senescence, and the p53 signaling pathway. In a constructed disease-compound-target-pathway network, the three hub targets occupied central positions, with degree values of 12 for TOP2A, 24 for CCNB2, and 18 for CCNA2. Database validation using UALCAN and TCGA confirmed that all three genes are expressed at significantly higher levels in gastric cancer tissues than in normal gastric tissue, and Human Protein Atlas data showed high protein expression in gastric adenocarcinoma across stages. Gene set enrichment analysis found that tumors with high expression of these genes were enriched in cell cycle-related pathways such as HALLMARK_E2F_TARGETS and HALLMARK_G2M_CHECKPOINT, along with proliferation and metabolism signatures including HALLMARK_MITOTIC_SPINDLE, HALLMARK_MYC_TARGETS_V1/V2, and HALLMARK_MTORC1_SIGNALING.

Notably, Kaplan-Meier survival analysis showed no significant association between the expression levels of the three genes and overall survival in gastric cancer patients. The authors suggest these cell cycle genes may primarily participate in the development and malignant phenotype of gastric cancer rather than serving as independent prognostic predictors, and they caution that clinical prognostic value requires validation in larger cohorts. An immune infiltration analysis based on TIMER2.0 also found that expression of the three hub targets was significantly negatively correlated with most of six major immune cell types, particularly CD4+ T cells, CD8+ T cells, and macrophages, hinting at a possible link to the tumor immune microenvironment—though the authors stress that the underlying regulatory mechanisms and any causal relationship require further experimental validation.

Molecular docking provided a theoretical estimate of whether artemisinin could bind the three proteins. Using the crystal structure of TOP2A (PDB ID: 1ZXM) and AlphaFold-predicted structures for CCNB2 and CCNA2, since no experimentally determined crystal structures were available for those two, the docking produced binding energies of −6.83 kcal/mol for TOP2A, −7.30 kcal/mol for CCNB2, and −6.87 kcal/mol for CCNA2—all below the −5.0 kcal/mol threshold generally considered to suggest potential binding affinity. Artemisinin’s energy with CCNB2 was the lowest, and comparisons with reference compounds showed artemisinin’s TOP2A score was comparable to that of the known inhibitor etoposide (−6.645 kcal/mol in a prior study), while its CCNB2 score was weaker than that of the CDK1 inhibitor seliciclib (−8.7 kcal/mol). The authors are explicit that molecular docking is a theoretical prediction and does not directly demonstrate physical binding; biophysical assays such as CETSA, SPR, or DARTS would be needed to confirm direct drug-protein interactions.

The in vitro experiments used two human gastric cancer cell lines, SGC7901 and SNU-5, at relatively high artemisinin concentrations of 150 μM and 200 μM respectively, chosen from preliminary gradient screening because they inhibited viability without causing complete cell death. The authors acknowledge that such high micromolar concentrations raise concerns about off-target effects, since artemisinin’s endoperoxide bridge can be cleaved in an iron-dependent reaction to generate reactive radicals that may cause non-specific cytotoxicity. They argue, however, that several findings support a targeted action rather than pure non-specific toxicity: artemisinin reduced phosphorylated ERK without changing total ERK levels, downregulated TOP2A, CCNB2, and CCNA2 mRNA, and its effects were partially reversed by the ERK pathway activator C16-PAF. Functionally, CCK-8 assays showed time- and dose-dependent inhibition of viability, EdU staining showed reduced DNA synthesis, colony formation assays showed fewer and smaller colonies, and Transwell assays showed significantly reduced migration through membranes and invasion through Matrigel-coated chambers.

The rescue experiments tied the phenotype to ERK/MAPK signaling. When C16-PAF was added alongside artemisinin, EdU-positive cells increased again, clonogenic capacity was restored, migration and invasion recovered, and Western blotting showed that p-ERK levels and the downstream proteins Ki-67, Vimentin, MMP2, and MMP9 returned to higher levels, while total ERK remained unchanged. The partial reversal suggests ERK/MAPK is one main pathway mediating artemisinin’s effects, but the authors note that other parallel pathways may also be involved, and that the study has not yet demonstrated whether the ERK/MAPK pathway directly regulates transcription of the three hub genes. The limitations are considerable: the work lacks in vivo animal validation, the downregulation of the three targets was confirmed only at the mRNA level and not at the protein level, and direct physical binding was not demonstrated. The authors conclude that the findings are restricted to in vitro models and that mouse xenograft studies are required before any translational claims about artemisinin’s anti-gastric cancer efficacy can be made. Within those bounds, the study offers a template for combining network pharmacology, machine learning, docking, and cell experiments to map how a natural product might act on cancer cells.

Subject of Research: Potential mechanisms of artemisinin against gastric cancer identified through network pharmacology, machine learning, molecular docking, and in vitro validation

Article Title: Exploring the potential mechanism of artemisinin against gastric cancer based on network pharmacology, machine learning and experimental validation

Article References: Hu, C., Zhao, Y., Yun, M., Shen, Y., Lian, W., Liu, B., Li, Y., Deng, C., Hou, B., & He, J. (2026). Exploring the potential mechanism of artemisinin against gastric cancer based on network pharmacology, machine learning and experimental validation. Frontiers in Oncology, 16, Article 1931227. https://doi.org/10.3389/fonc.2026.1931227

Image Credits: AI Generated

DOI: 10.3389/fonc.2026.1931227

Keywords: artemisinin, gastric cancer, ERK/MAPK pathway, network pharmacology, machine learning, molecular docking, TOP2A, CCNB2, CCNA2, cell cycle, in vitro, Artemisia annua