Plasma protein markers and application thereof in ovarian cancer early diagnosis kit
By screening with high-throughput LC-MS and constructing a machine learning model, an early diagnostic kit for ovarian cancer was developed using 11 plasma protein biomarkers. This solved the problem of low efficiency in the early diagnosis of ovarian cancer in existing technologies and achieved a high-sensitivity and specificity early diagnostic effect.
Patent Information
- Application Number
- CN202511590936.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-03
- Publication Date
- 2026-01-16
AI Technical Summary
Existing ovarian cancer screening markers, such as CA125, have a high false positive rate, moderate sensitivity, and low specificity, making it difficult to achieve efficient diagnosis of early ovarian cancer, resulting in most patients being diagnosed at an advanced stage.
By detecting plasma proteins using high-throughput LC-MS, more than 50 biomarkers important for the early diagnosis of ovarian cancer were screened. A machine learning model was constructed, and combined with 11 specific plasma protein biomarkers (HBD, CRP, PF4, APOA4, CA3, BPGM, BLVRB, HBB, HBA1, SELENBP1, CILP2, etc.), a kit for the early diagnosis of ovarian cancer was developed.
The model achieved high-precision differentiation between early-stage ovarian cancer and benign cysts. In independent external validation, the diagnostic model achieved an AUC of 0.96 and an accuracy of 0.89, demonstrating high sensitivity and specificity, and significantly improving the early detection rate of ovarian cancer.
Smart Images
Figure CN121347714A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of biological detection technology, and in particular to a group of plasma protein markers and their application in an early ovarian cancer diagnosis kit. BACKGROUND
[0002] Ovarian cancer is one of the most common malignant tumors in gynecology. Statistical data shows that ovarian cancer is the third most common cancer in women, ranking only behind cervical cancer and uterine body malignancy, and it is the first fatal cancer in gynecology. Its prognosis is extremely poor. According to the statistics of the National Institutes of Health, by 2024, the 5-year survival rate of patients diagnosed with ovarian cancer is only 50.9%. About 140,000 people die from ovarian cancer every year. The poor survival rate of ovarian cancer is closely related to the lack of rapid and effective diagnostic tools, especially the lack of specific symptoms in the early stage of the disease. Most patients are diagnosed with local or distant metastasis, and about 60% of ovarian cancer patients are diagnosed with advanced ovarian cancer.
[0003] The most commonly used screening marker for ovarian cancer in clinical practice is carbohydrate antigen 125 (CA125). It is a large transmembrane glycoprotein derived from the body cavity (pericardium, pleura and peritoneum) and Mϋller tube (fallopian tube, endometrium and endocervical epithelial cells). Therefore, it can be abnormally elevated in many diseases, including but not limited to ovarian cancer, cervical cancer, breast cancer, lung cancer, endometriosis, ovarian cyst, cirrhosis, cardiovascular disease, etc. However, its high false positive rate, moderate sensitivity and low specificity are still far from the ideal screening index. Even the use of CA125 for early ovarian cancer screening has not improved survival rates, and the Food and Drug Administration (FDA) has only approved its use for detecting the progression of epithelial ovarian cancer patients after treatment and prognosis evaluation. Therefore, markers for early screening of ovarian cancer still need to be further explored.
[0004] The present application measures hundreds of proteins by mass spectrometry detection of plasma proteins by high-throughput LC-MS, and applies machine learning to preliminarily screen more than 50 markers important for early diagnosis of ovarian cancer. Further verification of target proteins will update the model and external verification of proteins with consistent trends from previous mass spectrometry measurements, successfully constructing an excellent mathematical model. These markers combined with the model are expected to become an important tool for ovarian cancer screening. SUMMARY
[0005] The present application aims at the deficiencies of the prior art, and provides a group of plasma protein markers and application of the group of plasma protein markers in preparation of an ovarian early diagnosis kit, The present application is realized by the following technical scheme: a group of plasma protein markers, the markers including the following 11 proteins: HBD (Uniprot code: P02042), CRP (P02741), PF4 (P02776), APOA4 (P06727), CA3 (P07451), BPGM (P07738), BLVRB (P30043), HBB (P68871), HBA1 (P69905), SELENBP1 (Q13228) and CILP2 (Q8IUL8).
[0006] In a second aspect, the present application further provides a method for obtaining the group of plasma protein markers, and the specific steps are as follows: (1) Collecting clinical patient plasma samples and performing pretreatment, including removing high-abundance proteins, lysis, alkylation, trypsin digestion, desalination and TMT-16 labeling, and the like. Subsequently, high-throughput LC-MS technology is used to obtain plasma protein mass spectrum data.
[0007] (2) Data conversion and extraction: the original raw file is converted into mzXML format by using open source software MSconvert, and unpacking and extraction operations are performed by using fragpipe.
[0008] Further, the protein expression profile data of the training set population cohort is obtained, and the following methods are used for quality control, bioinformatics analysis and machine learning modeling: (1) The proteins with a number of peptide spectrum matches (PSM) less than or equal to 2 in mass spectrum detection are removed, and the proteins that are not detected in more than 50% of the samples in the total training cohort are also removed. For protein missing values, 80% of the lowest abundance between samples is used for filling treatment. Subsequently, the Combat algorithm in the SVA R package is used to correct the batch effect that may exist between samples; (2) After the above processing, the batch effect of the protein expression profile, the correlation between the Pool mixed samples and the Batch ordinary sample variation coefficient CV value, and the correlation between the Pool mixed samples are systematically analyzed, and it is confirmed that the data quality of the training cohort meets the requirements of subsequent modeling analysis; (3) After obtaining the protein expression profile through the above steps, the rank sum test method is used in R language to find the differentially expressed proteins between the early ovarian cancer OV-S1 and the benign cyst CYS two groups of patients, and between the late ovarian cancer OV and the benign cyst CYS two groups, and the Benjamin-Hochberg method is used for multiple hypothesis correction of statistical test results, and finally two candidate protein feature lists are obtained as input for entering the machine learning feature screening process; (4) For the candidate protein feature list of the two classification targets, first, the original expression value and the method difference are normalized, the logistic regression is used as the benchmark model, and the recursive feature elimination (RFE) method is used for feature screening, and different screening standards are set according to the characteristics of different classification model tasks: for the classification of early ovarian cancer and benign patients, the F1-score is used as the screening index, aiming to find as many suspected early ovarian cancer cases as possible, and taking into account the two core goals of "finding all" and "finding accurate"; and for the classification of late stage ovarian cancer and benign patients, the binary classification evaluation index AUC is used as the model optimization standard, aiming to achieve the best binary classification performance; finally, the protein feature subsets screened in the two classification tasks are obtained, and then the grid search method is used to optimize the hyperparameters of the logistic regression model, and the cross-validation is used to evaluate the generalization ability and stability of the model; (5) For the two ovarian cancer-benign cyst classification models and their features obtained above, parallel reaction monitoring (PRM) is used to verify in the external verification set queue; in the protein expression matrix in the external verification set, first, the unqualified protein features due to sample quality problems are deleted; then, the features with the same expression trend as the training set are selected; through the intersection analysis of the features selected by the early ovarian cancer classification model and the ovarian cancer classification model of all stages, 11 proteins are obtained as the final features, and are used to verify the classification ability in the early ovarian cancer samples and all ovarian cancer populations.
[0009] In a third aspect, the application further provides a use of the plasma protein marker according to claim 1 in the preparation of a kit for early diagnosis of ovarian cancer.
[0010] Further, based on the 11 protein features screened by the plasma protein marker acquisition method, the tumor tissue samples, endoscopic biopsy samples or plasma / serum samples are analyzed by using the proteomics sequencing method, a mathematical model for diagnosing ovarian cancer is constructed by using the logistic regression method, and the model effect is evaluated by using the ROC curve AUC. In the external verification, the AUC of the model of the application reaches 0.96, and the accuracy is 0.89, showing high sensitivity and specificity.
[0011] The beneficial effects of the application are: based on the markers of the application, a mathematical model for diagnosing colon cancer is constructed; the model has high sensitivity and good specificity, and in the independent external verification set of the application, the AUC of 0.96 and the accuracy of 0.89 are achieved for the diagnosis and classification of early ovarian cancer and benign cysts. In summary, the protein molecular markers disclosed by the application have good diagnostic index characteristics, can be effectively used for early diagnosis of ovarian cancer, and have high clinical application and popularization value. BRIEF DESCRIPTION OF DRAWINGS
[0012] Figure 1 is a flowchart of an embodiment of the application.
[0013] Figure 2 is the feature screening stage of the application, and the volcano plot of differentially expressed proteins. The horizontal axis is the differential expression fold (logFC), and the vertical axis is the statistical test significance. In which A is the differentially expressed protein between OV-S1 (early ovarian cancer) and CYS (benign cyst), and B is the differentially expressed protein between OV (late stage ovarian cancer) and CYS.
[0014] Figure 3 is the classification of OV-S1 (early ovarian cancer) and CYS (benign cyst) and OV (late stage ovarian cancer) in the training set and external verification set of the application. In which A is the ROC curve of the application in the internal training set (3-fold cross-validation); B is the ROC curve in the external verification set.
[0015] Figure 4 is the classification feature importance (SHAP interpretability algorithm) of the 11 key protein features finally selected by the application between the whole OV (all ovarian cancer patients) and CYS (benign cyst). DETAILED DESCRIPTION
[0016] The application will be further described below through specific embodiments, and it should be noted that the following description is only for the purpose of explaining the application and does not limit its content.
[0017] As Figure 1As shown, this invention provides a set of plasma protein biomarkers and their application in the preparation of a kit for early ovarian diagnostics. The kit is used to detect a set of specific plasma protein biomarkers, combined with a machine learning model, to achieve high-precision differentiation between ovarian cancer, particularly in its early stage (OV-S1), and benign cysts (CYS). This set of protein biomarkers includes the following 11 proteins: HBD (Uniprot code: P02042), CRP (P02741), PF4 (P02776), APOA4 (P06727), CA3 (P07451), BPGM (P07738), BLVRB (P30043), HBB (P68871), HBA1 (P69905), SELENBP1 (Q13228), and CILP2 (Q8IUL8).
[0018] The method for obtaining a set of plasma protein biomarkers and the process for early diagnosis of ovarian cancer provided by this invention are as follows: 1. Clinical Sample Collection and Pretreatment: Peripheral blood samples were collected from patients with pathologically confirmed ovarian cancer (including early OV-S1 and late OV stages) and patients with benign cysts (CYS). Plasma was immediately separated by centrifugation after collection and stored at -80°C. All samples underwent pretreatment under uniform conditions: First, high-abundance proteins such as IgG and albumin were removed from the plasma using a High-Select™ removal centrifugation column (ThermoFisher Scientific, USA, catalog number: A36372); then, protein lysis, reduction, and alkylation were performed; followed by two-stage digestion using mass spectrometry-grade trypsin (ThermoFisher Scientific, USA, catalog number: 90059); the digestion products were purified by desalting using a desalting column, labeled with TMT-16plex isotope tags, and mixed in equal volumes. After desalination using a C18 solid desalination column, fractionation was performed in a weakly alkaline environment, resulting in 60 fractions. These fractions were then combined in the order of 1+60, 2+59, and 3+58 to synthesize 30 peptide fragment mixtures, which were then rotary evaporated for subsequent mass spectrometry detection.
[0019] 2. High-throughput mass spectrometry data acquisition: The peptide mixture obtained in step (1) was detected by high-throughput liquid chromatography-tandem mass spectrometry (LC-MS / MS). An Orbitrap 480 Exploris mass spectrometer was used, with a reversed-phase C18 column for peptide separation and a gradient elution program of 60 minutes. Data-dependent acquisition (DDA) was performed after acquiring precursor ions, with a dynamic exclusion time of 30 seconds.
[0020] 3. Data Conversion and Protein Quantitative Extraction: The raw mass spectrometry data file (.raw format) obtained in step (2) was converted to the standard mzXML format using the open-source tool MSconvert (ProteoWizard suite). Subsequently, database searching and quantitative analysis were performed using the FragPipe bioinformatics workflow. Search parameter settings: Referencing the human Uniprot / Swiss-Prot database (2023 version), a maximum of two missed cleavage sites were allowed. Fixed modifications were TMT-16 tags (lysine and N-terminus) and cysteine alkylation (+57.02 Da), with variable modifications including methionine oxidation (+15.99 Da) and asparagine / glutamine deamidation (+0.98 Da). The FDR for both peptide and protein levels was controlled within 1%, and the quantitative results were calculated based on the normalized TMT reporter ion intensity.
[0021] 4. Quality Control and Standardization of Protein Expression Profiles: The protein expression matrix output in step (3) was subjected to strict quality control. Proteins that were not detected in more than 50% of the total training queue (missing rate > 50%) were removed, along with low-confidence proteins with a peptide support number (PSM) ≤ 2. For the remaining protein abundance data with missing values, the strategy of "lowest fold of detection mean (0.8 times)" was used to fill in the missing values. To eliminate systematic bias between different TMT batches, the ComBat algorithm (R package sva) was used for batch effect correction to ensure data comparability.
[0022] 5. Differentially expressed protein screening and candidate biomarker identification: such as... Figure 2 As shown, in the R language environment, the Wilcoxon rank-sum test was used to identify differentially expressed proteins (DEPs) for the two classification tasks "OV-S1 vs CYS" and "OV vs CYS". The p-values were corrected for multiple hypothesis testing using the Benjamini-Hochberg method, and proteins with adjusted p-values (adj. p-value) < 0.05 and |log2 fold change| > 0.25 (i.e., fold change > 1.189) were selected as candidate differentially expressed proteins to construct an initial feature list.
[0023] 6. Machine Learning Model Construction and Feature Optimization: Based on the candidate protein features obtained in step (5), a logistic regression classification model is constructed. Recursive Feature Elimination (RFE) combined with cross-validation is used for feature selection, gradually eliminating redundant or low-contribution variables, and finally determining the optimal feature subset. The model hyperparameters (such as regularization strength C) are optimized through grid search, and the model performance is evaluated using 5-fold cross-validation on the training set to ensure model stability and generalization ability.
[0024] 7. External Validation and Biomarker Confirmation: The characteristic proteins and their corresponding models selected in step (6) were validated in a separate external validation cohort. This validation cohort consisted of multicenter clinical samples, and the expression levels of the target proteins were requantified using parallel reaction monitoring (PRM) targeted mass spectrometry. Figure 4 As shown, a biomarker combination consisting of 11 proteins—HBD, CRP, PF4, APOA4, CA3, BPGM, BLVRB, HBB, HBA1, SELENBP1, and CILP2—exhibited excellent performance in differentiating early-stage ovarian cancer from benign cysts. Figure 3 As shown, the area under the ROC curve (AUC) reached 0.97, the accuracy was 0.89, and the sensitivity and specificity were 0.91 and 0.87, respectively.
[0025] 8. Diagnostic Model Construction and Performance Evaluation: Based on the expression levels of the aforementioned 11 protein biomarkers, a final logistic regression mathematical model for the early diagnosis of ovarian cancer was established. The model output is a disease probability value, which can be used to achieve binary classification decisions by setting a threshold. In multiple independent validation sets, the model demonstrated good stability and high discriminative ability (AUC range: 0.94–0.97), confirming its potential for clinical translation. Additionally, the importance of the 11 protein biomarkers in the mathematical model was evaluated using the SHAP interpretive algorithm, providing prior results for subsequent downstream mechanism studies.
[0026] In summary, the set of plasma protein biomarkers and the machine learning diagnostic model constructed based on them disclosed in this invention exhibit high sensitivity, high specificity, and good reproducibility. They can effectively distinguish ovarian cancer patients (especially early-stage patients) from individuals with benign lesions, significantly improving the early detection rate of ovarian cancer. This achievement can be applied to the development of clinical blood testing kits, possessing significant medical value and broad industrialization prospects.
[0027] The above embodiments are used to explain and illustrate the present invention, but not to limit the present invention. Any modifications and changes made to the present invention within the spirit and scope of the claims shall fall within the protection scope of the present invention.
Claims
1. A panel of plasma protein markers characterized in that, The UniprotKB of the markers is as follows: P02042, P02741, P02776, P06727, P07451, P07738, P30043, P68871, P69905, Q13228, Q8IUL8, and the corresponding coding genes are HBD, CRP, PF4, APOA4, CA3, BPGM, BLVRB, HBB, HBA1, SELENBP1, and CILP2, respectively.
2. A method for obtaining the plasma protein marker according to claim 1, characterized in that, The method comprises the following steps: (1) After collecting the clinical patient plasma samples, the samples are subjected to the pretreatment processes of high-abundance removal, lysis, alkylation, trypsin digestion, desalting, and TMT-16 labeling, and then high-throughput LC-MS is performed to obtain plasma protein mass spectrum data; (2) Data conversion, unpacking, and extraction: the raw files are converted into mzXML files by using MSconvert software, and then unpacking and extraction are performed by using fragepipe.
3. The method for obtaining the plasma protein markers according to claim 2, wherein the protein expression profile data of the training set population cohort is obtained, and the following methods are used for quality control, bioinformatics analysis, and machine learning modeling: (1) The proteins with a number of peptide spectrum matches (PSM) less than or equal to 2 in the mass spectrum detection are removed, and the proteins that are not detected in more than 50% of the samples in the total training cohort are also removed. For the missing values of the proteins, 80% of the lowest abundances among the samples are used for filling treatment. Then, the Combat algorithm in the SVA R package is used to correct the batch effects that may exist among the samples; (2) After the above processing is completed, the batch effects of the protein expression profile, the correlation between the Pool mixed samples and the Batch ordinary samples, and the correlation between the Pool mixed samples are systematically analyzed to confirm that the data quality of the training cohort meets the requirements of subsequent modeling analysis; (3) After the protein expression profile is obtained through the above steps, the rankSum test method in the R language is used to find the differentially expressed proteins between the early ovarian cancer OV-S1 and the benign cyst CYS groups, and between the advanced ovarian cancer OV and the benign cyst CYS groups, and the Benjamin-Hochberg method is used to correct the multiple hypothesis testing results, and finally two candidate protein feature lists are obtained as inputs for entering the machine learning feature screening process. (4) For the candidate protein feature list of two classification targets, first, the original expression value and method difference are normalized, the logistic regression is used as the benchmark model, and the recursive feature elimination (RFE) method is used for feature screening. According to the characteristics of different classification model tasks, different screening standards are set: for the classification of early ovarian cancer and benign patients, F1-score is used as the screening index, aiming to find as many suspected early ovarian cancer cases as possible, and to balance the two core goals of "finding all" and "finding accurate"; for the classification of late stage ovarian cancer and benign patients, the binary classification evaluation index AUC is used as the model optimization standard, aiming to achieve the best binary classification performance; finally, the protein feature subsets of the two classification tasks are obtained after screening, and then the grid search method is used to optimize the hyperparameters of the logistic regression model, and the cross-validation is used to evaluate the generalization ability and stability of the model; (5) For the two ovarian cancer-benign cyst classification models and their features obtained above, parallel reaction monitoring (PRM) is used to verify the external validation set. In the protein expression matrix in the external validation set, first, the unqualified protein features due to sample quality problems are deleted; then, the features with the same expression trend as the training set are selected; through the intersection analysis of the selected features of the early ovarian cancer classification model and the all-stage ovarian cancer classification model, 11 proteins are obtained as the final features, which are used to verify the classification ability in early ovarian cancer samples and all ovarian cancer populations.
4. Use of the plasma protein markers of claim 1 in the preparation of a kit for early diagnosis of ovarian cancer.
5. Use according to claim 4, characterized in that, Based on the 11 protein features screened by the plasma protein marker acquisition method, the proteomic sequencing method is used to analyze tumor tissue samples, endoscopic biopsy samples or plasma / serum samples, the logistic regression LR method is used to construct a mathematical model for the diagnosis of ovarian cancer, and the ROC curve and AUC are used to evaluate the effect.