ICC is a highly lethal malignancy that has an increasing incidence [34,35,36]. Advances in surgical management, locoregional therapy, as well as the development of some targeted therapies have improved the treatment of patients with ICC. For example, for the subset of patients with localized disease, surgical resection may be a curative-intent treatment option [37,38]. Unfortunately, most patients present with advanced disease at the time of diagnosis [39], and patients who undergo resection frequently experience early recurrence [16]. As such, improving outcomes for patients with ICC will require advances in systemic therapy. To date, several targeted therapies have been approved for the treatment of advanced ICC. In particular, targeted therapy has been approved for the roughly 15–20% of patients with FGFR2 fusions/rearrangements/deletions, as well as a smaller subset of patients with IDH1 mutations or dMMR [39,40,41]. Unfortunately, only a small minority of patients with ICC have these specific targetable genetic perturbations and are eligible for these therapies. As such, there is an ongoing need to identify pathogenic mechanisms involved in ICC tumorigenesis to inform future novel systemic treatment options.

The rapid advance of sequencing technology has allowed a deeper analysis and understanding of the molecular mechanism of ICC, thus promoting progress in the research on molecular diagnosis and targeted therapy. To that end, in the current study, we used a bioinformatics gene mapping technique to identify deleterious genes specific to ICC. In addition, we then screened drug candidates to identify potential pathways that warrant further investigation of therapeutic effect in ICC. Specifically, differentially expressed genes between ICC samples and normal biliary epithelial cells in GSE32225 and TCGA-CHOL datasets were identified; these data were then used to construct a weighted co-expression gene network by WGCNA. The function annotation and enrichment analysis of the GSE-DEG were examined. Overlapping genes among the three blocks were picked as key genes to establish a PPI network. A total of 10 hub genes were calculated from the network and their association with OS and DFS of ICC patients was examined. In addition, the gene–clinical trait connections were also validated in protein level by the IHC staining of ICC and normal cholangiocytes in HPA. Hub genes were then selected based on interacting perturbagens to generate a list of drug candidates. Of note, three of the top scored drug candidates were tyrosine kinase inhibitors (dasatinib, NVP-BHG712, tivantinib) and two were cannabinoid receptor agonists (palmitoylethanolamide, arachidonamide). Other drug candidates included two antibiotics (moxifloxacin, amoxicillin), one estrogen receptor agonist (levonorgestrel), and one serine/threonine protein kinase inhibitor (MK-2206). The other two were related to cellular biochemical regulation processes, involving one lipid peroxidase inhibitor (anisodamine) and one dihydrofolate reductase inhibitor (pyrimethamine).

The tyrosine kinase family is a large multigene subtype with particular relevance to multiple human diseases, especially cancer. Dasatinib is a multikinase inhibitor, which has been demonstrated to have a favorable therapeutic effect in many malignant diseases [42]. In fact, dasatinib has been reported to have hypersensitivity in ICC with isocitrate dehydrogenase (IDH1/IDH2) mutations [43]. This study identified SRC as a critical dasatinib target in IDH mutant ICC. Collectively, data from the current study, as well as early clinical information from previous reports, suggest that additional dasatinib may be a promising agent in the treatment of ICC [43]. In separated studies, tivantinib, an antiproliferative agent, has been proposed as second-line therapy for hepatocellular carcinoma following disease progression after sorafenib treatment based on receptor overexpression analyses [44]. In fact, data from phase I and II trials have noted promising results [45,46]. In a phase II randomized placebo-controlled trial (NCT00988741), patients treated with tivantinib not only had an increased median OS (7.2 vs. 3.8 months) and RFS (2.2 vs. 1.4 months) versus placebo, but also had an increased median time to progression (TTP) (2.7 vs. 1.4 months) [46]. However, in a phase III study, tivantinib did not lead to improved overall survival [47]. Given the findings of the current study, as well as preliminary HCC data, the clinical application of tivantinib for ICC warrants additional studies to determine its therapeutic effect relative to different targeted biomarkers.

Cannabinoids (CB) have demonstrated the effect to reduce tumor-associated symptoms such as neuropathic pain [48]. Some reports have explored the role of CBs in tumor proliferation and progression, suggesting that CBs interact with CB-Rs, thereby inducing an anti-cancer effect [49]. Endocannabinoids work via particular G-protein-coupled receptors (GPRs), CB-Rs (CB1-R and CB2-R), in which CB2-R has an established role in carcinogenesis and cancer progression [50]. In fact, CBs have been proposed as novel anti-tumor targets in patients with non-small cell lung cancer (NSCLC) and breast cancer [50,51]. While the mechanism remains poorly defined, palmitoylethanolamide (PEA) is an endogenous fatty acid amide related to CB-Rs that facilitates an anti-inflammatory effect [52]. In fact, ultra-micronized PEA (um-PEA) has been shown to inhibit tumor cell proliferation and tumor cell migration both in vitro and in vivo [52]. As such, further investigation of CB-R agonists and their therapeutic anti-proliferative effect relative to ICC may be a potential area to explore.

Estrogen receptors (ERs) are widely expressed in many cell types involved in both innate and adaptive immune responses. ERs are composed of two subtypes: ERα and ERβ. Both receptors interact similarly with endogenous estrogens and play a role in the development and progression of multiple malignant diseases, including breast, prostate, endometrial, ovarian, colorectal and lung cancers [53,54]. Levonorgestrel is usually administered by an intrauterine device that releases consistent localized progestin for the treatment of endometrial cancer [55]. Many studies have shown that ERs are generally expressed in cholangiocytes and demonstrate up-regulation during cell proliferation [56]. The fact that female sex has been previously reported as risk factor for ICC is consistent with the role of estrogen in ICC pathogenesis [57,58]. As such, the targeting of ER pathways may mitigate ICC oncogenesis and future studies will be necessary to examine further the inner connection between ERs and ICC progression.

Our study yielded a new hypothesis that some drug families including tyrosine kinase inhibitor, cannabinoid receptor inhibitor and estrogen receptor inhibitor, may be investigated as candidate drugs for ICC treatment. Due to the heterogeneity and complexity of ICC as well as many other tumors, lack of response to targeted therapies, and frequent existence of drug resistance, the need for novel drug candidate has been largely increased [59]. Ideally, both computational methods and experimental methods should be applied to screen potential drugs. In our study, WCGNA and drug-drug interactions derived computational methods were utilized to help study molecular signature differences between tumor and normal tissue, as well as different classes of drugs in the context of ICC targeted therapy. The approaches we applied in our study can serve as more rapid and effective means for screening and identifying candidate drug combinations for ICC. To achieve better data quality, a series of bioinformatic and statistical methods have been applied in sequencing data interpretation and connectivity map to overcome data noise in biological networks and literature mining results [7,59].

Interpretation of data from the current study should be interpreted in light of several limitations. While data from the current study are important to focus screening and identification of candidate drug combinations to treat ICC, development of any new drug therapies requires further experimental validation and preclinical tests before it can be applied to a clinical scenario. While both computational methods and experimental methods ultimately need to be applied to screen potential drugs, the current study provides the framework for future experiments. In particular, computational screening allows for the identification of potential therapies and better direction/application of future experiments in an informative and cost-saving way. For example, WCGNA and drug-drug interactions derived from computational methods can direct future studies aimed at molecular signature differences between different classes of drugs in specific disease contexts.