Prediction of chemoresistance trait of cancer cell lines using machine learning algorithms and systems biology analysis

NCBI GEO database was used to obtain data from datasets [25]. Five different platforms of microarray including GPL13667, GPL6947, GPL6480, GPL6104, and GPL6244 have been used. A total of 85 samples of gene expression datasets of microarray data were selected from different platforms. The selected datasets were related to various cell lines such as ovarian, pancreatic and non-small cell lung cancer (NSCLC) both resistance and sensitive to Cisplatin drug. Sample numbers as well as cancer types are indicated in Table 1.

The data has been normalized using the LIMA package in the R software [26]. Average has been taken between the expressed values of repetitive probes in each dataset to obtain a unique expression value for each probe. A total 7621 genes were harvested after the isolation of common genes between platforms. The Combat software was then used to eliminate batch effects between different platforms and experiments [27]. Also, an average has been taken between replicates of each platform (sensitive and resistance separately) which reduces the total number of samples from 85 to 14 sensitive and 14 resistances to Cisplatin samples.

As the data was collected from various sources, it was necessary to somehow remove the discrepancies known as “Batch effects” between different samples. The batch effects of 85 different samples, each containing 7620 genes, were corrected by the SVA package. The SVA package comprises functions to remove the batch factors and other undesirable conversion in high-throughput examination. Specifically, the SVA package comprises functions to identify and build surrogate variables for high-dimensional datasets. Surrogate variables are covariates built directly from high-dimensional data (such as gene expression and RNA sequencing data) that can be utilized in subsequent analyses to adjust for unknown, unmodeled, or latent sources of noise. Moreover, the t-SNE algorithm was then used in MATLAB software to make the data presentable before and after the batch effect correction [28].

High dimensional DNA microarray has presented serious challenges to the existing machine learning and classification methods. In other words, in many of medical and microarray datasets, it is possible that many genes are irrelevant or redundant for machine learning algorithm [29‐32]. Feature selection or gene selection is a popular and powerful approach in medical datasets to overcome this shortcoming [33‐35]. In gene selection, to decrease the microarray data dimensions, by eliminating the irrelevant and similar genes, only a subset of relevant and dissimilar genes that are strongly related to the objective function are selected [36]. This is a powerful strategy to increase the efficiency of the machine learning algorithm, reduce time complexity, build more general classification algorithm, and reduce storage requirements [37, 38]. Gene selection approaches have been successfully employed in many medical applications including; gene expression [39], cancer classification [40], medical diagnosis [32], etc. In other words, a fundamental problem in machine learning algorithms is the high dimensional datasets, in which the size of the feature subset is much higher than the size of the patterns. Therefore, the classification accuracy is significantly reduced. As a result, it is necessary to reduce the initial features using dimension reduction techniques [41, 42]. One efficient way to reduce the dimension is feature selection (gene selection in DNA microarray datasets). In feature selection, an attempt is made to choose a set of initial features that satisfy two targets simultaneously: the minimum similarity between the selected genes and the maximum relevance of these genes with the target class [43, 44]. The main goal of this step is to select appropriate and important feature from original features [45]. To do this purpose, Fisher Score (FS) feature selection algorithm is used to select the features that are most relevant to the target class. Fisher Score is a supervised filter method that selects a feature subset such that the samples in the specific class are most similar to each other and the samples in the different classes are less similar. As a result, this measure Scores higher value to genes with higher separation characteristic. The FS of gene f_i is defined as follows:

where \(\overline{f}_{k}\) is the mean value of all the samples regarding the feature f_k, V is a set of all classes in a dataset, n_ν is the number of pattern on the class ν, and σ (f_k) and \(\overline{f}_{k}^{\nu }\) shows the variance and average of feature f_k on class ν, correspondingly.

The Fisher Score method was performed for 14 folds and in each step, 100 features that were most consistent with the normal distribution were selected so that a total of 1400 features were obtained. In the next step, in order to reduce the sample size, the PCA [46] method was performed for 14 folds on the output features of the Feature Score method and reduced set of features were selected for machine training.

In order to measure the flexibility of the developed method on different classifiers, in the designed experiments, the efficiency of the various approaches on three widely used six machine learning algorithms including Naïve Bayes [47], SVM [7], KNN [48], Decision tree [49], Random Forest [50] and Neural Network [51] algorithms is examined for machine training to detect Cisplatin sensitive and resistant samples in cancer patients undergoing chemotherapy. These machine learning algorithms are one of the most well-known and widely used machine learning algorithms that are used by researchers in various prediction and classification problems.

In these experiments the developed approach and other compared methods implemented using Python language programming on an Intel Core-i9 CPU with 16 GB of RAM.

To train these algorithms and due to the small number of available samples (14 pairs of samples including 14 sensitive and 14 resistant samples), the Leave one out method (LOO) was used [52]. In this method, at each step of the machine training, 13 pairs of samples were used as training data and one pair was used as test data. The training process continued until all pairs of samples were once used as the test data. The training was performed twice, once by considering 1400 features extracted from Fisher Score method and once by considering 210 features extracted by PCA method as the training samples. Finally, the performance of the trained machines was evaluated according to Accuracy [53], Specificity [54], Sensitivity [54], Precision [53], MCC [55] and F1 scores [55].

73 genes were selected using the CountIF filter in Excel (Refer to online resource 1). These genes were repeated in 50% or more of the 14 pairs sample extracted from the GEO. The LIMA package was used to calculate the DEGs between all sensitive and resistance samples. The result indicated that 34 of 73 gens which were chosen in the feature selection phase were down-regulated and the rest were up-regulated. The genes were enriched in biological pathways using the Enrichr database [56]. To obtain the mirs related to the 73 harvested genes, miRCancer [57] and miRDB [58] databases were used and the common mirs between the harvested mirs from the two aforementioned databases were selected based on the specific cancer cell line as well as the direction of the regulation. mirs were enriched using DIANA mirPath v.3 database [59] (refer to online resource 2) and pathways related to chemoresistance were selected among the obtained pathways. Furthermore, the transcription factors which regulate the obtained DE genes were harvested using the TRRUST v2 database [60] and were enriched in pathways related to chemoresistance by Enricher database.

Using the obtained transcription factors and mirs, two types of network including a TF-gene network and a mir-gene network were constructed using Cytoscape software [61]. Topological analyses were then performed using degree and betweenness centralities to report the networks hub genes.

The obtained DEGs were studied through literature review. All of the 73 genes were checked in articles and are indicated in a table based on the mechanism of chemo resistance. Some of these genes were not mentioned in the related articles as chemo resistance agents and therefore, they were nominated for further resistance studies which indicated that these genes have main roles in cells and are included in important biological pathways.

To correct batch effects between different samples, t-SNE algorithm was used (Fig. 1). This algorithm reduces data dimension and makes the data easier to picture and therefore, easier to comprehend.

In order to select appropriate features for machine learning algorithms, feature selection was performed for 14 folds using the Fisher Score method. In each fold, 100 genes which were most similar to the normal distribution were selected out of 7620 genes. Furthermore, to reduce the dimension of the selected features, the PCA algorithm was performed for 14 folds on the extracted features and several genes were selected in each fold (Table 2).

Performance of the machines trained by using reduced features extracted from the PCA algorithm (Table 2) are shown in Fig. 2. Among the developed machines, the Decision Tree algorithm with the average Accuracy of 50% has the weakest performance in terms of accuracy. On the other hand, KNN shows the highest accuracy with an average of 67%. The best performance based on accuracy criteria also belong to KNN and Decision Tree algorithms according to the obtained box plots (Fig. 2c).

Among the developed machines, the Naïve Bayes algorithm is the weakest machine in terms of negative sample detection with a 50% Specificity criteria. Decision Tree algorithm, on the other hand, has the highest average Specificity criterion of 69%, followed by Random Forest and KNN algorithms. The KNN algorithm has the best performance based on the specificity criterion based on the extracted box plots (Fig. 2c).

Based on the results, the weakest performance in terms of Sensitivity criteria is related to the Naïve bayes algorithm, which has not correctly detected any positive samples. On the other hand, KNN and Decision Tree algorithms have the highest Sensitivity criteria with an average of 78 and 67%, respectively (Fig. 2). The Decision Tree algorithm has the best performance in terms of sensitivity criteria based on the obtained box plots (Fig. 2C).

Among these algorithms, Decision Tree and Random Forest algorithms with an average precision of 71% have the highest average precision. The Decision Tree algorithm has the best performance in terms of precision according to the extracted box plots (Fig. 2C).

Similarly, a new set of six machines was trained only this time, 1400 features extracted from the Fisher Score algorithm were used in the training process. Performance results of these machines are shown in Fig. 3. In the new set, the KNN algorithm with an average of 67% accuracy has the highest percentage of correct sample detection compared to other algorithms. Random Forest, Naïve Bayes and SVM algorithms come afterward with an average accuracy of 64%. The KNN algorithm is also the best machine in terms of accuracy based on extracted box plots (Fig. 3c).

Naïve Bayes and Random Forest algorithms, with an average specificity criterion of 67%, are the best machines to correctly detect negative samples. On the other hand, the Decision Tree algorithm with the average specificity of 45% has the weakest performance in this regard (Fig. 3). Furthermore, the KNN algorithm has the best performance in terms of specificity based on the obtained box plots (Fig. 3c). In addition, the KNN algorithm with 78% average sensitivity is the best machine to correctly detect positive samples. The SVM algorithm comes afterward with an average sensitivity of 70% (Fig. 3). According to the obtained box plots, the Naïve Bayes and KNN algorithms are the best choices in terms of Sensitivity criteria, respectively. In addition, among the above algorithms, Naïve Bayes and Random Forest algorithms with an average precision of 71% are the most precise machines. Similarly based on the calculated box plots, the Naïve Bayes algorithm performed better than other algorithms in terms of precision criteria. Finally, according to the MCC criteria, the KNN algorithm and according to the F1 Score criterion, the Random Forest and Naïve Bayes algorithms have the best performances (Fig. 3).

The specific mirs for the extracted 73 DEGs were harvested from the miRCancer and miRDB databases for related cancer types (Table 1). These results determine that the expression profile of mirs are down for upregulated genes and are up for down regulated ones (Table 3; Fig. 4).

The mir target network has been topologically analyzed using the degree centrality and the results revealed the hub genes which should be evaluated for their performance in the chemoresistance process (Table 4).

The upregulated and down regulated mirs in the pathways related to chemoresistance were enriched using DIANA mirPath v.3 database. The related pathways were extracted from KEGG database using the standard P-value of 0.05 (The extracted related pathway: Additional file 1).

Three factors including the in-degree, the out-degree and the betweenness centralities have been noted in TF-gene interaction network for topological analysis. The detected hub genes are related to the in-degree centrality and are listed in Table 5.

The network illustrated in Fig. 5 has been constructed by merging upregulated and down regulated genes and their corresponding regulatory TFs. Topology analysis has been performed based on this network. According to the results obtained from the trrust database, the TFs were down for upregulated genes and were up for down regulated genes (Table 6).

The transcription factors regulating the genes harvested from the feature selection step were enriched using Enrichr database in the oncogenes and chemoresistance pathways. In the Enrichr database the pathways were harvested from KEGG and the adjusted P-value was significant. The results are shown in Table 7.

A literature review was performed on the 73 DE genes which were harvested by the CountIF filter in the feature selection step. A group of these genes were reported in the literature as chemoresistance genes (Additional file 2). The other ones were identified to have a vital role in the oncogenesis pathway and other important cell functions. Therefore, although they were not reported as chemoresistance genes we propose that these genes might be potentially chemoresistance (Table 8). Future studies can be performed to validate these results.