RETRACTED: Polygenic risk score and prostate specific antigen predict death from prostate cancer in men with intermediate aggressive cancer

Abstract

Polygenic risk scores (PRS) and the prostate specific antigen (PSA) test have been shown to be successful tools for predicting prostate cancer (PCa) incidence. In this study, we assessed the potential of combining PRS and PSA as biomarkers for PCa aggressiveness and subsequent mortality from patients with low to intermediate risk PCa from the TAPG‐TURP cohort. Targeted sequencing of 140 PCa‐related genes was performed using 162 prostate samples from PCa patients with a Gleason score of 6 or 7, 80 of whom died from the disease. An additional 305 genome samples from healthy participants of the 1000 Genomes Project phase 3 were selected as controls. Two novel PRSs were developed using 21 single nucleotide polymorphisms (SNPs) selected from those differentiated between alive (n = 82) and dead (n = 80) PCa patients. The first PRS was used in decision tree‐based models, such as random forest (rf) able to accurately distinguish cancer from healthy samples (sensitivity = 100%, specificity = 100%, AUC = 1). The second PRS was used together with Gleason score and PSA in an artificial neural network model able to determine the aggressiveness of PCa by predicting PCa mortality with intermediate to high accuracy (sensitivity = 90%, specificity = 68.8%, AUC = 0.718). Further work must be done using our two machine learning classifiers to validate them further and apply them in the clinic, bypassing the necessity of invasive and more expensive approaches. Their application will potentially guide the clinical decision‐making process and reduce costs of the clinical management of PCa patients.

What's New?

Using 21 SNPs, two novel PRS were constructed and used to develop two new machine‐learning classifiers, one for the detection of prostate cancer and the other for the prediction of its aggressiveness and subsequent mortality. The classifier for disease detection is built using the PRS as the sole feature, whereas the one for disease aggressiveness includes Gleason score and PSA as additional features. Their clinical application may potentially guide disease management strategies in the future.

Abbreviations

4K

four‐kallikrein

ANN

artificial neural network

AUC

area under the curve

BP

biological processes

CAPRA

cancer of the prostate risk assessment

CC

cellular components

CCP

cell cycle progression

COSMIC

catalogue of somatic mutations in cancer

FDR

false discovery rate

FFPE

formalin‐fixed paraffin wax‐embedded

GO

gene ontology

GWAS

genome‐wide association study

HR

hazard ratio

INDELS

INsertions or DELetionS

kNN

k‐nearest neighbors

KS

Kolmogorov–Smirnov

MF

molecular functions

NCBI

National Center for Biotechnology Information

NYGC

New York Genome Center

OR

odds ratio

PCa

prostate cancer

PHI

prostate health index

PRS

polygenic risk score

PSA

prostate specific antigen

QDA

quadratic discriminant analysis

QMUL

Queen Mary University of London

ROC

receiver operating characteristic

SNP

single nucleotide polymorphism

TAPG

Trans‐Atlantic Prostate Group

TURP

transurethral resection of prostate

1. INTRODUCTION

Prostate cancer (PCa) is the second most common cancer in men worldwide and the most common among men in the Western world. ¹ However, it is usually a slowly progressing type of cancer where most men die with rather than from the disease itself. ² Currently, the greatest challenge is the ability to distinguish biologically indolent PCa from that which will progress. Several tools have been developed to provide a qualitative measure of the risk for disease progression ³ such as prostate‐specific antigen (PSA) isoforms (four‐kallikrein (4K) panel and prostate health index (PHI) tests), multifactorial risk‐calculator nomograms, urinary analysis, cell cycle progression (CCP) score, Cancer of the Prostate Risk Assessment (CAPRA) score, Gleason score, and age, among others.^{4, 5}

Despite the multitude of tools, there is still a need to increase their predictive accuracy or to develop alternative scores from quantifiable biomarkers that can be used clinically to improve disease management and stratification. For example, the PSA test is one of the most commonly used methods for early treatment; however, its ability to discriminate high‐risk disease is still poor due to lack of specificity.^{3, 6}

Another type of approach has been based on the fact that a significant proportion of PCa is associated with genetic factors. ⁷ Thus, besides the methods mentioned above, genome‐wide association studies (GWASs) have successfully identified single‐nucleotide polymorphisms (SNPs) associated with susceptibility to PCa. ⁸ However, they were not able to explain the PCa phenotype with any singular SNP, which led to the use of polygenic risk scores (PRS). The PRS is a measure of the cumulative contribution of individual SNPs to a disease, basically a weighted sum of the number of risk alleles an individual carries. ⁹ Despite some commercial PRS models for the detection of PCa having been developed,^{10, 11} the development of well‐validated PRS models is yet to be undertaken on a larger scale to be used successfully in the clinical field.

Machine learning (ML) methods have shown successful predictive capabilities in various medical domains. ¹² For this reason, they have been used in PCa research, especially by combining image (MRI), genetic‐based features, and clinical variables to predict clinically significant PCa, which could help to guide the diagnosis and treatment of the disease.^{13, 14, 15, 16, 17, 18} Even though the potential of PSA and PRS is evident, the former's low specificity and the fact that the latter's reliability is yet to be validated with higher certainty, their combination is yet to be explored as a potential biomarker for detection or aggressiveness assessment of PCa, which could be overcome by the use of ML methods. Even with the successful use of ML‐based techniques, the current inability to distinguish biologically indolent from aggressive PCa (i.e., that which will progress and potentially lead to mortality) is still a concern as it may lead to overdiagnosis and overtreatment. Currently, the most common way to predict aggressiveness and guide clinical management strategies is by using the Gleason grading system on prostate tissue samples. This system can provide a score technically ranging from 2 to 10, but with scores 6–10 only being reported in modern clinical practice and being associated with the severity of the cancer. In this scoring system, samples with a Gleason score ≥8 are considered aggressive, whereas scores 7 to 6 are considered intermediate and low risk, respectively, leading to an uncertain prognosis.

In the present study we aimed to explore the potential of ML methods to prognose the PCa aggressiveness from such uncertain cases using the Trans‐Atlantic Prostate Group (TAPG) transurethral resection of prostate (TURP) cohort. ² The TAPG study is the long‐term follow‐up that started in 1999 and comprises more than 3300 men diagnosed with PCa between 1990 and 2006 and includes follow‐up information such as mortality. The newly proposed classifier is trained using patients in the intermediate state with a Gleason score of 6 or 7 and thus will allow a more robust classification to segregate men of low risk who can be safely followed by active surveillance (indolent) from those who require active intervention (aggressive).

2. MATERIALS AND METHODS

2.1. Study subjects

Our study was conducted on prostate patients' samples (n = 2335) with low to high risk PCa based on Gleason scores from 4 to 10, from the TAPG‐TURP cohort. Samples were split into DEAD (n = 565) and ALIVE (n = 1770) groups, where DEAD were defined as patients who died from PCa and ALIVE were patients who progressed slowly and were alive at censoring (the end of the follow‐up) or died of other causes. Follow‐up of ALIVE was longer than follow‐up of DEAD. ALIVE and DEAD were matched according to their age, with a maximum of 4 years difference. Only samples from patients with PSA levels, Gleason score, CAPRA score, and matched ALIVE and DEAD were considered in our study (Supplementary Table S1).

2.2. Targeted sequencing panel design

To identify biomarkers associated with death from PCa, a customized sequencing panel for SNPs and other mutational biomarkers was developed. A literature search was performed in 2017, using PubMed and Web of Science, focusing on studies that reported genes associated with PCa prognosis, progression, or mortality. Furthermore, susceptibility SNPs with odds ratios (OR) equal to or greater than 1 were also downloaded from the GWAS catalog. ¹⁹ Both approaches resulted in 79 SNPs (GWAS catalog, Supplementary Table S2) and 140 genes that were used for targeted deep sequencing (Supplementary Table S3). The 140 genes included 93 genes from the literature review, six genes that were used to develop a methylation score that could predict death from PCa in men with low to intermediate clinical risk, ²⁰ 19 genes selected from a pilot study conducted in the Wolfson Institute of Population Health at Queen Mary University of London (QMUL), and 22 from a meta‐analysis of published data on genes from existing models for detecting PCa. Details of the selection of the genes and SNPs for the sequencing panel are described in the section TAPG pilot study in Supplementary Material and in Supplementary Table S3. A total of 195 samples were randomly selected from the 1735 samples of the TAPG‐TURP cohort with a Gleason score of 6 or 7, and used for DNA extraction, library preparation, and sequencing using the targeted sequencing panel.

2.3. DNA extraction and library preparation

DNA was extracted from formalin‐fixed paraffin wax‐embedded (FFPE) transurethral prostate resection (TURP) tissues. Hematoxylin and eosin‐stained sections previously annotated by an expert pathologist for cancerous and normal areas were used as a guide for macrodissection. Cancerous areas from six 5 μm FFPE sections were dissected for each patient, and DNA was extracted using the QIAamp DNA FFPE tissue kit (Qiagen, Germany) following the manufacturer's protocol. The DNA quality was assessed with LabChip® GX Touch™ Nucleic Acid Analyzer using the Genomic DNA Reagent Kit (PerkinElmer, UK). Prior to the library preparations, 40 ng of DNA was treated with NEBNext FFPE DNA Repair Mix (New England Biolabs, UK) as an optimization step to repair FFPE DNA in the next generation sequencing workflow. Repaired DNA was then quantified with the Qubit fluorometer using the DNA High Sensitivity Kit (Thermo Fisher Scientific, UK).

For library preparation, 10 ng of repaired genomic DNA per primer pool was amplified in 2× primer pools multiplex PCR reaction using the Ion Ampliseq Library Kit 2.0‐96LV (Thermo Fisher Scientific, UK) following the manufacturer's protocol. The resulting amplicons were treated with FuPa Reagent to remove excess primers and ligate to the sequencing Ion P1 Adapters and Ion Xpress barcodes. After AMPure beads purification, barcoded libraries were quantified using the Ion Library TaqMan Quantification kit (Thermo Fisher Scientific, UK) and QuantStudioTM 5 real‐time PCR system. Library concentrations were normalized to 50 pM with TE buffer, and an enrichment reaction was performed with the Ion PI‐Hi‐Q Chef kit (Thermo Fisher Scientific, UK) on the Ion Chef system. In a final step, libraries were loaded on the Ion P1 chip and sequenced on the Ion Proton platform using the Ion PI Hi‐Q Sequencing 200 kit (Thermo Fisher Scientific, UK).

2.4. Quality control

Initial quality control on raw sequence reads of the TAPG TURP samples (bam file format) was conducted using FastQC v0.11.9. ²¹ Due to poor quality (all reads with Phred quality score <20), 90 samples were re‐sequenced and subsequently merged to their previous raw run with SAMtools merge v1.8. ²² A new quality check was applied to verify the highest sequence content and quality (Phred quality score >20). A total of 162 samples (82 ALIVE and 80 DEAD) passed the quality control and were used for downstream analysis. The sequencing coverage and quality statistics for each sample are summarized in Supplementary Table S4.

2.5. Variant calling and imputation

The high‐quality reads of each sample were aligned to the human reference genome build GRCh38.p13 (NCBI Annotation Release 109) using the BWA‐backtrack algorithm of the BWA software 0.7.17. ²³ The resulting alignments were sorted, and the duplicates were removed by SAMtools. SAMtools mpileup and BCFtools call v1.8 ²⁴ were used to call variants across all individuals. BCFtools filter was used on the merged vcf file to retain variant sites with SNPs and insertions or deletions (INDELS) that passed the following filter parameters: (i) variants within three base pairs of an INDEL and clusters of INDELs separated by 10 base pairs allowing only one to pass; (ii) reads with quality above 20 (QUAL >20), with coverage higher than 10 (DP >10) and lower than 100 (DP< 100), high genotype quality, and allele counts above 3 (AC >3). VCFtools v0.1.16 ²⁵ was additionally used to remove sites with minor allele frequency <0.01 and minor allele counts <1 as well as variants with missing genotype rate >50%.

Genotype imputation was then performed using the New York Genome Center release as the reference panel. ²⁶ Eagle v2.4.1 ²⁷ was used for phasing and Beagle 5.1 (version 29May21.d6d) ²⁸ for imputation. The SNPeff and SnpSift v5.0 ²⁹ were used to characterize the effects of the individual variants on the genes. PLINK v1.9 ³⁰ was used to perform logistic regression between aggressive PCa and every variant and to calculate the OR.

2.6. Gene ontology (GO) term enrichment and pathway analysis

GO terms were assigned to each protein‐coding gene where variants were identified. GO annotations for the human reference genome were obtained from BioMart/Ensembl (Genome build GRCh38.p13, genomes release 101). The genes with identifier variants and assigned GO terms were sorted by raw p‐value due to edge effects associated with adjusted p‐values; then a rank‐based test was performed for each GO term. GO over‐representation analysis was performed using topGO v2.42.0 ³¹ with the Kolmogorov–Smirnov (KS) test and the “weight01” algorithm, while focusing on GO terms for which at least 20 genes were annotated.

A pathway over‐representation analysis was performed with Reactome browser version 3.7 (database release 74). ³² The NCBI gene identifiers for all genes on which variants were found were submitted to the Reactome web browser and tested for over‐representation of gene pathways (Benjamini–Hochberg adjusted p‐value [FDR] <0.05). ³³

2.7. Polygenic risk score (PRS) and CAPRA score

For the PRS derivation, summary statistics of patient samples from a recent GWAS study on more than 140,000 men of European ancestry for PCa ⁸ were used as reference data. Only variants with an imputation score >0.7 were used. A linkage disequilibrium reference panel of European samples ³⁴ was also used as a reference for the calculation. Duplicate variants and those with ambiguous strands (A/T or C/G) and very low or very high heterozygosity (highly correlated variants), as well as low Hardy–Weinberg equilibrium (p‐value <10⁻⁶) were removed. For the score derivation, European men from the 1000 Genomes Project phase 3 release of variant calls ³⁵ were randomly selected and used as healthy controls, as they have no history of cancer. The same quality check and filtering applied for the summary statistics from cancer samples were also performed on the healthy samples. A total of 305 samples passed the quality check and the filtering. As an additional step, regions not sequenced in the targeted sequencing panel of the cancer samples were removed from the healthy samples.

PRSs were calculated in two comparisons: between healthy and cancer samples, and between ALIVE and DEAD within cancer samples from the TAPG cohort. PRS were calculated using the penalized regression method, called lassosum v0.4.5, ³⁶ adjusting for the effect of population stratification for the comparison between healthy and cancer, and for age and the effect of population stratification for the comparison between ALIVE and DEAD. The best PRS in both comparisons was chosen based on the best lambda and the best‐fit PRS. The correlation between these newly derived PRS and disease events was tested with the point‐biserial correlation test.

For the CAPRA score calculation, missing T‐stage values were predicted by fitting a multivariate logistic regression with Gleason, prostate‐specific antigen (PSA), log(1 + Extent), and age as predictors according to Ahmad et al. ²⁰ Once T‐stage information was obtained for all patient samples, the CAPRA score was computed.

2.8. Cross validation and deep learning model

To evaluate the capability of PRS to discriminate between cancer and healthy samples, we performed k‐fold cross validation (k = 10) using a decision tree‐based ML method with the recursive partitioning (rpart) statistical method. The data set of 467 samples (305 healthy and 162 cancer) were split into train (70%) and test (30%) and the error of predicting the test was evaluated. With the best trained model, we predicted healthy and cancer samples in the test data set and determined the accuracy, sensitivity, and specificity of the model. The model with the highest accuracy and kappa value was selected, and its predictive ability was evaluated using the receiver operating characteristic (ROC) curve. The area under the curve (AUC) and Youden index were also calculated from the best model.

A deep‐learning algorithm based on artificial neural networks (ANN) was used to distinguish indolent from aggressive forms of PCa. The resilient backpropagation algorithm with weight backtracking was selected to calculate the neural network in the R package neuralnet v1.44.2. The data set of 162 samples was split into training (70%, n = 113) and test data (30%, n = 49). The run was performed with 1 to 3 hidden layers and 5 repetitions for the neural network on the training data. The measured performance metrics of the model were accuracy, sensitivity, specificity, Kappa, and AUC. The best model was chosen based on the highest sensitivity, and in case of equal sensitivities between models, specificity was used. The best results in terms of performance were achieved with one and two hidden layers. Prediction on the test data was performed with the repetition that presented the lowest error rate. The model was obtained with input layers represented by a combination of the derived PRS between ALIVE and DEAD groups, the PSA levels of the patients, the Gleason and CAPRA scores, and the age.

Both approaches (k‐fold cross validation and ANN) were also performed on the whole data set of 2335 patients to evaluate if PSA levels, CAPRA and Gleason scores, and age are capable of distinguishing indolent from aggressive PCa. For each model, the ROC curve was constructed, and the respective AUC was obtained. The Youden index was calculated for the different PSA levels that corresponded to the points of the ROC curve.

2.9. Survival analysis

Cox proportional hazards models were used to analyze the effects of the variables (PRS, CAPRA score, Gleason score, and PSA) on death from PCa. PCa survival probabilities were predicted from fitted univariate Cox proportional hazards models. Schoenfeld residuals were examined to determine if Cox modeling was an appropriate analytical approach. Kaplan–Meier method was used to verify differences in survival when grouping the patients as follows: PRS (group 1 = mean PRS of ALIVE and group 2 = mean PRS of DEAD); CAPRA score (group 0 = low risk, group 1 = intermediate risk, and group 2 = high risk); Gleason score (group 1 to group 10 according to the Gleason score values); PSA (group 1 = mean PSA of ALIVE and group 2 = mean PSA of DEAD). A p‐value <0.05 was considered significant for all applied statistical tests. No adjustment for multiple comparisons was made. All steps described in each section and all analyses were performed on QMUL's Apocrita High Performance Computing Cluster. ³⁷ Statistical analyses were performed in R version 4.1.0.

3. RESULTS

3.1. Genes with variants significantly associated with the aggressiveness of PCa

A total of 7,303,437 variants were called across 162 individuals having PCa with a Gleason score of 6 or 7. After QC, 524 SNPs and INDELs remained and were used for phasing and imputation. After imputation, a total of 14,115,957 variants were obtained across chromosomes: 1, 2, 3, 5, 6, 7, 10, 12, 13, 17, and 19. Only variants with an imputation quality score >0.7 were kept. The individual post‐imputed chromosomes were concatenated into one vcf file and then merged with the called variants in non‐imputed chromosomes using BCFtools concat. After QC, 2,685 variants (minor allele frequency − MAF >0.01 and missing genotype rate <0.5) remained, affecting 269 genes. Principal component analysis on genotypes showed no clustering due to differences in condition (DEAD and ALIVE, Supplementary Figure S1). Out of the 2685 variants, 6 (0.22%) were private to ALIVE, and 41 were fixed differences between DEAD and ALIVE. Among all variants, 24.18% affected coding sequences, with 34.79% changing the amino acid sequence (34.79% non‐synonymous vs. 65.21% synonymous changes). The variants in non‐coding sequences were identified in intergenic regions (0.007%), transcripts (73.734%), exons (0.145%), splice sites (0.014%), introns (1.245%), and upstream and downstream genes (0.671%) (Supplementary Figure S2).

After testing for association with aggressive PCa using logistic regression, adjusting for age, 106 variants across 10 genes (GPR19, CLCNKB, FAM131C, SCN11A, LOC399886, RNF123, GMPPB, IP6K1, CDHR4, UBA7) showed significant results (Supplementary Table S6, Figure 1). Their OR ranged from 0.28 to 3.478, and 27 of them presented values above 3. To determine which of the identified variants are in the ClinVar database ³⁸ and the Catalogue of Somatic Mutations In Cancer (COSMIC), ³⁹ two vcf files containing all known variants were downloaded from both databases. BCFtools isec was used to retrieve the identified variants that overlapped with ClinVar and COSMIC (Supplementary Figure S3). Of the 106 variants, 7 were found in the ClinVar database, while 10 were found in COSMIC.

FIGURE 1.

Manhattan plot of SNPs and INDELs in DEAD and ALIVE samples of the TAPG TURP cohort. Each point corresponds to a variant location throughout the genome against the –log10 of the unadjusted p‐values from logistic regression analysis for association with more aggressive PCa and adjusted for age. Colors gray and black differentiate the chromosomes for a better visualization. The horizontal line indicates significance level (−log10[0.05] ≈ 1.3).

3.2. Over‐representation of regulation‐related processes and immune‐related pathways

For the genes where variants were found, a total of 131 GO terms were significantly enriched for various biological processes (BP), molecular functions (MF), and cellular components (CC) according to the KS test (Supplementary Table S7). The majority of GO terms (45 out of the 131) were enriched for regulation‐related processes. The analysis of over‐representation in gene pathways showed 116 pathways with significant results after adjusting for multiple testing (FDR). Of the 116 significant pathways, 23 were associated with DNA damage or repair (Supplementary Table S8). The pathways with the highest number of occurring SNPs are involved in the cellular response to stimuli and stress (Supplementary Figure S4A, Supplementary Table S9). The same pathway analysis was performed only for the SNPs selected to calculate PRS (Supplementary Figure S4B, Supplementary Table S10).

3.3. PRS successfully predicts PCa

The PRS between healthy and cancer was derived from PCa samples (n = 162) and healthy samples (n = 305). After QC, the healthy samples were merged to the post‐imputed vcf file containing the cancer samples with BCFtools merge. PRS was initially computed using all the 106 SNPs significantly associated with aggressive PCa. However, when restricting the derivation to only those found in the ClinVar database (n = 7) and those present in DEAD but not in the ALIVE condition (n = 14), we obtained the best results in terms of OR and phenotypic variation explained. Therefore, out of the 106 variants, 21 were used to compute the PRS in both comparisons, healthy versus cancer and ALIVE versus DEAD (Table 1). The PRS OR (shrinkage parameter = 0.2, lambda = 0.004) ranged from 0.998 to 1.004. The phenotypic variation explained by the best‐fit PRS presented a high value (R ² = 0.695 [69.5%]).

TABLE 1.

SNPs used to calculate Polygenic risk scores for healthy versus cancer samples and ALIVE versus DEAD conditions. All SNPs were significantly associated with disease.

Chr	Position	SNP IDs	rs IDs	REF	ALT	p‐value	Genes	Odds ratio
chr12	12717761	chr12:12717761:T:C	rs34330	T	C	0.0464	GPR19	1.867
chr1	16054320	chr1:16054320:G:T	rs9442228	G	T	0.03514	CLCNKB	0.523
chr1	16056385	chr1:16056385:T:A	rs7512547	T	A	0.03592	CLCNKB	0.525
chr1	16056772	chr1:16056772:A:G	rs12728542	A	G	0.03592	CLCNKB	0.525
chr1	16056853	chr1:16056853:C:T	rs61769892	C	T	0.03592	CLCNKB	0.525
chr1	16057020	chr1:16057020:C:T	rs12402924	C	T	0.03592	CLCNKB	0.525
chr1	16057313	chr1:16057313:G:C	rs28690440	G	C	0.03628	FAM131C	0.526
chr1	16061134	chr1:16061134:A:G	rs9442198	A	G	0.03592	FAM131C	0.525
chr1	16061979	chr1:16061979:G:A	rs1889791	G	A	0.03514	FAM131C	0.523
chr1	16062039	chr1:16062039:C:T	rs2863454	C	T	0.03514	FAM131C	0.523
chr3	38917374	chr3:38917374:T:C	rs76952369	T	C	0.04353	SCN11A	0.504
chr3	38929877	chr3:38929877:C:T	rs11918480	C	T	0.04353	SCN11A	0.504
chr11	45376804	chr11:45376804:G:T	rs7930779	G	T	0.04156	LINC02687	2.453
chr3	49724138	chr3:49724138:G:A	rs4855850	G	A	0.02326	GMPPB	0.479
chr3	49724172	chr3:49724172:C:T	rs3811696	C	T	0.02326	GMPPB	0.479
chr3	49724180	chr3:49724180:C:T	rs3811695	C	T	0.02326	IP6K1	0.479
chr18	62895065	chr18:62895065:A:T	rs624821	A	T	0.01495	PHLPP1	0.409
chr16	72957816	chr16:72957816:A:G	rs4788682	A	G	0.004554	ZFHX3	0.28
chr8	104451462	chr8:104451462:A:G	rs2669429	A	G	0.008849	DPYS	2.986
chr9	134731755	chr9:134731755:C:G	rs3124310	C	G	0.009197	COL5A1	2.472
chr4	142260595	chr4:142260595:T:A	rs1907111	T	A	0.03628	INPP4B	3.478

Note: Current inability to distinguish biologically indolent from aggressive cases remains a key challenge in prostate cancer. Using 21 single nucleotide polymorphisms, the authors constructed two novel polygenic risk scores and machine‐learning classifiers for the detection of prostate cancer and the prediction of aggressiveness and mortality. The disease detection classifier was built using the polygenic risk score as the sole feature, whereas the disease aggressiveness classifier also considered Gleason score and prostate specific antigen level. The classifiers successfully distinguished cancer from healthy samples and predicted cancer mortality. Their clinical application may potentially guide disease management strategies in the future.

The results of the best trained model with k‐fold cross‐validation on the testing dataset showed: accuracy = 100%, kappa = 1, sensitivity = 100% and specificity = 100%; after validation on the test dataset. The decision tree (rpart and gbm) and random forest (rf) methods with PRS as predictor presented the best results over other models such as KNN, linear and support vector machines (Supplementary Table S5—tab “Healthy vs. Cancer”). The estimated AUC (1: 1–1) using the best model on the test dataset of healthy and cancer samples showed a high predictive performance for PCa onset (Supplementary Figure S5). Likewise, the logistic regression model applied to the test dataset showed that the PRS successfully predicts PCa (Pseudo R2 = 0.99, p = 0, AIC = 5.25, AUC = 1, Supplementary Figure S6). There was a significant positive association between PRS and PCa (Point‐biserial correlation = 0.738, p <.0001). For the Youden index analysis, the optimal cut‐off PRS value providing the best trade‐off between sensitivity and specificity for the identification of PCa was −0.0002 (sensitivity = 1, specificity = 1, PPV = 1, NPV = 1).

3.4. Polygenic risk score, PSA, and Gleason score discriminate between indolent and aggressive PCa

All DNA samples analyzed to compute this PRS came from PCa samples of the TAPG cohort (n = 162), which were divided into DEAD (n = 80) and ALIVE (n = 82), depending on whether they died or not from PCa, respectively. The PRS was calculated using the same SNPs used in the calculation between healthy and cancer. The best PRS OR (shrinkage parameter = 1, lambda = 0.007) between ALIVE and DEAD ranged from 0.99 to 1.00. The best‐fit PRS that explains the phenotypic variation was very low (R ² = 0.03 [3%]).

The ANN used to analyze the train data set showed: AUC = 0.718, accuracy = 53.57%, kappa = 0.404, sensitivity = 90% and specificity = 68.75%, after validation on the test data set. The best model was obtained using PRS + PSA + Gleason score as predictors (Supplementary Table S5). There were no improvements on the results when adding CAPRA score to the model, and a model where CAPRA was the only predictor showed lower performance (Supplementary Table S5). A model with either PRS, PSA, or Gleason score as the sole predictor showed lower performances as well (Supplementary Table S5). The logistic regression model applied to the test data set of PCa samples showed that the PRS, PSA, and Gleason moderately predict death from PCa (Pseudo R ² = 0.14, p = 0.026, AIC = 63.95, multi‐class AUC = 0.75, Supplementary Figure S7). For the Youden index analysis, the optimal cut‐off providing the best trade‐off between sensitivity and specificity for the identification of aggressive PCa for PRS was −3.202011e‐06 (sensitivity = 0.62, specificity = 0.58, PPV = 0.59, NPV = 0.61), for PSA was 7 (sensitivity = 0.76, specificity = 0.61, PPV = 0.65, NPV = 0.72).

3.5. PRS predicts PCa related death

PCa survival probabilities for low to intermediate risk cancer samples were predicted from fitted multivariate Cox proportional hazards model, built with PRS, PSA, Gleason, and CAPRA scores as predictors. By evaluating the importance of the features to the model using analysis of variance (ANOVA), we observed that only PRS and PSA were crucial for prediction (Supplementary Table S5). Therefore, we decided to predict survival using a model with PRS + PSA as predictors in a multivariate Cox model. Schoenfeld residuals demonstrated no violation of the assumption of proportional hazards for the features. PSA was significantly associated with a higher risk of PCa‐related death (HR = 1.016 [1.008–1.024], p <.001). PRS showed complete separation, resulting in an infinite hazard ratio (HR = Inf, p = .0239), suggesting a strong predictive effect. Penalized and bias‐reduced models were explored but did not affect the results. Figure 2 clearly shows the strong predictability of the multivariate Cox model; the higher the PRS + PSA, the lower the estimated survival probability. Kaplan–Meier curves were also generated to show unadjusted survival differences between ALIVE and DEAD groups for mean PRS, as it was the strongest predictor. The curves corroborate the results from the Cox model, showing that PRS presented subset differences in survival, with a decrease in survival as the time with PCa increases (Figure 3).

FIGURE 2.

Estimated survival probability from a multivariate Cox model with polygenic risk Score and PSA as predictors. The higher the PRS + PSA, the lower the survival.

FIGURE 3.

Estimated survival probability from Kaplan–Meier method with polygenic risk score used as a predictor. The mean PRS in the DEAD group shows a decrease in survival as the time with prostate cancer increases.

3.6. PSA and Gleason score predict PCa aggressiveness with low specificity

Among all 2335 PCa samples analyzed, 382 (331 ALIVE and 51 DEAD) had PSA levels lower than 4 ng/mL, and 1931 (1420 ALIVE and 511 DEAD) presented elevated levels, between 4 ng/mL and 98.1 ng/mL (Supplementary Figure S8A). In general, PSA levels were significantly higher in DEAD than in ALIVE (Mann–Whitney U test, W = 476.23, p <.0001). There was a significant association between PSA levels and PCa (Point‐biserial correlation = 0.226, p <.0001). In terms of CAPRA score, 603 (553 ALIVE and 50 DEAD) showed low risk values, 735 (602 ALIVE and 133 DEAD) intermediate risk and 997 (615 ALIVE and 382 DEAD) high risk (Supplementary Figure S8B). The DEAD group presented significantly higher CAPRA values than the ALIVE group (Mann–Whitney U test, W = 223.51, p <0.0001). There was also a significant association between CAPRA score and more aggressive PCa (Point‐biserial correlation = 0.304, p <0.0001). The majority of samples, 1735 (1423 ALIVE and 312 DEAD) out of 2335, presented a Gleason score of 6 and 7 (Supplementary Figure S8C), while DEAD showed significantly higher Gleason scores than ALIVE (Mann–Whitney U test, W = 247.29, p <.0001). The association between Gleason score and aggressive PCa was significant (Point‐biserial correlation = 0.32, p <.0001).

The Gradient Boosting Machine (gbm) (accuracy = 71.1%, kappa = 0.216, sensitivity = 79.7%, specificity = 42.6%), the Recursive Partitioning and Regression Trees (rpart) (accuracy = 70.7%, kappa = 0.216, sensitivity = 78.8%, specificity = 43.8%) and the Support Vector Machine (svmRadial) (accuracy = 70.9%, kappa = 0.215, sensitivity = 79.2%, specificity = 43.2%) models showed the best performances after validation on the test data set using PSA + Gleason as predictors (Supplementary Table S5). The best model was rpart based on AUC (0.613). Models with either PSA or Gleason score as a single predictor and the model with PSA and age as predictors showed lower performances (Supplementary Table S5). For the Youden index analysis, the optimal cut‐off PSA value providing the best trade‐off between sensitivity and specificity for the identification of higher risk PCa was 17.2 (sensitivity = 0.64, specificity = 0.62, PPV = 0.35, NPV = 0.84), and for the Gleason score, it was 7 (sensitivity = 0.80, specificity = 0.49, PPV = 0.33, NPV = 0.89).

3.7. PSA predicts PCa related death

Schoenfeld residuals demonstrated violation of the assumption of proportional hazards for CAPRA (chisq = 5.03, df = 1, p = .025) and Gleason scores (chisq = 25.9, df = 1, p <.0001). PSA showed to be a good predictor of PCa related death as the former is negatively correlated with the estimated survival probability (hazard ratio [HR] 1.02 (1 − 1), Wald test 187, p <.0001) (Figure 4). The Kaplan–Meier curve confirms such findings, showing that PSA presented subset differences in survival, with a decrease in survival as the time with cancer increases (Figure 5).

FIGURE 4.

Estimated survival probability from a univariate Cox model with PSA as the predictor. The higher the PSA level, the lower the survival.

FIGURE 5.

Estimated survival probability from Kaplan–Meier method with PSA level used as a predictor. The mean PSA level in the DEAD group shows a decrease in survival as the time with cancer increases.

4. DISCUSSION

Most of the SNPs with significant association with aggressive PCa were found in non‐coding regions, with 25% of them presenting an OR above 3% and 6.7% being found in the ClinVar database. However, no variant was directly associated with PCa. Some of the genes where the SNPs with significant association were identified are known to be indirectly involved in cancer metastasis, ⁴⁰ inhibition of PCa cell proliferation and metastasis ⁴¹ and regulation of cancer cell invasion, ⁴² among other biological processes. Nevertheless, while each SNP individually may only confer a modest increase in a patient's risk of being diagnosed with PCa, the combined effect can be truly substantial. ⁴³ For this reason, we built a PRS from studying a cohort of active surveillance patients, as Gleason 3 + 3 and 3 + 4 is an imperfect interface between active surveillance and immediate treatment, and as such, scores have been largely reported in studies of cancer prediction. ⁴⁴

Our results demonstrate that our PRS obtained from 21 SNPs significantly associated with aggressive PCa can successfully predict cancer, with strong power and high accuracy. Despite showing lower ORs in comparison to previous studies,^{45, 46, 47} the phenotypic variation explained by the PRS is relatively high (≈70%), which was enough to perfectly discriminate between cancer and non‐cancer samples. Our findings are in line with PRSs from studies based on a higher number of known risk SNPs.^{44, 48} Despite the PRS being used for clinical purposes since the 2000s, only a few are available commercially, either alone or in combination with other genetic and clinical information to improve the detection of PCa.^{48, 49} Such commercially available PRS were obtained with a higher number of SNPs, larger sample size for derivation and validation, and reported lower AUC (ranging from 0.64 to 0.86) than our study.

While hundreds of common genetic variants have been associated with the risk of developing PCa, leading to numerous PRS with moderate to high discriminatory power, ⁴³ the distinction between patients with indolent PCa and those who will progress is still a challenge faced by clinicians. This was the rationale behind our second application of PRS, which aimed to predict PCa mortality in patients with an intermediate Gleason score (6 or 7). Our results revealed that men who died from PCa had increased PRS compared to those who were ALIVE at censoring or died from other causes, and that these PRS are negatively associated with survival. Furthermore, PRS in combination with PSA levels and Gleason score successfully predicted PCa mortality with moderate specificity and high sensitivity.

In comparison to similar studies,^{11, 44, 50, 51, 52, 53} we found higher AUC with higher specificity, which is promising given the fact that low specificity is the major issue with the current PSA screening test. This indicates that our PRS has sufficient discriminatory power to differentiate between indolent and aggressive PCa in comparison to previous studies. Our study also demonstrates the improvement in PRS specificity in comparison to previous studies when combined with PSA and Gleason, which increases the discriminatory power.

Despite several reported PRS in PCa prediction studies and some being already commercially available, the PSA test is still the gold standard in PCa screening. Based on this test, decisions are taken on whether to proceed with further tests to confirm cancer and evaluate its aggressiveness. ⁵⁴ However, diagnosis and treatment based on PSA alone can lead to overdiagnosis and overtreatment of indolent cancer. Thus, there is a need for the development of a biomarker capable of minimizing overdiagnosis and taking into account tumor aggressiveness and therapy invasiveness to determine the best treatment for each patient. Currently, many tests have been proposed, mainly focusing on a combination of biomarkers and clinical information, so clinicians do not depend on a single biomarker that can be potentially as problematic as PSA. ⁵⁵ Nevertheless, no specific biomarker that can accurately distinguish indolent from aggressive PCa is available. ⁵⁶

Our findings demonstrate that PSA and Gleason score can successfully predict PCa mortality using an rpart model with an AUC of 0.613 and higher accuracy and higher sensitivity in comparison to current literature. Additionally, PSA levels are inversely proportional to survival in patients with aggressive cancer. Moreover, our ANN classifier using PRS, PSA, and Gleason score had an even higher AUC of 0.718.

The potential of ANN models for the prognosis of PCa has been largely explored as prediction tools that can improve the clinical decision‐making process for the treatment of PCa patients. ⁵⁷ Previous studies have shown that specificity and sensitivity range from 20% to 40% and 70% to 90%, respectively, and AUCs between 0.55 and 0.70 for the ability of PSA to identify patients with cancer. The performance of our classifiers demonstrated that the PRS can improve prediction models, making it worthwhile to explore the potential of combining PRS and PSA for developing prediction tools from large datasets. This can lead to novel deep learning‐based classifiers with high discriminatory power avoiding invasive and more expensive approaches, thus optimizing the clinical decision‐making process for PCa patients both in terms of speed and cost.

Despite their promising results, a potential limitation for our PRS is that it was derived from a small sample size; it was restricted to the European population, and the classifiers were tested on a small cohort. Additional tests on a more diverse population and larger cohorts are needed to confirm their broader utility and performance. Additionally, direct comparison with standard diagnostic methods is required to clinically verify the efficacy of the classifiers in relation to standard PCa screening tests. The current diagnostic methods for PCa in the United Kingdom involve either a high PSA level (>2.5 μg for patients under 50 years old, >3.5 μg for patients over 50 and >4.5 over 60 years old) or an abnormal digital rectal examination followed by a referral for MRI. ⁵⁸ Based on the results, biopsy may be necessary. Therefore, a comparison of how many patients could be more quickly and accurately diagnosed using our classifier, in comparison to the current diagnostic methods, would assess performance and the practical clinical application focusing on avoiding overdiagnosis. Since our genetic‐based classifiers were built from PRS, which are genome sequence‐based, any biological sample would be enough for the diagnosis, as the only requirement is a DNA sample, alongside the clinical information on PSA level and Gleason score. It would be similar to what has been done in a recent study that derived PRS using 130 variants from DNA extracted from saliva. ⁵⁹

This study's results exhibit the predictive potential of ML models that combine genetic‐based (PRS) and clinical‐based (PSA) features for both PCa presence and aggressiveness prediction. To further optimize the parameters of these models and incorporate them as modules of novel predictor algorithms, it is imperative to examine larger cohorts of both healthy and PCa participants from diverse populations. Furthermore, the manual curation of the genetic and/or clinical features in these models may provide additional insight on the biological mechanisms behind PCa and its progression. Planning is currently underway to verify with a larger cohort, and initial discussions with clinicians have already been made for future collaborations in grant applications that will allow us to do so. Once validated in a larger cohort, we expect this test to be a very useful tool for increased risk stratifications in patients with PCa, as PSA is not sufficient to discriminate those with higher risk for progression. After such a validation and the assessment of the risk–benefit trade‐offs of the classifier versus the current diagnostic methods, the proposed classifiers could be used as a protocol for clinical trials to help doctors more precisely identify men who would benefit from more aggressive treatments, favoring the screening of patients at highest genetic risk, thus reducing overdiagnosis and overtreatment.

AUTHOR CONTRIBUTIONS

Leandro Rodrigues Santiago: Writing – original draft; writing – review and editing; formal analysis; data curation. Efthymios Ladoukakis: Writing – review and editing; formal analysis; writing – original draft; data curation. Dorota Scibior‐Bentkowska: Methodology; writing – review and editing. Belinda Nedjai: Resources; supervision; conceptualization; investigation; funding acquisition; writing – original draft; writing – review and editing; project administration.

CONFLICT OF INTEREST STATEMENT

The authors have no conflict of interest to declare.

ETHICS STATEMENT

For the TURP specimens, UK national ethical approvals were obtained from the Northern Multicenter Research Ethics Committee (ethics number: MREC/99/3/40) followed by local ethics committee approvals from each of the collaborating hospital trusts. ²

Supporting information

Data S1: Supporting Information

Table S1: Gleason score and age for the 195 (DEAD and ALIVE) samples of the TAPG TURP cohort used in the present study.

Table S2: List of prostate cancer SNPs (OR >= 1) downloaded from GWAS catalog (Hg19 genome version) and included in the customized targeted sequencing panel.

Table S3: List of genes associated with prostate cancer used for the development of the customized targeted sequencing panel. Genes were selected based on literature review, TAPG pilot study conducted in the Wolfson Institute of Preventive Medicine at Queen Mary University of London and on meta‐analysis of published data.

Table S4: Sequencing coverage and quality statistics. Reference genome: GRCh38.

Table S5: Performance of the ML training models on the test data, based on accuracy, sensitivity, specificity, and kappa, used for the development of the classifiers to predict PCa (Table 1) and to differentiate aggressive from indolent PCa (Tables 2 and 3). Analysis of Deviance Table to verify if the removal of a feature would affect the results from the multivariate Cox model analysis (Table 4).

Table S6: Variants with significant association with aggressive prostate cancer among the 162 prostate cancer samples from the TAPG TURP cohort. The 162 samples were compared between them (ALIVE vs. DEAD) to examine aggressiveness of PCa.

Table S7: Significantly enriched GO terms identified for the 2685 variants across 269 genes with assigned GO terms among the 162 prostate cancer samples from the TAPG TURP cohort.

Table S8: Over‐represented gene pathways associated with DNA damage and/or DNA repair, for genes with occurring SNPs in prostate cancer samples with Gleason score 6 or 7 from the TAPG TURP cohort.

Table S9: Significantly enriched pathways found for the 2685 identified variants among the 162 prostate cancer samples from the TAPG TURP cohort.

Table S10: Significantly enriched pathways found for the 21 variants selected to calculate the polygenic risk score among the 162 prostate cancer samples from the TAPG TURP cohort.

DATA AVAILABILITY STATEMENT

The bam files from the targeted sequencing data of the 162 prostate cancer samples have been deposited in the Genome Sequence Archive (https://ngdc.cncb.ac.cn/gsa/) under the BioProject PRJCA033171. All source code is publicly available on GitHub (https://github.com/lrsantiago/Classifier_PCa). The GWAS catalog number for each of the SNPs downloaded from the GWAS database is available in Supplementary Table S2. Further information that supports the findings of this study is available from the corresponding author upon reasonable request.