Assessment of risk based on variant pathways and establishment of an artificial neural network model of thyroid cancer

The RNASeq data of thyroid cancer, including 482 thyroid cancer samples, were downloaded from TCGA (https://​portal.​gdc.​cancer.​gov/​) in April 2017, based on the thca_tcga_pub_rna_seq_v2_mrna platform. The clinical information of thyroid cancer samples downloaded from the TCGA database was shown in Additional file 1: Table S1. The downloaded data had been preprocessed.

Based on the clinical data and information, the thyroid cancer samples downloaded from TCGA database were divided into low-risk (including T0, T1, N0, and M0) and high-risk (including T2, N1, M1, and above stages) groups according to the International Union Against Cancer (UICC) tumor–node–metastasis (TNM) classification. Finally, 114 high-risk samples and 368 low-risk samples were distinguished.

To eliminate the inherent expression differences between genes, the low-risk group was used as control to normalize the expression value of all samples using Z-score transformation [20]. The expression values with more than 2.5-fold change in terms of the standard deviation were defined as abnormal and subjected to correction. The DEGs between high-risk samples and low-risk samples were then identified using the limma package (version 3.10.3, http://​www.​bioconductor.​org/​packages/​2.​9/​bioc/​html/​limma.​html) [21]. P < 0.05 and coefficient of variance (CV) > 33.8 or < − 35.1 were used as the cut-off values to identify the DEGs.

The coexpression analysis of different genes between low-risk and high-risk samples was evaluated using Pearson correlation coefficient. R > 0.5 was identified as positive correlation, whereas R < − 0.5 indicated negative correlation. In both low-risk and high-risk samples, the coexpressed genes were defined as stable pairs, whereas the coexpressed genes found only in one of the groups were considered specific pairs. The stable pairs were recognized as the key genes with important functions, whose coexpression was not markedly affected by environment or disease stimulation. Nevertheless, the coexpression of specific pairs would change during tumor initiation and progression, which could be used for evaluating the cancer risk. After identifying the stable and specific pairs, unsupervised hierarchical clustering analysis [22] for the genes in these pairs was performed using the heatmap2 package in R [23], for distinguishing the samples with different cancer risks. In addition, based on the coexpression relationships, the respective gene coexpression networks under the low-risk and high-risk status were constructed using Cytoscape 3.4.0 [24]. Two topological properties, including average shortest path (ASLP) and degree distribution, were then analyzed to measure the connectivity of the network.

To further analyze the biological functions of the genes in stable pairs and specific pairs, KEGG pathway enrichment analysis was carried out using Fisher’s exact test in the Database for annotation, visualization, and integrated discovery (DAVID) online tool [13]. A value of P < 0.05 was used as the cut-off value. As these DEGs were associated with the risk of cancer, these significantly enriched pathways might be used for the evaluation of gene functions related to cancer risk. Based on the expression of DEGs in each sample, the functional deviation score for significantly enriched pathways was calculated using the Euclidean distance algorithm according to the following formula:

Based on the DEGs and functional pathway analysis, an ANN model was established using the supervised classification method, with multilayer perceptron (MLP) algorithm [25]. Meanwhile, the ROC curve was drawn with five-time cross validation to evaluate the classification efficiency of ANN model. Meanwhile, logistic regression as a reference was also analyzed to evaluate the efficiency of ANN model, based on variant pathways for risk assessment of thyroid cancer.

Using the above ANN model, 482 thyroid cancer samples obtained from the TCGA database was classified into two groups with different cancer risks. Using the survival data for these samples, survival analysis for the low-risk and high-risk groups was conducted using the survival package in R, and significant differences in survival times between two groups were determined using log-rank test. To further validate the efficiency of the ANN model for risk assessment, an independent dataset, with accession number GSE34289 deposited by Alexander et al. [19], was downloaded from the Gene Expression Omnibus (GEO, https://​www.​ncbi.​nlm.​nih.​gov/​geo/​query/​acc.​cgi), generated on the GPL14961 platform ([HuEx-1_0-st] Affymetrix Human Exon 1.0 ST Array [transcript (gene) version]). A total of 318 thyroid cancer samples were used in this study, including 40 benign, 233 indeterminate, and 45 malignant samples. The clinical information of thyroid cancer samples downloaded from the GEO database was shown in Additional file 2: Table S2. The downloaded data had been preprocessed. The risk for these 318 thyroid cancer samples was then predicted using this ANN model.

Pearson correlation coefficient was used to evaluate the correlation between a given pair of DEGs. Figure 1 displays the heat map of the top 30 risk-related DEGs with the highest correlation. According to the coexpression relationships between the gene pairs, the respective gene coexpression networks for the DEGs between low-risk and high-risk groups were constructed (Fig. 2a and b). From the topological perspective of coexpression networks, we found that these risk-related DEGs had a high degree of aggregation and central. Compared with the low-risk group, the high-risk group had more low-degree nodes (Fig. 2c), suggesting that, with increasing risk of thyroid cancer, the network node degree distribution gradually decreased and the association between the genes was gradually lost. Meanwhile, compared with the low-risk group, the average shortest path length in the high-risk group was higher, indicating a decreased capacity of information transfer through the network with increasing risk of thyroid cancer (Fig. 2d). In addition, supervised hierarchical clustering analysis was performed for these coexpressed DEGs, and the results showed that the samples with different cancer risk could be successfully distinguished (Fig. 3).

KEGG pathway enrichment analysis for risk-related DEGs revealed 21 significantly enriched pathways, including those for Graft-versus-host disease, Allograft rejection, Type I diabetes mellitus, Autoimmune thyroid disease, Viral myocarditis, and Herpes simplex infection (Table 1). In order to quantify these significantly enriched pathways, the functional deviation score for each pathway was calculated for the identification of variant pathways associated with cancer risk. Using Student’s t-test with a significance threshold of P < 0.01, 10 risk-related variant pathways with significant differences between low-risk and high-risk groups were identified, including those for Measles, Antigen processing and presentation, Rheumatoid arthritis, Phagosome, Systemic lupus erythematosus, Herpes simplex infection, Inflammatory bowel disease IBD, Tuberculosis, Type I diabetes mellitus, and Toxoplasmosis (Table 2), most of which were related to inflammatory and immune responses. An ANN model for risk assessment was then constructed, for which the above 10 risk-related pathways served as the input variables (Fig. 4). This ANN model consisted of an input layer, two hidden layers, and an output layer, respectively, corresponding to 15, 8, 5, and 1 neuron, respectively. To assess the performance of this ANN model, ROC curves for this model and logistic regression model (control) were prepared (Fig. 5). We found that the area under the receiver operating curve (AUC) for the ANN model and logistic regression model was 0.85 and 0.73, respectively, indicating that the ANN model based on variant pathways had a better prediction accuracy for risk assessment.

The 482 thyroid cancer samples were classified into low-risk and high-risk groups using ANN model. The results of survival analysis showed that the survival time of thyroid cancer samples in the low-risk group was significantly greater than that of the high-risk group samples (P = 0.0166, Fig. 6), indicating that this model could effectively distinguish the samples with different risks and yielded accurate predictions for the risk of thyroid cancer. Furthermore, the efficiency of this model was validated using the microarray dataset GSE34289, containing 318 samples representing different stages of thyroid cancer (benign, indeterminate, and malignant). The results showed that the accuracy of this model for low-risk (benign) samples and high-risk (indeterminate and malignant) samples was 77.5 and 86.0%, respectively.