The potential prognostic value of Cartilage acidic protein 1 and secretoglobin family 1D member 1 in prostate cancer

doi:10.21203/rs.3.rs-3557321/v1

Download PDF

Article

The potential prognostic value of Cartilage acidic protein 1 and secretoglobin family 1D member 1 in prostate cancer

https://doi.org/10.21203/rs.3.rs-3557321/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Prostate cancer is a prevalent malignancy with a broad range of clinical outcomes, necessitating improved prognostic biomarkers for precise patient stratification and personalized treatment. This study leverages machine learning techniques to identify and validate novel prognostic biomarkers using DNAseq and RNAseq data for prostate cancer.

Methods

Whole genome sequencing and gene expression profiling in patients were used from TCGA to identify DEGs and genetic alterations. Next, deep learning was utilized to determine key genes. Gene Ontology, Reactom, GSEA, and Human Disease Ontology were employed to study the involved biological process and pathways. Survival analysis of patients with prostate cancer with reference on dysregulated genes was conducted using Cox regression and Kaplan-Meier analysis. The STRING database was used to build a protein-protein interaction (PPI) network. Moreover, candidate genes were subjected to machine learning -based analysis and the Receiver operating characteristic (ROC) curve.

Results

We identified a total of 609 DEGs in patients, of which 358 were downregulated and 251 were upregulated. Deep learning results identified 20 genes, and these were combined with the analysis of DNA-seq. Survival analysis of patients with prostatic cancer showed that dysregulated expression of ASB12, BLOC1S1, CRTAC1, KCNQ1, KISS1, M2T2A, RNF207, SCGB1D1, SLC13A2, SORBS1, TGFBR3, WSCD2, ANFKFY1, CRYBA4, MIR204, QRFP, SNX15, and YWHAH genes were related with a poor clinical prognosis. The combio-ROC curve analysis reveals ed that TGFRB3, SCGB1D1 and CRTAC1 were potential diagnostic biomarkers with a great higher sensitivity and AUC than currently available biomarkers. Also, the combination of SCGB1D1 and CRTAC1 demonstrated the greatest accuracy, sensitivity, and specificity towards diagnostic applications. The potential value of these markers was validated in six other datasets.

Conclusion

Our findings demonstrated the potential value of SCGB1D1 and CRTAC1 as novel biomarkers and therapeutic targets in prostate cancer which had a higher AUC, sensitivity, and specificity compared to PSA, indicating further functional investigations on the potential value of emerging markers in prostate cancer.

Biological sciences/Biotechnology

Biological sciences/Cancer

Biological sciences/Genetics

Health sciences/Biomarkers

prostate cancer

diagnostic biomarkers

prognostic biomarkers

DNAseq and RNAseq

Prostate cancer (PCa) is the second most prevalent malignancy among men worldwide. It is associated with a reasonable prognosis when it is diagnosed at early stages so accurate diagnostic approaches are of importance (1). Currently, prostate specific antigen (PSA) screening coupled with some biomarkers such as PCA3, AMACR or gene fusion events (e. g., TMPRSS2-ERG) are used for the diagnosis of PCa (2, 3). However, these have limitations and there remains a need to identify novel diagnostic biomarkers with more sensitivity and specificity. Although, some progress has been made in terms of diagnosis, prognosis and management of PCa remains a challenge (4). Current stratification methods include clinical staging based on PSA testing, digital rectal exam (DRE), Gleason Score and pathological staging based on results of prostate tissue examination after surgery. However, it is difficult to distinguish between low-risk PCa and prostate cancer with aggressive clinical behaviour. The distinction is important so as to choose the best treatment strategy because PCa is a heterogeneous cancer and may lead to overtreatment of an indolent prostate cancer or neglecting a PCa patient at risk of aggressive disease. Therefore, novel prognostic biomarkers are urgently required (5).

Transcriptome analysis have been widely used to identify biomarkers through high-throughput RNA-sequencing. Bioinformatics analysis of gene expression data could help to unravel underlying mechanisms of prostate cancer (6). Recently, Kumar et al. used PCa expression data on The cancer genome Atlas (TCGA) dataset and performed bioinformatics analysis to identify prognostic biomarkers. They suggested GNL3 and PA2G4 as novel prognostic biomarkers (7). Also, some researchers have used bioinformatics analysis and proposed a panel of miRNAs as a diagnostic tool to increase accuracy and specificity of prostate cancer diagnosis (8, 9). Artificial intelligence has emerged as a powerful approach to make predictive models that accurately identify most significant genes involved in cancer pathogenesis. There are a few studies that have applied machine learning to predict novel biomarkers in PCa. Hou et al. performed a meta-analysis to determine differentially expressed genes (DEGs) in the expression data of TCGA and GEO. Then they established a prognostic and diagnostic model to investigate the candidate genes among DEGs. They found C1QTNF3 as a promising biomarker for PCa(10). With all this, available diagnostic, and prognostic biomarkers such as PSA, prostate cancer antigen 3 (PCA3) gene, 4k score blood test, prostate health index (PHI) lacks enough sensitivity and specificity. There is a need for precise diagnostic and prognostic biomarkers identified by a more comprehensive manner(11). We used both DNA-seq and RNA-seq data of TCGA and applied a deep learning assisted approach to identify novel and reliable diagnostic and prognostic biomarkers in PCa. We identified a group of prognostic biomarkers of which further analysis showed that CRTAC1 and SCGB1D1 are potent novel diagnostic and prognostic biomarkers in PCa.

Data sets preparation and processing

Prostate DNA-seq and whole exome sequencing data, and RNA-seq data were obtained from the TCGA database (http://tcga-data.nci.nih.gov/tcga/). Expression data and clinical data of 549 samples including 498 cancer and 52 non-cancer samples were downloaded. RNA-seq data set was filtered to remove zero gene expressions and duplicates.

DNA-seq and whole exome sequencing data analysis

Data from the TCGA database was retrieved and processed using the R programming language for subsequent analysis. The data was obtained in the mutation annotation format (MAF), a standardized format utilized by TCGA to store and study various somatic mutations in cancer. The Maftools package in R programming was then utilized to analyse the MAF data of 490 patients.

Normalization of RNA-seq data and differential gene expression analysis

RNA-seq data set was normalized using DESEQ2 package in R 4.2.1 software. Data quality control was performed by plotting the principal component analysis (PCA) plot (Fig. 2B).

Then DESEQ2 package was used to get differentially expressed genes (DEGs) in PCa samples relative to non-cancer samples of TCGA data set. DEG were identified by filtering those with a p value < 0.05 and − 1.5 <|Log2FC (fold change) | < 1.5. visualization of DEGs were performed using volcano plot (Fig. 2A) and heatmap (supplementary Fig. 1). These plots were generated by R package.

Machine learning for identification of potential biomarkers

Deep learning is a superior subtype of machine learning that works based on artificial neural networks and can process huge amount of multi-variable data and identify the most effective features by itself. Deep learning utilizes backpropagation algorithms to efficiently calculates the gradient of the error function about the network weights(12). This method has been applied in cancer precision medicine for diagnosis, prognosis, and treatment decisions. For instance, Park et al. developed a deep learning prediction algorithm based on multi-omics data of non-small-cell lung cancer (NSCLC) (13). We used deep learning to identify most important biomarkers in 609 DEGs of TCGA data set to further analysing for prognostic and diagnostic functions of them.

Rapidminer(version 9.10) and R 4.2.3 were used to create the model. We set the parameters of activation function = Rectifier, epochs = 20, and rate = 0.01 for deep learning. In the following, we explain the standard workflow that was utilized to create, evaluate, and optimize methods.

As a common method, we first spilt the data source into the two types including training and test data to provide kind of independent evaluation levels. The created model was optimized using the training data and then was independently assessed using the test data.

We determined the train/test ratios of 70/30 for deep learning models. Model hyperparameter was tuned and a fixed optimal hyperparameter value model was trained again for each workflow, then sampled arbitrary from the data set. Further evaluation of model was done using the remained data.

Evaluation of model performance using a test set

We assessed the performance of model by 6 metrics including accuracy, MSE, RMSE, R², AUC, and PR-AUC. Accuracy is the most common used parameter in evaluation of algorithm efficiency and performance. Accuracy = (TP + TN) / (TP + TN + FP + FN)

Where:

TP - true positive

FP - false positive

TN - true negative

FN - false negative

MSE (Mean Squared Error) is another common metrics of algorithm performance which is indicative of the amount of error in a model.

MSE = (1/n) * Σ (actual – forecast)2

where:

Σ – a symbol that means “sum”

n – sample size

actual – the actual data value

forecast – the predicted data value

RMSE (Root Mean Square Error): It represents the proximity of predictions to the original values using Euclidean distance.

AUC (Area Under the Curve): It shows the degree of separability. It demonstrates the capability of the model to distinguish classes.

R² (R-Squared) is the main parameter in evaluation of a model

R² = 1 − Unexplained Variation / Total Variation. It is the coefficient of determination and tells us the percentage variation in y explained by x-variables.

PR-AUC (Precision recall-AUC): It illustrates if the model predicts the positive and negative classes correctly or not.

Functional enrichment analysis of candidate genes

Further functional enrichment analysis was accomplished to unravel the function of identified genes by deep learning. We used a R package, ‘clusterProfiler’, to do functional enrichment analysis based on Gene Ontology (GO), GSEA signalling pathways, Human Disease Ontology (DO), and Kyoto Encyclopedia of Genes and Genomes (KEGG). In Go analysis cellular components (CC), biological process (BP) and molecular function (MF) are described. Statistically significant results were defined by using the criteria, adjust p < 0.05, score > 1 and False discovery rate (FDR) < 0.25.

Protein-protein interaction network analysis

Candidate genes identified by deep learning methods were applied to the STRING online database to evaluate the interactions between them. We set the high confidence of 0.7 for interaction score to export the PPI network. Then, Cytoscape 3.10.0 software was used to obtain hub genes with important roles.

Survival analysis

Survival analysis was performed using gene expression profile of TCGA PCa and non-cancer samples related to identified DEGs to assess potential prognostic biomarkers. We considered to perform Cox regression analysis in two ways. First, we applied Cox to evaluate the effects of identified DEGs on survival using patient vital-status and days of last follow-up. Then we used Cox regression to investigate the effects of 7 confounding factors including age, race, Gleason-score, case-control, and serum PSA-value on survival of patients. Furthermore, we first plotted Kaplan-Meier for all 609 PCa DEGs and then for 20 key genes identified by deep learning to visualize the potential prognostic biomarkers. Plotted genes with p value < 0.05 were considered as possible prognostic markers.

Unravelling potential diagnostic biomarkers

Novel diagnostic biomarkers are in demand to aid early diagnosis of PCa. We performed combined receiver operating characteristic (combio-ROC) curve analysis using R 4.2.3 package ‘combioROC’ to unravel the diagnostic potential of identified genes or combination of them. For this purpose, we first performed combio-ROC analysis for all 20 identified key genes by deep learning. Then, we applied it for those with variable importance of 0.95 or higher. Moreover, to evaluate the potential diagnostic biomarkers, parameters including sensitivity (SE), specificity (SP), area under the ROC curve (AUC), positive predictive value, negative predictive value, and cut-off value were assessed.

Correlation of clinical/demographic characteristics and genes with PCa

We investigated the correlation between clinical data including age, race, Gleason-score, case-control, clinical -T, clinical -M, and serum PSA value with identified genes using R.4.2.3. Moreover, we performed correlation matrix analysis to evaluate the association between PCa and clinical data. The correlation coefficient is scaled between − 1 to 1. The correlation was determined weak when less than ± 0.3, moderate when between ± 0.3 and ± 0.5, and significant when more than ± 0.5.

Analysis of DNA-seq

The Maftools package in R programming was used to analyze MAF data from 490 patients. Figure 1 displays various plots, such as oncoplots, plotmafSummary, Transition, and Transversions reports, Plotting VAF (Variant Allele Frequencies), Drug-Gene Interactions, Somatic Interactions reports, and Oncogenic Signalling Pathways, to picture the MAF distribution between different groups. Based on Fig. 1A, missense mutations were the most frequent type of mutations. Furthermore, mutations in the TTN, TP53, and SPOP genes were present in 10–11% of patients. Many variants were involved in protein-homeostasis/ubiquitination, genome integrity, and chromatin-SWI/SNF complex (Fig. 1B). The Variant Allele Frequencies plot can estimate the clonal status of the top mutated genes, where clonal genes have an mean allele frequency of about 50% in pure samples.

TP53 was observed to have clonal status in tumor tissue in Fig. 1C. Somatic Interactions analysis showed exclusive or co-occurring events in Fig. 1D. Mutually exclusive events occur when mutations in one gene avoid the arising of mutations in another gene while co-occurring events happen when mutations in two or more genes arise together more commonly than would be anticipated by chance. Identifying mutually exclusive genes shows that these genes may take apart in the same pathway or process, and there could be a functional overlap between them. On the other side, recognizing genes that co-occur may suggest that they collaborate to aid tumour growth, or that their cumulative impact is necessary for cancer development. Figure 1E illustrates the interaction between genes and drugs that target Tyrosine kinase, serine-threonine-kinase, tumour suppressor, and other related processes. The involvement of mutated genes in Oca across different Oncogenic Signalling Pathways, including Wnt, RTK-RAS, Hippo, Notch, and others, is depicted in the Fig. 1F. Figure 1G provided an instance of genes that function as either tumour suppressors or oncogenes within the WNT pathway.

Identification of differentially expressed genes (DEGs) in PCa

To find dysregulated genes in PCa, TCGA data set containing 52 non-cancer samples and 498 tumour samples were used to obtain DEGs with DESEQ2 package. We identified a total of 609 DEGs which 358 of them were downregulated and 251 of them were upregulated. To further show the distribution of DEGs, a heatmap was plotted using heatmap package in R (supplementary Fig. 3) and a volcano plot was plotted using EnhancedVolcano package (Fig. 2A).

Determination of key DEGs in PCa by deep learning

Deep learning method was used to identify the most important DEGs of PC that were obtained before. We spilt the TCGA prostate cancer dataset with 549 samples with the 70/30 ratio into training set and test set. As depicted in table (1), 20 genes of 609 PC DEGs were identified as most significant genes in PCa. The SCNN1B gene was determined as most important gene in PCa with the variable importance of 1. SLC13A2 and SCGB1D1 with variable importance of 0.996 and 0.985 were the next important genes. To further visualize the results, heatmap of 20 identified key genes were plotted using R. 4.2.3(Fig. 2C). The performance of the model was confirmed with the AUC = 1, R² = 0.96, RMSE = 0.054, MSE = 0.002, PR-AUC = 1, and Accuracy = 93.8%. We consider 20 identified genes as candidate genes for further analysis.

Evaluation of candidate genes in DNA-seq data

Identified genes by deep learning were assessed in DNA-seq data analysis results to evaluate them at the genomic level. For this purpose, we identified and visualized 26 different genetic variants of 20 candidate genes as depicted in the venn diagram (Fig. 2D). As represented in Fig. 2E, the most frequent mutation among candidate genes were missense mutations. SCNN1B and CRTAC1 were mutated in about 17% of patients. Notably, SCNN1B mutations have been identified in people with pseudohypoaldosteronism type 1 (PHA1) also, 16 mutations of SCNN1B is related with Liddle syndrome(14). Also, COSMIC data base showed that CRTAC1 represent point mutations in 5.05% of prostate tissue samples and copy number variation in 0.11 of samples(15). Moreover, SORBS1 and C11orf24 mutations were presented in 11% of patients in our study (Fig. 2F). Figure 2G shows druggable genes that target kinase, serine-threonine-kinase, cell surface, and other related processes.

Functional enrichment analysis of candidate genes

To further explore the function of candidate genes and different pathways they contribute, functional enrichment analysis was performed using ‘clusterProfiler’ package in R.

ClusterProfiler enable us to find shared function and pathways of genes through different biological ontologies such as gene ontology (GO) that annotates genes to biological process (BP), cellular component (CC), and molecular function (MF), also Kyoto Encyclopedia of Genes and Genomes (KEGG) that annotates genes to pathways, and Human Disease Ontology (DO) that annotates genes to diseases. As shown in Fig. 3, GO enrichment of 20 identified genes showed that terms like dicarboxylate symporter activity, transforming growth factor beta receptor activity, and kisseptin receptor binding were significantly enriched in molecular function (MF). In biological process (BP), most genes were involved in response to response to peptide hormone, response to peptide. Also, terms like myeloid cell homeostasis and erythrocyte homeostasis were significantly enriched (Fig. 3A-3C). Do enrichment results showed that candidate genes are involved in hypertension, ovarian dysfunction, hypergonadism, renal carcinoma, vascular diseases and other diseases that are shown in Fig. 3D .Also, KEGG pathway enrichment of key genes demonstrated that Hippo signalling pathway, Aldosterone-regulated sodium reabsorption, Vibrio cholera infection, and GnRH secretion pathways were significantly enriched which is indicative that identified genes are significantly involved in pathogenesis of prostate cancer (Fig. 3E).

STRING database was utilized to construct PPI network of 20 identified genes. Then the PPI network was submitted to the Cytoscape software to visualize the hub genes. Cytohubba app was used to obtain 10 top hub genes. Hub genes were imported into the STRING to reanalyse the network. The constructed network of hub genes is depicted in Fig. 4A.

Identification of prognostic biomarkers

Cox regression analysis was utilized to assess the effects of differentially expressed genes on survival of patients with PCa. Results showed meaningless relation between DEGs and survival of patients in general. Also, we applied Cox analysis to evaluate the impact of confounding factors such as age, race, Gleason-score, clinical -M, clinical -T, case-control, and serum PSA-value on survival of patients which outcomes showed no meaningful relation between these factors and survival of patients with prostate cancer.

We conducted Cox regression analysis for 609 DEGs. Then Kaplan-Meier analysis was used to identify potential diagnostic biomarkers. Results showed that six genes including ANKFY1, CRYBA4, MIR204, QRFP, SNX15, YWHAH are correlated with survival of patients with PCa and could function as prognostic biomarkers (supplementary Fig. 2). Moreover, Kaplan-Meier analysis were performed separately for all 20 key genes identified by deep learning to discover potential prognostic biomarkers. Results showed that, dysregulated expression of 12 genes including ASB12, BLOC1S1, CRTAC1, KCNQ1, KISS1, M2T2A, RNF207, SCGB1D1, SLC13A2, SORBS1, TGFBR3, and WSCD2 genes are correlated with poor survival of PCa and identified as potential prognostic biomarkers (supplementary Fig. 4). Then, total of 18 identified potential biomarkers were searched for any previous report;` six of them including KCNQ1, M2T2A, SORBS1, MIR204, RNF207, and QRFP were reported in different studies as prognostic biomarkers. Therefore, other identified genes were considered as novel prognostic biomarkers. Furthermore, we assessed the function of novel diagnostic biomarkers and their role in other diseases which are represented in Table 2. These results show critical role of identified biomarkers in PCa progression and pathogenesis.

Table 1

List of 20 identified genes by deep learning and their coefficients.
Target category	Variable name	Coefficients of each variable	Performance of model
Important genes in prostate cancer Identified by Deep learning	SCNN1B	1	MSE: 0.002 R^2: 0.96 AUC: 1 Accuracy:93.8% RMSE:0.054 PR-AUC:1
	SLC13A2	0.996
	SCGB1D1	0.985
	CRTAC1	0.967
	MOB1B	0.963
	TGFBR3	0.942
	RNF207	0.941
	KERA	0.941
	SORBS1	0.931
	KCNQ1	0.930
	RASSF4	0.760
	C11orf24	0.759
	WSCD2	0.751
	MZT2A	0.744
	BLOC1S1	0.741
	ASB12	0.739
	REEP4	0.739
	CKAP2	0.730
	CEP170P1	0.695
	KISS1	0.676

Table 2

The identified novel prognostic biomarkers and their functions.
Gene name	Gene function	Function in other diseases
ASB12	couple suppressor of cytokine signalling (SOCS) proteins/ mediates the ubiquitination /expressed abundantly in testes(36)	Laryngotracheomalacia(35)
BLOC1S1	play distinct roles in endosome and lysosome biology	Liver regeneration disorder(35) prognostic function in adult acute myeloid leukemia(AML)(37)
CRTAC1	May involved in cell-cell or cell-matrix interactions. up regulated in mesenchymal stem cells	Biomarker of osteoarthritis severity and progression(34) Neurofibromatosis(21)
MZT2A	Enhances NSCLC survival and invasion by increasing Akt phosphorylation	MZT2A plays an important role in the progress of some tumours/ prognostic value in non–small-cell lung cancer (NSCLC)(38, 39)
SCGB1D1	ortholog of prostatein, probably is transcriptionally regulated by steroid hormones, ability to bind androgens, other steroids, may bind and concentrate estramustine, a chemotherapeutic drug used for prostate cancer.	Ovarian carcinoma (33)
SLC13A2	sodium-coupled citrate transporter	Upregulated in colorectal cancer(40)
WSCD2	Probably is an integral component of membrane	Breast cancer(41)
ANKFY1	essential for retinal endothelial cell proliferation and migration	Genetic Steroid-Resistant Nephrotic Syndrome and Human Monocytic Ehrlichiosis(42)
CRYBA4	structural components of the vertebrate eye lens	Prognostic gene in renal cell carcinoma(43)
SNX15	Involved in intracellular trafficking	Genetic marker in cholangiocarcinoma / Gastric cancer prognosis (44, 45)
YWHAH	mediate signal transduction by binding to phosphoserine-containing proteins	Oncogene in several types of cancers/involved in cell growth, apoptosis, cell growth, migration, and invasion (46)

ROC analyses were performed to unravel the potency of identified prognostic biomarkers and to assess their potential in diagnosis of PCa. Results demonstrated that the CRTAC1(AUC = 0.89, SE = 0.78, SP = 0.9), SCGB1D1 (AUC = 0.85, SE = 0.85, SP = 0.71), and TGFBR1 (AUC = 0.89, SE = 0.81, SP = 0.84) are potential diagnostic biomarkers. The ROC curves and related Kaplan-Meier of these genes are depicted in Fig. 4B and Fig. 4C. Also, the combination of CRTAC1 and MOB1B with coefficients of -1.29 and − 1.22 (AUC = 0.9, SE = 0.93, and SP = 0.75) have diagnostic capability. Combination of CRTAC1 and SCGB1D1 with coefficients of -1.22 and − 1.69, AUC = 0.91, SE = 0.94, and SP = 0.76 also, combination of C11orf24 and SORBS1 with coefficients of -0.34 and − 2.25, AUC = 0.9, SE = 0.93, and SP = 0.75 identified as potential diagnostic biomarkers. Moreover, combination of three genes including CRTAC1, SCGB1D1, and SCNN1B were determined as potential biomarker for diagnosis of PCa. The coefficients of genes were − 1.27, -2.03. 0.34 with AUC = 0.92, SE = 0.9, SP = 0.8. Also, combination of CRTAC1, MOB1B, SCGB1D1, and SCNN1B with coefficients of -1.24, -0.27, -1.89, and 0.42, AUC = 0.92, SE = 0.91, SP = 0.8 have diagnostic ability in PCa (Supplementary table 2).

Correlation analysis of clinical/demographic characteristics and genes with PCa

Correlation matrix analysis results showed no significant association between identified genes and clinical/demographic data. Also, correlation analysis of clinical data and PCa showed no meaningful association between them (Fig. 9).

Potential value of CRTAC1 and SCGB1D1 as prognostic biomarkers

ROC analysis data showed that among 18 identified prognostic biomarkers three genes including CRTAC1, TGFBR3, and SCGB1D1 are valuable prognostic biomarkers also potential diagnostic markers with highest AUC, sensitivity, and specificity (Fig. 4B). Surprisingly, they have a higher sensitivity and AUC than most of those reported for current diagnostic biomarkers such as(PSA: AUC = 0.58, SE = 0.09–0.33, PCA3: AUC = 0.73, SE = 0.69, PHI: AUC = 0.815, SE = 0.77)(16, 17). Especially, one of the controversions about these PCa biomarkers is lack of specificity as PSA has specificity of 0.17–0.19, PHI has specificity of 0.36–0.42 for diagnosis of PCa(18, 19). However, both of our novel potential biomarkers including CRTAC1 and SCGB1D1 have specificity of 0.9 and 0.71, respectively. Therefore, we selected these genes for further evaluation. We evaluated identified biomarkers in relevant DNA-seq data and visualized the variants in detail.

As indicated in Fig. 5A, CRTAC1 is mutated in 17% of patients and had 5 variants including two missense mutations, two silent mutations and one nonsense mutation. Also, we assessed TGFRB3 in DNA-seq data which showed one missense variant for this gene (Fig. 5B). Although, SCGB1D1 did not represent any variant in the current PCa DNA-seq dataset. Then, a literature review was performed to assess any previous report of identified biomarkers. TGFRB3 was reported by a group of researchers as a potential prognostic biomarker which validate our data(20). Also, we showed perfect sensitivity and specificity of TGFBR3 as a potential diagnostic biomarker. CRTAC1 and SCGB1D1 considered as two novel candidate biomarkers for more validation.

Regard to function of identified genes and related pathways, SCGB1D1 is a member of the lipophilin protein subfamily, also as an ortholog of prostatein, the main secretory glycoprotein of the rat ventral prostate gland. Probably, this gene is transcriptionally regulated by steroid hormones which bind to androgens and other steroids. SCGB1D1 may bind and concentrate estramustine, a chemotherapeutic drug commonly used for prostate cancer. Also, In GO ontology, protein binding molecular function (MF) was identified and extracellular matrix was showed for cellular component (CC). CRTAC1 gene encodes an extracellular matrix protein that is up regulated in mesenchymal stem cells used as a marker to distinguish chondrocytes from mesenchymal stem cells. Possibly, this protein may be involved in cell-cell or cell-matrix interactions.

In GO ontology, molecular function (MF) related to CRTAC1 is calcium ion binding and protein binding. Also, in biological process(BP) term, this gene is involved in axonal fasciculation, also CRTAC1 is a extracellular protein and located in extracellular exosome in cellular component(CC) term(21). These results show the potential value of these biomarkers in prostate cancer prognosis and diagnosis especially, as they are extracellular exosome proteins and may be present in body fluids.

Validation of CRTAC1 and SCGB1D1 in other PCa datasets

We further evaluated the expression of SCGB1D1 and CRTAC1 as novel biomarkers in other PCa datasets using PCaDB database(22). Downregulation of CRTAC1 in primary PCa was confirmed in other datasets including Cambridge, CIT, GSE59745, GSE32269, GSE8218, GSE29079, ERP00077, SRP119917, GSE30521, GSE133626, GSE80609, GSE22260, GSE114740, GSE32571. Bar plots of validated expression of CRTAC1 is shown in Fig. 6. Downregulation of SCGB1D1 in primary PCa was validated in other datasets such as GSE59745, GSE35988-GPL6848, GSE32269, GSE30521. However, In GSE2443 dataset it was shown that the expression of SCGB1D1 is significantly higher in androgen dependent primary PCa in comparison with androgen independent primary PCa which justify the upregulation of SCGB1D1 in some datasets such as GSE79021. Also, upregulation of SCGB1D1 was observed in metastatic PCa in GSE3325.

Deep learning is a subtype of machine learning which has several advantages over the traditional machine learning(23). It is based on artificial neural networks which could automatically process huge amount of complex data to identify the most important features. This method has been applied in different studies for cancer diagnosis, prognosis, and treatment. In the current study, 20 genes were identified as key genes in PCa among 609 DEGs by deep learning approach. The performance of the model was confirmed with the AUC= 1, R²= 0.96, RMSE= 0.054, MSE=0.002, PR-AUC=1, and Accuracy=93.8%. Also, we evaluated genetic variants of 20 identified genes by integration of DNA-seq results and visualized their mutations. Twenty-six genetic variants were identified for our 20 candidate genes in patients (figure 2D). Moreover, these genes were assessed for functional enrichment analysis to find over-represented biological functions or pathways in our gene list. pathway analysis results demonstrated that 4 main pathways including Hippo signalling pathway, Aldosterone-regulated sodium reabsorption, Vibrio cholera infection, and gonadotropin-releasing hormone (GnRH) secretion pathways are significantly enriched. Hippo signalling pathway is a conserved signalling pathway that controls apoptosis, cell proliferation, and stem cell self-renewal. Aldosterone-regulated sodium reabsorption pathway plays role in sodium and potassium metabolism regulation in renal collecting duct cells. Also, GnRH secretion pathway is the key regulator of reproduction in mammals.

Gonadotropin-releasing hormone (GnRH) triggers synthesis and secretion of the gonadotropins including luteinizing hormone and follicle-stimulating hormone. Therefore, identified genes incorporate significantly in PCa progress and pathogenesis. Also, we performed PPI network analysis to assess the interactions between genes and to find hub genes. Cox regression was performed to assess the relation between survival of patients and confounding factors such as age, race, Gleason-score, clinical -M, clinical -T, case-control, and serum PSA-value which results showed no meaningful relation between survival of patients and these factors. Also, Cox regression showed no general relation between DEGs and survival.

To disclose prognostic biomarkers, we first executed Kaplan-Meier analysis for 20 identified genes by deep learning. Also, cox regression was done for all 609 identified genes then Kaplan-Meier analysis performed to not miss any potential prognostic biomarkers. In total 18 genes were significantly (p < 0.05) related with survival of patients with PCa. KCNQ1, M2T2A, SORBS1, MIR204, RNF207, and QRFP were reported in other studies as prognostic biomarkers (24-30). Other genes considered as novel prognostic biomarkers for further evaluation. We determined biological functions and roles of candidate biomarkers in other diseases which indicated their critical role in cancer progression and pathogenesis (Table 2). Combio-ROC analysis was accomplished for key genes to discover diagnostic biomarkers. ROC results showed that CRTAC1 (AUC=0.89, SE=0.78, SP=0.9), SCGB1D1 (AUC=0.85, SE=0.85, SP=0.71), and TGFBR1 (AUC=0.89, SE=0.81, SP=0.84) are also potent diagnostic biomarkers in PCa. Also, combination of CRTAC1 and SCGB1D1 with AUC= 0.91, SE=0.94 and SP=0.76 identified as a diagnostic biomarker. Surprisingly, these genes demonstrated significantly higher sensitivity and AUC in comparison with commercial biomarkers such as PCA3 (AUC=0.73, SE=0.69), PHI (AUC=0.69, SE=0.77), and serum PSA ( AUC=0.58, SE=0.09-0.33) (17, 31, 32).

Moreover, combination of CRTAC1, SCGB1D1, and SCNN1B showed high AUC=0.92, SE=0.9, and SP=0.8 as a potent diagnostic marker which have notably higher AUC, sensitivity, and specificity than 4k score (AUC=0.8, SE=89, SP=0.29). TGFBR3 was reported in another study as potential prognostic biomarker which validate our data which is a potent diagnostic and prognostic biomarker(26). CRTAC1 and SCGB1D1 were considered as novel biomarkers for further evaluation. We further assessed alterations of these genes in DNA-seq and whole exome sequencing dataset. As depicted in figure 5, five variants were observed for CRTAC1 which two of them were missense mutations, two were silent mutations, and the other one was a nonsense mutation. Noteworthy, CRTAC1 mutations was observed in the 17% of patients. However, there were no mutations for SCGB1D1 in our patients. As described in the Table 2, SCGB1D1 gene is an ortholog of prostatein, the main secretory glycoprotein of the rat ventral prostate gland and probably is transcriptionally regulated by steroid hormones and bind to androgens and other steroids. Also, this gene was detected as a prognosis marker in ovarian cancer(33). Moreover, CRTAC1 takes a part in cell-cell or cell-matrix interactions and is a biomarker of osteoarthritis progression (34). We employed PCa data base and validated down regulation of CRTAC1 gene in 14 other PCa data set(22). Then we validated our results using other datasets available on PCaDB(22). Down regulation of SCGB1D1 was validated in 4 other datasets. In one dataset, significantly higher expression of SCGB1D1 was showed in androgen dependent primary PCa compared with androgen independent primary PCa which justify the upregulation of SCGB1D1 in some datasets. Also, in some data sets upregulation of SCGB1D1 was observed in metastatic PCa which show that SCGB1D1 could be used as biomarker in stratification of PCa.

In conclusion, we identified about 11 novel prognostic biomarkers which further ROC analysis showed that TGFRB3, SCGB1D1 and CRTAC1 are potent prognostic biomarkers and have diagnostic ability in PCa. Dysregulation of SCGB1D1 and CRTAC1 as novel prognostic and diagnostic biomarkers were validated in other PCa datasets. Also, we introduced combination of CRTAC1 and SCGB1D1 with AUC= 0.91, SE=0.94 and SP=0.76 as potent biomarkers which have notably higher AUC, SE, and SP than the current biomarkers of prostate cancer such as 4k score. However, more investigations are needed to evaluate these candidate biomarkers for clinical applications as prognostic markers for prostate cancer

Data Availability

The datasets analysed during the current study are available in the TCGA data base (TCGA PRAD cohort ) with the accession number of phs000178 ( http://tcga-data.nci.nih.gov/tcga/).

Fund:

This work was supported by Mashhad University of Medical Sciences (Amir Avan)

Descotes JL. Diagnosis of prostate cancer. Asian J Urol. 2019;6(2):129-36.
Porzycki P, Ciszkowicz E. Modern biomarkers in prostate cancer diagnosis. Cent European J Urol. 2020;73(3):300-6.
Madu CO, Lu Y. Novel diagnostic biomarkers for prostate cancer. J Cancer. 2010;1:150-77.
Silberstein JL, Pal SK, Lewis B, Sartor O. Current clinical challenges in prostate cancer. Transl Androl Urol. 2013;2(3):122-36.PMID: 37561382
Terada N, Akamatsu S, Kobayashi T, Inoue T, Ogawa O, Antonarakis ES. Prognostic and predictive biomarkers in prostate cancer: latest evidence and clinical implications. Ther Adv Med Oncol. 2017;9(8):565-73.
Bulten W, Kartasalo K, Chen P-HC, Ström P, Pinckaers H, Nagpal K, et al. Artificial intelligence for diagnosis and Gleason grading of prostate cancer: the PANDA challenge. Nature Medicine. 2022;28(1):154-63.
Kumar S, Shuaib M, AlAsmari AF, Alqahtani F, Gupta S. GNL3 and PA2G4 as Prognostic Biomarkers in Prostate Cancer. Cancers. 2023;15(10):2723.
Matin F, Jeet V, Moya L, Selth LA, Chambers S, Yeadon T, et al. A Plasma Biomarker Panel of Four MicroRNAs for the Diagnosis of Prostate Cancer. Scientific Reports. 2018;8(1):6653.
Moya L, Meijer J, Schubert S, Matin F, Batra J. Assessment of miR-98-5p, miR-152-3p, miR-326 and miR-4289 Expression as Biomarker for Prostate Cancer Diagnosis. International Journal of Molecular Sciences. 2019;20(5):1154.
Hou Q, Bing ZT, Hu C, Li MY, Yang KH, Mo Z, et al. RankProd Combined with Genetic Algorithm Optimized Artificial Neural Network Establishes a Diagnostic and Prognostic Prediction Model that Revealed C1QTNF3 as a Biomarker for Prostate Cancer. EBioMedicine. 2018;32:234-44.
Khatami F, Aghamir SMK, Salmaninejad A, Shivarani S, Khorrami MH. Biomarkers for Prostate Cancer Diagnosis from Genetic Perspectives. Translational Research in Urology. 2020;2(2):51-8.
Sidey-Gibbons JAM, Sidey-Gibbons CJ. Machine learning in medicine: a practical introduction. BMC Medical Research Methodology. 2019;19(1):64.
Park MK, Lim JM, Jeong J, Jang Y, Lee JW, Lee JC, et al. Deep-Learning Algorithm and Concomitant Biomarker Identification for NSCLC Prediction Using Multi-Omics Data Integration. Biomolecules. 2022;12(12).
MedlinePlus [Internet] Bethesda (MD): National Library of Medicine (US); [updated Jun 24; cited 2020 Jul 1] Available from: https://medlineplusgov/.
COSMIC, the Catalogue of Somatic Mutations in Cancer, is a project of the Wellcome Sanger Institute (WSI) WSI is operated by Genome Research Limited (GRL), a charity registered in England with the number 1021457 and a company registered in England with number 2742969, whose registered office is 215 Euston Road, London, NW1 2BE.
Loeb S, Partin AW. PCA3 Urinary Biomarker for Prostate Cancer. Rev Urol. 2010;12(4):e205-6.
Loeb S, Catalona WJ. The Prostate Health Index: a new test for the detection of prostate cancer. Ther Adv Urol. 2014;6(2):74-7.
Othman H, Abu Yamin AH, Md Isa N, Bahadzor B, Syed Zakaria SZ. Diagnostic performance of prostate health index (PHI) in predicting prostate cancer on prostate biopsy. Malays J Pathol. 2020;42(2):209-14.
de la Calle C, Patil D, Wei JT, Scherr DS, Sokoll L, Chan DW, et al. Multicenter Evaluation of the Prostate Health Index to Detect Aggressive Prostate Cancer in Biopsy Naïve Men. J Urol. 2015;194(1):65-72.
Liang Luo1 WT, Tao He1,liao-yuan li1. Prediction of potential prognostic biomarkers in metastatic prostate cancer based on a competing endogenous RNA regulatory network. 2020.
Stelzer G RR, Plaschkes I, Zimmerman S, Twik M, Fishilevich S, Iny Stein T, Nudel R, Lieder I, Mazor Y, Kaplan S, Dahary, D, Warshawsky D, Guan - Golan Y, Kohn A, Rappaport N, Safran M, and Lancet D. The GeneCards Suite: From Gene Data Mining to Disease Genome Sequence Analyses (PMID: 27322403 ; Citations: 2,262).
Li R, Zhu, J., Zhong, W.-D., and Jia, Z. . PCaDB - a comprehensive and interactive database for transcriptomes from prostate cancer population cohorts (2021)
Kourou K, Exarchos TP, Exarchos KP, Karamouzis MV, Fotiadis DI. Machine learning applications in cancer prognosis and prediction. Computational and Structural Biotechnology Journal. 2015;13:8-17.
Cagle P, Qi Q, Niture S, Kumar D. KCNQ1OT1: An Oncogenic Long Noncoding RNA. Biomolecules. 2021;11(11).
Wang H, Jones J, Turner T, He QP, Hardy S, Grizzle WE, et al. Clinical and biological significance of KISS1 expression in prostate cancer. Am J Pathol. 2012;180(3):1170-8.
Luo L, Zhang L-L, Tao W, Xia T-L, Li L-Y. Prediction of potential prognostic biomarkers in metastatic prostate cancer based on a circular RNA-mediated competing endogenous RNA regulatory network. PLOS ONE. 2021;16(12):e0260983.
Zhang X, Chen Y, Li Z, Han X, Liang Y. TGFBR3 is an independent unfavourable prognostic marker in oesophageal squamous cell cancer and is positively correlated with Ki-67. Int J Exp Pathol. 2020;101(6):223-9.
Kawan MA, Kyrou I, Ramanjaneya M, Williams K, Jeyaneethi J, Randeva HS, et al. Involvement of the glutamine RF‑amide peptide and its cognate receptor GPR103 in prostate cancer. Oncol Rep. 2019;41(2):1140-50.
Todorova K, Metodiev MV, Metodieva G, Mincheff M, Fernández N, Hayrabedyan S. Micro-RNA-204 Participates in TMPRSS2/ERG Regulation and Androgen Receptor Reprogramming in Prostate Cancer. Horm Cancer. 2017;8(1):28-48.
Li Y, Wang H, Pan Y, Wang S, Zhang Z, Zhou H, et al. Identification of bicalutamide resistance-related genes and prognosis prediction in patients with prostate cancer. Front Endocrinol (Lausanne). 2023;14:1125299.
Borque-Fernando Á, Rubio-Briones J, Esteban LM, Dong Y, Calatrava A, Gómez-Ferrer Á, et al. Role of the 4Kscore test as a predictor of reclassification in prostate cancer active surveillance. Prostate Cancer Prostatic Dis. 2019;22(1):84-90.
Zappala SM, Scardino PT, Okrongly D, Linder V, Dong Y. Clinical performance of the 4Kscore Test to predict high-grade prostate cancer at biopsy: A meta-analysis of us and European clinical validation study results. Rev Urol. 2017;19(3):149-55.
Bignotti E, Tassi RA, Calza S, Ravaggi A, Rossi E, Donzelli C, et al. Secretoglobin expression in ovarian carcinoma: lipophilin B gene upregulation as an independent marker of better prognosis. J Transl Med. 2013;11:162.
Szilagyi IA, Vallerga CL, Boer CG, Schiphof D, Ikram MA, Bierma-Zeinstra SMA, et al. Plasma proteomics identifies CRTAC1 as a biomarker for osteoarthritis severity and progression. Rheumatology. 2022;62(3):1286-95.
DisGeNET v6 and v7 [Internet].
Zhang R, Xu J, Shen C, Zhang X, Li S, Lv J, et al. Testis-enriched Asb12 is not required for spermatogenesis and fertility in mice. Transl Androl Urol. 2022;11(2):168-78.
Chen CT, Wang PP, Mo WJ, Zhang YP, Zhou W, Deng TF, et al. Expression profile analysis of prognostic long non-coding RNA in adult acute myeloid leukemia by weighted gene co-expression network analysis (WGCNA). J Cancer. 2019;10(19):4707-18.
Wang H, Jiang X, Cheng Y, Ren H, Hu Y, Zhang Y, et al. MZT2A promotes NSCLC viability and invasion by increasing Akt phosphorylation via the MOZART2 domain. Cancer Sci. 2021;112(6):2210-22.
zhan S, Wang W, Yan M, Li D, Fan X, Zhang Q, et al. Correlation of MZT2A Expression with Immunity and Prognosis in Pan-Cancer Analysis. 2022.
Wang S, Zhang, C., Zhang, Z., Qian, W., Sun, Y., Ji, B., Zhang, Y., Zhu, C., Ji, D., Wang, Q., & Sun, Y. . Transcriptome analysis in primary colorectal cancer tissues from patients with and without liver metastases using next-generation sequencing. Cancer Medicine, 6(8), 1976-1987. 2017.
Bushel PR, Ward J, Burkholder A, Li J, Anchang B. Mitochondrial-nuclear epistasis underlying phenotypic variation in breast cancer pathology. Sci Rep. 2022;12(1):1393.
Stelzer G RR, Plaschkes I, Zimmerman S, Twik M, Fishilevich S, Iny Stein T, Nudel R, Lieder I, Mazor Y, Kaplan S, Dahary, D, Warshawsky D, Guan - Golan Y, Kohn A, Rappaport N, Safran M, and Lancet D. The GeneCards Suite: From Gene Data Mining to Disease Genome Sequence Analyses (PMID: 27322403 ; Citations: 2,262) [Internet].
Zeng M-H, Qiu J-G, Xu Y, Zhang X-H. IDUA, NDST1, SAP30L, CRYBA4, and SI as novel prognostic signatures clear cell renal cell carcinoma. Journal of Cellular Physiology. 2019;234(9):16320-7.
Hu B, Yin G, Sun X. Identification of specific role of SNX family in gastric cancer prognosis evaluation. Sci Rep. 2022;12(1):10231.
Lawal B, Kuo YC, Tang SL, Liu FC, Wu ATH, Lin HY, et al. Transcriptomic-Based Identification of the Immuno-Oncogenic Signature of Cholangiocarcinoma for HLC-018 Multi-Target Therapy Exploration. Cells. 2021;10(11).
Gan Y, Ye, F., He, X.X. . The role of YWHAZ in cancer: A maze of opportunities and challenges. . Journal of Cancer, 11(8), 2252-2264 https://doiorg/107150/jca41316 2020.

No competing interests reported.

SupplementaryPC.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

The potential prognostic value of Cartilage acidic protein 1 and secretoglobin family 1D member 1 in prostate cancer

Status:

Version 1

Abstract

Background

Methods

Results

Conclusion

Figures

Introduction

Materials and Methods

Data sets preparation and processing

DNA-seq and whole exome sequencing data analysis

Normalization of RNA-seq data and differential gene expression analysis

Machine learning for identification of potential biomarkers

Evaluation of model performance using a test set

Functional enrichment analysis of candidate genes

Protein-protein interaction network analysis

Survival analysis

Unravelling potential diagnostic biomarkers

Correlation of clinical/demographic characteristics and genes with PCa

Results

Analysis of DNA-seq

Identification of differentially expressed genes (DEGs) in PCa

Determination of key DEGs in PCa by deep learning

Evaluation of candidate genes in DNA-seq data

Functional enrichment analysis of candidate genes

Identification of prognostic biomarkers

Correlation analysis of clinical/demographic characteristics and genes with PCa

Potential value of CRTAC1 and SCGB1D1 as prognostic biomarkers

Validation of CRTAC1 and SCGB1D1 in other PCa datasets

Discussion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1