Predicting pathogenic DNA damage repair gene mutations in prostate cancer patients: a multi-center magnetic resonance imaging radiomics study

Abstract

Background

Genetic testing for pathogenic DNA damage repair gene (pDDRg) mutations has clinical benefits for prostate cancer (PCa) patients, but its real-world application faces challenges due to its high associated costs. We sought to develop a magnetic resonance imaging (MRI)-based radiomics model capable of assessing the likelihood of PCa patients harboring pDDRg mutations. We then rigorously validated its predictive value in two external validation cohorts.

Methods

A total of 225 patients with both multiparametric MRI data before prostate biopsy and genetic testing information for pDDRg mutations were included in this study. The radiomics features were extracted from the T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) sequences of the MRI images in the training cohort (N=101) using the least absolute shrinkage and selection operator (LASSO) algorithm. The area under the curve (AUC) values of the receiver operating characteristic (ROC) curves and a decision curve analysis (DCA) were used to validate the predictive value of the model in both the internal (N=41) and external (N=83) validation cohorts.

Results

In total, 48 of the 225 (21.3%) patients in our cohort were identified by genetic testing as having positive pDDRg mutations, including BRCA1/2 (N=13), CDK12 (N=15), ATM (N=9), and other pDDRg mutations (N=17). Thirteen radiomics features from T2WI (N=7) and ADC sequences (N=6) were extracted to develop a model predicting pDDRg mutation carriers. The radiomics-based model had AUC values of 0.824 [95% confidence interval (CI): 0.677–0.923] in the internal validation dataset and 0.836 (95% CI: 0.738–0.908) in the external validation dataset. Notably, setting the cut-off value as “zero misseddignoses” resulted in a potential reduction of around 25% in unnecessary gene testing across both the internal and external validation datasets.

Conclusions

Our MRI radiomics-based predictive model is a promising pre-testing tool for pDDRg mutation prediction in patients with PCa. Prospective studies need to be conducted to further validate the power of this predictive model before its clinical application.

Introduction

The presence of pathogenic DNA damage repair gene (pDDRg) mutations has significant prognostic value and predictive utility in prostate cancer (PCa) management. There is substantial clinical evidence that pDDRg alterations drive disease aggressiveness and portend unfavorable patient outcomes. Current epidemiological data reveal that approximately 23% of men with metastatic PCa harbor somatic DNA damage repair (DDR) mutations, while 12% carry germline DDR mutations (1). These frequencies are substantially higher than those observed in localized PCa patients (1-3).

Emerging evidence from prospective trials highlights the therapeutic potential of poly ADP-ribose polymerase inhibitors (PARPis) and platinum-based regimens in the treatment of advanced PCa patients with pDDRg mutations (4-8). For instance, the PROfound trial established that olaparib significantly enhanced the progression-free survival of metastatic castration-resistant prostate cancer (mCRPC) patients carrying BRCA1/2 or ATM alterations, reinforcing the clinical relevance of these genetic biomarkers. Similarly, the TALAPRO-2 trial (5) confirmed that patients with CDK12 mutations also showed a positive response to talazoparib, another PARPi. Additionally, the MAGNITUDE trial found that mCRPC patients and pDDRg mutations, including BRCA1/2, ATM, and others, had improved outcomes with niraparib treatment, further supporting the clinical utility of PARPis in this patient population.

Given these therapeutic advances, both the National Comprehensive Cancer Network (NCCN) guidelines and expert consensus now mandate DDR mutation testing to enable the formulation of precision treatment strategies for advanced/metastatic PCa patients. According to the NCCN guidelines, patients with high-risk or very high-risk localized PCa, locally advanced PCa, or metastatic PCa (both hormone-sensitive and castration-resistant) should undergo DDR genetic testing, as should those who have ≥1 first-, second-, or third-degree relative with high-risk or very high-risk localized PCa, locally advanced disease or metastatic PCa, and those with ≥2 first-, second-, or third-degree relatives with PCa (any stage). However, pDDRg mutations in PCa are not as prevalent as they are in other solid tumors, such as endometrial, bladder, and ovarian malignancies. The approximate mutation rate of pDDRg mutations in PCa has been reported to range from 8% to 27% (average mutation rate: 14.1%) (9-11).

Current risk-based recommendations for genetic testing, which frequently depend on age, clinical patterns, and family history, may not be sufficient in clinical practice because they can exclude the testing of more than one-third of males with germline DDR alterations (12). In the current era of precision medicine, a non-invasive pre-testing predictive tool urgently needs to be developed to identify potential patients harboring pDDRg mutations. Any such tool would reduce unnecessary genetic testing and could eventually also achieve the cost-effectiveness of other genetic screening procedures.

Radiomics is a rapidly developing field in cancer research that aims to quantify the qualitative information collected in medical imaging. This approach can be used to unveil the underlying biochemical intricacies of tumors. Notably, previous studies have successfully demonstrated the feasibility of using radiomic features extracted from magnetic resonance imaging (MRI) to predict certain molecular events in diverse malignancies, including glioma, rectal cancer, and non-small cell lung cancer (13-16). Radiomic models leverage pre-existing imaging data and therefore incur no additional costs for patients. This positions radiomics as a cost-effective triage mechanism within diagnostic methods, enabling the stratification of patients by mutation probability prior to confirmatory genetic testing, and thereby avoiding unnecessary expenditures in low-risk cohorts while prioritizing high-risk candidates. However, the relationship between the radiomic features of multiparametric MRI and the possibility of pDDRg mutations in PCa is still unknown.

In this study, we sought to establish a non-invasive radiomics tool to identify PCa patients likely to harbor pDDRg mutations, thereby optimizing genetic testing strategies. We present this article in accordance with the TRIPOD + AI reporting checklist (17) (available at https://qims.amegroups.com/article/view/10.21037/qims-2024-2708/rc).

Methods

The study population for model development, validation, and external testing

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of West China Hospital [No. 2022(78)], and the requirement of individual consent for this retrospective analysis was waived. All participating hospitals/institutions were informed of and approved the study.

The training and internal validation datasets comprised consecutive patients diagnosed with PCa who had undergone genetic testing for DNA damage repair gene (DDRg) mutations from May 2015 to December 2021 at the West China Hospital. These patients were randomized into the training and internal validation groups at a 7:3 ratio using a stratified sampling approach to ensure balanced rates of pDDRg mutations between the cohorts. The external validation dataset comprised patients treated at the Fudan University Shanghai Cancer Center from May 2015 to December 2021 and Sun Yat-sen University Cancer Center from May 2015 to December 2021 (Figure 1A and Figure S1). Detailed inclusion and exclusion criteria are provided in Appendix 1.

Figure 1.

Study design, radiomics workflow, and landscape of DDR gene mutations of the included prostate cancer patients. (A) The study design and radiomics workflow. (B) The OncoPrint visualizes the clinical-pathologic and genomic variables of all the included patients. The columns signify individual patients, while the rows indicate the frequencies of all the analyzed clinical-pathologic traits and the frequent pathogenic DDR gene mutations. 3D, three-dimension; ADC, apparent diffusion coefficient; DCA, decision curve analysis; DDR, DNA damage repair; DICE, dice similarity coefficient; GLCM, gray level co-occurence matrix; GLDM, gray level dependence matrix; GLRLM, gray level run-length matrix; GLSZM, gray level size zone matrix; GS, Gleason score; H, high-pass filtering; ICC, intraclass correlation coefficient; L, low-pass filtering; M stage, metastatic stage; NA, not applicable; NGTDM, neighbouring gray tone difference matrix; PCa, prostate cancer; pDDRg, pathogenic DNA damage repair gene; PSA, prostate specific antigen; ROC, receiver operating characteristic; T2WI, T2-weighted imaging; VOI, volume of interest; WCH, West China Hospital.

Clinicopathological information was collected from all of the enrolled patients, including (I) clinical characteristics: age, weight, and body mass index; (II) laboratory indices: levels of serum prostate specific antigen (PSA); and (III) histopathologic features: tumor-node-metastasis stage, Gleason score, tumor size, and tumor number.

Genetic testing procedure

Germline DNA was extracted from the patients’ mononuclear cells in the peripheral blood. Following a hematoxylin and eosin stained histological evaluation to confirm ≥20% tumor content, genomic DNA was extracted from diagnostically validated formalin-fixed paraffin-embedded tissue sections for the downstream analyses. Targeted capture was performed using a customized next-generation sequencing panel, which specifically targets 21–733 cancer-related genes, including 18 DDRgs, at three medical centers (West China Hospital, Fudan University Shanghai Cancer Center and Sun Yat-sen University Cancer Center). Unlike whole-exome sequencing, this panel was focused on a predefined set of genes relevant to cancer and DDR, and thus had high sensitivity in the detection of mutations in these critical genes. All the Genetic testing platforms covered the 18 DDRgs (i.e., ATR, MSH6, NBN, BRCA2, CDK12, ATM, MSH3, RAD50, PALB2, FANCA, BRIP1, SLX4, BRCA1, FANCD2, MSH2, MLH3, BLM, and FANCE). Variants, including single-nucleotide variants and insertions/deletions, were detected as described previously (3). Sequencing achieved a mean coverage depth of 1,000×, and the read alignment to the hg19 reference genome was performed using Burrows-Wheeler Aligner (version 0.7.12).

The germline variant interpretation adhered to the American College of Medical Genetics and Genomics and the Association for Molecular Pathology guidelines. We developed an in-house analysis pipeline incorporating population frequencies, computational predictions, and functional annotations, with quarterly updates to maintain its interpretation accuracy and clinical relevance. The updates incorporated newly published research, revised population frequency databases, advancements in computational prediction models, and improved functional annotation data. Two certified geneticists independently assessed each variant interpretation while monitoring newly published evidence. If any inconsistencies in the classifications arose, a senior genetic consultant reviewed the classification and made the final determination. The pathogenic/likely pathogenic variants of DDRgs were also analyzed in this study.

MRI technique and evaluation

The MRI parameters used by the three institutions (West China Hospital, Fudan University Shanghai Cancer Center and Sun Yat-sen University Cancer Center) are summarized in Table S1 in Appendix 1. All the scans were performed with a 16-channel surface coil. No endorectal coil was used. No fat suppression technique was used in the T2-weighted imaging (T2WI) sequence. The B values for diffusion scanning were 0/800/1,400 s/m². Apparent diffusion coefficient (ADC) mapping was calculated based on b values of 800 s/m². All the imaging examinations were performed at West China Hospital for blinded review by two genitourinary radiologists (Y.C. and L.Y.), each with >5 years of subspecialty experience. Strict blinding protocols prevented access to patients’ clinicopathological and genetic profiles during the image analysis.

VOI segmentation and radiomic feature extraction

Two board-certified radiologists specializing in genitourinary imaging (with a minimum of five years of experience each) independently reviewed all the MRI scans and annotated the largest detectable lesions. Lesions in the peripheral zone and transitional zone were measured in the T2WI and ADC maps, respectively. Any inconsistencies in opinion were resolved by consultation. The two radiologists would then make a three-dimensional volume of interest (VOI) of the marked lesions on the T2WI and ADC sequences in the Digital Imaging and Communications in Medicine images using ITK-SNAP software (version 2.01). The dice value was calculated by different radiologists. If one case’s dice value was <0.90, re-segmentation after discussion with a genitourinary imaging expert (J.Y.) was performed until the dice value was >0.90. To mitigate the effects of variations in image intensities on the stability and reproducibility of radiomic features, Z-score intensity normalization was implemented before feature extraction. The quantitative radiomic analysis yielded 1,675 distinct features per lesion, with identical feature sets extracted independently from the T2WI and ADC mapping sequences. Details of the radiomics procedures and radiomic features are provided in Appendix 1.

Model construction and testing

Least absolute shrinkage and selection operator (LASSO) logistic regression, a method particularly suited for high-dimensional data, was employed to identify treatment response-associated features (non-zero coefficients) in the training cohort. The resulting radiomics signature, derived from a weighted linear combination of selected features, quantified the DDR mutation probability for each prostate lesion. The predictive performance of the radiomics signature was initially evaluated in the training set using the area under the receiver operating characteristic (ROC) curve and area under the curve (AUC) values, and subsequently validated in the independent internal and external cohorts.

Statistical associations with DDR mutations were analyzed using the Mann-Whitney U test, complemented by comprehensive subgroup analyses. Details of the radiomics procedures and radiomic features are presented in Appendix 1. To evaluate whether adding the radiomics signature improved the performance of the clinical predictors, ROC analyses were performed in all patients to compare the discriminatory efficacy of the radiomics model with that of each selected clinical predictor and the combined clinical predictors. A decision curve analysis (DCA) was performed to determine the clinical utility of the radiomics model by quantifying the net benefits at different threshold probabilities.

Statistical analysis

All the statistical analyses were conducted using R software (version 3.4.2, R Foundation for Statistical Computing). Our analytical workflow incorporated specialized R packages for each methodological component as follows: (I) the LASSO logistic regression was implemented via “glmnet”; (II) the ROC curve analysis was conducted using “pROC”; (III) the multicollinearity assessment was performed using “car” [variance inflation factor (VIF) threshold =10]; (IV) calibration was evaluated using the “rms” package; (V) the Hosmer-Lemeshow goodness-of-fit test was performed using “vcdExtra”; and (VI) the DCA was conducted using the custom “dca.R script”. A two-sided P value <0.05 was considered statistically significant.

Results

Patients’ characteristics

The study cohort comprised 225 eligible participants. Table 1 sets out the comprehensive clinicopathological characteristics of these participations at the baseline. In total pDDRg mutations were detected in 48 of the 225 patients (21.33%), of whom 15 had germline mutations and 33 had DDR mutations (Figure 1B). The most frequently altered genes were CDK12 (N=15; 6.7%), BRCA1/2 (N=13; 5.8%), and ATM (N=9; 4.0%). Seven patients carried more than one pDDRg mutation (Table S1). The baseline factors, including metastatic status, age, International Society of Urological Pathology (ISUP) grading, and the baseline serum PSA level, were not associated with the occurrence of the pDDRg mutations (Table S2).

Table 1. Baseline characteristics of the patients in the training, internal validation, and external validation dataset.

Characteristics	Whole cohort, n (%)	Training dataset, n (%)	Internal validation dataset, n (%)	External validation dataset, n (%)	P
M stage					0.131
M0	103 (45.78)	45 (44.55)	14 (34.15)	44 (53.01)
M1	122 (54.22)	56 (55.45)	27 (65.85)	39 (46.99)
Age (years)					0.629
<70	126 (56.00)	53 (52.48)	24 (58.54)	49 (59.04)
≥70	99 (44.00)	48 (47.53)	17 (41.46)	34 (40.96)
PSA (ng/mL)					<0.001
<20	49 (21.78)	12 (11.88)	13 (31.71)	24 (28.92)
≥20	131 (58.22)	61 (60.40)	18 (43.90)	52 (62.65)
Unknown	45 (20.00)	28 (27.72)	10 (24.39)	7 (8.43)
ISUP grading group					0.361
1–3	84 (37.33)	41 (40.59)	17 (41.46)	26 (31.33)
4–5	141 (62.67)	60 (59.41)	24 (58.54)	57 (68.67)
pDDR mutation					0.507
Negative	177 (78.67)	76 (75.25)	34 (82.93)	67 (80.72)
Positive	48 (21.33)	25 (24.75)	7 (17.07)	16 (19.28)
Mutation type					0.484
Negative	177 (78.67)	76 (75.25)	34 (82.93)	67 (80.72)
Somatic	33 (14.67)	15 (14.85)	5 (12.20)	13 (15.66)
Germline	15 (6.67)	10 (9.90)	2 (4.88)	3 (3.61)

ISUP, International Society of Urological Pathology; M, metastasis; pDDR, pathogenic DNA damage repair; PSA, prostate specific antigen.

To explore the correlation between the radiomic features on the MRI sequences and the pDDRg mutations, we randomly divided the total cohort from our center (Westchina hospital) into the training (N=101, 70%) and internal validation (N=41, 30%) cohorts. The external validation was conducted using data from a cohort of 83 patients obtained from two additional medical centers (Fudan University Shanghai Cancer Center and Sun Yat-sen University Cancer Center). Among the baseline characteristics, only the PSA level exhibited significant differences across the cohorts, such that a higher proportion of patients had unknown PSA levels in the training and internal validation cohorts. The other baseline characteristics and pDDRg mutation rates were well-balanced among the three cohorts (Table 1).

The development of an MRI-based radiomic model predicting pDDRg mutations

The quantitative radiomic analysis yielded 1,675 distinct features extracted from VOIs across three MRI sequences: T2WI, ADC maps, and contrast-enhanced scans (https://pyradiomics.readthedocs.io/en/latest/features.html). Among these features, 201 had a high reproducibility, which was defined as an intraclass correlation coefficient ≥0.7 between the two radiologists. The LASSO logistic regression algorithm was used to evaluate the ability of these reliable features to predict pDDRg mutations. Finally, a combined predictive model comprising 13 radiomic features (seven T2WI features and six ADC features) was constructed by linear regression (Appendix 1).

We then examined the radiomics-based predictive model using different validation approaches. Compared to models that exclusively incorporated ADC or T2WI features alone (Table S3 and Figure 2), the combined model showed improved performance, and had AUC values of 0.849 [95% confidence interval (CI): 0.746–0.902] and 0.828 (95% CI: 0.677–0.923) in the training and internal validation cohorts, respectively (Figure 2A,2B). The result of the Hosmer-Lemeshow test was not significant (P=0.484), indicating that the model achieved a close fit with the actual data. Further, the agreement between the model-predicted probabilities and the observed frequencies of pDDRg mutations was quantitatively assessed using calibration curves. The predicted rates of the pDDRg mutations were found to be positively correlated with the actual rates observed (Figure 2G,2H). Moreover, the DCA results also exhibited satisfactory positive net benefits across a range of threshold probabilities for the models predicting pDDRg mutations, indicating the potential clinical benefit of this combined model (Figure 2D,2E).

Figure 2.

Comparative performance of various radiomic signatures in the training and validation datasets. (A-C) AUCs for the receiver operating characteristic curves of T2WI, ADC, and the combined radiomic model in the training set (A), internal validation set (B), and external validation dataset (C). (D-F) Decision curves for T2WI, ADC, and the combined radiomic model for the training set (D), internal validation set (E), and external validation dataset (F). The light grey curve symbolizes the assumption that all the patients received genetic testing, while the dark grey curve indicates the assumption that no patients were subjected to genetic testing. The x-axis signifies the threshold probability, which is the point at which the anticipated benefit of the treatment matches the expected advantage of forgoing treatment. (G-I) Calibration curves for the combined model across the training set (G), internal validation set (H), and external validation dataset (I). ADC, apparent diffusion coefficient; AUC, area under the curve; T2WI, T2-weighted imaging.

The efficacy of this novel MRI-based radiomics pDDRg mutation predictive model for PCa was further assessed using an external dataset comprising data from the other two medical centers. The external validation indicated that the MRI-based radiomics model had an AUC of 0.836 (95% CI: 0.738–0.908), which closely resembled the performance achieved in the aforementioned internal validation dataset (Figure 2C). Additionally, the calibration and decision curves also showed that the radiomics model performed well in the external validation (Figure 2F,2I).

Clinical feasibility of the constructed radiomics model

We also explored the optimal cut-off value of this model. The likelihood of a pDDRg mutation increased substantially as the MRI radiomics score increased (Figure 3A). The threshold for maximum sensitivity (zero missed pDDRg mutations) was set based on the ROC curve, and we found that the number of patients who underwent genetic testing could be reduced by an average of 25% without missing any patients with positive pDDRg mutations in both the internal and external cohorts (Figure 3B).

Figure 3.

Efficacy of the combined radiomics model across various thresholds and patient subgroups. (A) Visualization of the risk scores assigned to each patient. The x-axis delineates the individual patients, while the y-axis displays their respective risk scores as computed by the combined radiomics model. The columns are color-coded based on the DDR status of each patient, with the red arrow indicating the presence of a germline mutation. (B) The model’s potential to decrease genetic testing across different thresholds. “Performed” signifies the number of patients recommended for genetic testing according to the radiomics risk score, while “Reduced” indicates the number of patients who can forgo testing based on the same score. “Detected” represents the count of patients with pDDRg mutations among those selected for genetic testing, and “Missed” shows the count of patients with pDDRg mutations excluded from testing. The x-axis introduces various thresholds for genetic testing decisions related to pDDRg mutations, and the y-axis enumerates the patients. (C) Evaluation of the combined radiomics model’s effectiveness in distinct patient subgroups. AUC, area under the curve; CI, confidence interval; ISUP, International Society of Urological Pathology; M, metastasis; pDDRg, pathogenic DNA damage repair gene; PSA, prostate specific antigen.

Additionally, we evaluated the predictive power of the radiomics model in patients with different baseline characteristics. The results revealed that the model’s predictive accuracy was concordant, regardless of metastatic status, age, ISUP grading, baseline serum PSA level, or somatic/germline pDDRg mutations (Figure 3C). Notably, while most radiomic features included in the model were equally effective at predicting the BRCA, ATM, and CDK12 gene mutations, the “shape flatness original” feature from T2WI was significantly negatively correlated with the ATM mutation (Figure S2).

To further facilitate the practical application of our findings, we also developed a dedicated website (https://research.infervision.com/v2/) featuring an interactive calculator that operationalizes our radiomics model. This tool will enable the feasible prediction of pDDRg mutations in clinical settings (Figure 4).

Figure 4.

MRI illustrations of lesions used for radiomic evaluation. (A) Patient 17: a 61-year-old male with a PSA level of 25.0 ng/mL. The biopsy indicated a Gleason score of 4+5 (ISUP 5 group). Bone metastasis was identified by the bone scan. MRI depicted a low ADC value and T2 hyperintensity signal spanning the entire TZ and portions of the PZ. Radiomics risk score: 70.17%; categorized as high risk for pDDRg mutation. The genetic test identified a somatic BRAC2 mutation. (B) Patient 45: a 77-year-old male with a PSA level of 14.0 ng/mL. The biopsy indicated a Gleason score of 4+4 (ISUP 4 group). No metastasis was observed, and MRI revealed a low ADC value along with a T2 hypointensity signal in the right TZ and right PZ. Radiomics risk score: 66.43%; categorized as high risk for pDDRg mutation. The genetic test identified a somatic ATM mutation. (C) Patient 44: a 75-year-old male with a PSA level of 57.7 ng/mL. The biopsy indicated a Gleason score of 4+4 (ISUP 4 group). Both bone and lymph node metastases were detected via PET/CT. MRI showed a low ADC value and a T2 hyperintensity signal in the PZ. Radiomics risk score: 62.66%; categorized as high risk for pDDRg mutation. The genetic test revealed a somatic CDK12 mutation. (D) Patient 72: a 74-year-old male with a PSA level of 51.2 ng/mL. The biopsy indicated a Gleason score of 4+4 (ISUP 4 group). Bone and lymph node metastases were revealed by PET/CT, with MRI indicating a low ADC value and T2 hyperintensity signal in the left TZ and left PZ. Radiomics risk score: 14.44%; categorized as low risk for pDDRg mutation. The genetic testing did not detect any pDDRg alteration. ADC, apparent diffusion coefficient; DDR, DNA damage repair; DWI, diffusion-weighted imaging; ISUP, International Society of Urological Pathology; M, metastasis; MRI, magnetic resonance imaging; PET/CT, positron emission tomography/computed tomography; PSA, prostate specific antigen; PZ, peripheral zone; pDDRg, pathogenic DNA damage repair gene; T2 SAG, sagittal T2-weighted imaging; T2 TRA, transverse T2-weighted imaging; TZ, transition zone.

Discussion

Genetic testing for PCa is driven by the principles of precision medicine. The detection of pDDRg mutations is of great significance in both predicting clinical outcomes (18,19) and assisting in treatment decision making (4-8) in PCa patients. Radiomics has recently gained prominence as a transformative approach that enables precision medicine through quantitative image analysis (20). This innovative methodology, which leverages existing medical imaging data as the foundation for comprehensive analysis, stands out because of its non-invasive nature and cost-effectiveness. In the present study, we conducted an in-depth investigation of the ability of MRI-based radiomic features to predict the likelihood of pDDRg mutations in PCa patients. Our approach involved the development of a sophisticated radiomics model, comprising 13 radiomic features extracted from T2WI and ADC sequences. This radiomics features-based model demonstrated remarkable performance in the external validation across multi-centers, reaffirming its robust predictive capability across different patient subgroups.

Radiomics is a fast-developing field that holds considerable promise in cancer research, encompassing domains like pathological diagnosis, molecular subtyping, and prognosis (16,21-23). Notably, various studies have shown the potential of computed tomography (CT)- or MRI-based radiomic features in predicting certain molecular events among patients with different tumors (24). In terms of pDDRg mutation prediction, several studies have shown the potential of CT-based radiomics models in predicting BRCA mutations in ovarian cancer (25-27). Further, two recent studies have underscored the efficacy of MRI radiomics in estimating the homologous recombination deficiency score and BRCA mutation status of individuals with breast cancer (28,29). However, the image diagnosis of breast cancer is mainly based on the T1WI sequence, while the diagnosis of PCa is mostly dependent on T2WI, diffusion-weighted imaging, and ADC sequences. Obviously, the T1WI-based radiomic features cannot be used to predict pDDRg mutations in PCa. Further, previous radiomics studies have primarily focused on predicting BRCA gene mutations, overlooking other important DDR-related gene mutations, such as CDK12 and ATM mutations.

Using advanced machine-learning techniques, we established a radiomic signature derived from MRI data for precise pDDRg mutation prediction in PCa patients. More importantly, we assessed the predictive capacity of the novel model in patients with various pDDRg alterations. No discernible difference was found between the model’s predictions for somatic and germline pDDRg mutations. These findings show the robustness of our radiomics model in predicting mutations in different DDR-related genes.

To assess the practical feasibility of our novel radiomics model, we conducted a validation study using external cohorts from other medical centers with varying MRI machines, which suggests that changes in the MRI machines did not significantly affect the predictive accuracy of the model, underscoring its robustness. Moreover, to make the radiomics model more convenient for clinical use, we determined the risk threshold at which the negative predicted value could remain at 100% and how many cases could be eliminated at this point. When the threshold for a pDDRg mutation test was set to zero ‘missed diagnoses’ our findings revealed that nearly 25% of patients in both the internal and validation cohorts could avoid unnecessary gene testing.

The current study also assessed the power of the predictive model among patients with different clinicopathological factors. Our findings indicated that this radiomics model showed similarly favorable efficacy in predicting the possibility of pDDRg mutations, regardless of age, the baseline PSA level, the ISUP grading, and metastatic (M) stage. Previous research (30) has reported that a young age is related to the pDDRg mutation; however, we found no evidence of any such relationship. It is important to note that our dataset lacked information on the family history of the patients. To address this limitation, future research should seek to examine the associations between clinical parameters, including family history, and pDDRg mutations in PCa patients. In the future, we intend to incorporate these clinical parameters into our radiomics model to potentially augment its predictive accuracy.

Our results also showed that tumors with the ATM mutation might exhibit specific radiomic features in contrast to those harboring the BRCA2 and CDK12 mutations. The observed negative correlation implies that ATM-mutant PCa may exhibit more elongated morphological patterns than wild-type tumors, potentially reflecting distinct invasive behaviors; however, further study needs to be conducted. This observation suggests the existence of potentially unique radiomic features associated with specific genetic alterations. Further research into the differential radiomic characteristics among different pDDRg mutations may extend our understanding of the underlying biological mechanisms of PCa with different type of pDDRg mutation.

This study had several limitations. First, the retrospective design might have introduced recall bias into the study, and it also limited our control over the data collection processes. Second, while the external validation showed the robustness of our novel radiomics model, the results need to be validated in cohorts of patients from diverse ethnic backgrounds (beyond Asian populations). Third, we were unable to analyze the association between the patients’ outcomes and the model’s predictive ability due to follow-up loss in some cases and the retrospective nature of the study. Finally, this study only incorporated MRI radiomic features into the model and did not include any clinical factors associated with pDDRg mutations. In the future, we intend to incorporate clinical factors into the model in an attempt to enhance its predictive performance.

Conclusions

The current study investigated the potential of radiomic features extracted from T2WI and ADC MRI sequences to detect the presence of pDDRg mutations in PCa patients. A MRI-based radiomics model was established that exhibited satisfactory performance in predicting the possibility of pDDRg mutation carriers among PCa patients. Our model may serve as a valuable tool for precise genetic testing and curtail superfluous genetic screening in PCa patients.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of West China Hospital [No. 2022(78)], and the requirement of individual consent for this retrospective analysis was waived. All participating hospitals/institutions were informed of and approved the study.