Datasets

The RNA-Seq data (HTSeq-FPKM, 500 THCA samples, and 58 normal or adjacent non-tumorous samples obtained from 500 PTC patients) was retrieved from the Genomic Data Commons (GDC) data portal (https://portal.gdc.cancer.gov/). Notably, there were adjacent non-tumorous or normal samples available for 58 patients out of 500 patients only. In addition, manifest, metadata, clinical data, biospecimen files were also downloaded from the GDC data portal to obtain clinical information of the patients using Biospecimen Core Resource (BCR) IDs of patients. For every sample, mRNA expression of 60,483 RNA transcripts was reported in terms of FPKM (Fragments Per Kilobase Million) values. To ascertain the importance of the different types of transcripts, we segregated transcripts into subtypes like protein-coding, LincRNA, snoRNA, snRNA, and miRNA, etc. transcripts using annotation from GENCODE v22 (S1 Table).

Datasets for prediction models

Data processing

The contribution of the batch effect in the TCGA-THCA expression data was checked using TCGA Batch Effects Viewer [29]. Further, the FPKM values had a wide range of variation; the values were log2 transformed after the addition of 1.0 as a constant number to each of the values. Adding the constant one ensured that all of the transformed values would be positive. This approach is common in the literature related to the analysis of RNA expression [20, 30]. Thereafter, features with low variance (less than 0.25) were removed by employing caret package in R, followed by the Z-score normalization of data. The Eqs (1) and (2) were used for log transformation and normalization of data, respectively. (1) (2) In Eq (2), Z-score is the normalized scaled and centered score, x is the log-transformed transcript expression, μ is the mean of transcript’s expression in the training dataset, and σ is the standard deviation of a transcript in the training dataset. The log2 transformed validation data was Z-score normalized by taking the mean and standard deviation of training features.

Features filtering using threshold-based models

In the current study, we have considered the expression values of RNA transcripts as features for the analysis. In order to identify the classification potential of each RNA transcript, we employed a simple expression threshold-based approach similar to our previous study [21]. Briefly, in this approach, for every transcript, we designated a threshold, which determines whether a sample is in the early or late stage of cancer. The threshold was selected by iterating from the minimum to maximum expression of that transcript across all the patients. The threshold which gives maximum AUROC of classification between early and late-stage sample was reported. Briefly, if the mean expression of a transcript is greater in the early-stage than late-stage, and the log2 FPKM value of that transcript is found to be higher than the selected threshold for a given sample, then we assign that sample as early-stage otherwise late-stage. While, if the transcript’s average log2 FPKM value is greater in late-stage as compared to the early-stage, and the log2 FPKM of that transcript is greater than the threshold for a given sample, then we assign that sample as the late-stage sample otherwise as to the early-stage. Subsequently, threshold-based models were developed for each feature. Eventually, they were ranked based on their performance in the segregation of samples into different classes. Using the method mentioned above, AUROC was also calculated for cancer vs. normal samples.

Feature selection

To further improve the classification accuracy and to develop multiple-genes classification models, we used state-of-the-art techniques to select relevant features. First, we performed feature selection by employing an attribute evaluator named ‘SymmetricalUncertAttributeSetEval’ with the search method of ‘FCBFSearch’ of WEKA. The FCBF (Fast Correlation-Based Feature) algorithm uses mainly correlation to identify important discriminating features in high-dimensional datasets in reduced feature space [31]. Secondly, we employed the sklearn.feature_selection. F_ANOVA method of feature selection using the Scikit-learn package [32]. This method selects the features by computing F-statistics.

Third, we applied two more advanced feature selection methods for the features (both protein-coding and non-coding transcripts together), which performed best in comparison to other features. One was the (Support Vector Classifier) SVC with the L1 penalty using Scikit-learn [33], and the other was a Wrapper approach for feature selection. In a wrapper-based approach, human opinion dynamics optimizer has been used as a search algorithm to search through the space of possible feature subsets with the objective of maximizing MCC on the training set. This is an iterative algorithm in which each candidate solution represents a feature subset. The solution is encoded using a 100-bit binary vector where 1 and 0 indicate the presence and absence of a feature in a subset, respectively. The quality of features selected is evaluated using support vector machine (SVM) and 10-fold cross-validation. The details of the algorithm can be found in publications [34, 35]. The algorithm has been implemented in MATLAB® using LIBSVM and CODO (an open-source library hosted on https://github.com/rishemjit/CODO).

Implementation of machine learning techniques

We have developed machine learning models using two software; Scikit-learn package and Waikato Environment for Knowledge Analysis (WEKA) [36]. We employed SVC using Scikit-learn and used the Radial Basis Function (RBF) kernel of SVC at different parameters; g ∈ [10⁻³–10], c ∈ [1–10] using grid search for optimizing the SVC performance. In addition, random forests, sequential minimal optimization (SMO), Naïve Bayes, and J48 were employed using WEKA software.

Cross-validation technique

The validation is an indispensable part of evaluating the performance of a prediction method. In this direction, the ten-fold cross-validation technique is exploited to calculate the performance of early vs. late-stage and cancer vs. normal classification models. Here, the dataset is randomly divided into ten sets, from which nine sets are used as training sets and the leftover tenth set as a testing dataset. This process is repeated ten times in such a manner that each set is used once as a testing dataset.

Performance measures

In the present study, the performance of different models was measured by employing threshold-dependent and threshold-independent parameters. In case of threshold-dependent parameters, sensitivity (Sens), specificity (Spec), overall accuracy (Acc (%)), Matthews correlation coefficient (MCC), Precision, Recall and F1 score were calculated by using Eqs (3)–(9), respectively: (3) (4) (5) (6) (7) (8) (9) where FP, FN, TP, and TN are false positive predictions, false negative predictions, true positive and true negative, respectively.

We also calculated a threshold-independent parameter called AUROC value, which is computed from the receiver operating characteristic (ROC) plot in this study. The ROC curve is produced by plotting true positive rate against the false positive rate at different thresholds. In addition, to ascertain the reliability of prediction, we also calculated PPV (Positive Predictive Value) and NPV (Negative Predictive Value) at various thresholds using the probability score obtained by employing SVC.

Multiclass classification

The multiclass classification was implemented using the approach of one vs. rest multiclass SVC classifier employing the Scikit-learn package [37].

Functional enrichment of genes

Enrichment of genes was done using the Enrichr tool [38, 39]. Only those terms were selected for which adjusted p-value was less than 0.05. The Enrichr tool applies the Fisher’s exact test along with the adjustment using Bonferroni correction to give adjusted p-values.

Single RNA transcript-based stage classification

To rank the classification potential of each RNA transcript for segregating early and late-stage samples, we developed a stage classification method using the expression threshold of each RNA transcript (see methods). Briefly, in this approach, for every transcript/feature from 60,483 RNA transcripts, we designated a threshold, which determines whether a sample is in the early or late stage of cancer. This threshold was chosen by iterating from the minimum to maximum value of the expression for that transcript across all the patients. Here, a stage was assigned to a sample if the expression of an RNA transcript was more than the threshold; in case the RNA transcript was overexpressed in samples of that stage. Subsequently, RNA transcripts were ranked based on the discriminatory power in terms of AUROC of the threshold model. Eventually, we obtained 179 transcripts which had an AUROC score greater than equal to 0.60 and named as THCA-EL-AUROC. We have selected a cut off value of 0.60 to select the maximum features which performed better than random features (AUROC = 0.50). The THCA-EL-AUROC panel contains key transcripts that can help to discriminate early-stage samples from late-stage samples. Thus, they can be further explored for their potential biomarker capability in stage identification of PTC samples (S2 Table and S3 Fig), although more research is needed to confirm this hypothesis. The DCN protein (overexpressed in late-stage) coding transcript shows the highest AUROC of 0.66. It is a proteoglycan whose role is well established in discriminating benign and metastatic thyroid and other tumors [40, 41]. Out of 179 transcripts, 166 are protein-coding, six are lincRNA, and the other seven belong to other classes of non-coding transcripts (S2 Table). The 34 out of 179 RNA transcripts have a significant adjusted p-value of less than 0.05 (S2 Table).

The 166 protein-coding transcripts are significantly enriched in nine oncogenic signatures from MSigDB Database [42] (S3 Table) that points out these genes have also been previously implicated in many cancers. Additionally, the 166 protein-coding transcripts are enriched in many pathways of the KEGG database, such as Focal adhesion pathway (5% genes, adjusted p-value = 4.0e-5), PI3K-Akt signaling pathway (3.5% genes, adjusted p-value = 0.001), Proteoglycans in cancer (3% genes, adjusted p-value = 0.007). Further, the enriched terms of gene ontology for 166 protein-coding transcripts are mainly related to matrix organization and collagen binding (S4 Fig).

Stage classification model using multiple RNA transcripts

As shown in the above section and S2 Table, individual 179 RNA transcripts (THCA-EL-AUROC) have limited ability to classify early and late-stage samples with maximum AUROC 0.66. Therefore, to develop a model that can classify the stage of samples with high precision, we used the expression of multiple RNA transcripts. The stage classification models based on THCA-EL-AUROC (179 RNA transcripts) features were developed implementing a number of machine learning techniques. As shown in S4 Table, the SVC model achieved F1 score of 0.69 with AUROC of 0.72 (95% CI: 0.67–0.78) on training data and F1 score of 0.67 on the validation data with AUROC of 0.70 (95% CI: 0.59–0.82). There is a marginal improvement in the performance of the models developed using multiple transcripts (AUROC 0.70, S4 Table) as compared to the single gene model (AUROC 0.66 for DCN gene). We further tried few other intricate feature selection methods taking different subsets of transcripts in the following sections to further improve the performance of our models.

Protein-coding transcripts

From the previous results, it is evident that protein-coding transcripts were the major type of transcripts in THCA-EL-AUROC signature, therefore in this analysis, at first, we selected 19,814 protein-coding transcripts from 60,483 transcripts. Subsequently, different feature selection techniques like FCBF-WEKA (Fast correlation-based feature selection method present in WEKA) [43] and F_ANOVA [32] applied on these protein-coding transcripts. Thereafter, the prediction models were developed based on the selected set of features employing different machine learning techniques like SVC, RF, SMO, Naïve Bayes, and J48. The SVC model based on 37 features (selected by FCBF-WEKA) is the top performer with F1 score of 0.75 and 0.79 AUROC (95% CI: 0.74–0.84) on training data and F1 score of 0.72 with 0.66 AUROC (95% CI: 0.54–0.77) on the validation (THCA-EL-PC, Table 1) using 37 features obtained using SVC. There was a marginal increase in the performance in terms of accuracy, but the number of features was reduced to a reasonable extent as compared to THCA-EL-AUROC. The performance using other algorithms like Random Forest, Naïve Bayes, SMO, and J48 was lower on the validation dataset, as shown in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The performance measures of the prediction models developed based on 37-protein-coding mRNA feature set (THCA-EL-PC) selected by FCBF-WEKA feature selection method on training and validation dataset by implementing various machine-learning algorithms.

https://doi.org/10.1371/journal.pone.0231629.t001

To understand the interaction network among the key identified protein-coding genes, interaction analysis was performed in the STRING database [44] (S5 Fig) with THCA-EL-PC transcripts. On adding less than ten indirect nodes, we observed three important clusters enriched in different pathways. HIST1H2BJ, the transcript present in our signature forms a cluster, and this cluster is enriched in nucleosome cellular component. This cluster has also shown to be related to the progression of prostate cancer [45]. Another cluster of three genes is enriched in the dihydrolipoyl dehydrogenase complex (FDR <0.01), out of which DBT is present in our original signature. In addition, one more cluster of three genes is a part of the checkpoint clamp complex, out of which RAD1 is present in the original signature, and is involved in DNA damage response [46].

In addition, the top 100 features (S5 Table) were selected using the F_ANOVA feature selection method. The SVC based model has achieved F1 score of 0.71 with AUROC of 0.73 (95% CI: 0.68–0.79) on the training data, and F1 score of 0.68 with 0.71 AUROC (95% CI: 0.60–0.82) is obtained on the validation data (THCA-EL-F-PC, S6 Table).

Cancer hallmark based transcripts

Hanahan and Weinberg uncovered the importance of eight biological processes that played a vital role in tumor growth and metastatic propagation and called them as cancer hallmark processes [47]. Thus, genes involved in these processes could also act as key signature markers. Here, we have tried to ascertain only relevant transcripts from this subset of cancer hallmark genes.

To develop prediction models based on the cancer hallmark genes, initially, 4,814 cancer hallmark specific genes extracted from 60,483 transcripts. Subsequently, the number of features was reduced using various features selection techniques followed by the development of prediction models. Towards this, the 15 transcripts selected by FCBF-WEKA (THCA-EL-H) from cancer hallmark genes were used to develop models. The F1 score of 0.68 with AUROC of 0.71 (95% CI: 0.66–0.77) is attained on the training data, while the F1 score of 0.69 with AUROC of 0.73 (95% CI: 0.61–0.85) is obtained on the validation dataset (S7 Table). Out of 15 transcripts, two transcripts PROC and NLK (adjusted p-value = 0.002) are involved in the developmental pathway of the Wnt signaling, and are shown to be dysregulated in cancer [48]. The other two genes CYSLTR1 and ADRB1, are enriched in GPCRs terms (adjusted p-value = 0.04). CYSLTR1 is upregulated in colon cancer patients and associated with poor prognosis [49]. A similar performance is obtained on 50 genes selected using the F_ANOVA method (THCA-EL-FH, S8 Table).

Protein-coding and non-coding transcripts

Further to ascertain the role of both coding and non-coding transcripts, we explored all the 60,483 transcripts to identify relevant features that can segregate early and late stage samples. The 78 transcripts (THCA-EL-All-WEKA) were chosen using FCBF-WEKA based feature selection algorithm. The SVC model based on the THCA-EL-All-WEKA panel performed well and attained F1 score of 0.78 and 0.86 AUROC (95% CI: 0.83–0.90) on the training dataset and F1 score of 0.70 with 0.73 AUROC (95% CI: 0.63–0.84) on the validation dataset (S9 Table). Among the 78 selected features, 28 are protein-coding transcripts, 12 are long non-coding RNA, 12 are antisense transcripts, 11 are processed pseudogenes, and others are different non-coding RNAs (S10 Table).

Nest we applied another feature selection method called F_ANOVA to select top 100 features. The SVC model using these 100 features achieved F1 score of 0.72 with 0.77 AUROC (95% CI: 0.66–0.77) on the training data and F1 score of 0.63 with 0.68 AUROC (95% CI: 0.56–0.79) on the validation data (THCA-EL-All-F, S12 Table).

Additionally, the top 100 features selected using the F_ANOVA were further subjected to the second stage of feature selection. In this stage, a wrapper-based approach combining human opinion dynamics optimizer and SVC has been employed (see methods for details). The number of features was reduced to 27 (S13 Table, THCA-EL-CODO). It achieved F1 score of 0.59 and 0.72 AUROC (95% CI: 0.67–0.78) on the training set and F1 score of 0.58 and 0.73 AUROC (95% CI: 0.62–0.84) on the validation set using the Naïve Bayes Classifier (S14 Table).

From the above analysis, it has been observed that the prediction models based on both protein-coding and non-coding transcripts gave higher performance as compared to protein-coding and cancer hallmark protein-coding transcripts alone. One of our recent studies has shown that the prediction model based on the SVC-L1 feature selection method achieved higher performance with the minimum number of features [50]. Hence, we performed feature selection using the SVC with L1 penalty (see methods). SVC-L1 method resulted in 36 transcripts (shown in S15 Table). The SVC classifier based on the THCA-EL-SVC-L1 features attained F1 score of 0.75 with 0.73 AUROC (95% CI: 0.62–0.84) (Table 2) on the validation data. Notably, the prediction model based on 36 features is the best model among all the prediction models developed using different feature sets in classifying early and late-stage samples in terms of the number of features, accuracy, and F1 score on the validation dataset.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. The performance measures of the prediction models developed based 36-full feature set (THCA-EL-SVC-L1) selected by SVC-L1 on training and validation dataset by implementing various machine-learning algorithms.

https://doi.org/10.1371/journal.pone.0231629.t002

Further, we also calculated PPV and NPV on various thresholds of the SVC probability score (Table 3). On the training data, for the SVC score greater than 0.90, 161 early-stage samples are correctly predicted out of a total of 170 samples predicted as early-stage samples (PPV = 94.71%). In the case of late-stage samples, 60 out of 64 late-stage predicted samples are correct (NPV = 93.75%). In the case of validation data, the PPV is 85.71%, and the NPV is 66.67% (Table 3). This shows that at the SVC score of 0.90, there is a high probability of correct positive (early-stage) and negative (late-stage) prediction. At the threshold of 0.70, at which we presented other performance measures in Table 2, the PPV for training data is 93.03%, and NPV is 94.51%, while, in case of validation, the PPV is 83.87%, and NPV is 63.64% (Table 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. The performance of SVC based model at the different threshold in term of the probability of correct prediction, developed using 36-full feature set (THCA-EL-SVC-L1) on training and validation dataset.

https://doi.org/10.1371/journal.pone.0231629.t003

One of the advantages of this prediction model is that it resulted in a more balanced sensitivity and specificity along with higher F1 score on a smaller number of features (36 features, THCA-EL-SVC-L1) as compared to 78 features (THCA-EL-All-WEKA) selected by WEKA. These 36 transcripts consist of 17 protein-coding genes, six long non-coding RNAs, and rest other types of non-coding RNA transcripts (S15 Table). The TERT gene in this signature has been an important oncogene in the case of PTC [51]. Overexpression of TERT induced by the MAP pathway has shown to aggravate tumor development [52].

This model is deemed as the paramount model in our analysis for binary classification for early and late-stage samples. This also points out that both protein-coding and non-coding transcripts play an important role in tumor development.

Independent validation.

Eventually, to assess the classification potential of these features, validation performance is also evaluated on the external validation dataset, i.e., GSE48953, in addition to the independent dataset from TCGA. The GSE48953 dataset contains only 18 common features with that of 36-features signature. Therefore, we developed a model based on only those common 18 features. This model correctly predicted 70.6% (12 samples out of 17 samples) of early samples and 100% (3 samples) of late-stage samples of external validation dataset, as shown in Table 4. These results further strengthen and validate the classification potential of our signature for the segregation of early and late-stage samples.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. The performance measures of the SVC based model, developed using 18-features set on training (TCGA dataset) and external validation (GSE48953) dataset.

https://doi.org/10.1371/journal.pone.0231629.t004

The performance of different prediction models based on various feature sets for the segregation of early-stage and late-stage samples of training and validation datasets in terms of AUROC curves is depicted in Fig 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2.

The AUROC plot comparing the performance of prediction models based on different feature sets for segregating early and late-stage tissue samples of (A) Training dataset and (B) Validation dataset.

https://doi.org/10.1371/journal.pone.0231629.g002

Multiclass classification

One of the limitations of binary classification is that it would force even normal samples into either early or late-stage samples. Therefore, we implemented the multiclass classification by considering the normal samples available in TCGA for thyroid cancer.

From the above analysis, it has been observed that the prediction model developed based on features selected by the SVC-L1 method has higher performance with the minimum number of features. Therefore, we employed the SVC-L1 to select 107 RNA-transcripts (THCA-NEL-M, S16 Table) from 60,483 transcripts and classified normal, early and late-stage samples. This model obtained F1 score of 0.99, 0.88, 0.77 (Normal, Early, and Late) on the training dataset with AUROC of 0.99 (95% CI: 0.98–0.99), 0.93 (95% CI: 0.92–0.94), and 0.91 (95% CI: 0.90–0.93), respectively. The same model on the validation data attained F1 score of 0.88, 0.78 and 0.55 (Normal, Early and Late) with AUROC of 0.95 (95% CI: 0.84–1), 0.76 (95% CI: 0.66–0.85), and 0.72 (95% CI: 0.61–0.83), respectively (Table 5).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. The performance measures of the multiclass prediction model developed based on 107 features selected by SVC-L1 on training and validation dataset by implementing SVC.

https://doi.org/10.1371/journal.pone.0231629.t005

Discrimination of cancer vs. normal samples

Single gene-based models.

In this section, we ranked RNA transcripts on this basis of their AUROC for categorizing the PTCs samples from the non-tumorous samples. First, out of 60,483 features, those features with a variance of less than 0.02 were removed; thus, the number of transcripts reduced to 24,334. Among 24,334 transcripts, there are 8,180 RNA transcripts, which have AUROC of 0.6 or higher. Further, to identify highly discriminatory RNA transcripts, we selected 426 RNA transcripts having AUROC of 0.85 or greater (S17 Table). The overlapping sense transcript RP11-363E7.4 and protein-coding transcript FAM84A show AUROC of 0.96 and 0.95, respectively, for classifying cancer and normal samples. Gene Ontology pathway enrichment analysis for 386 protein-coding transcripts from the top 426 transcripts is shown in S6 Fig. Gene enrichment analysis revealed that 11 genes from the 386 transcripts are involved in the axon guidance pathway from KEGG. These genes comprising axon guidance pathway have been shown to play an important role in tumorigenesis [53]. Enrichment analysis shows that eight genes from 386 signatures have been also found in lung cancer-specific 86 genes defined in KEGG. Further, many of the genes are enriched in biological processes that regulate the expression of non-coding RNAs. The enriched cellular components’ terms are mostly related to interleukin receptor complexes, T-cell receptor complexes, and plasma membrane components. Serum IL-2 has been shown to discriminate patients with active thyroid cancer from the healthy with a sensitivity of 98%, and specificity of 58% [54]. Besides the protein-coding genes, 17 long non-coding RNAs were identified with AUROC above 0.85.

Among 17 lincRNAs, the PTCSC3 has been reported as highly thyroid-specific, and found to be downregulated in thyroid tumor tissues and thyroid cell lines [55]. LINC00936, RP11-774O3.3 and LINC00205 have been observed to be involved in other cancers [56, 57]. Taken together, it points out the literature validation of the key signatures identified in our study.

Protein-coding RNA transcript-based signatures. Our next goal was to develop a prediction model based on the least number of only protein-coding RNA features to classify cancer and normal samples with high precision. Therefore, we selected the top five features RELN, RASSF9, PLA2R1, MMRN1, and RPS6KA5 using F_ANOVA (THCA-CN-F). The prediction models were developed based on these five RNA-transcripts using various machine-learning algorithms. The SVC model attained F1 score of 0.98 and 0.97 AUROC (95% CI: 0.93–1) on the training dataset and F1 score of 0.97 and 0.99 AUROC (95% CI: 0.91–1) on the validation dataset (Table 6). As there was a large difference in the number of cancer samples and normal samples, we down-sampled the larger dataset and selected only 58 cancer samples corresponding to the 58 normal samples (THCA-CN-P). Subsequently, leave-one-out SVC model was developed using the same five features and obtained 0.99 AUROC. Also, we selected random 58 cancer samples (THCA-CN-R) and developed a leave-one-out cross-validation SVC model and obtained 0.96 AUROC (S7 Fig). As there is no outsized change in the performance by down-sampling the cancer samples, we incorporated the model with all the 500 samples as the final model.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. The performance measures of prediction models developed based on 5-protein coding transcripts (THCA-CN-F) feature set selected by F_ANOVA feature selection method for discriminating cancer and normal patients on training and independent validation dataset.

https://doi.org/10.1371/journal.pone.0231629.t006

Gene enrichment analysis revealed that RELN and RPS6KA5 are associated with activation of cyclic AMP (cAMP) response element-binding protein (CREB) transcription factor (adjusted p-value<0.01), which is responsible for tumor initiation, progression, and metastasis [58]. RELN is an extracellular glycoprotein that plays a vital role in neuronal migration and has been shown to be downregulated in many cancers [59, 60].

Web server implementation

We established a web server, CancerTSP (Thyroid cancer stage prediction), that implements models developed in the present study for investigation and estimation of cancer stage from the transcripts’ expression data. The CancerTSP consists of two key modules consisting of prediction and analysis.

The prediction module consists of two modules for predicting the stage of the THCA cancer sample. One of the models is based on protein-coding features (THCA-EL-PC), and the other model is based on both protein-coding and non-protein coding transcripts (THCA-All-SVC-L1). We also provide a third prediction module based on the THCA-NEL-M features, which can predict whether the sample is normal, early-stage, or late-stage sample. The results are displayed on the score thresholds that exhibited a minimum difference in recall and specificity with maximum accuracy. The user can change the threshold for less or more stringent results in terms of recall and specificity. The lower threshold will increase the recall but decrease the specificity and a higher threshold will do the opposite to prediction outcomes (S8 Fig).

The user needs to provide transcript expression (FPKM values) of potential biomarker genes for every patient. The number of patients corresponds to the number of columns in a file. The output includes a list of patients and corresponding predicting stage of cancer (early or late-stage) along with the prediction score (probability value).

Another module is dedicated for analysis which is helpful in evaluating the role of each transcript in discrimination of early-stage from the late-stage. This module gives p-value (calculated using Wilcoxon rank test) for each transcript that signifies whether the transcript’s expression varies in the early and late-stage significantly. It also gives expression threshold and classifying AUROC of each transcript along with the average expression of that gene in the early and late-stage of cancer. The CancerTSP webserver is available from URL http://webs.iiitd.edu.in/raghava/cancertsp/ for public use.

Limitation of the study

In this study, we have scrutinized potential transcriptomic signatures to distinguish early and late-stage samples of PTC. One of the limitations associated with these signatures is that they are derived from tissue samples only, which is an invasive technique for biomarker discovery. Further, the external dataset used in the current study contains only 20 samples. Thus, field can be advanced by adopting non-invasive biomarkers from specimens like blood, urine, cell-free DNA, etc. along with the validation on large sampled cohorts to confirm their clinical utility.