Development and validation of the utLIFE-PC algorithm for noninvasive detection of prostate cancer in urine: A prospective, observational study Authors: Sujun Han, Mingshuai Wang, Yong Wang, Junlong Wu, Zhaoxia Guo, Huina Wang, Ranlu Liu, Xiaofu Qiu, Linjun Hu, Jianbin Bi, Weigang Yan, Hengqing An, Gejun Zhang, Yi Zhi, Zhiyuan Chen, Libin Chen, Lei Liu, Huanqing Cheng, Shuaipeng Zhu, Meng Wang, Yanrui Zhang, Xiao Liu, Feng Lou, Shanbo Cao, Dingwei Ye, Yuanjie Niu, Nianzeng Xing Published: 2024 ABSTRACT SummaryOverbiopsy is a serious health issue in prostate cancer (PCa) diagnostics. We have developed a urine tumor DNA multidimensional bioinformatic algorithm, utLIFE, to avoid unnecessary biopsy. The objective is to recognize all or clinically significant PCa. Of the 801 participants recruited in our study, 630 are selected for subsequent analysis. In the training cohort (n = 237), utLIFE-PC gets an area under the receiver operating characteristic curve (AUC) of 0.967 and a sensitivity of 85.57% at 95% specificity. In the independent prospective validation cohort (n = 343), utLIFE-PC has an AUC of 0.929, sensitivity of 84.24%, and specificity of 93.26%. Notably, in patients with ≥grade group (GG)2 and ≥GG3, the assay’s sensitivity is still excellent (85.33% and 87.10%, respectively). The model shows better performance than prostate-specific antigen (PSA) (p < 0.001) or the single-dimensional biomarkers (methylation, p < 0.001; copy-number variations [CNVs], p < 0.001; mutation, p < 0.001). The utLIFE-PC model can potentially optimize the PCa diagnostic process and avoid unnecessary biopsies. This study was registered at Chinese Clinical Trial Registry: ChiCTR2300071837. Introduction Prostate cancer (PCa) is the most common solid tumor and the fifth leading cause of cancer death in men worldwide, with 1.4 million new cases and approximately 375,304 deaths in 2020.1 The mortality-to-incidence rate ratio (MR/IR) is much higher in China (0.44) than in America (0.15) since most PCa patients in China have high Gleason scores (GSs) upon diagnosis and a short survival time after that. As the PCa incidence increases at a rate of 2.6%/year in China,2 an efficient method to screen PCa and more precise measures for identifying all-grade PCa is critical. Currently, serum prostate-specific antigen (PSA) is a widely used PCa diagnosis biomarker, as the major international guidelines recommend. However, in the PSA gray zone (4–10 ng/mL), the positive biopsy rate is 25% in China3 and 40% in America.4 Even in patients with PSA above 20 ng/mL, the positive biopsy rate was only 70%.3,4 The low specificity leads to many unnecessary biopsies. Therefore, novel biomarkers for PCa are urgently needed. We previously developed a noninvasive method named the utLIFE-UC model that was applied for urothelium carcinoma (UC), which included larger copy-number variations (CNVs) and mutations and showed superior sensitivity (92.8%) and specificity (96.0%) compared with other models.5 This suggests that combinational profiling of multidimensional biomarkers may increase the performance of cancer detection. CNVs and mutations are associated with various carcinomas and the hallmarks of almost all human cancers, including PCa. However, the application of multidimensional biomarkers in PCa diagnosis based on DNA has yet to be investigated. The prevalence of mutations and CNVs was much lower in PCa than in UC,6 indicating that combining mutations and CNVs may not achieve outstanding performance in PCa. Aberrant changes in DNA methylation, such as the hypermethylation of CpG islands, are another classic biomarker in the early detection of tumors.7 Many methylated genes are observed in PCa by assessing tissue biopsies, serum, or urine; these genes include GSTP1, MGMT, CD44, EDNRB, APC, TIMP2, CDKN2A, RASSF1A, and AR.8 The use of these genes may increase the accuracy of PCa detection. In this study, we developed a urine DNA multiple-dimensional model (utLIFE-PC) by combining genetic mutations, CNVs, and methylation profiles for the early detection of PCa and assessed the performance of utLIFE-PC through a prospective cohort. This utLIFE-PC score is highly associated with PCa and can accurately detect PCa. Results Patient characteristics Of the 801 participants recruited in this study, 630 participants were selected for the subsequent analysis (Figure 1). The median age was 69 in both the training and validation cohorts for PCa patients, and the median age was 68 in the training and validation cohorts, for biopsy-negative (BxN) patients. The median PSA in the training and validation cohorts was 23.64 ng/mL and 16.93 ng/mL in the PCa patients, and 10.05 ng/mL and 10.00 ng/mL in the BxN patients, respectively (Tables 1 and 2). More than 95% of PCa and BxN patients underwent ≥12 core systematic biopsies. The proportion of abnormal digital rectal examination (DRE) for PCa and BxN patients was 40.20% and 51.51% in the training cohort, and 36.36% and 36.52% in the validation cohort, respectively. The ratio of PCa patients in these two cohorts reflected the percentages observed in the real-world clinical setting (40%–60%). Of all the PCa patients, 9.54% had grade group (GG) 1, 13.74% had GG2, and 76.72% had GG ≥ 3 (Tables 1 and 2). Study design and participant enrollment In the discovery phase, 824 tissue samples and 237 urine samples were used to screen the mutation biomarkers. In all, 138 genes shared in tissue and urine were selected as candidate biomarkers. For methylation biomarkers screening, 10 pairs of urine samples and 531 tissue samples from TCGA were used. Two prostate cancer-related biomarkers shared in urine and TCGA were selected as candidate biomarkers. In the model construction and validation phase, the study recruited 801 participants. 171 cases (61 without biopsy results, 101 failed QC, 9 other cancers) were excluded according to the exclusion criteria, sample QC. Finally, 50 healthy participants were used for baseline construction; 237 participants were included in the training cohort; and 343 participants were included in the validation cohort. QC, quality control; PCa, prostate cancer; BxN, biopsy negative. Table 1Subject baseline demographics and clinical characteristics in the training cohortCharacteristicsCaseControlPCa (n = 97)BxN (n = 66)Healthy (n = 64)GUC (n = 10)AgeMedian (IRQ)69 (64,75)68 (62,73)53 (50,61)67 (55,75)PSAMedian (IRQ)23.64 (10,100.00)10.05 (7,14.7)2.49 (1.12,3.31)2.75 (2.4,3,22)GleasonGG110–––GG210–––GG312–––GG432–––GG533–––BxN–66––Core of biopsyMedian (IRQ)13 (12,16)13 (12,19)––DRENegative103––Positive3934––Not reported48296410Renal carcinoma–––5Urothelium carcinoma–––5Healthy––64– Subject baseline demographics and clinical characteristics in the training cohort Table 2Subject baseline demographics and clinical characteristics in the validation cohortCharacteristicsBiopsy positive (n = 165)Biopsy negative (n = 178)p valueAge––0.631Median (IQR)69 (64,75)68 (64,74)–PSA––<0.001Median (IQR)16.93 (9.18,63)10.00 (7.00,17.00)–Gleason–––GG115––GG226––GG324––GG453––GG547––BxN–178–Core of biopsy––0.474Median (IQR)13 (12,16)13 (12,16)–DRE–––Negative1111–Positive6065–Not reported94102–p value: Mann-Whitney U was used to test the significance between biopsy-positive/negative groups. Subject baseline demographics and clinical characteristics in the validation cohort p value: Mann-Whitney U was used to test the significance between biopsy-positive/negative groups. Biomarkers for PCa in urine To identify a subset of candidate biomarkers for PCa detection, mutations, CNVs, and methylation biomarkers were analyzed in the training cohorts. First, the genetic profiles of PCa and control samples were analyzed to find candidate genes. In all, 138 genes were chosen as candidate markers, including KMT2D, TP53, TET2, DNMT3A, ATM, FOXA1, KMT2C, BRCA2, CDK12, KMT3B, and PTEN (Table S1). At least one candidate mutation was found in 68.04% (66/97) of PCa patients but not in any of the noncancerous controls (Figure 2A; Figure S1). This mutation was strongly correlated with PCa (p < 0.01) (Table S2). For the CNV features, the CNV score of each sample was calculated as described in the methods. Most of the CNV scores of the 19 selected genome regions were also significantly related to PCa (p < 0.01), and all were frequently (≥20%) identified in tumor cases but not in controls (≤10%) (Figures 2B and S1; Tables S2 and S3). Next, hypermethylated APC and GSTP1 were identified in PCa samples retrieved from the The Cancer Genome Atlas (TCGA) database and in-house urine samples from PCa according to the method. The methylation levels of APC and GSTP1 were significantly higher in the urine sample of PCa patients than in the BxN patients (Figures 2C and S1), and public data (Figure S1) imply that high methylation levels of these two genes in the urine sample could be seen as risk factors for PCa (Table S2). Stacked bar plots showing the proportions of mutation/methylation/CNV features in the case-control group of the training cohort (A) The blue box means the percentage of samples with mutations. The red box means the percentage of samples without mutations. The p values were calculated using the Fisher exact test. (B) CNV scores of 19 selected CNV biomarkers in urine. The blue boxplots represent the CNV scores in healthy populations and BxN patients. The red boxplots represent the CNV scores in PCa patients. Mann-Whitney U test was performed in statistical analysis. (C) Methylation levels of APC and GSTP1 in urine discriminate cases from controls. The blue boxplots represent the methylation levels in healthy populations and BxN patients. The red boxplots represent the methylation levels in PCa patients. PCa, prostate cancer; BxN, biopsy negative. Data are represented as mean ± SD. Mann-Whitney U test was performed in statistical analysis. A p value ≤0.05 was considered statistically significant (∗). A p value ≤0.01 was considered highly statistically significant (∗∗). A p value ≤0.001 was considered extremely statistically significant (∗∗∗). See also Figure S1 and Tables 1 and 2. All experiments have no repeats. Construction of a multidimensional model for PCa detection To reduce the number of false positives and unnecessary biopsies, an utLIFE-PC model was constructed by a systematic machine learning (ML) framework using combined assays for mutations, CNVs, and methylation for PCa detection (Figure 3; Table S4). The detailed procedure used to build the utLIFE-PC model is described in the following. (1)The feature matrix contains the absence/presence of mutation features, the Z score of CNV features, and the ΔCT values of methylation features. The feature matrix contains the absence/presence of mutation features, the Z score of CNV features, and the ΔCT values of methylation features. (2)In the training phase, a random forest model was considered. The dataset was first randomly sampled n times, and then the sub-dataset was used as the training data; each branch of the sub-training set is called a tree so that n trees can be obtained. Each tree obtains the result of sample prediction through the distribution of features. The prediction results of n trees were calculated, and the most frequent prediction result will be voted as the final prediction result. In the training phase, a random forest model was considered. The dataset was first randomly sampled n times, and then the sub-dataset was used as the training data; each branch of the sub-training set is called a tree so that n trees can be obtained. Each tree obtains the result of sample prediction through the distribution of features. The prediction results of n trees were calculated, and the most frequent prediction result will be voted as the final prediction result. (3)The clinical performance of the utLIFE-PC model was validated in a prospective initial biopsy population. The clinical performance of the utLIFE-PC model was validated in a prospective initial biopsy population. Schematic illustration of the utLIFE-PC algorithm Urine samples were collected from initial biopsy patients, as well as GUC participants and healthy controls. The ucfDNA was then extracted from the urine supernatant samples and subjected to target sequencing and whole-genome sequencing. The uexDNA was extracted from the urine sediment samples and subjected to qMSP. Mutation, CNV, and methylation features were extracted, and a base model was constructed. The DNA features were then calculated into a large matrix, which was subsequently trained by random forest. Then the utLIFE-PC algorithm was validated in a prospective independent cohort. See also Tables S3–S6. The specificity of detecting PCa among healthy participants, BxN patients, and other genitourinary cancer (genitourinary cancer [GUC], kidney cancer, and bladder cancer) patients was 96.88%, 93.94%, and 90.00%, respectively (Table S5). To better screen tumor characteristics and further enhance the specificity of PCa identification, the aforementioned three groups were all included in the training cohort as part of the control group. Overall, the utLIFE-PC model achieved a sensitivity of 85.57% (95% confidence interval [CI], 76.97%–91.88%) at 95.00% (95% CI, 89.97%–97.97%) specificity with a specific cutoff of 0.402 in the training cohort (Table S6). This model showed a superior area under the receiver operating characteristic curve (AUC) (0.967; 95% CI, 0.947–0.988) to any other model tested, including mutations (0.861; 95% CI, 0.810–0.902, p < 0.001), CNVs (0.860; 95% CI, 0.813–0.908, p < 0.001), mutations and CNVs (0.948; 95% CI, 0.918–0.977, p < 0.05), methylation status (0.761; 95% CI, 0.695–0.828, p < 0.001), or PSA (0.841; 95% CI, 0.790–0.892, p < 0.001) (Figure 4A). In addition, the performance of utLIFE-PC was rather robust among different PSA levels (Table S6). In the decision curve analysis (DCA) analysis, utLIFE-PC provided a higher net benefit across a wide range of risk acceptance levels compared to the separate features and PSA (Figure 4B). Clinical performance of the utLIFE-PC model validation cohorts The receiver operating characteristic analysis (ROC) plots detailing AUC of different risk models: (A) any cancer in the training cohort, (C) any cancer in the validation cohort, (E) detection of ≥GG2 of validation cohort, and (G) detection of ≥GG3 of validation cohort. Decision curve analysis (DCA) plots detail the standardized net benefit (sNB) of adopting different risk models. Panels show the sNB based upon the detection of varying levels of disease severity: (B) any cancer of training cohort, (D) any cancer of validation cohort, (F) ≥GG2 PCa of validation cohort, and (H) ≥GG3 PCa of validation cohort. All these analyses are shown to compare the performances of the utLIFE-PC model, methylation features, CNV features, mutation features, and PSA. utLIFE-PC, utLIFE-PC model; Mut, mutation feature; CNV, CNV feature; Mut+CNV, model combined mutation and CNV; Meth, methylation feature; PSA, total prostate-specific antigen. (I) Analysis of the utLIFE-PC scores in relation to prostate biopsy outcomes in the validation cohort. utLIFE-PC score is shown on the y axis. Each colored bar represents an individual patient’s utLIFE-PC score and true biopsy diagnosis, increasing left to right (0–1 scale). Light green indicates benign biopsy; purple, biopsy GS 3 + 3; blue, biopsy GS 3 + 4; orange, dominant GG 3 or higher. Gray horizontal lines represent the cutoff point. A data statistic of different GGs is shown below the waterfall plot. utLIFE-PC score = risk score of utLIFE-PC model; BxN, biopsy negative; GG, grade group. See also Figures S2 and S3 and Tables S7–S10. Validation of the model for cancer diagnosis To further validate the diagnostic performance of the utLIFE-PC assay, we assessed it in an independent prospective cohort that only included the initial biopsy populations. The utLIFE-PC model still maintained a stable accuracy, with an AUC of 0.929 (95% CI: 0.897–0.961), sensitivity of 84.24% (95% CI: 77.77%–89.44%), specificity of 93.26% (95% CI: 88.52%–96.47%), positive predictive value (PPV) of 92.05% (95% CI: 86.98%–95.26%), and negative predictive value (NPV) of 86.46% (95% CI: 81.74%–90.11%) (Figure 4C; Table S7) and a higher net benefit (Figure 4D). The posttest probability of PCa for positive prediction was 90.33% (95% CI: 84.35%–94.19%), and that for negative prediction was 11.22% (95% CI: 8.14%–15.27%) (Table S7). At a decision threshold of 0.4, the utLIFE-PC could result in up to 38% fewer unnecessary biopsies of men presenting with a suspicion of PCa, without missing substantial numbers of men with aggressive disease (Figure S2A). Notably, even in patients with gray-zone PSA levels (4–10 ng/mL), the utLIFE-PC model still achieved an accuracy of 88.62% (95% CI: 81.64%–93.64%), a sensitivity of 84.44% (95% CI: 70.54%–93.51%), and a specificity of 91.03% (95% CI: 82.38%–96.32%) (Table S7). Moreover, at the decision threshold of 0.4, the utLIFE-PC could result in up to 51% fewer unnecessary biopsies in suspicious prostate males with gray-zone PSA levels (4–10 ng/mL) (Figure S2B). Additionally, the diagnostic specificity was also rigorously assessed using an additional validation cohort of 22 GUC patients, in which the specificity reached 90.90% (Table S5). Identification of clinically significant cancer It is of great significance to develop an accurate test to detect clinically significant PCa (GS ≥ 3 + 4, csPCa), and therefore, we further evaluated whether utLIFE-PC could be used to identify csPCa from benign patients. In the training cohort, utLIFE-PC could discriminate 86.21% (95% CI, 77.15%–92.67%) of patients with ≥GG2 and 87.01% (95% CI, 77.41%–93.59%) of patients with ≥GG3 disease at 95.00% (95% CI, 89.97%–97.97%) specificity (Table S8). Moreover, the assay still exhibited a robust performance in the validation cohort, with an AUC of 0.934 (95% CI: 0.902–0.966), sensitivity of 85.33% (95% CI: 78.64%–90.57%), specificity of 93.26% (95% CI: 88.52%–96.47%), PPV of 90.44% (95% CI: 84.52%–94.26%), and NPV of 89.48% (95% CI: 85.23%–92.61%) in detecting ≥GG2 PCa (Figure 4E; Table 3) and with an AUC of 0.938 (95% CI: 0.902–0.973), sensitivity of 87.10% (95% CI: 79.89%–92.44%), specificity of 93.26% (95% CI: 88.52%–96.47%), PPV of 90.62% (95% CI: 84.78%–94.37%), and NPV of 90.63% (95% CI: 85.93%–93.86%) in detecting ≥GG3 PCa (Figure 4G; Table S9). Additionally, the good performance of utLIFE-PC was also observed in distinguishing ≥GG2 PCa patients from GG1 PCa patients plus BxN patients with an AUC of 0.919 (95% CI: 0.885–0.953), sensitivity of 85.33% (95% CI: 78.64%–90.57%), specificity of 88.08% (95% CI: 82.66%–92.29%), PPV of 84.77% (95% CI: 79.04%–89.15%), and NPV of 88.54% (95% CI: 83.96%–91.94%) in detecting ≥GG2 PCa (Figure S3A; Table S10). Moreover, the performance of utLIFE-PC in detecting ≥GG3 PCa patients from GG2, GG1, and BxN patients was analyzed, with an AUC of 0.893 (95% CI: 0.853–0.933), sensitivity of 87.10% (95% CI: 79.89%–92.44%), specificity of 80.37% (95% CI: 74.48%–85.41%), PPV of 71.52% (95% CI: 65.58%–76.81%), and NPV of 91.67% (95% CI: 87.39%–94.58%) (Figure S3C; Table S11). As expected, in the DCA analysis, utLIFE-PC still exhibited a higher net benefit compared with the separate features and PSA in patients with ≥GG2 (Figures 4F and S3B) and patients with ≥GG3 of the validation cohort (Figures 4H and S3D). At a decision threshold of 0.4, utLIFE-PC could result in up to 41% fewer unnecessary biopsies of initial biopsy men in csPCa (≥GG2) (Figure S2C), while in ≥GG3, the same decision threshold of 0.4 could save 49% of men from receiving an unnecessary biopsy (Figure S2D). Table 3The clinical performance of utLIFE-PC in different PSA levels of high-grade cancer (≥GG2) in the validation cohortAllPSA < 44 ≤ PSA ≤ 10PSA > 10–≥GG2BxN≥GG2BxN≥GG2BxN≥GG2BxNPredict positive1281211347934Predict negative221660117711584Sensitivity (%)85.33 (78.64–90.57)–100 (2.50–100.00)–82.93 (67.94–92.85)–86.11 (78.13–92.01)–Specificity (%)93.26 (88.52–96.47)–91.67 (61.52–99.79)–91.03 (82.38–96.32)–95.46 (88.77–98.75)–AUC0.934 (0.902–0.966)–––0.888 (0.802–0.974)–0.959 (0.930–0.987)–Positive likelihood ratio12.65 (7.30–21.95)–12.00 (1.84–78.37)–9.24 (4.50–18.99)–18.94 (7.25–49.50)–Negative likelihood ratio0.16 (0.11–0.23)–0–0.19 (0.10–0.37)–0.15 (0.09–0.23)–Disease prevalence0.4278–0.1854–0.2581–0.5296–PPV (%)90.44 (84.52–94.26)–73.20 (29.49–94.69)–76.2861.01–86.86)–95.52 (89.08–98.24)–NPV (%)89.48 (85.23–92.61)–100.00 (71.51–100.00)–93.87 (88.61–96.79)–85.93 (79.21–90.73)–Accuracy (%)89.87 (86.08–92.91)–93.21 (65.17–99.89)–88.94 (81.88–93.95)–90.51 (85.51–94.22)– The clinical performance of utLIFE-PC in different PSA levels of high-grade cancer (≥GG2) in the validation cohort Clinical utility of the utLIFE-PC model The distribution of risk scores derived from the utLIFE-PC model in the validation cohort is shown with the International Society of Urological Pathology GG of each patient (Figure 4I). The utLIFE-PC model could detect PCa with a positive rate of 84.24%. Among the 124 patients with GG3 or higher, utLIFE-PC could correctly identify 87.10% of these patients. The utLIFE-PC model could also detect 73.33% of the patients with GG1 and 76.92% of the patients with GG2, showing a higher classification efficiency with increasing grade. Furthermore, the utLIFE-PC model could avoid 55.98% (192/343) of prostate biopsies, providing another cue for the decision of whether to perform a biopsy. Discussion In this study, we implemented the utLIFE algorithm in PCa to discriminate any grade of PCa for initial biopsy populations. The utLIFE-PC achieved the highest AUC compared with any other separate feature and showed the highest sensitivity and specificity in detecting PCa, especially in PCa with GS ≥4 + 3. Additionally, the performance of the utLIFE-PC model was stable across patients with different PSA levels, even in the population with the so-called gray zone of PSA level in the 4.0–10.0 ng/mL range, who generally has a relatively low PCa detection rate.9 PCa exhibits a pronounced reshaping of its genomic landscape when it progresses through advanced tumorigenic stages.10 Mutations are screened by using local PCa, indicating the mutation feature was more sensitive in early-stage PCa. Conversely, methylation alterations, particularly those with diagnostic and prognostic implications, tend to become more pronounced and discriminatory in late-stage PCa, specifically in castration-resistant prostate cancer (CRPC).11 It demonstrated the enhanced discriminatory power of specific methylation features in distinguishing CRPC from earlier stages or more responsive forms of PCa. Hoeck et al. reported that in prostate carcinomas, significantly increased levels of genomic instability were induced in later evolutionary stages,10 which means the CNV feature may be easier to detect in late-stage PCa. These findings underscore the complementary nature of genetic and methylation profile changes across the progression spectrum of the disease, which also explains why multidimensional detection is needed in PCa detection. The methylation and mutation features were detected in both Chinese PCa samples and the TCGA or ClinVar/ANNOVAR database, indicating that these features could be widely used in different ethnic groups. In our training cohort, healthy men with PSA less than 4 ng/mL (standard PSA cutoff with 93.6% specificity) and men who underwent an initial biopsy that was negative or were diagnosed with other GUCs were used as controls. This ensured that all the features selected were PCa specific. In our initial biopsy prospective validation cohort, the utLIFE-PC model could accurately discriminate PCa of any grade and additional PSA levels with a high PPV and NPV, confirming the robustness and reliability of the model. Currently, most cases of PCa are diagnosed by prostate biopsy. The prostate biopsy decision process is multifactorial and encompasses the integration of clinical variables, including past PSA levels, DRE exam, family history, MRI-PRAD, computed tomography, ExoDx Prostate (IntelliScore) (EPI), 4K-score test, Prostate Health Index (PHI) test, and patient anxiety. Of all noninvasive methods for PCa diagnosis, PSA with the standard cutoff of 4.0 ng/mL (sensitivity 74%, specificity 44%),12 EPI (sensitivity 91%, specificity 34%), the 4K-score test (sensitivity 81%–95%, specificity 28%–62%), SelectMDx (sensitivity 94%, specificity 46%), and PHI (sensitivity 96%, specificity 36%)13 all have unsatisfactory specificity in initial biopsy patients. Unnecessary biopsies remain a burden in clinical practice. Prostate biopsy could result in more complications, such as postoperative infection, hematuria, and hematochezia, which cause suffering to patients and make the treatment expensive and time-consuming.14,15 The risk score of the utLIFE-PC model was strongly correlated with an increased risk for PCa, which could provide an individual risk assessment of PCa at the initial biopsy. In the 343 men undergoing an initial biopsy, the utLIFE-PC model showed outstanding sensitivity (84.24%; 95% CI, 77.77%–89.44%) and specificity (93.26%; 95% CI, 88.25%–96.47%) in distinguishing PCa patients from benign patients. Furthermore, it can be utilized as an effective and noninvasive method to detect clinically significant cancer to aid in treatment decisions and active cancer surveillance. Among patients with a PSA level of 4.0–10.0 ng/mL, the utLIFE-PC model showed comparable sensitivity in detecting GS ≥3 + 4 with the EPI test (82.93% vs. 87%) and avoided more biopsies than the EPI test (58.49% vs. 27%). The utLIFE-PC model demonstrated an overall net benefit when compared with PSA across a wide range of decision threshold probabilities in DCA analysis, showing good extensibility. Thus, it can be used in different clinical settings. At an alternative cutoff (0.545), the utLIFE-PC model achieved a 99.44% specificity and a 99.22% PPV with a sensitivity of 76.97% in the validation cohort. For patients with lower utLIFE-PC scores, there is a smaller chance of positive PCa findings on biopsy, and a close follow-up may be sufficient. For patients with higher scores, stronger evidence is provided for surgeons to make a surgical decision, but a large population study is still needed. In clinical application, the threshold for utLIFE-PC will be decided based on a shared decision-making process between the patient and the urologic surgeon, which requires validation in an intended-use population from the real world. Notably, the percentage of PCa patients with GG1 was lower in some Chinese cohorts (6.75%–9.85%) than in Caucasian cohorts (11.59%–31.84%). Another retrospective study compared 2,343 PCa patients in a Chinese cohort and 2,626 PCa patients in a Canadian cohort, finding more GS 8–10 PCa patients in the Chinese cohort (38% vs. 10%, p < 0.001).16 Consistently, the proportion of GG1 patients was 4.37%, and 36.15% were GG ≥3 patients in our study. These differences may be caused by distinct risk factors between China and Western countries. The high percentage of GS 8–10 PCa also led to a high MR/IR in China,17 indicating the urgent need for more precise measures for PCa screening in China. The utLIFE-PC model showed outstanding sensitivity (84.24%) in detecting PCa, especially in GG ≥3 patients (87.10%). Thus, it provides an accurate method for detecting PCa. The consistency of the comprehensive analyses in our study documents the robustness, reproducibility, and reliability of utLIFE-PC model. In conclusion, our data suggest that this model is useful for the noninvasive and efficient detection of individuals at risk for PCa with different PSA levels. The utLIFE-PC model is a promising method to detect PCa, showing great potential in improving the PCa diagnostic process and decreasing unnecessary biopsies in a convenient and noninvasive way. Limitations of the study There are potential limitations to our study. As patients with PSA >10 ng/mL are likely to proceed with biopsy according to guidelines, the study design should validate in a more needed biomarker population with PSA 4–10 ng/mL. The number of GG1 and GG2 patients in the validation dataset was small, and the age of the control group in the training cohort was lower than that of the case group. A larger and prospective population with PSA range of 4–10 ng/mL will be performed in the future. Resource availability Lead contact Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Nianzeng Xing (nianzeng2006@vip.sina.com). Materials availability This study did not generate new unique reagents. Data and code availability The custom code for training and validation models was written in Python with Numpy, Matplotlib, and Seaborn packages. The code is available in Zenodo (https://doi.org/10.5281/zenodo.13986239). The data are available in dbGAP repository under accession number phs003836.v1.p1 (http://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs003836.v1.p1). Any additional information required to reanalyze the data reported in this work is available from the lead contact upon request. The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request. Acknowledgments We thank all the participants for their contribution to this study. This study was funded by National Key Research and Development Program of China (grant no. 2022YFE0200800). Author contributions Conceptualization, N.X., Y.N., D.Y., S.C., and F.L.; resource, S.H., Mingshuai Wang, Y.W., J.W., R.L., L.H., X.Q., J.B., W.Y., H.A., G.Z., Y. Zhi, Z.C., and Y. Zhang; investigation, Z.G. and H.W.; methodology, L.C., L.L., S.Z., and Meng Wang; writing – original draft, S.H., Mingshuai Wang, Y.W., J.W., Z.G., and H.W.; writing – review and editing, N.X., Y.N., D.Y., S.C., F.L., and H.C.; visualization, Meng Wang, Z.G., and X.L.; supervision, N.X., Y.N., D.Y., S.C., and F.L.; funding acquisition, N.X. Declaration of interests Z.G., H.W., L.C., L.L., H.C., S.Z., M.W., Y. Zhang, and X.L. are employees of Acornmed Biotechnology Co., Ltd. F.L. and S.C. are founders of Acornmed Biotechnology Co., Ltd. and members of its scientific advisory board. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERCritical commercial assaysEZ DNA Methylation-Lightning KitZYMO ResearchCat#D5031Zymo-Spin IC columnZymo ResearchCat#C1004EpiTect MethyLight PCR +ROX Vial KitQiagenCat#59496KAPA HyperPrep KitRocheCat#KK8500urine DNA Storage TubeCWBIOCat# CW2657MMagMAX™ Cell-Free DNA Isolation KitThermo Fisher ScientificCat#A29319TIANamp Genomic DNA KitTianGenCat#4992199Deposited dataAnalyzed dataThis paperphs003836.v1.p1 (http://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs003836.v1.p1)Software and algorithmsutLIFE-PC modelThis paperZenodo: https://doi.org/10.5281/zenodo.13986239FASTP (v0.19.1)Chen et al.18RRID:SCR_016962BWA (v0.7.12)Aligning sequence reads, clone sequences and assembly contigs with BWA-MEM. arXiv:1303.3997v2RRID:SCR_010910Picard–https://broadinstitute.github.io/picard/WisecondorX (v 0.2.1)Reman et al.19N/Amachine learning package–https://scikit-learn.org/stable/index.htmlNumpy package (v 1.26)–https://numpy.org/Matplotlib (v3.82)–https://matplotlib.org/stable/index.htmlSeaborn–https://seaborn.pydata.org/index.htmlANNOVAR (20160201)Wang et al.20RRID:SCR_012821Genome Analysis Toolkit (GATK; v3.8)–https://gatk.broadinstitute.org/hc/en-usOtherMicro DropBIODLCat#58601001Qubit 4.0Thermo Fisher ScientificCat# Q33238Novaseq 6000IlluminaCat# 200128507500Fast ABI systemThermo FisherCat#4351104A non-invasive, multi-dimensional study of early detection of prostate cancer with urinehttps://www.chictr.org.cn/bin/project/edit?pid=194060(ChiCTR2300071837) Experimental model and study participant details Study population A multicenter cohort (https://www.chictr.org.cn/bin/project/edit?pid=194060) study consists 801 participants (362 participants in the first stage and 439 participants in the second stage) which were recruited from 11 sites in China from April 26, 2021, to March 30, 2023.was used for model construction (a case-control study, n = 362) and validation (a prospective study, n = 439). Voided urine samples were meticulously collected from male participants across 11 sites in China, from April 26, 2021, to March 30, 2023. These participants were candidates for initial prostate biopsies due to presenting with PSA levels exceeding 4.0 ng/mL or exhibiting PSA levels under 4.0 ng/mL coupled with abnormal findings from digital rectal examinations (DRE). Additionally, the cohort included healthy males as controls. The inclusion criteria were as follows: (1) for healthy males: PSA <4.0 ng/mL and negative DRE; no suspicious tumoral lesion in the GU tract and prostate by ultrasound; and provided physical examination reports and urine sample; (2) for males who underwent an initial biopsy: PSA ≥4.0 ng/mL, positive DRE or suspicious computed tomography (CT) scan/magnetic resonance imaging (MRI)/Prostate Health Index (PHI); scheduled for an initial prostate needle biopsy; and provided a confirmed pathological result by two experienced pathologists. The exclusion criteria were as follows: history of prior genitourinary cancer (GUC) and history of invasive treatments for benign prostatic hypertrophy or lower urinary tract symptoms within 6 months of study enrollment. Participants who were healthy during these routine health checkups were enrolled in the analysis. The males who underwent an initial biopsy who met the preliminary assessment received subsequent following prostate biopsy and were divided into 3 groups: PCa, biopsy negative (BxN), or other GUC groups. All PCa patients were grouped based on the grading system recommended by the International Society of Urological Pathology (ISUP) as follows: GS2 - 6 = Grade Group (GG)1, GS3 + 4 = GG2, GS4 + 3 = GG3, GS8 = GG4 and GS9 -10 = GG5. All BxN patients were followed until at least 6 months. All centers are level A tertiary hospitals in China, ensuring high-quality physical examination or pathology review. Approximately 100 mL of non-DRE voided urine samples were collected before any examination or biopsy. Samples which didn’t pass quality control (QC, insufficient extraction amount of DNA, sequence failure, or qMSP failure) were eliminated as technical failure. Written informed consent was obtained from each patient. The study was approved by the Ethics Committee of each center. Method details DNA extraction and quality control Urine supernatant was collected using a urine DNA Storage Tube (CWBIO) for urinary cell-free DNA (ucfDNA) extraction. The remaining urine was also collected using a sterile tube containing urine conditioning buffer (UCB, ZYMO) for exfoliated cell DNA (uexDNA) extraction. All the samples were transported to the laboratory at 2°C–8°C within 72 h. Urine supernatants were centrifuged at 1600 × g and 16000 × g for 10 min separately, while urine sediment was centrifuged at 1600 × g for 10 min and then stored at −80°C until DNA extraction. Quality control was measured by Micro Drop (BIO-DL) and Qubit 4.0 (Thermo Fisher Scientific) devices. Library preparation and sequencing At least 30 ng of ucfDNA were used for NGS libraries construction using KAPA HyperPrep Kit (Roche). Target enrichment was performed with a custom design panel (Twist Bioscience) which covered whole exons of 808 cancer-related genes. Whole genome sequencing (WGS) and target enrichment sequencing were performed on Novaseq 6000 (Illumina) to generate 150bp paired-end reads. All experimental steps were carried out following the manufacturer’s instructions. Bisulfite treatment and qMSP analysis A total of 30 ng uexDNA were used for quantitative methylation special PCR (qMSP) analysis. The bisulfite treatment of uexDNA was carried out by EZ DNA Methylation-Lightning Kit (Catalog Number: D5031, ZYMO Research, CA, USA) according to the manufacturer’s instructions. The bisulfite-treated uexDNA was used immediately for qMSP analysis after elution. The EpiTect MethyLight PCR + ROX Vial Kit was used to analyze DNA methylation with 4μL bisulfite-treated DNA according to the manufacturer’s instructions. The reactions were performed in the 7500Fast ABI system (Applied Biosystems, Foster City, CA, USA). The qMSP assays were used to detect the methylation levels of GSTP1 and APC, and ACTB was defined as an internal reference. Cohort definition In all the 801 participants, 171 cases (61 without biopsy results, 101 failed QC, 9 other cancers) were excluded according to the exclusion criteria, or sample QC. For model construction, 362 participants (128 healthy participants and 234 initial biopsy participants) were recruited. After excluding 32 initial biopsy participants without subsequent following biopsies and 43 urine samples that failed QC, 50 healthy patients were used in baseline construction, 97 PCa patients, 66 BxN patients, 10 other GUC patients, and 128 64 healthy participants were enrolled in a case‒control cohort for algorithm development (Figure 1). While in the validation cohort, 439 initial biopsy patients were recruited prospectively. After excluding 29 participants without subsequent following biopsy results, 9 with other GUCs, and 48 58 failed QC, 343 males who underwent an initial biopsy (PCa, n = 165; BxN, n = 178) were included in an independent prospective validation cohort (Figure 1). All the biopsy pathological examinations in the validation cohort were blinded to the test results. Sample size calculation An alternative area under the receiver operating characteristic (ROC) curve (AUC) of at least 0.90 was expected, which estimated a difference of 0.1 from the current methods (AUC ≤0.80).12,13,21,22,23 We compared AUCs under the null hypothesis of 0.8 and an AUC under the alternative hypothesis of 0.9 using PASS 2021, v21.0.324,25. The power for the alternative AUC is calculated based on a two-sided z test with a significance level of 5%. With a sample size of 97 in the case group and 140 in the control group, the training cohort will have more than 90% power to detect a difference of 0.1 in the current methods. Mutation biomarker screening We firstly identified mutation genes from 824 tissue samples of primary PCa in Chinese population. Fastq (version 0.19.1) was used to quality control statistics and preprocessing of raw sequence data. Burrows–Wheeler Aligner software (BWA, version 0.7.12) were used to aligned reads to the hg19 version of the human genome. The MarkDuplicates tool in Picard PCR was used to mark the duplicates. IndelRealigner and BaseRecalibrator in the Genome Analysis Toolkit (GATK; version 3.8) were used for realignment and recalibration of the BWA alignment results, respectively. Somatic mutations were were detected by Sentieon TNhaplotyper. Softclip type bases and bases with sequencing quality less than 20, mutations with PopFreqMax >0.001 in GnomAD or ExAC_EAS, mutant allele frequency (MAF) > 0.01, and allele base depth <5 was filtered out. Truncating mutations in tumor suppressor genes and pathogenic/likely pathogenic point mutations annotated with ANNOVAR and Clinvar database were selected. These genes were then filtered in the urine samples of the training cohort. Mutation genes shared in tissue and urine samples were chosen as candidate markers. Finally, 138 genes were selected for model construction. Methylation biomarker screening We first compared 10 paired (PCa and BxN) sets of targeted methylation sequencing data from in-house urine samples. Then, we compared the methylation data derived from 482 PCa and 49 adjacent tissue samples from the cancer genome atlas (TCGA)TCGA. The following filters were used to screen the differentially methylated regions (DMRs) between cancer samples and control samples: fold change >4, FDR <0.05; median difference >0.4, cancer median >0.4, normal median <0.1 (TCGA data) or median difference >0.09, cancer median >0.05, normal median <0.005 (in-house data). The PCa-related DMRs (APC and GSTP1) that were shared in the urine samples and TCGA (APC and GSTP1) were chosen for further analysis (Figure S1). CNV biomarker screening The CNV biomarkers were extracted from WGS data (median depth 7×) of ucfDNA. The CNV Z score was standardized and calculated in 1M bins of each sample by WisecondorX software (version 0.2.1). The sensitivity of each genome region was calculated at 95% specificity. Nineteen genome regions with a sensitivity greater than 20% were chosen for further analysis. utLIFE-PC algorithm development The datasets of methylation, mutation, and CNV features were integrated and evaluated together through a random forest (RF) machine learning model. A 10-fold cross-validation procedure was used to estimate model hyperparameters in the training cohort. The independent validation cohort was used to verify the utLIFE-PC algorithm. ROC and decision curve analysis (DCA) were used to assess the performance of utLIFE-PC, including AUC, accuracy, sensitivity, specificity, and net benefit. Quantification and Statistical analysis The Chinese Prostate Cancer Consortium (CPCC) data from 19124 individuals were used to estimate the PCa distribution at different PSA levels.3,26 The utLIFE-PC algorithm was developed by the Numpy, Matplotlib, and Seaborn packages of Python 3.8.0. The performance of utLIFE-PC was calculated by sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), with 95% CIs. The ROC was constructed by using the pROC package of R 4.0.4, and the delong test was used for AUC significance testing. The DCA was constructed by using the rmda package of R 4.0.4. The risk factors were analyzed by univariate analysis. The posttest probabilities were calculated by using positive and negative likelihood ratios for PCa at different PSA levels from pretest probabilities based on the prevalence reported. Fisher’s exact test, Student’s t test or Mann Whitney U test was performed with GraphPad Version 8.0. Data are represented as mean ± SD. A p-value ≤ 0.05 was considered statistically significant (∗). A p-value ≤ 0.01 was considered highly statistically significant (∗∗).A p-value ≤ 0.001 was considered extremely statistically significant (∗∗∗). A p-value ≤ 0.0001 was considered super extremely statistically significant (∗∗∗∗). Additional resources This study is registered on the Chinese Clinical Trials website (https://www.chictr.org.cn/bin/project/edit?pid=194060, ID: ChiCTR2300071837). Published: December 9, 2024 Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.xcrm.2024.101870. Contributor Information Feng Lou, Email: fenglou@acornmed.com. Shanbo Cao, Email: shanbocao@acornmed.com. Dingwei Ye, Email: dwyeli@163.com. Yuanjie Niu, Email: niuyuanjie9317@163.com. Nianzeng Xing, Email: nianzeng2006@vip.sina.com. Supplemental information Document S1. Figures S1–S3 and Tables S1–S11 Document S2. Article plus supplemental information References Associated Data Supplementary Materials Document S1. Figures S1–S3 and Tables S1–S11 Document S2. Article plus supplemental information Data Availability Statement The custom code for training and validation models was written in Python with Numpy, Matplotlib, and Seaborn packages. The code is available in Zenodo (https://doi.org/10.5281/zenodo.13986239). The data are available in dbGAP repository under accession number phs003836.v1.p1 (http://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs003836.v1.p1). Any additional information required to reanalyze the data reported in this work is available from the lead contact upon request. The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.