We have used public databases in this study and the details are listed as follows:

PubChem (https://pubchem.ncbi.nlm.nih.gov/, accessed on 10 April 2024; CID for AG: 64981).

SwissTargetPrediction (http://www.swisstargetprediction.ch/, accessed on 10 April 2024; version: 2023 release).

Pharmmapper (http://www.lilab-ecust.cn/pharmmapper/, accessed on 10 April 2024; pharmacophore model version: 2014 update).

Uniprot (https://www.uniprot.org/, accessed on 10 April 2024; database release: 2024.01).

TCGA (https://portal.gdc.cancer.gov/, accessed on 10 April 2024; project: TCGA-BRCA, data release: v36).

RCSB Protein Data Bank (PDB) (http://www.rcsb.org/, accessed on 14 April 2024; data retrieved from release archive version dated 10 April 2024).

A PubChem search was conducted using “Arctigenin” to obtain its structural formula and retrieve the SMILES. Subsequently, the MOL2 file was imported into PharmMapper and the SMILES file into the SwissTargetPrediction database to predict the corresponding target proteins of AG. After that, the acquired drug targets were annotated using the UniProt databases.

Transcriptome and clinical data samples of breast cancer patients were sourced from the TCGA database, including 1,109 tumor and 113 normal breast tissue samples. After filtering samples with negative results for estrogen receptor (ER), progesterone receptor (PR), and human epidermal growth factor receptor 2 (HER2), a total of 124 TNBC samples and 113 normal samples were selected. Differentially expressed genes (DEGs) associated with TNBC were identified using the DESeq2 package in R software (version 4.1.3), applying the criteria of: |log2 fold change (FC)| ≥ 1 and adjusted p-value <0.05 (Zhu et al., 2024). The DEGs were visualized through a volcano plot and a heat map, created using the Pheatmap and ggplot2 packages, respectively.

Weighted gene co-expression network analysis (WGCNA) was performed to identify co-expression modules (Langfelder and Horvath, 2008). To enhance result robustness, the top 25% of the most significant DEGs were included in the WGCNA (Zeng et al., 2024). The expression data was first normalized through the normalizeBetweenArrays function, and genes with low variance, specifically those with a standard deviation below 0.5, were excluded (Bourgon et al., 2010). To enhance network reliability, hierarchical clustering and static tree cutting techniques were applied to remove outlier samples. The pickSoftThreshold function was then employed to assess the fit of the scale-free topology model (R²) and the average connectivity across various soft-thresholding powers. The ideal soft-thresholding power was determined using the criterion of R² ≥ 0.9, which was applied to create a weighted adjacency matrix (Zhang and Horvath, 2005). Following this, both a weighted adjacency matrix and topological overlap matrix (TOM) were constructed, leading to hierarchical clustering and dynamic tree cutting to delineate gene modules. Modules were merged based on eigengene correlation threshold of 0.25. The correlation between module eigengenes and clinical traits was evaluated via Pearson correlation, and genes with high module membership (MM) and gene significance (GS) were retained for further analysis (Miller et al., 2011; Yip and Horvath, 2007). By using a Venn diagram to intersect DEGs, WGCNA-derived TNBC-related targets, and AG drug targets, several genes were identified as promising therapeutic candidates for TNBC treatment.

To elucidate the biological functions and key signaling pathways implicated in AG-mediated treatment of TNBC, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses were performed on the overlapping target genes using the clusterProfiler package (version 4.1.3) in R software (Ashburner et al., 2000; Kanehisa and Goto, 2000). Significantly enriched terms were selected by applying a threshold of p-value <0.05 and subsequently ranked in descending order based on their enrichment scores.

Three machine learning algorithms—least absolute shrinkage and selection operator (LASSO), support vector machine-recursive feature elimination (SVM-RFE), and random forest (RF)—were employed to identify hub genes among overlapping target genes (Bao et al., 2023; Sanz et al., 2018; Tang et al., 2023), aiming to selecting genes capable of distinguishing TNBC patients from healthy controls.

For SVM-RFE, a recursive feature elimination framework based on a linear kernel was executed using the e1071 package in R (Huang et al., 2018). The gene expression data underwent z-score normalization, and binary labels were assigned to the groups (Normal vs. TNBC). Feature ranking was conducted using linear SVMs, with the cost parameter set to 10 and the scaling turned off. The elimination process was directed by the squared weight coefficients. A 10-fold stratified cross-validation scheme was employed, batch elimination of 50% was applied when the number of remaining features exceeded 50. Hyperparameter optimization was performed via grid search across gamma = 2^ (−12:0) and cost = 2^ (−6:6) (Huang et al., 2017). The selection of the optimal feature subset was based on achieving the lowest classification error during cross-validation.

In our analysis using LASSO regression, we employed the glmnet package in R (v4.1.3) to construct a binomial logistic model featuring with L1 regularization (alpha = 1) (Friedman et al., 2010). Z-score normalization was applied to gene expression values, and sample groups were binarized. 10-fold cross-validation was used to identify the optimal regularization parameter (lambda.min), defined as the value yielding the lowest mean cross-validated deviance. Genes with non-zero coefficients at lambda. min were selected as potential biomarkers.

The randomForest package (Garge et al., 2013) was utilized to implement the RF model. The model was initially set up with 1,000 trees (ntree = 1,000) using the default parameters. To evaluate the model’s effectiveness, out-of-bag (OOB) error estimates were used, and the ideal number of trees was identified by finding the lowest OOB error. The significance of genes was assessed through the mean decrease in the Gini index, leading to the selection of the highest-ranked genes for further analysis.

Hub genes were identified through the intersection of results derived from three distinct machine learning algorithms. This was followed by an analysis using Receiver Operating Characteristic (ROC) curves to assess their predictive capabilities, along with correlation and differential expression studies to confirm their significance in AG interactions.

The 2D structure of AG was drawn using ChemBioDraw 2014, and its energy minimization was optimized by Chem 3D 2014. The 3D structure of the human kinase proteins were downloaded from the RCSB Protein Data Bank (PDB IDs: 4MXO, 3OP3,8JG8 and 3DB6) (Elling et al., 2008; Levinson and Boxer, 2014; Tang et al., 2024; Tsytlonok et al., 2019). Ligand preparation was performed using the Schrödinger LigPrep module (Schrödinger, LLC, New York, NY, 2018) with the OPLS_2005 force field. The docking box center was defined as the centroid of the original ligand. For the docking task, the standard precision (SP) docking mode of Glide (Friesner et al., 2004; Friesner et al., 2006; Halgren et al., 2004; Yang et al., 2021) was used with the default docking parameters.

Molecular dynamics (MD) was stimulated by AMBER 22 (Case et al., 2005) for the selected ligand-target complex. The AMBER14SB force field (Maier et al., 2015) was used for protein, and the force field parameters of AG were generated based on the general AMBER force field (Wang et al., 2004) by antechamber. The TIP3P water model was used in an octahedral model, with a minimum 12 Å distance between the protein surface and box boundary. Sodium were then added to neutralize the system. Each simulation system was initially energy-minimized to optimize unreasonable atomic contacts and stereochemical conflicts by applying position restraints (force constant of 5.0, 1.0, and 0 kcal·mol⁻¹·Å⁻², respectively) on the backbone atoms and AG. Subsequently, the solvent was heated to 300 K in 50 ps with all solute atoms restrained with a force constant of 10.0 kcal·mol⁻¹·Å⁻². Next, two 50 ps equilibration steps were done in the NPT ensemble with temperature and pressure (1 bar) control by the Langevin thermostat and Berendsen barostat method. Periodic boundary conditions were applied to eliminate boundary effects. All solute atoms were restrained with force constants of 1.0 and 0.5 kcal·mol⁻¹·Å⁻². The production run, with no restraints, was performed for 200 ns. Electrostatic interactions were calculated by the particle mesh Ewald (PME) (Poier et al., 2023). The cutoff distance for nonbonded interactions was 8 Å. The SHAKE algorithm was used to constrain bonds involving hydrogens. The integration time step was set to 2 fs, and conformations were sampled every 10 ps for subsequent analysis.

To acquire more statistically significant results, the binding free energy between SRC and AG was calculated for the 50 ns of MD trajectories every 0.5 ns via the conventional MM/GBSA approach in AMBER tools according to procedures in our previous work (Guo et al., 2012). For each frame, the free energy was calculated for each molecular species (SRC-AG complex, SRC, and AG), and the binding free energy was computed as below (Chong et al., 2009; Kollman et al., 2000): Δ G b i n d = G c o m p l e x − G r e c e p t o r + G l i g a n d (1) Δ G b i n d = Δ H − T Δ S ≈ Δ G g a s + Δ G s o l (2) Δ G g a s = Δ E M M − T Δ S (3) Δ E M M = Δ E int + Δ E e l e + Δ E v d w (4) Δ G s o l = Δ G p o l , s o l + Δ G n p o l , s o l (5) where Δ G b i n d (Equations 1, 2) denotes the binding free energy, and it can be decomposed into two terms: (1) The free energy in a vacuum, Δ G g a s is decomposed into the molecular mechanical energy ( Δ E M M ) and the configurational entropy ( − T Δ S (Equation 3). Δ E M M is the summation of the intramolecular energy ( Δ E int , including bond, angle, and dihedral energies, which is 0 in this study), electrostatic energy ( Δ E e l e ), and van der Waals energy ( Δ E v d w ) (Equation 4). The entropic contribution ( − T Δ S ), which is associated with the conformational entropy loss when a free-state ligand binds to the corresponding unbound-state receptor; (2) the solvation energy ( Δ G s o l ), which is composed of the polar ( Δ G p o l , s o l ) and non-polar contributions ( Δ G n p o l , s o l ) (Equation 5).

SRC protein (MedChemExpress, New Jersey, United States) was diluted to 2 μg/mL with acetate solutions at pH4.5, pH5.0, and pH5.5, respectively. The protein was immobilized on the chip with acetate solutions at pH4.5 and a concentration of 18 μg/mL. One channel of SRC was immobilized before coupling the ligand, the chip was activated for 420 s. The coupling was stopped when the amount reached 13800 RU. Ligand coupling was completed after blocking the chip for 420 s. The actual immobilized amount was 11700 RU. Analyte AG was prepared as the solution with concentrations of 800, 400, 200, 100, 50, 25, 12.5, 6.25, 3.125, and 1.5625 μM by gradient dilution. Multi-cycle kinetics runs were performed with a period of association of 120 s and dissociation of 240 s. The buffer contained 1xPBS (pH 7.4), 0.05% Tween-20, and 5% DMSO. The raw data was imported into BiacoreTM Insight Evaluation Software 4.0, and the multi-cycle kinetics evaluation method was selected for calculating kinetic rate constants. The curve was fitted by the 1:1 binding model with data from 5 chosen concentrations. The association rate constant (K_a), dissociation constant (K_d), and affinity constant (KD) were acquired.

MDA-MB-453 cell line and MDA-MB-231 cell line purchased from the National Infrastructure of Cell Line Resource (NICR, Beijing, P.R. China). The cells lines were cultured in RPMI-1640 medium, supplemented with 10% (v:v) of fetal bovine serum (FBS; Biological Industries, Kinneret, Israel) and 1% penicillin-streptomycin solution (Hyclone|Cytiva, Marlborough, United States) at 37 °C in a 5% CO₂ incubator.

Cell viability was evaluated by Cell Counting Kit-8 (CCK8 assay; MedChemExpress, New Jersey, United States). MDA-MB-453 and MDA-MB-231 cells were seeded into 96-well plates (3 × 10⁴ per well) and pre-cultured for 24 h, then incubated in 200 μL complete medium containing AG (0, 100, 200, 300, 400, and 500 μM) for 24 h, 48 h or 72 h. Subsequently, CCK8 solution was added to each well and incubated in the dark at 37 °C for 1 h. The absorbance was measured at 450 nm on a microplate reader (Eon Microplate Spectrophotometer, BioTek, United States). All experiments were repeated thrice independently. Cell viability (%) = A 450  of drug − A 450  of blank / A 450  of control  –   A 450  of blank × 100   % .

The cell cycle was assessed using a PI staining kit (KeyGEN BioTECH, Nanjing, P.R. China) following the manufacturer’s instructions via flow cytometry. After incubating with PI/RNase A for 45 min in the dark, the cells were conducted using a flow cytometer (Cytoflex, Beckman, United States). Software Modfit LT was employed for cell cycle distribution analysis. A total of 20,000 events from each cell sample were obtained. All experiments were repeated thrice independently.

GreenNuc™ caspase-3 Assay Kit for Live Cells (Beyotime, Shanghai, P.R. China) was subjected to caspase-3 activity analysis following the manufacturer’s instructions. After incubation at room temperature for 30 min, a fluorescence microscope (OBSERVER D1/AX10 cam HRC, Zeiss, Germany) was used to observe green fluorescence.

MMP assay kits with JC-1 (Beyotime, Shanghai, P.R. China) were purchased to detect MMP of MDA-MB-453. The fluorescence signals of AG-treated or untreated cells were detected via flow cytometry. All experiments were repeated thrice independently.

Apoptosis was assessed by an annexin V-FITC/PI staining kit (KeyGEN BioTECH, Nanjing, P.R. China) following the manufacturer’s instructions. After incubating with 5 μL Annexin V-FITC and 5 μL PI for 15 min in the dark, the cells were conducted using a flow cytometer (Cytoflex, Beckman, United States) and analyzed by software FlowJo V10. A total of 20,000 events from each cell sample were obtained. All experiments were repeated thrice independently.

Total proteins were lysed from AG-treated cells (100, 300, and 500 μM) or untreated cells in RIPA buffer with the protease inhibitor cocktail and PMSF (Beyotime, Shanghai, P.R. China). The concentration of proteins in each group was tested by the BCA Protein Assay Kit (Beyotime, Shanghai, P.R. China). Protein extracts were quantitated and loaded on 8%–12% sodium dodecyl sulfate-polyacrylamide gel, then electrophoresed and transferred onto a PVDF membrane (Beyotime, Shanghai, P.R. China), which was blocked in 5% skimmed milk for an hour. The membranes were incubated with primary antibody overnight at 4 °C. The primary antibodies used were anti-caspase-3, anti-caspase-9, anti-Bax, anti-Bcl-2, anti-CDK2, anti-cyclin A2, Anti-P27, anti-SRC, anti-AKT1, anti-pAKT1, anti-ERK1/2, anti-pERK1/2, anti-PI3K(p85), anti-p-PI3K(p85), anti-MEK1/2, anti-p-MEK1/2and anti-β-actin antibodies (human anti-rabbit, 1:1,000; CST, Massachusetts, United States). Then, the membranes were washed and incubated with horseradish peroxidase (HRP)-conjugated secondary antibody (Goat anti-rabbit, 1:100; ZSGB-BIO, Beijing, P.R.China) for 1 h. The positive signal on the membranes was detected by an enhanced chemiluminescence detection kit (Beyotime, Shanghai, P.R. China). Band intensities were scanned and quantified by NIH ImageJ software.

Data were presented as the mean ± SD. Differences between data groups were evaluated using the one-way analysis of variance followed by the Dunnett test, using GraphPad Prism 8 software. P < 0.05 was considered as a statistically significant result.

The molecular structure is depicted in Figure 2a, and a total of 183 potential targets for AG were identified on the PharmMapper and Swiss Target Prediction database (Supplementary Table S1). Differential expression analysis of the TCGA dataset revealed 5193 DEGs, which were visualized using a heatmap and a volcano plot (Figures 2b,c; Supplementary Table S1). To explore the molecular mechanisms underlying TNBC, we constructed a gene co-expression network via WGCNA. The scale-free topology fit index (R²) and mean connectivity were assessed with the pickSoftThreshold function across a range of powers (β = 1–20). As shown in Figure 2d, the R² value surpassed the recommended threshold of 0.90 at a power of 3, while maintaining acceptable mean connectivity, justifying its selection as the ideal soft-threshold for network construction. Genes were clustered and partitioned into modules using the dynamic tree cut method (Figure 2e), resulting in 12 distinct gene modules, with the cluster dendrogram shown in Figure 2f. Additionally, a network heatmap was created to visualize the correlations among genes within each module (Figure 2g). Subsequently, the analysis of module-trait relationships indicated that the turquoise module had the strongest association with tumor/normal control phenotypes (Figure 2h), showing a significant positive correlation (cor = 0.997) between gene significance for TNBC and module membership (Figure 2i). These results suggest that the 6173 genes in the turquoise module are associated with TNBC and may serve as a reservoir of candidate genes for further prioritization (Supplementary Table S1). By intersecting the DEGs, WGCNA key module genes, and the predicted targets of AG, we identified 28 overlapping genes (Figure 2j; Supplementary Table S1), proposed as potential therapeutic targets for AG in TNBC treatment.

We performed GO and KEGG pathway enrichment analyses on these 28 intersecting target genes. The GO analysis results are presented in Figure 3a. The potential target genes were predominantly enriched in the following biological process (BP) terms: positive regulation of the cell cycle, G2/M transition of mitotic cell cycle, cell cycle G2/M phase transition, regulation of G2/M transition of mitotic cell cycle, regulation of nuclear division, and regulation of the cell cycle G2/M phase transition. Regarding cellular components (CC), the enriched entries included the spindle pole, spindle microtubule, spindle, spindle midzone, and mitotic spindle pole. For molecular functions (MF), the enriched terms included protein serine/threonine kinase activity, protein serine kinase activity, transmembrane receptor protein tyrosine kinase activity, histone kinase activity, and protein tyrosine kinase activity. In addition, based on the KEGG pathway enrichment analysis, these intersecting target genes were mainly enriched in signaling pathways such as the Cell cycle, Progesterone-mediated oocyte maturation, Oocyte meiosis, EGFR tyrosine kinase inhibitor resistance, and Gap junction (Figure 3b).

To further determine the critical hub genes in TNBC treatment using AG, we set the capability to discriminate between TNBC samples and non-TNBC samples in the TCGA dataset as the evaluation criterion and filtered the 28 intersecting target genes using three machine learning algorithms. We narrowed down 28 overlapping target genes through the application of three machine learning techniques (Supplementary Table S2). The SVM-RFE method revealed 23 core target genes (Figures 4a,b). Meanwhile, LASSO analysis highlighted 13 of the 28 genes as significant core targets (Figures 4c,d). The RF algorithm provided variable importance scores for all potential target genes, leading to the identification of 10 core genes based on these scores (Figures 4e,f). By computing the intersection of these machine-learning-predicted core target genes, 4 genes (AURKA, SRC, PLK1, and CDC25C) were identified as the target hub genes for TNBC treatment with AG (Figure 4g). Subsequently, we conducted gene expression analyses on different samples for these five target hub genes (Figure 4h). The results showed that these genes are closely interrelated, with their expression levels significantly higher in TNBC tissues compared to non-TNBC tissues. To explore the diagnostic efficacy of the 4 hub genes, ROC curve analysis was performed, with hub genes exhibiting an AUC value >0.7 considered diagnostic markers. In the TCGA dataset, the AUC values were 0.833 for SRC, 0.994 for AURKA, 0.983 for CDC25C, and 0.989 for PLK1 (Figure 4i).

Molecular docking analysis was performed to validate the interaction between AG and the identified hub genes, with the most prominent binding interactions visualized in Figures 5a-d. According to established criteria, ligand-receptor binding affinities are considered biologically significant when the binding energy falls below −5 kcal/mol. Remarkably, AG demonstrated strong binding affinity with three critical targets: SRC (−7.777 kcal/mol), PLK1 (−7.690 kcal/mol), and AURKA (−7.685 kcal/mol), as detailed in Table 1. By integrating both docking scores and Glide emodel scores, SRC emerged as the most promising therapeutic target and was therefore prioritized for experimental validation. To validate the docking protocol, known SRC inhibitors were redocked, and the root mean square deviation (RMSD) values of the redocked structures compared to their original conformations in the protein data bank (PDB) were calculated. The RMSD values for PDB IDs: 1Y57, 2H8H, 3EL8, and 4MXO were found to be 1.94 Å, 1.08 Å, 0.80 Å, and 2.50 Å, respectively (Supplementary Table S3). These RMSD values can be used to assess the reliability of the docking method; generally, lower RMSD values (considered reliable when less than 2 Å, though it depends on the specific research system) indicate that the docking results are relatively credible, providing a methodological validation basis for subsequent molecular docking.

To investigate the binding pattern of SRC and AG, MD simulations and SPR were conducted. The overall binding mode is shown in Figure 6a. The oxygen atom of AG forms hydrogen bonds with the oxygen atom in the M341 and the nitrogen atom in the G344 of SRC. The formation of these hydrogen bonds may influence the affinity of AG to SRC, thereby affecting its biological activity. The binding free energy provided showed the relative binding strengths of SRC-AG complex (Table 2). Using the MM/GBSA methods, the ΔEvdw was calculated to be −34.85 ± 5.19 kcal/mol for the SRC-AG complex, which contributed to the main part. On the other hand, the electrostatic energy was calculated to be −17.97 ± 5.35 kcal/mol. The total free binding energy was calculated to be −26.66 ± 3.75 kcal/mol for the complex. The binding free energy decomposition analysis identified key residues, as shown in Figures 6b,c. The analysis revealed specific key residues (per-residue contribution less than −1 kcal·mol⁻¹) (7 residues), highlighting the role of amino acids such as G274, V281, M314, S345, G344, and L393 in recognizing AG.

For a deeper investigation of the interaction between AG and SRC, SPR-based binding analysis was performed. The full SPR sensorgrams and corresponding fitting residuals are shown in Figures 6d,e, respectively. A 1:1 Langmuir binding model was employed for kinetic fitting, with a Chi-squared (χ²) value of 0.604 RU² indicating a good fit. The kinetic parameters, detailed in Table 3, revealed that AG binds to SRC with an equilibrium dissociation constant (KD) of 71.8 μmol/L, an association rate constant (ka) of 6.95 × 10² M⁻¹·s⁻¹, and a dissociation rate constant (kd) of 4.99 × 10⁻² s⁻¹. These values suggest a moderately strong binding affinity and validate the specific interaction between AG and SRC.

The cell viability of MDA-MB-453 cells incubated with a series of concentrations of AG for 12 h, 24 h, and 48 h was examined (Figure 7a). Compared with the control, the viability of AG-treated cells was negatively correlated with the concentrations of AG, as well as the duration of incubation. The AG-treated group showed a low G0/G1 ratio and an increased ratio of S-phase cells. No significant change was observed in the ratio of G2/M phase cells (Figure 7b), suggesting AG might induce cell cycle arrest to inhibit the proliferation of tumor cells by blocking the process of DNA replication. Similar inhibitory effects were noted in MDA-MB-231 cells as seen in MDA-MB-453. A marked decline in cell viability was observed with higher AG concentrations and extended treatment durations (Supplementary Figure S1a). Cell cycle analysis revealed that AG treatment led to a decrease in the G0/G1 phase cell population while increasing the S phase, with no significant changes in the G2/M phase (Supplementary Figure S1b). Expression of cyclins, CDK2, and Cyclin A2 decreased, while the level of P27 showed no significant change, suggesting AG-induced DNA replication arrest was CDK2/Cyclin A2-targeted (Figure 7c). Interestingly, we also found increased Cyclin E1 expression after AG treatment, which might have accounted for apoptosis (Figure 7d).

To verify the AG-induced cytotoxicity, MDA-MB-453 cells were employed to apoptosis indications after AG incubation for 48 h. Detected MMP showed depolarization, andcaspase-3 activity was promoted, which both were AG concentration-dependent, suggesting the apoptosis event was AG-relevant (Figures 8a,b). The Annexin-FITC/PI-staining flow cytometry results showed that AG treatment increased death rate in a concentration-dependent manner (Figure 8c). It is worth noting that the number of cells treated with low concentrations of AG showed no significant differences in early apoptosis (100 and 200 μM) but mostly varied in late apoptosis or necrosis. Apoptosis-related protein indicators showed that Bcl-2, pro-caspase-9, and pro-caspase-3 were downregulated, yet cleaved caspase-9 and cleaved caspase-3 increased in a concentration-dependent manner (Figure 8d). In MDA-MB-231 cells, the evaluation of caspase-3 activity showed a positive correlation between the number of apoptotic cells and AG concentration (Supplementary Figure S1c). Further confirmation from Annexin V-FITC/PI staining flow cytometry indicated that AG treatment also increased apoptosis in a concentration-dependent manner (Supplementary Figure S1d). Apoptosis-related protein analysis showed significant downregulation of Bcl-2, pro-caspase-9, and pro-caspase-3 with increasing AG concentrations, while Bax expression remained unchanged between treatment and control groups (Supplementary Figure S1e). The results suggested AG might trigger intracellular caspase cascades. Combined with depolarized MMP, AG-induced apoptosis in TNBC cell lines might be achieved via a caspase-dependent mitochondrial apoptosis pathway.

To investigate the AG-mediated SRC-related protein expression, MDA-MB-453 cells were incubated with AG for 48 h. The results showed that the levels of t-AKT and p-AKT decreased. The level of p-ERK1/2 attenuated, hence indicating an entirely decreased ERK1/2 phosphorylation in the cytosol and effective AG-evoked blockage on ERK1/2 activation (Figure 8e). Meanwhile, as shown in Figure 8f, there were no notable changes in the levels of PI3K and MEK1/2 following AG treatment compared to the control group, while p-PI3K and p-MEK1/2 levels were markedly lower. A decrease in SRC levels was also noted, suggesting that AG modulates the PI3K-AKT and MEK/ERK pathways by downregulating SRC expression. Furthermore, we assessed the expression of SRC, AKT, p-AKT, ERK1/2, p-ERK1/2, MEK1/2, p-MEK1/2, PI3K, and p-PI3K in MDA-MB-231 cells. The results were largely consistent with those observed in MDA-MB-453 cells, showing significant reductions in SRC, p-PI3K, p-AKT, p-MEK1/2, and p-ERK1/2 in the AG-treated groups (Supplementary Figures S1f,g).