Plasma proteome-based colorectal cancer precancerous lesion diagnosis model construction method

Abstract

The invention discloses a plasma proteome-based colorectal cancer precancerous lesion diagnosis model construction method, and aims to solve the problem that a current colorectal cancer diagnosis kit is insufficient in sensitivity and specificity when detecting an early stage (precancerous lesion / early cancer) through a proximity extension technology (PEA) and selecting a specific detection panel. Protein biomarkers in plasma samples of 80 patients with precancerous lesion (progression stage adenoma) and early cancer (Tis / T1 stage) in a discovery queue are identified. By adjusting potential clinical confounding factors, 15 key protein biomarkers are identified. In the discovery queue, four machine learning algorithms (Lasso regression, random forest, SVM and XGBoost) are adopted to diagnose early colorectal cancer lesions, and high diagnosis efficiency is realized through multiple machine learning algorithms. In a verification queue sample, the diagnosis model also shows an important diagnosis capability for patients with progression-stage adenoma or early-stage colorectal cancer, and has a potential clinical application value.

Description

Technical Field

[0001] The present invention relates to the field of medical diagnosis, and in particular to a method for constructing a colorectal precancerous lesion diagnosis model based on plasma proteomes. Background Art

[0002] As a major global public health problem, colorectal cancer ranks third in incidence and second in mortality among all tumors. Early detection of colorectal cancer significantly affects the overall survival rate, and the five-year survival rate of early tumors reaches 90%. The development of sporadic colorectal cancer follows the adenoma-carcinoma sequence, providing a critical window of 5-10 years for early intervention. In the field of early screening, colonoscopy pathological diagnosis is the primary "gold standard". Secondly, non-invasive screening methods based on stool and blood samples, such as FIT, multi-target stool DNA detection and methylated septin9 detection, have shown strong sensitivity and specificity in early detection. In addition, in recent years, a large number of studies have identified many biomarkers for the diagnosis of colorectal cancer through liquid biopsy, including circulating tumor cells, various non-coding RNAs, exosomes and plasma proteins. Among them, the plasma proteome is a hot topic in early diagnosis research due to its specific and stable characteristics, which is significantly correlated with clinical chemistry and disease phenotype.

[0003] The conventional mass spectrometry detection method for diagnosing plasma proteins is currently unable to detect different proteins with huge differences in concentration in the blood, especially most low-abundance proteins (fg / ml). At the same time, the current diagnostic methods (FIT / multi-target fecal DNA testing) have low sensitivity and specificity (less than 0.60) for diagnosing the early stages of colorectal cancer (precancerous lesions / early cancer). Summary of the invention

[0004] The purpose of the present invention is to solve the problem of insufficient sensitivity and specificity in the diagnosis of early stage colorectal cancer and to provide a method for constructing a colorectal precancerous lesion diagnosis model based on plasma proteome.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] A method for constructing a colorectal precancerous lesion diagnosis model based on plasma proteome, comprising the following steps:

[0007] S1. Plasma was collected before endoscopy, stored at -80°C or in liquid nitrogen, and analyzed using Olink Explore 384 Cardiometabolic Panel based on PEA technology;

[0008] S2. 1 mL of serum was mixed with 369 pairs of antibody probes, and markers were generated by DNA polymerase extension, quantified by NGS, and presented by NPX. At the same time, some proteins were verified by ELISA;

[0009] S3. Determine differentially expressed proteins by testing and identify proteins after correcting clinical factors;

[0010] S4. Train LASSO, random forest, SVM, and XGBoost algorithms to build models, evaluate using cross-validation and other methods, calculate the AUC index, and build models in combination with clinical factors.

[0011] Furthermore, in step S2, Olink Explore 384 Cardiometabolic Panel was used for analysis based on proximity extension technology (PEA). 1 mL of serum was mixed with 369 pairs of antibody probes linked to oligonucleotides. After the probes bound to the target antigens, they were extended by DNA polymerase to generate surrogate markers, which were then quantified by next-generation sequencing (NGS). The data were presented as normalized protein expression values ​​(NPX), and randomized sample distribution and internal plate control normalization were performed. Plasma proteins were further verified by enzyme-linked immunosorbent assay (ELISA).

[0012] Furthermore, in step S3, based on the results of the homogeneity of variance test and the normality test, the protein levels were compared using the t test or the Wilcoxon rank sum test, and the plasma proteins with an adjusted p value (FDR) less than 0.05 were defined as differentially expressed proteins. The multivariate linear regression model was used to correct the control group and incorporate clinical factors, and then the proteins were identified.

[0013] Furthermore, in step S4, LASSO regression, random forest, support vector machine (SVM), and XGBoost machine learning algorithms were trained as binary classification models to predict colorectal lesions, and 5-fold cross validation and 0.632+ bootstrap methods were used to evaluate model performance, enhance robustness, and avoid overfitting. The area under the receiver operating characteristic curve (ROC) (AUC) and its 95% confidence interval (CI) were calculated. The sensitivity, specificity, positive predictive value (PPV), and F1 score were calculated at the optimal threshold. The diagnostic efficacy of individual proteins was evaluated by drawing ROC curves. Finally, the key protein markers were combined with the selected clinical predictors to construct clinical risk models and joint models.

[0014] Furthermore, the models were constructed using the scikit-learn 1.3.2 package in Python and implemented in Jupyter Notebook. The key protein markers were combined with the selected clinical predictors to construct the models, specifically:

[0015] LASSO regression model: 5 markers were screened: TGM2, MMP7, GDF15, RNASE3, and REG1B;

[0016] SVM model: 7 markers were screened: RNASE3, TGM2, MMP7, REG1A, REG1B, GDF15, and MFAP5;

[0017] Random forest model: 5 markers were screened: GDF15, REG1B, MMP7, TFF3, and RNASE3;

[0018] XGBoost model: 7 markers were screened out: GDF15, MMP7, TFF3, RNASE3, REG1B, MFAP5, and TGM2.

[0019] The beneficial effects of the present invention are:

[0020] (1) Using the proteomics technology with high sensitivity and wide detection range, proximity extension assay (PEA), to discover diagnostic markers for colorectal precancerous lesions (advanced adenomas) and early cancers (Tis / T1 stage);

[0021] (2) Use a variety of high-precision machine learning models to establish a machine learning diagnostic model with high diagnostic efficacy, and have high diagnostic efficacy in the validation cohort. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 is a flowchart of steps of an embodiment of the present invention;

[0023] Figure 2 Diagnostic model diagrams were established using Lasso regression, random forest, SVM, and XGboost based on diagnostic markers (AUC > 0.90) in the discovery cohort. A, B, C, and D are Lasso regression, random forest, SVM, and XGboost, respectively;

[0024] Figure 3 To verify that the model has a good diagnostic data for early lesions of colorectal cancer in the validation cohort, A is advanced adenoma and B is early cancer. DETAILED DESCRIPTION

[0025] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0026] A method for constructing a colorectal precancerous lesion diagnosis model based on plasma proteome is provided, the steps comprising:

[0027] S1. Plasma was collected before endoscopy, stored at -80°C or in liquid nitrogen, and analyzed using Olink Explore 384 Cardiometabolic Panel based on PEA technology;

[0028] S2. 1 mL of serum was mixed with 369 pairs of antibody probes, and markers were generated by DNA polymerase extension, quantified by NGS, and presented by NPX. At the same time, some proteins were verified by ELISA;

[0029] S3. Determine differentially expressed proteins by testing and identify proteins after correcting clinical factors;

[0030] S4. Train LASSO, random forest, SVM, and XGBoost algorithms to build models, evaluate using cross-validation and other methods, calculate indicators such as AUC, and build models based on clinical factors.

[0031] See also Figure 1 , which are steps of a specific embodiment:

[0032] This example specifically emphasizes and studies advanced adenomas and early colorectal cancer so that they can be detected in time for early cure through endoscopy. The study recruited two cohorts of hospital cases, including a discovery cohort and a validation cohort. The inclusion criteria are as follows: (1) individuals over 18 years old; (2) those who meet the NCCN guidelines for early colorectal cancer and precancerous lesions; (3) cases with gastrointestinal endoscopy and pathology reports. The exclusion criteria are as follows: (1) cases who have received radiotherapy or chemotherapy; (2) cases with severe chronic diseases; (3) cases with other cancers; (4) cases with hemolysis in plasma samples.

[0033] Data collection and histopathological evaluation: The general data of the participants were obtained from the HIS electronic medical record system, including gender, age, BMI, smoking history, drinking history, and other clinical information. In addition, the following clinical data were collected: (1) family history of colorectal adenoma and cancer; (2) lipid levels (cholesterol and triglycerides); (3) recurrence status after endoscopic resection; and (4) histological classification.

[0034] Histological evaluation was based on the Vienna and WHO classification criteria. Advanced adenoma, as a precursor to colorectal cancer, was defined as: (1) adenoma ≥1 mm in diameter; (2) adenoma with tubulovillous or villous histological features; (3) adenoma with high-grade intraepithelial neoplasia. Early colorectal cancer was defined as an invasive tumor confined to the mucosa or submucosa, including Tis and T1 tumors in the AJCC staging system. Advanced neoplasia was defined as early colorectal cancer and advanced adenoma.

[0035] Sample collection and multiplex protein testing: Plasma samples from participants were collected before endoscopy, processed according to standard procedures, and stored at -80°C or in liquid nitrogen. Subsequently, plasma samples were analyzed using the OlinkExplore 384 Cardiometabolic Panel provided by Olink Proteomics. The measurement is based on the proximity extension technique (PEA), where 1 mL of serum is mixed with 369 pairs of antibodies (probes) linked to oligonucleotides. After the probe binds to the target antigen, the oligonucleotide is extended by DNA polymerase to generate a surrogate marker for the specific antigen, which is quantified by next-generation sequencing (NGS). Data are presented as normalized protein expression values ​​(NPX), arbitrary units on a logarithmic (log2) scale, which is used to normalize the data and minimize inter- and intra-assay variability. Samples were randomized on the PEA test plate and normalized using an internal plate control as recommended by the manufacturer. In addition, plasma proteins with high performance significance (MMP7 and GDF15) were further validated by enzyme-linked immunosorbent assay (ELISA).

[0036] Marker discovery and model building: In the discovery cohort, the protein levels between the disease group and the control group were compared using the t-test or Wilcoxon rank sum test based on the results of the variance homogeneity test and the normality test. Plasma proteins with an adjusted p-value (FDR) less than 0.05 were defined as differentially expressed proteins. In order to clarify the association between each biomarker and high-grade colorectal lesions, a multivariate linear regression model was used to correct the control group and incorporate clinical factors (age, sex, BMI, and lipid levels). A total of 15 proteins were identified in the study.

[0037] Subsequently, four machine learning algorithms—LASSO regression, random forest, support vector machine (SVM), and XGBoost—were trained as binary classification models for the prediction of colorectal lesions. To evaluate the performance of the diagnostic models and enhance their robustness while avoiding overfitting, 5-fold cross validation and 0.632+ bootstrap were used. The area under the receiver operating characteristic (ROC) curve (AUC) and its 95% confidence interval (CI) were calculated, and the sensitivity, specificity, positive predictive value (PPV), and F1 score were calculated at the optimal threshold determined by the 0.632+ bootstrap method. In addition, the diagnostic efficacy of individual proteins was evaluated by plotting ROC curves. Finally, the key protein markers were combined with selected clinical predictors to construct clinical risk models and joint models. The models were evaluated in terms of their discrimination ability (ROC curve), calibration ability (calibration curve), and clinical utility (decision curve analysis, DCA).

[0038] All machine learning algorithms in this embodiment are constructed using the scikit-learn 1.3.2 software package in Python and implemented in Jupyter Notebook.

[0039] LASSO regression model: 5 markers were screened out: TGM2, MMP7, GDF15, RNASE3, REG1B

[0040] SVM model: 7 markers were screened: RNASE3, TGM2, MMP7, REG1A, REG1B, GDF15, MFAP5

[0041] Random forest model: 5 markers were screened: GDF15, REG1B, MMP7, TFF3, RNASE3

[0042] XGBoost model: 7 markers were screened out: GDF15, MMP7, TFF3, RNASE3, REG1B, MFAP5, and TGM2.

[0043] See also Figure 1 and Figure 2 ,, the test efficiency of the discovery cohort and the validation cohort was above 0.85. In the discovery cohort, based on diagnostic markers, a diagnostic model (AUC>0.90) was established using Lasso regression, random forest, SVM and XGboost. In the validation cohort, the validation model had good diagnostic efficiency for early lesions of colorectal cancer (advanced adenomas and early cancers) (AUC>0.85).

[0044] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A method for constructing a colorectal precancerous lesion diagnosis model based on plasma proteome, characterized in that the steps include: S1. Plasma was collected before endoscopy, stored at -80°C or in liquid nitrogen, and analyzed using Olink Explore 384 Cardiometabolic Panel based on PEA technology; S2. 1 mL of serum was mixed with 369 pairs of antibody probes, and markers were generated by DNA polymerase extension, quantified by NGS, and presented by NPX. At the same time, some proteins were verified by ELISA; S3. Determine differentially expressed proteins by testing and identify proteins after correcting clinical factors; S4. Train LASSO, random forest, SVM, and XGBoost algorithms to build models, evaluate using cross-validation and other methods, calculate the AUC index, and build models in combination with clinical factors.

2. The method for constructing a colorectal precancerous lesion diagnosis model based on plasma proteome according to claim 1, characterized in that: In step S2, the Olink Explore 384 Cardiometabolic Panel was used for proximity extension-based analysis. 1 mL of serum was mixed with 369 pairs of antibody probes linked to oligonucleotides. After the probes bound to the target antigens, they were extended by DNA polymerase to generate surrogate markers, which were then quantified by next-generation sequencing. The data were presented as normalized protein expression values, and sample randomization and internal plate control normalization were performed. Plasma proteins were further verified by enzyme-linked immunosorbent assay.

3. The method for constructing a colorectal precancerous lesion diagnosis model based on plasma proteome according to claim 1, characterized in that: In step S3, based on the results of the homogeneity of variance test and the normality test, the protein levels were compared using the t-test or the Wilcoxon rank sum test, and the plasma proteins with an adjusted p-value less than 0.05 were defined as differentially expressed proteins. The multivariate linear regression model was used to correct the control group and incorporate clinical factors, and then the proteins were identified.

4. The method for constructing a colorectal precancerous lesion diagnosis model based on plasma proteome according to claim 1, characterized in that: In step S4, LASSO regression, random forest, SVM, and XGBoost machine learning algorithms were trained as binary classification models to predict colorectal lesions, and the sensitivity, specificity, positive predictive value, and F1 score were calculated. The diagnostic efficacy of individual proteins was evaluated by plotting ROC curves, and finally, the key protein markers were combined with the selected clinical predictors to construct a diagnostic model.

5. The method for constructing a colorectal precancerous lesion diagnosis model based on plasma proteome according to claim 4, characterized in that: The diagnostic models were constructed using the scikit-learn 1.3.2 package in Python and implemented in Jupyter Notebook. The key protein markers were combined with selected clinical predictors to construct the models, specifically: LASSO regression model: 5 markers were screened: TGM2, MMP7, GDF15, RNASE3, and REG1B; SVM model: 7 markers were screened: RNASE3, TGM2, MMP7, REG1A, REG1B, GDF15, and MFAP5; Random forest model: 5 markers were screened: GDF15, REG1B, MMP7, TFF3, and RNASE3; XGBoost model: 7 markers were screened out: GDF15, MMP7, TFF3, RNASE3, REG1B, MFAP5, and TGM2.

Images