A novel diagnostic model for early-stage high-grade serous ovarian cancer based on serum tRNA-derived fragments

Sample collection

Plasma samples were prospectively collected from patients prior to any therapy at the First Affiliated Hospital of Sun Yat-sen University between May 2023 and July 2024. Detailed criteria for participant inclusion and exclusion are provided in Supplementary Materials. All participants provided written informed consent, and the study was conducted in accordance with the principles of the Declaration of Helsinki. The study protocol was approved by the Institutional Review Boards of all participating institutions.

A total of 158 samples were obtained from the First Affiliated Hospital of Sun Yat-sen University. These samples were randomly allocated into the following cohorts: a sequencing cohort (n = 15; including 5 normal controls [NC], 5 early-stage ovarian cancer [EOC], and 5 late-stage ovarian cancer [LOC]), a discovery cohort (n = 30; 10 NC, 10 EOC, 10 LOC), a model construction cohort (n = 93; 44 NC, 49 OC), and an external validation cohort (n = 20; 10 NC, 10 OC). Additionally, 60 independent samples were collected from two external medical centers: 30 from Xinhua Hospital Affiliated to Shanghai Jiaotong University School of Medicine (10 NC, 20 OC) and 30 from Hangzhou People’s Hospital (10 NC, 20 OC), to further validate the model for multi-center validation. There were no significant differences in key clinical parameters among the different cohorts, as confirmed by appropriate statistical tests (p > 0.05), ensuring the comparability of groups for subsequent analyses. The patient selection process is presented in Figure 1.OPEN IN VIEWER

RNA sequencing data were obtained from The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov/) database for HGSOC patients (n=422) and the fallopian tube RNA sequencing of normal controls (n=180) was acquired from a professional website-The Genotype-Tissue Expression (GTEx, https://www.gtexportal.org/) portal. Every dataset from the TCGA and GTEx databases was available to the general public. And unnecessarily for ethics committee approval for those data.

Plasma RNA extraction

Peripheral venous blood samples (300 μL) were collected in EDTA anticoagulant tubes and immediately centrifuged at 1750 × g for 10 minutes at 4°C. The plasma was then aseptically aliquoted and stored at -80°C. RNA extraction was performed using the TRIzol method, specifically targeting small RNAs. Plasma was mixed with three times the volume of TRIzol, vortexed immediately, and after standing for a brief period, 1/5 volume of chloroform was added and vortexed again. The mixture was left at room temperature for 3-5 minutes, followed by centrifugation at 16000g for 20 minutes at 4°C. The supernatant was collected and mixed with two volumes of isopropanol, and RNA was precipitated at -20°C for 12 hours. After another round of centrifugation at 16000 × g for 20 minutes at 4°C, the supernatant was discarded, and 1 mL of 75% ethanol was added, followed by centrifugation at 16000 × g for 20 minutes at 4°C. The liquid was decanted, and the RNA was dissolved in an appropriate volume of DEPC-treated water. The RNA concentration was adjusted to 400 ng/μL using a Nanodrop One (Thermo Fisher Scientific, USA).

Sequencing and primer design for tsRNAs

Sequencing was performed using PANDORA-seq, which employs a combinatorial enzymatic treatment to remove key RNA modifications that block adapter ligation and reverse transcription in sncRNA. The sequencing was conducted by Guangzhou Epigenetics Co., Ltd. ²⁰ The comparison and identification of mature or precursor tRNA sequences were based on the small RNA annotation software SPORTS1.1 (available on GitHub at https://github.com/junchaoshi/sports1.1). Primer design and synthesis were carried out by Guangzhou RiboBio Co., Ltd., using the stem-loop method. This technique is covered by the company’s patent, and as per the company’s policy, the primer sequences are not disclosed.

Quantitative reverse transcription PCR experiment

Each RNA sample was reverse-transcribed into cDNA using the Bulge-loop™ qRT-PCR primer set (RiboBio Co., Ltd., Guangzhou, China), specifically designed for tsRNAs according to the manufacturer’s instructions. qRT-PCR was conducted on a real-time fluorescence quantitative PCR instrument (Bio-Rad CFX Connect Real-Time PCR System 1855201, Bio-Rad Laboratories, Inc., USA) following the manufacturer’s protocol with universal reagents. The relative expression levels were normalized using RNU6B (U6) (RiboBio, Guangzhou, China) as an internal control. The comparative expression level was evaluated by 2^^-ΔΔCt method.

Statistical analysis

Sequencing data differential analysis, model construction, and plotting were performed using R software (version 4.3.2; https://www.Rproject.org) and its associated packages. For analysis of biomarkers and construction of diagnostic model, logistic regression algorithm was applied and calculated by the formula: Logit(P) = b₀ + b₁*ΔCt₁ + b₂*ΔCt₂ … + b_n*ΔCt_n, where the b_n meant the nth regression coefficients and the ΔCt_i meant the relative cycle threshold of each biomarker.^21,22 Statistical analysis of ΔCt and graphing were conducted using SPSS 26.0 and Origin 2023. For normally distributed data, differences were analyzed using the Student's t-test, while for non-normally distributed data, the Mann–Whitney test was used. A fold change threshold of 2 was set, and P < 0.05 was considered statistically significant. This partitioning of training and validation subsets was performed using the “createDataPartition” function from the “caret” package in R.

tsRNAs of different abundance in plasma were screened out between HGSOC and healthy controls

To preliminarily assess the potential differences in tsRNA expression between HGSOC and normal fallopian tube epithelial tissues, we analyzed sequencing data from the GTEx (normal tissues) and TCGA (HGSOC tissues) databases. Analysis revealed differential expression of tsRNA-related genes, among which QTRT1 and SLFN13 showed significant differences in gene expression (Figure 2(A)). Based on this, we hypothesized that there may be differences in tsRNA levels between the plasma of HGSOC patients and that of healthy individuals. Consequently, we collected pre-treatment patients’ plasma from 5 normal controls (NC), 5 early-stage HGSOC (EOC), and 5 late-stage HGSOC (LOC) from our hospital. After RNA extraction and high-throughput sequencing, Uniform Manifold Approximation and Projection (UMAP) dimensionality reduction analysis was conducted to evaluate inter- and intra-group variability. Based on proximity in reduced-dimensional space, 4 NC, 4 EOC, and 4 LOC samples were selected for further analysis (Figure 2(B)).OPEN IN VIEWER

Based on the results of the differential analysis, we identified 15 fragments with statistically significant differences between NC and EOC (Figure 2(C)), 97 fragments between EOC and LOC (Figure 2(D)), and 166 fragments between NC and LOC (Figure 2(E)). To enhance diagnostic consistency across different clinical stages and compensate for the limited sample size of EOC specimens in subsequent large-scale modeling, for biomarker screening we selected fragments that exhibited statistically significant differences between the control group and early-stage patients, but no significant differences between early-stage and late-stage patients. A total of 8 fragments met these criteria.

Biomarkers were identified from the tsRNAs that met the specified criteria

Given the accessibility and widespread use of biomarker detection, further screening of biomarkers for model construction was performed using qRT-PCR technology. Due to the limitations of current techniques, only four fragments were suitable for specific primer design. These fragments are as follows:

1. tsRNA-5003c-GlyGCC; tsRNA-5004c-GlyGCC/GlyCCC (labeled as 3c1)

The sequences of the fragments are provided in Supplementary Material.

For preliminary screening, we selected 10 NC and 10 EOC samples. The fragment labeled 3c1 did not follow the same trend observed in the sequencing data within the screening set. Although the fragment labeled 3c2 showed consistency with the sequencing results in plasma samples, the cycle threshold (Ct) values in qRT-PCR were greater than 33, indicating suboptimal reverse transcription feasibility. Fragments 3c3 and 20b demonstrated consistency with the sequencing data in plasma samples, along with good reverse transcription feasibility (Figure 3(A)). Following the designing logic for the target discovery process, we further included 10 LOC samples from the discovery cohort to evaluate the enrichment levels of the 3c3 and 20b fragments in the plasma of late-stage patients. The 3c3 fragment continued to exhibit the expected trend in LOC samples, whereas the 20b fragment did not meet the anticipated levels (Figure 3(B)). In a larger-scale validation (n = 93), 3c3 showed statistically significant abundance differences between NC and LOC samples, while 20b did not (Figure 3(C)). Consequently, we selected the 3c3 as the biomarker for model construction.OPEN IN VIEWER

A diagnostic model was constructed based on a large sample size

We used the ΔCt values from qRT-PCR of plasma samples of NC and HGSOC patients for model construction. Given that the independent variable is a normally distributed continuous variable and the outcome is a binary variable, a logistic regression algorithm was applied. The modeling dataset was randomly divided, with 2/3 used as the training set and 1/3 as the internal validation set for model development. Ultimately, a model was successfully constructed based on the maximum area under the curve (AUC) of the receiver operating characteristic (ROC) curve from the training dataset (Figure 3(D)). The formula of derived model was as follows: Logit(P) = -0.066021143-0.550878495*ΔCt_(3c3). According to the calculations, the AUC of the ROC curve was 0.859 for the training set and 0.832 for the internal validation set (Figure 3(E)), indicating that the model demonstrated good performance in both sets.

The model was further validated and exhibited excellent performance in early diagnosis of OC

We further tested the model in the whole model construction set and the overall AUC for the model construction set was 0.854 (Figure 3(F)). Additionally, the model exhibited strong performance across various indicators (including Sensitivity, Specificity, Positive Predictive Value, Negative Predictive Value and Accuracy) used to evaluate biomarker efficacy in the training set, internal validation set, and the overall model construction set (Table 1). To further assess the accuracy of the model, we prospectively collected another set of pre-treatment plasma samples from NC patients and those clinically diagnosed with EOC post-surgery at our center for external validation. The AUC for this external validation was 0.860 (Figure 3(G)). To compare the diagnostic performance of this model with other commonly used screening methods for EOC, specifically, “Risk of Ovarian Malignancy Algorithm (ROMA)” and ultrasound examination, we further collected clinical data from 10 NC and 10 EOC cases in the external validation. It turned out that the AUC of our model in external validation for early diagnosis was 0.860, higher than that of ROMA (AUC = 0.850) and ultrasound examination (AUC = 0.800) (Figure 4(A),4(B)).

OPEN IN VIEWER

Table 1.

Model performance across datasets.

​	SE	SP	PPV	NPV	Accuracy
Training Set	0.9062	0.7667	0.8056	0.8846	0.8387
Internal Validation Set	0.8824	0.6429	0.75	0.8182	0.7742
Overall Set	0.898	0.7273	0.7857	0.8649	0.8172

OPEN IN VIEWER

Figure 4.

Performance evaluation and detection of tumor tissues.

In addition, to evaluate the impacts of other clinical characteristics on the model’s score, we stratified the samples in the model construction set by different clinical parameters. The results indicated that there was no statistically significant difference in risk scores between these groups divided by age and menstrual status (Figure 4(C)). Furthermore, no significant difference was detected in risk scores across different clinical stages (Figure 4(D)). Taken together, these findings suggest that our model not only demonstrates strong performance in early diagnosis but is also independent of other clinical characteristics.

To evaluate the generalizability and diagnostic accuracy of our identified biomarker, we further validated the model using an independent multi-center cohort comprising 60 plasma samples collected from two additional hospitals (each providing 10 normal controls and 20 HGSOC patients). The validation results showed that the model achieved an AUC of 0.874 in the ROC analysis, demonstrating its robust generalizability and potential clinical utility for the early detection of ovarian cancer (Figure 4(E)).

The involved tsRNAs in the plasma are derived from cancerous tissue

To investigate whether the 3c3 fragment originates from cancerous tissue and to exclude the potential influence from high tumor burden, such as cachexia in advanced stages, we collected fallopian tube epithelial tissues from 10 NC patients and cancerous tissues from 10 EOC patients to perform qRT-PCR for tissue validation. The results showed that the abundance of the 3c3 fragment was significantly higher in EOC tissues compared to NC tissues (Figure 4(F)). This preliminary evidence suggests that the biomarker fragment in the diagnostic model is derived from tumor tissues.