Literature-based translation from synthetic lethality screening into therapeutics targets: CD82 is a novel target for KRAS mutation in colon cancer.

Synthetic lethality (SL) is an emerging therapeutic paradigm in cancer. We introduced a different approach to prioritize SL gene pairs through literature mining and RAS-mutant high-throughput screening (HTS) data. We matched essential genes from text-mining and mutant genes from the COSMIC and CCLE HTS datasets to build a prediction model of SL gene pairs. CCLE gene expression data were used to enrich the essential-mutant SL gene pairs using Spearman's correlation coefficient and literature mining. In total, 223 essential trigger terms were extracted and ranked. The threshold of the essential gene score ( S g ) was set to 10. We identified 586 genes essential for the SL prediction model of colon cancer. Seven essential RAS-mutant SL gene pairs were identified in our model, including CD82-KRAS/NRAS, PEBP1-NRAS, MT-CO2-HRAS, IFI27-NRAS/KRAS, and SUMO1-HRAS gene pairs. Using RAS-mutant HTS data validation, we identified two potential SL gene pairs, including the CD82 (essential gene)-KRAS (mutant gene) pair and CD82-NRAS pair in the DLD-1 colon cancer cell line (Spearman's correlation p-values = 0.004786 and 0.00249, respectively). Based on further annotations by PubChem, we observed that digitonin targeted the complex comprising CD82, especially in KRAS-mutated HCT116 cancer cells. Moreover, we experimentally demonstrated that CD82 exhibited selective vulnerability in KRAS-mutant colorectal cancer. We used literature mining and HTS data to identify candidates for SL targets for RAS-mutant colon cancer.

Introduction

KRAS is the most frequently mutated oncogene in human cancers. Most KRAS-mutant cancers depend on the sustained expression and signaling of KRAS, making this gene a high-priority therapeutic target. Nevertheless, developing small-molecule direct inhibitors of the KRAS function remains an ongoing challenge [1]. An alternative therapeutic strategy for KRAS-mutant malignancies involves targeting codependent vulnerabilities or synthetic lethal partners essential for oncogenic KRAS [2]. Synthetic interactions between mutations in two different genes were first identified in Drosophila by Dobzhansky [3] in 1946. Synthetic lethality (SL) is defined as the biological consequence for a pair of viable genes if a cell with a mutation of either gene remains alive, but mutations or malfunctions of both genes would lead to cell death. SL has recently emerged as a novel strategy for the treatment of cancer. The interpretation of SL stipulates the downstream effects of two mutually expressed genes capable of performing the same essential function. In this regard, inhibiting a single gene is viable, whereas inhibiting both is lethal. The first clinical trial of SL-based treatment investigated BRCA1/2-deficient cancers, which have an SL interaction with poly (ADP-ribose) polymerase (PARP) [4], [5]. Small PARP inhibitors are currently being designed for breast cancer treatment and are undergoing evaluation in clinical trials for BRCA1/2-deficient cancers [6]. Another example is Werner syndrome helicase (WRN) and high microsatellite instability (MSI-H) in cancer. Defects in DNA mismatch repair promote a hypermutable state referred to as microsatellites, which are unstable. Microsatellite instability contributes to the occurrence of several cancers, including colon (15 %), gastric (22 %), endometrial (20 %–30 %), and ovarian cancers [7]. WRN, a synthetic lethal partner of MSI-H, was identified using the DepMap database [8] and in an in vitro study [9].

KRAS activates numerous effector pathways that contribute to many potential synthetic lethal partners. A major challenge is the selection of appropriate candidate targets from thousands of SL gene pairs to determine the mechanism and potency of anti-cancer drugs. Many public SL databases have been employed to identify appropriate SL gene pairs. The dominant method, high-throughput screening (HTS) data surveys [10], enables the identification of many SL interactions in human cancer. However, the screening-based approach in human cancer is fraught with technical issues, including inconsistent cell lines and mechanistic interpretations, leading to false-positive SL candidates. In addition to these obstacles, most known SL cases have been discovered in yeast cells, but only a few SL gene pairs have been identified in humans. Indeed, it is challenging to select drug target candidates from these HTS-derived SL gene pairs simply based on their priority by ranking the score or signal of a reporter gene. Although more than 12 000 HTS-derived human SL gene pairs have been listed in the SynLethDB database [11], very few cases have become drug targets for cancer therapeutic discovery. Instead of identifying new SL gene pairs, we adopted a different approach by translating HTS-derived SL gene pairs into potential anti-cancer drug targets at the cellular level.

In this study, we aimed to identify novel therapeutic targets of essential genes for RAS-mutant colon cancer based on the concept of SL. We designed and implemented a text-mining model and experimental biological data to predict the critical SL gene pairs for specific cancer types. Moreover, we validated the anti-cancer potential of SL-based therapeutic targets through RAS-mutant colon cell line testing and a bioassay database [12], [13].

Materials and methods

Study design and workflow

Our study comprised four main themes based on literature mining and HTS data. We designed an SL prediction model based on a text-mining method, which was modified in our previous study [14]. Fig. 1 depicts the proposed model architecture, which consists of four parts. (A) Essential genes were first identified by text-mining and HTS. We extracted a candidate list of essential genes using databases and biomedical literature mining from PubMed. (B) Essential SL gene pairs for cancer-specific types were matched and predicted. Since SL was identified by linking phenotype with genotype [10], we hypothesized that cell death could be mediated by an essential gene that was dysregulated at conduction, whereby another gene was mutated. Therefore, an SL gene pair could be modeled by building an association between an essential gene and a particular mutant gene. Cell lines were selected in the SL experimental screening approach based on whether they had one mutated gene (gene X). Next, the importance of other genes in these cell lines was tested. A potential SL (gene X and gene Y) gene pair could be predicted if one gene (gene Y) was essential in the cancer cell lines that had mutations in a specific gene (gene X) but was not essential in other cell lines. (C) SL gene pairs were enriched and filtered using CCLE data and literature. (D) Appropriate candidate (i.e., CD82-KRAS) SL gene pairs were selected using RAS-mutant HTS and PubChem bioassay data [13].

Fig. 1

Overview of the synthetic lethality (SL) prediction model. A. Identification of essential and mutant SL genes using text-mining and high-throughput screening (HTS) data (COSMIC and CCLE). B. Matching and prediction of essential-mutant SL gene pairs for cancer-specific types. C. Enrichment and filtering of SL gene pairs by gene co-expression and co-occurrences. D. Selection of candidate (CD82-KRAS) SL gene pairs using RAS-mutant HTS and PubChem bioassay data.

Trigger term mining and ranking

To extract the biological relationships between genes and diseases from the literature, we identified essential genes using trigger terms as the main concept behind our text-mining method. First, we extracted trigger terms regarding “essentiality.” A word was considered a trigger term if it was a common ancestor of two entities (referring to gene and disease herewith) in the dependency parse tree. We used the Stanford natural language processing (NLP) tool [15] to parse the dependency tree in the present study. COLT-Cancer [16], including breast, ovarian, and pancreatic cancers, was used as our initial essential gene database.

To rank the trigger terms, we divided sentences containing one gene name and one disease name into two sets: (i) essential relations, comprising sentences in which the gene was annotated as an essential gene in the disease; and (ii) other relations, comprising sentences in which the gene was not annotated as an essential gene in the disease. Subsequently, we used the following equation to compute the reliability score of each trigger term (): represents the set of all cancer types in the essential database, is the number of co-occurrences of trigger term and cancer type in essential relations, is the number of co-occurrences of trigger term and cancer type in other relations, is the maximum number of co-occurrences in essential relations for disease , and is the maximum number of co-occurrences in other relations for disease .

Essential gene extraction and scoring

After extracting the essential trigger terms, we extracted potential essential genes with the essential trigger terms. Each potential essential gene had a common essential trigger term with other diseases. To filter out false essential genes, we computed and assigned a score () to each essential gene candidate using the following equation:where denotes the score of the essential gene candidate in disease , denotes the set of trigger terms, is the score of trigger term , and is the number of co-occurrences when trigger term is the common ancestor of gene and disease .

Matching essential-mutant SL gene pairs for colon cancer

We matched essential-mutant SL gene pairs in colon cancer after identifying cancer-specific essential genes using the text-mining method. We derived our mutant gene data from the well-known COSMIC [17] and CCLE datasets [18]. Using the COSMIC dataset, we selected gene names and tumor locations (e.g., the large intestine). We then identified each tumor location corresponding to the cancer type (e.g., colon cancer). Using the CCLE dataset, we obtained the pre-processed dataset containing mutant genes in specific cells using the cell line named entity recognition tool [19]. We then identified each cell line name corresponding to the cancer type and matched the essential-mutant SL gene pairs.

Enrichment and filtering of SL gene pairs

To extract more reliable predicted SL gene pairs, we used three criteria to rank the predicted SL gene pairs: (i) essential gene candidate score (), (ii) gene co-expression, and (iii) number of co-occurrences (Fig. 2). We ranked each SL gene pair according to its essential gene candidate score () for the first criterion. We enriched each SL gene pair for the second criterion according to the gene co-expression obtained from the CCLE gene expression data; Spearman’s correlation coefficient and p-value were used. For the third criterion, we filtered each SL gene pair according to the number of co-occurrences of two genes in one article.

Fig. 2

Workflow of enrichment and filtering of synthetic lethality (SL) gene pairs. To extract and rank more reliable SL gene pairs, we used three criteria: (i) Essential gene candidate score (), gene co-expression, and co-occurrences to enrich and filter SL gene pairs. In total, 586 essential genes were identified based on the threshold of the (ii) essential gene candidate score () > 10. Subsequently, we enriched and set Spearman’s correlation (iv) p-value < 0.05 according to gene co-expression obtained from (iii) CCLE gene expression data. Next, we filtered each SL gene pair by co-occurrences (v) in the literature. The threshold of (vi) co-occurrences of two genes in one article was > 0 or ≥ 0. , essential gene score.

RAS-mutant HTS database

Two RAS-mutant HTS datasets, including the HCT-116 cell line (KRAS mutation) with the NAE1 gene study [20] and the DLD-1 cell line with RAS genes (HRAS, NRAS, and KRAS) study [21], were used to validate the SL gene pairs. The DLD-1 cell line is a genome-wide RNAi screen that permits the identification of multiple synthetic lethal interactions with the RAS oncogene. HTS experimental scores with log2 ratio differences were calculated to determine the correlation with RAS gene status [22].

PubChem bioassay database

PubChem is an open database of chemical and biological functional assays maintained by the National Institute of Health (NIH) that allows users to search for chemicals by name and identify chemical and physical properties biological activities safety and toxicity information. The literature search identified compounds targeting proteins (reporters) of the pathway or phenotype in a cell line under a particular culture condition. We annotated the compounds for SL (i.e. CD82-KRAS) candidate targets in the PubChem database [13] using the search keywords “digitonin” and “colon cancer.” Using the COSMIC database we further identified the type of mutations in the KRAS gene in colon cancer cell lines.

Cell culture and cell viability assay

The isogenic colon cancer cell lines DLD-1 W/− (W: RAS wild-type, HD 105–002) and DLD-1 W/M (M: KRAS p.Gly13Asp mutation, HD PAR-086) [22] were obtained from Horizon Discovery Ltd. (Cambridge, UK). Cells were seeded in 96-well plates at a density of 100 cells/well in a complete medium and incubated overnight. A series of 5-FU concentrations (0 μM, 0.5 μM, 1 μM, and 10 μM) was added to the cultures, and the cells were incubated for a further 72 h. Cell viability was analyzed using the MTT assay and ATP bioluminescence assay kit (Promega, Madison, WI, USA).

Results

Identification of essential genes by trigger term mining in colon cancer

For the literature collection, we first downloaded literature published before 1 February 2016 from PubMed using Entrez Programming Utilities (E-utilities). We identified 783 975 articles that mentioned genes and 637 932 articles that mentioned specific cancers.

Our approach involved mining essential genes with trigger terms and parsing the dependency tree using the Stanford NLP tool [15]. Supplementary Fig. S1A presents an example sentence, “Her-2/neu gene amplification in familial vs sporadic breast cancer” (PMID: 14671981). In this example, “neu” is an essential gene in “breast cancer,” and “amplification” is the common ancestor in the dependency parse tree. A total of 223 essential trigger terms were extracted and ranked (Supplementary Table S1). In total, 57, 52, and 99 essential trigger terms were extracted and ranked for breast, ovarian, and pancreatic cancers, respectively (Supplementary Tables S2, S3, and S4, respectively). Supplementary Fig. S1B presents the top seven trigger terms for breast, ovarian, and pancreatic cancers.

The was considered high if the trigger term appeared more in essential relations than in other relations. The frequency of the trigger term was considered higher in other relations than in essential relations when < 1. This study used  > 1 to reduce any false-positive errors.

Our essential gene database consisted of three cancer types: breast, pancreatic, and ovarian. We selected two cancers as the training data and the third as the test data in this experiment. Supplementary Fig. S1B presents the precision/recall/F-score results for pancreatic, ovarian, and breast cancers. Supplementary Fig. S1B depicts the results tested on breast cancer and trained on two other cancers with different thresholds of the , whereby a higher threshold reflects higher precision but a substantially lower recall. To reduce the number of false essential genes extracted by trigger terms and remove nonsignificant genes, we set the threshold of the to 10 (Fig. 2).

For colon cancer-specific text-mining, Fig. 3A presents the dependency parse tree for “significance of the member of TM4SF (MRP-1/CD9, KAI1/CD82, and CD151) in human colon cancer” (PMID: 12838318). In this example, “KAI1/CD82” is an essential gene in “colon cancer,” and “significance” is the common ancestor in the dependency parse tree. Fig. 3B depicts the top seven trigger terms in colon cancer, including amplification, testing, alpha, brca1, evaluation, xenografts, and tumors. We identified 586 essential genes for SL gene pairs with  > 10 in colon cancer (Supplementary Table S5).

Fig. 3

Dependency parse tree of an example sentence and trigger term extraction. A. The dependency parse tree for “significance of the member of TM4SF (MRP-1/CD9, KAI1/CD82, and CD151) in human colon cancer.” “KAI1/CD82” is an essential gene in “colon cancer,” and “significance” is the common ancestor in the dependency parse tree. B. The top seven trigger terms in colon cancer. The nmod, dep, and conj. are Universal Stanford Dependencies representing grammatical relations between words. Abbreviations. nomd, nominal modifier; the nmod relation is used for nominal dependents of another noun or noun phrase and functionally corresponds to an attribute or genitive complement. Dep, unspecified dependency; a dependency can be labeled as dep when it is impossible to determine a more precise relation. Conj, conjunct; a conjunct is a relation between two elements connected by a coordinating conjunction, such as and, or, etc. The links for the main organizing principles of the Universal Dependencies taxonomy were as follows: https://universaldependencies.org/u/dep/all.html (accessed on 2 Feb 2022). , trigger term score.

Enrichment and filtering of matched essential-mutant SL gene pairs by gene co-expression and co-occurrence in colon cancer

We selected the appropriate essential-mutant SL gene pairs from our SL prediction model and matched the COSMIC/CCLE genetic mutation datasets with specific cancer types (Fig. 1). After selecting a suitable threshold, we used the CCLE dataset to enrich the SL gene pairs by co-expression. The SL gene pairs were identified by a cutoff Spearman’s p-value < 0.05 (Fig. 2), and each SL pair was ranked by co-occurrence in the literature. Fig. 4 presents the number of gene pairs between our prediction model and the screening data in Venn diagram format with different thresholds. For colon cancer, we set a Spearman’s correlation p-value < 0.05 and the threshold of the number of co-occurrences of two genes in one article ≥ 0 (Fig. 2), as shown in Fig. 4A (left). Our SL model predicted 16 SL pairs. However, we only set Spearman’s correlation p-value < 0.05 and the threshold of the number of co-occurrences of two genes in one article > 0, as shown in Fig. 4B. We identified seven appropriate candidates for SL gene pairs in our model (Supplementary Table S6), including CD82-KRAS/NRAS, PEBP1-NRAS, MT-CO2-HRAS, IFI27-NRAS/KRAS, and SUMO1-HRAS SL gene pairs. The appropriate SL candidates were selected by enriching and filtering.

Fig. 4

Identification of KRAS-mutant synthetic lethality (SL) gene pairs. The synthetic lethal relationships of these genes were identified using RAS-mutant high-throughput screening (HTS) data and text-mining analysis. Predicted SL gene pairs were compared with HTS data using a Venn diagram with the threshold of gene co-expression and co-occurrences in colon cancer (A and B). The left circle (red) denotes the number of predicted SL gene pairs; the right circle (blue) represents the number of SL gene pairs recorded in the screening data. A. Stricter thresholds were set for gene co-expression (Spearman correlation p-value < 0.05) and co-occurrence (number of co-occurrences of two genes in the literature ≥ 0). B. Stricter thresholds were set for gene co-expression (Spearman correlation p-value < 0.05) and co-occurrence (number of co-occurrences of two genes in the literature > 0). C. Two potential SL gene pairs in colon cancer, including CD82 (essential)-KRAS (mutant) and CD82 (essential)-NRAS (mutant) gene pairs, were identified. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Validation of CD82-KRAS/NRAS SL gene pairs using RAS-mutant HTS data

Focusing on RAS mutations, we investigated SL in colon cancer using HTS screening data from the HCT116 cell line (KRAS mutation) with the NAE1 gene study [20] and the DLD-1 cell line with RAS genes (HRAS, NRAS, and KRAS) study [20]. We validated notable cases in the intersection between our prediction model and the RAS-mutant HTS data. We identified two potential SL gene pairs, including the CD82 (essential gene)–KRAS (mutation gene) and CD82 (essential gene)–NRAS (mutation gene) pairs in colon cancer (Spearman correlation p-values = 0.004786 and 0.00249, respectively) (Fig. 4C and Supplementary Table S6). For gene co-expression, Spearman’s correlation coefficients between CD82 and KRAS mutations and between CD82 and NRAS mutations were −0.329 and −0.355, respectively. The HTS experimental score for CD82 and KRAS/NRAS was −1.23 in RAS-mutant HTS data [21]. The data revealed negative co-expression of CD82 and KRAS/NRAS mutation genes (Fig. 4C and Supplementary Table S6). As KRAS and NRAS are oncogenes and CD82 is a tumor suppressor gene, the results imply that the CD82-KRAS/NRAS mutation is a synthetic dosage-lethality pair. In this regard, the CD82-KRAS/NRAS genetic interaction, involving an under-expression of CD82 combined with KRAS/NRAS mutations, could lead to cancer cell death.

Our approach enabled matching essential-mutant SL gene pairs from the biomedical literature and COSMIC and CCLE datasets. For instance, KRAS is a mutant gene in colon cancer, and CD82 is an essential gene extracted using the text-mining method. Thus, we predicted and verified that KRAS and CD82 constituted an SL pair in colon cancer. Various unknown SLs were extracted in screening experiments. We selected key SLs from the screening data, and our prediction system permitted the selection of the most important SL gene pairs.

Identification of KRAS mutation drug candidates using a bioassay database

To investigate whether CD82-KRAS SL gene pairs had previously been reported in bioassay studies, we searched the bioassay database in PubChem (Fig. 5A), with CD82 as the target. The search revealed a tetraspanin-cholesterol interaction via digitonin [12]. “Digitonin” was identified as an active substance in three bioassays, in which HT29 (AID: 417513, IC50 (half maximal inhibitory concentration) = 5.6 µM), CC20 (AID: 417511, IC50 = 9.6 µM), and HCT116 (AID: 417512, IC50 = 8.7 µM) were the target cell lines [13], [23] (Fig. 5B). Among the cell lines, HCT116 carries the KRAS-mutant (G13D) gene (reported in the COSMIC database) (Fig. 5B), and HT29 carries the BRAF mutation (V600E), a downstream gene of KRAS within the same pathway. These results implied that depletion of tetraspanin family members (such as CD9, CD82, and CD151) by digitonin suppressed KRAS/BRAF-mutant colon cancer cell lines and that CD82 was a target of digitonin, especially in KRAS-mutated HCT116 cancer cells.

Fig. 5

PubChem bioassay of the CD82-KRAS synthetic lethality (SL) gene pair in colon cancer. A. A PubChem database search revealed that a chemical compound screen assay (fluorometric microculture cytotoxicity assay (FMCA)) demonstrated that digitonin was used to precipitate tetraspanins (a protein complex including CD82) and significantly suppressed the growth of a KRAS/BRAF-mutant colon cancer cell line (https://pubchem.ncbi.nlm.nih.gov/bioassay/471512). B. The half-maximal inhibitory concentration (IC50) values for three cancer cell lines are presented.

Knockdown of CD82 expression in KRAS-mutant isogenic cell lines

The RAS-mutant DLD-1 isogenic cell line was used to validate the biological relevance of CD82 in SL. CD82, a tetraspanin family member, contributes to the chemoresistant phenotype of cancers [24]. To experimentally corroborate the CD82 dependency of KRAS-mutant cancer cells, we applied siRNA-mediated knockdown of CD82 and assessed the interaction of CD82 and KRAS based on MEK activation in CRC cell lines (Fig. 6). Knockdown of CD82 resulted in a modest reduction in MEK phosphorylation in KRAS-mutant cells compared to that in KRAS wild-type and control cells. CRC cell lines were transfected with specific siRNA targeting CD82, and the impact on cell viability following 5-FU chemotherapy was determined using the MTT assay. Treatment with CD82 siRNA significantly reduced the level of p-MEK in DLD-1 W/M cells compared to that in cells treated with scrambled siRNA (Fig. 6A). DLD-1 W/− cells transfected with CD82 siRNA exhibited a significant increase in p-MEK levels. A greater reduction in cell viability was observed in CD82 siRNA-treated DLD-1 W/M cells than in DLD-1 W/− cells (Fig. 6B and Supplementary Table S7). In the bar graph, siRNA targeting CD82 affects the viability of isogenic CRC cells following 5-FU chemotherapy using the MTT assay. A series of doses of 5-FU was administered to cancer cells (0.5 μM, 1 μM, and 10 μM). The viability of DLD-1 W/M cells treated with CD82 siRNA was significantly lower than that of scrambled siRNA-treated cells. The p-values for 5FU concentrations of 0.5 μM and 1 μM are 1.06E-07 and 1.63E-05, respectively (Supplementary Table S7). There was no significant difference in cell viability at high doses of 5-FU (10 μM). CD82 depletion by siRNA selectively impaired the viability of KRAS-mutant cancer cell lines. These results indicated that CD82 dependency was associated with KRAS mutation status in cancer cells.

Fig. 6

Effects of siRNA-mediated CD82 depletion on cell viability of KRAS-mutant cancer cell lines. siRNA-mediated knockdown of CD82 resulted in the downregulation of p-MEK and decreased cell viability following chemotherapy. A. Western blot (WB) analysis of DLD-1 W/M (KRAS mutant) and DLD-1 W/− (KRAS wild-type). CD82 expression was higher in DLD-1 W/M cells than in DLD-1 W/− cells. The colon cancer cell lines were transiently transfected with control siRNA or CD82 siRNA. The WB results of siRNA-CD82 in DLD-1 W/− and W/M cells. B. Results of the MTT assay after CD82 knockdown and treatment with 5-FU for 72.

Discussion

We conducted a comprehensive SL prediction model and identified CD82-KRAS SL gene pairs in colon cancer. Our results highlighted the following important points: (i) We designed and implemented a text-mining model to predict more potential SL gene pairs in cancer-specific types. (ii) We compared potential SL gene pairs with screening data and identified the CD82 gene as a novel target for colon cancer with KRAS/NRAS mutation by synthetic lethality. (iii) We demonstrated digitonin as a potential therapeutic agent via the bioassay database for KRAS-mutant colon cancer cell lines. (iv) Finally, the CD82-KRAS is an essential SL gene pair validated in vivo.

KRAS and NRAS are well-known mutated genes in most cancers, including leukemia and colon, pancreatic, and lung cancers. However, no studies to date have provided direct evidence that these CD82 mutations constitute SL gene pairs. Functional genetic screening approaches, including RNAi or CRISPR-Cas9, have been used for KRAS synthetic lethal targets [2]. Previous studies used this approach to identify therapeutic targets for a cancer type with a specific mutant gene using HTS, resulting in SL databases and information resources. Several screening technologies have been developed to detect SL interactions in yeast cells [25], human cell lines [26], [27], and malignant tissues [28]. Several published algorithms use cancer genomic and multiomic data to predict SL interactions [29], [30]. However, their scope remains insufficient to include the comprehensive functional and experimental studies of genetic interactions that need to be evaluated. In this regard, a literature-based translation from SL screening to therapeutic targets may provide more functional and experimental data for SL candidates.

In order to identify and find relationships between essential genes and specific cancers, we used text-mining to analyze their co-occurrence frequency in biomedical literature.

The literature mining and screening-based approach for identifying SL gene pairs in human cancer have several issues that should be addressed, such as false-positive results and inconsistencies for different cancer cell types. Many SL gene pairs in HTS experimental data have been identified and ranked; however, there is a paucity of appropriate SL candidates for cancer therapeutic discovery. As an alternative to the HTS experimental score, we introduced literature mining as a distinct approach to prioritize SL candidates and identified three SL gene pairs. With further annotations based on public data sources of gene expression and bioactivity, a CD82-KRAS SL gene pair in colon cancer was identified. This permitted the selection of novel SL gene pairs, although they were not included in the top-ranking candidates in the HTS experimental data. Moreover, we experimentally demonstrated that the knockdown of CD82 increased the sensitivity of KRAS-mutant colon cells in response to regular cancer drug treatment with 5-FU. Our approach will enable researchers to narrow the size of notable SL pairs. Indeed, our literature-based identification of anti-cancer targets from HTS data and bioassays provides an alternative method to unveil novel modes of action for modern cancer therapeutics.

Compared with the data mining SL identification pipeline (DAISY), another data-driven SL prediction system [31], our approach permitted the identification of more SL gene pairs in different cancers because we extended essential genes with diseases using text-mining in the published literature. Limited experimental data were available on essential genes (147 cell lines) compared to mutant genes (approximately 1600 cell lines). We filtered with a stricter threshold regarding co-occurrence in the literature to remove incorrect SL gene pairs and enriched with gene co-expression using the CCLE database. With advances in treatable KRAS mutations for cancer therapy, cancer-specific SL should be considered in SL prediction [32]. DAISY predicted potential SL using mutant genes, essential genes, and gene expression in non-specific cancers. In addition to the quantity of experimental data, only three cancer types (breast, ovarian, and pancreatic cancers) have been examined in experiments on essential genes. Owing to the lack of experimental data on essential genes in colon cancer, we utilized the trigger term-based method to extend the quantity of potential essential genes and identified two notable SL gene pairs at the intersection of RAS-mutant screening data in colon cancer. Furthermore, this method used the co-occurrence of entities and trigger terms to identify additional relationships between essential genes and specific cancers.

Recently, several drugs that directly target KRAS have been investigated in clinical trials. Sotorasib can be used to treat patients with non-small cell lung cancer with a particular KRAS mutation subtype, referred to as the G12C mutation [33]. However, a single-arm phase II trial reported a lack of efficacy in CRC patients with the KRAS (G12C) mutation [34]. In this regard, the same mutations in cancer cells in the colon or lung may lead to different therapeutic outcomes. The spectrum and distribution of KRAS mutation subtypes also differ for different cancer types. Despite the clinical success of targeting KRAS mutations in patients with lung cancer, additional efforts are warranted to identify drugs to treat CRC patients with KRAS mutations.

In this study, we identified seven appropriate candidates for RAS-mutant SL gene pairs and verified that CD82-KRAS/NRAS gene pairs had SL in a bioassay database of colon cancer cell lines. Digitonin (CID: 25444), a cholesterol-precipitating reagent, can inhibit tetraspanins CD9, CD81, and CD82 [11] and be used to verify CD82-KRAS as an SL gene pair via bioassay. We searched the PubChem database with the keywords “digitonin,” “KRAS,” and “colon cancer,” leading to three digitalis glycosides candidate bioassays as the output. The three digitalis glycosides include digitonin, digitoxin (CID: 441207), and digoxin (CID: 2724385) (Fig. 5A) [35], [36]. Digitonin, digitoxin, and digoxin were identified as active substances in HCT116 cancer cells (IC50 = 8.7 µM, 0.74 µM, and 0.27 µM, respectively). As the three digitalis glycosides share similar 2D structures, they may be able to target CD82 leading to SL in KRAS-mutant colon cancer cells. However, no literature reports an association between CD82 and digitoxin or digoxin. Furthermore, digitonin is an active compound in RAS-mutant colon cancer cell lines [13]. Lower CD82 expression and high miR-633 are biomarkers of a worse prognosis for melanoma combined with CRC cancers [37]. The miR-633 molecule is another possible CD82 target. According to bioinformatics analyses and in vitro studies, miR-633 could target and regulate CD82 expression [38]. Therefore, the inhibition of miR-633 results in decreased cell viability and migration, indicating that it may be a potential target for CD82.

CD82, a tumor suppressor gene, is downregulated during tumor progression in human cancers and can be activated by p53 [39]. CD82 functions with BCL2L12 via the AKT/STAT5 signaling pathway in acute myelogenous leukemia cells [40]. CD82 may interact with the KRAS gene through p53 or may be directly associated with growth factors [41]. Except for bioassay data in PubChem [12], no literature to date has directly demonstrated the CD82-KRAS SL pair in colon cancer. To confirm our SL prediction model and test the sensitivity of cancer cells to drug inhibition, we established isogenic mutation model cell lines, including DLD-1 KRAS wild-type (W/−) as a positive control and DLD-1 KRAS mutation (W/M) siRNA for gene knockdown in stabilized cancer cell lines. We administered a drug to inhibit receptor tyrosine kinase signaling in KRAS-mutated cells and knocked down essential genes to identify the proteins involved in downstream signaling pathways using western blotting. Finally, based on cell viability, we demonstrated CD82-KRAS SL in DLD-1 colon cancer cells.

This study aims to identify SL pairs through different angles to prioritize the SL candidates. To demonstrate the co-dependency of CD82 and KRAS mutation in colorectal cancer cells, we used isogenic DLD-1 cancer cell lines, including DLD-1 W/M and DLD-1 W/−, in our study. Fig. 5A compares baseline CD82 expression in scrambled siRNA-treated DLD-1 W/M and DLD-1 W/− cancer cells. When KRAS mutations are knocked out in DLD-1 W/− isogenic cancer cells, CD82 expression and p-MEK decrease compared to KRAS-mutant colorectal cancer cells DLD-1 W/M. Using DLD-1 isogenic cell lines, we demonstrated the co-dependency between CD82 and KRAS mutations. However, our study has several limitations. Firstly, we did not use more cancer cell lines as positive or negative controls to validate our findings. Through RNAi screening of different cell lines, Barbie et al. demonstrate that TBK1 is required for oncogenic KRAS-driven cancers [42]. We used the RAS-mutant HTS database with a genome-wide RNA interference screen in DLD-1 W/M to validate the SL gene pairs. Secondly, the performance of our text-mining method may be improved using other methods or data resources. Thirdly, some SL pairs are presented comprehensively, such as the BCL2L1-KRAS-mutant SL pair in colorectal cancer [2]. However, we briefly introduced text-mining into the literature rather than comprehensively; therefore, our study may not have identified additional SL pairs. Lastly, in most cases, automatically generated trigger terms were shown to yield higher recalls but lower precisions.

Conclusions

This study developed an SL prediction system based on a text-mining method. We identified and validated CD82-KRAS SL using cell-based experiments. Our findings highlight CD82 as a novel target for KRAS mutations in colon cancer and demonstrated digitonin as a potential therapeutic agent via the bioassay database for KRAS-mutant colon cancer cell lines. Accordingly, further studies using digitonin are warranted.

Declarations

Consent for publication: All authors agree to publication.

Availability of data and material: The data generated in this study are available in Supplementary Table S1-S7. Publicly available data used in the study are listed in the key resources table. The code used for the analysis in this paper is available online at https://github.com/imwilly3737/SL_prediction_by_literature. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Funding: This work was supported in part by the , Taiwan, under Research Grant of MOST 111-2634-F-006-002, MOST 111-2634-F-006-007, and MOST 110-2634-F-006-020, Ministry of Health and Welfare (MOHW111-TDU-B-221-014005), and (NCKUH-11102061).

Authors’ contributions: Conception and study design: HT Yang, MY Chien, JH Chiang, and PC Lin; Development of methodology: HT Yang and MY Chien; Acquisition of data: HT Yang, MY Chien, and PC Lin; Statistical and computational analysis: HT Yang, MY Chien, and PC Lin; Writing, review, and/or revision of the manuscript: HT Yang, MY Chien, and PC Lin; Study supervision: HT Yang and PC Lin.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.