Minimally invasive diagnosis of precancerous cervical lesions using single-cell peripheral immune atlas

Summary

Cervical cancer remains a major global health concern for women. Current screening methods are either invasive or lead to low participation and over-referral for colposcopy, particularly among high-risk human papillomavirus (HPV)-positive women. This study analyzes 613 participants with varying cervical lesions using mass cytometry by time-of-flight (CyTOF) to identify disease-specific peripheral immune signatures. A diagnostic model based on 23 immune features achieves ∼91% sensitivity and specificity for detecting precancerous and cancerous lesions. A separate model for HPV-positive women shows even higher accuracy (∼93% sensitivity, ∼95% specificity), especially in HPV16/18-positive cases (99% sensitivity, 100% specificity). In an independent validation cohort (n = 105), the model distinguishes cervical intraepithelial neoplasia (CIN) 2+ from ≤CIN1 with 86.5% sensitivity and 85.3% specificity (area under the curve [AUC] = 0.89). These findings support peripheral immune profiling as a minimally invasive and accurate biomarker strategy for early cervical cancer screening, particularly in HPV16/18-positive women.

Graphical abstract

Highlights

• •CyTOF profiling reveals immune signatures for early cervical cancer detection
• •CIN2+ model shows ∼91% sensitivity and specificity for CIN2/3 and cancer detection
• •HPV+ model detects HPV16/18+ cases with 99% sensitivity and 100% specificity
• •Both models outperform TCT screening in independent cohort validation

This study presents a CyTOF-based peripheral immune profiling method that accurately detects cervical precancer and cancer. Diagnostic models show high accuracy, especially in HPV16/18-positive women, and are validated in an independent cohort. The findings support a non-invasive strategy to improve cervical cancer screening and triage.

Introduction

Cervical cancer (CC) remains a significant global health issue, listed fourth in cancer-related deaths among women. Over 600,000 new cases and 340,000 deaths from CC were reported annually.¹ The progression of this disease is often driven by human papillomavirus (HPV) infection, making prevention through vaccination and early screening the most effective treatment.^2,3,4 Effective CC prevention relies on screening programs that aim to detect and treat precancerous lesions before they become invasive. However, traditional cytological screening, while convenient, lacks accuracy,⁵ with a sensitivity of 50%–70%. The World Health Organization (WHO) recommends high-risk HPV (hrHPV) testing,^6,7,8 which has a high sensitivity (over 90%). This method usually requires the referrals for colposcopy, a procedure that can cause significant discomfort and psychological distress for patients.⁹

HPV testing, although invasive, has the additional drawback of low specificity due to the transient nature of most HPV infections. This results in unnecessary colposcopy referrals, particularly for HPV16/18-positive women. Studies indicate that only around 20% of HPV16-positive patients have an immediate risk of developing significant cervical intraepithelial neoplasia (CIN) 2 or more severe conditions (including CIN2, CIN3, and CC, referred to as CIN2+). This percentage is even lower (∼10%) for HPV18-positive women.¹⁰ HPV-positive women aged 30 and above often show atypical squamous cells of undetermined significance (ASC-US) or low-grade squamous intraepithelial lesion (LSIL) on cytology at the initial screening, but many may not have clinically significant cervical lesions, such as CIN1 or lower grades.¹¹ Avoiding unnecessary colposcopies is crucial to prevent psychosocial and physical discomfort and avoid overtreatment of regressive CIN.^11,12,13,14 Recent research has noted an increase of CIN risk for patients with non-HPV16/18 hrHPV. However, these individuals with normal cytology do not require immediate colposcopy but should undergo follow-up after 12 months. This process may delay treatment in the case of CC, suggesting the need for additional triage tests to distinguish clinically significant lesions from transient infections.^15,16,17,18 This will reduce missed detections and highlight the urgent need for innovative, non-invasive or minimally invasive diagnostic approaches to improve CC detection accuracy while addressing the limitations of traditional screening methods.

Mass cytometry by time-of-flight (CyTOF) technology is emerging as a promising tool for comprehensive immune profiling.¹⁹ It provides insights into the immune composition, alteration, and responses in the progression of various cancers. Understanding the immune response to HPV infection and cervical carcinogenesis is crucial because persistent inflammation can promote cancer development.^20,21 In CC, prior research has shown that a decrease in cytotoxic CD8⁺ T cells, coupled with an increase in regulatory T cells (Tregs), can independently predict a poorer survival outcome.^22,23,24 Moreover, evidence supports the idea that the emergence and advancement of tumors are accompanied by broader disruptions in systemic immunity, along with alterations in peripheral immune cell populations.²⁵ An elevated ratio of Tregs to CD3⁺ T cells detected in the peripheral blood of patients with head and neck squamous cell carcinoma has been linked to an unfavorable prognosis.^26,27 CD4⁺ lymphocytes exhibiting a regulatory phenotype characterized by subdued proliferative responsiveness have been identified not only within the primary site but also in lymph node metastases and peripheral blood of individuals with CC, suggesting a systemic nature of immune tolerance induction.²⁸ Changes in peripheral immune cell subsets may aid in tumor detection and offer a readily applicable detection model for hepatocellular carcinoma and pancreatic ductal adenocarcinoma.²⁹ Moreover, recent study has developed an immunoprofiling platform using multiparameter flow cytometry and unsupervised clustering to identify five distinct immunotypes in patients with cancer, providing a potential tool for stratifying patients into therapy response groups based on systemic immunity.³⁰ Therefore, the profiling of immune components circulating in the bloodstream holds promise as a potential indicator to facilitate the diagnosis of tumors.

This study aims to leverage CyTOF technology to depict the peripheral immune atlas for patients with varying degrees of cervical lesions and construct diagnostic models based on these peripheral immune profiles. This would enable the minimally invasive and accurate identification of individuals with a high risk of developing precancerous lesions or CC. Such models have potential as triage tests, particularly for HPV16/18-positive women or those infected with non-HPV16/18 hrHPV types. This offers a tailored approach to CC screening that optimizes both clinical performance and resource utilization. By integrating advanced diagnostic modalities like CyTOF into existing screening protocols, we aspire to enhance the effectiveness and practicality of CC prevention strategies, ultimately advancing toward the WHO’s goal of eliminating CC as a public health concern on a global scale.

Results

Study population characteristics

A consort style diagram of our study is shown in Figure 1. Between January 2021 and February 2023, we aimed to collect peripheral blood samples from 800 women who visited the Women’s Hospital, Zhejiang University, for outpatient consultations or were admitted for inpatient treatment. A 1:1 enrollment strategy was used to include an equal number of CIN2+ and ≤CIN1 cases to enhance the diagnostic model’s accuracy. Based on the inclusion and exclusion criteria (STAR Methods), 650 women were enrolled, among whom 37 participants were further excluded due to the abnormal separation of the white membrane layer during peripheral blood mononuclear cell extraction. As a result, a total of 613 (94.3%) women were involved for the study (STAR Methods). The age of involved patients ranged from 15 to 84 years, with a median of 45.7 years. In consideration of ethical principles, women with double-negative results for HPV and thinprep cytologic test (TCT) do not have available cervical biopsy pathology results. The diagnosis for other individual patients was based on pathological findings from surgical samples or cervical biopsies. The participants included 164 cases (26.8%) of normal cervix without HPV infection or abnormal TCT results, 120 cases (19.6%) of normal cervix along with hrHPV infection, 52 cases (8.5%) identified as CIN1, 123 cases (20.1%) identified as CIN2/3 (CIN2 and CIN3), and 154 cases (25.1%) diagnosed with CC. As individuals with CIN1 may spontaneously regress to a normal state, individuals with CIN1 and those with a normal cervix, regardless of their HPV infection status, were grouped together as the “≤CIN1” group. CIN2/3 and CC were grouped together as the “CIN2+” group. Therefore, our study included 336 cases in the ≤CIN1 group and 277 cases in the CIN2+ group. The clinicopathological features of all participants are summarized in Table 1.

Figure 1.

A CONSORT-style flowchart of the study

Table 1.

Distribution of HPV and TCT tests in three different cervical groups

Group	Subgroup	≤CIN1	CIN2/3	CC	Total	Age (average, range)	Menopausal status (1: yes; 0: no)
HPV infection	HPV positive	172	115	141	428	46.5, 23–75	1: 162; 0: 266
	HPV16/18 infection	30	55	102	187	47.5, 24–75	1: 73; 0: 114
	non-HPV16/18 hrHPV infection	138	46	31	215	45.3, 23–72	1: 76; 0: 139
	HPV negative	164	8	13	185	43.7, 15–84	1: 46; 0: 139
TCT subtype	abnormal	84	86	122	292	47.6, 23–75	1: 121; 0: 171
	normal	252	37	32	321	43.9, 15–84	1: 87; 0: 234
TCT subtype 1	HSIL or worse	4	48	104	156	49.7, 25–75	1: 66; 0: 90
	LSIL or less	232	75	50	357	44.3, 15–84	1: 142; 0: 215
TCT subtype 2	LSIL or worse	32	63	109	204	47.7, 25–75	1: 80; 0: 124
	ASC-US or less	304	60	45	409	44.7, 15–84	1: 128; 0: 281

Peripheral immune compositions undergo significant changes during the progression of CC

We conducted the CyTOF analyses with the pre-designed antibody panel (Figure 2A; Table S1) on 613 peripheral blood samples, all of which passed quality control (QC). After performing the preprocessing step to exclude the dead, debris, doublets, and non-immune cells, the number of valid immune cells per sample ranged from a minimum of 11,093 to a maximum of 225,753. The cell number distribution for 613 samples was shown in Figure S1 with the median as 89,164, the 25 percentile as 68,142, and the 75 percentile as 108,825. By utilizing well-established cell surface markers, we identified 9 main cell types: CD4⁺ T cells (CD3⁺CD4⁺), CD8⁺ T cells (CD3⁺CD8⁺), double-negative T (DNT) cells (CD3⁺CD4⁻CD8⁻), gamma-delta (γδ) T cells (CD3⁺, γδTCR+), B cells (CD19/20/21+HLA-DR+), natural killer (NK) cells (CD56⁺), monocytes (CD14+HLA-DR+ CD11b+), dendritic cells (DCs, CD123+HLA-DR+), and basophils (CD123+HLA-DR−) (Figures 2B and S2). Comparing the variations of these main cell types across different cervical lesion groups, we found that certain subsets displayed alterations linked to the severity of the disease, progressing from ≤CIN1 to CC (Figures 2C and 2D). Specifically, CD4⁺ T cells showed a notable increase along with the cervical lesion progression. Conversely, CD8⁺ T cells, DNT cells, and plasmacytoid DCs (pDCs) showed a gradual reduction as the severity of the cervical lesion increased (Figure 2D).

Figure 2.

Single-cell CyTOF analysis revealing the peripheral immune profile in cervical cancer development

(A) Experimental design and analysis workflow for single-cell CyTOF.

(B) t-distributed stochastic neighbor embedding (t-SNE) plots depicting immune cells randomly selected from the pool of 613 cases, color-coded to represent major immune cell types.

(C) t-SNE density plot illustrating the immune cell landscape in different cervical groups.

(D) Cell frequency comparisons of identified immune subsets in patients with different cervical conditions: patients with ≤CIN1, patients with CIN2/3, and patients with cervical cancer. The line in the middle of the box represents the median, and the upper and lower limits of the box represent the upper quartile and the lower quartile of the data, respectively. The two additional lines above and below the box represent the upper quartile plus 1.5 times the interquartile range (IQR) and the lower quartile minus 1.5 times the interquartile range, respectively. Non-parametric Wilcoxon test was applied for statistical analysis.

Statistical significance is indicated as follows: ∗p < 0.05; ∗∗p < 0.01; ∗∗∗∗p < 0.0001.

To gain a deeper understanding of disease-related immune features, we further clustered these immune cells into 26 subsets (Figures 3A and 3B; Table S2). Among the 17 identified T cell clusters, five T cell subsets were notably more pronounced in the CIN2/3 and CC groups compared to the ≤CIN1 group. These subsets included C02_Th1-like CD4⁺ T, C04_Th2-like CD4⁺ T, C05_Tph, and C12_CD57+ CD8⁺ effector memory T (Tem) cells. In contrast, C09_Naive CD8⁺ T cells, C10_CD161+ CD8⁺ Tem cells, C22_CD57− NK, and C25_pDC cells showed decreased frequencies in the CIN2/3 and CC groups (Figure 3C). Additionally, we noticed that clusters C03–C06 exhibited similar marker expression patterns with high expression of CD4, CD28, CD95, and CD45RO but low expression of CD45RA (Figures 3D and S3). We also observed in peripheral blood an elevation of CD4^highCD28^highCD95^highCD45RO^highCD45RA^low T lymphocytes in the CIN2/3 group (Figure 3E). Moreover, this increase was more pronounced in the CC group compared to the ≤CIN1 controls (Figure 3E), implying a potential role of HPV in related cervical carcinogenesis.

Figure 3.

Characterization of peripheral immune cells in patients with different cervical conditions: patients with ≤CIN1, CIN2/3, and CC

(A) t-SNE plots displaying distinct cell clusters color-coded for each cluster.

(B) Heatmap displaying the normalized expression levels of markers for identified immune clusters. The proportions of various immune cell clusters are shown to the right of the heatmap.

(C) Cell frequency comparisons of the main differentially expressed cell clusters across the three groups.

(D) Expression patterns of the differentially expressed cell markers for the clusters C03–06 in patients with different cervical conditions: patients with ≤CIN1, patients with CIN2/3, and patients with CC (cervical cancer).

(E) Cell frequency comparison of the combined clusters C03–06 within the overall population of T cells varies across different cervical groups, including patients with ≤CIN1, CIN2/3, and those diagnosed with CC.

(F) Cell frequency comparisons of the main differentially expressed cell clusters between 120 cases of hrHPV-infected normal cervical samples and 164 cases of normal cervical samples without HPV infection.

(G) Cell frequency comparisons of the differentially expressed cell clusters in the peripheral blood of patients infected with HPV16/18 infections compared to those with non-HPV16/18 hrHPV genotypes. The line in the middle of the box represents the median, and the upper and lower limits of the box represent the upper quartile and the lower quartile of the data, respectively. The two additional lines above and below the box represent the upper quartile plus 1.5 times the IQR and the lower quartile minus 1.5 times the interquartile range, respectively. Non-parametric Wilcoxon test was applied for statistical analysis.

Statistical significance is denoted as follows: ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001; ∗∗∗∗p < 0.0001.

Immunological variations associated with hrHPV infection and among different hrHPV types

To further investigate the impact of HPV infection on peripheral blood immune cells, we conducted a comparative analysis of immune subsets in 120 cases of hrHPV-infected normal cervical samples and 164 cases of normal cervical samples without HPV infection (Figure 3F). The results revealed a notable decrease in C02_Th1-like CD4⁺ T cells and C26_Basophils, while there was an increase in C14_CD8+ NKT cells and C22_CD57− NKs with HPV infection. These findings suggest that these subsets may be initially relevant to the virus-induced immune responses. Interestingly, when stratified further by HPV infection types, we found that more subsets were enriched in the HPV16/18 infection group compared to the non-HPV16/18 hrHPV group (Figure 3G). The enriched subsets included C02_Th1-like CD4⁺ T cells, C04_Th2-like CD4⁺ T cells, C05_Tph cells, C06_Tregs, C11_CD57− CD8⁺ Tem cells, C12_CD57+ CD8⁺ Tem cells, and C20_Plasma cells. However, three clusters, C22_CD57− NKs, C25_pDCs, and C26_Basophils, showed opposite trends, decreasing in the HPV16/18 infection group. These findings demonstrate that hrHPV infection can influence the peripheral blood immune system, and there are distinct immune responses differently mediated by HPV16/18 infection or non-HPV16/18 hrHPV infections.

Prediction of precancerous/cancerous lesions based on peripheral immune signatures

To predict the presence of precancerous (CIN2/3) or cancerous (CC) lesions (collectively referred to as CIN2+), we developed a diagnostic model (CIN2+ model) that incorporates the peripheral immune features identified through single-cell profiling. The model was constructed using the random forest algorithm²⁹ (STAR Methods). In brief, we randomly divided the dataset, which included samples with CIN2+ and CIN1 or lower grades (collectively referred to as ≤CIN1), into a training set 1 and a testing set 1 (Figure 1). The training set 1, which accounted for 70% of the total samples (220 for ≤CIN1 and 220 for CIN2+), was used to select important peripheral immune features and train the model using a 10-fold cross-validation approach.²⁹ As a result, we identified 23 features, including eight marker expressions and the cell frequencies of 15 immune subsets (Table S3), which were used to calculate the disease risk for each individual sample. Statistical analysis revealed significant differences in disease risk between the CIN2+ and ≤CIN1 groups in both the training and testing sets (Figure S4).

The testing set 1, which consisted of 116 ≤CIN1 and 57 CIN2+ samples that were not involved in the model construction process, was then used to evaluate the model performance. In addition, we established the binary classifiers by using only the clinical indicators, such as TCT and HPV test values (STAR Methods). Average receiver operating characteristic (ROC) curves were then generated to compare the performance of these binary classifiers using clinical features as well as our CIN2+ model using immune features on the same training and testing sets. First, we observed that the CIN2+ model using the 23 peripheral immune features demonstrated high specificity and sensitivity in discriminating CIN2+ from ≤CIN1 in both the training (specificity = 90.9%, sensitivity = 91.4%) and the testing sets (specificity = 81.9%, sensitivity = 89.5%) (Figure 4A). Furthermore, comparing the area under the ROC curves (AUC) across these models revealed the highest AUC value for the CIN2+ model using the peripheral immune feature set (AUC = 0.97 for training data, AUC = 0.89 for testing data) (Figure 4B), indicating the superior performance of the diagnostic model using these peripheral immune features compared to other models.

Figure 4.

Development and validation of CIN2/3 and CC detection models

(A) Scatterplots displaying the sensitivity and specificity of disease risk using the 23 peripheral immune features for patients with CIN2+ (CIN2/3 and CC) from ≤CIN1 in both training and testing sets.

(B) The ROC curves for HPV tests, TCT tests, and peripheral immune features for CIN2+ detection in both training and testing sets.

(C) Scatterplots depicting the sensitivity and specificity of disease risk for patients with CIN2+ from ≤CIN1 in three distinct subgroups categorized by their HPV test status: the HPV-negative group, the HPV16/18 group, and the non-HPV16/18 hrHPV group.

(D) ROC curves illustrating the performance of HPV tests, TCT tests, and peripheral immune features for detecting CIN2+ from ≤CIN1 in the HPV-negative group, the HPV16/18 group, and the non-HPV16/18 hrHPV group within the testing set, respectively. AUC, area under the curve; ROC, receiver operating characteristic.

We divided the testing set 1 (116 ≤CIN1 and 57 CIN2+) into three subgroups based on their HPV test status: the HPV-negative group (n = 65; 60 ≤CIN1 and 5 CIN2+), the HPV16/18-positive group (n = 39; 10 ≤CIN1 and 29 CIN2+), and the non-HPV16/18 hrHPV group (n = 62; 45 ≤CIN1 and 17 CIN2+). Our findings showed that the CIN2+ model using the peripheral immune feature set achieved the best performance in the non-HPV16/18 subgroup (specificity = 91.1%, sensitivity = 88.2%, AUC = 0.95) (Figures 4C and 4D). This suggests that patients in this subgroup may directly use this model to assist with their diagnosis. However, for the HPV16/18 subgroup, the performance of the current model is slightly inferior to that of the other two subgroups, indicating the need for additional models to improve the diagnostic performance of patients with HPV16/18 infection.

In order to minimize clinical burden and avoid reliance on invasive sampling, the proposed CIN2+ model was developed without incorporating HPV/cytology data, aiming to provide a convenient and minimally invasive diagnostic tool for CC. As expected, when combining immune features with clinical data, the model’s performance could be further enhanced. Specifically, we constructed additional random forest models by using both immune signatures and HPV/cytology data and trained them on the same training set 1. The diagnostic results showed improved specificity and sensitivity for the models with combined features when evaluated on the same testing set 1 (Figures S5A and S5B), suggesting that integrating additional clinical features can further optimize diagnostic accuracy.

The peripheral immune feature-based diagnostic model focusing on HPV16/18-positive patients

As HPV16/18 infection is a high-risk factor for CC, patients with this infection often undergo additional tests for referral. To improve the accuracy of discriminating between ≤CIN1 and CIN2+ cases and reduce unnecessary referrals for ≤CIN1 patients, we developed a separate random forest model specifically for HPV16/18-positive patients. However, due to the significant imbalance in the number of CIN2+ and ≤CIN1 samples in the HPV16/18 group (≤CIN1: 16%; CIN2+: 84%), we included non-HPV16/18 HPV-positive samples in the model construction to balance the relative number of ≤CIN1 and CIN2+ cases. Consequently, we collected all HPV+ samples, which consisted of 172 ≤CIN1 cases (40%) and 256 CIN2+ cases (60%), and divided them into training set 2 (270 samples; 135 ≤CIN1 and 135 CIN2+) and testing set 2 (158 samples; 37 ≤CIN1 and 121 CIN2+) (Figure 1). Using a similar modeling strategy aforementioned (STAR Methods), we identified 20 relevant peripheral immune features, including two marker expressions and the cell frequencies of 18 different immune cell subsets, using the training set (Table S4). The calculated disease risk, based on these selected peripheral immune features, showed significant differences between CIN2+ and ≤CIN1 groups in both the training and testing sets (Figure S4). Subsequently, a diagnostic model for HPV-positive samples was constructed (referred to as HPV+ model) using the random forest algorithm. This model demonstrated high specificity (∼95%) and sensitivity (∼93%) for discriminating CIN2+ from ≤CIN1 cases in the training set. It performed similarly well in the testing set, with a specificity of 95% and sensitivity of 92% (Figure 5A). Furthermore, when compared to other prediction models constructed with clinical indicators (e.g., TCT test), the HPV+ model using the 20 peripheral immune features exhibited the highest AUC value in both the training and testing sets (Figure 5B).

Figure 5.

Development and validation of CIN2/3 and CC detection in women specifically infected with HPV16 or 18

(A) Scatterplots displaying the sensitivity and specificity of disease risk for hrHPV-infected CIN2+ patients compared to hrHPV-infected ≤ CIN1 patients using the HPV+ model in both the training and testing sets.

(B) ROC curves for TCT tests and the 20 peripheral immune features in hrHPV-infected CIN2+ detection for both the training and testing sets.

(C) Sensitivity and specificity scatterplots for disease risk in HPV16/18-infected CIN2+ patients compared to HPV16/18-infected ≤ CIN1 in the testing set using the HPV+ model.

(D) ROC curves for TCT and the peripheral immune features in HPV16/18-infected CIN2+ detection from HPV16/18- infected ≤CIN1 in the testing set based on the HPV+ model.

We divided the testing set further into two subgroups: HPV16/18 (n = 80; 6 ≤CIN1 and 74 CIN2+) and non-HPV16/18 (n = 63; 31 ≤CIN1 and 32 CIN2+) (Figures 5C, 5D, and S6) and compared the model’s performance on different subgroups. The results showed that the model performed exceptionally well (sensitivity = 99%, specificity = 100%, AUC = 1.0) in distinguishing CIN2+ from ≤CIN1 cases, especially in patients with HPV16/18. These results were significantly better than what was predicted by models built using clinical indicators (Figure 5D). Overall, these findings indicate that the HPV+ model, constructed using peripheral immune features, has the potential to accurately identify women infected with HPV16 or 18 who require immediate colposcopy referral, as opposed to those who can undergo retesting for hrHPV and assessment of peripheral immune cell activity during the follow-up period.

The performance of peripheral immune feature-based diagnostic models in an independent validation cohort of cervical samples

To confirm the aforementioned immune feature-based diagnostic models identified by CyTOF as minimally invasive biomarkers for early CC screening, we performed an independent cohort validation for the CIN2+ and HPV+ models. We expanded our dataset with 105 new samples. This additional cohort comprises 68 cases of ≤CIN1 and 37 cases of CIN2+, including 7 CIN2/3 cases and 30 CC cases (Table S5). Our validation results demonstrate that our CIN2+ model effectively distinguishes CIN2+ from ≤CIN1 with a sensitivity of 86.5% and specificity of 85.3%, yielding an AUC of 0.89 (Figures 6A and 6B). These results significantly outperform clinical index-based methods, including HPV-based and TCT-based diagnostic models (Figure 6B). We further performed subgroup analyses across different HPV infection statuses, and both the sensitivity and specificity exceeded 75% in all cases (Figure 6C). We observed a slight performance decrease in the non-HPV16/18 hrHPV group, which may be due to its sample size bias after stratification. Focusing on the HPV+ model, the analysis for the newly added cohort (47 cases) demonstrated a sensitivity of 96.6%, a specificity of 72.2%, and an AUC of 0.93 (Figures 6D and 6E), superior to the diagnostic performance for using TCT method. Our HPV+ model showed particularly high performance in the HPV16/18 subgroup (29 cases), with a sensitivity reaching 100% and a specificity of 83.3% (Figure 6F), demonstrating the robustness and effectiveness of our HPV+ model across multiple validation sets.

Figure 6.

Validation of peripheral immune feature-based models in an independent cohort

(A) Scatterplots displaying the sensitivity and specificity of disease risk using the 23 peripheral immune features for patients with CIN2+ from ≤CIN1 in an independent validation cohort of 105 new cervical samples.

(B) The ROC curves for HPV tests, TCT tests, and peripheral immune features for CIN2+ detection in the independent validation cohort.

(C) Scatterplots depicting the sensitivity and specificity of disease risk for patients with CIN2+ from ≤CIN1 in three distinct subgroups categorized by their HPV test status: the HPV-negative group, the HPV16/18 group, and the non-HPV16/18 hrHPV group.

(D) Scatterplots displaying the sensitivity and specificity of disease risk for hrHPV-infected CIN2+ patients compared to hrHPV-infected ≤ CIN1 patients using the HPV+ model in the independent validation cohort.

(E) ROC curves for TCT tests and the 20 peripheral immune features in hrHPV-infected CIN2+ detection for the independent validation cohort.

(F) The patients are categorized into two subgroups based on their HPV test status, HPV16/18 and non-HPV16/18 hrHPV, with scatterplots depicting the sensitivity and specificity of disease risk in the independent validation cohort using the HPV+ model.

Discussion

CC remains a significant global health concern. Early detection and screening are vital in reducing its mortality and morbidity rates.¹ In recent years, there has been increasing interest in exploring non-invasive or minimally invasive methods for cancer screening.^29,30,31,32 Our study aimed to investigate the feasibility of utilizing an auxiliary diagnostic model based on single-cell profiled peripheral immune features to assist in screening and triaging precancerous lesions and CC related to hrHPV infection. The results of this study provide valuable insights into the potential of minimally invasive diagnostic models and their implications for improving the early detection of CC.

The immune system plays a crucial role in surveilling and defending against cancer cells.^33,34,35 Extensive research has been conducted on the interactions between the tumor microenvironment and immune cells in various cancer types, including CC.^36,37,38,39 In this study, we utilized the high-dimensional profiling capability of CyTOF to analyze the composition and characteristics of peripheral immune cells in relation to CC development. Our approach is based on the assumption that the systemic immune response to cancer might be reflected in the peripheral blood, offering minimally invasive screening opportunities.^29,40 Our findings revealed significant alterations in the distribution of different immune cell subsets in individuals with precancerous lesions (CIN2/3) or CC compared to ≤CIN1. Notably, we observed an increased proportion of CD4⁺ T cells, including Tregs, alongside a decreased proportion of CD8⁺ T cells, DNT cells, and pDCs in patients with CIN2/3 and CC. This shift in immune cell composition suggests potential immune evasion mechanisms employed by cervical tumors. Moreover, previous research has revealed a CD4⁺CD25⁺ T cell subset, characterized by a phenotype featuring CD45RO^high, CD45RA^low, CD28^high, CD62L^high, and CD95^high, that possesses the ability to suppress hepatitis C virus-specific T cell responses, thus contributing to viral persistence.⁴¹ Consistent with this, we identified distinct CD4⁺ T cells characterized by a CD4^highCD28^highCD95^highCD45RO^highCD45RA^low phenotype, which exhibited heightened levels in the CIN2/3 group. The abundance of this subtype further increased in the CC group compared to the ≤CIN1 group, indicating its consistent involvement in the development of CC.

To distinguish precancerous and cancerous lesions (CIN2+) from CIN1 or lower grades (≤CIN1), we developed multiple models using either clinical indicators (HPV test or TCT test) or peripheral immune features. Their diagnostic performance was compared by using the same testing set. Interestingly, the model based on HPV test had the lowest AUC value. The models using different TCT stratification methods showed moderate AUC values ranging from 0.75 to 0.77, which were consistent with previous research findings.^42,43 However, the model utilizing peripheral immune features demonstrated the best diagnostic performance (AUC = 0.89). The validation results of this peripheral immune feature-based diagnostic model in an independent cohort strongly support their utility as minimally invasive biomarkers for early CC screening. Our model’s ability to distinguish CIN2+ from ≤CIN1 with a sensitivity of 86.5% and specificity of 85.2%, achieving an AUC of 0.89, demonstrates superior performance compared to traditional clinical diagnostic methods such as HPV-based and TCT-based models. This highlights the potential of immune profiling as a more accurate and reliable screening tool for precancerous and cancerous cervical lesions. In this study, we initially aimed to create a diagnostic tool based solely on peripheral immune signatures, independent of invasive testing methods such as HPV/cytology, to reduce clinical burden. As anticipated, incorporating HPV/cytology data into the model could result in improved performance; it also introduces increased complexity and potentially higher patient burden due to the additional clinical data required. These trade-offs should be carefully considered when designing future diagnostic strategies for CC.

HrHPV infection, particularly HPV16 and HPV18, is strongly associated with CC and other malignancies.⁴⁴ Therefore, in clinical practice, all patients with HPV16 or 18 infection are directly recommended for colposcopy examination.⁴⁵ However, the incidence of CIN2+ infection with HPV16/18 decreased from 2008 to 2016, whereas no trend was observed for non-HPV16/18 hrHPV-infected CIN2+ among young women.⁴⁶ Therefore, adopting alternative triage methods is crucial for improving the management of HPV16/18-infected patients, and the impact of other hrHPV genotypes on health cannot be overlooked.^47,48 In HPV16/18 infection, some individuals can clear the virus without developing cancer, while others progress to malignancies.⁴⁹ Therefore, it is important to understand the immune responses associated with different patient outcomes. Here, we also leveraged CyTOF technology to investigate the immune cell profiles in peripheral blood of patients infected with different hrHPV genotypes. Our investigation revealed notable distinctions in immune cell profiles between individuals infected with non-HPV16/18 hrHPV genotypes and those with HPV16/18 infections. This observation implies that distinct hrHPV genotypes may elicit different interactions with the immune system, resulting in specific immune cell responses. Understanding these distinct patterns is crucial for developing patient-specific management strategies. Subgroup analysis based on HPV test status revealed that the peripheral immune feature-based diagnostic model performed exceptionally well in the non-HPV16/18 subgroup, suggesting its potential utility as a diagnostic tool for patients in this category. However, the slight decrease in performance in the non-HPV16/18 hrHPV group of the independent validation cohort could be attributed to sample size limitations after stratification, suggesting the need for larger sample sizes in future studies to further refine the model’s accuracy in this subgroup. As for the HPV16/18 subgroup, the proposed HPV+ model also exhibited superior performance in discriminating CIN2+ from ≤CIN1 cases. Notably, in the independent validation cohort, our HPV+ model exhibited exceptional diagnostic power, particularly in the HPV16/18 subgroup, where sensitivity reached 100% and specificity 83.33%. These results underscore the robustness of the HPV+ model in identifying high-risk cases, especially in HPV16/18 infections, which are the most oncogenic types and require immediate colposcopy referral.

In conclusion, our study provides valuable insights into the immunological variations associated with hrHPV infection, highlighting distinct responses among different hrHPV types. The developed diagnostic model based on peripheral immune features, particularly the HPV+ model, shows promising potential for improving the accuracy of distinguishing precancerous and cancerous lesions. Peripheral immune features may serve as minimally invasive biomarkers for screening and triaging hrHPV-infected patients, offering a valuable alternative to traditional clinical indicators. Further validation and clinical implementation of these findings may contribute to more personalized and effective management strategies for individuals at risk of CC.

Limitations of the study

While our study demonstrates the utility of the CyTOF approach in providing high-resolution immune profiling with exceptional specificity and sensitivity, several limitations must be acknowledged. One of the primary challenges of this technology is its cost and technical complexity, which currently hinder its widespread use in routine clinical setting. The requirement for specialized equipment, expert operators, and substantial infrastructure investment makes CyTOF impractical for routine screening or triage, particularly in resource-limited environments. An important consideration is whether a smaller, more accessible panel of markers could offer similar performance using conventional flow cytometry, which is increasingly available in many hospitals. While CyTOF enables simultaneous analysis of over 20 markers, it is not yet clear whether a reduced marker set could maintain the same level of sensitivity and specificity necessary for clinical decision-making. Future research should focus on developing a core set of markers that could be effectively analyzed using routine flow cytometry. This would not only reduce costs but also make the technology more feasible for broader clinical implementation. Further clinical studies are needed to assess whether such a reduced marker panel can yield comparable diagnostic accuracy to the full CyTOF panel. If successful, this approach could serve as a bridge to make high-dimensional immune profiling more accessible and cost-effective, ultimately advancing the clinical applications of immune monitoring in diverse healthcare settings.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Dr. Hui Wang (wang71hui@zju.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •The mass cytometry data have been deposited in the OMIX database of the National Genomics Data Center, China National Center for Bioinformation (https://ngdc.cncb.ac.cn/omix/), under the accession number OMIX009788.
• •This paper does not report original code.
• •Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.

Acknowledgments

We acknowledge Dr. Xun Zeng for his advice in immunological interpretation of this study and acknowledge the use of BioRender for creating Figure 2A and the graphical abstract. The study was supported by the National Key R&D Program of China (grant no. 2021YFC2701204 to H.W.), the National Natural Science Foundation of China (grant no. 82373260 to H.W. and 82072855 to J.X.), the Key Research and Development Program of Zhejiang Province, China (grant nos. 2022C03013 to H.W., 2022C03012 to Q.J., and 2023C03169 to X.L.).

Author contributions

H.W. and J.X. conceived the project. J.X., B.Z., X.H., Z.C., and P.X. collected all clinical samples and data. Q.J., Q.K., and L.L. performed the bioinformatics analyses. J.X., Q.J., X.L., W.Y., and H.W. conducted statistical analyses. J.X. wrote the original draft with input from all authors. J.X., Q.J., X.L., W.Y., and H.W. revised the paper. H.W., W.Y., and J.X. supervised the project.

Declaration of interests

The authors have filed two invention patent applications based on this study (application numbers: 202410088246.2 and 202410481951.9).

STAR★Methods

Key resources table

REAGENT OR RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-Human CD45 (HI30) - 89Y	BioLegend	Cat# 304002; RRID: AB_314390
Anti-Human CD3 (UCHT1) - 115In	BioXcell	Cat# BE0231; RRID: AB_2687713
Anti-Human CD56 (NCAM16.2) - 141Pr	BD Biosciences	Cat# 559043; RRID: AB_397180
Anti-Human TCRγδ (5A6.E9) - 142ND	In House	N/A
Anti-Human CD196 (G034E3) - 143ND	BioLegend	Cat# 353402; RRID: AB_10918625
Anti-Human CD14 (M5E2) - 144ND	BioLegend	Cat# 301810; RRID: AB_314192
Anti-Human IgD (IA6-2) - 145ND	BioLegend	Cat# 348202; RRID: AB_10550095
Anti-Human CD123 (6H6) - 146ND	BioLegend	Cat# 306002; RRID: AB_2562834
Anti-Human CD85j (GHI/75) - 147Sm	BioLegend	Cat# 333722; RRID: AB_2814224
Anti-Human CD19 (HIB19) - 148ND	BioLegend	Cat# 302202; RRID: AB_314232
Anti-Human CD25 (24212) - 149Sm	R&D Systems	Cat# MAB1020; RRID: AB_357376
Anti-Human CD274 (29E.2A3) - 150ND	BioLegend	Cat# 329710; RRID: AB_2275581
Anti-Human CD278 (C398.4A) - 151Eu	BioLegend	Cat# 313502; RRID: AB_416326
Anti-Human CD39 (A1) - 152Sm	BioLegend	Cat# 328202; RRID: AB_940438
Anti-Human CD27(O323) - 153Eu	BioLegend	Cat# 302802; RRID: AB_314294
Anti-Human CD24 (ML5) - 154Sm	BioLegend	Cat# 311102; RRID: AB_314851
Anti-Human CD45RA (HI100) −155Gd	BioLegend	Cat# 304102; RRID: AB_314406
Anti-Human CD86 (Fun-1) - 156Gd	BD Biosciences	Cat# 555655; RRID: AB_396010
Anti-Human CD28 (CD28.2) - 157Gd	BioLegend	Cat# 302914; RRID: AB_314316
Anti-Human CD197 (G043H7) - 158Gd	BioLegend	Cat# 353222; RRID: AB_10945157
Anti-Human CD11c (BU15) - 159Tb	BioLegend	Cat# 337202; RRID: AB_1236381
Anti-Human CD33 (WM53) - 160Gd	BioLegend	Cat# 303410; RRID: AB_2074243
Anti-Human CD152 (14D3) - 161Dy	eBioscience	Cat# 14-1529-82; RRID: AB_467512
Anti-Human TCR Vδ2 (B6) - 162Dy	BioLegend	Cat# 331402; RRID: AB_1089226
Anti-Human CD161 (HP-3G10) - 163Dy	BioLegend	Cat# 339902; RRID: AB_1501090
Anti-Human CD185 (RF8B2) - 164Dy	BD Biosciences	Cat# 552032; RRID: AB_394324
Anti-Human CD95 (DX2) - 165Ho	BioLegend	Cat# 305614; RRID: AB_314552
Anti-Human CD183 (G025H7) - 166Er	BioLegend	Cat# 353718; RRID: AB_11150594
Anti-Human CD94 (HP-3D9) - 167Er	BD Biosciences	Cat# 555887; RRID: AB_396199
Anti-Human CD57 (HNK-1) - 168Er	BioLegend	Cat# 359602; RRID: AB_2562403
Anti-Human CD45RO (UCHL1) - 169Tm	BioLegend	Cat# 304202; RRID: AB_314418
Anti-Human CD127 (A019D5) - 170Er	BioLegend	Cat# 351302; RRID: AB_10718513
Anti-Human CD279 (EH12.2H7) - 171Yb	BioLegend	Cat# 329926; RRID: AB_11147365
Anti-Human CD38 (HIT2) - 172Yb	BioLegend	Cat# 303502; RRID: AB_314354
Anti-Human CD194 (L291H4) - 173Yb	BioLegend	Cat# 359402; RRID: AB_2562364
Anti-Human CD20 (2H7) - 174Yb	BioLegend	Cat# 302302; RRID: AB_314250
Anti-Human CD16(3G8) - 175Lu	BioLegend	Cat# 302014; RRID: AB_314214
Anti-Human HLA-DR (L243) −176Yb	BioLegend	Cat# 307612; RRID: AB_314690
Anti-Human CD4 (RPA-T4) - 197Au	BioLegend	Cat# 300516; RRID: AB_314084
Anti-Human CD8a (RPA-T8) - 198Pt	BioLegend	Cat# 301018; RRID: AB_314136
Anti-Human CD11b (M1/70) - 209Bi	BioLegend	Cat# 101202; RRID: AB_312785
Chemicals, peptides, and recombinant proteins
Ficoll-Paque™ PLUS	GE Healthcare	Cat# 17-1440-03
Bovine Serum Albumin (BSA)	Sigma-Aldrich	Cat# V900933-1KG
Fc Receptor Blocking Solution	In House	N/A
BioStab Antibody Stabilizer	MERCK	Cat# 55514
Maxpar® Antibody Labeling Kit	Standard BioTools	N/A
Maxpar® Fix and Perm Buffer	Standard BioTools	Cat# 201067
Cell-ID™ Cisplatin-194Pt	Standard BioTools	Cat# 201194
Cell-ID™ Intercalator-Ir	Standard BioTools	Cat# 201192B
EQ Four Element Calibration Beads	Standard BioTools	Cat# 201078
Tuning Solution	Standard BioTools	Cat# 201072
Washing Solution	Standard BioTools	Cat# 201070
Palladium isotopes	TRACE	N/A
Deposited data
Mass cytometry data of peripheral immune cells	This paper	OMIX database: OMIX009788; https://ngdc.cncb.ac.cn/omix/
Experimental models: Organisms/strains
Peripheral blood	This paper	Women’s Hospital, Zhejiang University School of Medicine
Software and algorithms
FlowJo v10.0.7	BD Biosciences	RRID:SCR_008520; https://www.flowjo.com/
R v3.6.1	R Core Team (2019)	RRID:SCR_001905; https://www.R-project.org/
Python (3.10)	Python Software Foundation	RRID:SCR_008394; https://www.python.org/

Experimental model and study participant details

Ethics approval

This study was performed in accordance with Declaration of Helsinki and approved by the Institutional Review Board (IRB) of Women’s Hospital of Zhejiang University (IRB-20200347-R).

Study population

Between January 2021 and February 2023, we aimed to collect peripheral blood samples from 800 women either visited the hospital for outpatient consultations or were admitted for inpatient treatment at the Women’s Hospital, Zhejiang University School of Medicine. To enhance the diagnostic model’s accuracy, a 1:1 enrollment strategy was employed, with an equal number of CIN2+ and ≤CIN1 cases. The eligibility criteria for inclusion in the study were as follows: no surgical history of cervix uteri, and informed consent obtained. Exclusion criteria included: (1) a history of cancer-related treatment; (2) pregnancy or being within 2 months of the postpartum period; (3) the use of immunosuppressive therapy; (4) having acute infections; (5) having immunodeficiency syndrome; (6) having confirmed coagulopathy; (7) poorly controlled diabetes; (8) tobacco use; (9) chronic alcohol use; (10) recent vaccination; (11) autoimmune diseases (such as rheumatoid arthritis, lupus, or multiple sclerosis); (12) severe or uncontrolled comorbidities; (13) current or recent steroid use; (14) chronic HBV/HCV infections; (15) recent blood transfusion. Due to some patients not meeting the inclusion criteria or having at least one of the exclusion criteria, 650 participants were finally enrolled and their blood samples were collected, among which 37 participants were further excluded due to abnormal separation of the white membrane layer during PBMC extraction. As a result, a total of 613 participants were obtained with CyTOF data and involved in the study analyses, including 336 with ≤CIN1 and 277 with CIN2+ cases. Throughout the study, all information pertaining to the participants, including their diagnoses, was kept confidential and concealed from both the experimenters and the analysts. Due to ethical considerations, women with double-negative results for HPV and TCT lack cervical biopsy pathology results. The diagnosis of other samples was confirmed through pathology from surgical procedures or cervical biopsies in accordance with clinical guidelines. The independent validation cohort, consisting of 105 new cervical samples, was collected between January 2024 and May 2024 from the Women’s Hospital, Zhejiang University School of Medicine, following the same inclusion and exclusion criteria.

The cytologic findings have been categorized following the 2014 Bethesda System,⁵⁰ which includes the following results: negative for intraepithelial lesion or malignancy (NILM), atypical squamous cells of undetermined significance (ASCUS), low-grade squamous intraepithelial lesion (LSIL), atypical squamous cells-cannot exclude high-grade squamous intraepithelial lesion (ASC-H), high-grade squamous intraepithelial lesion (HSIL), atypical glandular cells (AGC), adenocarcinoma in situ (AIS), and cancer cells. Abnormal cytology is defined as ASCUS or more severe abnormalities (ASCUS+). LSIL or worse (LSIL+) includes LSIL, ASC-H, HSIL, AGC, AIS, and cancer cells, while HSIL or worse (HSIL+) encompasses ASC-H, HSIL, AGC, AIS, and cancer cells. For diagnostic model analysis in this study, three distinct TCT result categories were utilized: TCT subtype (abnormal vs. normal), TCT subtype 1 (HSIL+ vs. LSIL-), and TCT subtype 2 (LSIL+ vs. ASC-US-).

Method details

Sample collection and processing

A 5 mL peripheral blood sample was collected from each participant prior to any treatments. These collected samples were then transported to the laboratory within 12 h and maintained at room temperature. Importantly, none of the samples were cryopreserved, and all analyses were conducted using fresh blood samples. To isolate PBMCs, we employed density gradient centrifugation with Ficoll. Following centrifugation, buffy coat was collected and washed using 5 mL of pre-cooled FACS buffer (PBS with 0.5% BSA) at 400×g for 5 min at 4°C. Subsequently, the cell precipitates were resuspended in FACS buffer. For quality control purposes, the samples needed to meet the following criteria: The cell count ≥3×10⁶, and the viability rate should be ≥85%.

Cell staining and CyTOF data acquisition

For mass cytometry analysis, we obtained purified antibodies from various sources, including BioLegend, R&D Systems, BD Biosciences, eBioscience, and BioXcell, and used the clones listed in Table S1. Antibody labeling with the specified metal tags was carried out using the MaxPAR antibody labeling kit from Fluidigm. Conjugated antibodies were titrated to determine the optimal concentration for usage. In total, we have done CyTOF experiments for over 600 blood samples. In order to ensure the intra-batch consistency of the staining antibodies throughout the study, we prepared a 100-test antibody dose each time with every staining antibody kept separate and not mixed with others. Prior to each batch’s use, we performed the titration tests by comparing its staining capability with the antibody from previous batch on the same PBMC sample to ensure the staining consistency across different batches. After the 100-test antibodies were used up, another 100-test antibodies would be prepared, and so on until all samples have been tested.

For each blood sample, a maximum of 3×10⁶ cells were initially washed once with 1xPBS and subsequently stained with 100μL of 250nM cisplatin (Fluidigm) for 5 min on ice to exclude dead cells. Following this step, they were incubated in a Fc receptor blocking solution before being stained with a cocktail of surface antibodies for 30 min on ice. After staining, the cells were washed twice with FACS buffer and fixed in 200μL of a MaxPar Fix and Perm Buffer (Fluidigm, Cat# 201067), containing 250nM of 191/193Ir, and incubated overnight.

For running one CyTOF experiment with multiple samples, we used the strategy of selecting two palladium isotopes from five available isotopes (Pd-104, Pd-105, Pd-106, Pd-108, Pd-110) to barcode the samples. As a result, a maximum of 10 unique barcodes could be used for each batch, allowing up to 10 samples examined together in one CyTOF running. For the barcoding step, each sample was resuspended in 100μL of the barcode solution and incubated on ice for 30 min. After barcoding each sample with the designated unique combination of palladium isotopes, we took maximumly 300,000 cells per sample, pooled them together, and then mixed them with 20% EQ beads (Fluidigm) before being acquired on a mass cytometer (Helios, Fluidigm), aiming to collect approximately 150,000 events per sample on average.

CyTOF data analysis

After collecting the raw data, the normalization process based on EQ beads⁵¹ was applied, and then the de-barcoding process⁵² was implemented to obtain the individual fcs file for each sample.

Due to the large size of enrolled samples, we used automated data cleanup instead of manual gating for preprocessing CyTOF data. The automated data cleanup method was based on the probability state modeling theory (PSM theory),⁵³ and was implemented in Python (3.10). The good performance of this automatic data cleanup method has already been validated by Dr. Vladimir Baranov and his colleagues.⁵⁴ After applying the automated data cleanup method, we also used the software FlowJo (v10.0.7) to double check the automatically processed data. To use one sample as a representative, we used both the manual gating and automated data cleanup methods for preprocessing. The strategy for manual gating was displayed by the first row in Figure S7A. Specifically, we used the 140Ce, DNA1/DNA2, 194Pt-cisplatin and Event-length to remove beads, debris, dead and doublet cells sequentially. The data separately preprocessed by manual gating (the second row in Figure S7A) and the automatic data cleanup method (the third row in Figure S7A) displayed very similar distributions in each gating step, demonstrating the similar performance for two methods. Besides, we also randomly selected 30 samples from our cohort, and used both methods for preprocessing. The obtained cell numbers from both preprocessing methods also showed well consistency (Figure S7B). After preprocessing, the obtained CyTOF data were arcsinh-transformed with a cofactor 5 before applying downstream analyses.

The PARC clustering algorithm⁵⁵ was then applied to partition all cells into distinct phenotypes based on their marker expression levels. It was implemented in Python (3.10) with the relevant parameters set as default. A maximum of 50,000 cells were randomly sampled from each sample file and pooled together before clustering, resulting in a total of approximately 30 million cells for clustering. All 41 markers were used for running PARC algorithm, with a resolution set as 0.8. As a result, 26 clusters were identified. Based on the expressions of typical lineage markers, including CD3, CD4, CD8, CD14, CD19, CD56, CD123, γδ TCR and HLA-DR (Figure S2), the clusters were manually merged into 9 meta-clusters (Figure 2B). To visualize the high-dimensional data in two dimensions and depict the distribution of each cluster and marker expression, as well as differences among various groups or cell types, the dimensionality reduction algorithm t-distributed stochastic neighbor embedding (t-SNE) was employed. For t-SNE plots, a maximum of 1,000 cells were randomly sampled from each sample file and pooled together, with a total of 613,000 cells involved in the t-SNE analysis. Similarly, all 41 markers were used for t-SNE algorithm that was implemented by “fastTSNE” function in Python (3.10) with default setting. The colors in the t-SNE plots were either based on the meta-cluster ID (Figure 2B) of the cells or on the cluster ID derived from PARC clustering (Figure 3A). Lastly, non-parametric Wilcoxon test was performed to compared the frequencies of annotated cell populations across different groups.

Model construction

We had a total of 613 valid samples that included 277 CIN2+ cases and 336 cases of ≤CIN1. We had set up two random forest models separately. The divisions of the training set and validation set for each model were illustrated as follows:(1) The CIN2+ model was to discriminate CIN2+ from ≤CIN1 cases. The training set 1 for CIN2+ Model included 440 cases (220 CIN2+ cases and 220 ≤CIN1 cases). The remaining cases in each group were considered as the testing set 1, which included 116 ≤CIN1 cases and 57 CIN2+ cases. (2) The HPV+ model was designed specifically for HPV16/18 positive patients. We had a total of 428 HPV+ samples that included 172 ≤CIN1 cases (40%) and 256 CIN2+ cases (60%). We used 270 samples (135 ≤CIN1 and 135 CIN2+) as the training set 2 and the remaining 158 samples (37 ≤CIN1 and 121 CIN2+) as the testing set 2.

The two random forest models were constructed separately, used different sets for model training, but shared the same modeling strategy, which included two main steps.

Firstly, the important immune features for each model were selected based on the training set. To create the immune feature pool, we randomly selected 50 samples and annotated the positive and negative cells for each marker in these samples by using FlowJo (v10.0.7). The annotated cells were then used as the training set to train a BP neural network classifier for each marker. In total, 41 BP neural network classifiers were set up and used to label the negative and positive cells for each marker in the remaining samples. Then, for each individual sample, according to the predefined marker combinations of 99 immune subsets (Table S6), the cell number and relative cell frequency of 99 immune subsets were calculated. Besides, the cell frequency of positive and negative cells for every marker was also calculated. As a result, the relative cell frequencies of 99 immune subsets and the positive/negative cell frequency of 41 markers were collected together as a pool of immune features. Then, the feature selection was implemented as follows: Step1, we randomly selected 80% of the training data as a training subset, based on which a random forest model was built up via a 10-fold cross-validation strategy and the corresponding feature importance was computed. Step2, we repeated the Step1 process for 1000 times, resulting in 1000 feature importance values for each feature. Step3, the features with its importance value ≥ 0.01 for over 350 times are finally selected. Using this process, CIN2+ model selected 23 immune features as listed in Table S3. While, HPV+ model selected 20 immune features as listed in Table S4.

Secondly, with the selected features, the final random forest model was constructed using the entire training set with a 10-fold cross-validation strategy. It was implemented in Python (3.10) by using the Scikit-learn library. The parameters were set as the following: Boostrap: False, max_depth: 9, min_sample_leaf: 7, n_estimator: 500, and others as default setting.

Besides, we have also built up the binary classifiers using only TCT or HPV test values with the following criteria.

• (1)Using HPV test record as the indicator: the binary classifier treated HPV positive as CIN2+ (labeled as 1) and HPV negative as ≤ CIN1 (labeled as 0).
• (2)Using TCT subtype (abnormal vs. normal) as the indicator: the binary classifier treated TCT abnormal as CIN2+ (labeled as 1) and TCT normal as ≤ CIN1 (labeled as 0).
• (3)Using TCT subtype1 (HSIL+ vs. LSIL-) as the indicator: the binary classifier treated TCT HSIL+ as CIN2+ (labeled as 1) and TCT LSIL-as ≤ CIN1 (labeled as 0).
• (4)Using TCT subytype2 (LSIL+ vs. ASC-US-) as the indicator: the binary classifier treated TCT LSIL+ as CIN2+ (labeled as 1) and TCT ASC-US- as ≤ CIN1 (labeled as 0).

In order to compare the performances of random forest model as well as the binary classifiers, we used the specificity, the sensitivity and the ROC curve and evaluated these metrics on the same training and testing sets. The specificity was calculated as the rate of true negatives among negative samples, and the sensitivity were calculated as the rate of true positives among positive samples. The ROC was plotted with the False Positive Rate (FPR) on the x axis and the True Positive Rate (TPR) on the y axis. And the area under the ROC curve was calculated as the AUC (Area Under the Curve).

For the random forest model, a probability score of being positive were predicted for each sample, and 0.5 was used as the cutoff for the positive label. The ROC curve was generated by the following steps: Based on the model’s output probability scores on the training (or testing) set, we sort the scores in descending order, and sequentially used each score value as the threshold. Based on each threshold, we calculated the corresponding False Positive Rate (FPR) and True Positive Rate (TPR) on the given dataset, and finally used these series of FPR and TPR coordinates to plot the ROC curve.

For the binary classifiers, labels (0, or 1) were directly predicted for each sample on the training (or testing) set. As they are hard classifiers, we firstly used the predicted labels to train a logistic regression model, and then used the probability score output by the regression model to create the ROC curves according to above mentioned steps.

Quantification and statistical analysis

The boxplots shown in Figures 2D, 3C–3G, and S1 are described as follows: The line in the middle of the box represents the median of the data; the upper and lower limits of the box respectively represent the upper quartile and the lower quartile of the data; the two additional lines above and below the box represent the upper quartile plus 1.5 times the interquartile range (IQR) and the lower quartile minus 1.5 times the interquartile range, respectively. Non-parametric Wilcoxon test was applied for statistical analysis in Figures 2D and 3C–3G. Statistical significance is indicated as follows: ∗, p < 0.05; ∗∗, p < 0.01; ∗∗∗, p < 0.001; ∗∗∗∗, p < 0.0001. All these statistical analyses were implemented in R v3.6.1.

Published: May 23, 2025