Integrative machine learning frameworks to uncover specific protein signature in neuroendocrine cervical carcinoma

Abstract

Objective

Neuroendocrine cervical carcinoma (NECC) is a rare but highly aggressive tumor. The clinical management of NECC follows neuroendocrine neoplasms and cervical cancer in general. However, the diagnosis and prognosis of NECC remain dismal. The aim of this study was to identify a specific protein signature for the diagnosis of NECC.

Methods

Protein and gene expression data for NECC and other cervical cancers were retrieved or downloaded from self-collected samples or public resources. Eleven machine-learning algorithms were packaged into 66 combinations, of which we selected the optimal algorithm, including randomForest, SVM-RFE, and LASSO, to select key NECC specific dysregulated proteins (kNsDEPs). The diagnostic effect of kNsDEPs was validated by a set of predictive models and immunohistochemical staining method. The dysregulation patterns of kNsDEPs were further investigated in other neuroendocrine carcinomas.

Results

Our results showed that NECC displays distinctive biological characteristics, such as HPV18 infection, and exhibits unique molecular features, particularly an enrichment in cytoskeleton-related functions. Furthermore, secretagogin (SCGN), adenylyl cyclase-associated protein 2 (CAP2), and calcyclin-binding protein (CACYBP) were identified as kNsDEPs. These kNsDEPs play a central role in cytoskeleton protein binding and showcase robust diagnostic ability and specificity for NECC. Moreover, the concurrent upregulation of SCGN and CACYBP, along with the downregulation of CAP2, represents a unique feature of NECC, distinguishing it from other neuroendocrine carcinomas.

Conclusions

This study uncovers the significance of kNsDEPs and elucidates their regulated networks in the context of NECC. It highlights the pivotal role of kNsDEPs in NECC diagnosis, thus offering promising prospects for the development of diagnostic biomarkers for NECC.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12885-025-13454-z.

Introduction

Neuroendocrine cervical carcinoma (NECC) is a rare but aggressive tumor of the female genital tract, representing 1–2% of all cervical cancers [1, 2]. Considering the global annual incidence of cervical cancer is approximately 604,127 cases, it is estimated that NECC accounts for 6,000 to 12,000 new cases per year worldwide [3, 4]. The grading of NECC follows a similar pattern as neuroendocrine neoplasia found in other locations, such as the lungs or digestive system. Specifically, neuroendocrine tumors predominantly occur in the lungs, and the aggressive nature of NECC resembles that of small cell lung cancer, which is often found to be locally advanced or metastasized [1, 5]. Positive immunohistochemical staining for neuroendocrine markers, including synaptophysin (SYP), chromogranin (CHGA), neural cell adhesion molecule (NCAM, also named CD56), and neuron-specific enolase (ENO2) are also applied for the diagnostic confirmation of NECC [6]. However, it is worth noting that in certain cases of NECC, the expression of neuroendocrine markers may be negative, suggesting the need to explore the cell type-specific biomarker for the detection of NECC.

Cervical cancer continues to be a significant contributor to female mortality on a global scale [7]. The most common subtypes of cervical cancer are cervical squamous cell carcinoma (CSCC) and endocervical adenocarcinoma (ECA), comprising up to 70% and 25% of all cases, respectively [8–10]. NECC exhibits distinct biological characteristics compared to CSCC and ECA. For instance, NECC has a higher propensity to invade the lymph-vascular space and metastasize to the regional lymph nodes upon diagnosis. Additionally, NECC is more prone to both local and distant relapses, resulting in a significantly lower 5-year overall survival rate of approximately 30% compared to the > 65% survival rate observed in CSCC and ECA [9–11]. Treatment strategies for NECC, such as surgery, radiotherapy, and chemotherapy, are often extrapolated from CSCC and ECA [5, 12, 13]. However, the mean recurrence-free survival of NECC is 16 months, and the median overall survival of NECC only falls within the range of 22 and 25 months [5, 6].

The purpose of this study was to investigate the specific protein expression profiles of NECC. We utilized quantitative proteomic analysis on fresh-frozen tissue samples of NECC and paired paracancerous tissues. A total of 211 differentially expressed proteins were identified in NECC specimens. We also conducted a comprehensive comparison of the proteomic heterogeneity between NECC and common cervical cancers, including CSCC and ECA. We found that a total of 142 proteins were dysregulated specifically in NECC but not in CSCC and ECA. In addition, to identify the key proteins involved in NECC, we integrated multiple machine learning algorithms and identified SCGN, CAP2, and CACYBP. The specific dysregulation patterns of these key proteins in NECC were further validated by immunohistochemical samples. Expression analyses further revealed the differential expression patterns in NECC—where SCGN and CACYBP were upregulated, and CAP2 was downregulated—was distinct with other neuroendocrine carcinomas. Based on these key proteins, we constructed a novel predictive model that demonstrated good accuracy for NECC diagnosis. These results will expand our understanding of the molecular landscape and provide potential diagnostic biomarkers and therapeutic targets for NECC.

Methods

Clinical samples

NECC specimens comprising both paired paracancerous tissues and tumor samples were retrieved from East China Normal University Wuhu Affiliated Hospital (The Second People’s Hospital of Wuhu City). The paracancerous speimens were retrieved from a distance of more than 2 centimeters from the cancer edge according to previous research [14]. Both paracancerous and tumor speimens were further validated by HE staining to assess cellular morphology and tissue architecture by three experienced pathologists from East China Normal University Wuhu Affiliated Hospital (The Second People’s Hospital of Wuhu City) and The First Affiliated Hospital of Wannan Medical College (Yijishan Hospital of Wannan Medical College). Collectively, we acquired sixteen paired paracancerous tissue and tumor samples from eight patients. As the tumor type of patient eight was identified as NECC mixed with ECA by immunohistochemical identification, and the protein expression pattern of patient eight was dramatically discrepant from other NECC samples, we excluded patient eight for the following analyses (Figures S1, 2 and Table S1). In addition, we collected three normal cervical samples from hysteromyomectomy patients, and three paired paracancerous and tumor samples from CSCC and NECC samples as validation cohort. The research protocol was approved by the ethical committee for human experimentation of East China Normal University Wuhu Affiliated Hospital (The Second People’s Hospital of Wuhu City, approve number 2024-KY-009) and Anhui Normal University (approval number AHNU-ET2023080). All patients included in this research have granted written informed consent for the use of their data.

Sample preparation

NECC samples were taken from 10 to 15 paraffin sections and deparaffinized according to the procedure of the deparaffinizing agent. After deparaffinization, samples were centrifuged and the supernatant was discarded, and the samples were lyophilized in a lyophilizer for 30 min until they were completely dry. The samples were ground thoroughly and 30ul of SDT lysate was added to the precipitate and heated at 95 °C for 90 min, then frozen at room temperature, sonicated in an ice water bath for 5 min, and centrifuged at 15,000 g. The supernatant protein solution was collected. The total protein concentration was determined by BCA protein quantification analysis. An aliquot of the protein solution was taken according to its concentration and the volume was adjusted with 8 M urea to 200 µL. Subsequently, the solution was reduced with 10 mM dithiothreitol for 45 min at room temperature and alkylated with 50 mM iodoacetamide for 15 min at room temperature under dark conditions. For precipitation, pre-cooled acetone was used and added to the protein solution in fourfold volume. This precipitation process was carried out at -20 °C for 2 h. After centrifugation, the protein precipitate was collected and resuspended in a 200 µL solution containing 25 mM amine bicarbonate solution and 3 µL trypsin. The mixture was digested at 37 °C overnight. After digestion, peptides from each sample were desalted using a C18 column, concentrated by vacuum centrifugation, and then redissolved in 0.1% (v/v) formic acid solution.

Quantitative proteomic analysis

The proteomic analysis of the NECC specimens mentioned above was performed by Wuhan Metware Biotechnology Inc. (Table S2). Quantitative proteomics analysis was conducted using 4D-DIA methods. 4D-DIA quantitative proteomics technology is the use of timsTOF Pro2 mass spectrometer diaPASEF acquisition mode to achieve differential quantitative proteomic analysis. diaPASEF acquisition mode is a combination of DIA (data-independent acquisition) acquisition mode and PASEF technology, first based on the mass-to-charge ratio (m/z) The whole scanning range of mass spectrometry is divided into several windows, and then all the parent ions in each window are simultaneously accumulated and continuously fragmented, m/z is detected, and the retention times, mass-to-charge ratios, ion intensities, and ion mobilities of the fragmented ions of all the parent ions are collected, and finally protein characterization and quantification are carried out based on this information. The library search software employed in the study was DIA-NN (v1.8.1). The database used was the uniprot_proteomeUP000005640_human_20230504.fasta database (82492 entries). The parameters derived from deep learning were employed to predict a spectral library. The Match Between Runs option was selected, a spectral library was created and for the reanalysis of the DIA data with the objective of protein quantification. The false discovery rate was filtered at 1% for both the precursor ion and protein levels. Total protein was determined through SDS-PAGE and trypsin enzymolysis. After desalting, the samples were identified using liquid chromatography-tandem mass spectrometry (LC-MS/MS). In this study, the quantitative proteomics data were organized and the screening conditions were set as p-value < 0.05. Additionally, the proteomics data from NECC samples was also used as training cohort in machine learning and predictive model analyses.

Immunohistochemical analysis

Paraffin sections were prepared and IHC staining were performed as described previously [15]. In detail, three normal slides, three paired CSCC slides, and three paired NECC slides and eight NECC slides from East China Normal University Wuhu Affiliated Hospital (The Second People’s Hospital of Wuhu City) were collected for IHC experiments. The rabbit polyclonal antibody targeting CACYBP was obtained from Proteintech, USA (11745-1-AP), with a primary antibody dilution of 1:200 for CACYBP. Similarly, the rabbit polyclonal antibody targeting CAP2 was obtained from Proteintech, USA (15865-1-AP), with a primary antibody dilution of 1:50 for CAP2. Additionally, the rabbit monoclonal antibody targeting SCGN was obtained from Abcam, UK (ab138493), with a primary antibody dilution of 1:250 for SCGN. Finally, these sections were visualized by light microscopy, and the results of IHC were analyzed using ImageJ software.

Differential analysis of protein expression

We compared the expression profiles between paracancerous and tumor samples of NECC to identify the differentially expressed proteins (DEPs) using the R package “limma”, with a p-value < 0.05 as the criterion. The p values were calculated using the Wilcoxon rank sum test (Table S3).

Public proteomic and transcriptomic data acquisition and preprocessing

Proteomic datasets related to CSCC and ECA were obtained from literature reports [16–21]. Redundancies were eliminated after extracting the DEPs that pertain to CSCC and ECA from the aforementioned works (Table S4). Transcriptomic datasets related to CSCC and ECA were obtained from the TCGA database (https://portal.gdc.cancer.gov/). To perform an unbiased analysis, all CSCC and ECA patients, comprising 248 Squamous Cell Neoplasms, 31 Adenomas, and Adenocarcinomas, and 3 Normal cervix tissues, available on the TCGA database with gene expression information were included. The differentially expressed genes (DEGs) were obtained by using the R package “limma” in view of |log2 fold change (FC)| ≥ 2 and p-value < 0.001. Omics data pertaining to 23 non-CSCC and ECA patients were retrieved from the TCGA database (https://portal.gdc.cancer.gov/), which was further applied as validation cohort in machine learning and predictive model analyses.

Identification of kNsDEPs using multiple machine learning algorithms

Since many types of machine learning have unique algorithm designs, it would be inappropriate to rely on specific machine learning algorithms to construct models without systematic comparisons. In this study, a total of 11 independent machine learning algorithms and their combinations were used to form a machine learning framework for the identification of kNsDEPs, which include Random forest (RF), Lasso regression (Lasso), Linear discriminant analysis (LDA), k-Nearest Neighbor (KNN), Decision tree (DT), ridge regression (RR), Elastic Net regression (ENR), Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), Naive Bayesian algorithm (NaiveBayes) [22–26]. These algorithms play a fundamental role in the initial stages of kNsDEPs identification and work in concert with other algorithms. Both individual algorithms and combinations of both algorithms contribute to this comprehensive framework. We evaluated 66 permutations of these 11 algorithms in a 10-fold cross-validation framework using a training cohort for variable selection and model construction. We ranked the performance of the models based on average C-index and selected combinations of algorithms with robust performance and clinical translational significance, i.e., the model with the highest average C-index across all cohorts was considered the best model.

Validation of diagnostic markers

Nomogram is a graphical prediction tool that can combine the expression of key genes to predict the risk of disease development [27, 28]. The purpose of using the Nomogram algorithm in this study was to evaluate the diagnostic effect of kNsDEPs on NECC compared with other commonly used neuroendocrine markers, CSCC and ECA markers that have been applied in NECC clinical diagnosis. Specifically, to determine the importance of kNsDEPs and the indicated markers in the diagnosis of NECC, we used the R package “rms” to construct a column chart of identified characterized genes. The column plot consists of “points”, which represent the scores of the candidate genes, and “total score”, which shows the sum of the scores of all the genes. The area under the curve of the receiver operating characteristic (ROC) was then plotted to evaluate the performance of kNsDEPs and column maps in the diagnosis of NECC. The ROC analysis generated the area under the curve (AUC) and 95% confidence intervals (CIs), with an AUC value of > 0.7 considered to be highly diagnostic.

Construction and validation of the artificial neural network (ANN) model

Artificial Neural Networks (ANN), as one of the major types of artificial intelligence, can mimic the structure and function of the brain’s neural network, derive a set of classification rules from complex and irregular data, and thus produce highly accurate diagnostic models, which are now widely used in clinical medicine for diagnosis and treatment [29–31]. The purpose of using the ANN algorithm in this study was to assess the diagnostic effect of kNsDEPs on NECC compared with other commonly used neuroendocrine markers, CSCC, and ECA markers that have been applied in NECC clinical diagnosis. Specifically, we constructed an ANN diagnostic model based on the expression level of kNsDEPs and the indicated markers using the R package “neuralnet”. The seed was set to 123 after we normalized the expression data using the min-max normalization method.

The constructed ANN model consisted of predominantly three layers:

1) Input layer, which mainly consists of the protein expression of the markers.

2) Hidden layer, which mainly includes the protein expressions of the markers and the weights of the markers.

3) Output layer, which indicates the result of judging whether the sample belongs to the Normals or the Tumors.

Each of these nodes employed a softmax activation fiction to ensure the model in the configuration of the key parameters of the model. The error fiction was defined using the sum of squared errors, the nonlinear fiction “sigmoid” is regarded as the activation function. We selected the cross-entropy loss function and implemented the efficient Adam optimization method to fine-tune the neural networks weights. Finally, we use the AUC to evaluate the diagnostic ability of the ANN model.

PPI network construction and functional enrichment analysis

We submitted the three kNsDEPs and all 211 DEPs in NECC to the STRING database [32]. Proteins with a score ≥ 0.4 were selected from the STRING database to construct a network model, which was visualized using Cytoscape (v3.7.2) [33]. The resulting network consisted of 38 nodes and 94 edges, all of which were linked to kNsDEPs. We comprehensively analyzed the biological functions of these DEPs and kNsDEPs using GO enrichment analysis, including cellular component (CC), molecular function (MF), and biological process (BP). All enrichment analyses were conducted using the R packages clusterProfiler, org.Hs.eg.db, and enrich plot (https://bioconductor.org/packages/clusterProfiler/, https://bioconductor.org/packages/org.Hs.eg.db/, https://bioconductor.org/packages/enrichplot/), with both P and FDR values less than 0.05 considered statistically significant.

Statistical analysis

All statistical analyses (analytical program) and visualizations were performed by the R language, and P < 0.05 was considered statistically significant.

Result

Global profiling of NECC proteomics

Of the eight patients, the median age of diagnosis was 53 years old (range, 43–75). 7 of 8 patients had histologically pure NECC. Staging was conducted in accordance with the FIGO staging system (2018) (Table S1) [34]. Patient 8 was found to have mixed cervical cancers and was therefore excluded from the subsequent analysis (Figures S1, 2). High-risk HPV infections were detected in all cases, with HPV18 being the most common type found in six cases. Only one case tested positive for HPV16, indicating that HPV18 is the most common subtype in NECC, which was consistent with previous findings (Table S1) [35].

To investigate the protein expression profiles of these NECC samples and their respective para-carcinoma tissues, we employed a 4D-DIA quantitative proteomic strategy. In total, the quantification of proteins with one or more unique peptides resulted in the identification of 8812 proteins. Specifically, 8771 proteins were found in the NECC group, and 8542 proteins were detected in the paracancer group (Fig. 1A, Table S2). In addition, a two-dimensional principal component analysis was conducted to illustrate the variation in protein abundance within and between the two groups. Based on the log2-ratio of each sample over the average of all samples, we observed a complete separation of the NECC group from the paracancer group (Fig. 1B). In comparison to the paracancer group, the NECC group exhibited significant differential expression in a total of 211 quantified proteins, consisting of 84 upregulated proteins and 127 downregulated proteins (Fig. 1C, D, and Table S3).

Fig. 1.

Proteomic profiles of NECC samples. (A) Statistical chart of protein peptide identification results. (B) Principal component analysis (PCA) for the NECC patients, each dot represents a single sample. (C) Volcano plot of protein expression profiles of NECC. (D) Heatmap shows the differentially expressed proteins (DEPs) between paracancerous and tumor samples of NECC

Identification of the NECC-specific dysregulated proteins

To identify NECC-specific dysregulation in proteins, we conducted a systematic analysis by first collecting protein and gene expression profiles from both TCGA-retrieved and literature-reported samples of CSCC and ECA, as well as their respective paracancerous controls [16–21]. Subsequently, we focused our comparison on differentially expressed proteins (DEPs) and differentially expressed genes (DEGs) derived from two distinct cohorts: ‘NECC versus NECC paracancerous control’ and ‘CSCC and ECA versus their respective paracancerous controls’. This allowed us to isolate NECC-specific DEPs and DEGs by contrasting them with those found in other cervical carcinoma types (CSCC and ECA). Specifically, we compared NECC-DEPs with CSCC and ECA-DEPs, along with CSCC and ECA-DEGs (Fig. 2A, B, Tables S3 and S4). Our approach ensures that any identified NECC-specific proteins or genes are truly distinct from those observed in other related cancer types, thus providing insights into the unique protein expression landscape of NECC.

Fig. 2.

Identification of NECC-specific DEPs. (A) Schematic view of the procedures for analyzing NECC-specific DEPs. (B) Integrating the amount of differentially expressed proteins in CSCC and ECA of cervical cancers from literature reports. (C) Upset diagram shows the intersection of NECC-DEPs with CSCC and ECA-DEPs. Chromosome map shows the chromosomal location and up- or down-regulation of NDEPs. (D) Volcano plot of genes expression profiles of CSCC and ECA of cervical cancer. (E) Venn diagram of NDEPs and DEGs in CSCC and ECA of cervical cancer

In Fig. 2C, it was observed that a total of 142 DEPs were exclusively displayed in NECC samples, distinct from CSCC and ECA specimens, designated as proteins only differentially expressed in NECC (NDEPs). Moreover, we also analyzed DEGs from TCGA-retrieved CSCC and ECA samples, as several studies have also highlighted the potential of transcriptional biomarkers as diagnostic signatures and therapeutic targets for cervical cancer [36–39], and compared them with the aforementioned NDEPs (Fig. 2D, Table S5). Among them, 83 NDEPs also did not exhibit dysregulation at the mRNA level in TCGA-retrieved CSCC and ECA samples (Fig. 2E). Together, we identified 83 NECC-specific DEPs that did not dysregulate in both mRNA and protein levels in CSCC and ECA, namely NECC-specific DEPs (NsDEPs).

By utilizing the Gene Ontology database for enrichment analysis, we observed that the function of the NsDEPs was distinct from the specific DEPs or DEGs in CSCC and ECA samples. Biological processes and molecular functions enriched for NsDEPs mainly included “structure molecule activity”, “biological adhesion”, “cell adhesion”, “calcium ion binding”, and “cytoskeletal protein binding”. As a contrast, biological processes and molecular function enriched for specific DEPs in CSCC and ECA mainly consisted of “cell adhesion molecule binding”, “cell death”, “programmed cell death”, “cellular response to stress”, “epithelial cell differentiation”, “cellular development process”, “ATPase activity”, “cell differentiation”, “RNA binding”, “cadherin binding”, “transport”, “immune effector process”, “mRNA metabolic process”, “mRNA catabolic process”, “protein localization”, “RNA catabolic process”, “nucleic acid binding”, “activation of immune response”. And biological processes and molecular functions enriched for specific DEGs in CSCC and ECA are mainly comprised of “tissue development”, “epithelium development”, “epithelial cell differentiation”. “mitotic cell cycle”, “cell division”, “epidermal cell differentiation”, “mitotic cell cycle process”, “cell cycle”, “cell proliferation”, “DNA replication”, “cell cycle process”, “cell adhesion”, “cell cycle phase transition”, “mitotic nuclear division”, “DNA-dependent DNA replication”, “regulation of chromosome segregation”, “regulation of epithelial cell differentiation”, “sequence-specific DNA binding”, “cell cycle G1/S phase transition”, “G1/S transition of the mitotic cell cycle” (Figure S3).

SCGN, CAP2, and CACYBP were identified as key NECC-specific DEPs

To further identify the pivotal NsDEPs contributing to NECC, the aforementioned NsDEPs were integrated into a machine learning-based integration framework. The present study strategically combined multiple algorithms by employing a 10-fold cross-validation approach with the assistance of C-index performance metrics in order to identify the combination with the best results. As shown in Figure S4 and Table S6, the SVM-RFE, RF, and Lasso combined model performed best, with the highest average C-index among the 66 combinations. Next, we applied this combination to characterize key NsDEPs. Initially, we utilized the SVM-RFE model to identify proteins with the minimum cross-validation error, leading to the selection of 26 candidate proteins (Fig. 3A). Additionally, the NsDEPs were incorporated into the random forest model, resulting in a minimized cross-validation error of 39 trees. Consequently, 23 candidate proteins with important points greater than 0.15 were identified (Fig. 3B). Moreover, we employed the LASSO algorithm to identify variations in the regression coefficients of the NsDEPs. This screening process resulted in the identification of 6 candidate proteins (Fig. 3C). To summarize, the SVM-RFE algorithm identified 26 candidates, the random forest algorithm identified 23 candidates, and the LASSO algorithm identified 6 candidates (Fig. 3D). Upon intersecting all the candidate proteins, it was observed that SCGN, CAP2, and CACYBP were consistently identified by all three machine learning approaches and therefore defined as key NsDEPs (kNsDEPs).

Fig. 3.

Identification of key NECC-specific DEPs (kNsDEPs). (A) SVM–RFE algorithm identified 26 candidates with an accuracy of 1 (left panel) and an error of 0 (right panel). (B) RandomForest algorithm identified 23 candidates. RandomForest error rate versus the number of classification trees (left panel) and protein importance scores (right panel). (C) LASSO coefficient profiles of the indicated NECC-specific DEPs (left panel). The lambda.min parameter indicates that 6 candidates are preserved (right panel). (D) Venn plot shows the overlapped kNsDEPs

Validation of the specific dysregulated patterns of kNsDEPs in NECC

To further confirm the specific dysregulated patterns of SCGN, CAP2, and CACYBP in NECC, we collected normal cervix specimens from patients with hysteromyomectomy and paracancerous tissues adjacent to CSCC and NECC as additional controls for CSCC and NECC. Then, we detected the expressions of these kNsDEPs in the indicated samples. Our results showed that the expression levels of SCGN, CAP2, and CACYBP remain unchanged in normal cervix specimens and paracancerous tissues but are dramatically altered in NECC specimens (Fig. 4, and S5). In addition, compared with the expression of these kNsDEPs in normal cervix specimens and paracancerous tissues, IHC results also demonstrated that the upregulated expression of SCGN and CACYBP, along with the downregulation of CAP2, were exclusively existed in NECC, other than CSCC (Fig. 4, and S5). Collectively, we observed a NECC-specific tumoral high-expression of SCGN and CACYBP, and tumoral low-expression of CAP2.

Fig. 4.

IHC staining of kNsDEPs in normal cervix specimens, paracancerous and tumor tissues of CSCC and NECC. (A-C) Representative IHC images of SCGN (A), CACYBP (B), and CAP2 (C) in the indicated specimens (n = 3). (D-F) Quantitative statistical analysis of protein levels of SCGN (D), CACYBP (E), and CAP2 (F) according to (A-C) and Figure S5. (n = 3) (*p < 0.05, **p < 0.01, ***p < 0.001; ns: not significant)

kNsDEPs have the potential to be diagnostic biomarkers for NECC

To investigate the diagnostic potential of kNsDEPs, we assessed the receiver operating characteristic (ROC) curve of SCGN, CAP2, and CACYBP in the evaluation of diagnostic tests for NECC patients. Our analysis revealed that all three kNsDEPs could effectively differentiate NECC from normal controls, displaying high accuracy as indicated by the AUC scores of SCGN = 0.816, CACYBP = 1.000, and CAP2 = 1.000 (Figure S6A). In addition to kNsDEPs, we also collected other commonly used diagnostic biomarkers for cervical cancer. These include neuroendocrine markers such as SYP, NCAM1, CHGA, and ENO2, CSCC markers such as SCC antigen, including SERPINB3 and SERPINB4, and ECA markers such as carcinoembryonic antigens (CEACAMs) and CA125 (also known as MUC16) [6, 40–42]. We then applied these markers for the evaluation of diagnostic tests for NECC patients by ROC analyses. Our findings demonstrated that the ROC scores based on currently used cervical cancer diagnostic markers generally had lower accuracy than kNsDEPs (Figures S6B-D), suggesting the better performance of kNsDEPs in NECC diagnosis.

To further validate the diagnostic roles of kNsDEPs, we also constructed a diagnostic model based on the expression of kNsDEPs through logistic regression modeling. The AUC score based on kNsDEPs was 0.959, which is higher than that based on other commonly used diagnostic biomarkers for cervical cancer (Fig. 5A, B and Figures S7A-F). In order to verify the diagnostic effect of kNsDEPs in a broader scenario, we also downloaded omic data from 23 independent non-CSCC and ECA samples in TCGA. The results showed that the kNsDEPs still had a high AUC value of 0.967, which is higher than other commonly used diagnostic biomarkers for cervical cancer, suggesting the promising value of kNsDEPs in non-CSCC and ECA diagnosing (Fig. 5C, D and Figures S7G-L).

Fig. 5.

Nomogram models based on kNsDEPs. (A) Nomogram model for NECC prediction based on SCGN, CAP2, and CACYBP. (B) The AUC score of (A). (C) Nomogram model for NECC prediction from TCGA based on SCGN, CAP2, and CACYBP. (D) The AUC score of (C)

In addition, artificial neural network modeling, as a machine learning algorithm that can simulate the brain, is widely useful in the clinical diagnosis process [43]. Therefore, we also constructed an artificial network model based on the expression of the identified kNsDEPs to validate the above findings. The results showed that the kNsDEPs still had a high AUC value of 0.929 (Figures S8A, B). And the AUC values based on other commonly used diagnostic biomarkers for cervical cancer were 0.786, 0.714, and 0.714, respectively (Figures S8C-H). This finding still holds true in the independent external dataset from TCGA, with AUC values of 0.900 based on kNsDEPs and AUC values of 0.800, 0.790, and 0.755 based on other commonly used diagnostic biomarkers for cervical cancer (Figures S8I-P). In conclusion, we demonstrated that the kNsDEPs are promising diagnostic biomarkers for NECC through multiple predictive models.

kNsDEPs play roles in regulating cytoskeleton-related functions of NECC

To further explore the potential regulatory roles of the kNsDEPs, we established a protein-protein interaction (PPI) network among these kNsDEPs and other proteins that were confirmed to be differentially expressed in NECC tissues (NECC-DEPs), utilizing the STRING database. This network, which comprises 38 nodes (representing unique proteins) and 94 edges (indicating interactions between proteins), provides a visual representation of the potential interactions among these proteins in the context of NECC (Fig. 6). Importantly, all proteins included in this PPI network were based on empirical proteomic data demonstrating their differential expression in NECC, thereby mitigating concerns about the inclusion of proteins without evidence of expression in NECC. Subsequently, we performed functional enrichment analysis to gain insights into the biological processes and pathways potentially regulated by these interacting proteins. Consequently, our results identified multiple pathways enriched for the kNsDEPs, including “actin filament binding”, “actin binding”, “protein homodimerization activity”, “cytoskeletal protein binding”, “actin-mediated cell contraction”, “positive regulation of exit from mitosis”, “positive regulation of neuron projection development”, “cellular protein-containing complex assembly”, “regulation of exit from mitosis”, “positive regulation of cellular component organization”, “actin filament-based process”, “regulation of protein polymerization”, “actin cytoskeleton organization”, “protein dimerization activity”, “cell cycle phase transition”, “estrogen receptor binding”, “G1/S transition of mitotic cell cycle”, “cell-matrix adhesion”, “cytoskeleton organization”, “regulation of macromolecule biosynthetic process”, “regulation of actin filament-based process” (Fig. 6).

Fig. 6.

Gene ontology analysis of kNsDEPs. Protein-protein interaction (PPI) network of kNsDEPs and other DEPs in NECC (left panel). Functional enrichment of kNsDEPs and their associated DEPs (right panel)

Next, we searched the literature to find more evidence of protein-protein associations that have been documented in the context of biological systems, such as a variety of cancer cell lines, in order to support the aforementioned prediction results by STRING. This search allowed us to compile a refined set of kNsDEPs-associated NECC-DEPs, for which there is existing experimental or literature-based evidence that support the direct or indirect interactions, co-expressions, and/or regulatory networks among kNsDEPs and NECCC-DEPs. Based on this refined association dataset, we performed a new functional enrichment analysis to further validate the biological processes and molecular functions potentially regulated by the kNsDEPs. Remarkably, the updated functional enrichment analysis revealed that several terms, including a common pathway “cytoskeletal protein binding”, were significantly enriched, reflecting the pivotal role of kNsDEPs in cytoskeleton-related functions (Figure S9). Cytoskeleton is an intracellular structure made up of microtubules, microfilaments, and intermediate filaments. Together, they determine the mechanical properties of cells [44]. In normal cells, these components are highly integrated, and their functions are well-coordinated. However, mutations and abnormal expression of cytoskeletal and cytoskeletal-associated proteins can contribute to cancer cells’ resistance to chemotherapy and ability to metastasize [45, 46]. The functional enrichment of kNsDEPs in multiple cytoskeleton-related pathways indicates SCGN, CAP2, and CACYBP might contribute to NECC development through regulating cytoskeleton-related functions.

Gene expression profiling of kNsDEPs in other neuroendocrine cancer types

To explore whether the expression signature of kNsDEPs is specific to NECC or shared across other neuroendocrine cancer types, we extended our investigation by analyzing gene expression data from other neuroendocrine tumor in Gene Expression Omnibus database (GEO, https://www.ncbi.nlm.nih.gov/geo/). After filtering datasets to exclude those lacking extractable data or appropriate normal sample controls, we included four neuroendocrine cancer datasets: pancreatic neuroendocrine tumors (PanNET, GSE73338), and small intestine neuroendocrine tumors (SI-NETs, GSE65286), small cell lung cancer (SCLC, GSE43346), lung cancer large cell neuroendocrine (LCNEC, GSE30219). Our results revealed, while individual genes displayed some degree of dysregulation in these neuroendocrine carcinoma, the distinct expression pattern observed for the three kNsDEPs in NECC did not replicate across these types of neuroendocrine carcinoma (Fig. 7A-D). This suggests that the differential expression of kNsDEPs might be specific to NECC rather than a generalized feature of neuroendocrine tumors. It implies that these genes may play pivotal roles in the underlying molecular mechanisms driving NECC pathogenesis, thus again presenting themselves as promising diagnostic tools.

Fig. 7.

The differential expression analyses of kNsDEPs in neuroendocrine carcinomas. (A-D) The results of kNsDEPs in pancreatic neuroendocrine tumors (PanNET) (Normal, n = 5, PanNET, n = 63) (A), small intestine neuroendocrine tumors (SINET) (Normal, n = 10, SINET, n = 33) (B), small cell lung cancer (SCLC) (Normal, n = 1, SCLC, n = 21) (C), and lung cancer large cell neuroendocrine (LCNEC) (Normal, n = 14, LCNEC, n = 56) (D). (*p < 0.05, **p < 0.01, ***p < 0.001)

Discussion

Tumors of the female reproductive tract represent a significant area of focus for the advancement of diagnostic, prognostic, and therapeutic strategies [47]. In recent research, proteomics has emerged as a promising approach to achieving more effective diagnostic outcomes. Gynecological oncology already utilizes several valuable protein biomarkers, including human epididymis protein 4 (HE4), chorionic gonadotropin β (βhCG), carbohydrate antigen 125 (CA125), soluble Fas (sFas), tumor-associated glycoprotein 72 (CA72-4), and serum levels of immunosuppressive acidic protein [47–49]. However, the current diagnostic tools used for detecting NECC, a rare form of cervical cancer, are not adequately effective, lacking the necessary sensitivity and specificity. Therefore, clinical proteomics presents a promising avenue for acquiring comprehensive knowledge of the underlying mechanisms of NECC neoplastic processes. This, in turn, will facilitate the discovery and characterization of novel, specific, and sensitive markers. In this study, we collected NECC tissue and para-carcinoma tissue samples. Using proteomic strategies, we identified 211 differentially expressed proteins in NECC specimens. Of these, 83 were specifically dysregulated in NECC rather than in other common forms of cervical cancer, indicating their potential as NECC-specific biomarkers.

Currently, specific and effective diagnostic or treatment guidelines available for NECC are lacking due to its rarity [11]. Several authors have reported various multimodality approaches, primarily derived from neuroendocrine tumors and cervical cancer in general [5, 12, 13]. However, NECC displays unique biological characteristics in comparison to CSCC and ECA [11]. In our analyses, we also found the molecular landscapes are totally distinct between NECC and cervical cancer in general. The dysregulated proteins specific to NECC were primarily enriched in molecular functions related to “actin filament binding”, “regulation of protein polymerization”, and “cytoskeletal protein binding”, whereas the dysregulated proteins and genes associated with CSCC and ECA were mainly enriched in molecular functions such as “cell adhesion molecular binding”, “RNA binding”, “cadherin binding”, “nucleic acid binding”, “sequence-specific DNA binding”.

To identify the kNsDEPs, an integrative machine learning algorithm was applied. We identified SCGN, CAP2, and CACYBP as central proteins in NECC. Functional enrichment analyses also identified these kNsDEPs as regulating cytoskeleton-related functions. Abnormal cytoskeleton dynamics are associated with a range of diseases, such as cancer, as well as immunological and neurological disorders [50]. For instance, SCGN can regulate actin dynamics and focal adhesion, thereby affecting insulin secretion in pancreatic β-cells. Silencing SCGN can inhibit multiple classic tumor-related activities, such as the activation of FAK, paxillin, ERK1/2 and AKT induced by glucose and hydrogen peroxide [51]. CAP2 is a highly conserved protein and plays essential roles in regulating the actin cytoskeleton and signal transduction [52, 53]. In the context of epithelial ovarian cancer, dysregulation of CAP2 has been observed to correlate with aggressive histologic types and poorer outcomes [54]. The tumorigenic role of CACYBP has been widely reported in a variety of tumors, including hepatocellular carcinoma, clear cell renal cell carcinoma, pancreatic cancer, and colorectal cancer [55–59]. And relatively limited reports have shown the relationship between CACYBP and cytoskeleton regulation by Rho GTPase, suggesting the potential of CACYBP to regulate cytoskeleton-related functions in tumors [60].

Additionally, we also explored the potential of kNsDEPs as diagnostic biomarkers. In the present study, we integrated the aforementioned kNsDEPs into a set of predictive models with the objective of distinguishing between paracancerous and NECC samples. The accuracy of the prediction based on the kNsDEPs-generated predictive model was markedly superior to that achieved using the currently utilized biomarkers in neuroendocrine carcinoma, CSCC and ECA. Moreover, an independent cohort of 23 patients was evaluated to assess the diagnostic efficacy of kNsDEPs. The results demonstrated that kNsDEPs exhibited superior accuracy in diagnosing non-SCC and ECA patients compared to other prevalent diagnostic biomarkers for cervical cancer. Combined with the expression specificity of kNsDEPs in NECC, authenticated by immunohistochemical staining comparisons between NECC and SCC samples, our results showed SCGN, CAP2, and CACYBP expression signature is a promising index for NECC diagnosing.

Certainly, the current study has some limitations. Firstly, due to the rarity of NECC, we were only able to obtain 14 pure NECC samples for constructing the diagnostic prediction model based on protein expression signatures. In the future, it will be necessary to include a larger cohort of patients to validate our findings. Secondly, since there is a shortage of commercial NECC cell and animal models, the functions and regulatory roles of kNsDEP revealed in this study remain at the level of scientific speculation. Despite the identification of potential biomarkers and regulatory mechanisms, we have not yet validate these findings in a biological context. Therefore, further in-depth investigation into the regulatory role of kNsDEPs on NECC requires the development of biological models and the conduct of related functional experiments in the future.

Conclusion

In summary, this paper presents proteomics analyses that unveil the protein landscape of patients with NECC. By integrating multiple proteomics and transcriptome data, we successfully discriminated dysregulated proteins specific to NECC (referred to as NsDEPs). Through the utilization of various machine learning algorithms, we further identified key NsDEPs as key proteins (referred to as kNsDEPs) and validated their potential for diagnostic purposes. Additionally, we investigated the regulatory effects of these kNsDEPs on the pathogenesis of NECC and revealed potential therapeutic drugs. These findings not only contribute to our understanding of NECC’s pathogenesis but also provide a new perspective on diagnostic and treatment strategies for NECC.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

T.S. was responsible for Writing - Original Draft, Writing - Review & Editing, Conceptualization, and Methodology. T.D., H.W., J.Z., and X.Z. were responsible for Formal analysis, Investigation, and Visualization. B.T., F.J., S.W., and Q.W. were responsible for Resources and methodology. Y.D., Y.D., W.C., and Y.W. were responsible for Data Curation and methodology. All authors reviewed the manuscript.

Funding

This study was supported by grants from the National Natural Science Foundation of China (32300450), the Health Research Program of Anhui Province (AHWJ2023BBa20033), the 2023 Wuhu Science and Technology Plan Project (2023jc14). National Undergraduate Training Program for Innovation and Entrepreneurship (202310370158), and the platforms and resources provided by the Outstanding Innovative Research Team for Molecular Enzymology and Detection in Anhui Provincial Universities (2022AH010012).

Data availability

Data is provided within the manuscript or supplementary information files.

Declarations

Ethics approval and consent to participate

This study was performed in line with the principles of Declaration of Helsinki. The research protocol was approved by the ethical committee for human experimentation of East China Normal University Wuhu Affiliated Hospital (The Second People’s Hospital of Wuhu City) (Approval number 2024-KY-009) and Anhui Normal University (Approval number AHNU-ET2023080). All patients included in this research have granted written informed consent for the use of their data.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.