Multimodal deep learning model for predicting homologous recombination deficiency in prostate cancer: an international multi-cohort study

Multimodal deep learning model for predicting homologous recombination deficiency in prostate cancer: an international multi-cohort study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Multimodal deep learning model for predicting homologous recombination deficiency in prostate cancer: an international multi-cohort study Zijian Song, Qianwen Zhang, Wenhui Zhang, Na Ta, Yan Zhu, Longxin Deng, and 15 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6488233/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Accurate prediction of homologous recombination deficiency (HRD) status is crucial for effective management of prostate cancer (PCa). However, genetic testing is expensive and inaccessible. We propose a multimodal deep learning approach that integrates clinical information, Hematoxylin & Eosin (H&E)-stained whole-slide images (WSIs), and multi-parameter MRI images to predict the HRD status of patients with PCa. Patients from the Cancer Genome Atlas (n = 387) and three Chinese hospitals (n = 179) were used to establish and validate the prediction models. The Pathology Signature, Radiology Signature, and Patho-Radiology Signature could accurately predict the HRD status in external validation or 5-fold cross validation (AUC = 0.815, 0.833, and 0.933, respectively). Notably, four interpretative pathological features were identified in the Pathology Signature. These signatures could serve as a prescreening tool to select patients for confirmatory genetic testing. We suggest that this multi-modal approach could be applied to the prediction of molecular alterations in other malignancies. Biological sciences/Cancer/Urological cancer/Prostate cancer Health sciences/Pathogenesis/Clinical genetics Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 INTRODUCTION Prostate cancer (PCa) is the most commonly diagnosed malignancy and the fifth leading cause of cancer-related death in men worldwide 1 . PCa is a malignancy that is characterized by diverse molecular alterations 2 . Homologous recombination deficiency (HRD) is a frequently observed molecular alteration in PCa 3 . HRD leads to defective DNA break repair, increased somatic copy number alterations, genomic instability, and oncogenesis[4,5]. HRD has emerged as a potent biomarker for selecting patients for Poly(ADP-Ribose)-polymerase (PARP) inhibitor (PARPi) treatment of various cancers, including advanced-stage PCa, as demonstrated in the PROfound study 6 . However, HRD detection typically relies on genetic sequencing, which is associated with limited accessibility, high cost, and a long detection period 7 . Considering that the prevalence of HRD is approximately 5–10% in primary PCa 8 and approximately 13% in advanced-stage PCa, ten genetic tests are required to identify one patient with HRD 3 . Therefore, there is an urgent need for a rapid, low-cost, and accessible method to detect HRD status or, at the very least, prescreening patients with a higher risk of HRD status. Artificial intelligence (AI), particularly deep learning (DL), has shown great promise in identifying pathological subtypes, molecular alterations, and prognostic outcomes in cancer patients 9 . Prediction of Microsatellite Instability (MSI) in gastrointestinal cancer is one of the most investigated applications. Kather et al. proposed the application of deep learning in the analysis of conventional Hematoxylin & Eosin (H&E) histology to predict microsatellite instability based on multiple international cohorts 10 . Yamashita et al. proposed DL models that surpassed the performance of experienced gastrointestinal pathologists in predicting MSI on H&E-stained whole-slide images (WSIs) in colorectal cancer 11 . However, despite the progress made in utilizing deep learning methods for predicting molecular alterations in various cancer types, a comprehensive review highlighted the limited exploration in this field in the context of PCa 12 . Chinnaiyan et al. reported the application of DL in the analysis of conventional pathological slides to predict the presence of ETS-related gene (ERG) fusions in a cohort of 392 cases 13 . Mark et al. reported the application of DL-based methods for speckle-type POZ protein (SPOP) mutation prediction in 177 PCa patients, but the study was not officially published 14 . The main limitation of these studies may be the small sample size due to poor data accessibility, as there were a limited number of PCa cases with available pathology images and molecular alteration data. Furthermore, studies aimed at predicting molecular alterations in PCa using Magnetic Resonance Imaging (MRI) data are still in the preliminary stages and involve a limited number of patients 15 . Recently, multimodal data were analyzed using DL algorithms in the risk stratification of high-grade serous ovarian cancer 16 , in the prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer 17 , and in the prediction of molecular classification of endometrial cancer 18 . In a previous study, we presented the Chinese Prostate Cancer Genetic and Epigenetic Atlas (CPGEA) cohort, which is one of the largest multi-omics datasets of PCa, encompassing genomic, transcriptome, and epigenetic data 19 . In this study, we for the first time, collected and analyzed the pathological and imaging data of patients in the CPGEA cohort. By combining internal cross-validation and external validations, we intend to establish and validate multimodal models for predicting the HRD status in patients with PCa. RESULTS Cohorts and characteristics After data pre-processing, 387 patients from the TCGA-PRAD cohort and 179 patients from the CPGEA cohort were included in this study. The CPGEA cohort was sequenced and reported in 2020 by our team, and we collected and integrated the pathological, radiological, and genomic data for the first time. Patient characteristics are summarized in Table 1 . Clinical-genomic data were available for all cases, whereas pathological data were available for 157 and 386 cases in the CPGEA and TCGA-PRAD cohorts, respectively. Radiological data were available for 109 cases in the CPGEA cohort and eight cases in TCGA-PRAD (two cases with only T2 images were excluded from further analysis). HRD was detected in 36 (20.7%) and 73 (18.9%) patients in the CPGEA and TCGA-PRAD cohorts, respectively ( Figure 1 ). The distribution of cases across training and testing sets for different modalities, as well as their distribution in the database, can be found in Supplementary Figure S1. Table 1. Patient characteristics. CPGEA TCGA-PRAD Total HRP (N=143) HRD (N=36) P value HRP (N=314) HRD (N=73) P value HRP (N=457) HRD (N=109) P value AGE 0.974 0.092 0.134 Mean (SD) 68.8 (6.24) 68.8 (5.30) 60.4 (6.82) 62.0 (7.25) 63.0 (7.69) 64.2 (7.38) PSA 0.260 0.483 0.493 Median(IQR) 18.9 [11.3, 37.6] 23.1 [13.3, 35.1] 7.2 [5.0, 10.1] 7.9 [5.2, 12.0] 8.5 [5.6, 16.8] 10.7 [5.8, 21.9] Missing 0 (0%) 0 (0%) 7 (2.2%) 3 (4.1%) 7 (1.5%) 3 (2.8%) pT 0.258 0.694 0.459 T2 65 (45.5%) 19 (52.8%) 127 (40.4%) 26 (35.6%) 192 (42.0%) 45 (41.3%) T3 73 (51.0%) 14 (38.9%) 176 (56.1%) 44 (60.3%) 249 (54.5%) 58 (53.2%) T4 5 (3.5%) 3 (8.3%) 2 (2.7%) 11 (2.4%) 5 (4.6%) Tx 0 (0%) 0 (0%) 5 (1.6%) 1 (1.4%) 5 (1.1%) 1 (0.9%) pN 1.000 0.544 0.471 N0 104 (72.7%) 25 (69.4%) 217 (69.1%) 52 (71.2%) 321 (70.2%) 77 (70.6%) N1 18 (12.6%) 5 (13.9%) 41 (13.1%) 13 (17.8%) 59 (12.9%) 18 (16.5%) Nx 21 (14.7%) 6 (16.7%) 56 (17.8%) 8 (11.0%) 77 (16.8%) 14 (12.8%) pM 0.097 0.477 0.192 M0 84 (58.7%) 13 (36.1%) 146 (46.5%) 38 (52.1%) 230 (50.3%) 51 (46.8%) M1 16 (11.2%) 7 (19.4%) 0 (0%) 1 (1.4%) 16 (3.5%) 8 (7.3%) Mx 43 (30.1%) 16 (44.4%) 168 (53.5%) 34 (46.6%) 211 (46.2%) 50 (45.9%) Residual_T 0.324 1.000 0.825 YES 57 (39.9%) 18 (50.0%) 4 (1.3%) 1 (1.4%) 61 (13.3%) 19 (17.4%) NO 84 (58.7%) 17 (47.2%) 290 (92.4%) 67 (91.8%) 374 (81.8%) 84 (77.1%) Missing 2 (1.4%) 1 (2.8%) 20 (6.4%) 5 (6.8%) 22 (4.8%) 6 (5.5%) ISUP 0.590 0.074 0.051 1 8 (5.6%) 2 (5.6%) 35 (11.1%) 4 (5.5%) 43 (9.4%) 6 (5.5%) 2 41 (28.7%) 7 (19.4%) 102 (32.5%) 15 (20.5%) 143 (31.3%) 22 (20.2%) 3 28 (19.6%) 5 (13.9%) 65 (20.7%) 22 (30.1%) 93 (20.4%) 27 (24.8%) 4 24 (16.8%) 6 (16.7%) 38 (12.1%) 9 (12.3%) 62 (13.6%) 15 (13.8%) 5 42 (29.4%) 15 (41.7%) 74 (23.6%) 23 (31.5%) 116 (25.4%) 38 (34.9%) Missing 0 (0%) 1 (2.8%) 0 (0%) 0 (0%) 0 (0%) 1 (0.9%) BCR at 3 years 1.000 0.211 0.197 YES 98 (68.5%) 25 (69.4%) 256 (81.5%) 63 (86.3%) 354 (77.5%) 88 (80.7%) NO 45 (31.5%) 11 (30.6%) 24 (7.6%) 2 (2.7%) 69 (15.1%) 13 (11.9%) Missing 0 (0%) 0 (0%) 34 (10.8%) 8 (11.0%) 34 (7.4%) 8 (7.3%) Abbreviations: SD=standard deviation; IQR=interquartile range; HRP=homologous recombination proficient; HRD=homologous recombination deficient; pT=pathological T stage; pN=pathological N stage; pM=pathological M stage; Residual_T=residual tumor presence; ISUP=International Society of Urological Pathology grade; BCR=biochemical recurrence. Clinical Signature Clinical variables were not significantly associated with the HRD status ( Supplementary Figure S2 ). Although age and International Society of Urological Pathology (ISUP) grade may be associated with HRD, the predictive value of clinical variables is limited. In the TCGA-PRAD cohort, clinical variables such as age, PSA level, pathological TNM stage, residual tumor, and ISUP grade were not significantly associated with HRD status in either the univariate or multivariate logistic regression analyses. We constructed the Clinical Signature based on a K-Nearest Neighbor (KNN) model with data from the TCGA-PRAD cohort and validated it in the CPGEA cohort, yielding an area under the curve (AUC) of 0.563. The results indicated a moderate association between Clinical Signature and HRD status. Radiology Signature Owing to the limitation of available MRI data in the TCGA-PRAD cohort, we combined TCGA-PRAD and CPGEA cohorts and applied 5-fold cross-validation to establish and validate the Radiology Signature ( Supplementary Figure S1 ). We extracted shape features, first-order features, and texture category radiomic features that contained 1132 handcrafted features using Pyradiomics ( Supplementary Table S1 ). A total of 2264 features were included after concatenating the DWI and T2WI features. Nine features were selected based on the Least Absolute Selection and Shrinkage Operator (LASSO) regression feature-selection procedure ( Supplementary Table S2 ). Correlations among the features were explored using Spearman analysis and hierarchical clustering of the nine features. The Radiology Signature was established using the best-performing algorithm in the test cohort, the Support Vector Machine (SVM), which achieved an AUC of 0.833 ( Supplementary Table S3 ). Some linear models achieved higher results than nonlinear models, which may be related to the interpretability and physical meaning of handcrafted imaging features. Waterfall plots depicted the distribution of the HRD status in Radiology Signature ( Figure 2 ). Pathology Signature First, we divided the WSI into 512×512 patches, and only patches with an overlap of over 80% with the ROI were used for the subsequent analysis. Patch-level tumor-region prediction was performed based on the ResNet50 model. The ResNet50 model demonstrated high efficacy in distinguishing tumor regions from non-tumor regions, with an AUC of 0.936 (95%CI 0.935-0.936) in the test cohort (Figure 3, Supplementary Table S4) . Subsequently, the patch labels and corresponding probabilities were applied to generate 106 pathological WSI-level features using the PLH and BoW pipelines ( Supplementary Table S5 ). Then WSI-level features were selected by Lasso regression. Finally, 42 features were applied for model construction by a series of algorithms ( Supplementary Table S6 ), including Logistic Regression (LR), Support Vector Machine (SVM), Random Forest, XGBoost, LightGBM, to identify the HRD status of each patch. The LightGBM model (Pathology Signature) achieved an AUC of 0.815 in external validation. Applying the cutoff value of 0.395, we could detect the 100% (72/72) and 83.3% (15/18) HRD cases in the training and test cohorts, while reducing 77.6% (243/313) and 66.4% (83/125) genetic tests in the training and test cohorts, respectively ( Figure 3 and Supplementary Table S7 ). In the eleven patients received neoadjuvant ADT in the test cohort (all actual HRP), four cases were predicted as HRD and seven cases were predicted as HRP. Treatment-associated pathological changes may be associated with features of HRD. Multimodal Signatures Multimodal prediction is commonly believed to enhance the prediction accuracy. When combining radiology and pathological features to build the Patho-Radiology model, the predictive accuracy improved compared with the performance of the single-modality models, with an AUC of 0.933 in the Naïve Bayes algorithm model, 0.800 in the logistic regression model, 0.833 in the KNN model, and 0.867 in the LightGBM model, respectively ( Figure 4, Supplementary Table S8) . The model based on the Naïve Bayes algorithm was selected as the Patho-Radiology Signature. In addition, we compared the predictive performance of the Patho-Clinical, Patho-Radiology, and Patho-Radio-Clinical models and found that the predictive accuracy of the Patho-Radiology models was more potent and stable among different DL-based algorithms (Figure 4) . The Patho-Radiology Signature showed the highest AUC of 0.933 and the Patho-Radio-Clinical Signature, based on MLP, yielded an AUC of 0.867 in 5-fold cross validation. The Patho-Clinical models achieved limited AUCs in all models, and it is inconclusive whether the clinical model, whether used alone or in combination with other modalities, can predict HRD status. Therefore, we suggest that both the Pathology Signature and the Radiology Signature could predict HRD status; however, the Patho-Radiology Signature yielded improved predictive accuracy, although direct comparison was not possible owing to differences in the test dataset. We propose that these three signatures could help predict HRD in clinical scenarios. We subsequently selected cases with data available in all modalities to investigate the correlation and possible interaction between the Pathology Signature and the Radiology Signature (Figure 4E) . The heatmap showed no significant correlation between the Pathology Signature and the Radiology Signature (Figure 4F) . The scatter plot showed that the predictive value of the Pathology Signature and the Radiology Signature were distributed differently among all cases (Figure 4G) . The Patho-Radiology Signature exhibited enhanced predictive power in terms of both AUC and ridge plot metrics. Molecular characterization of patients with actual and Pathology-predicted HRD The PAM50 molecular subtypes, genetic alterations, and RNA expression of HRR-associated genes in patients with actual HRD and Pathology Signature-predicted HRD are summarized (Figure 5) . For the PAM50 classification, the basal-like, Luminal A and luminal B subtypes were observed in 18.9%, 37.8% and 43.3% of cases with actual HRD and in 21.1%, 36.9%, and 42.0% of cases predicted as HRD by the Pathology Signature, respectively. There was no significant difference in the distribution of the different subtypes between the actual and predicted HRD patients (P= 0.95). The ETS fusion was the most prevalent genetic alteration, followed by the FOXA1 mutation, SPOP mutation and RB1 deletion. In patients with actual HRD, the ETS gene fusion, RB1 deletion, HDAC2 deletion, FOXA1 mutation, and ROS1 deletion were the most common alterations (37.8%, 26.7%, 21.1%, 18.9%, and 18.9%, respectively), whereas these alterations were observed in 30.7%, 18.7%, 16.7%, 24.7%, and 14.0%, predicted HRD. Both the actual and predicted HRD cases exhibited higher incidences of RB1 deletion (predicted HRD vs. predicted HRP, p<0.001; actual HRD vs. actual HRP, p<0.001), HDAC2 deletion (predicted HRD vs. predicted HRP, p=0.002; actual HRD vs. actual HRP, p<0.001), ROS1 deletion (predicted HRD vs. predicted HRP, p<0.001; actual HRD vs. actual HRP, p<0.001), and PMS2 mutations (predicted HRD vs. predicted HRP, p<0.001; actual HRD vs. actual HRP, p=0.003) when compared to their actual and predicted HRP counterparts. There was no significant difference in the frequency of genetic alterations between the predicted and actual patients with HRD (P=0.37). The predicted and actual HRD cases had similar RNA expression levels for all the genes. In conclusion, there was no significant difference in the molecular subtypes, genetic alterations, and RNA expression of HRR-associated genes in patients with actual and predicted HRD. Interpretability of the Pathology Signature To improve the interpretability of the Pathology Signature, the Grad-CAM method was applied to illustrate the pathological characteristics of the images at the patch level ( Figure 6 ). After viewing all the Grad-CAM figures, two pathologists discussed and summarized four HRD-associated pathological features, including high-grade tumors (enlarged nuclear-to-cytoplasmic ratio, vacuolated appearance of cells, presence of pathological mitotic figures, poor formation of glandular follicles arranged in sheets, and small nests), fibrosis of the stroma (increased proliferation of fibroblasts with significant collagen degeneration observed around high-grade tumor cells), inflammatory cell infiltration (intensive infiltration of lymphoplasmacytic cells, usually observed around the tumor), and cribriform structures (presence of irregular glandular spaces resembling a sieve or honeycomb pattern), which were possible predictors of the presence of HRD by pathologists. Furthermore, we confirmed that the Grad-CAM figures could focus on the region of higher likelihood in patches at different magnifications (e.g. ×12.5, ×20, ×50, ×100, and ×200; Supplementary Figure S3 ). To evaluate whether the features of the Pathology Signature could be understood and identified by pathologists, nine pathologists with different clinical experiences were included in the following analysis: specialized urologic pathologists (pathologists 1 to 3), attending general pathologists (pathologists 4 to 6), and junior general pathologists (pathologists 7 to 9). First, nine pathologists were trained by a senior pathologist to understand the HRD-associated pathological features using the selected Grad-CAM figures. Furthermore, the nine pathologists were asked to classify 20 WISs (10 WSIs with the highest Pathology Signature scores and 10 WSIs with the lowest Pathology Signature scores) into the HRD or HRP group independently. After four weeks, the pathologists were informed about the predicted Pathology Signature score of each WSI and were asked to reclassify these WSIs. While there was no significant difference in the AUC between specialized urologic and attending general pathologists (p = 0.59), a notable difference was observed when compared with junior general pathologists (p = 0.01). Specialized urologic pathologists tended to score “very certain” more than junior general pathologists (26/60 vs. 14/60, p=0.02). After being informed about the Pathology Signature score, most pathologists could achieve a higher AUC and accuracy. The perceived certainty of the pathologists' classifications (defined as “very certain”, “fairly certain”, “relatively uncertain”) were also improved after informed about the Pathology Signature score in the three group of pathologists. However, based on repeated measures ANOVA, only junior general pathologists showed a statistically significant improvement (p=0.04). In contrast, specialized urologic and attending general pathologists did not demonstrate significant changes (p=0.33 and p=0.93, respectively). Although preliminary, these results illustrated that the pathological features could be understood by pathologists and helped address the interpretability of the Pathology Signature. Association between HRD and biochemical recurrence (BCR) In patients with actual HRD, the 3-year BCR-free survival rate was 77.5%, while the 3-year BCR-free survival rate was 80.7% in patients with HRP (p=0.1969). The Kaplan-Meier curve showed that there was no significant difference in BCR-free survival between actual HRD and HRP. The HRD status predicted by the Clinical Signature could stratify patients’ risk of BCR in the training and test cohorts. However, the HRD status predicted by the Pathology Signature and Radiology Signature could not stratify the risk of BCR ( Supplementary Figure S4 ). DISCUSSION PARP inhibitors have been approved by the United States Food and Drug Administration (FDA) for the treatment of patients with advanced PCa and HRD. Thus, predicting HRD is crucial for the treatment of patients with advanced PCa; however, current prediction methods based on genetic testing are usually time-consuming, commonly inaccessible, and very expensive 20 . For instance, the FDA approved three assays for the detection of HRD in PCa (BRACAnalysis CDx testing blood 21 , FoundationOne CDx testing tissue 22 , and FoundationOne Liquid CDx testing plasma 23 ); but these tests are commonly inaccessible for patients in underdeveloped regions. Additionally, the cost of these assays is as high as $4,800 to $5,800 per test. Patients with HRD are relatively infrequently observed among PCa patients and require genetic tests of approximately ten patients to identify one patient with HRD, which means it costs $48,000 to $58,000 to identify one patient with HRD 24 . These logistic and financial challenges discouraged patients from undergoing genetic testing, resulting in missed opportunities for the accurate application of PARP inhibitors in PCa patients with HRD. Recently, AI has been integrated into cancer management. Recently, the FDA approved the application of Paige-AI for the diagnosis of prostate biopsy specimens 25 . In this study, we developed an AI-based HRD status prediction tool for patients with PCa using multimodal data including pathological, radiological, and clinical features. The Pathology Signature, Radiology Signature, and Patho-Radiology Signature could effectively predict the HRD status in PCa patients. Our findings suggest that the HRD status of PCa patients can be predicted by an AI-based algorithm using WSIs and MRI images. This prediction tool does not require additional examinations other than the clinical routine. Pathological slides and MRI images were routinely collected. This made it possible to collect data retrospectively for prediction. In addition, only a very limited amount of time is required to generate the predicted results, thereby saving waiting time using genetic tests. Clinically, we can apply this prediction tool as a pre-screening tool to select high-risk cases, thus reducing the overall cost. In single-modal prediction models, Pathology Signature and Radiology Signature showed high accuracy, whereas the Clinical Signature showed a limited predictive performance (AUC=0.6). This is in accordance with a previous study indicating low relevance between HRD and clinical features 26,27 . In contrast, Gerstung et al. illustrated that a series of genetic alterations are associated with pathological features 28 . The prediction of molecular alterations using radiological data has also been confirmed by recent studies on gliomas and non-small cell lung cancer 29,30 . In addition, there are reports on multimodal methods for genetic alteration prediction or treatment response prediction 27,31–33 . In this study, the Patho-Radiology Signature yielded a high predictive performance (AUC=0.933 in NavieBayes). The model integrated with pathological features and radiology features could achieve high predictive performance in achieving a higher AUC ( Figure 4C, 4D ) and a more distributed predicted risk ( Figure 4E ). It could be interpreted that the Pathology Signature and Radiology Signature were not strongly corelated ( Figure 4F) , but complementary to each other ( Figure 4G ). However, the predictive performance decreased after introducing clinical features into the Patho-Radiology model (CPR model AUC=0.850 in the NaiveBayes model and the highest AUC=0.867 in the MLP model). The possible reason for the decreased performance may be the low predictive accuracy of the Clinical Signature itself. In addition, the learning strategy or optimization methods of the model cannot effectively handle the newly added clinical modality. Owing to the limited number of cases with multimodal data, the actual reason for this could not be identified in this study. Future studies may consider prototypical modality rebalance 34 or modality balance networks 35 to improve the performance of multimodal models. In analyzing the pathological data, previous studies suggested that a weakly supervised approach, only label tumor or non-tumor by slide level, can yield satisfied results 36 . Initially, we attempted to apply weakly supervised methods; however, we found that the use of non-annotated slides failed to yield satisfactory results. We attributed the differences between our study and previous studies using weakly supervised methods to the nature of PCa, in which tumor heterogeneity played a significant role. Unlike tumors with distinct boundaries between the tumor and normal tissues, cancer, normal, and prostate stromal tissues are intertwined in PCa. The tumor proportion in one slide varied from 5% to 95% among different slides 37 . Without defining the tumor region, it was difficult for the model to capture the features of the tumor. Therefore, in this study, we employed annotation of the tumor region, defined as the area with the major proportion of tissue consisting of tumor cells and intratumoral stromal tissue. We did not strictly annotate every region of the adenocarcinoma at the cellular level because we believe that the stromal structures within the tumor might be relevant to molecular changes. In particular, the pathological grade of PCa is closely related to tissue structure rather than the characteristics of individual cancer cells, which is different from many other malignancies 38 . Annotating only the cancerous region, but not the stromal region of PCa, would result in the loss of important structural information. In contrast, for cancer types with pathology slides predominantly consisting of tumor tissue, weakly supervised methods can improve the prediction 39 . Our findings demonstrated that the supervised approach may be more applicable for predicting molecular alterations in prostate cancer. To enhance interpretability, a Grad-CAM plot was used to analyze the pathology images. Pathologists have identified features including high-grade tumors, fibrosis of the stroma, inflammatory cell infiltration, and cribriform structures. This is in accordance with a pre-print paper by Jakob et al., which identified that high grade, fibrosis, and lymphocytic infiltration were associated with HRD in different types of tumors 40 . However, hemorrhage was associated with HRD by Jakob et al.; however, this association was not confirmed in this study because there were very few hemorrhages in PCa. In addition, cribriform structures were associated with HRD in this study but not in a previous study. We suppose that this is because cribriform structures are more often observed in high-grade PCa than in other malignancies. To provide further interpretations of the Patho-Radiology Signature based on the Naïve Bayes algorithm, a Probability Density Function was introduced. We illustrated the mean and variance of the Gaussian distribution for each feature (Supplementary Figures S5-S6) . The top radiological features exhibiting pronounced differences between the two groups were predominantly associated with texture. This observed pattern may be associated with high-grade tumor and intra-tumor heterogeneity. Nevertheless, considering the methodology employed, which involved the leave-one-out selection of features and subsequent dimensionality reduction, an in-depth analysis and validation are imperative to derive definitive and unbiased conclusions. The limitations of this study include the limited MRI data in the TCGA-PRAD cohort, leading to difficulties in performing external validation of the radiological and multimodal models. This highlights one of the obstacles in current studies on PCa, namely multi-omics data rareness. In addition, the sample size of this study was relatively small. Finally, it is important to note that the modeling was conducted in a Caucasian-dominant cohort and validated in an Asian cohort, thereby limiting the assessment of ethnic disparities in the study. Future studies should include cohorts with diverse genetic backgrounds to validate the findings of this study. METHODS This study complied with relevant ethical regulations, and its protocols were approved by the Institutional Review Board of Shanghai Changhai Hospital (CHEC2022-151). This study was registered with the Chinese Clinical Trial Registry (ChiCTR2200064329). The requirement for informed consent was waived for this retrospective study, and the participants were not compensated for. Two cohorts of hormone-sensitive PCa patients were included in this study. The study design is illustrated in Figure 1 . Cohort description The TCGA-PRAD cohort consisted of 387 patients with PCa 41 , for which we downloaded diagnostic HE-stained WSIs from the Genomic Data Commons (GDC) portal (https://portal.gdc.cancer.gov/). The CPGEA cohort was our in-house cohort, consisting of 210 pathologically confirmed PCa patients sequenced and reported in a previous study 42 . WSIs and MRI images of patients in the CPGEA cohort were collected and analyzed for the first time. All clinicopathological information related to the medical history was deposited in a prospectively collected database. Patients in the two cohorts were eligible for this study if they had at least one of the following: (1) WSIs on H&E slides of radical prostatectomy. (2) multi-parametric prostate MRI or pelvic MRI before radical prostatectomy. Two patients in the TCGA-PRAD cohort and the CPGEA cohort, respectively. Study design For the analysis of pathology and clinical data, we selected the TCGA-PRAD cohort as the training cohort and the CPGEA cohort as the test cohort. However, for the MRI images, we randomly sampled 80% of all patients with MRI images for training, whereas the remaining patients were used as the test cohort. This sampling strategy was employed to ensure fair comparisons between unimodal and multimodal models, thereby avoiding any spurious differences in test concordance indices that may arise due to selective patient exclusion in some models but not in others (Supplementary Figure S1) . The proportion of data annotated by each pathologist and radiologist in each cohort and the different types of equipment was not significantly different between the two cohorts (Supplementary Figure S7) . Inferring genetic alteration status The HRD status was ascertained based on alterations in 14 homologous recombinant repair (HRR)-related genes, as previously demonstrated (including ATM, BRCA1, BRCA2, BARD1, BRIP1, CDK12, CHEK1, CHEK2, FANCL, PALB2, RAD51B, RAD51C, RAD51D, and RAD54L) 43 . Patients were assigned to the HRD subtype if they had at least one nonsynonymous mutation or deep deletion in the HRR-related genes. Radiology pipeline A flowchart of the radiology pipeline is shown in Figure 2 . Prostate MRI was performed on a 3.0T MR scanner with an abdominal phase array coil, following a 4h fasting period and enema treatment with glycerin (20 ml). Routine sequences included sagittal T2WI, axial high-resolution T2WI, and axial DWI. Supplementary Table S9 shows the axial T2WI and DWI parameters used for the machine learning. Each radiologist traced the outer contour of the prostate lesions on the tumor-containing axial section using Insight Segmentation and Registration Toolkit-SNAP v.3.8.0 software. Handcrafted features were extracted using Pyradiomics (https://pyradiomics.readthedocs.io/en/latest/). Pathology pipeline A flowchart of the pathology pipeline is shown in Figure 3 . All the WSIs were digitally captured using a 20× objective lens. The task of annotating the H&E WSIs to identify the tumor region was performed by two fellowship-trained uropathologists using the QuPath software (https://github.com/qupath). The WSIs were partitioned into 512×512-pixel patches. The ResNet50 algorithm was then used to predict the tumor region based on annotations by pathologists. The patch label and corresponding probability were extracted and applied to generate WSI-level features via Patch Likelihood Histogram (PLH) and bag-of-words (BoW) pipelines. WSI-level features were applied to construct pathological models for HRD status prediction. Multimodal integration An early fusion strategy, which involved normalizing numeric features and concatenating them with categorical features to create new integrated features was applied. Based on these selected and optimized features, we input them into a predictive pipeline composed of ten machine learning models, providing a prediction probability for each case. Visualization of Pathology Signature prediction with Gradient-weighted Class Activation Mapping ( Grad-CAM) Grad-CAM class localization maps are created by visualizing the gradients flowing into the final convolutional layer of the network immediately before the fully connected layers. Because the convolutional layers contain class-specific spatial information from the input image that is lost in the fully connected layers, this is the optimal point for map generation. Grad-CAM requires neither any modifications to the existing model architecture nor any retraining of the model. Survival analysis Linear Cox proportional hazards models with L2 regularization (c=0.5) and without L1 regularization were applied to all the multimodal and unimodal models. The homologous recombination potent (HRP) subtype was designated as high-risk (risk score 1), and patients assigned to the HRD subtype were designated as low-risk (risk score 0); no interaction terms were used. Conclusion We illustrated that the Pathology Signature, Radiology Signature, and Patho-Radiology Signature obtained high accuracy in predicting HRD status in PCa patients. The framework of this study can be applied to other malignancies and highlights the promising applications of AI for treatment selection in cancer management. In clinical settings, AI-based prediction tools can be applied as pre-screening tools to reduce the number of genetic analyses required. Declarations Data availability Data with identifiers in the CPEGA cohort will be made available to research partners upon reasonable request to the corresponding author R-C or the Shanghai Changhai Hospital, China, which is a data transfer agreement approved by the legal departments of the requesting researcher and all legal departments of the institutions that provided data for the study and an ethics clearance. All images and patient data from TCGA-PRAD cohort used in this study are publicly available at http://portal. gdc. cancer. gov/. Code availability The code for the model development is available online after the article has been published. Acknowledgments Funding: This study was supported by grants from the National Natural Science Foundation of China (NSFC) (grant number:82272905), Rising Star Program of Shanghai Science and Technology Commission (grant number:21QA1411500), and the Shanghai Action Plan for Technological Innovation Grant (No. 22ZR1478000) and Guhai Project of Shanghai Changhai Hospital. Author information: Conceptualization: R.C., G.X., and L.W. Methodology: Z.S., Q.Z., N.T., Y.Z., W.Z., and L.D. Investigation: Z.S., Q.Z., N.T., Y.Z., W.Z., L.D., Y.W., M.Q., Z.D., H.Y., Y.L., X.W., Y.Y., H.W., L.Z., Y.G., W.Z., and Y.L. Software: Z.S., Q.Z., N.T., Y.Z., W.Z., L.D., Y.W., M.Q., Z.D., and H.Y. Visualization: W.Z., L.D., Y.W., M.Q., Z.D., H.Y., H.W., L.Z., Y.G., W.Z., and Y.L. Supervision: R.C., G.X., and L.W. Writing—original draft: Z.S., Q.Z., N.T., Y.Z., W.Z., and L.D. Writing—review and editing: Y.W., M.Q., Z.D., H.Y., H.W., L.Z., Y.G., W.Z., Y.L., R.C., G.X., and L.W. Resources: Z.S., Q.Z., N.T., Y.Z., Y.L., X.W., and Y.Y. Validation: R.C., G.X., and L.W. Project Administration: R.C., G.X., and L.W. Funding Acquisition: R.C., G.X., and L.W. Competing interests: The authors declare no conflicts of interest. References Rl, S., Kd, M., Ns, W. & A, J. Cancer statistics, 2023. CA: a cancer journal for clinicians 73 , (2023). Mitchell, T. & Neal, D. E. The genomic evolution of human prostate cancer. Br J Cancer 113 , 193–198 (2015). Nguyen, L., W. M. Martens, J., Van Hoeck, A. & Cuppen, E. Pan-cancer landscape of homologous recombination deficiency. Nat Commun 11 , 5584 (2020). Hoeijmakers, J. H. Genome maintenance mechanisms for preventing cancer. Nature 411 , 366–374 (2001). Groelly, F. J., Fawkes, M., Dagg, R. A., Blackford, A. N. & Tarsounas, M. Targeting DNA damage response pathways in cancer. Nat Rev Cancer 23 , 78–94 (2023). ESMO 2019: PROfound: Phase 3 Study of Olaparib vs. Enzalutamide or Abiraterone for Metastatic Castration-Resistant Prostate Cancer with Homologous Recombination Repair Gene Alterations. https://www.urotoday.com/conference-highlights/esmo-2019/esmo-2019-prostate-cancer/115401-esmo-2019-profound-phase-3-study-of-olaparib-vs-enzalutamide-or-abiraterone-for-metastatic-castration-resistant-prostate-cancer-with-homologous-recombination-repair-gene-alterations.html. de Luca, X. M. et al. Using whole-genome sequencing data to derive the homologous recombination deficiency scores. npj Breast Cancer 6 , 1–8 (2020). Sztupinszki, Z. et al. Detection of Molecular Signatures of Homologous Recombination Deficiency in Prostate Cancer with or without BRCA1/2 Mutations. Clin Cancer Res 26 , 2673–2680 (2020). Coudray, N. & Tsirigos, A. Deep learning links histology, molecular signatures and prognosis in cancer. Nat Cancer 1 , 755–757 (2020). Kather, J. N. et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nat Med 25 , 1054–1056 (2019). Yamashita, R. et al. Deep learning model for the prediction of microsatellite instability in colorectal cancer: a diagnostic study. Lancet Oncol 22 , 132–141 (2021). Cifci, D., Foersch, S. & Kather, J. N. Artificial intelligence to identify genetic alterations in conventional histopathology. J Pathol 257 , 430–444 (2022). Dadhania, V. et al. Leveraging artificial intelligence to predict ERG gene fusion status in prostate cancer. BMC Cancer 22 , 494 (2022). Schaumberg, A., Rubin, M. & Fuchs, T. H&E-stained Whole Slide Image Deep Learning Predicts SPOP Mutation State in Prostate Cance. Preprint at https://doi.org/10.1101/064279 (2016). Houlahan, K. E. et al. Molecular hallmarks of multiparametric MRI visibility in prostate cancer. Eur Urol 76 , 18–23 (2019). Boehm, K. M. et al. Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer. Nat Cancer 3 , 723–733 (2022). Vanguri, R. S. et al. Multimodal integration of radiology, pathology and genomics for prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer. Nat Cancer 3 , 1151–1164 (2022). Fremond, S. et al. Interpretable deep learning model to predict the molecular classification of endometrial cancer from haematoxylin and eosin-stained whole-slide images: a combined analysis of the PORTEC randomised trials and clinical cohorts. Lancet Digit Health 5 , e71–e82 (2023). Li, J. et al. A genomic and epigenomic atlas of prostate cancer in Asian populations. Nature 580 , 93–99 (2020). Wu, K. et al. Evaluation of the Efficacy of PARP Inhibitors in Metastatic Castration-Resistant Prostate Cancer: A Systematic Review and Meta-Analysis. Front Pharmacol 12 , 777663 (2021). De Bono, J. S. et al. PROfound: A randomized Phase III trial evaluating olaparib in patients with metastatic castration-resistant prostate cancer and a deleterious homologous recombination DNA repair aberration. JCO 35 , TPS5091–TPS5091 (2017). Milbury, C. A. et al. Clinical and analytical validation of FoundationOne®CDx, a comprehensive genomic profiling assay for solid tumors. PLoS One 17 , e0264138 (2022). Woodhouse, R. et al. Clinical and analytical validation of FoundationOne Liquid CDx, a novel 324-Gene cfDNA-based comprehensive genomic profiling assay for cancers of solid tumor origin. PLoS One 15 , e0237802 (2020). Sztupinszki, Z. et al. Detection of Molecular Signatures of Homologous Recombination Deficiency in Prostate Cancer with or without BRCA1/2 Mutations. Clin Cancer Res 26 , 2673–2680 (2020). Commissioner, O. of the. FDA Authorizes Software that Can Help Identify Prostate Cancer. FDA https://www.fda.gov/news-events/press-announcements/fda-authorizes-software-can-help-identify-prostate-cancer (2021). Feng, C. et al. Relationship between homologous recombination deficiency and clinical features of breast cancer based on genomic scar score. Breast 69 , 392–400 (2023). Boehm, K. M. et al. Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer. Nat Cancer 3 , 723–733 (2022). Fu, Y. et al. Pan-cancer computational histopathology reveals mutations, tumor composition and prognosis. Nat Cancer 1 , 800–810 (2020). Kim, M. et al. Diffusion- and perfusion-weighted MRI radiomics model may predict isocitrate dehydrogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade glioma. Eur Radiol 30 , 2142–2151 (2020). Zhang, J. et al. Value of pre-therapy 18F-FDG PET/CT radiomics in predicting EGFR mutation status in patients with non-small cell lung cancer. Eur J Nucl Med Mol Imaging 47 , 1137–1146 (2020). Sammut, S.-J. et al. Multi-omic machine learning predictor of breast cancer therapy response. Nature 601 , 623–629 (2022). Vanguri, R. S. et al. Multimodal integration of radiology, pathology and genomics for prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer. Nat Cancer 3 , 1151–1164 (2022). Lipkova, J. Artificial intelligence for multimodal data integration in oncology. OPEN ACCESS . Fan, Y., Xu, W., Wang, H., Wang, J. & Guo, S. PMR: Prototypical Modal Rebalance for Multimodal Learning. Preprint at https://doi.org/10.48550/arXiv.2211.07089 (2022). Improving Multispectral Pedestrian Detection by Addressing Modality Imbalance Problems | SpringerLink. https://link.springer.com/chapter/10.1007/978-3-030-58523-5_46. Muti, H. S. et al. Development and validation of deep learning classifiers to detect Epstein-Barr virus and microsatellite instability status in gastric cancer: a retrospective multicentre cohort study. Lancet Digit Health 3 , e654–e664 (2021). Mohanty, S. K., Lobo, A. & Cheng, L. The 2022 revision of the World Health Organization classification of tumors of the urinary system and male genital organs: advances and challenges. Hum Pathol 136 , 123–143 (2023). van Leenders, G. J. L. H. et al. The 2019 International Society of Urological Pathology (ISUP) Consensus Conference on Grading of Prostatic Carcinoma. Am J Surg Pathol 44 , e87–e99 (2020). Bilal, M. et al. Development and validation of a weakly supervised deep learning framework to predict the status of molecular pathways and key mutations in colorectal cancer from routine histology images: a retrospective study. Lancet Digit Health 3 , e763–e772 (2021). Lavinia Loeffler, C. M. et al. Direct prediction of Homologous Recombination Deficiency from routine histology in ten different tumor types with attention-based Multiple Instance Learning: a development and validation study. medRxiv 2023.03.08.23286975 (2023) doi:10.1101/2023.03.08.23286975. Cancer Genome Atlas Research Network. The Molecular Taxonomy of Primary Prostate Cancer. Cell 163 , 1011–1025 (2015). Li, J. et al. A genomic and epigenomic atlas of prostate cancer in Asian populations. Nature 580 , 93–99 (2020). Clarke, N. et al. Olaparib combined with abiraterone in patients with metastatic castration-resistant prostate cancer: a randomised, double-blind, placebo-controlled, phase 2 trial. Lancet Oncol 19 , 975–986 (2018). Additional Declarations No competing interests reported. Supplementary Files SupplementaryTableS1.docx SupplementaryTableS2.xlsx SupplementaryTableS3.xlsx SupplementaryTableS4.docx SupplementaryTableS5.docx SupplementaryTableS6.xlsx SupplementaryTableS7.xlsx SupplementaryTableS8.docx SupplementaryTableS9.docx SupplementaryFigure1.pdf SupplementaryFigure2.pdf SupplementaryFigure3.pdf SupplementaryFigure4.pdf SupplementaryFigure5.pdf SupplementaryFigure6.pdf SupplementaryFigure7.pdf Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6488233","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":448900889,"identity":"d057d242-6014-4d2a-9091-242ce537bbc1","order_by":0,"name":"Zijian Song","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zijian","middleName":"","lastName":"Song","suffix":""},{"id":448900890,"identity":"e73cd4e7-babd-4d13-a204-8d7495aa6033","order_by":1,"name":"Qianwen Zhang","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Qianwen","middleName":"","lastName":"Zhang","suffix":""},{"id":448900891,"identity":"e4539c98-ef2d-4f28-bf99-23ca67db95d8","order_by":2,"name":"Wenhui Zhang","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Wenhui","middleName":"","lastName":"Zhang","suffix":""},{"id":448900892,"identity":"caa25e07-71fa-4e83-a415-9b647fc8cda7","order_by":3,"name":"Na Ta","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Na","middleName":"","lastName":"Ta","suffix":""},{"id":448900893,"identity":"79dc59d4-86e4-448b-a349-bbe9d1d4d9dd","order_by":4,"name":"Yan Zhu","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"Zhu","suffix":""},{"id":448900894,"identity":"eeddc2e5-43ae-49f5-a38f-6f7b5524ce46","order_by":5,"name":"Longxin Deng","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Longxin","middleName":"","lastName":"Deng","suffix":""},{"id":448900895,"identity":"08161a4e-d8dd-411f-9bd3-061eb9894c5b","order_by":6,"name":"Han Wu","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Han","middleName":"","lastName":"Wu","suffix":""},{"id":448900896,"identity":"226d06c4-f882-4b8c-8f11-a076b909a067","order_by":7,"name":"Lingxuan Zhu","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Lingxuan","middleName":"","lastName":"Zhu","suffix":""},{"id":448900897,"identity":"2450579d-f1b2-492e-b0e9-6a198d32e9dc","order_by":8,"name":"Yanan Gao","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yanan","middleName":"","lastName":"Gao","suffix":""},{"id":448900898,"identity":"ed16d7cb-9d03-413f-85fd-73663eebd649","order_by":9,"name":"Yaoming Li","email":"","orcid":"","institution":"Army Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yaoming","middleName":"","lastName":"Li","suffix":""},{"id":448900901,"identity":"cd3508ac-33a5-49c1-9b62-e80942a259e8","order_by":10,"name":"Xiao Wang","email":"","orcid":"","institution":"Zhejiang University","correspondingAuthor":false,"prefix":"","firstName":"Xiao","middleName":"","lastName":"Wang","suffix":""},{"id":448900904,"identity":"59dde99d-f52e-4c6d-ae1d-9a0b85715f14","order_by":11,"name":"Yang Yang","email":"","orcid":"","institution":"Nanjing University School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Yang","middleName":"","lastName":"Yang","suffix":""},{"id":448900908,"identity":"17d20361-cbf7-46a0-85ff-de0b5a0bbb43","order_by":12,"name":"Wei Zhang","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Wei","middleName":"","lastName":"Zhang","suffix":""},{"id":448900910,"identity":"3bd49e6b-b400-432d-a525-ffd14ee89fd2","order_by":13,"name":"Yancheng Lai","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yancheng","middleName":"","lastName":"Lai","suffix":""},{"id":448900912,"identity":"621155d5-f278-42e3-b87a-810c139060b4","order_by":14,"name":"Yan Wang","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Yan","middleName":"","lastName":"Wang","suffix":""},{"id":448900916,"identity":"72283d6e-4d8a-46c7-9b9f-dbec69762687","order_by":15,"name":"Min Qu","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Min","middleName":"","lastName":"Qu","suffix":""},{"id":448900917,"identity":"7fd20898-a022-4f85-aa54-8afffdc99da4","order_by":16,"name":"Zhenyang Dong","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Zhenyang","middleName":"","lastName":"Dong","suffix":""},{"id":448900918,"identity":"872f08ff-6901-44aa-aaa3-605ee698fd01","order_by":17,"name":"Huamao Ye","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Huamao","middleName":"","lastName":"Ye","suffix":""},{"id":448900921,"identity":"e3f66a71-c11a-4add-9298-1d9615043cd6","order_by":18,"name":"Linhui Wang","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Linhui","middleName":"","lastName":"Wang","suffix":""},{"id":448900924,"identity":"389b1235-7f92-4119-a053-6c76c655e765","order_by":19,"name":"Gao Xu","email":"","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":false,"prefix":"","firstName":"Gao","middleName":"","lastName":"Xu","suffix":""},{"id":448900925,"identity":"1cd39e75-7de6-47c4-9a62-48d0afa4ed8a","order_by":20,"name":"Rui Chen","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA0ElEQVRIiWNgGAWjYDCCAwxsYJqNvfEBiObhI14Lz2EDsBY2orUwSCQbQPQS0sF3+/CzBx933Evsk3zM+Jg3x06GjYH54aMbeLRInkszN5x5pjixTTqZ2XDmtmSgw9iMjXPwaDE4w8MmzduWANSSf0zi4zZmoBagCEEtf0FaJA+zSSRuqydSCyNIiwQzG9CWw4S1SJ5hM5PsbUswbuMB++U4DxszAb/wnWF+JvGzLUF2fvthYIhtq7bnZ29++BifFiyAmTTlo2AUjIJRMAqwAAAXiT/QrgGg0gAAAABJRU5ErkJggg==","orcid":"","institution":"Shanghai Changhai Hospital, Second Military Medical University","correspondingAuthor":true,"prefix":"","firstName":"Rui","middleName":"","lastName":"Chen","suffix":""}],"badges":[],"createdAt":"2025-04-20 08:38:07","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6488233/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6488233/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":82147323,"identity":"42077e3e-9cf0-4157-ac2d-df6b0effa82b","added_by":"auto","created_at":"2025-05-07 07:00:26","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":438356,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSchematic outline of the study and\u003c/strong\u003e \u003cstrong\u003emultimodal cohort characteristics.\u003c/strong\u003e (A) Work-flow of this study. The multimodal models incorporate clinical, genomic, pathological, and radiology features for predicting homologous recombination deficiency (HRD) status in prostate cancer patients. (B) The Venn diagram depicts the distribution of patients across clinical, genomic, pathological, and radiomic categories. A multimodal cohort heatmap displays clinical, genomic, pathological, and radiomic features for the patients in (C) the TCGA-PRAD cohort and (D) the CPGEA cohort. The stacked bar chart reveals the type of genetic mutation and the presence of corresponding clinical, genomic, pathological, and radiomic features in each patient. HRD=homologous recombination deficiency. TCGA=The Cancer Genome Atlas Prostate Cancer. CPGEA= Chinese Prostate Cancer Genetic and Epigenetic Atlas. BCR=Biochemical Recurrence. ISUP=International Society of Urological Pathology.\u003c/p\u003e","description":"","filename":"Figure100.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/09dc9619194a26f3cee4854b.jpg"},{"id":82147349,"identity":"92846c0b-cede-4cea-bde9-0903aa8b3f64","added_by":"auto","created_at":"2025-05-07 07:00:27","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":308130,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEstablishment and validation of Radiology Signature. \u003c/strong\u003e(A) Radiology pipeline across image pre-processing, feature generation, and model establishment. (B) Pre-processing of the MRI images. (C) Feature generation and selection. Number, ratio, and p-value of handcrafted features generated by the Pyradiomics were summarized. Nine features of nonzero coefficients were selected by the least absolute shrinkage and selection operator (LASSO) logistic regression model to establish the radiology models. (D) Heatmap indicating the association among the nine selected radiology features. (E) The AUC of different radiology models in 5-fold cross-validation. (F) The performance of the Radiology models in the test cohort. (G) The waterfall plot of the Radiology Signature based on SVM indicting predicted Radiology Signature score and HRD status. AUC=area under the curve. SVM=Support Vector Machine. LASSO=least absolute shrinkage and selection operator. HRD=Homologous Recombination Deficiency.\u003c/p\u003e","description":"","filename":"Figure200.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/bc3051102bdc17e3cfa80fdf.jpg"},{"id":82150480,"identity":"9f8d6352-3010-4c32-84ef-f65f22906763","added_by":"auto","created_at":"2025-05-07 07:16:26","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":379957,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEstablishment and validation of the Pathology Signature.\u003c/strong\u003e (A) Flow-chart of pathology pipeline, including image pre-processing, feature generation, and model establishment. (B) Patch-level tumor region prediction was performed based model established with ResNet50. Examples from the TCGA-PRAD and the CPGEA cohort were illustrated, indicating high coherence of actual and predicted tumor regions. (C) The confusion matrix plot of the ResNet50 model in differentiating tumor and non-tumor patches in the training and test cohort. (D) The ROC plot of the ResNet50 model in differentiating tumor and non-tumor patches in the training and test cohort. (E) The performance of the pathological models in the test cohort. Pathology Signature was established based on LightGBM. (F) The waterfall plot depicted the association between the Pathology Signature score and the HRD status in the training and test cohort. (G) Distribution of the Pathology Signature score and the HRD status in the training and test cohort. HRD=Homologous Recombination Deficiency. TCGA=The Cancer Genome Atlas Prostate Cancer. CPGEA= Chinese Prostate Cancer Genetic and Epigenetic Atlas.\u003c/p\u003e","description":"","filename":"Figure300.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/7568d5f9f9fb736aa9618632.jpg"},{"id":82148746,"identity":"f92b9a42-fb76-4c92-b3ed-33179a2f0fdd","added_by":"auto","created_at":"2025-05-07 07:08:27","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":308105,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eThe predictive performance of multimodal prediction models and the association between the Pathology Signature and the Radiology Signature. \u003c/strong\u003eThe ROC curve of the (A) Patho-Radiology Signature, (B) Patho-Clinical Signature, (C) Patho-Radio-Clinical Signature in prediction of HRD status in the test cohorts. (D) The AUC of different multimodal models based on different algorithms. (E) The distribution of predicted probability by induvial and combined signature scores. The x-axis indicated the different signatures. The y-axis indicated the count of cases according to the predicted probabilities. (F) Spearman’s rank correlation coefficient of the risk quantile across pairs of the individual modalities, indicating low mutual ordering information between individual modalities. (G) Radiology Signature score and Pathology Signature score of individual patients and their predicted HRD status by Pathology Signature only (purple), Radiology Signature only (yellow), both modalities (pink) and none modalities (grey). (H) Radiology Signature score and Pathology Signature score of individual patients and their multimodal prediction results: correct predicted HRD (pink), correct predicted HRP (purple), wrongly predicted HRD (pink with a cross) and wrongly predicted HRP (purple with a cross). ROC=receiver operating characteristic. AUC=area under the curve. HRD= Homologous Recombination Deficiency. HRP= Homologous Recombination Potent.\u003c/p\u003e","description":"","filename":"Figure400.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/67244b3bce52f7e180ca2e32.jpg"},{"id":82148716,"identity":"61d39eda-dc38-4b5d-9769-a99b59f10855","added_by":"auto","created_at":"2025-05-07 07:08:27","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":235563,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMolecular Characterization of patients with actual and Pathology Signature-predicted HRD Status. \u003c/strong\u003eMolecular subtypes for both actual and predicted HRD and HRP cases are exhibited in the overall cohort (A), CPEGA cohort (C), and TCGA-PRAD cohort (F). Genetic alterations for these cases are depicted in the overall cohort (B), CPEGA (D), and TCGA-PRAD (G). RNA expression for both actual and predicted HRD and HRP cases, is showcased in CPEGA cohort (E) and TCGA-PRAD cohort (H).\u003c/p\u003e","description":"","filename":"Figure500.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/db9fd235301d867e41a6b8e0.jpg"},{"id":82147327,"identity":"66824e8d-67a4-4370-9b69-bb90bc49cddb","added_by":"auto","created_at":"2025-05-07 07:00:26","extension":"jpg","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":323708,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisualization and interpretability of the Pathology Signature. \u003c/strong\u003e(A) Four pathological features including high-grade tumors, fibrosis of the stroma, inflammatory cell infiltration, and cribriform structures were identified by pathologists to be associated with HRD status in Gradient-weighted Class Activation Mapping (Grad-CAM). (B) The predictive performance of the Pathology Signature compared with 9 pathologists with distinctive clinical experience in grouping 20 cases of patients with highest or lowest Pathology Signature scores to HRD or HRP group. (D) The predictive accuracy rate of pathologists with different clinical experience with or without knowing the Pathology Signature score of each case. (E) The perceived certainty of pathologists’ classification, classified by “very certain=3”, “fairly certain=2”, and “relatively uncertain=1”. Grad-CAM, Gradient-weighted Class Activation Map. Error bars represent standard deviation.\u003c/p\u003e","description":"","filename":"Figure600.jpg","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/3f3a3b275b30576baf053b0f.jpg"},{"id":88188754,"identity":"cd1651ff-14e5-49bf-accf-5837dfa1c95c","added_by":"auto","created_at":"2025-08-03 12:46:43","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3537954,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/26ceee7a-c04e-48db-8f40-f3a20f577e7b.pdf"},{"id":82148707,"identity":"77ce7dcd-8039-4a9b-bb79-8471f6d5d867","added_by":"auto","created_at":"2025-05-07 07:08:26","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":13590,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS1.docx","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/4c761d62b3a25af41b777c62.docx"},{"id":82147331,"identity":"1e21b842-c6b0-4d43-b85c-028ebdb27303","added_by":"auto","created_at":"2025-05-07 07:00:26","extension":"xlsx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":2862638,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/2fb6710a4ebca7afb870bd88.xlsx"},{"id":82147332,"identity":"065f5215-fc49-4a83-ada6-0e03a8c9394e","added_by":"auto","created_at":"2025-05-07 07:00:26","extension":"xlsx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":44486,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS3.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/abb1315e5548b976233479fb.xlsx"},{"id":82148713,"identity":"62355f6f-c231-4313-8690-fdb6332eba4e","added_by":"auto","created_at":"2025-05-07 07:08:26","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":17662,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS4.docx","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/453c92ac28dfcac15dfb91f7.docx"},{"id":82148715,"identity":"7464179a-b7a9-4702-a729-08c09f23baa4","added_by":"auto","created_at":"2025-05-07 07:08:26","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":12871,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS5.docx","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/dcb93914c85a24a2d28a9821.docx"},{"id":82147351,"identity":"34edaba6-702b-4c88-b2d1-c9a7c072dd2f","added_by":"auto","created_at":"2025-05-07 07:00:27","extension":"xlsx","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":689711,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS6.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/5012d64327eb8247f08ab64c.xlsx"},{"id":82148721,"identity":"9db7865f-0418-40ca-bc3c-424f25b735bc","added_by":"auto","created_at":"2025-05-07 07:08:27","extension":"xlsx","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":272951,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS7.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/ff250754def68ddfd9dc3419.xlsx"},{"id":82147347,"identity":"6e448f0b-8461-4cb9-bd60-a6e6fac0e132","added_by":"auto","created_at":"2025-05-07 07:00:27","extension":"docx","order_by":8,"title":"","display":"","copyAsset":false,"role":"supplement","size":16803,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS8.docx","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/95226eb0dc4ccd31d999614a.docx"},{"id":82147338,"identity":"e86e7c37-e6da-46db-a649-0aaf07adb221","added_by":"auto","created_at":"2025-05-07 07:00:26","extension":"docx","order_by":9,"title":"","display":"","copyAsset":false,"role":"supplement","size":17387,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS9.docx","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/3082204117d6b69a32ed5e6b.docx"},{"id":82148743,"identity":"82cb148b-deb1-4381-aa8b-15f8c3680208","added_by":"auto","created_at":"2025-05-07 07:08:27","extension":"pdf","order_by":10,"title":"","display":"","copyAsset":false,"role":"supplement","size":421500,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/ec666679ba39e76713ce1f51.pdf"},{"id":82148749,"identity":"e7a96fad-b8b7-4f87-8559-840596264910","added_by":"auto","created_at":"2025-05-07 07:08:27","extension":"pdf","order_by":11,"title":"","display":"","copyAsset":false,"role":"supplement","size":420924,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure2.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/ea7d919a0a798eda5cf841a9.pdf"},{"id":82147342,"identity":"71467774-4383-458b-831f-30e57ac4f87b","added_by":"auto","created_at":"2025-05-07 07:00:27","extension":"pdf","order_by":12,"title":"","display":"","copyAsset":false,"role":"supplement","size":639671,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/97de2cca357cc0abfd06e787.pdf"},{"id":82147405,"identity":"1c6ea935-f3b6-461e-8382-af8429d89837","added_by":"auto","created_at":"2025-05-07 07:00:30","extension":"pdf","order_by":13,"title":"","display":"","copyAsset":false,"role":"supplement","size":47851522,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure4.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/c673419cf8f3cbc30d23a227.pdf"},{"id":82148755,"identity":"2bbd3936-357f-462c-9e00-c740ff945772","added_by":"auto","created_at":"2025-05-07 07:08:28","extension":"pdf","order_by":14,"title":"","display":"","copyAsset":false,"role":"supplement","size":944294,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure5.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/3077593cf7fb002f025838bd.pdf"},{"id":82148720,"identity":"d14e7ed7-76e6-4ba1-a960-7c2a95578743","added_by":"auto","created_at":"2025-05-07 07:08:27","extension":"pdf","order_by":15,"title":"","display":"","copyAsset":false,"role":"supplement","size":1193891,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure6.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/f19f35af487f405b19554aee.pdf"},{"id":82148739,"identity":"6dd74377-5770-40d5-89a8-deb94fb62dde","added_by":"auto","created_at":"2025-05-07 07:08:27","extension":"pdf","order_by":16,"title":"","display":"","copyAsset":false,"role":"supplement","size":636609,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigure7.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6488233/v1/98b6ced9b0cb99566e4ab73e.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Multimodal deep learning model for predicting homologous recombination deficiency in prostate cancer: an international multi-cohort study","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eProstate cancer (PCa) is the most commonly diagnosed malignancy and the fifth leading cause of cancer-related death in men worldwide\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. PCa is a malignancy that is characterized by diverse molecular alterations\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Homologous recombination deficiency (HRD) is a frequently observed molecular alteration in PCa\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. HRD leads to defective DNA break repair, increased somatic copy number alterations, genomic instability, and oncogenesis[4,5]. HRD has emerged as a potent biomarker for selecting patients for Poly(ADP-Ribose)-polymerase (PARP) inhibitor (PARPi) treatment of various cancers, including advanced-stage PCa, as demonstrated in the PROfound study\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. However, HRD detection typically relies on genetic sequencing, which is associated with limited accessibility, high cost, and a long detection period\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. Considering that the prevalence of HRD is approximately 5\u0026ndash;10% in primary PCa\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e and approximately 13% in advanced-stage PCa, ten genetic tests are required to identify one patient with HRD\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. Therefore, there is an urgent need for a rapid, low-cost, and accessible method to detect HRD status or, at the very least, prescreening patients with a higher risk of HRD status.\u003c/p\u003e \u003cp\u003eArtificial intelligence (AI), particularly deep learning (DL), has shown great promise in identifying pathological subtypes, molecular alterations, and prognostic outcomes in cancer patients\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. Prediction of Microsatellite Instability (MSI) in gastrointestinal cancer is one of the most investigated applications. Kather et al. proposed the application of deep learning in the analysis of conventional Hematoxylin \u0026amp; Eosin (H\u0026amp;E) histology to predict microsatellite instability based on multiple international cohorts\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Yamashita et al. proposed DL models that surpassed the performance of experienced gastrointestinal pathologists in predicting MSI on H\u0026amp;E-stained whole-slide images (WSIs) in colorectal cancer\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eHowever, despite the progress made in utilizing deep learning methods for predicting molecular alterations in various cancer types, a comprehensive review highlighted the limited exploration in this field in the context of PCa\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Chinnaiyan et al. reported the application of DL in the analysis of conventional pathological slides to predict the presence of ETS-related gene (ERG) fusions in a cohort of 392 cases\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Mark et al. reported the application of DL-based methods for speckle-type POZ protein (SPOP) mutation prediction in 177 PCa patients, but the study was not officially published\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. The main limitation of these studies may be the small sample size due to poor data accessibility, as there were a limited number of PCa cases with available pathology images and molecular alteration data. Furthermore, studies aimed at predicting molecular alterations in PCa using Magnetic Resonance Imaging (MRI) data are still in the preliminary stages and involve a limited number of patients\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eRecently, multimodal data were analyzed using DL algorithms in the risk stratification of high-grade serous ovarian cancer\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, in the prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e, and in the prediction of molecular classification of endometrial cancer\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. In a previous study, we presented the Chinese Prostate Cancer Genetic and Epigenetic Atlas (CPGEA) cohort, which is one of the largest multi-omics datasets of PCa, encompassing genomic, transcriptome, and epigenetic data\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. In this study, we for the first time, collected and analyzed the pathological and imaging data of patients in the CPGEA cohort. By combining internal cross-validation and external validations, we intend to establish and validate multimodal models for predicting the HRD status in patients with PCa.\u003c/p\u003e"},{"header":"RESULTS","content":"\u003cp\u003e\u003cstrong\u003eCohorts and characteristics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAfter data pre-processing, 387 patients from the TCGA-PRAD cohort and 179 patients from the CPGEA cohort were included in this study. The CPGEA cohort was sequenced and reported in 2020 by our team, and we collected and integrated the pathological, radiological, and genomic data for the first time. Patient characteristics are summarized in \u003cstrong\u003eTable 1\u003c/strong\u003e.\u0026nbsp;Clinical-genomic data were available for all cases, whereas pathological data were available for 157 and 386 cases in the CPGEA and TCGA-PRAD cohorts, respectively. Radiological data were available for 109 cases in the CPGEA cohort and eight cases in TCGA-PRAD (two cases with only T2 images were excluded from further analysis).\u0026nbsp;HRD was detected in 36 (20.7%) and 73 (18.9%) patients in the CPGEA and TCGA-PRAD cohorts, respectively (\u003cstrong\u003eFigure 1\u003c/strong\u003e). The distribution of cases across training and testing sets for different modalities, as well as their distribution in the database, can be found in \u003cstrong\u003eSupplementary Figure S1.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. Patient characteristics.\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"107%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" style=\"width: 31px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCPGEA\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" style=\"width: 27px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTCGA-PRAD\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"3\" style=\"width: 27px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHRP\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(N=143)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHRD\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(N=36)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eP value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHRP\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(N=314)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHRD\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(N=73)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eP value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHRP\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(N=457)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHRD\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e(N=109)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eP value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eAGE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.974\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.092\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e0.134\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eMean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e68.8 (6.24)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e68.8 (5.30)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e60.4 (6.82)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e62.0 (7.25)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e63.0 (7.69)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e64.2 (7.38)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003ePSA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.260\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.483\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e0.493\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eMedian(IQR)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e18.9 [11.3, 37.6]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e23.1 [13.3, 35.1]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e7.2 [5.0, 10.1]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e7.9 [5.2, 12.0]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e8.5 [5.6, 16.8]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e10.7 [5.8, 21.9]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eMissing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e7 (2.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e3 (4.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e7 (1.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e3 (2.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003epT\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.258\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.694\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e0.459\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eT2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e65 (45.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e19 (52.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e127 (40.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e26 (35.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e192 (42.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e45 (41.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eT3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e73 (51.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e14 (38.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e176 (56.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e44 (60.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e249 (54.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e58 (53.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eT4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e5 (3.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e3 (8.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e2 (2.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e11 (2.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e5 (4.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eTx\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e5 (1.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e1 (1.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e5 (1.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e1 (0.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003epN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e1.000\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.544\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e0.471\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eN0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e104 (72.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e25 (69.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e217 (69.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e52 (71.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e321 (70.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e77 (70.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eN1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e18 (12.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e5 (13.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e41 (13.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e13 (17.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e59 (12.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e18 (16.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eNx\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e21 (14.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e6 (16.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e56 (17.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e8 (11.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e77 (16.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e14 (12.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003epM\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.097\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.477\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e0.192\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eM0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e84 (58.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e13 (36.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e146 (46.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e38 (52.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e230 (50.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e51 (46.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eM1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e16 (11.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e7 (19.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e1 (1.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e16 (3.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e8 (7.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eMx\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e43 (30.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e16 (44.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e168 (53.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e34 (46.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e211 (46.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e50 (45.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eResidual_T\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.324\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e1.000\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e0.825\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eYES\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e57 (39.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e18 (50.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e4 (1.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e1 (1.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e61 (13.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e19 (17.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eNO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e84 (58.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e17 (47.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e290 (92.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e67 (91.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e374 (81.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e84 (77.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eMissing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e2 (1.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e1 (2.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e20 (6.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e5 (6.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e22 (4.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e6 (5.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eISUP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.590\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.074\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e0.051\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e8 (5.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e2 (5.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e35 (11.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e4 (5.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e43 (9.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e6 (5.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e41 (28.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e7 (19.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e102 (32.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e15 (20.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e143 (31.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e22 (20.2%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e28 (19.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e5 (13.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e65 (20.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e22 (30.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e93 (20.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e27 (24.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e24 (16.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e6 (16.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e38 (12.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e9 (12.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e62 (13.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e15 (13.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e42 (29.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e15 (41.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e74 (23.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e23 (31.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e116 (25.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e38 (34.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eMissing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e1 (2.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e1 (0.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eBCR at 3 years\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e1.000\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e0.211\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e0.197\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eYES\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e98 (68.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e25 (69.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e256 (81.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e63 (86.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e354 (77.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e88 (80.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eNO\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e45 (31.5%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e11 (30.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e24 (7.6%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e2 (2.7%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e69 (15.1%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e13 (11.9%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 13px;\"\u003e\n \u003cp\u003eMissing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 12px;\"\u003e\n \u003cp\u003e0 (0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e34 (10.8%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 9px;\"\u003e\n \u003cp\u003e8 (11.0%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 7px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e34 (7.4%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 10px;\"\u003e\n \u003cp\u003e8 (7.3%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 6px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eAbbreviations: SD=standard deviation; IQR=interquartile range; HRP=homologous recombination proficient; HRD=homologous recombination deficient; pT=pathological T stage; pN=pathological N stage; pM=pathological M stage; Residual_T=residual tumor presence; ISUP=International Society of Urological Pathology grade; BCR=biochemical recurrence.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eClinical Signature\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eClinical variables were not significantly associated with the HRD status (\u003cstrong\u003eSupplementary Figure S2\u003c/strong\u003e). Although age and International Society of Urological Pathology (ISUP) grade may be associated with HRD, the predictive value of clinical variables is limited. In the TCGA-PRAD cohort, clinical variables such as age, PSA level, pathological TNM stage, residual tumor, and ISUP grade were not significantly associated with HRD status in either the univariate or multivariate logistic regression analyses. We constructed the Clinical Signature based on a K-Nearest Neighbor (KNN) model with data from the TCGA-PRAD cohort and validated it in the CPGEA cohort, yielding an area under the curve (AUC) of 0.563. The results indicated a moderate association between Clinical Signature and HRD status.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRadiology\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eSignature\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOwing to the limitation of available MRI data in the TCGA-PRAD cohort, we combined TCGA-PRAD and CPGEA cohorts and applied 5-fold cross-validation to establish and validate the Radiology Signature (\u003cstrong\u003eSupplementary Figure S1\u003c/strong\u003e). We extracted shape features, first-order features, and texture category radiomic features that contained 1132 handcrafted features using Pyradiomics (\u003cstrong\u003eSupplementary Table S1\u003c/strong\u003e). A total of 2264 features were included after concatenating the DWI and T2WI features. Nine features were selected based on the Least Absolute Selection and Shrinkage Operator (LASSO) regression feature-selection procedure (\u003cstrong\u003eSupplementary Table S2\u003c/strong\u003e). Correlations among the features were explored using Spearman analysis and hierarchical clustering of the nine features. The Radiology Signature was established using the best-performing algorithm in the test cohort, the Support Vector Machine (SVM), which achieved an AUC of 0.833 (\u003cstrong\u003eSupplementary Table S3\u003c/strong\u003e). Some linear models achieved higher results than nonlinear models, which may be related to the interpretability and physical meaning of handcrafted imaging features. Waterfall plots depicted the distribution of the HRD status in Radiology Signature (\u003cstrong\u003eFigure 2\u003c/strong\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePathology Signature\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFirst, we divided the WSI into 512\u0026times;512 patches, and only patches with an overlap of over 80% with the ROI were used for the subsequent analysis. Patch-level tumor-region prediction was performed based on the ResNet50 model. The ResNet50 model demonstrated high efficacy in distinguishing tumor regions from non-tumor regions, with an AUC of 0.936 (95%CI 0.935-0.936) in the test cohort \u003cstrong\u003e(Figure 3, Supplementary Table S4)\u003c/strong\u003e. Subsequently,\u0026nbsp;the patch labels and corresponding probabilities\u0026nbsp;were applied to generate 106 pathological WSI-level features\u0026nbsp;using\u0026nbsp;the PLH and BoW pipelines (\u003cstrong\u003eSupplementary Table S5\u003c/strong\u003e). Then\u0026nbsp;WSI-level features were selected\u0026nbsp;by Lasso regression. Finally, 42 features were\u0026nbsp;applied for model construction by a series of algorithms (\u003cstrong\u003eSupplementary Table S6\u003c/strong\u003e), including Logistic Regression (LR), Support Vector Machine (SVM), Random Forest, XGBoost,\u0026nbsp;LightGBM, to identify the HRD status of each patch. The LightGBM model (Pathology Signature) achieved an AUC of 0.815 in external validation. Applying the cutoff value of 0.395,\u0026nbsp;we could detect the 100% (72/72) and 83.3% (15/18) HRD cases in the training and test cohorts, while reducing 77.6% (243/313) and 66.4% (83/125) genetic tests in the training and test cohorts, respectively\u0026nbsp;(\u003cstrong\u003eFigure 3 and Supplementary Table S7\u003c/strong\u003e). In the eleven patients received neoadjuvant ADT in the test cohort (all actual HRP), four cases were predicted as HRD and seven cases were predicted as HRP. Treatment-associated pathological changes may be associated with features of HRD.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMultimodal Signatures\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMultimodal prediction is commonly believed to enhance the prediction accuracy. When combining radiology and pathological features to build the Patho-Radiology model, the predictive accuracy improved compared with the performance of the single-modality models, with an AUC of 0.933 in the Na\u0026iuml;ve Bayes\u0026nbsp;algorithm model, 0.800 in the logistic regression model, 0.833 in the KNN model, and 0.867 in the LightGBM model, respectively (\u003cstrong\u003eFigure 4, Supplementary Table S8)\u003c/strong\u003e. The model based on the Na\u0026iuml;ve Bayes algorithm was selected as the Patho-Radiology Signature.\u003c/p\u003e\n\u003cp\u003eIn addition, we compared the predictive performance of the Patho-Clinical, Patho-Radiology, and Patho-Radio-Clinical models and found that the predictive accuracy of the Patho-Radiology models was more potent and stable among different DL-based algorithms \u003cstrong\u003e(Figure 4)\u003c/strong\u003e. The Patho-Radiology Signature showed the highest AUC of 0.933 and the Patho-Radio-Clinical Signature, based on MLP, yielded an AUC of 0.867 in 5-fold cross validation. The Patho-Clinical models achieved limited AUCs in all models, and it is inconclusive whether the clinical model, whether used alone or in combination with other modalities, can predict HRD status.\u003c/p\u003e\n\u003cp\u003eTherefore, we suggest that both the Pathology Signature and the Radiology Signature could predict HRD status; however, the Patho-Radiology Signature yielded improved predictive accuracy, although direct comparison was not possible owing to differences in the test dataset. We propose that these three signatures could help predict HRD in clinical scenarios. We subsequently selected cases with data available in all modalities to investigate the correlation and possible interaction between the Pathology Signature and the Radiology Signature \u003cstrong\u003e(Figure 4E)\u003c/strong\u003e. The heatmap showed no significant correlation between the Pathology Signature and the Radiology Signature \u003cstrong\u003e(Figure 4F)\u003c/strong\u003e. The scatter plot showed that the predictive value of the Pathology Signature and the Radiology Signature were distributed differently among all cases \u003cstrong\u003e(Figure 4G)\u003c/strong\u003e. The Patho-Radiology Signature exhibited enhanced predictive power in terms of both AUC and ridge plot metrics.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMolecular characterization of patients with actual and Pathology-predicted HRD\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe PAM50 molecular subtypes, genetic alterations, and RNA expression of HRR-associated genes in patients with actual HRD and Pathology Signature-predicted HRD are summarized \u003cstrong\u003e(Figure 5)\u003c/strong\u003e. For the PAM50 classification, the basal-like, Luminal A and luminal B subtypes were observed in 18.9%, 37.8% and 43.3% of cases with actual HRD and in 21.1%, 36.9%, and 42.0% of cases predicted as HRD by the Pathology Signature, respectively. There was no significant difference in the distribution of the different subtypes between the actual and predicted HRD patients (P= 0.95). The ETS fusion was the most prevalent genetic alteration, followed by the FOXA1 mutation, SPOP mutation and RB1 deletion. In patients with actual HRD, the ETS gene fusion, RB1 deletion, HDAC2 deletion, FOXA1 mutation, and ROS1 deletion were the most common alterations (37.8%, 26.7%, 21.1%, 18.9%, and 18.9%, respectively), whereas these alterations were observed in 30.7%, 18.7%, 16.7%, 24.7%, and 14.0%, predicted HRD. Both the actual and predicted HRD cases exhibited higher incidences of RB1 deletion (predicted HRD vs. predicted HRP, p\u0026lt;0.001; actual HRD vs. actual HRP, p\u0026lt;0.001), HDAC2 deletion (predicted HRD vs. predicted HRP, p=0.002; actual HRD vs. actual HRP, p\u0026lt;0.001), ROS1 deletion (predicted HRD vs. predicted HRP, p\u0026lt;0.001; actual HRD vs. actual HRP, p\u0026lt;0.001), and PMS2 mutations (predicted HRD vs. predicted HRP, p\u0026lt;0.001; actual HRD vs. actual HRP, p=0.003) when compared to their actual and predicted HRP counterparts. There was no significant difference in the frequency of genetic alterations between the predicted and actual patients with HRD (P=0.37). The predicted and actual HRD cases had similar RNA expression levels for all the genes. In conclusion, there was no significant difference in the molecular subtypes, genetic alterations, and RNA expression of HRR-associated genes in patients with actual and predicted HRD.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInterpretability of the Pathology Signature\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo improve the interpretability of the Pathology Signature, the Grad-CAM method was applied to illustrate the pathological characteristics of the images at the patch level (\u003cstrong\u003eFigure 6\u003c/strong\u003e). After viewing all the Grad-CAM figures, two pathologists discussed and summarized four HRD-associated pathological features, including high-grade tumors (enlarged nuclear-to-cytoplasmic ratio, vacuolated appearance of cells, presence of pathological mitotic figures, poor formation of glandular follicles arranged in sheets, and small nests), fibrosis of the stroma (increased proliferation of fibroblasts with significant collagen degeneration observed around high-grade tumor cells), inflammatory cell infiltration (intensive infiltration of lymphoplasmacytic cells, usually observed around the tumor), and cribriform structures (presence of irregular glandular spaces resembling a sieve or honeycomb pattern), which were possible predictors of the presence of HRD by pathologists. Furthermore, we confirmed that the Grad-CAM figures could focus on the region of higher likelihood in patches at different magnifications (e.g. \u0026times;12.5, \u0026times;20, \u0026times;50, \u0026times;100, and \u0026times;200; \u003cstrong\u003eSupplementary Figure S3\u003c/strong\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eTo evaluate whether the features of the Pathology Signature could be understood and identified by pathologists, nine pathologists with different clinical experiences were included in the following analysis: specialized urologic pathologists (pathologists 1 to 3), attending general pathologists (pathologists 4 to 6), and junior general pathologists (pathologists 7 to 9). First, nine pathologists were trained by a senior pathologist to understand the HRD-associated pathological features using the selected Grad-CAM figures. Furthermore, the nine pathologists were asked to classify 20 WISs (10 WSIs with the highest Pathology Signature scores and 10 WSIs with the lowest Pathology Signature scores) into the HRD or HRP group independently. After four weeks, the pathologists were informed about the predicted Pathology Signature score of each WSI and were asked to reclassify these WSIs.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eWhile there was no significant difference in the AUC between specialized urologic and attending general pathologists (p = 0.59), a notable difference was observed when compared with junior general pathologists (p = 0.01). Specialized urologic pathologists tended to score \u0026ldquo;very certain\u0026rdquo; more than junior general pathologists (26/60 vs. 14/60, p=0.02). After being informed about the Pathology Signature score, most pathologists could achieve a higher AUC and accuracy. The perceived certainty of the pathologists\u0026apos; classifications (defined as \u0026ldquo;very certain\u0026rdquo;, \u0026ldquo;fairly certain\u0026rdquo;, \u0026ldquo;relatively uncertain\u0026rdquo;) were also improved after informed about the Pathology Signature score in the three group of pathologists. However, based on repeated measures ANOVA, only junior general pathologists showed a statistically significant improvement (p=0.04). In contrast, specialized urologic and attending general pathologists did not demonstrate significant changes (p=0.33 and p=0.93, respectively). Although preliminary, these results illustrated that the pathological features could be understood by pathologists and helped address the interpretability of the Pathology Signature.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAssociation between HRD and biochemical recurrence (BCR)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn patients with actual HRD, the 3-year BCR-free survival rate was 77.5%, while the 3-year BCR-free survival rate was 80.7% in patients with HRP (p=0.1969). The Kaplan-Meier curve showed that there was no significant difference in BCR-free survival between actual HRD and HRP. The HRD status predicted by the Clinical Signature could stratify patients\u0026rsquo; risk of BCR in the training and test cohorts. However, the HRD status predicted by the Pathology Signature and Radiology Signature could not stratify the risk of BCR (\u003cstrong\u003eSupplementary Figure S4\u003c/strong\u003e).\u003c/p\u003e"},{"header":"DISCUSSION","content":"\u003cp\u003ePARP inhibitors have been approved by the United States Food and Drug Administration (FDA) for the treatment of patients with advanced PCa and HRD. Thus, predicting HRD is crucial for the treatment of patients with advanced PCa; however, current prediction methods based on genetic testing are usually time-consuming, commonly inaccessible, and very expensive\u003csup\u003e20\u003c/sup\u003e. For instance, the FDA approved three assays for the detection of HRD in PCa (BRACAnalysis CDx testing blood\u003csup\u003e21\u003c/sup\u003e, FoundationOne CDx testing tissue\u003csup\u003e22\u003c/sup\u003e, and FoundationOne Liquid CDx testing plasma\u003csup\u003e23\u003c/sup\u003e); but these tests are commonly inaccessible for patients in underdeveloped regions. Additionally, the cost of these assays is as high as $4,800 to $5,800 per test. Patients with HRD are relatively\u0026nbsp;infrequently observed among PCa patients and require genetic tests of approximately ten patients to identify\u0026nbsp;one patient with HRD, which means it costs $48,000 to $58,000 to identify one patient with HRD\u003csup\u003e24\u003c/sup\u003e. These logistic and financial challenges discouraged patients from undergoing genetic testing, resulting in missed opportunities for the\u0026nbsp;accurate application of PARP inhibitors in PCa patients with HRD.\u003c/p\u003e\n\u003cp\u003eRecently, AI has been integrated into cancer management. Recently, the FDA approved the application of Paige-AI for the diagnosis of prostate biopsy specimens\u003csup\u003e25\u003c/sup\u003e. In this study, we developed\u0026nbsp;an AI-based HRD status prediction tool\u0026nbsp;for patients with PCa using multimodal data including pathological, radiological, and clinical features. The Pathology Signature, Radiology Signature, and Patho-Radiology Signature could effectively predict the HRD status in PCa patients. Our findings suggest\u0026nbsp;that the HRD status of PCa patients can be predicted by an AI-based algorithm using WSIs and MRI images. This prediction tool does not require additional examinations other than\u0026nbsp;the clinical routine. Pathological slides and MRI\u0026nbsp;images were\u0026nbsp;routinely collected. This made it possible to collect data retrospectively for prediction. In addition, only a very limited amount of time\u0026nbsp;is required to generate the predicted results, thereby saving\u0026nbsp;waiting time using genetic tests. Clinically, we can\u0026nbsp;apply this prediction tool as a pre-screening tool to select high-risk cases, thus reducing the overall cost.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn single-modal prediction models, Pathology Signature and Radiology Signature showed high accuracy, whereas the Clinical Signature showed a limited predictive performance (AUC=0.6). This is in accordance with a previous study indicating low relevance between HRD and clinical features\u003csup\u003e26,27\u003c/sup\u003e. In contrast, Gerstung et al. illustrated\u0026nbsp;that a series of genetic alterations\u0026nbsp;are associated with pathological features\u003csup\u003e28\u003c/sup\u003e. The prediction of molecular alterations using radiological data has also\u0026nbsp;been\u0026nbsp;confirmed by recent studies on gliomas and non-small cell lung cancer\u003csup\u003e29,30\u003c/sup\u003e. In addition, there are reports on multimodal methods for genetic alteration prediction or treatment response prediction\u003csup\u003e27,31–33\u003c/sup\u003e. In this study, the Patho-Radiology Signature yielded a high predictive performance (AUC=0.933\u0026nbsp;in NavieBayes). The model integrated with pathological features and radiology features could achieve high predictive performance in achieving\u0026nbsp;a higher AUC (\u003cstrong\u003eFigure 4C, 4D\u003c/strong\u003e) and a more distributed predicted risk (\u003cstrong\u003eFigure 4E\u003c/strong\u003e). It could be interpreted that the\u0026nbsp;Pathology Signature and Radiology Signature were not strongly corelated (\u003cstrong\u003eFigure 4F)\u003c/strong\u003e, but complementary to each other (\u003cstrong\u003eFigure 4G\u003c/strong\u003e). However, the predictive performance decreased after introducing clinical features into the Patho-Radiology model (CPR model AUC=0.850 in the NaiveBayes model and the highest AUC=0.867 in\u0026nbsp;the MLP\u0026nbsp;model). The possible reason for the decreased performance may be the low predictive accuracy of\u0026nbsp;the Clinical Signature itself. In addition,\u0026nbsp;the learning strategy or optimization methods of the model cannot effectively handle the newly added clinical modality. Owing to the limited number of cases with multimodal data, the actual reason\u0026nbsp;for this could not be identified in this study. Future studies may consider\u0026nbsp;prototypical modality rebalance\u003csup\u003e34\u003c/sup\u003e or modality balance networks\u003csup\u003e35\u003c/sup\u003e to improve the performance of multimodal models.\u003c/p\u003e\n\u003cp\u003eIn analyzing the pathological data, previous studies suggested that a weakly supervised approach, only label tumor or non-tumor by slide level, can yield satisfied results\u003csup\u003e36\u003c/sup\u003e. Initially, we attempted to apply weakly supervised methods; however, we found that the use of non-annotated slides failed to yield satisfactory results. We attributed the differences between our study and previous studies using weakly supervised methods to the nature of PCa, in which tumor heterogeneity played a significant role. Unlike tumors with distinct boundaries between the tumor and normal tissues, cancer, normal, and prostate stromal tissues are intertwined in PCa. The tumor proportion in one slide varied from 5% to 95% among different slides\u003csup\u003e37\u003c/sup\u003e. Without defining the tumor region, it was difficult for the model to capture the features of the tumor. Therefore, in this study, we employed annotation of the tumor region, defined as\u0026nbsp;the area with the major proportion of\u0026nbsp;tissue consisting of tumor cells and intratumoral stromal tissue. We did not strictly annotate every region of\u0026nbsp;the adenocarcinoma at the cellular level because we believe\u0026nbsp;that the stromal structures within the tumor might be relevant to molecular changes.\u0026nbsp;In\u0026nbsp;particular, the pathological grade of PCa is closely related to\u0026nbsp;tissue structure rather than the characteristics of individual cancer cells, which is different\u0026nbsp;from many other malignancies\u003csup\u003e38\u003c/sup\u003e. Annotating only the\u0026nbsp;cancerous region, but not the stromal region of PCa, would result in the loss of important structural information. In contrast, for cancer types with pathology slides predominantly consisting of tumor tissue,\u0026nbsp;weakly supervised methods can improve\u0026nbsp;the prediction\u003csup\u003e39\u003c/sup\u003e. Our findings demonstrated that the supervised approach may be more applicable\u0026nbsp;for predicting molecular alterations\u0026nbsp;in prostate cancer.\u003c/p\u003e\n\u003cp\u003eTo enhance interpretability, a Grad-CAM plot was used to analyze the pathology images. Pathologists have identified features including high-grade tumors, fibrosis of the stroma, inflammatory cell infiltration, and cribriform structures. This is in accordance with a pre-print paper by Jakob et al., which identified that high grade, fibrosis, and lymphocytic infiltration were associated with HRD in different types of tumors\u003csup\u003e40\u003c/sup\u003e. However, hemorrhage was associated with HRD by Jakob et al.; however,\u0026nbsp;this association was not confirmed in this study because there were very few hemorrhages in PCa. In addition, cribriform structures were associated with HRD in this study but not in a previous study. We suppose that this is because cribriform structures are more often observed in high-grade PCa than in other malignancies.\u0026nbsp;To provide further\u0026nbsp;interpretations\u0026nbsp;of the Patho-Radiology Signature based on the Naïve Bayes algorithm, a Probability Density Function was introduced. We illustrated the mean and variance of the Gaussian distribution for each feature\u0026nbsp;\u003cstrong\u003e(Supplementary Figures S5-S6)\u003c/strong\u003e. The top radiological features exhibiting pronounced differences between the two groups were predominantly associated with texture. This observed pattern may be associated with high-grade tumor and intra-tumor heterogeneity. Nevertheless, considering the methodology employed, which involved the leave-one-out selection of features and subsequent dimensionality reduction, an in-depth analysis and validation are imperative to derive definitive and unbiased conclusions.\u003c/p\u003e\n\u003cp\u003eThe limitations of this study include the limited MRI data in the TCGA-PRAD cohort, leading to difficulties in performing external validation of the radiological and multimodal models. This highlights one of the obstacles in current studies on PCa, namely multi-omics data rareness. In addition, the sample size of this study was relatively small. Finally, it is important to note that the modeling was conducted in a Caucasian-dominant cohort and validated in an Asian cohort, thereby limiting the assessment of ethnic disparities in the study. Future studies should include cohorts with diverse genetic backgrounds to validate the findings of this study.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cp\u003eThis study complied with relevant ethical regulations, and its protocols were approved by the Institutional Review Board of Shanghai Changhai Hospital (CHEC2022-151). This study was registered with the Chinese Clinical Trial Registry (ChiCTR2200064329). The requirement for informed consent was waived for this retrospective study, and the participants were not compensated for. Two cohorts of hormone-sensitive PCa patients were included in this study. The study design is illustrated in\u003cstrong\u003e\u0026nbsp;Figure 1\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCohort description\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe TCGA-PRAD cohort consisted of 387 patients with PCa\u003csup\u003e41\u003c/sup\u003e, for which we downloaded diagnostic HE-stained WSIs from the Genomic Data Commons (GDC) portal (https://portal.gdc.cancer.gov/). The CPGEA cohort was our in-house cohort, consisting of 210 pathologically confirmed PCa patients sequenced and reported in a previous study\u003csup\u003e42\u003c/sup\u003e. WSIs and MRI images of patients in the CPGEA cohort were collected and analyzed for the first time. All clinicopathological information related to the medical history was deposited in a prospectively collected database. Patients in the two cohorts were eligible for this study if they had at least one of\u0026nbsp;the following: (1) WSIs on H\u0026amp;E slides of radical prostatectomy. (2) multi-parametric prostate MRI or pelvic MRI before radical prostatectomy. Two patients in the TCGA-PRAD cohort and the CPGEA cohort, respectively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStudy design\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor the analysis of pathology and clinical data, we selected the TCGA-PRAD cohort as the training cohort and the CPGEA cohort as the test cohort. However, for the MRI images, we randomly sampled 80% of all patients with MRI images for training, whereas the remaining patients were used as the test cohort. This sampling strategy was employed to ensure fair comparisons between unimodal and multimodal models, thereby avoiding any spurious differences in test concordance indices that may arise due to selective patient exclusion in some models but not in others \u003cstrong\u003e(Supplementary Figure S1)\u003c/strong\u003e. The proportion of data annotated by each pathologist and radiologist in each cohort and the different types of equipment was not significantly different between the two cohorts \u003cstrong\u003e(Supplementary Figure S7)\u003c/strong\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eInferring genetic alteration status\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe HRD status was ascertained based on alterations in 14 homologous recombinant repair (HRR)-related genes, as previously demonstrated (including ATM, BRCA1, BRCA2, BARD1, BRIP1, CDK12, CHEK1, CHEK2, FANCL, PALB2, RAD51B, RAD51C, RAD51D, and RAD54L)\u003csup\u003e43\u003c/sup\u003e. Patients were assigned to the HRD subtype if they had at least one nonsynonymous mutation or deep deletion in the HRR-related genes.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRadiology pipeline\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA flowchart of the radiology pipeline is shown in \u003cstrong\u003eFigure 2\u003c/strong\u003e. Prostate MRI was performed on a 3.0T MR scanner with an abdominal phase array coil, following a 4h fasting period and enema treatment with glycerin (20 ml). Routine sequences included sagittal T2WI, axial high-resolution T2WI, and axial DWI. \u003cstrong\u003eSupplementary Table\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003eS9\u003c/strong\u003e shows the axial T2WI and DWI parameters used for the machine learning. Each radiologist traced the outer contour of the prostate lesions on the tumor-containing axial section using Insight Segmentation and Registration Toolkit-SNAP v.3.8.0 software. Handcrafted features were extracted using Pyradiomics (https://pyradiomics.readthedocs.io/en/latest/).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePathology pipeline\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA flowchart of the pathology pipeline is shown in \u003cstrong\u003eFigure 3\u003c/strong\u003e.\u0026nbsp;All the WSIs were digitally captured using a 20\u0026times; objective lens. The task of annotating the H\u0026amp;E WSIs to identify the tumor region was performed by two fellowship-trained uropathologists using the QuPath software (https://github.com/qupath). The WSIs were partitioned into 512\u0026times;512-pixel patches. The ResNet50 algorithm was then used to predict the tumor region based on annotations by pathologists. The patch label and corresponding probability were extracted and applied to generate WSI-level features via Patch Likelihood Histogram (PLH) and bag-of-words (BoW) pipelines. WSI-level features were applied to construct pathological models for HRD status prediction.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMultimodal integration\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAn early fusion strategy, which involved normalizing numeric features and concatenating them with categorical features to create new integrated features was applied. Based on these selected and optimized features, we input them into a predictive pipeline composed of ten machine learning models, providing a prediction probability for each case.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVisualization of Pathology Signature prediction with Gradient-weighted Class Activation Mapping (\u003c/strong\u003e\u003cstrong\u003eGrad-CAM)\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGrad-CAM class localization maps are created by visualizing the gradients flowing into the final convolutional layer of the network immediately before the fully connected layers. Because the convolutional layers contain class-specific spatial information from the input image that is lost in the fully connected layers, this is the optimal point for map generation. Grad-CAM requires neither any modifications to the existing model architecture nor any retraining of the model.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSurvival analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLinear Cox proportional hazards models with L2 regularization (c=0.5) and without L1 regularization were applied to all the multimodal and unimodal models. The homologous recombination potent (HRP) subtype was designated as high-risk (risk score 1), and patients assigned to the HRD subtype were designated as low-risk (risk score 0); no interaction terms were used.\u0026nbsp;\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eWe illustrated that the Pathology Signature, Radiology Signature, and Patho-Radiology Signature obtained high accuracy in predicting HRD status in PCa patients. The framework of this study can be applied to other malignancies and highlights the promising applications of AI for treatment selection in cancer management. In clinical settings, AI-based prediction tools can be applied as pre-screening tools to reduce the number of genetic analyses required.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eData with identifiers in the CPEGA cohort will be made available to research partners upon reasonable request to the corresponding author R-C or the Shanghai Changhai Hospital, China, which is a data transfer agreement approved by the legal departments of the requesting researcher and all legal departments of the institutions that provided data for the study and an ethics clearance. All images and patient data from TCGA-PRAD cohort used in this study are publicly available at http://portal. gdc. cancer. gov/.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe code for the model development is available online after the article has been published.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding:\u0026nbsp;\u003c/strong\u003eThis study was supported by grants from the National Natural Science Foundation of China (NSFC) (grant number:82272905), Rising Star Program of Shanghai Science and Technology Commission (grant number:21QA1411500), and the Shanghai Action Plan for Technological Innovation Grant (No. 22ZR1478000) and Guhai Project of Shanghai Changhai Hospital.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor information:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eConceptualization: R.C., G.X., and L.W. Methodology: Z.S., Q.Z., N.T., Y.Z., W.Z., and L.D. Investigation: Z.S., Q.Z., N.T., Y.Z., W.Z., L.D., Y.W., M.Q., Z.D., H.Y., Y.L., X.W., Y.Y., H.W., L.Z., Y.G., W.Z., and Y.L. Software: Z.S., Q.Z., N.T., Y.Z., W.Z., L.D., Y.W., M.Q., Z.D., and H.Y. Visualization: W.Z., L.D., Y.W., M.Q., Z.D., H.Y., H.W., L.Z., Y.G., W.Z., and Y.L. Supervision: R.C., G.X., and L.W. Writing\u0026mdash;original draft: Z.S., Q.Z., N.T., Y.Z., W.Z., and L.D. Writing\u0026mdash;review and editing: Y.W., M.Q., Z.D., H.Y., H.W., L.Z., Y.G., W.Z., Y.L., R.C., G.X., and L.W. Resources: Z.S., Q.Z., N.T., Y.Z., Y.L., X.W., and Y.Y. Validation: R.C., G.X., and L.W. Project Administration: R.C., G.X., and L.W. Funding Acquisition: R.C., G.X., and L.W.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests:\u0026nbsp;\u003c/strong\u003eThe authors declare no conflicts of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eRl, S., Kd, M., Ns, W. \u0026amp; A, J. Cancer statistics, 2023. \u003cem\u003eCA: a cancer journal for clinicians\u003c/em\u003e \u003cstrong\u003e73\u003c/strong\u003e, (2023).\u003c/li\u003e\n\u003cli\u003eMitchell, T. \u0026amp; Neal, D. E. The genomic evolution of human prostate cancer. \u003cem\u003eBr J Cancer\u003c/em\u003e \u003cstrong\u003e113\u003c/strong\u003e, 193\u0026ndash;198 (2015).\u003c/li\u003e\n\u003cli\u003eNguyen, L., W. M. Martens, J., Van Hoeck, A. \u0026amp; Cuppen, E. Pan-cancer landscape of homologous recombination deficiency. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 5584 (2020).\u003c/li\u003e\n\u003cli\u003eHoeijmakers, J. H. Genome maintenance mechanisms for preventing cancer. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e411\u003c/strong\u003e, 366\u0026ndash;374 (2001).\u003c/li\u003e\n\u003cli\u003eGroelly, F. J., Fawkes, M., Dagg, R. A., Blackford, A. N. \u0026amp; Tarsounas, M. Targeting DNA damage response pathways in cancer. \u003cem\u003eNat Rev Cancer\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e, 78\u0026ndash;94 (2023).\u003c/li\u003e\n\u003cli\u003eESMO 2019: PROfound: Phase 3 Study of Olaparib vs. Enzalutamide or Abiraterone for Metastatic Castration-Resistant Prostate Cancer with Homologous Recombination Repair Gene Alterations. https://www.urotoday.com/conference-highlights/esmo-2019/esmo-2019-prostate-cancer/115401-esmo-2019-profound-phase-3-study-of-olaparib-vs-enzalutamide-or-abiraterone-for-metastatic-castration-resistant-prostate-cancer-with-homologous-recombination-repair-gene-alterations.html.\u003c/li\u003e\n\u003cli\u003ede Luca, X. M. \u003cem\u003eet al.\u003c/em\u003e Using whole-genome sequencing data to derive the homologous recombination deficiency scores. \u003cem\u003enpj Breast Cancer\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 1\u0026ndash;8 (2020).\u003c/li\u003e\n\u003cli\u003eSztupinszki, Z. \u003cem\u003eet al.\u003c/em\u003e Detection of Molecular Signatures of Homologous Recombination Deficiency in Prostate Cancer with or without BRCA1/2 Mutations. \u003cem\u003eClin Cancer Res\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 2673\u0026ndash;2680 (2020).\u003c/li\u003e\n\u003cli\u003eCoudray, N. \u0026amp; Tsirigos, A. Deep learning links histology, molecular signatures and prognosis in cancer. \u003cem\u003eNat Cancer\u003c/em\u003e \u003cstrong\u003e1\u003c/strong\u003e, 755\u0026ndash;757 (2020).\u003c/li\u003e\n\u003cli\u003eKather, J. N. \u003cem\u003eet al.\u003c/em\u003e Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 1054\u0026ndash;1056 (2019).\u003c/li\u003e\n\u003cli\u003eYamashita, R. \u003cem\u003eet al.\u003c/em\u003e Deep learning model for the prediction of microsatellite instability in colorectal cancer: a diagnostic study. \u003cem\u003eLancet Oncol\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 132\u0026ndash;141 (2021).\u003c/li\u003e\n\u003cli\u003eCifci, D., Foersch, S. \u0026amp; Kather, J. N. Artificial intelligence to identify genetic alterations in conventional histopathology. \u003cem\u003eJ Pathol\u003c/em\u003e \u003cstrong\u003e257\u003c/strong\u003e, 430\u0026ndash;444 (2022).\u003c/li\u003e\n\u003cli\u003eDadhania, V. \u003cem\u003eet al.\u003c/em\u003e Leveraging artificial intelligence to predict ERG gene fusion status in prostate cancer. \u003cem\u003eBMC Cancer\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 494 (2022).\u003c/li\u003e\n\u003cli\u003eSchaumberg, A., Rubin, M. \u0026amp; Fuchs, T. H\u0026amp;E-stained Whole Slide Image Deep Learning Predicts SPOP Mutation State in Prostate Cance. Preprint at https://doi.org/10.1101/064279 (2016).\u003c/li\u003e\n\u003cli\u003eHoulahan, K. E. \u003cem\u003eet al.\u003c/em\u003e Molecular hallmarks of multiparametric MRI visibility in prostate cancer. \u003cem\u003eEur Urol\u003c/em\u003e \u003cstrong\u003e76\u003c/strong\u003e, 18\u0026ndash;23 (2019).\u003c/li\u003e\n\u003cli\u003eBoehm, K. M. \u003cem\u003eet al.\u003c/em\u003e Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer. \u003cem\u003eNat Cancer\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 723\u0026ndash;733 (2022).\u003c/li\u003e\n\u003cli\u003eVanguri, R. S. \u003cem\u003eet al.\u003c/em\u003e Multimodal integration of radiology, pathology and genomics for prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer. \u003cem\u003eNat Cancer\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 1151\u0026ndash;1164 (2022).\u003c/li\u003e\n\u003cli\u003eFremond, S. \u003cem\u003eet al.\u003c/em\u003e Interpretable deep learning model to predict the molecular classification of endometrial cancer from haematoxylin and eosin-stained whole-slide images: a combined analysis of the PORTEC randomised trials and clinical cohorts. \u003cem\u003eLancet Digit Health\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, e71\u0026ndash;e82 (2023).\u003c/li\u003e\n\u003cli\u003eLi, J. \u003cem\u003eet al.\u003c/em\u003e A genomic and epigenomic atlas of prostate cancer in Asian populations. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e580\u003c/strong\u003e, 93\u0026ndash;99 (2020).\u003c/li\u003e\n\u003cli\u003eWu, K. \u003cem\u003eet al.\u003c/em\u003e Evaluation of the Efficacy of PARP Inhibitors in Metastatic Castration-Resistant Prostate Cancer: A Systematic Review and Meta-Analysis. \u003cem\u003eFront Pharmacol\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 777663 (2021).\u003c/li\u003e\n\u003cli\u003eDe Bono, J. S. \u003cem\u003eet al.\u003c/em\u003e PROfound: A randomized Phase III trial evaluating olaparib in patients with metastatic castration-resistant prostate cancer and a deleterious homologous recombination DNA repair aberration. \u003cem\u003eJCO\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, TPS5091\u0026ndash;TPS5091 (2017).\u003c/li\u003e\n\u003cli\u003eMilbury, C. A. \u003cem\u003eet al.\u003c/em\u003e Clinical and analytical validation of FoundationOne\u0026reg;CDx, a comprehensive genomic profiling assay for solid tumors. \u003cem\u003ePLoS One\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, e0264138 (2022).\u003c/li\u003e\n\u003cli\u003eWoodhouse, R. \u003cem\u003eet al.\u003c/em\u003e Clinical and analytical validation of FoundationOne Liquid CDx, a novel 324-Gene cfDNA-based comprehensive genomic profiling assay for cancers of solid tumor origin. \u003cem\u003ePLoS One\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, e0237802 (2020).\u003c/li\u003e\n\u003cli\u003eSztupinszki, Z. \u003cem\u003eet al.\u003c/em\u003e Detection of Molecular Signatures of Homologous Recombination Deficiency in Prostate Cancer with or without BRCA1/2 Mutations. \u003cem\u003eClin Cancer Res\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 2673\u0026ndash;2680 (2020).\u003c/li\u003e\n\u003cli\u003eCommissioner, O. of the. FDA Authorizes Software that Can Help Identify Prostate Cancer. \u003cem\u003eFDA\u003c/em\u003e https://www.fda.gov/news-events/press-announcements/fda-authorizes-software-can-help-identify-prostate-cancer (2021).\u003c/li\u003e\n\u003cli\u003eFeng, C. \u003cem\u003eet al.\u003c/em\u003e Relationship between homologous recombination deficiency and clinical features of breast cancer based on genomic scar score. \u003cem\u003eBreast\u003c/em\u003e \u003cstrong\u003e69\u003c/strong\u003e, 392\u0026ndash;400 (2023).\u003c/li\u003e\n\u003cli\u003eBoehm, K. M. \u003cem\u003eet al.\u003c/em\u003e Multimodal data integration using machine learning improves risk stratification of high-grade serous ovarian cancer. \u003cem\u003eNat Cancer\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 723\u0026ndash;733 (2022).\u003c/li\u003e\n\u003cli\u003eFu, Y. \u003cem\u003eet al.\u003c/em\u003e Pan-cancer computational histopathology reveals mutations, tumor composition and prognosis. \u003cem\u003eNat Cancer\u003c/em\u003e \u003cstrong\u003e1\u003c/strong\u003e, 800\u0026ndash;810 (2020).\u003c/li\u003e\n\u003cli\u003eKim, M. \u003cem\u003eet al.\u003c/em\u003e Diffusion- and perfusion-weighted MRI radiomics model may predict isocitrate dehydrogenase (IDH) mutation and tumor aggressiveness in diffuse lower grade glioma. \u003cem\u003eEur Radiol\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 2142\u0026ndash;2151 (2020).\u003c/li\u003e\n\u003cli\u003eZhang, J. \u003cem\u003eet al.\u003c/em\u003e Value of pre-therapy 18F-FDG PET/CT radiomics in predicting EGFR mutation status in patients with non-small cell lung cancer. \u003cem\u003eEur J Nucl Med Mol Imaging\u003c/em\u003e \u003cstrong\u003e47\u003c/strong\u003e, 1137\u0026ndash;1146 (2020).\u003c/li\u003e\n\u003cli\u003eSammut, S.-J. \u003cem\u003eet al.\u003c/em\u003e Multi-omic machine learning predictor of breast cancer therapy response. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e601\u003c/strong\u003e, 623\u0026ndash;629 (2022).\u003c/li\u003e\n\u003cli\u003eVanguri, R. S. \u003cem\u003eet al.\u003c/em\u003e Multimodal integration of radiology, pathology and genomics for prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer. \u003cem\u003eNat Cancer\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 1151\u0026ndash;1164 (2022).\u003c/li\u003e\n\u003cli\u003eLipkova, J. Artificial intelligence for multimodal data integration in oncology. \u003cem\u003eOPEN ACCESS\u003c/em\u003e.\u003c/li\u003e\n\u003cli\u003eFan, Y., Xu, W., Wang, H., Wang, J. \u0026amp; Guo, S. PMR: Prototypical Modal Rebalance for Multimodal Learning. Preprint at https://doi.org/10.48550/arXiv.2211.07089 (2022).\u003c/li\u003e\n\u003cli\u003eImproving Multispectral Pedestrian Detection by Addressing Modality Imbalance Problems | SpringerLink. https://link.springer.com/chapter/10.1007/978-3-030-58523-5_46.\u003c/li\u003e\n\u003cli\u003eMuti, H. S. \u003cem\u003eet al.\u003c/em\u003e Development and validation of deep learning classifiers to detect Epstein-Barr virus and microsatellite instability status in gastric cancer: a retrospective multicentre cohort study. \u003cem\u003eLancet Digit Health\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, e654\u0026ndash;e664 (2021).\u003c/li\u003e\n\u003cli\u003eMohanty, S. K., Lobo, A. \u0026amp; Cheng, L. The 2022 revision of the World Health Organization classification of tumors of the urinary system and male genital organs: advances and challenges. \u003cem\u003eHum Pathol\u003c/em\u003e \u003cstrong\u003e136\u003c/strong\u003e, 123\u0026ndash;143 (2023).\u003c/li\u003e\n\u003cli\u003evan Leenders, G. J. L. H. \u003cem\u003eet al.\u003c/em\u003e The 2019 International Society of Urological Pathology (ISUP) Consensus Conference on Grading of Prostatic Carcinoma. \u003cem\u003eAm J Surg Pathol\u003c/em\u003e \u003cstrong\u003e44\u003c/strong\u003e, e87\u0026ndash;e99 (2020).\u003c/li\u003e\n\u003cli\u003eBilal, M. \u003cem\u003eet al.\u003c/em\u003e Development and validation of a weakly supervised deep learning framework to predict the status of molecular pathways and key mutations in colorectal cancer from routine histology images: a retrospective study. \u003cem\u003eLancet Digit Health\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, e763\u0026ndash;e772 (2021).\u003c/li\u003e\n\u003cli\u003eLavinia Loeffler, C. M. \u003cem\u003eet al.\u003c/em\u003e Direct prediction of Homologous Recombination Deficiency from routine histology in ten different tumor types with attention-based Multiple Instance Learning: a development and validation study. \u003cem\u003emedRxiv\u003c/em\u003e 2023.03.08.23286975 (2023) doi:10.1101/2023.03.08.23286975.\u003c/li\u003e\n\u003cli\u003eCancer Genome Atlas Research Network. The Molecular Taxonomy of Primary Prostate Cancer. \u003cem\u003eCell\u003c/em\u003e \u003cstrong\u003e163\u003c/strong\u003e, 1011\u0026ndash;1025 (2015).\u003c/li\u003e\n\u003cli\u003eLi, J. \u003cem\u003eet al.\u003c/em\u003e A genomic and epigenomic atlas of prostate cancer in Asian populations. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e580\u003c/strong\u003e, 93\u0026ndash;99 (2020).\u003c/li\u003e\n\u003cli\u003eClarke, N. \u003cem\u003eet al.\u003c/em\u003e Olaparib combined with abiraterone in patients with metastatic castration-resistant prostate cancer: a randomised, double-blind, placebo-controlled, phase 2 trial. \u003cem\u003eLancet Oncol\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 975\u0026ndash;986 (2018).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6488233/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6488233/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eAccurate prediction of homologous recombination deficiency (HRD) status is crucial for effective management of prostate cancer (PCa). However, genetic testing is expensive and inaccessible. We propose a multimodal deep learning approach that integrates clinical information, Hematoxylin \u0026amp; Eosin (H\u0026amp;E)-stained whole-slide images (WSIs), and multi-parameter MRI images to predict the HRD status of patients with PCa. Patients from the Cancer Genome Atlas (n\u0026thinsp;=\u0026thinsp;387) and three Chinese hospitals (n\u0026thinsp;=\u0026thinsp;179) were used to establish and validate the prediction models. The Pathology Signature, Radiology Signature, and Patho-Radiology Signature could accurately predict the HRD status in external validation or 5-fold cross validation (AUC\u0026thinsp;=\u0026thinsp;0.815, 0.833, and 0.933, respectively). Notably, four interpretative pathological features were identified in the Pathology Signature. These signatures could serve as a prescreening tool to select patients for confirmatory genetic testing. We suggest that this multi-modal approach could be applied to the prediction of molecular alterations in other malignancies.\u003c/p\u003e","manuscriptTitle":"Multimodal deep learning model for predicting homologous recombination deficiency in prostate cancer: an international multi-cohort study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-05-07 07:00:21","doi":"10.21203/rs.3.rs-6488233/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"be662767-cd3d-4f03-ac32-60c6884b2566","owner":[],"postedDate":"May 7th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":47765917,"name":"Biological sciences/Cancer/Urological cancer/Prostate cancer"},{"id":47765918,"name":"Health sciences/Pathogenesis/Clinical genetics"}],"tags":[],"updatedAt":"2025-08-03T12:38:35+00:00","versionOfRecord":[],"versionCreatedAt":"2025-05-07 07:00:21","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6488233","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6488233","identity":"rs-6488233","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}