Surgical biopsies were collected from 16 patients who underwent total/subtotal or distal gastrectomy because of GC since 2012 at St. Olav’s Hospital, Trondheim, Norway (Table 1). Four biopsies per patients were taken from tumor and normal tissue and used for clinical pathological evolution and gene expression profiling. The study was approved by the Regional Committees for Medical and Health Research Ethics Central Norway (REK 2012-1029). Furthermore, seven independent datasets of human GC from the TCGA database were used (Table 2).

The mouse model of GC, i.e., the transgenic INS-GAS mice which spontaneously develop gastric cancer, was used (Zhao et al., 2014). Stomachs were collected from 26 mice, i.e., six females and four males INS-GAS mice at age of 15 months and eight females and eight males wild-type (WT) mice at age of 12 months, for gene expression profiling. In addition, 31 INS-GAS mice, i.e., 12 females and nine males at age of 10 months, and 10 WT mice, i.e., 6 females and four males at age of 10 months, were used for in vivo testing.

Animals were housed as four to five mice per cage on wood chip bedding with a 12 h light/dark cycle in a specific pathogen free environment with room temperature of 22°C and 40–60% relative humidity. Animals were inspected daily by investigators and authorized veterinarian using a scoring sheet. Animals should be euthanized at score of 10 if they are emaciated, underconditioned in five consecutive days, or show poor clinical signs (e.g., body weight, appearance, and behavior) before end of the study. This was done according to the Directive 2010/63/EU in which human primary endpoint are defined as “the earliest indicator in an animal experiment of (potential) pain and/or distress that, within the context of moral justification and scientific endpoints to be met, can be used to avoid or limit pain and/or distress by taking actions such as humane killing or terminating or alleviating the pain and distress (Hendriksen and Morton, 1999)” (https://www.humane-endpoints.info/en/council-directive-2010-63-eu). The study was approved by The Norwegian Food Safety Authority (Mattilsynet).

Total RNA was extracted from the surgical biopsies of patients and harvested stomachs of mice. RNA quality and quantity were obtained using NanoDrop One (Thermo Scientific, Norway) and Agilent Bioanalyser. For human samples, RNA microarray of GC samples, including 24 tumors of intestinal, diffuse and mixed types from seven patients and 37 normal tissue from six patients, was performed using Illumina platform as described earlier (Zhao et al., 2014). Illumina microarray data was analyzed using Lumi on the log₂ scale and analyzed using the empirical Bayesian method implemented in Limma. The data is accessible via Mendeley Data repository with DOI link at http://dx.doi.org/10.17632/hzmfshy7hp.1. Illumina identifiers (ILMN) were uploaded to Ingenuity Pathway Analysis (IPA, QIAGEN, Hilden, Germany) together with log₂-fold change, p-values and q-values (false discovery rates). For mouse samples, RNA sequencing was performed using Illumina HiSeqNS500 instrument (NextSeq 500) at 75 bp with paired end (PE) reads using NS500H flow cells with 25M reads/sample. Paired end forward read length (R1): 81, reverse read length (R2): 81. Downstream processing and analysis of the data was performed in the Bioconductor environment in R. For humans, a total of 47,323 transcripts was assigned to analysis in which 37,489 transcripts were mapped and 9,834 transcripts unmapped by Ingenuity Pathway Analysis (IPA) (QIAGEN, Hilden, Germany). For mice, a total of 54,460 transcripts was loaded in which 54,162 were mapped/298 transcripts were unmapped in IPA. For mouse GC after ivermectin treatment, 54,416 transcripts were loaded in which all were mapped in IPA. Filtering of datasets included species (mouse or human) and p-value cut-off (p < 0.05). Gene expression was analyzed using a t-test between tumor and normal tissue in patients, between INS-GAS and WT mice and between INS-GAS mice with and without ivermectin. Genes with a p-value of less than 0.05 were considered to be differentially expressed. Transcriptomics datasets were analyzed using IPA. Molecular networks and canonical pathways were algorithmically constructed based on known connectivity and relationships among genes/proteins/metabolites using Ingenuity Knowledge Base. Local and regulatory z-scores for canonical pathways and diseases and biofunctions that overlapped with the experimental data of the present study were calculated using the formula described previously (Sitarz et al., 2018). IPA has sophisticated algorithms to calculate predicted functional activation/inhibition of canonical pathways, diseases and functions, transcription regulators and regulators based on their downstream molecule expressions (QIAGEN Inc., https://www.qiagenbioinformatics.com/products/ingenuitypathway-analysis). Fischer’s exact test was used to calculate a p-value determining the probability that the association between the genes in the datasets from human GC and mouse GC and the canonical pathway or disease/function by chance alone.

The concept of a Connectivity Map (cMap) was recently developed, whereby genes, drugs, and disease states are connected by virtue of common gene expression signatures (Qu and Rajpal, 2012; Subramanian et al., 2017; Musa et al., 2018). To identify candidate drugs, the gene expression signature of GC was generated based on the gene expression profile of human GC. A positive cMap score indicates there is a positive similarity between a given perturbagen’s signature, i.e., genes that are increased by treatment (in reference datasets) are also upregulated in the human GC dataset, while a negative score indicates that the two signatures are opposing. cMap was performed using the gene expression signature of human GC (n = 7 GC vs. n = 6 normal tissue).

Data mining was performed using the gene expression profile data of 61 samples from 16 patients, 26 samples from 26 mice, and 324 samples from seven independent datasets from the TCGA database. Furthermore, knowledge-based pathway mining was used based on previous studies that showed WNT/β-catenin signaling pathway as one of the important pathways in gastric tumorigenesis (Zhao et al., 2014; Rabben et al., 2021). Custom-made molecular networks were generated using the Ingenuity Knowledge Base. Networks were then algorithmically generated based on their interrelationships. WNT/β-catenin signaling pathway was constructed based on the transcriptomic data of INS-GAS mice and were then entered into the “Pathway” module of the IPA to obtain the nodes in every corresponding signaling pathway. Nodes from pathways were added as entries into the “My list”-function and all entries in the list were added to the “My pathway” in IPA. My pathway was used to produce a network of nodes/genes from the WNT/β-catenin signaling pathway that matched with our experimental data from INS-GAS vs. WT mice. The build-tool was used to connect nodes using edges, i.e., relationships including both direct and indirect interactions like chemical-protein interactions, ubiquitination, molecular cleavage, translocation, localization, phosphorylation, expression, protein-protein interactions, activation, regulation of binding, inhibition, membership, reaction, protein-DNA interactions, transcription and modification. The Canonical Pathway overlay-tool was used to arrange the entries into clusters based on pathway. Local z-scores were calculated in IPA based on the dataset’s correlation with the activated state. Negative z-scores indicate a decrease in activity, positive z-scores indicate an increase in activity. Canonical pathways were identified using statistical cut-offs at p < 0.05.

The expression data from mouse GC was compared to all genes in the pathway. The molecule activity predictor (MAP)-function was used to predict activation/inhibition between the nodes in the network. The in silico tool integrated with the MAP-function was employed to predict effects on the network after gene inhibition and/or stimulation in the ivermectin cluster. Connections between genes were then algorithmically generated based on their interrelationships including both direct and indirect interactions like chemical-protein interactions, ubiquitination, molecular cleavage, translocation, localization, phosphorylation, expression, protein-protein interactions, activation, regulation of binding, inhibition, membership, reaction, protein-DNA interactions, transcription and modification. Network clusters of WNT/β-catenin pathway was constructed based on the transcriptomic data of INS-GAS mice (i.e., limited to and built on genes from the dataset). The build-tool was used to connect nodes using edges, i.e., relationships. Categorical values were set to each gene/node using a semi-quantitative method to quantify the color-change resulting from in silico inhibition. Local z-scores were calculated in IPA based on the dataset’s correlation with the activated state. Negative z-scores indicate a decrease in activity, positive z-scores indicate an increase in activity. Canonical pathways were identified using statistical cut-offs at p < 0.05.

GC cell lines included human gastric cancer cells MKN74 (intestinal type) and KATO-III (diffuse type) (for detailed information on molecular characteristic, see Yokozaki, 2000). It should be noticed that both cell lines overexpress β-catenin (Asciutti et al., 2011). Cells were maintained in RPMI-1640 medium (Sigma Aldrich, Oslo, Norway) supplemented with fetal bovine serum (10%, FBS), Sodium pyruvate and penicillin streptomycin solution (1%) (Sigma Aldrich, Oslo, Norway) in a humidified incubator holding 5% CO₂ and 37°C. For proliferation assay, MKN74 and KATO-III were seeded in 96-well plates (2,500 cells/well and 3,000 cells/well, respectively) and incubated overnight. Ivermectin (MW: 875.09 g/mol) was dissolved in DMSO (100%) to 50 mM stock solution. Cells were treated with ivermectin (0–50 µM) or vehicle control (0.45% v/v DMSO) for 24, 48, and 72 h. Proliferation was measured using a commercial CCK-8 Kit (Sigma Aldrich, Oslo, Norway) with absorbance read at 450 nm. For cell cycle analysis, KATO-III cells were seeded as 3.0 × 10⁵ cells/well in 6-well plates and incubated for 72 h with medium change after 48 h. Ivermectin was added to the wells as final concentrations of 12, 15 or 18 µM for 24 h. Cells were harvested by trypsin, washed twice in room tempered PBS, resuspended in ice cold ethanol (70%) and kept at −20°C for minimum 15 min. Cells were washed twice in cold PBS and centrifuged (1,500 rpm, 5 min, 4°C), and resuspended in freshly prepared PI staining solution (0.25% Triton- X-100, 50 μg/ml propidium iodide (PI) and 200 μg/ml RNAase) for minimum 30 min. Cell cycle distribution was analyzed using FACS. Single cells were gated to exclude doublets and clustered cells. 2.0 × 10⁴ cells were counted per sample, and percentage cell distribution was derived from obtained histograms in the FACSDiva software. Results are presented as means of n = 3 replicates/treatments. Data was analyzed using Microsoft Excel 2010.

Thirty-one INS-GAS mice were randomly divided into two groups: ivermectin treatment (12 females and nine males at age of 10 months) and controls (no treatment, six females and four males at age of 10 months). Ivermectin was reconstituted from lyophilized powder in DMSO to 50 mM solution and then diluted in saline before use. The treatment regimen was designed to let the mice tolerate the procedure easily, i.e., intraperitoneal injection at a dose of 10 mg/kg in a volume of about 0.5 ml/mouse with 27G needle once per day for 5 days, followed by no treatment for 5 days and then injection once per day for 10 days. This regimen was repeated 10 days later. The total duration of treatment was 2 months (2 × 30 days). Vehicle treatment was not performed because neither vehicle per se nor procedure would lead to any significant stress response. The mice were euthanized under isoflurane inhalation anesthesia (2–3%), and stomachs were collected as described previously (Zhao et al., 2014). Tumor volume density (% of glandular area of the stomach occupied by tumor) was measured using a point count method (Zhao et al., 2014). The tissue samples were collected for transcriptomics as aforementioned.

Values are expressed as means ± SEM or SD (stated in individual figure legend). For comparison of two independent groups, student independent t-test was used. For comparison of multiple groups, one-way ANOVA with Tukey’s or Dunnett’s post-hoc tests were used (stated in individual figure legend). SPSS version 26.0 for Windows (SPSS Inc., Chicago, IL, United States) was used and a p-value < 0.05 was considered to be statistically significant. Other methods are stated in corresponding figure legends.

A cMap was created according to the gene expression signature (Figures 2A,B). A total of 2,428 drugs were categorized into 47 groups of inhibitors for, e.g., DNA synthesis, murine double minute (MDM) and lactate dehydrogenase, including ivermectin. Additionally, drugs we have demonstrated previously, including gemcitabine, paclitaxel, everolimus, and scopolamine, were also found (Zhao et al., 2014; Rabben et al., 2021).

Data/pathway mining revealed activation of the WNT/β-catenin signaling in human as well as mouse GC (Table 2). In addition to our own human GC and mouse GC data, seven independent datasets of human stomach adenocarcinoma (STAD) were found to have activation of the WNT/β-catenin (Figure 3A and Table 2). Using the knowledge-based repositioning strategy, nine targets were identified in connection with WNT/β-catenin signaling pathway and proliferation in both human and mouse GC (Figure 3B), i.e., ATP-dependent translocase (Abcb1b), retinol-binding proteins (Rbp), ATP binding cassette subfamily B member 1 (ABCB1), ATP Binding cassette subfamily G member 2 (ABCG2), cytochrome P450 family 3 subfamily A member 4 (CYP3A4), P-glycoprotein (also known as multi-drug resistant protein, MDRP), ATP binding cassette subfamily B member 4 (ABCB4), cytokine, and P2X purinoceptor 7 (P2RX7). Each gene/protein connected to a subset of algorithmically chosen genes based on the Ingenuity Knowledge Base. These genes were collectively activating both WNT/β-catenin signaling and proliferation, resulting in locally activated z-scores (shown in orange) (Table 3).

An in silico interaction network of ivermectin with WNT/β-catenin signaling pathway was constructed (Figure 4A). Inhibition of the targets of ivermectin led to the inhibition of downstream nodes in a “dose-dependent manner” (Figure 4B). It should be noticed that the inhibition of single molecules was not enough to have inhibitory or stimulatory effects on the signaling pathway.

Testing of ivermectin in MKN74 and KATO-III cells showed that there were time- and concentration-depended inhibitions of proliferation by the drug with similar IC₅₀ values for the periods of 24, 48 and 72 h (Figures 5A,B). Furthermore, ivermectin induced cell cycle arrest in a concentration-depended manner (Figure 5C). It should be noticed that ivermectin at IC₅₀ did not affect the cells in S-phase but increased percentage of cells in G₁ while reducing percentage of cells in G₂/M phases. By contrast, higher concentration of ivermectin increased percentage of cells in S phase while reducing the percentage of cells in G₁ and G₂/M phases, suggesting that ivermectin arrested cells at the G₁ phase at IC₅₀ and higher dose of the drug shifted cells to S phase.

A treatment regimen using ivermectin at 10 mg/kg for 2 months was established based on the in silico and pilot experiments. Mice tolerated the treatment well, although some mice had weight loss during treatment (<15%, p > 0.05, two-tailed). The mice had no serious side effects of ivermectin and no mice that were treated with ivermectin were killed according to the human primary endpoints which include stressful behavior, abdominal pain and impaired physical activity. The tumor size was reduced by ivermectin treatment (Figure 6A). Comparison analysis between mouse GC with and without ivermectin treatment revealed 4,112 differentially expressed genes (Figure 6B). The genes involved in WNT/β-catenin signaling pathway were particularly inhibited by ivermectin treatment, as shown by a change in z-scores from 1.151 (mouse GC without treatment) down to −1.789 (mouse GC after ivermectin treatment) (Figure 6C and Table 2) and log₂ fold-changes (Figure 6D vs. 6E).

Expression analysis in IPA revealed that cell proliferation was activated in mouse GC without treatment and inactivated in mouse GC with treatment. On the other hand, cell death including apoptosis was inactivated in mouse GC without treatment but activated in mouse GC with treatment (Figures 7A–D).