XGBoost-based and tumor-immune characterized gene signature for the prediction of metastatic status in breast cancer

Tumor expression data for modeling in this research were based upon data generated by the Cancer Genome Atlas (TCGA) database. All paired clinical data and transcript profiles of breast cancer (BRCA) samples were obtained and trimmed from the TCGA Data Portal by R package “GDCRNATools” [18]. The original data consisted of 1,097 samples in total. According to the pathologic_M column in the clinical information table, data rows with M0 and M1 status remained unchanged, and data with MX status (ambiguous metastatic status) were removed. A total of 923 BRCA samples, including 901 non-metastatic samples and 22 metastatic samples, were finally retained. Then, we grouped the BRCA samples into 2 groups depending on the status of pathologic metastasis. The metastatic and non-metastatic groups were labeled M1 and M0, respectively. M0 and M1 were used as labels for binary samples before classification by the following machine learning algorithms. Single-cell sequencing data were achieved from the Gene Expression Omnibus (GEO) database, GEO accession number was GSE162726 [19]. The R package “Seurat” [20] was used for the quality control and integration of the single-cell RNA-seq data.

After data were downloaded and integrated, we grouped the BRCA samples into 2 groups according to the status of pathologic metastasis. The metastatic and non-metastatic groups were labeled M1 and M0, respectively. Samples with unclear metastatic status were removed. The R package “DESeq2” [21] was adopted to generate DEGs between the two groups based on a negative binomial distribution. The significance criteria for DEGs was P value < 0.05, and the |log2 Fold-change| ≥ 1. The R package “EnhancedVolcano” was used to conduct the volcano plot and visualize the results of differential expression analyses. Then, clustering and visualization of the non-redundant biological terms of genes in a functionally grouped network was conducted with the Cytoscape (V3.8.0) desktop application and the “ClueGO” plug-in.

Different machine learning models were adopted to decide which model was the best one suitable for the present study. The XGBoost, DT, SVM, KNN, LR, and RF classifiers were used to establish the classification model. Tenfold cross-validation was performed for each model, and the ROC curve was plotted to calculate the mean area under the curve (AUC). The model with the highest mean AUC value was selected for modeling. We used Jupyter Notebook (version 6.1.4), a web-based application for interactive computing in Anaconda Navigator (anaconda3), to implement different machine learning algorithms. Scikit-learn module in Python (version 3.9) programming was adopted.

XGBoost classifier is a gradient boosting method that combines the regression tree [14]. The goal function of the XGBoost algorithm model is \(obj(\theta ) = L(\theta ) + \Omega (\theta )\), where \(L\left( \theta \right)\) is the training loss function, and \(\Omega \left( \theta \right)\) is the complexity function of the tree. \(L(\theta ) = \sum\nolimits_{i = 1}^{{\text{n}}} {l(y_{i} ,\hat{y}_{i} )}\), \(l(y_{i} ,\hat{y}_{i} )\) corresponds to the training loss function for each sample, where \(y_{i}\) represents the true value of the ith sample, and \(\hat{y}_{i}\) represents the estimated value of the ith sample. \(\hat{y}_{i} = \sum\nolimits_{k = 1}^{K} {f_{k} (x_{i} )} ,f_{k} \in F\), where K represents the number of trees, F represents all possible DT, and f denotes a specific CART tree. \(\Omega (f) = \gamma T + \frac{1}{2}\lambda \sum\nolimits_{i = 1}^{T} {w_{i}^{2} }\), where w_i is the score on the ith leaf node, and T is the number of leaf nodes in the tree. By adjusting parameters, the objective function was continuously optimized, and optimal results were obtained. The grid search algorithm was used to optimize the hyper-parameters, including max_depth, min_child_weight, gamma, subsample, colsample_bytree and learning_rate in each iteration.

The lentivirus construction to knockdown SQSTM1, GDF9, LINC01125, PTGS2, GVINP1, and TMEM64 was purchased from Genepharma (Shanghai, China). Breast cancer cell line MCF-7 was plated in six-well dishes at 50% confluence and then infected with the above 6 lentiviruses (termed as shSQSTM1, shGDF9, shLINC01125, shPTGS2, shGVINP1, and shTMEM64), or control (termed as shCtrl) in MCF-7 cell, respectively. Stable cell lines were generated by selection using puromycin at a concentration of 4 µg/mL for 2 weeks. The cell transfection protocol described above was in accordance with the manufacturer’s instructions.

In MTT (3-[4,5-dimethylthiazol-2-yl]-2,5 diphenyl tetrazolium bromide) assay, the shSQSTM1, shGDF9, shLINC01125, shPTGS2, shGVINP1, shTMEM64 and shCtrl MCF-7 cells were seeded into 96-well plates (Cat. # 3599, Cornning) at a density of 2 × 10³ cells per well over night. Before adding 150 μL of DMSO, 20 μL of MTT at a concentration of 5 mg/mL was added to each well and incubated for 4 h. Then, a microplate reader was used to measure the optical density at 490 nm. Colony formation assay and transwell assay were performed according to our previous research [22]. Cell migration was observed using a wound healing assay. Transfected MCF-7 cells were maintained in 6-well plates and upon reaching 90% confluence, scratches were created using micropipette tips. Cells were washed 3 times using sterile PBS to wash off non-adherent cells generated by the scratch, and fresh serum-free medium was replaced to continue culturing the cells. The wound status was observed at 0 h and 24 h after scratching with an X71 inverted microscope (Olympus). The means of intercellular distances were calculated using the ImageJ software. All experiments were performed in triplicates.

We analyzed the correlation between the expression of the selective gene signature and several immune cell markers to determine the association of infiltrating immune cells with our proposed gene signature. Immune gene markers were selected from the website of R&D Systems or from the GEPIA 2.0 [23] recommendations, including markers of B cell, naïve T cell, effector T cell, resident memory T cell, T helper 1 (Th1) cell, regulatory T cell (Treg), T cell exhaustion, macrophage, tumor-associated macrophage (TAM), monocyte, natural killer (NK) cell, neutrophil, and dendritic cell (DC). Gene expression correlation analysis was performed of BRCA tumor datasets of TCGA expression data by GEPIA 2.0. Correlation coefficients were determined by the Spearman method.

After we grouped the BRCA samples into 2 groups depending on the status of pathologic metastasis, the metastatic and nonmetastatic groups were labeled M1 and M0. There were 901 nonmetastatic breast tumor samples and only 22 metastatic breast tumor samples, representing an imbalanced-class dataset. To recognize distinct patterns between subgroups of metastatic and nonmetastatic breast cancer, we conducted a DEG analysis. A total of 117 mRNAs passed the threshold screening, including 37 upregulated genes and 80 downregulated genes, as exhibited in the volcano plot (Fig. 1a). Based on these 117 genes, Gene Ontology (GO) analyses were further conducted, indicating that these differentially expressed mRNAs were mainly enriched in biological processes associated with chemokine-mediated signaling pathways and regulation of humoral immune responses (Fig. 1b). These results suggested that the antitumor immune response could have an immunological effect on the metastasis of breast cancer, and the stemness of tumor cells might also be responsible since the GO term “positive regulation of stem cell differentiation” was also enriched significantly. Survival analysis by Kaplan–Meier (KM) plot differed significantly in survival outcome between the M0 and M1 groups (Fig. 1c), with significantly worse survival in the M1 group (i.e., the metastatic group), consistent with a previous report showing poor survival in metastatic breast cancer [8, 19].

To precisely predict tumor metastatic status in breast cancer patients using gene expression profiling data, we sought to develop an effective classification model that would identify metastatic cases from nonmetastatic cases. Based on 117 DEGs screened out above as features or modeling, machine learning classification algorithms, including DT, support vector machine, KNN, LR, RF and XGBoost, were used to establish the classification model (Fig. 2). Since the accuracy could still reach 98% when all the metastatic samples were classified as nonmetastatic, we chose area under the ROC curve (AUC) instead of accuracy (ACC) as the evaluation index. Ten-fold cross-validation was performed for each model, and the ROC curve was plotted to calculate the mean AUC. The results showed that the classification model based on XGBoost performed best, with the highest mean AUC, reaching 0.64. The PR (precision-recall) curves was also plotted in Additional file 1: Fig. S1. This might be because XGBoost is a machine learning technique featuring significant improvements in efficiency and performance relative to other classifiers.

Compared with 901 nonmetastatic breast tumor samples, only 22 metastatic breast tumor samples were found in this study, indicating an imbalanced binary classification problem. When training imbalanced-class data, oversampling the minority class or undersampling the majority class is often used to alleviate the positive–negative sample ratio in datasets [24]. XGBoost provides an additional method to handle imbalanced data, with the scale_pos_weight parameter set to give samples of the minority class a certain weight. Therefore, we manually adjusted two hyperparameters, setting the parameter objective to binary: logistic based on our purpose and setting the parameter scale_pos_weight to 200 based on the positive and negative sample ratio. Then, we used the grid search algorithm to optimize other important relevant hyperparameters, including max_depth, min_child_weight, gamma, subsample, colsample_bytree and learning_rate. The grid search algorithm attempted to maximize the average AUC score in each iteration. After a given number of iterations were completed, the model with the highest mean AUC score was selected for the following prediction. The order in which the parameters are tuned and the final parametric results are presented in Fig. 3a. Figure 3b shows that using the current settings of XGBoost hyperparameters, the mean AUC score obtained by ten-fold cross-validation increased to 0.8, and the prediction performance of the optimized XGBoost model was greatly improved, the PR curves was plotted in Additional file 1: Fig. S1a. However, there were still many redundant features among the 117 features, which may cause overfitting and difficult clinical application. To improve the generalization capability of classifiers and reduce the time for training the classifier, we calculated the importances of the features (Fig. 3c) and selected the top 6 features higher than 100 ranked by feature importance score for subsequent modeling.

Next, the selected 6 features were fed into the optimized XGBoost predictive model. The results showed that through parameter optimization and feature selection, the average AUC value increased from 0.64 to 0.82, indicating that the deletion of redundant features was beneficial to improving the model’s accuracy (Fig. 3d). It should be noted that the long noncoding RNA LINC01125 plays the most important role in differentiating metastatic and nonmetastatic breast cancers. LINC01125 was previously reported to suppress the proliferation of breast cancer cells via in vitro experiments [25], consistent with our study that the expression of LINC01125 decreased in metastatic tissues.

Detailed information on the 6 selected genes is illustrated in Table 1, including Ensembl ID, gene symbol, log2 fold change, standard error, Wald statistic, Wald test P value and BH adjusted P value, calculated by the R package “DESeq2”. The highest ranking for classification importance was not necessarily the one with the greatest fold change and vice versa. This was also the advantage of the feature selection algorithm over traditional statistical methods, favoring the selection of biomarkers that could serve as a distinction between metastatic and nonmetastatic breast cancer. The feature selection before modeling was also called informative gene selection when handling the RNA-seq data. Table 1 also shows that only SQSTM1 is upregulated in metastatic breast cancer tissues, while the rest are downregulated to some extent, indicating that SQSTM1 is a risk factor, while LINC01125, GDF9, PTGS2, GVINP1 and TMEM64 are protective factors. Although bulk-tissue RNA-seq is frequently used to illustrate transcriptomic variations under case-specific conditions such as metastatic status, understanding the composition and proportion of cell types in intact tissues is important because of their different properties [26]. Therefore, we then explored the role of the selective gene signature in metastatic breast cancer from both the tumor cell and immune cell perspectives.

To explore the role of the selective gene signature from the tumor cell perspective, first, we utilized public breast cancer single-cell sequencing data to probe the gene expression levels of the selective gene signature in fast-moving migratory breast cancer cells compared to those in non-migratory cancer cells. The R package “Seurat” was used for quality control and integration of the single-cell RNA-seq data. Uniform Manifold Approximation and Projection (UMAP) [27] clustering demonstrated the distinct gene expression profiles of migratory and non-migratory breast cancer cells, with each dot representing a cell (Fig. 4a). Migratory and nonmigratory populations of the same cell line were easily distinguished, suggesting differential gene expression patterns consistent with previous research using t-SNE clustering [19]. We explored the expression of the selective gene signature for each cluster and found that among all 6 informative genes, only the expression of SQSTM1 was generally increased in migratory GUM36 cells compared to non-migratory GUM36 cells (Fig. 4b, Additional file 2: Fig. S2). As can also be seen from the chart label in Additional file 2: Fig. S2, in addition to the gene SQSTM1, the expressions of the other 5 genes in breast cancer cells are relatively low. Based on this, we infered that SQSTM1 might play a role to breast cancer cell migration from the tumor cell perspective.

Next, we did some in vitro experiments to verify the above assumption. The MCF-7 cell line is by far the most commonly used xenograft model of breast cancer. To elucidate the biological functions of SQSTM1, GDF9, LINC01125, PTGS2, GVINP1, and TMEM64 in breast tumor cells, we knocked down the expression of the 6-gene signature using shRNA or the negative control in MCF-7 cell lines to assess cell proliferation, migration and invasion in vitro. The MTT assay demonstrated that tumor proliferation was significantly inhibited in the shSQSTM1 group compared to the shCtrl group (Fig. 5a, b). In MCF-7 knockdown cells, the number of cell clones decreased in the shSQSTM1 group compared with that in the shCtrl group (P < 0.05, Fig. 5c, d). Transwell assays revealed that MCF-7 cell invasion was significantly reduced after downregulation of SQSTM1 (Fig. 5e, f). Finally, cell migration was evaluated by wound-healing assay, and decreased expression of SQSTM1 significantly inhibited the migration of MCF-7 cells (Fig. 5g, h). Taken together, the above data indicated that knockdown of SQSTM1 could inhibit the proliferation, migration and invasion of breast tumor cells in vitro. In other words, SQSTM1 functioned from the perspective of tumor cells since it was significantly upregulated in metastatic breast cancer, and its knockdown attenuated the ability of tumor cells to invade metastases.

Since tumor cells and tumor infiltrating immune cells, especially T cells account for the highest proportion of cells in breast tumor tissues [28], and considering the expression of SQSTM1 was upregulated in metastatic breast cancer tissues, while the expression levels of the remaining genes were downregulated. We speculated whether these 5 genes, LINC01125, GDF9, PTGS2, GVINP1, and TMEM64, contributed to breast cancer metastasis because of immune dysfunction and conducted an exploration from the immune cell perspective. First, we investigated the correlation between the selective gene signature and different types of immune cells based on breast tumor expression data in TCGA. GEPIA 2.0 was used to explore the correlation between the selective gene signature above and the indicated gene markers, including the markers of B cells, naïve T cells, effector T cells, resident memory T cells, Th1 cells, Tregs, T cell exhaustion, macrophages, TAMs, monocytes, NK cells, neutrophils, and DCs. As shown in Table 2 and Fig. 6a–d, these 5 genes may play a role in the metastasis of breast cancer from an immunological point of view through naïve T cell, effector T cell, resident memory T cell and DC populations. Because the expression of these genes and immune cells represented a significantly positive correlation (correlation coefficient > 0.3) and these 5 genes were downregulated in metastatic breast cancer tissues, it could be speculated that these genes contributed to breast cancer metastasis by attenuating the immune response. Results also showed that there was no correlation between fibroblasts and these 5 genes in tumors, although there was a weak correlation in normal tissues. Next, we performed a correlation analysis of each gene in the selective gene signature with the indicated immune cell marker genes. As shown in Fig. 6e, GVINP1 was significantly correlated with immunity, with high correlation coefficients above 0.7 with CD69 and CCR7. PTGS2 also showed some correlation with immunity, mainly reflected by the correlation coefficients of more than 0.4 with CD1C and CD69. SQSTM1 was poorly immune-related, validating previous results that SQSTM1 regulates breast cancer metastasis from a tumor cell perspective. LINC01125, GDF9 and TMEM64 exhibited significant but very weak immune correlations.