Plasma exosomal miR-150-3p, NMT2, and PRDM1 as predictive biomarkers of acute tumor response in patients with cervical cancer undergoing chemoradiotherapy

Abstract

Locally advanced cervical cancer (LACC) is primarily treated with weekly cisplatin-based concurrent chemoradiotherapy (CCRT); however, predicting acute tumor response remains challenging. This study aimed to identify plasma exosomal microRNAs (miRNAs) and messenger RNAs (mRNAs) that could predict rapid tumor regression in patients with LACC undergoing CCRT. Overall, 41 patients with stage IB-IVB cervical cancer were included. All patients received CCRT, and plasma exosomal RNA samples were collected before treatment and 2 weeks after radiation therapy (RT). Acute tumor response (AR) was defined as the regression rate of tumor volume (TV) (cm³) measured at the fourth week of treatment compared with the initial TV (iTV). The log₂ fold change of miRNA and mRNA was calculated by comparing RNA read counts before and after the second week of CCRT for each patient. A correlation matrix identified RNAs associated with AR. The selected RNAs were validated through linear regression and Wilcoxon rank-sum tests. Leave-one-out cross-validation was performed in subgroups based on iTV. miR-150-3p, NMT2, and PRDM1 were identified as key predictors of AR, demonstrating significant associations with immune-mediated tumor responses. A decrease in post-RT levels of these RNAs was significantly associated with poor AR, particularly in patients with large iTVs. The predictive model combining miR-150-3p, NMT2, and PRDM1 showed strong correlation with AR (R² = 0.831, P < 0.0001) in the test dataset and was validated in an independent cohort (R² = 0.496, P = 0.006). Cross-validation indicated the robustness of these biomarkers in predicting AR across varying TVs. These findings highlight the potential of plasma exosomal miR-150-3p, NMT2, and PRDM1 are promising biomarkers for predicting AR in patients with LACC undergoing CCRT. These findings could facilitate personalized RT strategies and improve patient outcomes. Further multicenter studies are warranted to validate these biomarkers in larger, diverse cohorts.

Introduction

Locally advanced cervical cancer (LACC) is commonly treated with weekly cisplatin-based concurrent chemoradiotherapy (CCRT), comprising external beam radiation therapy (EBRT) followed by intracavitary brachytherapy (ICBT) [1]. However, inadequate regression of large cervical tumors after EBRT can complicate the effective application of ICBT for residual disease eradication. Conversely, even with sufficient regression of primary tumors or metastatic lymph nodes, EBRT doses exceeding 45-50 Gy are often administered. This underscores the need for predictive markers of acute tumor response to tailor radiation therapy (RT) strategies in patients with cervical cancer.

As most cervical cancer cases are human papillomavirus (HPV)-positive, predicting tumor response to EBRT based on HPV status alone presents limited utility [2,3]. Therefore, identifying alternative biomarkers for predicting treatment response is crucial. Recent evidence suggests that combining pembrolizumab with CCRT significantly improves progression-free survival (PFS) in patients with LACC compared to placebo and CCRT alone (hazard ratio: 0.7, 95% confidence interval: 0.55-0.89, P = 0.002) [4]. This finding highlights the pivotal role of lymphocyte-mediated tumor eradication in treatment response, demonstrating the importance of the immune system in cervical cancer therapy.

Immune checkpoint inhibitors (ICIs) targeting cytotoxic T-lymphocyte associated protein-4, programmed cell death protein 1, or programmed death-ligand 1 have demonstrated delayed tumor responses in various cancers, necessitating evaluation criteria beyond standard radiologic assessments [5]. Additionally, some studies suggest that lymphocytes may influence acute tumor responses, as evidenced by sustained absolute lymphocyte counts (ALCs) in patients achieving pathologic complete response following neoadjuvant CCRT for rectal cancer [6,7]. Therefore, identifying biomarkers associated with immune response is essential for predicting tumor regression after RT.

Exosomes, extracellular vesicles measuring 40-160 nm in size, are secreted by both cancer and normal cells and are involved in intercellular communication and regulation of various physiological processes, including cancer progression, immune modulation, and inflammation [8,9]. Exosomal microRNAs (miRNAs) play a critical role in modulating inflammatory responses by regulating messenger RNAs (mRNAs) [10]. The abundance of immune cell-derived exosomes in blood enables the identification of treatment-related miRNAs and mRNAs through differential gene expression analysis of exosomal RNA before and after RT in individual patients [11,12]. By examining mRNAs associated with specific RNAs, the potential biological significance of exosomal RNA can be inferred [12].

In light of the points above, this study aimed to identify and validate biomarkers that predict acute tumor response to RT by analyzing post-RT plasma exosomal RNA levels associated with immune response in patients with cervical cancer.

Material and methods

Study participants

This study was approved by the Institutional Review Board of Ajou University Hospital (approval number: DB-2024-482), and the requirement for informed consent was waived owing to its retrospective design. Initially, 42 patients diagnosed with stage IB-IVB cervical cancer, according to the 2018 International Federation of Gynecology and Obstetrics (FIGO) criteria, were included. The inclusion criteria were receipt of weekly cisplatin-based CCRT and the availability of plasma exosomal RNA data collected from blood samples taken before and 2 weeks after RT (Figure S1). One patient was excluded owing to significant dissimilarity in miRNA expression in the post-RT sample, as observed in hierarchical clustering, resulting in 41 patients being included in the final cohort (Figures 1 and S2). The diagnosis was histologically confirmed by biopsy, and regional lymph node (LN) and distant metastases (DM) were assessed using magnetic resonance imaging (MRI) and positron emission tomography-computed tomography (PET-CT). Treatment commenced within 1 month following the imaging studies. EBRT was administered using intensity-modulated RT targeting the pelvis or para-aortic LNs. The pelvic RT dose was 45 Gy delivered in 25 fractions, with a simultaneous integrated boost of 120%-130% to regional LN metastases. RT response was evaluated via MRI during the fourth week of pelvic EBRT (36-45 Gy in 20-25 fractions). All patients received a weekly cisplatin regimen (30-70 mg/m²) for six cycles during RT. The details of three patients with one or two distant metastatic lesions who received EBRT in metastatic sites are described in the Supplementary Methods. Of the 41 patients, 39 underwent ICBT, with one patient substituting ICBT with an EBRT boost and another who declined further treatment (Iridium-192; Gamma Medplus iX; Varian, Palo Alto, CA, USA). After completing treatment, patients were followed up every 3 months. Evaluation of primary cervical tumors, regional LN, and DM included pelvic examination, pap smear tests, tumor markers, MRI, and CT.

Figure 1.

Flowchart illustrating the relationship between plasma exosomal RNAs and acute tumor response to radiation therapy in patients with cervical cancer.

Acute tumor response

Acute tumor response to RT (AR) was defined as the regression rate of tumor volume (TV) (cm³) measured at the fourth week of treatment compared with the initial TV (iTV) measured at diagnosis. The regression rate was calculated as

$$AR=\left(\frac{TV \left({\mathrm{cm}}^3\right) at 4th week}{TV \left({\mathrm{cm}}^3\right) at diagnosis}\right)$$

TV measurements were obtained using the Eclipse™ version 18.0 Treatment Planning System (Varian). TV was measured using MRI for the primary site, while LN metastatic lesions were assessed using CT or MRI (Figure S1). Measurement of TV was limited to pelvic and abdominal LNs.

Datasets

After excluding two patients with iTVs of < 4 cm³ from the 41 patients, the remaining 39 patients were divided into a test dataset (n = 19) and a validation dataset (n = 20) based on the chronological order of their diagnostic dates. Plasma exosomal RNAs were selected to predict AR in the test dataset through statistical and biological screening methods. The selected RNAs were validated in another dataset. Cross-validations of subgroups based on iTV were performed using these 39 patients. For survival analysis, 40 patients were included, excluding one who refused further treatment. This process is detailed in Figure 1.

Log₂ fold change

The log₂ fold change (Log₂FC) was calculated using next-generation sequencing data from RNAs, including both miRNAs and mRNAs, isolated from plasma exosomes isolated via a polyethylene glycol/dextran aqueous two-phase system (ATPS) (Exo2D; ExosomePlus, Seoul, Republic of Korea). Log₂FC values were determined by comparing RNA read counts before (control) and after the second week of CCRT (treatment) for each patient in the cohort, using a trimmed mean of M-value normalization with edgeR after excluding RNAs that were not detected in at least 50% of the samples. This analysis was conducted separately for each dataset. Plasma exosomal RNA sequencing and profiling were conducted by Macrogen (www.macrogen.com) and ROKIT Genomics (www.rokitgenomics.com). Further details are available in the Supplementary Methods.

RNA screening in the test dataset

Pearson’s correlations matrix between all variables, including AR and RNAs, was calculated using the “rcorr” function in the “Hmisc” package for R programming (R Foundation for Statistical Computing, Vienna, Austria). miRNAs associated with AR (|R| > 0.6) were initially selected. The optimal model incorporating the selected miRNAs was identified through an exhaustive search using the “regsubsets” function in the “leaps” package for R. This function evaluates all possible combinations of variables and selects the model with the highest Adjusted R² for each variable count. This method was consistently applied throughout the study. To assess the relevance of the sum and difference of the suggested miRNAs to AR, linear regression and the Wilcoxon rank-sum test were performed. A network was constructed using miRNAs related to the selected miRNAs (|R| > 0.4), and the miRNA most strongly associated with AR was identified. Two networks were then created: (1) using mRNAs linked exclusively to the selected miRNA (|R| > 0.6) and (2) incorporating miRNAs within the network, mRNAs associated with AR (|R| > 0.6), and those correlated with both the selected miRNA (|R| > 0.6) and AR (|R| > 0.4). Gene ontology analyses for biological processes and cell types were performed for each network. Among the mRNAs most relevant to AR (|R| > 0.7) within the second network, those demonstrating significant biological relevance based on gene ontology analysis were selected. Key components among the selected RNAs in the network were identified through an exhaustive search using the “regsubsets” function. The relevance of the sum of the finally selected miRNAs and mRNAs to AR was assessed using linear regression and the Wilcoxon rank-sum test. Additional networks were constructed using mRNAs relevant to the selected RNAs (|R| > 0.6), and gene ontology analyses were performed on the RNAs within these networks. Another network was constructed using mRNAs relevant to all the selected RNAs (|R| > 0.4), and gene ontology analyses were conducted on the RNAs within this network. Comparisons were made on the differences in TV at diagnosis (iTV), TV at the 4th week, age at diagnosis, pathology, 2018 FIGO stage, RT field, total dose (TD), ICBT, ALC0 (pretreatment ALC), ALC1 (ALC measured 1 week after RT), ALC2 (ALC measured 2 weeks after RT), and the Log₂FC values of the selected RNAs based on an AR value of 0.2.

Validation dataset

Linear regression and the Wilcoxon rank-sum test were performed to assess the relevance of the sum of the finally selected miRNA and mRNAs to AR in the validation dataset. Comparisons were made on the differences in iTV, TV at the 4th week, age at diagnosis, pathology, 2018 FIGO stage, RT field, TD, ICBT, ALC0, ALC1, ALC2, and the Log₂FC values of the selected RNAs based on an AR value of 0.3.

Cross-validation according to TV

Leave-one-out cross-validations (LOOCVs) were performed using the “trainControl” function from the “caret” package in subgroups based on iTV. The subgroups included patients with iTV values greater than 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, and 48 cm³.

Survival analysis

The endpoint of the survival analysis was progression-free survival (PFS). The follow-up period was measured from the end of treatment to the last visit or event date. The 3-year PFS (3PFS) between the two patient groups, categorized based on the median value of the selected RNAs, was compared using the Kaplan-Meier method and log-rank tests.

Network

Network analyses were performed using Prim’s algorithm of minimum spanning tree in the “igraph” package for R. Positively and negatively correlated edges are shown in red and blue, respectively, and were calculated as Pearson’s correlations.

Gene ontology

Enrichment analysis was performed for gene ontology annotation in Enrichr (https://maayanlab.cloud/Enrichr) using the mRNAs that significantly changed based on the log₂FC of the selected RNAs and the RNAs from the network [13].

All data analyses and visualizations were performed using R version 4.4.1 (https://www.r-project.org).

Results

Table 1 presents the characteristics of the patients included in this study.

Table 1.

Patients’ clinical characteristics (all patients)

	All
	N or median [IQR]
	(N = 41)
Tumor volume at diagnosis (cm3)	26.20 [16.5; 81.4]
Tumor volume at 4th week (cm3)	5.50 [1.8; 14.9]
Age (years) [IQR]	50 [46.0; 58.0]
≥ 50	25 (61.0%)
< 50	16 (39.0%)
FIGO stage 2018
IB	6 (14.6%)
IIB-IIIC1	23 (56.1%)
IIIC2-IVB	12 (29.3%)
Pathology
Adenocarcinoma	7 (17.1%)
Adenosquamous cell carcinoma	1 (2.4%)
Unknown carcinoma	1 (2.4%)
Squamous cell carcinoma	32 (78.1%)
Radiotherapy field
Pelvis with paraaortic region	11 (26.8%)
Pelvis	30 (73.2%)
Total dose (EQD2)	76.25 [72.25; 81.75]
≥ 76.25	28 (68.3%)
< 76.25	13 (31.7%)
Intracavitary brachytherapy
24 Gy in 4 fractions	14 (34.2%)
24 Gy in 6 fractions	10 (24.4%)
25 Gy in 5 fractions	1 (2.4%)
28 Gy in 7 fractions	1 (2.4%)
30 Gy in 6 fractions	13 (31.7%)
EBRT replacement	1 (2.4%)
Refusal	1 (2.4%)
ALC0 (cells/ul) [IQR]	1699.2 [1463.0; 2025.0]
ALC1 (cells/ul) [IQR]#	921.45 [615.2; 1242.55]
ALC2 (cells/ul) [IQR]	509.60 [371.2; 670.0]
miR-150-3p [log2FC]	-0.67 [-1.58; 0.18]
PRDM1 [log2FC]	0.13 [-0.7; 1.24]
NMT2 [log2FC]	-0.17 [-0.71; 0.31]
miR-150-3p+PRDM1 [log2FC]	-0.36 [-2.17; 0.74]
miR-150-3p+NMT2+PRDM1 [log2FC]	-0.48 [-3.23; 1.05]

IQR: interquartile range, FIGO: International Federation of Gynecology and Obstetrics, EQD2: equivalent dose in 2 Gy fractions, EBRT: External Beam Radiotherapy, ALC: absolute lymphocyte counts, ALC0: pretreatment ALC, ALC1: ALC 1 week after EBRT, ALC2: ALC 2 week after EBRT, NMT2: N-myristoyltransferase 2, PRDM1: PR/SET domain 1.

^#ALC1 has 1 missing value.

miR-150-3p as a key component for predicting AR

Figure 2 illustrates the process of selecting miR-150-3p as a key component for predicting AR in patients with cervical cancer who underwent CCRT in the test dataset.

Figure 2.

Identification of miR-150-3p as a key microRNA (miRNA) associated with acute tumor response to radiation therapy in the test dataset. (A) Among all possible combinations of three miRNAs, two were selected based on adjusted R² values from multiple regression analysis, indicating the strongest association with acute tumor response. (B) A negative linear correlation was observed between the log₂ fold change (log₂FC) difference of the selected two miRNAs and acute tumor response. (C-E) Boxplots comparing the difference between patients with a poor response (acute tumor response ≥ 0.2) and those with a good response (acute tumor response < 0.2): (C) miR-150-3p and miR-424-3p (log₂FC), (D) miR-150-3p (log₂FC), and (E) miR-424-3p (log₂FC). (F) A network analysis of miRNA-miRNA interactions highlighting the central role of miR-150-3p.

Among the subsets of combinations using miR-150-3p, miR-424-3p, and miR-3940-3p, which showed significant correlations with AR (|R| > 0.6), all optimal models included miR-150-3p (Figure 2A). The seven possible combinations are as follows: miR-150-3p, miR-3940-3p, miR-424-3p; miR-150-3p, miR-424-3p; miR-150-3p, miR-3940-3p; miR-3940-3p, miR-424-3p; miR-150-3p; miR-424-3p; and miR-3940-3p. The models with the highest Adjusted R² were selected for each number of explanatory variables: miR-150-3p, miR-3940-3p, and miR-424-3p for three variables; miR-150-3p and miR-424-3p for two variables; and miR-150-3p for one variable. Among these, the combination of miR-150-3p and miR-424-3p was identified as the optimal model due to its highest Adjusted R² value. The combination of miR-150-3p and miR-424-3p (log₂FC) showed a strong linear relationship with AR and demonstrated statistically significant differences between patients with poor AR and those with good AR (Figure 2B and 2C). However, when miR-150-3p and miR-424-3p were analyzed separately (Figure 2D and 2E), miR-150-3p was identified as the primary component distinguishing patients with poor response (AR < 0.2) from those with good response (AR ≥ 0.2). Network analysis of miRNAs associated with miR-150-3p or miR-424-3p (miR-miR network) revealed that miR-150-3p was central within this network (Figure 2F).

Given that miR-150-3p was the primary miRNA for predicting AR in the test dataset, while miR-424-3p played a supplementary role, and since the goal was to identify miRNAs that can be universally applied to predict acute tumor response, it was appropriate to focus solely on evaluating miR-150-3p.

Selection of lymphocyte-enriched RNAs relevant to AR

The network analysis of mRNAs associated with miR-150-3p (miR-150-3p-mR network) revealed a centralized structure with miR-150-3p at its core (Figure 3A). Biological process and cell type annotations of the RNAs within this network revealed strong associations with natural killer (NK) cell and T cell-mediated immunity (Figure 3B), with a stronger association with mature NK cell-mediated immunity than T cell-mediated immunity. The miR-mR network, defined as the network incorporating miRNAs within the miR-miR network, mRNAs associated with AR, and those correlated with both miR-150-3p and AR, is shown in Figure 3C. The network structure comprised three groups of RNAs based on KMO, MYDGF, and TLN2 (blue boxes in Figure 3C). The biological process and cell type annotations of the RNAs in this network revealed associations with mature T cells, mature NK cell-mediated adaptive immune response, and B cells (Figure 3D). The miR-mR network was strongly associated with lymphocyte-mediated immune responses. Among the 17 mRNAs (CRTAC1, DIABLO, DNAJB5, DPY19L4, E2F3, ERAP2, ESPN, FCER2, MORN3, NMT2, NUP107, PRDM1, PTCD3, QPRT, SLC39A3, TLN2, TRABD) highly relevant to AR within the miR-mR network (|R| > 0.7), FCER2, NMT2, and PRDM1, along with miR-150-3p, were additionally selected due to their enrichment in lymphocytes from immune cell specificity and blood single-cell data in the Human Protein Atlas (https://www.proteinatlas.org/) [14]. The selection criteria were based on their strong association with AR and biological relevance within the miR-mR network, which showed significant links to lymphocyte-mediated immune responses. The biological process ontology and cell types associated with acute tumor response and miR-150-3p-related mRNAs highlighted robust connections to immune response mechanisms and blood lymphocytes (Figure 3D). This indicates that exosome RNAs associated with AR likely originate from lymphocytes. Among the 17 mRNAs highly associated with AR, FCER2, NMT2, and PRDM1 were specifically selected due to their abundance in lymphocytes, ensuring universality and statistical and biological relevance in the context of immune-related mechanisms.

Figure 3.

Network analyses of messenger RNAs (mRNAs) associated with miR-150-3p and acute tumor response, highlighting lymphocyte-related ontologies and cell types. A. Network of mRNAs linked to miR-150-3p (red circle). B. Ontologies related to biological processes (upper) and cell types (lower). C. Network of mRNAs associated with both miR-150-3p (|R| > 0.6) and acute tumor response (|R| > 0.4), or with acute tumor response (|R| > 0.6), highlighting lymphocyte-enriched RNAs among them (|R| > 0.7, red circles). D. Ontologies for associated biological processes (upper) and cell types (lower) in the network.

Identification and validation of RNAs for predicting AR

Figure 4 illustrates the process of selecting miR-150-3p, NMT2, and PRDM1 as the final predictors of AR. The combined log₂FC values of miR-150-3p, NMT2, and PRDM1 showed a strong negative correlation with AR (Figure 4A and 4B). The selection process was as follows: the possible combinations were the following 15 subsets: FCER2, NMT2, PRDM1, miR-150-3p; NMT2, PRDM1, miR-150-3p; FCER2, NMT2, PRDM1; FCER2, NMT2, miR-150-3p; FCER2, PRDM1, miR-150-3p; NMT2, PRDM1; NMT2, miR-150-3p; FCER2, NMT2; PRDM1, miR-150-3p; FCER2, miR-150-3p; FCER2, PRDM1; NMT2; FCER2; miR-150-3p; and PRDM1. The values shown in Figure 4B represent the models with the highest Adjusted R² for each number of explanatory variables: when the model included four variables, FCER2, NMT2, PRDM1, and miR-150-3p were selected; for three variables, NMT2, PRDM1, and miR-150-3p were selected; for two variables, NMT2 and PRDM1 were selected; and for one variable, NMT2 was selected. Among these, the combination of NMT2, PRDM1, and miR-150-3p exhibited the highest Adjusted R² value and was ultimately chosen as the final model. Additionally, these values showed a significant difference between patients with poor response and those with good response in the test dataset (Figure 4C, 4D and Table S1). In the validation dataset, similar trends were observed (Figure 4E, 4F and Table S2), although the R² value for linear regression and the P-value for comparison between the two groups were less significant than those of the test dataset (R²: 0.4957 vs. 0.8305, P-value: 0.006 vs. < 0.0001, respectively). The median iTV of the test dataset was approximately three times larger than that of the validation dataset, as shown in Tables S1 and S2 (66.7 vs. 23 cm³, P = 0.006, respectively). In the test dataset, there was a significant difference in ALC1 between patients with poor AR and those with good AR (median value: 615.2 vs. 1,001.3 cells/μL, P = 0.013, respectively). However, no significant difference was observed in the validation dataset (median value: 784.4 vs. 1,121.55 cells/μL, P = 0.494, respectively).

Figure 4.

Identification and validation of combinations of miR-150-3p, PRDM1, and NMT2. A. Simplified network highlighting selected four RNAs. B. Three combinations, including one miRNA and three mRNAs, were identified through multiple regression analysis of all possible combinations, demonstrating the strongest association with acute tumor response based on adjusted R² values. C. A negative correlation was observed between the combined log₂ fold change (log₂FC) of miR-150-3p, PRDM1, and NMT2 and acute tumor response. D. Boxplot showing a significant difference in miR-150-3p + PRDM1 + NMT2 (log₂FC) between patients with a poor response (acute tumor response ≥ 0.2) and a good response (acute tumor response < 0.2). E. The validation dataset confirmed a negative correlation between miR-150-3p + PRDM1 + NMT2 (log₂FC) and acute tumor response. F. Validation dataset boxplot showing a significant difference in miR-150-3p + PRDM1 + NMT2 (log₂FC) between patients with a poor response (acute tumor response ≥ 0.3) and good response (acute tumor response < 0.3).

Cross-validations of subgroups based on TV

Figure 5 illustrates the cross-validation of subgroups based on iTV and survival analysis of finally selected RNAs. The R² values and root mean squared errors (RMSE) of LOOCV for miR-150-3p + NMT2 + PRDM1 (log₂FC) tended to increase and decrease with the increase in iTV of subgroups but remained consistent between 0.39 and 0.48 for R² and between 0.06 and 0.1 for RMSE (Figure 5B and 5C). The R² values and RMSE of LOOCV for miR-150-3p (log₂FC) showed a similar pattern, increasing and decreasing with the increase in iTV of subgroups. However, they remained consistent between 0.07 and 0.23 for R² and 0.11 and 0.16 for RMSE. For NMT2 (log₂FC), the R² values and RMSE tended to decrease with increasing iTV of subgroups but were stable between 0.08 and 0.16 for R² and 0.12 and 0.15 for RMSE. The R² values and RMSE of LOOCV for PRDM1 (log₂FC) also tended to increase and decrease with the increase in iTV of subgroups but were consistent between 0.28 and 0.44 for R² and 0.09 and 0.13 for RMSE. There was no significant difference in 3PFS between patients with lower log₂FC values than the median and those with higher log₂FC values for miR-150-3p, NMT2, and PRDM1, as shown in Figure 5D-F, respectively. The 3PFS rates were 69.6% vs. 80% for miR-150-3p (P = 0.5), 69.6% vs. 80% for NMT2 (P = 0.54), and 79% vs. 70% for PRDM1 (P = 0.42).

Figure 5.

Cross-validation of subgroups based on initial tumor volume and survival analysis of finally selected RNAs. (A) The number of patients stratified into subgroups based on initial tumor volume. (B) R² values. (C) Root mean squared error (RMSE) from leave-one-out cross-validation (LOOCV) of selected RNAs in subgroups of patients with initial tumor volumes above specific thresholds. (D-F) Kaplan-Meier plots and log-rank tests comparing progression-free survival between two patient groups divided by the median value of log₂ fold change (log₂FC) values of (D) miR-150-3p, (E) NMT2, and (F) PRDM1, respectively.

Evaluation of biological functions of selected RNAs

To explore the potential biological significance of PRDM1 and NMT2, two networks were constructed using mRNAs relevant to PRDM1 and NMT2. The biological processes and cell type annotations of the RNAs within each network were analyzed (Figure S3). The potential biological implications of PRDM1 and NMT2 include their roles in T cell-mediated adaptive immune response and migration from lymph nodes and cell killing by CD8+ T cells or NK cells, respectively. The 54 mRNAs relevant to miR-150-3p, NMT2, and PRDM1 suggest a potential involvement in immune activation processes mediated by NK cells or T cells (Figure S4).

Discussion

This study demonstrated that a combined decrease in post-RT levels of miR-150-3p, NMT2, and PRDM1 in plasma exosomes can predict poor AR and may be associated with reduced activation of NK and T cells. This predictive capability is more significant in patients with large iTVs, with PRDM1 emerging as the strongest predictor, potentially due to its role in T cell activation, highlighting its potential as a key biomarker for AR.

Exosome isolation using ATPS is a rapid and cost-effective method that yields high purity and minimal disruption but requires careful optimization to avoid contaminant co-isolation, which may necessitate additional purification steps [15]. The ATPS used in this study was selected for exosome isolation because it offers rapid results and ease of use, along with a higher recovery rate (70% vs. 5%-20%) and similar purity (77.5% vs. 70%-90%) compared to ultracentrifugation [16]. Using log₂FC to quantify RNA level changes facilitates detailed gene ontology analysis, overcoming the limitations of traditional group comparisons [17].

The decreased log₂FC value of miR-150-3p consistently predicted poor AR across the test, validation, and LOOCV datasets. Gene ontology analysis of related mRNAs suggested that miR-150-3p was more associated with mature NK cell-mediated immune responses rather than T cell responses. Additionally, the LOOCV results showed that miR-150-3p exhibited a stronger predictive ability in patients with iTV > 25 cm³, contributing to the increased combined log₂FC values of miR-150-3p, NMT2, and PRDM1. Previous studies have indicated that elevated miR-150 levels can promote mature NK cells and suppress immature NK cells, while miR-150-5p can reduce NK cell cytotoxicity by targeting perforin-1 [18,19]. These findings suggest that miR-150-5p promotes the development of mature NK cells but suppresses their cytotoxicity against tumor cells. However, our study indicated that higher post-RT levels of miR-150-3p, another mature miRNA derived from the miR-150 precursor, may enhance both the number and cytotoxicity of NK cells, especially in patients with large iTV. Notably, no studies to date have examined miR-150-3p’s role in lymphocytes, including NK cells.

Myristoylation, mediated by the enzyme N-myristoyltransferase (encoded by NMT1 or NMT2), involves attaching a myristoyl group to the N-terminal glycine of a protein, crucial for its membrane targeting and function [20]. This process plays a key role in T cell viability, activation, and the induction of extrinsic apoptosis in target cells by cytotoxic T cells [21]. During apoptosis, cleaved BID undergoes N-myristoylation by caspase 8, targeting BID to the mitochondria and enhancing its pro-apoptotic function [22]. Therefore, the decreased log₂FC values of NMT2 observed in this study may indicate a reduced capacity of CD8 T cells to induce apoptosis. This role related to the cytotoxicity of NMT2 may also apply to the extrinsic apoptosis pathway of target tumor cells induced by mature NK cells [23].

PRDM1, identified as the primary predictor of AR among the three selected plasma exosomal RNAs, plays a critical role in T cell migration and maturation, likely originating from lymph nodes or the thymus [24]. This suggests that CD8 T cells attack tumor cells through adaptive immune responses mediated by PRDM1. Notably, patients with increased post-RT levels of PRDM1 tended to have lower PFS than those with decreased levels, contrasting with patterns observed for miR-150-3p and NMT2. Blimp-1, encoded by the PRDM1 gene, is essential for CD8 T cell differentiation and migration to inflammatory sites while directing CD4 T cells toward Th2 differentiation and inhibiting other lineages [24,25]. Furthermore, Blimp-1 is significantly upregulated in exhausted CD4 and CD8 T cells under chronic antigen stimulation from infections or cancer [24,26]. Increased PRDM1 levels post-RT may initially indicate active cancer cell eradication by CD8 T cells, as RT enhances the exposure of immunogenic antigens [27]. However, chronic stimulation can lead to Th2 differentiation and T cell exhaustion, disrupting normal immune function. The presence of residual tumors after CCRT could further facilitate disease progression, especially with elevated PRDM1 levels. Conversely, low PRDM1 levels may promote a delayed but sustained immune response against residual tumors despite poor AR. This suggests that ICIs like pembrolizumab may be more effective for patients having LACC with low post-RT PRDM1 levels, while those with high levels may need alternative strategies to address T cell exhaustion. A recent study demonstrated that PRDM1-knockout chimeric antigen receptor (CAR) T cells can enhance persistence and therapeutic response by promoting early memory T cell expression and preventing terminal differentiation under repeated stimulation [28]. This highlights the potential of PRDM1-ablated CAR T cells as a breakthrough in CAR T cell therapy for solid tumors, especially in cases where residual tumors are resistant to ICIs following disease progression after rapid tumor regression and increased PRDM1 levels post-RT or chemotherapy [29].

The combination of post-RT levels of miR-150-3p, NMT2, and PRDM1 enhances the predictive ability for AR, with accuracy tending to increase in patients with large iTV. These findings suggest that the coordination of innate immunity by NK cells, cytotoxicity, and adaptive immunity by T cells, mediated by the three selected plasma exosomal RNAs as key regulators of these immune responses, can influence rapid regression following RT. This effect appears to be more significant in larger tumors. PRDM1 stands out as the most predictive among the three RNAs. This may be because it reflects the extent of cytotoxic T-cell activation and migration in response to increased antigen exposure following RT. Larger tumors are likely to present a greater variety of tumor antigens than smaller ones, necessitating CD8 T cell activation through the recognition of new, unmemorized antigens. NK cells can directly kill tumor cells or indirectly support this process by enhancing antigen presentation and promoting the maturation and activation of dendritic cells [30].

Employing miR-150-3p, NMT2, and PRDM1 clinically to predict AR in patients with LACC may allow for a more flexible RT schedule. Elevated post-RT levels of these RNAs could enable a reduction in RT dose to metastatic LNs or pelvic wall lesions or early initiation of ICBT before completing planned pelvic EBRT. Conversely, lower post-RT levels may support dose escalation to ensure adequate coverage of primary and metastatic lesions without delaying the initiation of ICBT. This approach can be particularly beneficial for large cervical tumors, especially in settings without MRI-guided ICBT, where cervical motion during EBRT can be managed using advanced technologies like setup control, non-invasive ultrasound tracking, and MRI-based online adaptive planning [31,32].

Despite these promising findings, this study had some limitations. The small sample size may limit the generalizability of results, and larger, multi-institutional studies are required to confirm these findings. The ATPS isolation method, while effective, requires refinement to ensure purity for clinical applications. Additionally, further research is needed to elucidate the causal mechanisms of these RNA changes and their impact on immune modulation and tumor regression. Prospective clinical trials are necessary to validate the utility of miR-150-3p, NMT2, and PRDM1 as biomarkers for AR prediction and personalized treatment strategies in cervical cancer.

Conclusions

Decreased post-RT levels of miR-150-3p, NMT2, and PRDM1 in plasma exosomes serve as predictors of poor AR, particularly in patients with large tumors. The role of PRDM1, the most significant predictor, was highlighted in T cell activation. These findings suggest that these RNAs could be valuable biomarkers for personalizing RT strategies. Moreover, potentially combining these biomarkers with immunotherapy could improve patient outcomes. Future research should focus on validating these biomarkers in larger cohorts and clinical trials.

Disclosure of conflict of interest

None.