Comparison of discovery rates and prognostic utility of [⁶⁸Ga]Ga-PSMA-11 PET/CT and circulating tumor DNA in prostate cancer—a cross-sectional study

In this retrospective single-center study with prospective sample collection conducted at the Medical University of Vienna (March 2019–August 2021), 187 men with confirmed PCa referred for [⁶⁸Ga]Ga-PSMA-11 PET/CT underwent PET/CT imaging and blood sample collection. An all-comer recruitment strategy was employed. All patients gave their written informed consent for imaging, blood sample collection, and associated analysis. This study was approved by the ethics committee of the Medical University of Vienna (ID: 1649/2016).

For this analysis, patients with histologically proven PCa, known PSA levels, and castration status were included, while patients with active or a history of concomitant malignancies other than PCa (N = 11), unknown PSA values (N = 31), and unknown castration status (N = 15) were excluded (Fig. 1).

Clinical data, such as PSA levels, castration status, and pre-, concurrent, and post-imaging therapy data, were gleaned retrospectively from the medical records. Follow-up and overall survival (OS) data (censorization 13th August 2023) were sourced from the National Health Statistical Service. The primary endpoints of this study were (a) ctDNA and PSMA PET/CT tumor signal discovery rates according to castration status and PSA levels, (b) the relationship of ctDNA concentrations and the PSMA-TV in all patients and according to their respective castration status, and (c) the prognostic value of ctDNA and PSMA-TV levels with regard to overall survival (OS).

Before tracer injection, blood samples were collected in Cell-Free DNA BCT tubes (Streck Inc., Nebraska, USA). cfDNA was extracted from the plasma using the QIAamp Circulating Nucleic Acid Kit (QIAGEN, Venlo, Netherlands), following the manufacturer’s procedure. cfDNA sample quantities were assessed using the Fragment Analyzer system and the HS NGS Fragment Kit (respectively Agilent, California, USA), as per the manufacturer’s guidelines. Next, Fragment Analyzer results were analyzed with the PROSize software (v2.0, Agilent, California, USA) for automatic DNA concentration calculation. cfDNA quantities are expressed as ng/µL.

DNA sequencing libraries were prepared from 19.5 µl of isolated DNA using the xGEN EZ UNI Library preparation kit combined with stubby adaptors (IDT, Iowa, USA, respectively) containing 3 bp random sequence used as UMI. Library PCR amplification was carried out with the xGEN EZ UNI Library preparation kit in combination with KAPA UDI primers (Roche, Switzerland). Samples were sequenced on NovaSeq 6000 (Illumina, California, USA) in a paired-end 2 × 60 bp setting.

We developed an in-house method to analyze the ctDNA fraction in blood samples using low-coverage WGS sequencing. Raw sequencing reads were initially mapped to the human genomic reference GRCh38 using the BWA tool [26]. Next, mapped raw sequencing reads were counted in 500 kb bin intervals. These bin counts underwent normalization based on sample size and GC content to address biases and variations. From this data, we determined an initial, approximate ctDNA fraction using the density plots of the bin sizes. To call CNVs, we employed an algorithm using the normalized binned read counts, incorporating a negative binomial distribution for individual bin counts to handle overdispersion similar to the ichorCNA methodology [20]. This was followed by using a dynamic Bayesian network model for holistic CNV predictions. The procedure was iterative, with CNVs re-called based on the updated ctDNA fraction and the ctDNA fractions recalculated using the new CNV predictions. In cases where no CNVs were discerned, we assigned the ctDNA fraction a default value of 0.05 and repeated the CNV calling. We measured the quality of our modelled CNVs and ctDNA by examining the residual difference between the actual bin sizes and the sizes predicted post-CNV and ctDNA adjustments. To test the significance of our predicted ctDNA for each sample, we compared residuals from our primary model to those from a noise model created using the same bin count data but with randomly permuted bins, employing the Kolmogorov–Smirnov Test for this purpose. Our final results excluded samples without called CNVs, those with a Kolmogorov–Smirnov Test P-value greater than 0.05, and samples predominantly predicting deletions around chromosome centromeres—a potential sign of an unidentified technical bias (Supplementary Material—Methods).

Patients were given an intravenous injection of 184.8 MBq (± 19.7 SD) of [⁶⁸Ga]Ga-PSMA-11 and scanned on Biograph TruePoint PET/CT scanner (Siemens Healthineers, Erlangen, Germany) from the skull base to the upper femur an hour post-injection. First, CT scans were taken at 120 kv and 230 mAs and intravenous contrast, except contraindications for contrast applications existed, followed by PET scan acquisition in 3–4 bed positions and iterative reconstruction.

Two nuclear medicine physicians analyzed the images on a dedicated workstation using the Hybrid 3D software (v4.0.0, Hermes Medical Solutions, Stockholm, Sweden), delineating and labelling all PSMA-expressing primary and secondary tumor lesions by their anatomical locations. Lesion identification was performed qualitatively, informed by liver uptake, followed by semiautomatic delineation using a region-growing algorithm (Hybrid 3D software, v4.0.0). The PSMA-TV and standardized uptake values (SUV) normalized to body weight were extracted both from an aggregated master lesion and per anatomic region. The anatomic tumor lesion region that contributed most to the overall PSMA-TV was defined as the dominant tumor fraction (Supplementary Material—Methods).

Continuous variables are reported as mean (± SD), and categorical outcomes as frequencies (%). ctDNA and PSMA PET tumor signal discovery rates and dominant fraction’s association with castration status, PSA ranges, and disease extent were evaluated using Chi-squared and Fisher’s tests. Non-normalized and PSMA-TV-normalized ctDNA concentrations were compared using the Kruskal–Wallis test after assessing normality and heteroskedasticity with Shapiro–Wilk and Levene’s tests. Post-hoc analysis was conducted using Dunn-Bonferoni’s test if the null hypothesis was rejected. The correlation between ctDNA concentrations and PSMA-TV was determined using Spearman’s coefficient. PSMA-TV’s predictive value for ctDNA discovery was analyzed using the area under the receiver-operating-characteristic curves (AUC). OS probabilities were estimated using Kaplan–Meier estimates, and differences in survival distributions were evaluated with the non-parametric Logrank test. The relationship between OS and ctDNA concentration and PSMA-TV was examined using multivariate Cox regression analysis after checking data for multicollinearity and proportional hazard with Belsley-Kuh-Welsch and Schoenfeld residuals. All analyses assumed a 5% alpha risk. All confidence intervals (CI) are 95% CIs. Statistical analyses were conducted using EasyMedStat software (v3.24, EasyMedStat, Paris, France) (Supplementary Material—Methods). For details on the exploratory machine learning analysis of imaging- and plasma-derived features in their predictive ability for OS in single and combined use, see Supplementary Material—Methods.

In total, 130 men with confirmed PCa (age 70.5 ± 8.0 years, PSA 96.35 ± 438.44) who underwent [⁶⁸Ga]Ga-PSMA-11 PET/CT imaging and plasma sample collection were analyzed. The demographic and clinical characteristics are presented in Table 1.

The PSMA PET and ctDNA analysis detected tumor signals (discovery rates) in 56.52% and 13.04%, 45.45% and 18.18%, 80.0% and 40.0%, and 88.89% and 35.8% of patients at the PSA ranges (0, 0.5), (0.5, 1.0), (1.0, 2.0), and (2.0, 3689.0) ng/mL, respectively (P < 0.001 and P = 0.123, respectively) (Table 2, Fig. 2A, B).

The PSMA PET discovery rates between hsPC (67.8%) and CRPC (87.32%) did differ significantly (odds ratio (OR) = 0.31; 95% CI = [0.13, 0.74]; P = 0.013), while the ctDNA discovery rates between castration statuses (hsPC, 25.42%; CRPC, 35.21%) did not (P = 0.311) (Supplementary Fig. 1A-B).

The ctDNA discovery rates were 14.81%, 15.00%, and 39.76% in patients with no lesions, localized, and metastatic disease on PSMA PET imaging, respectively (P = 0.013) (Fig. 2C).

Respective median levels of ctDNA according to PSMA PET imaged disease extent differed significantly (P = 0.006) and are displayed in Fig. 2E. Post-hoc adjusted, pairwise analyses revealed differences for the metastatic versus no lesion (P = 0.006, mean difference CI = [− 0.37, 0.1]) and metastatic versus localized groups (P = 0.029, CI = [− 0.4, 0.15]).

The dominant lesion fraction in patients with metastatic disease on imaging was in 6.02% prostate, 37.35% lymph node, 53.01% bone, and 3.61% organ (Fig. 2D). Respective median PSMA-TV-normalized ctDNA levels (normctDNA) according to the dominant fraction of PSMA PET differed significantly (P = 0.036) and are shown in Fig. 2F. Pairwise post-hoc analyses revealed only normctDNA differences for the prostate versus bone (P = 0.003, CI = [− 0.0021, 0.0051]) and prostate versus lymph node groups (P = 0.036, CI = [− 0.0006, 0.0053]) (Fig. 2F).

Significant and positive correlations were found between ctDNA levels and the PSMA-TV in all (r = 0.42, P < 0.001) and CRPC patients (r = 0.65, P < 0.001), while no significant correlation was observed in hsPC patients (r = 0.15, P = 0.249) (Fig. 3A, B).

In the overall cohort, the optimal threshold for PSMA-TV to predict ctDNA discovery was 19.66 [cm³] according to Youden [27] with an area under the curve (AUC) of the receiver operator curve (ROC) of 0.722 (CI = [0.618, 0.826]), while in metastatic patients, the optimal threshold was 34.88 [cm³] with an AUC of 0.719 (CI = [0.592, 0.846]) (Fig. 3C, D).

There were significant differences between survival distributions of the PSMA-TV and ctDNA high and low groups (P < 0.0001 and P < 0.0001, respectively) (Fig. 4A, B).

In the multivariate Cox regression analysis, there were significant hazard differences between the PSMA-TV (hazard ratio (HR) = 6.65, CI = [1.96, 22.52], P = 0.0023) and ctDNA high and low groups (HR 7.56, CI = [2.94, 19.42], P < 0.0001) (Table 3). For the results of the exploratory machine learning analysis of imaging- and plasma-derived features in their predictive ability for OS in single and combined use, see Supplemental Material—Results.