Cisplatin resistance-related multi-omics differences and the establishment of machine learning models

Details of cell lines information were downloaded from Cancer Dependency Map (Depmap, depmap.org) and Cancer Cell Line Encyclopedia (CCLE, https://​portals.​broadinstitute.​org/​ccle/​data), including IC50 value, cell line source, somatic mutation, mRNA expression, miRNA expression, and metabolite. The information of lung adenocarcinoma patients treated with cisplatin was downloaded from The Cancer Genome Atlas (TCGA, https://​gdc.​cancer.​gov/​) (TCGA-LUAD) with their gene expression data. As the CCLE, Depmap, TCGA databases are open to the public under specific guidelines, it confirms that all written informed consents were obtained before data collection.

Differential analysis of somatic mutation, RNA, miRNA, and metabolite data between low IC50 and high IC50 groups was performed with R (version 3.6.1). Maftools, the R package, was used to summarize, analyze and visualize the somatic mutation data. RNA, miRNA, and metabolite data were first normalized and standardized by constructing relevant expression matrices using edgfR after removing those without enough sequence fragments in the sample. Differential genes, somatic mutations, miRNAs (P < 0.05, false discovery rate (FDR) < 0.05) were sorted according to logFoldChange values (|logFC|> 1) to identify significantly different expressions. All the differential analyses were presented in a heat map and volcano plots.

The characteristic of LASSO regression is to consider both Variable Selection and Regularization when fitting a generalized linear model, for applying which, the "Glmnet" (Lasso and Elastic-Net Regularized Generalized Linear Models) R package via penalized maximum likelihood fitness was used, and the mRNAs and miRNAs mostly relative to the resistance to cisplatin were obtained.

Logistic regression, classification and regression tree (CART), and C4.5 decision tree classification algorithm, which use occurrence ratio to determine the category of the dependent variable, was based on the results of LASSO. Finally, we use stepwise regression to select variables, and contribute the predictable model.

NSCLC A549 and H358 cells were obtained from the American Type Culture Collection (Manassas, VA, USA). Cells were fostered in RPMI-1640 containing 10% fetal bovine serum (FBS) and 100 μg/mL of penicillin–streptomycin with or without DDP (Sigma-Aldrich, Merck KGaA, Darmstadt, Germany) added into the culture medium for incubation in a humid atmosphere containing 5% CO₂ at 37 °C.

Cell proliferation was evaluated by Cell Counting Kit-8(CCK-8; Dojindo, Kumamoto, Japan). Briefly, 2 × 10³ of A549 and H358 cells were plated in 96 well plates. They were incubated with 100 μL RPMI-1640 containing 10% fetal bovine serum (FBS) and 100 μg/mL of penicillin–streptomycin for 24 h, then with or without DDP (Sigma-Aldrich, Merck KGaA, Darmstadt, Germany) for another 24 h at 37 °C. After treatment, cells were incubated in 10% CCK-8 reagent. The OD value was measured after 2 h at 450 nm with a microplate reader from Bio-Rad (Microplate reader 3550-UV).

siRNAs targeting BATF3, IRF5, ZBTB38, and Silencer Negative Control siRNAs were purchased from Ribobio (sequences provided in Additional file 6: Table S1). We purchased two different siRNAs for each gene to avoid the off-target effects. A549 and H358 cells were seeded in 6-well plates for 24 h prior to transfection with siRNA targeting BATF3, IRF5, ZBTB38, and corresponding non-targeting controls. A total of 150 nM of siRNA was added to each experiment made up of the target siRNA and topped up with the appropriate concentration of non-targeted controls where appropriate. Transfections were carried out in OptiMem medium (Gibco) using Lipofectamine 8000 transfection reagent (Beyotime). 48 h post-transfection, cells were harvested and assayed for RNA and protein expression levels of the target of interest. At the same time, corresponding samples were treated as described in the text.

To detect the expression of BATF38, IRF5, ZBTB38 in A549 and H358 cell lines, RT-qPCR was carried out on an QuantStudio^® 5 real-time PCR system (Applied Biosystems) with proper PCR parameters.

Total RNAs were extracted by TRIzol (TIANGEN, Beijing, China). The first-strand cDNA was synthesized using Hifair® III 1st Strand cDNA Synthesis SuperMix for qPCR (gDNA digester plus) (YEASEN, Tokyo, Japan) according to the manufacturer's instructions. Then Hieff^® qPCR SYBR Green Master Mix (Low Rox Plus) (YEASEN) was used with the following PCR parameters, 1 cycle of 30 s at 95 °C, 40 cycles of 5 s at 95 °C and 34 s at 60 °C. β-actin was used as the reference. Primers used in this study are listed in Additional file 7: Table S2.

All the samples were repeated three times.

Proteins of A549 cells and H358 cells after RNA interference were extracted using RIPA (Beyotime, Shanghai, China) with protease and phosphatase inhibitor cocktail (Topscience). Then these proteins were quantified by Enhanced BCA Protein Assay Kit (Beyotime). Proteins were then resolved, separated, and finally transferred into PVDF membranes under the influence of an electric current in a procedure (Merck-Millipore, Burlington, MA, USA). Membranes were blocked, followed by incubation with specific primary antibodies [11].

Finally, we observed the protein bands by Moon chemiluminescence kit (Beyotime). The following antibodies were used: Rabbit anti-BATF3 (NBP2-41296, dilution 1:1,500, Novus Biologicals); Rabbit anti-IRF5 (CY5822, dilution 1:1,500, Abways); rabbit anti-ZBTB38 (21906–1-AP, dilution 1: 1,500, Proteintech), mouse β-ACTIN (1:3,000, AA128, Beyotime), horseradish peroxidase (HRP)-labeled goat anti-rabbit IgG (H + L) (1:3,000, A0208, Beyotime), and HRP-labeled goat anti-mouse IgG (H + L) (1:3,000, A0208, Beyotime).

We used the same methods described in our previous studies to test cisplatin-sensitivity-related genes [12].

This study was approved by the Ethics Committee of Zhongshan Hospital, Fudan University (B2019–137R). Patients had signed the informed consent at hospitalization.

661 cell lines were finally enrolled with the complete data of mRNA, microRNA expression, and metabolite. They were divided into 3 groups according to the IC50 value, the half inhibitory concentration, given by the CCLE database, which reflects the sensitivity to chemotherapeutic drugs. The lower the value, the stronger the sensitivity. In this study, the top 1/3 (220) with a low IC50 value (IC50 ≤ 10.4) were defined as the sensitive group, while the last 1/3 (220) with a low IC50 value (IC50 ≥ 26.2) were enrolled in the insensitive group (Fig. 1).

Obviously, the resistance to cisplatin varies among different cancers. In the Depmap database, we successfully classify different tumor cell lines according to their source of cancer, so as to obtain the average IC50 value of cisplatin of related cancers. As shown in Additional file 1: Fig. S1, the average IC50 value of 6 cancers were more than 100, including Lung adenocarcinoma (LUAD), Mesothelioma (MESO), Thyroid carcinoma (THCA), etc., which means patients with these types of cancer may suffer a higher possibility of resistance to cisplatin, one of the most common chemotherapy drugs.

The overall pattern of somatic mutation of the cell line in detail was described in Additional file 2: Fig. S2. After matching the somatic mutation data with the drug sensitivity data, the differences in the characteristics of somatic mutation were investigated between the low IC50 (218 in 220) and high IC 50 (220 in 220) groups. As shown in Fig. 2A and B, although two somatic mutation oncoplots and co-occurring sets of these genes are similar to a certain extent, it is not difficult to find that in the cisplatin-resistant group (high IC50), the mutation of TP53 ranked the first (10%), which is higher than the mutation rate in the cisplatin-sensitive group (low IC50 group). As previously reported by Aditya Bagrodia [13], TP53 mutations are only found in cisplatin-resistant tumors, especially in primary mediastinal non-seminoma. Therefore, we might conclude that TP53 mutation is closely related to cisplatin resistance. In contrast, MUT5 had a higher mutation rate in the low IC50 group. We also checked for drug-gene interactions compiled from Drug Gene Interaction database with "drugInteractions" function. Most drugs are related to TTN and TP53, which is less different between the two groups of cells.

To further discover the differences between the 2 groups, differential analysis between the two cell lines was applied from different levels. First, at the mRNA level, through edgfR, we found 1348 mRNA were differentially expressed between two groups (P < 0.01), including 1 downregulating in the cisplatin resistance group (SLFN1) and 35 upregulating (TMEM4SB, CLDN3, KLK6, IRF5, ZBTB38, etc.) with an obvious fold change (|logFC| > 1) (Fig. 3A). Similarly, at the miRNA level, as shown in Fig. 3B, 80 differentially expressed miRNA were discovered (P < 0.01) and 11 downregulating (miR-194, miR-206, miR-215, etc.,), 12 upregulating (miR-144, miR-16, miR-129-3p, etc.,) with an obvious fold change (|logFC| > 1). However, for the metabolizes, only the expression of 15 metabolites, such as urail, serine, carnosine, N-carbamoyl-β-alanine, cystathionine, etc., showed significant differences between the two groups, with no obvious fold change (|logFC| < 1) (Fig. 3C). For the methylation, none of the differential methylation sites with obvious fold change (|logFC| > 1) were closely related to the differentially expressed genes (Fig. 3D).

We firstly performed a Gene Ontology (GO) analysis of the 1348 mRNAs that were found differentially expressed before based on the CCLE data. Gene functional enrichment analysis showed that 20 GO functional groups exhibited significant differences between the low IC50 group and the high IC 50 group, including actin-binding, extracellular matrix structural constituent, GTPase binding, guanyl-nucleotide exchange factor activity, etc. (Fig. 5A). For Kyoto Encyclopedia of Genes and Genomes (KEGG) based on these mRNAs (Fig. 5B), 10 pathways differed between two groups: PI3K-Akt signaling pathway, which had the most significant difference and the highest gene ratio, followed by proteoglycans in cancer, focal adhesion, etc.

Next, to further conclude the difference of pathways and functions, we applied the Gene set variation analysis (GSVA), which takes specific gene sets as a characteristic expression matrix and quantifies the results of gene enrichment (Fig. 5C). In all, we concluded 643 gene pathways, most of which were related to drug resistance, including Riggins tamoxifen resistance, EGFR signaling 24HR, gefitinib resistance, KIM-MIC amplification targets, etc.

Firstly, the glmnet R package was applied with "family = binomial", which is suitable for binary discrete dependent variables to determine whether the DEGs were related to the high or low IC 50. 6 of the 1348 mRNAs, FOXA2, BATF3, SIX1, HOXA1, ZBTB38, IRF5, selected as the associated features with cisplatin resistance (Fig. 4A, B), and then related logistic model was established through glm [14] function (Fig. 4C).

Next, we applied J48 and rpart R package with ""formula = ic50 ~ FOXA2 + BATF3 + SIX1 + HOXA1 + ZBTB38 + IRF5"" and ""family = gaussian"", which is suitable for continuous univariate as we aimed to get the predicting value of IC50 value through this model (Additional file 3: Fig. S3). The step function is used to achieve stepwise regression. From the results, it is found that 3 in 6 variables had passed the significance test (p < 0.05) and became relatively important variables for constructing the model. While each variable in the model passed the significance test, it is necessary to ensure that the entire model is significant. Therefore, the Chi-Squared test was performed on the model. As the variables were added to the model one by one from the first to the last, the model finally passed the significance test, indicating that the model composed (Table 1) of these variables is meaningful and correct.

Finally, we testified whether these genes are not only related to cisplatin resistance, but also have connection with patient survival. In addition, according to the data in the TCGA database, their expression levels were related to the prognosis of different tumor patients, including breast cancer, ovarian cancer, lung cancer and gastric cancer to a certain extent (Fig. 5). Indeed, BATF3, IRF5 and ZBTB38 also contributed in evaluation of the response to chemotherapy of breast and ovarian cancer patients (Additional file 4: Fig. S4).

Similarly, we composed the miRNA predicting model. From the LASSO regression, 4 significant miRNAs were concluded, miR-203, miR-200c, miR-148a, miR-142-5p (Additional file 5: Fig. S5). After the step function, miR-203 and miR-200c left, both of which passed the significance test conducted by Chi-Squared test, indicating that the model composed (Table 1) of these variables is meaningful and correct.

To test the causal relationship between BATF3, IRF5, ZBTB38 expression and sensitivity to cisplatin on the cell level, we transfected BATF3-targeting, IRF5-targeting, and ZBTB38-targeting siRNAs separately in human cancer cell lines, A549 and H358 (non-small-cell lung cancer), which on the one hand expresses high BATF3, IRF5, ZBTB38 levels. These two cell lines in the result of previous sequencing analysis, and on the other hand, were constantly used in our lab. All siRNA-mediated silencings proved to be highly efficient and were sustained for 3d or more. IRF5 and ZBTB38-silenced cells exhibited at least a reduced sensitivity to cisplatin compared with mock-silenced cells, while the BATF3-silenced cells exhibited increased sensitivity to cisplatin (P ≤ 0.01, Fig. 6A, B).

A total of 6 ESCC (Esophageal Squamous Cell Carcinoma) patients were selected, who accepted complete cisplatin-containing neoadjuvant chemotherapy treatment, including 4 neoadjuvant-chemotherapy insensitive (NACT-NON-SEN) tumor samples and 2 neoadjuvant-chemotherapy sensitive (NACT-SEN, complete response [CRs]) tumor samples (Fig. 7A). RECIST standard was used to evaluate the effectivity of NACT effect.

Similarly, 5 LUAD patients were selected, 3 of which were NACT-SEN (3 CRs, 1 PRs) and 2 were NACT-NON-SEN (Fig. 7B).

In LUAD, BATF3 significantly down-regulated in the NACT-SEN group, while the expression of IRF5, ZBTB38 in the NACT-SEN group were significantly higher than the NACT-NON-SEN group in malignant epithelial cell cluster (marked with EPCAM and SOX4, Fig. 7D). In ESCC, BATF3, IRF5, ZBTB38 showed the same tendency in the 9801 malignant cells marked by LYZ and C1QB (Fig. 7C).