Skip to main content
NCBI home page
Journal of Cancer Research and Clinical Oncology logoLink to Journal of Cancer Research and Clinical Oncology
. 2023 Sep 7;149(17):15923–15938. doi: 10.1007/s00432-023-05358-x

A deep learning approach based on multi-omics data integration to construct a risk stratification prediction model for skin cutaneous melanoma

Weijia Li 1,#, Qiao Huang 1,#, Yi Peng 1, Suyue Pan 1, Min Hu 1, Pu Wang 1, Yuqing He 1,2,✉
PMCID: PMC11796813  PMID: 37673824

Abstract

Purpose

Skin cutaneous melanoma (SKCM) is a highly aggressive melanocytic carcinoma whose high heterogeneity and complex etiology make its prognosis difficult to predict. This study aimed to construct a risk subtype typing model for SKCM.

Methods

The study proposes a deep learning framework combining early fusion feature autoencoder (AE) and late fusion feature AE for risk subtype prediction of SKCM. The deep learning framework integrates mRNA, miRNA, and DNA methylation data of SKCM patients from The Cancer Genome Atlas (TCGA), and clusters the screened multi-omics features associated with survival prognosis to identify risk subtypes. Differential expression analysis and functional enrichment analysis were performed between risk subtypes, while SVM classifiers were constructed between differentially expressed genes (DEGs) obtained by Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression screening and risk subtype labels inferred from multi-omics data, and the predictive robustness of risk subtypes inferred from the risk subtype classification prediction model was validated using two independent datasets.

Results

The deep learning framework that combined early fusion feature AE with late fusion feature AE distinguished the two best risk subtypes compared to the multi-omics integration approach with single strategy AE or PCA. A promising C-index (C-index = 0.748) and a significant difference in survival (log-rank P value = 4.61 × 10–9) were found between the identified risk subtypes. The DEGs with the top significance values together with differentially expressed miRNAs provided the biological interpretation of risk subtypes on SKCM. Finally, the framework was applied to predict risk subtypes in two independent test datasets of SKCM patients, all of which showed good predictive power (C-index > 0.680) and significant survival differences (log-rank P value < 0.01).

Conclusion

The SKCM risk subtypes identified by integrating multi-omics data based on deep learning can not only improve the understanding of the molecular mechanisms of SKCM, but also provide clinicians with assistance in treatment decisions.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00432-023-05358-x.

Keywords: Skin cutaneous melanoma, Deep learning, Autoencoder, Multi-omics data integration, Subtyping, Prognosis prediction

Introduction

Skin cutaneous melanoma (SKCM) is the most aggressive type of melanoma of skin cancer (Schadendorf et al. 2018; Marie et al. 2020). SKCM accounts for roughly 80% of all skin cancer deaths worldwide (Rl et al. 2017). The occurrence of SKCM is associated with UV radiation-induced DNA mutations (Pavey et al. 2020). The main treatment approach for malignant melanoma is the local surgical removal combined with systemic therapies (e.g., radiotherapy, chemotherapy and immunotherapy) (Hao et al. 2020). Although early local surgical resection can improve the survival of patients with melanoma, post-surgical tumor metastasis can lead to a poor prognosis for many patients (Schadendorf et al. 2018). Malignant melanoma is known for its risk of early metastasis and poor sensitivity to radiotherapy, and the overall treatment outcome is not ideal (Avogadri et al. 2014). Although the advent of novel immune checkpoint inhibitors has changed the treatment of advanced melanoma (Chang et al. 2020), many patients with advanced melanoma still exhibit resistance to immune checkpoint inhibitors (Haymaker et al. 2021). Given the high heterogeneity of SKCM and the complex etiology, there is an urgent need for a new tool capable of identifying risk subtypes of cutaneous melanoma.

With advances in data collection techniques, more and more omics data are being used in cancer research, and molecular subtypes for melanoma prognosis have increasingly become an area of interest. Wan et al. used RNA-seq data from TCGA SKCM patients to construct robust subtypes associated with four cell death-associated genes (Wan et al. 2022). Zhang et al. defined new subtypes from a metabolic perspective and obtained new subtypes of SKCM after consistent clustering of genes for glycolysis and cholesterol and found that patients in the glycolytic metabolic subgroup had the worst prognosis (Zhang et al. 2021b). In addition to this, some studies have typed SKCM from other molecular perspectives, such as copper apoptosis-related genes (Lv et al. 2022), immune-related differential genes (Ji et al. 2022), necroptosis genes (Peng et al. 2022) versus immunogenic cell death-related genes (Chen et al. 2023), and other molecular perspectives. Although these molecular typing studies provide a different molecular perspective for the understanding of SKCM, such studies considering only single-omics provide only limited information, and the grouping results are also susceptible to noise. Therefore, a broad consideration of multi-omics data may help the mutual complementation of information from different data types, which puts forward high demand on the method of integration of multi-omics data. With the development of research in deep learning (DL) techniques such as autoencoder (AE) (Kramer 1991), cancer researchers can use this technique to integrate multi-omics data to identify significant risk subgroups. Chaudhary et al. first used the DL autoencoder computational framework for typing patients with hepatocellular carcinoma in their study, which integrated RNA, miRNA and DNA methylation expression data, and which yielded significant survival differences between robust molecular subtypes and withstood testing across study cohorts (Chaudhary et al. 2018). In studies of prostate cancer and high-risk neuroblastoma (Zhang et al. 2018; Wang et al. 2021), it was shown that DL-based autoencoders significantly outperformed similarity network fusion (SNF) algorithms, integrated cluster analysis (iCluster), and non-DL multi-omics fusion algorithms such as PCA, which also reflect the advantages of AE for disease typing. In addition to these diseases mentioned above, the computational framework of AE is also widely used in the subtyping of other diseases, such as colon adenocarcinoma (Lv et al. 2020) and bladder cancer (Poirion et al. 2018; Zhang et al. 2021c), but few studies have used the DL algorithm for the subtyping of SKCM multi-omics data.

Notably, Leng et al. evaluated 16 DL-based data fusion methods, named according to the structural features of the multi-omics data fusion model, and showed that the early feature fusion AE methods presented good performance in terms of classification performance (Leng et al. 2022). In this study, we not only used the early feature fusion approach described above, but also referred to the late feature fusion approach in the study of Poirion et al. (2018) to propose a new DL framework for integrating RNA, miRNA and DNA methylation expression data by combining early feature and late feature fusion. Robust samples of the same risk subtype in DL were trained using a support vector machine (SVM) classification algorithm to construct a SKCM risk subtype prediction model. Finally, two independent datasets were collected to validate this prediction model. A subtype classification method that reflects different survival status is of valuable for guiding the clinical application of treatment for SKCM patients.

Materials and methods

Data collection and preprocessing

446 SKCM samples with RNA, miRNA, and DNA methylation expression data with clinical prognostic information were obtained from The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov) database. The expression profiles of mRNA and miRNA were converted from probe id to gene symbols before analysis. The DNA methylation data were processed using the R package “ChAMP” (Tian et al. 2017). The CpG sites were mapped to gene symbols, and the methylation β values were averaged for 1500 bp upstream of the gene transcription start site. In addition, a study reported the inconsistency between the omics and clinical data of SKCM in the TCGA database and found that observed survival interval (OBS) was more suitable for correlation with omics data but not for correlation with clinicopathologic features compared with overall survival (OS) (Xiong et al. 2019). Therefore, in this study, OBS was used to correlate with omics data to avoid biased results, while OS was still used to correlate with clinicopathologic features.

Concerning the above omics data, after removing the probes that cannot be mapped to gene symbols, preprocessing is performed using the following steps: (1) Remove the probes or genes with more than half of the sample number of missing values. (2) The missing values were filled in using the KNN method of the R package “impute”. (3) If more than 20% of the patients had mRNA or miRNA expression values of 0, they were removed. (4) Since the mean methylation values were taken in the range of 0–1 interval, only log2(x + 1) was applied to mRNA and miRNA expression values to avoid the effect of extreme values on the data distribution, and min–max normalization was used to scale the mRNA and miRNA expression values to the range of 0–1 interval. After pre-processing, the dataset included a total of 15,770 mRNAs, 580 miRNAs, and 17,357 DNA methylation features.

Two mRNA test set data were downloaded from the Gene Expression Omnibus database (GEO, https://www.ncbi.nlm.nih.gov). They are GSE65904 (n = 210), and GSE54467 (n = 79), respectively. The preprocessing process of the above data was kept consistent with that of the TCGA dataset, and only samples with survival prognosis information were considered for these datasets.

Multi-omics data integration by deep learning

AE models are constructed on TCGA-SKCM data with two different strategies, the early feature fusion AE model (Fig. 1A) and the late feature fusion AE model (Fig. 1B), both of which implement autoencoders with three hidden layers. Among them, the bottleneck layer is the important features extracted from the multi-omics data. The AE model is trained using Tanh as the activation function and a gradient descent algorithm with epochs and dropout set to 10 and 50%, respectively, and l1 and l2 are set to 0.001 to prevent overfitting of the model. The nodes of each hidden layer of the early feature fusion AE model are 500,210,500. In the late feature fusion AE model consisting of three single-omics AE models, the number of hidden nodes in each model is 500,70,500. The sum of nodes in the three single-omics bottleneck layers in the late feature fusion AE model is kept consistent with the number of nodes in the bottleneck layers in the early feature fusion AE model. The data integration analysis of the AE model is implemented using the R package “h2o”.

Fig. 1.

Fig. 1

Open in a new tab

Different strategies are used to construct AE models. A, B AE models are constructed based on early and late feature fusion

Transformation feature extraction and clustering

For demonstrating the typing capability of the deep learning model, this study additionally used linear PCA dimensionality reduction methods to select multi-omics features. All these methods reduced the original multi-omics features to 210. Subsequently, a one-factor Cox proportional hazards (Cox-PH) model was used to select features associated with survival. K-means clustering of SKCM patients was performed using screened multi-omics features, and two subgroups of SKCM samples were named as high- and low-risk subgroups depending on prognosis. Samples from the two AE models predicting subgroups with the same risk subgroup label were constructed as joint feature fusion AE models (n = 353). At the same time, survival rates were compared between early feature fusion AE, late fusion feature fusion AE, combined feature AE and two subgroups extracted by PCA method, and the prognosis was assessed using C-index, Brier score and log-rank P value. PCA feature extraction was performed using the R package “FactorMineR” (Lê et al. 2008), single-factor Cox-PH models were processed using the R package “survival”, and K-means clustering was done using the R package “stats” (Team 2014).

Evaluation metrics for models

To evaluate the accuracy of the predicted risk subgroup labels, three assessment metrics, C-index, Brier score and log-rank P value, were applied in this study. The C-index is the percentage of pairs in the data for which the predicted output agrees with the actual result in all patient pairs (Longato et al. 2020). The Brier score is applied to assess the accuracy of the outcome prediction and reflects the mean squared error between the actual outcome observation and its predicted probability (Yang et al. 2022), and takes a value between 0 and 1. The lower the Brier score, the higher the accuracy of the probabilistic prediction. Log-rank P value was obtained from a log-rank test of the Cox-PH model for two risk subgroups. C-index and Brier score was generated with the R package “survcomp” (Schröder et al. 2011), and log-rank P value was obtained with the R package “survival”.

The risk subgroup is an independent prognostic factor in SKCM

After obtaining the risk subgroups for SKCM, multivariate Cox regression analysis was performed and forest plots were drawn for the clinical characteristics and risk subgroup characteristics of the patients. Additionally, a nomogram was constructed with the R package “rms” to predict the 3- and 5-year survival probabilities among SKCM patients. Calibration curves are used to assess the variance between the survival probability predicted by the nomogram and the actual survival probability.

Differential expression analysis and enrichment analysis

To characterize the differences between the two risk subtypes, differential expression analysis of mRNAs and miRNAs between subtypes was applied with the R package “limma” (Ritchie et al. 2015) and volcano maps were drawn. |log2 fold change (FC)|> 1 and false discovery rate (FDR) < 0.05 were selected as thresholds for identifying significantly differentially expressed genes (DEGs) and differentially expressed miRNAs. For the DEGs of the two risk subtypes, Gene Ontology (GO) functional enrichment analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis were performed with the R package “clusterProfiler” (Wu et al. 2021). FDR < 0.01 was used as the threshold for GO and KEGG analysis.

LASSO logistic regression was used to screen candidate DEGs associated with risk subtypes

Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression is an algorithm used to control model complexity to reduce overfitting and thus retain important variables, which retains important variables by reducing the coefficients of some insignificant variables in the original model to zero. In this study, a tenfold cross-validation LASSO logistics regression was performed on the risk subgroup-related DEGs, with the distribution of the dependent variable following a binomial distribution, and the minimum lambda value was used to filter the key DEGs. The LASSO logistics regression was implemented using the R package “glmnet” (Tay et al. 2023).

Construction of the SVM classifier based on the risk subgroup

Before constructing the SVM models for predicting risk subgroups, the TCGA data were divided into training set/validation set (60%/40%) to ensure that there were enough training samples for constructing the prediction models and validation set samples for testing the model prediction performance. The key DEGs extracted by the LASSO logistic regression were intersected with genes from the external test sets to ensure that the constructed model could be applied to the other test sets data, in which the screened key DEGs are the independent variables of the SVM model, and the multi-omics risk subgroup labels are the dependent variables of the SVM model, and a fivefold cross-validation approach was used to assess the accuracy of the training set SVM model. The risk subtype label prediction SVM model was constructed using the R package “e1071” (Meyer et al. 2014).

Confirmation using SKCM validation-set and two independent test-sets

The SKCM validation set was used along with two other independent GEO test sets for the assessment of risk subgroup prediction model robustness, and the evaluation metrics were calculated for each dataset.

Statistical analysis

Wilcoxon test was applied to test the difference between subgroups, which was significant at P < 0.05. All the above analysis methods were done in R software version 4.2.2.

Result

In constructing the risk stratification model for cutaneous melanoma (Fig. 2), the multi-omics features of the SKCM dataset were first integrated using two different strategies of AE, followed by the extraction of multi-omics features associated with survival outcomes in the bottleneck layer using Cox-PH, and K-means clustering of the SKCM samples based on the extracted features to identify subtypes associated with prognosis. To construct the SVM model, subtypes of both AE models predicted the same robust samples for gene differential expression analysis and data set partitioning.

Fig. 2.

Fig. 2

Open in a new tab

The overall workflow for constructing a skin cutaneous melanoma risk stratification model by integrating multi-omics data through deep learning

The SVM model was constructed as follows: first, genes contributing to risk subtype prediction among differentially expressed genes were extracted using LASSO-Logistic regression and intersected with genes from the GEO test set to obtain key DEGs. Finally, the SKCM robust samples extracted from the AE model were divided into training and validation sets, the key differentially expressed genes and predicted risk subtype labels in the training set samples were used to fit the SVM model, and the SKCM validation set and GEO test set were used for internal validation and external validation of the SVM model.

Multiple omics features were extracted by PCA linear dimension reduction technique to construct risk subtypes

The Elbow methods were first used to decide the optimal number of clusters for the multi-omics features of PCA. In the Elbow method (Fig. 3A), the reduction of intra-cluster distance at the inflection (K = 2) was significantly decreased. The SKCM samples were subsequently clustered into two classes using the K-means method, and the findings indicated that the samples in the two risk subgroups were indistinguishable from each other (Fig. 3B). Finally, Kaplan–Meier curves were plotted for the two risk subgroups (Fig. 3C), and although the statistically significant distinction was observed between the two risk subgroups (log-rank P value = 6.18 × 10–3), the C-index of this survival model was only 0.606 and the Brier score was 0.197, and the PCA linear dimensionality reduction method extracted survival subgroups had poor ability to distinguish SKCM samples.

Fig. 3.

Fig. 3

Open in a new tab

SKCM risk subgroups were obtained by K-means clustering of the multi-omics features extracted from PCA. A Determination of the optimal number of clusters using Elbow methods. B Clustering of SKCM samples based on the multi-omics features extracted by PCA. C Kaplan–Meier plot between two risk subgroups

The risk prediction subtype based on autoencoder was superior to PCA linear dimension reduction method

The risk subtypes predicted by early fusion feature AE and late fusion feature AE had a significant ability to discriminate between SKCM samples compared to the risk subtypes predicted by PCA. The Cox-PH survival model based on the AE risk subtype of early fusion characteristics (Fig. 4A) had a C-index of 0.716, a Brier score of 0.174, and a log-rank P value of 6.55 × 10–9. The Cox-PH survival model C-index for the late fusion characteristic AE risk subtype (Fig. 4B) was 0.711, and the Brier score was 0.183, with the same statistically significant distinction between the two risk subtypes (log-rank P value = 1.27 × 10–8). The results showed that the two AE models had a better ability to discriminate SKCM risk subtypes than the PCA linear dimensionality reduction method.

Fig. 4.

Fig. 4

Open in a new tab

Construction of early, late and combined early and late fusion feature AEs. A, B Sample clustering plots and Kaplan–Meier curves for early- and late-fusion AE. C, D Keeping the same samples for both AE models risk subtype prediction and plotting Kaplan–Meier curves

A total of 353 samples with the same risk subtype prediction in both AE models (Fig. 4C) were extracted for use in constructing the Cox-PH survival model (Fig. 4D), with a C-index of 0.748, Brier score of 0.174, and a log-rank P value of 4.61 × 10–9, which shows the statistical difference between the two risk subtypes. The deep learning framework with combined early and late feature fusion AEs outperformed single-strategy AEs in discriminating risk subtypes (Table 1).

Table 1.

Performance of different methods for extracting multi-omics features

Method C-index Brier score log-rank P value
PCA 0.606 0.197 6.18 × 10–3
Pre-Autoencoder 0.716 0.174 6.55 × 10–9
Late-Autoencoder 0.711 0.183 1.27 × 10–8
Union-Autoencoder 0.748 0.174 4.61 × 10–9
Open in a new tab

Risk subgroups were independent factors influencing SKCM

Multi-factor Cox regression analysis was performed on the risk subtype characteristics and clinical characteristics of the SKCM sample (Fig. 5A). The results showed that risk subgroups were independent influences on SKCM (P < 0.01) and that patients in the high-risk subgroup were 1.979 times more at risk than those in the low-risk subgroup (HR 1.979, 95% CI 1.256–3.118). Secondly, disease stage (Stage III and IV vs. Stage I and II, HR 1.899, 95% CI 1.223–2.949) was also influential factor for SKCM. Meanwhile, a nomogram model (Fig. 5B) was developed to predict the survival probability of SKCM based on the risk subgroup characteristics and clinical characteristics to predict the survival probability. Calibration curves show that the actual 3-year versus 5-year survival rates are close to those predicted by the nomogram (Fig. 5C, D).

Fig. 5.

Fig. 5

Open in a new tab

Multivariate Cox regression of risk subgroups versus clinical characteristics for SKCM and construction of nomogram models for predicting survival. A Forest plot showing the results of multivariate Cox analysis of clinical characteristics and risk subgroups in SKCM data. B Nomogram models were constructed to predict 3- and 5-year survival based on risk subgroups and clinical characteristics of SKCM. C, D Nomogram of 3-, 5-year survival Calibration curves

Differential expression analysis and enrichment analysis

To identify the differences between the two risk subtypes, the differentially expressed genes in the low and high risk subtypes of SKCM were visualized using volcano maps. A total of 464 DEGs were screened under the conditions of |log2 FC|> 1 and FDR < 0.05 (Table S1), of which 7 genes were upregulated and 457 genes were downregulated in the SKCM high-risk subtype. A total of 32 differentially expressed miRNAs were screened under the same conditions (Table S2), with 12 up-regulated differential miRNAs and 20 down-regulated ones. The top 10 most significant up-regulated and down-regulated genes in differential mRNAs and miRNAs were labeled in the volcano plot (Fig. 6A, C), and their expression heat maps were plotted (Fig. 6B, D), respectively.

Fig. 6.

Fig. 6

Open in a new tab

Differential expression analysis between SKCM risk subtypes with GO and KEGG enrichment analysis of DEGs. A, B Volcano plot of DEGs and heat map of differential expression values. C, D Volcano plot of differentially expressed miRNAs and heat map of differential expression values. E GO functional enrichment analysis shows the top 5 BPs, CCs, and MFs of GO terms. F KEGG enrichment analysis screening out the top 10 signaling pathways

The 464 DEGs associated with risk subtypes were analyzed for GO function and KEGG pathway enrichment, and 852 GO terms (Table S3) and 52 KEGG pathways (Table S4) were screened under the threshold condition of FDR < 0.01. In the GO analysis, the biological processes (BP), cellular localization (CC), and molecular function(MF) are 739, 58, and 55, respectively, and the bar chart shows the top 5 GO terms (Fig. 6E) and the top 10 KEGG pathways (Fig. 6F) with the smallest FDR under each category, respectively. GO analysis revealed that these DEGs were mainly enriched in cellular components such as the external side of plasma membrane, MHC protein complex and immunological synapse, biological processes were mainly enriched in regulation of T cell activation, leukocyte cell–cell adhesion and regulation of leukocyte cell–cell adhesion, and biological functions were mainly enriched in immune receptor activity, MHC protein complex binding and MHC class II receptor activity. KEGG pathway analysis showed that these DEGs are mainly associated with cell adhesion molecules, hematopoietic cell lineage and cytokine–cytokine receptor interaction pathways.

Candidate DEGs associated with risk subtypes were used to construct SVM models

To control the complexity of the risk typing model and to reduce the effect of multicollinearity on the results, LASSO logistic regression was used to reduce the coefficients of some DEGs variables to zero as a way to retain important DEGs for the construction of the SVM model. The simplified model enhanced sparsity and minimized binomial bias by reducing the 464 DEGs between SKCM risk subtypes to 34 DEGs by LASSO logistic regression (Fig. 7A, B). Analysis of these 34 candidate DEGs using single-factor Cox regression revealed that 30 low-risk candidate DEGs, including TMEM156, CLEC4E, and CD274, were significantly associated with the prognosis of SKCM. Notably, all of the candidate DEGs extracted by LASSO logistic regression were statistically significant (P < 0.05) except for TUBB4A, COL11A2, MMP12 and CCL18, indicating that these candidate DEGs' heterogeneity contributed significantly to the construction of subsequent SVM models (Fig. 7C).

Fig. 7.

Fig. 7

Open in a new tab

Screening of DEGs associated with SKCM risk subgroups using tenfold cross-validation LASSO logistic regression. A Number of variables with minimal cross-validation error in the retained LASSO model. B Distribution of coefficient changes for the DEGs variables. C One-factor Cox regression forest plots for 34 candidate DEGs

To make the SVM risk subgroup prediction model available for the test set, candidate DEGs were intersected with genes from the rest of the test set before constructing the SVM model, and a total of 27 candidate DEGs were obtained. A five-fold cross-validation SVM model was constructed using these candidate DEGs and the risk subtype features predicted by the AE model. The SKCM dataset is partitioned into a training set and a validation set by a ratio of 60%/40%. The average accuracy of the predictions from the fivefold cross-validation was 95.28% (95.24%, 92.86%, 97.67%, 95.24%, 95.35% for the five separate validations), and the AUCs values for the training and validation sets were 0.96 and 0.88, respectively. The risk subtype predicted by the training set (Fig. 8A) had a significant survival difference (log rank P value < 0.0001) with a C-index of 0.746 and a Brier score of 0.148.

Fig. 8.

Fig. 8

Open in a new tab

The TCGA SKCM training set, validation set and external test sets all exhibit significant survival differences. A–D TCGA SKCM training set, TCGA SKCM validation set, GSE65904 and GSE54467 test sets

Confirmation using SKCM validation-set and two independent test-sets

The SKCM validation set was used with the other two GEO test sets to assess the predictive capability of the SVM risk subgroup prediction model. The results showed (Fig. 8B–D) that the SVM model was able to distinguish well between the risk subtypes of this SKCM validation set and the GEO test set (log-rank P value < 0.05). The evaluation metrics of the above datasets are listed in Table 2. The C-index of all these datasets exceeded 0.680, the Brier score was mostly less than 0.200, and the log-rank P value was less than 0.05. These results suggest that the candidate DEGs obtained using the LASSO logistic regression algorithm are robust for predicting AE multi-omics risk subtypes.

Table 2.

Performance of SKCM training set, validation set and GEO external test sets

Dataset C-index Brier score log-rank P value
SKCM training(60%) 0.746 0.148 < 0.0001
SKCM validation(40%) 0.728 0.197 0.0014
GSE65904 0.693 0.202 0.0004
GSE54467 0.683 0.097 0.0052
Open in a new tab

Discussion

The high heterogeneity of SKCM poses a great challenge for prognostic prediction, and with the generation and public availability of large amounts of omics data, analysis using multi-omics data may provide better strategies and help in the study of SKCM heterogeneity. In this study, we constructed a SKCM risk subtype typing model with good predictive performance (C-index of 0.746) based on mRNA, miRNA and DNA methylation expression data, and defined two risk subtypes (log-rank P value = 7.31 × 10–7). The model was also applied to the risk subtype prediction of two independent SKCM patient test datasets, all of which showed good predictive power (C-index > 0.680) and significant survival differences (log-rank P value < 0.01), indicating the reliability of our inferred risk subtypes.

The results also showed that the multi-omics risk subtypes obtained from a deep learning framework based on combined early and lately features fused with AE were an independent influence on the prognosis of SKCM patients (HR 1.979, 95% CI 1.256–3.118). By comparing the differences with other subtyping methods, it was found that the model was able to distinguish risk subtypes of SKCM patients better than multi-omics subtyping methods such as early feature fusion AE (C-index of 0.716), late feature fusion AE (C-index of 0.711) and linear PCA (C-index of 0.606), indicating that the deep learning proposed in this study framework in SKCM disease risk typing has unique advantages over traditional risk subtype typing methods.

In addition, bioinformatics analysis revealed differences in gene expression between the two risk subtypes, and among the most significant genes obtained, down-regulation of three genes, IL2RG, IL-21R, and SP140, was associated with the high-risk subtype of SKCM. Among them, the downregulation of IL2RG gene expression in the high-risk group was significantly associated with poor prognosis. The IL2RG gene, which is considered a pivotal gene in melanoma metastasis, was significantly upregulated in its expression level compared with normal samples, and may function through the JAK-STAT signaling pathway (Zhang et al. 2021a). The above evidence of association results suggests that IL2RG may play a crucial role in SKCM progression. IL-21 induced antitumor responses require the presence of its receptors IL-21R (Xue et al. 2019), and IL-21 regulates tumor cell proliferation or apoptosis by binding to IL-21R and activating the JAK/STAT signaling pathway (Monteleone et al. 2005; di Carlo et al. 2007), and multicenter phase II studies have demonstrated the efficacy of IL-21 in metastatic melanoma (Petrella et al. 2012). In the current study, high expression levels of SP140 were found to be associated with better prognosis. Tanagala et al. found that high expression of SP140 in metastatic melanoma was associated with high immune scores, CD8 + T cells, M1 macrophages and gamma delta T cells and likewise correlated with higher survival (Tanagala et al. 2022).

GO analysis of DEGs between risk subtypes has shown that these DEGs are mainly engaged in the regulation of T cell activation, leukocyte cell–cell adhesion and its regulation and other BPs related to immune responses. The CC and MF in GO analysis suggest that these DEGs are mainly enriched on major histocompatibility complex (MHC) class II protein complexes, molecular functions such as MHC class II protein complex binding and its receptor activity. MHC class II molecules are a class of heterodimeric cell surface glycoproteins. They are present on antigen-presenting cells (e.g., macrophages, dendritic cells and B lymphocytes), and their main role is to deliver antigen fragments of pathogenic origin to CD4 + T cells, thereby triggering an immune response (Miller et al. 2017). The recognition and attack of pathogens by the body's immune system can be effectively elicited through the MHC II-mediated antigen presentation process. In the tumor environment, tumor antigens may be presented through MHC II class molecules that activate CD4 + T cells to participate in the antitumor immune response (Mondello et al. 2020). Low MHC II expression may be associated with exacerbation of disease and immunosuppression, while high MHC II expression may predict better antitumor immune response and prognosis. Therefore, this study explored the association between 13 genes encoding MHC II and SKCM risk subtypes. These genes were all found to be expressed at significantly lower levels in high-risk subtypes (Fig. 9), and most of these genes were identified as DEGs, suggesting that downregulation of these genes in high-risk subtypes is associated with poor prognosis of SKCM, and Ji et al. and Chen et al. also showed this trend in their studies (Chen et al. 2019; Ji et al. 2022). MHC II molecules may serve as novel immunotherapeutic targets that are expected to improve treatment outcomes and the quality of survival in patients with SKCM.

Fig. 9.

Fig. 9

Open in a new tab

Expression of MHC class II molecule-related genes in different subtypes of SKCM. ****Represents P < 0.0001

KEGG analysis showed that DEGs between risk subtypes were significantly enriched in the cell adhesion molecule pathway. Cell adhesion molecules (CAM) are a class of protein molecules present on the cell surface that binds to other cells or the extracellular matrix (ECM) (Moin et al. 2021) and regulate cellular physiological processes. In the early stages of melanoma, melanoma cells may lose or acquire new CAM and thus lose their close association with the surrounding keratin-forming cells or basement membrane, allowing melanoma cells to proliferate and invade the dermis. In contrast, after further mutations in melanoma, the type and number of adhesion molecules on the melanoma cell surface are altered, promoting ECM degradation and crossing the extracellular matrix, eventually entering the circulation and metastasizing to other sites (D’Arcy and Kiel 2021). Understanding the CAM changes in melanoma is also crucial in finding ways to prevent and treat melanoma.

The role played by miRNAs in various BPs related to melanoma development and progression should not be overlooked. The correlation between the downregulation of miR-150-5p, miR-142-3p and miR-142-5p and the poorer prognosis was verified by Sun et al., compared to patients with poor prognosis, these miRNAs were upregulated in patients with good prognosis and the fold of difference was more than 1.5 (Tembe et al. 2015), which is highly consistent with the findings of the present study. More research evidence supports this idea, with Sun et al. finding that miR-150 is downregulated in melanoma, and that downregulated miR-150 promotes proliferation, migration, and invasion of melanoma cells (Sun et al. 2019). The miR-142-3p is thought to be an intermediate link in the activation of the melanocyte autophagy pathway by Licochalcone A. Licochalcone A can exert anti-cancer effects by inducing melanocyte autophagy and inhibiting the growth of melanoma cells by upregulating the expression of miR-142-3p and suppressing its target genes Rheb and mTOR signaling pathway (Zhang et al. 2020). Low levels of miR-142-5p expression were associated with low survival in gastric cancer patients, and similar results were observed in melanoma patients in this study. miR-142-5p-regulated target genes are involved in most oncogenic signaling pathways associated with SKCM, such as cell cycle, cell proliferation, MAPK, Wnt, and VEGF (Zhang et al. 2011).

In addition, miR-342 and miR-155 have also been closely associated with SKCM. miR-342 expression is downregulated in melanoma, and its upregulation may exert tumor suppressive effects in part by suppressing the expression of its target gene, zinc finger E-box binding homeobox 1 (ZEB1), which may regulate cell proliferation, colony formation, migration, invasion, epithelial-mesenchymal transition, and tumorigenicity to promote melanoma development and progression (Shi et al. 2018). While miR-155 is transferred to cancer-associated fibroblasts through exosomes secreted by melanocytes, elevated miR-155 expression levels can inhibit the suppressor of cytokine signaling 1 (SOCS1) expression, activate the JAK2/STAT3 signaling pathway, trigger reprogramming of normal fibroblasts into pro-angiogenic cancer-associated fibroblasts, and promote melanoma angiogenesis (Zhou et al. 2018). In comparison to benign nevi, miR-155-5p was significantly expressed in melanoma biopsy tissues (Lunavat et al. 2015; Gencia et al. 2020), but interestingly, downregulation of its expression was instead observed in the high-risk subtype of this study. The relationship between miR-155 and melanoma proliferation and metastasis remains to be further investigated.

SKCM is the most lethal type of invasive melanocytic carcinoma in skin cancer. This study is the first to apply a deep learning framework combining early feature fusion AE and late feature fusion AE to distinguish SKCM subtypes and validate it in an independent dataset, and the integrated bioinformatics analysis provides a different perspective on the understanding of SKCM prognosis. Risk typing of SKCM not only improves our understanding of the molecular mechanisms, but also helps to elucidate the possible pathogenesis of SKCM.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary file1 (CSV 41 KB) (41KB, csv)
Supplementary file2 (CSV 3 KB) (3.2KB, csv)
Supplementary file3 (CSV 175 KB) (175.4KB, csv)
Supplementary file4 (CSV 11 KB) (11.3KB, csv)

Acknowledgements

We would like to thank the TCGA database and GEO database platform, and the providers who uploaded SKCM datasets in the platform. Thanks are also due to all developers of the R packages covered in this paper.

Author contributions

Conception and design: LWJ. Data collection and interpretation: LWJ, PY, HQ. Statistical analysis: LWJ and PY. Writing original draft: LWJ, HQ and PY. Writing review and editing: all authors. Supervision: HQ. Final approval of manuscript: all authors.

Funding

This work was supported by the National Science Foundation of China (no. 81773312); Natural Science Foundation of Guangdong Province (no. 2015A030313517); Talents Recruitment Grant of “Yangfan Plan” of Guangdong Province (no. 201433005); and Social Development Science and Technology Project of Dongguan (no. 20211800905552).

Data availability

All data can be retrieved from the TCGA database (https://portal.gdc.cancer.gov) and the GEO database (https://www.ncbi.nlm.nih.gov).

Declarations

Conflict of interest

The authors declare no conflict of interest.

Footnotes

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Weijia Li and Qiao Huang have share first authorship.

References

  1. Avogadri F, Zappasodi R, Yang A et al (2014) Combination of Alphavirus replicon particle-based vaccination with immunomodulatory antibodies: therapeutic activity in the B16 melanoma mouse model and immune correlates. Cancer Immunol Res 2:448–458. 10.1158/2326-6066.CIR-13-0220 [DOI] [PMC free article] [PubMed] [Google Scholar]
  2. Chang C-Y, Park H, Malone DC et al (2020) Immune checkpoint inhibitors and immune-related adverse events in patients with advanced melanoma. JAMA Netw Open 3:e201611. 10.1001/jamanetworkopen.2020.1611 [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Chaudhary K, Poirion OB, Lu L, Garmire LX (2018) Deep Learning based multi-omics integration robustly predicts survival in liver cancer. Clin Cancer Res 24:1248–1259. 10.1158/1078-0432.CCR-17-0853 [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Chen Y-Y, Chang W-A, Lin E-S et al (2019) Expressions of HLA class II genes in cutaneous melanoma were associated with clinical outcome: bioinformatics approaches and systematic analysis of public microarray and RNA-Seq datasets. Diagnostics (basel) 9:59. 10.3390/diagnostics9020059 [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Chen L, Wen Y, Xiong J et al (2023) An immunogenic cell death-related gene signature reflects immune landscape and predicts prognosis in melanoma independently of BRAF V600E status. Biomed Res Int 2023:1189022. 10.1155/2023/1189022 [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. D’Arcy C, Kiel C (2021) Cell adhesion molecules in normal skin and melanoma. Biomolecules 11:1213. 10.3390/biom11081213 [DOI] [PMC free article] [PubMed] [Google Scholar]
  7. di Carlo E, de Totero D, Piazza T et al (2007) Role of IL-21 in immune-regulation and tumor immunotherapy. Cancer Immunol Immunother 56:1323–1334. 10.1007/s00262-007-0326-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  8. Gencia I, Baderca F, Avram S et al (2020) A preliminary study of microRNA expression in different types of primary melanoma. Bosn J Basic Med Sci 20:197–208. 10.17305/bjbms.2019.4271 [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Hao Y, Chen Y, He X et al (2020) Near-infrared responsive 5-fluorouracil and indocyanine green loaded MPEG-PCL nanoparticle integrated with dissolvable microneedle for skin cancer therapy. Bioact Mater 5:542–552. 10.1016/j.bioactmat.2020.04.002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Haymaker C, Johnson DH, Murthy R et al (2021) Tilsotolimod with ipilimumab drives tumor responses in anti-PD-1 refractory melanoma. Cancer Discov 11:1996. 10.1158/2159-8290.CD-20-1546 [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Ji Z-H, Ren W-Z, Yang S et al (2022) Identification of immune-related biomarkers associated with tumorigenesis and prognosis in skin cutaneous melanoma. Am J Cancer Res 12:1727–1739 [PMC free article] [PubMed] [Google Scholar]
  12. Kramer MA (1991) Nonlinear principal component analysis using autoassociative neural networks. AIChE J 37:233–243. 10.1002/aic.690370209 [Google Scholar]
  13. Lê S, Josse J, Husson F (2008) FactoMineR: an R package for multivariate analysis. J Stat Softw 25:1–18. 10.18637/jss.v025.i01 [Google Scholar]
  14. Leng D, Zheng L, Wen Y et al (2022) A benchmark study of deep learning-based multi-omics data fusion methods for cancer. Genome Biol 23:171. 10.1186/s13059-022-02739-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Longato E, Vettoretti M, Di Camillo B (2020) A practical perspective on the concordance index for the evaluation and selection of prognostic time-to-event models. J Biomed Inform 108:103496. 10.1016/j.jbi.2020.103496 [DOI] [PubMed] [Google Scholar]
  16. Lunavat TR, Cheng L, Kim D-K et al (2015) Small RNA deep sequencing discriminates subsets of extracellular vesicles released by melanoma cells—evidence of unique microRNA cargos. RNA Biol 12:810–823. 10.1080/15476286.2015.1056975 [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Lv J, Wang J, Shang X et al (2020) Survival prediction in patients with colon adenocarcinoma via multiomics data integration using a deep learning algorithm. Biosci Rep. 10.1042/BSR20201482 [DOI] [PMC free article] [PubMed]
  18. Lv H, Liu X, Zeng X et al (2022) Comprehensive analysis of cuproptosis-related genes in immune infiltration and prognosis in melanoma. Front Pharmacol 13:930041. 10.3389/fphar.2022.930041 [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Marie KL, Sassano A, Yang HH et al (2020) Melanoblast transcriptome analysis reveals pathways promoting melanoma metastasis. Nat Commun 11:333. 10.1038/s41467-019-14085-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  20. Meyer D, Dimitriadou E, Hornik K, et al (2014) Misc Functions of the Department of Statistics (e1071), TU Wien
  21. Miller D, Tallmadge RL, Binns M et al (2017) Polymorphism at expressed DQ and DR loci in five common equine MHC haplotypes. Immunogenetics 69:145–156. 10.1007/s00251-016-0964-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Moin AT, Sarkar B, Ullah MA et al (2021) In silico assessment of EpCAM transcriptional expression and determination of the prognostic biomarker for human lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC). Biochem Biophys Rep 27:101074. 10.1016/j.bbrep.2021.101074 [DOI] [PMC free article] [PubMed] [Google Scholar]
  23. Mondello P, Tadros S, Teater M et al (2020) Selective inhibition of HDAC3 targets synthetic vulnerabilities and activates immune surveillance in lymphoma. Cancer Discov 10:440–459. 10.1158/2159-8290.CD-19-0116 [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Monteleone G, Monteleone I, Fina D et al (2005) Interleukin-21 enhances T-helper cell type I signaling and interferon-γ production in Crohn’s disease. Gastroenterology 128:687–694. 10.1053/j.gastro.2004.12.042 [DOI] [PubMed] [Google Scholar]
  25. Pavey S, Pinder A, Fernando W et al (2020) Multiple interaction nodes define the postreplication repair response to UV-induced DNA damage that is defective in melanomas and correlated with UV signature mutation load. Mol Oncol 14:22. 10.1002/1878-0261.12601 [DOI] [PMC free article] [PubMed] [Google Scholar]
  26. Peng J, Wang T, Yue C et al (2022) PGAM5: A necroptosis gene associated with poor tumor prognosis that promotes cutaneous melanoma progression. Front Oncol 12:1004511. 10.3389/fonc.2022.1004511 [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Petrella TM, Tozer R, Belanger K et al (2012) Interleukin-21 has activity in patients with metastatic melanoma: a phase II study. J Clin Oncol 30:3396–3401. 10.1200/JCO.2011.40.0655 [DOI] [PubMed] [Google Scholar]
  28. Poirion OB, Chaudhary K, Garmire LX (2018) Deep Learning data integration for better risk stratification models of bladder cancer. AMIA Jt Summits Transl Sci Proc 2017:197–206 [PMC free article] [PubMed] [Google Scholar]
  29. Ritchie ME, Phipson B, Wu D et al (2015) limma powers differential expression analyses for RNA-sequencing and microarray studies. Nucleic Acids Res 43:e47. 10.1093/nar/gkv007 [DOI] [PMC free article] [PubMed] [Google Scholar]
  30. Schadendorf D, Van Akkooi ACJ, Berking C et al (2018) Melanoma. Lancet 392:971–984. 10.1016/S0140-6736(18)31559-9 [DOI] [PubMed] [Google Scholar]
  31. Schröder MS, Culhane AC, Quackenbush J, Haibe-Kains B (2011) survcomp: an R/Bioconductor package for performance assessment and comparison of survival models. Bioinformatics 27:3206–3208. 10.1093/bioinformatics/btr511 [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Shi Q, He Q, Wei J (2018) MicroRNA-342 prohibits proliferation and invasion of melanoma cells by directly targeting zinc-finger E-box-binding Homeobox 1. Oncol Res 26:1447–1455. 10.3727/096504018X15193823766141 [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Siegel RL, Miller KD, Jemal A (2017) Cancer Statistics, 2017. CA Cancer J Clin 67:10. 10.3322/caac.21387 [DOI] [PubMed] [Google Scholar]
  34. Sun X, Zhang C, Cao Y, Liu E (2019) miR-150 suppresses tumor growth in melanoma through downregulation of MYB. Oncol Res 27:317–323. 10.3727/096504018X15228863026239 [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Tanagala KKK, Morin-Baxter J, Carvajal R et al (2022) SP140 inhibits STAT1 signaling, induces IFN-γ in tumor-associated macrophages, and is a predictive biomarker of immunotherapy response. J Immunother Cancer 10:e005088. 10.1136/jitc-2022-005088 [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Tay JK, Narasimhan B, Hastie T (2023) Elastic net regularization paths for all generalized linear models. J Stat Softw 106:1–31. 10.18637/jss.v106.i01 [DOI] [PMC free article] [PubMed] [Google Scholar]
  37. Team R (2014) R: A language and environment for statistical computing. MSOR connections
  38. Tembe V, Schramm S-J, Stark MS et al (2015) MicroRNA and mRNA expression profiling in metastatic melanoma reveal associations with BRAF mutation and patient prognosis. Pigment Cell Melanoma Res 28:254–266. 10.1111/pcmr.12343 [DOI] [PubMed] [Google Scholar]
  39. Tian Y, Morris TJ, Webster AP et al (2017) ChAMP: updated methylation analysis pipeline for Illumina BeadChips. Bioinformatics 33:3982–3984. 10.1093/bioinformatics/btx513 [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Wan Q, Wei R, Wei X, Deng Y (2022) Crosstalk of four kinds of cell deaths defines subtypes of cutaneous melanoma for precise immunotherapy and chemotherapy. Front Immunol 13:998454. 10.3389/fimmu.2022.998454 [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Wang T-H, Lee C-Y, Lee T-Y et al (2021) Biomarker identification through multiomics data analysis of prostate cancer prognostication using a deep learning model and similarity network fusion. Cancers (basel) 13:2528. 10.3390/cancers13112528 [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Wu T, Hu E, Xu S et al (2021) clusterProfiler 40: A universal enrichment tool for interpreting omics data. Innovation (camb) 2:100141. 10.1016/j.xinn.2021.100141 [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. Xiong J, Bing Z, Guo S (2019) Observed survival interval: a supplement to TCGA pan-cancer clinical data resource. Cancers (basel) 11:280. 10.3390/cancers11030280 [DOI] [PMC free article] [PubMed] [Google Scholar]
  44. Xue D, Yang P, Wei Q et al (2019) IL-21/IL-21R inhibit tumor growth and invasion in non-small cell lung cancer cells via suppressing Wnt/β-catenin signaling and PD-L1 expression. Int J Mol Med 44:1697–1706. 10.3892/ijmm.2019.4354 [DOI] [PMC free article] [PubMed] [Google Scholar]
  45. Yang W, Jiang J, Schnellinger EM et al (2022) Modified Brier score for evaluating prediction accuracy for binary outcomes. Stat Methods Med Res 31:2287–2296. 10.1177/09622802221122391 [DOI] [PMC free article] [PubMed] [Google Scholar]
  46. Zhang X, Yan Z, Zhang J et al (2011) Combination of hsa-miR-375 and hsa-miR-142-5p as a predictor for recurrence risk in gastric cancer patients following surgical resection. Ann Oncol 22:2257–2266. 10.1093/annonc/mdq758 [DOI] [PubMed] [Google Scholar]
  47. Zhang L, Lv C, Jin Y et al (2018) Deep learning-based multi-omics data integration reveals two prognostic subtypes in high-risk neuroblastoma. Front Genet 9:477. 10.3389/fgene.2018.00477 [DOI] [PMC free article] [PubMed] [Google Scholar]
  48. Zhang Y, Gao M, Chen L et al (2020) Licochalcone A restrains microphthalmia-associated transcription factor expression and growth by activating autophagy in melanoma cells via miR-142-3p/Rheb/mTOR pathway. Phytother Res 34:349–358. 10.1002/ptr.6525 [DOI] [PubMed] [Google Scholar]
  49. Zhang C, Dang D, Cong L et al (2021a) Pivotal factors associated with the immunosuppressive tumor microenvironment and melanoma metastasis. Cancer Med 10:4710–4720. 10.1002/cam4.3963 [DOI] [PMC free article] [PubMed] [Google Scholar]
  50. Zhang E, Chen Y, Bao S et al (2021b) Identification of subgroups along the glycolysis-cholesterol synthesis axis and the development of an associated prognostic risk model. Hum Genomics 15:53. 10.1186/s40246-021-00350-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  51. Zhang X, Wang J, Lu J et al (2021c) Robust prognostic subtyping of muscle-invasive bladder cancer revealed by deep learning-based multi-omics data integration. Front Oncol 11:689626. 10.3389/fonc.2021.689626 [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Zhou X, Yan T, Huang C et al (2018) Melanoma cell-secreted exosomal miR-155-5p induce proangiogenic switch of cancer-associated fibroblasts via SOCS1/JAK2/STAT3 signaling pathway. J Exp Clin Cancer Res 37:242. 10.1186/s13046-018-0911-3 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Citations

  1. Lv J, Wang J, Shang X et al (2020) Survival prediction in patients with colon adenocarcinoma via multiomics data integration using a deep learning algorithm. Biosci Rep. 10.1042/BSR20201482 [DOI] [PMC free article] [PubMed]

Supplementary Materials

Supplementary file1 (CSV 41 KB) (41KB, csv)
Supplementary file2 (CSV 3 KB) (3.2KB, csv)
Supplementary file3 (CSV 175 KB) (175.4KB, csv)
Supplementary file4 (CSV 11 KB) (11.3KB, csv)

Data Availability Statement

All data can be retrieved from the TCGA database (https://portal.gdc.cancer.gov) and the GEO database (https://www.ncbi.nlm.nih.gov).

Articles from Journal of Cancer Research and Clinical Oncology are provided here courtesy of Springer

RESOURCES