Discrimination method for esophageal cancer based on serum metabolism and polypeptide markers and application thereof

Abstract

The present application relates to the field of biotechnology, in particular to a serum metabolism and polypeptide marker-based esophageal cancer discrimination method and application thereof. The present application provides a non-invasive (minimally invasive), high-throughput and low-cost method, i.e. through MALDI TOF MS mass spectrometry technology, detection of metabolic and polypeptide markers in serum, effective screening of esophageal cancer and benign lesion patients from high-risk populations, and low, medium and high risk stratification management of the population, to replace or supplement the current endoscopic examination-based invasive screening method.

Description

Technical Field

[0001] This invention relates to the field of biotechnology, specifically to a method for esophageal cancer identification based on serum metabolism and polypeptide markers, and its application. Background Technology

[0002] Esophageal cancer is one of the most common malignant tumors worldwide. Due to the lack of specific clinical manifestations in its early stages, most patients are diagnosed at an advanced stage, resulting in limited treatment options and poor prognosis. Currently, the "gold standard" for esophageal cancer screening and diagnosis is upper gastrointestinal endoscopy combined with pathological biopsy. However, endoscopy has significant drawbacks, including invasiveness, reliance on professional personnel, high cost, and poor patient tolerance, making it unsuitable for mass screening in large asymptomatic populations. Liquid biopsy, as an emerging non-invasive detection method, shows promise in the early diagnosis of various tumors by detecting circulating tumor markers in peripheral blood. However, from a clinical perspective, the application of liquid biopsy in esophageal cancer screening still faces technical bottlenecks due to insufficient sensitivity.

[0003] Peptidomics is an omics technology that studies endogenous peptides in organisms. Unlike proteomics, which mainly focuses on intact proteins, the advantage of peptidomics lies in its ability to directly analyze naturally occurring active peptides produced by specific enzymatic cleavage within cells. These peptides often directly participate in signal transduction during disease processes, providing a more real-time reflection of physiological or pathological dynamics. Studying these peptides helps identify extremely low concentrations of specific signals in early lesions and reveals key enzyme activity characteristics in the tumor microenvironment. Metabolomics, on the other hand, involves comprehensive qualitative and quantitative analysis of small molecule metabolites (typically with a molecular weight less than 1500 Daltons and non-peptide substances) in biological samples. Metabolites are downstream of the central dogma, directly reflecting the current physiological or pathological state of an organism and representing a comprehensive "output" of changes in the genome, transcriptome, and proteome. Cancer causes significant "reprogramming" of cellular metabolism (such as the Wahlberg effect). Therefore, by comparing the metabolite profiles of cancer patients and healthy individuals, abnormal metabolic pathways related to cancer can be systematically identified, providing biomarkers for early diagnosis and personalized treatment. In esophageal cancer detection applications, peptidomics and metabolomics are complementary. Peptidomics directly reflects the functional processing of upstream proteins, while metabolomics reveals the final physiological effects following this series of events. Combining these two omics approaches can depict the molecular panorama of tumors from different dimensions, offering the possibility of finding more reliable combinations of esophageal cancer biomarkers.

[0004] Matrix-assisted laser desorption / ionization time-of-flight mass spectrometry (MALDI-TOF MS) is a powerful biomolecular analysis technique widely used in proteomics / peptideomics, metabolomics, microbial identification, and clinical diagnostics. Its principle involves using laser soft ionization to energize biomolecules, resulting in ionized ions that exhibit different flight times in the flight tube due to their varying mass-to-charge ratios (m / z), allowing for precise determination of the mass-to-charge ratio. MALDI-TOF MS, combined with silicon-based micro / nano particles and silicon-based chips, enables dual-omics detection on a single platform. This method offers significant advantages, including low sample volume, simple operation, fast analysis speed, high sensitivity, high throughput, and low cost. In contrast, LC-MS mass spectrometry requires sophisticated sample pretreatment and has low throughput, making it unsuitable for clinical applications.

[0005] Currently, widely used clinical diagnostic techniques for esophageal cancer include: 1. Esophagoscopy combined with pathological biopsy is the gold standard; 2. Multiple markers such as CEA, CA19-9, SCC, and CYFRA21-1 are detected together; 3. Based on MALDI-TOF MS combined with weak cation exchange magnetic beads (WCX-MB), serum proteins / peptides are purified, and differentially expressed peptides are screened by MALDI-TOF MS to construct esophageal cancer diagnostic models; 4. Serum metabolomics diagnostic models based on the LC-MS platform, where the markers have good sensitivity and specificity for esophageal cancer and can be used for esophageal cancer diagnosis, including early diagnosis; 5. Esophageal cancer diagnostic techniques based on ctDNA methylation.

[0006] The aforementioned esophageal cancer diagnostic techniques, which are widely used in clinical practice, have the following drawbacks: 1) Esophagoscopy combined with pathological biopsy. Endoscopic examination has significant disadvantages such as being invasive, requiring professional personnel to operate, being costly, and having poor patient tolerance, making it unsuitable for mass screening in large asymptomatic populations; 2) Combined testing of traditional esophageal cancer-related protein biomarkers such as CEA, CA19-9, SCC, and CYFRA21-1. Broad-spectrum biomarkers lack specificity and are not recommended for early screening or diagnosis of esophageal cancer. Their main value lies in postoperative monitoring, efficacy evaluation, and recurrence warning, rather than "early detection." 3) MALDI-TOF MS combined with weak cation exchange magnetic beads (WCX-MB) method; 4) Serum metabolomics detection based on LC-MS platform; 5) Esophageal cancer diagnostic technology based on ctDNA methylation; Disadvantages of methods 3), 4), and 5): 1. Both are single-agent biomarkers, lacking sufficient information dimension. Esophageal cancer is a complex disease driven by multiple molecular events, and relying solely on single-agent biomarkers (peptides or metabolites) is insufficient to provide enough information dimension to support highly sensitive early diagnosis. Peptides reflect protein / peptide information directly secreted by the tumor or in response to the tumor, while metabolites reflect overall physiological metabolic disorders caused by the tumor. The inherent inadequacy of single-agent biomarkers limits further improvements in detection sensitivity and specificity.

[0007] 2. Lack of risk stratification functionality hinders precise stratified management. Current technologies only provide a binary output of cancer or non-cancer, or are only used for prognostic prediction and treatment efficacy assessment (such as CCRT efficacy evaluation). However, the actual clinical needs of esophageal cancer screening go far beyond a simple "yes / no" diagnosis—healthcare professionals need to scientifically stratify high-risk groups to optimize the use of endoscopic resources, accurately allocating limited endoscopic screening resources to truly high-risk individuals and avoiding unnecessary invasive examinations for low-risk groups. Existing biomarker detection technologies generally lack a low, medium, and high risk stratification design, resulting in poor practicality in the clinical scenario of "stratified management of high-risk groups."

[0008] 3. Poor ability to differentiate between benign esophageal lesions. Most existing esophageal cancer serum testing methods use healthy individuals as negative controls for model construction and validation, lacking systematic inclusion studies of benign esophageal lesions (such as esophagitis, low-grade / high-grade intraepithelial neoplasia). Patients with benign esophageal lesions are the group most prone to false positives in esophageal cancer screening. If the testing method cannot effectively distinguish between esophageal cancer and benign lesions, it will lead to a large number of unnecessary endoscopic referrals and anxiety. Although some studies in existing technologies have included precancerous lesion analysis, they are usually only used as single-sample groups, without incorporating benign lesions as an independent denominator in model design and validation.

[0009] 4. LC-MS mass spectrometry detection methods have high requirements for sample pretreatment and low detection throughput, making them unsuitable for clinical needs; 5. ctDNA methylation markers can only be detected through tumor cell apoptosis and necrosis, and therefore do not have an advantage in the early detection of esophageal cancer.

[0010] Therefore, there is an urgent need in this field for a technical method based on a single MALDI-TOF MS platform to simultaneously detect serum metabolites and peptide markers, as well as an esophageal cancer discrimination model that can classify subjects into three risk levels: low, medium, and high. This model would be used for early screening of esophageal cancer and stratified management of high-risk populations, in order to overcome the technical deficiencies of existing technologies, such as insufficient information dimensions, lack of risk stratification, and poor ability to distinguish benign lesions. Summary of the Invention

[0011] In view of this, the technical problem to be solved by the present invention is to provide a combination of markers.

[0012] This invention provides a combination of biomarkers, including peptide groups and / or metabolite groups; The polypeptide group includes at least one of the following: a polypeptide with a mass-to-charge ratio of 1207.1678, a polypeptide with a mass-to-charge ratio of 1261.7113, a polypeptide with a mass-to-charge ratio of 1588.3147, a polypeptide with a mass-to-charge ratio of 1630.5943, a polypeptide with a mass-to-charge ratio of 1639.0169, a polypeptide with a mass-to-charge ratio of 1655.2586, a polypeptide with a mass-to-charge ratio of 1660.2189, a polypeptide with a mass-to-charge ratio of 1779.0734, and a polypeptide with a mass-to-charge ratio of 2010.6. The following peptides have mass-to-charge ratios: 646, 2105.0986, 2272.2004, 2363.8626, 2876.9920, 3057.8379, 3364.4412, 3474.8535, 3586.6758, or 4788.7358. The metabolite group includes at least one of the following: LPA 16:0, LPA 20:0, LPE 18:2, LPA 22:0, LPE 20:4, SM 33:1; O2, PE P-38:3, SHexCer 38:2; O3, SHexCer 40:0; O3, SHexCer 41:1; O3, SHexCer 42:1; O3, SHexCer 42:2; O4, FA 22:4, LPC 18:0, CE 16:1, PA 40:9, PG 36:4, Taurine, Thiodiacetic acid, p-Cresol sulfate, O-methoxycatechol-O-sulphate, Pyrogallol-1-O-sulphate, Heptadecanoic acid (FA 17:0), Hawkinsin, 5,8,11-Eicosatrienoic acid (FA 20:3), Docosahexaenoic acid (FA 22:6) or Adrenic acid (FA 22:4).

[0013] The above-mentioned biomarker combination was obtained by mass spectrometry identification. Experimental verification showed that, compared with tumor biomarkers in the prior art, the biomarker combination of the present invention can distinguish between esophageal cancer / precancerous lesions, inflammation, benign hyperplasia, and healthy individuals, significantly improving the detection performance of early esophageal cancer.

[0014] As a feasible example, the biomarker combination includes a peptide group and a metabolite group; The polypeptide group includes: a polypeptide with a mass-to-charge ratio of 1207.1678, a polypeptide with a mass-to-charge ratio of 1261.7113, a polypeptide with a mass-to-charge ratio of 1588.3147, a polypeptide with a mass-to-charge ratio of 1630.5943, a polypeptide with a mass-to-charge ratio of 1639.0169, a polypeptide with a mass-to-charge ratio of 1655.2586, a polypeptide with a mass-to-charge ratio of 1660.2189, a polypeptide with a mass-to-charge ratio of 1779.0734, and a polypeptide with a mass-to-charge ratio of 2010.6646. The peptides with mass-to-charge ratios of 2105.0986, 2272.2004, 2363.8626, 2876.9920, 3057.8379, 3364.4412, 3474.8535, 3586.6758, and 4788.7358 were identified. The metabolite group includes: LPA 16:0, LPA 20:0, LPE 18:2, LPA 22:0, LPE 20:4, SM 33:1; O2, PE P-38:3, SHexCer 38:2; O3, SHexCer 40:0; O3, SHexCer 41:1; O3, SHexCer 42:1; O3, SHexCer 42:2; O4, FA 22:4, LPC 18:0, CE 16:1, PA 40:9, PG 36:4, Taurine, Thiodiacetic acid, p-Cresol sulfate, O-methoxycatechol-O-sulphate, Pyrogallol-1-O-sulphate, Heptadecanoic acid (FA 17:0), Hawkinsin, 5,8,11-Eicosatrienoic acid (FA 20:3), Docosahexaenoic acid (FA 22:6) and Adrenic acid (FA 22:4).

[0015] In some embodiments, the polypeptide with a mass-to-charge ratio of 1207.1678 is a fragment of Caveolae-associated protein 2, the amino acid sequence of which is shown in SEQ ID NO: 1; The polypeptide with a mass-to-charge ratio of 1261.7113 is a fragment of the immunoglobulin heavy constant gamma 2 region, and its amino acid sequence is shown in SEQ ID NO: 2; The polypeptide with a mass-to-charge ratio of 1588.3147 is a fragment of the immunoglobulin λ variable region 3-10, and its amino acid sequence is shown in SEQ ID NO: 3; The polypeptide with a mass-to-charge ratio of 1630.5943 is a fragment of endoplasmic reticulum aminopeptidase 2, and its amino acid sequence is shown in SEQ ID NO: 4. The polypeptide with a mass-to-charge ratio of 1639.0169 is a fragment of apolipoprotein AI, and its amino acid sequence is shown in SEQ ID NO: 5. The polypeptide with a mass-to-charge ratio of 1655.2586 is a fragment of the immunoglobulin κ variable region 2-28, and its amino acid sequence is shown in SEQ ID NO: 6; The polypeptide with a mass-to-charge ratio of 1660.2189 is a fragment of apolipoprotein B-100, and its amino acid sequence is shown in SEQ ID NO: 7. The polypeptide with a mass-to-charge ratio of 1779.0734 is a fragment of the large ribosomal subunit protein uL22, and its amino acid sequence is shown in SEQ ID NO: 8. The polypeptide with a mass-to-charge ratio of 2010.6646 is a fragment of prothrombin, and its amino acid sequence is shown in SEQ ID NO: 9; The polypeptide with a mass-to-charge ratio of 2105.0986 is a fragment of phosphoglycerate kinase 1, and its amino acid sequence is shown in SEQ ID NO: 10. The polypeptide with a mass-to-charge ratio of 2272.2004 is a fragment of complement C1q subcomponent subunit A, and its amino acid sequence is shown in SEQ ID NO: 11. The polypeptide with a mass-to-charge ratio of 2363.8626 is a fragment of the fibrinogen beta chain, and its amino acid sequence is shown in SEQ ID NO: 12; The polypeptide with a mass-to-charge ratio of 2876.9920 is a fragment of α-2-macroglobulin, and its amino acid sequence is shown in SEQ ID NO: 13; The polypeptide with a mass-to-charge ratio of 3057.8379 is a fragment of fructose-bisphosphate aldolase A, and its amino acid sequence is shown in SEQ ID NO: 14. The polypeptide with a mass-to-charge ratio of 3364.4412 is a fragment of apolipoprotein AI, and its amino acid sequence is shown in SEQ ID NO: 15. The polypeptide with a mass-to-charge ratio of 3474.8535 is a fragment of heterogeneous nuclear ribonucleoprotein F, and its amino acid sequence is shown in SEQ ID NO: 16. The polypeptide with a mass-to-charge ratio of 3586.6758 is a fragment of the hemoglobin α subunit, and its amino acid sequence is shown in SEQ ID NO: 17. The polypeptide with a mass-to-charge ratio of 4788.7358 is a fragment of α-1-antitrypsin, and its amino acid sequence is shown in SEQ ID NO: 18.

[0016] The present invention also provides a method for screening combinations of markers as described above, comprising: Peptides and / or metabolites were extracted from the serum of the subjects as analytical samples; the subjects included healthy controls, patients with esophageal cancer, esophageal precancerous lesions, esophageal inflammatory / functional diseases, and esophageal proliferative diseases / benign tumors; Mass spectrometry was used to obtain peptide omics data and / or metabolomics data of the analyzed samples. The peptide omics data and / or metabolomics data are preprocessed and peak annotations are performed. After Boruta feature screening of significantly differentially expressed peptides and metabolites identified by differential analysis, biomarker combinations were determined using the Sequential Forward Floating Selection (SFFS) algorithm.

[0017] Experimental results demonstrate that, compared to commonly used single-feature selection methods in existing technologies, this invention employs a screening strategy of "differential analysis, Boruta global feature screening, and SFFS stepwise floating selection," which can accurately select the most discriminative biomarker combinations from hundreds or thousands of mass spectrometry peaks in a large sample cohort. This screening method exhibits stability in cross-validation with small AUC fluctuations, proving the robustness and reproducibility of the screening results.

[0018] In some embodiments, the conditions for the mass spectrometry detection include: The mass spectrometry platform is a MALDI-TOF MS mass spectrometry platform; The mass spectrometry detection mode for the metabolomics is a reflectance positive / negative ion mode, with a mass-to-charge ratio scanning range of 50–1000 Daltons and a laser energy setting range of 55%–63%. The mass spectrometry detection mode for the peptide omics is linear positive ion mode, the mass-to-charge ratio scanning range is 600~20000 Daltons, and the laser energy setting range is 20%~40%.

[0019] Through experimental screening, it was found that, compared with other detection conditions, the conditions used in this invention can efficiently acquire high-quality peptide and metabolite mass spectrometry data on the same MALDI-TOF MS platform, and the detection throughput is much higher than that of the LC-MS platform.

[0020] In some embodiments, the preprocessing includes baseline correction, normalization, peak identification, or peak alignment.

[0021] In some embodiments, the criteria for the difference analysis screening include: the difference fold meets a preset threshold and the false detection rate (FDR) is less than 0.05.

[0022] In some specific embodiments, the preset threshold is FC > 1.5 and P < 0.05 under multiple test correction.

[0023] This invention also provides the application of the aforementioned biomarker combinations in the preparation of esophageal cancer-related disease risk discrimination reagents, or in the construction of esophageal cancer risk stratification models.

[0024] Furthermore, this invention provides a method for constructing an esophageal cancer risk stratification model, including: Based on the peptidomics and metabolomics data of the analyzed samples, as well as the biomarker combinations described above, peptide models and metabolic models are constructed. Obtain the esophageal cancer prediction probability value of the polypeptide model and the esophageal cancer prediction probability value of the metabolic model corresponding to the analyzed sample; The esophageal cancer prediction probability values ​​of the polypeptide model and the metabolic model are fused together to obtain an esophageal cancer risk stratification model, which outputs an esophageal cancer risk score.

[0025] In some embodiments, the peptide and metabolic models are constructed using random forest (RF), LASSO linear regression, or XGBoost algorithms. Experimental verification shows that, compared to models built using clinical tumor markers, under the same modeling algorithm, the peptide and metabolic models constructed based on the biomarker combination of this invention exhibit significantly better predictive performance than models built using the aforementioned clinical tumor markers. Furthermore, both the peptide and metabolic models demonstrate good diagnostic efficacy, indicating that the biomarker combination obtained by this invention possesses excellent stability and reliability.

[0026] In some specific embodiments, the classification threshold for the LASSO linear regression peptide model is 0.3534, and the classification threshold for the metabolic model is 0.3028.

[0027] In some specific embodiments, the RF model calculates the Youden index based on the cross-validation ROC curve of the training set to determine the optimal classification threshold, which is 0.2365.

[0028] In some specific embodiments, the classification threshold for the XGBoost peptide model is 0.2252, and the classification threshold for the metabolism model is 0.3212.

[0029] Preferably, the peptide model and metabolic model are models constructed based on the Random Forest (RF) algorithm. By comparing the cross-validation results of different algorithm models, the Random Forest model performed best in terms of AUC and accuracy. Therefore, the RF peptide model and RF metabolic model were selected to construct the esophageal cancer risk stratification model.

[0030] In some embodiments, the esophageal cancer risk stratification model is a Stacking ensemble model, whose secondary learner uses a logistic regression algorithm to fuse the esophageal cancer prediction probability values ​​of the peptide model and the esophageal cancer prediction probability values ​​of the metabolic model as input features.

[0031] Specifically, the esophageal cancer prediction probability values ​​of the RF peptide model and the esophageal cancer prediction probability values ​​of the RF metabolism model are used as input features for fusion.

[0032] Furthermore, the present invention also provides an esophageal cancer risk stratification model, including: The data acquisition module is used to acquire peptide omics data and metabolomics data of the test subject samples, including omics data corresponding to the biomarker combinations as described above. The calculation module includes a peptide model and a metabolic model, which are used to obtain the esophageal cancer prediction probability value of the peptide model and the esophageal cancer prediction probability value of the metabolic model corresponding to the sample of the subject to be tested. The output module is used to fuse the esophageal cancer prediction probability value of the polypeptide model and the esophageal cancer prediction probability value of the metabolic model to obtain an esophageal cancer risk score and output the esophageal cancer discrimination result.

[0033] Experimental results show that, compared with single-mathematical models or clinical tumor marker models, the esophageal cancer risk stratification model (Stacking ensemble model) constructed in this invention has higher diagnostic performance and accuracy, significantly improving screening efficiency.

[0034] In some embodiments, the calculation formula for the esophageal cancer risk stratification model includes: Logit(P) = w0 + w1 × esophageal cancer prediction probability value of metabolic model + w2 × esophageal cancer prediction probability value of peptide model (Equation I). Where P represents the predicted risk probability of esophageal cancer by the model; w0 is -7.5 to -6.0, representing the intercept term of the model; w1 is 8.0 to 9.5, representing the weight coefficient corresponding to the predicted probability value of esophageal cancer by the metabolic model; and w2 is 7.0 to 8.5, representing the weight coefficient corresponding to the predicted probability value of esophageal cancer by the peptide model. Risk Score = 1 / [1+exp(-Logit(P))] (Equation II), where Risk Score represents the risk score for esophageal cancer.

[0035] For example, w0 is -7.5, -7.3, -7.0, -6.8, -6.5, -6.3, or -6.0; w1 is 8.0, 8.2, 8.5, 8.7, 8.9, 9.2, or 9.5; and w2 is 7.0, 7.2, 7.5, 7.8, 8.0, 8.2, or 8.5. Specifically, w0 is -6.774772, w1 is 8.701288, and w2 is 7.640766.

[0036] As a feasible example, the calculation formula for the esophageal cancer risk stratification model includes: Logit(P) = -6.774772 + 8.701288 × esophageal cancer prediction probability value of the metabolic model + 7.640766 × esophageal cancer prediction probability value of the peptide model (Equation I); Risk Score = 1 / [1+exp(-(-6.774772 + 8.701288×esophageal cancer prediction probability value of metabolic model + 7.640766×esophageal cancer prediction probability value of peptide model)] (Equation II).

[0037] In some embodiments, the boundary threshold of the model is determined based on the ROC curve, the maximum Yoden index principle, and a preset positive risk probability threshold. The boundary threshold includes the boundary threshold between low risk and medium risk, and the boundary threshold between medium risk and high risk. Among them, an esophageal cancer risk score below the low-risk and intermediate-risk threshold is output as low risk; an esophageal cancer risk score between the two thresholds is output as intermediate risk; and an esophageal cancer risk score above the intermediate-risk and high-risk threshold is output as high risk.

[0038] Optionally, 0.2365 can be set as the threshold between low risk and medium risk, which is determined by the ROC curve and the maximum Yoden index principle. A threshold of 0.6597 is set as the dividing line between medium and high risk, and this threshold is determined by the model risk score corresponding to a preset positive risk probability threshold; optionally, the preset positive risk probability threshold is 70%. An esophageal cancer risk score below 0.2365 indicates low risk, 0.2365-0.6597 indicates medium risk, and a score above 0.6597 indicates high risk.

[0039] The three molecular risk stratifications (low, intermediate, and high risk) established in this invention are intended to guide the prioritization and urgency of endoscopic examinations, but are not a substitute for routine clinical assessments or guideline-defined indications for endoscopy. Subjects classified as low-risk based on their molecular risk score do not require immediate endoscopic examination and can continue with routine clinical follow-up; however, if warning symptoms appear or other recognized clinical indications are present, endoscopic examination should still be performed. Intermediate-risk subjects are advised to complete diagnostic endoscopy as scheduled; high-risk subjects should be given priority for expedited endoscopic evaluation.

[0040] Furthermore, the present invention provides a risk assessment reagent for esophageal cancer-related diseases, including a detection reagent comprising the combination of biomarkers as described above.

[0041] In some embodiments, the reagents include reagents for detecting peptide markers in serum samples and / or reagents for detecting metabolite markers in serum samples.

[0042] In some embodiments, the detection reagent includes: (1) Reagents for enriching peptides and / or metabolites in serum samples; (2) Reagents used for mass spectrometry detection; (3) Standard products or quality control products.

[0043] In some embodiments, the reagent also includes an esophageal cancer risk stratification model or instructions for use, the instructions for use describing the steps of substituting data of biomarker combinations obtained by mass spectrometry detection into the esophageal cancer risk stratification model as described above, calculating an esophageal cancer risk score, and outputting a risk level based on a risk score threshold.

[0044] Furthermore, the present invention also provides a computer program product, including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement the biomarker combination screening method as described above and output the biomarker combination screening results, or implement the esophageal cancer risk stratification model construction method as described above and output an esophageal cancer risk score.

[0045] Furthermore, the present invention also provides an electronic device, including at least one processor and a memory connected to said processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program so that the electronic device can implement the biomarker combination screening method as described above and output the biomarker combination screening results, or implement the esophageal cancer risk stratification model construction method as described above and output the esophageal cancer risk score.

[0046] Furthermore, the present invention also provides a computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the biomarker combination screening method as described above and output the biomarker combination screening results, or to implement the esophageal cancer risk stratification model construction method as described above and output an esophageal cancer risk score.

[0047] Furthermore, the present invention also provides a method for identifying esophageal cancer-related diseases, including: Using the previously described combination of biomarkers, the previously described esophageal cancer risk stratification model, the previously described reagents, the previously described electronic devices, computer-readable storage media, or computer program products, the subjects to be tested are identified as having esophageal cancer-related diseases.

[0048] In some embodiments, the esophageal cancer-related diseases include patients with esophageal cancer, esophageal precancerous lesions, esophageal inflammatory / functional disorders, or esophageal proliferative disorders / benign tumors.

[0049] The discrimination method described in this invention can be used for diagnostic purposes or for non-diagnostic purposes.

[0050] When the aforementioned discrimination method is used for diagnostic purposes, it is targeted at subjects suspected of having esophageal cancer-related diseases. By detecting the mass spectrometry data of the combination of biomarkers as described above in the subject's serum sample, and combining it with the esophageal cancer risk stratification model to output a risk score, it assists clinicians in determining whether the subject has esophageal cancer-related diseases or assessing the subject's risk level of having esophageal cancer-related diseases, providing a reference for clinical diagnosis and treatment decisions.

[0051] When used for non-diagnostic purposes, the discrimination method can be used for large-scale population screening of esophageal cancer, risk stratification management of high-risk groups, basic medical research related to esophageal cancer, biomarker monitoring in the drug development process, or sample classification in the study of the mechanism of esophageal cancer development.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Richer detection information dimensions, resulting in higher diagnostic performance. By combining metabolite and peptide omics, complementary disease information can be obtained. Metabolites reflect global metabolic pathway alterations caused by tumor growth, while peptides reflect the expression and degradation products of tumor-specific proteins. This combination overcomes the limitations of single-omics in terms of information dimensions, significantly improving the detection performance of early esophageal cancer.

[0053] 2. The risk stratification output model is more adapted to clinical needs, improving screening efficiency and optimizing resource allocation.

[0054] Compared to the traditional binary output of "cancer / non-cancer," this system establishes three risk levels: low, medium, and high, providing clear stratified management for clinical practice: high-risk individuals are promptly recommended for endoscopic examination and pathological biopsy; medium-risk individuals are advised to have regular follow-up examinations or further confirmation using other auxiliary testing methods; and low-risk individuals can maintain routine health management. This design precisely meets the clinical needs for risk stratification and precise prevention and control among high-risk groups for esophageal cancer.

[0055] In a specific embodiment, in the real-world validation cohort, all 931 participants met the traditional high-risk population criteria. If screening were based solely on questionnaires and high-risk factors, all would, in principle, require further upper gastrointestinal endoscopy. Following a pre-defined model-guided triage strategy, only intermediate- and high-risk participants were referred for endoscopy, reducing the number of participants requiring endoscopy to 75, or 8.1% of the total population—a potential reduction of 91.9% in the burden of endoscopy. Simultaneously, 14 out of 16 cases of esophageal cancer or precancerous lesions were classified into the intermediate- and high-risk groups, achieving a detection sensitivity of 87.5%. The detection rate of esophageal cancer or precancerous lesions in the intermediate- and high-risk population reached 18.7%, significantly higher than the overall detection rate of 1.7% when endoscopy was performed directly without triage, achieving approximately a 10.9-fold enrichment of positive cases. The above results indicate that the esophageal cancer risk stratification model (serum dual-omics risk stratification model) can serve as a fine screening tool between traditional high-risk questionnaires and endoscopy. While maintaining a high ability to identify clinically significant esophageal lesions, it prioritizes the allocation of endoscopic resources to the population most likely to benefit, thereby potentially significantly reducing the examination burden on examinees and medical institutions and improving the efficiency of esophageal cancer screening.

[0056] 3. High throughput, low cost, and easy operation

[0057] MALDI-TOF MS offers high-throughput detection capabilities, requiring only a small amount of serum sample for testing. It is non-invasive or minimally invasive, making it suitable for large-scale testing of high-risk populations.

[0058] This invention relates to a method for esophageal cancer discrimination based on serum metabolism and peptide biomarkers and its application. This invention screens biomarker combinations for esophageal cancer discrimination and provides a technical method for simultaneously detecting serum metabolites and peptide biomarkers using a single MALDI-TOF MS platform. Through dual-omics information fusion and machine learning modeling, an esophageal cancer discrimination model is constructed that can classify subjects into low, medium, and high risk levels. This model is used for early screening of esophageal cancer and stratified management of high-risk populations, overcoming the technical deficiencies of existing technologies such as insufficient information dimensions, lack of risk stratification, and poor ability to distinguish benign lesions. Attached Figure Description

[0059] Figure 1 A method for screening marker combinations according to one embodiment of this application; Figure 2 Another embodiment of this application describes a method for constructing an esophageal cancer risk stratification model; Figure 3 A schematic diagram of the composition structure of an electronic device provided in an embodiment of this application; Figure 4 This is a risk stratification assessment diagram for the esophageal cancer risk stratification model (A is the distribution of esophageal cancer risk scores for each disease subgroup; B is the fitted curve of risk scores for each subgroup). Figure 5 ROC curves for the cross-validation and test sets of the peptide model, metabolic model, and tumor marker model are shown (A and D are the ROC curves for the cross-validation and test sets of the three peptide models (RF, LASSO, and XGBoost); B and E are the ROC curves for the cross-validation and test sets of the three metabolic models (RF, LASSO, and XGBoost); C and F are the ROC curves for the cross-validation and test sets of the three tumor marker models (RF, LASSO, and XGBoost). Figure 6 To cross-validate the performance comparison between the esophageal cancer risk stratification model and tumor markers, and to test the performance comparison results of the esophageal cancer risk stratification model and tumor markers (A and B cross-validate the performance comparison between the esophageal cancer risk stratification model and tumor markers; C and D test the performance comparison between the esophageal cancer risk stratification model and tumor markers). Figure 7 DCA curve for a real-world queue; Figure 8 This is a flowchart of the technical solution of this application. Detailed Implementation

[0060] This invention provides a method for esophageal cancer identification based on serum metabolism and polypeptide markers, and its application. Those skilled in the art can refer to the content of this document and appropriately modify the process parameters to achieve the desired result. It should be particularly noted that all similar substitutions and modifications are obvious to those skilled in the art and are considered to be included in this invention. The method and application of this invention have been described through preferred embodiments. Those skilled in the art can obviously make modifications or appropriate alterations and combinations to the method and application described herein without departing from the content, spirit, and scope of this invention to implement and apply the technology of this invention.

[0061] Unless otherwise defined in this invention, scientific and technical terms related to this invention shall have the meanings understood by one of ordinary skill in the art.

[0062] The terms “comprising,” “including,” and “having” are used interchangeably to indicate the inclusiveness of a scheme, meaning that the scheme may contain elements other than those listed. It should also be understood that the use of “comprising,” “including,” and “having” herein also provides for schemes “consisting of…”.

[0063] When used herein, the term “and / or” includes the meaning of “and,” “or,” and “all or any other combination of elements linked by the term.”

[0064] The term "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or multiple items.

[0065] It should be understood that in the various embodiments of this application, the order of the above processes does not imply the order of execution. Some or all steps may be executed in parallel or sequentially. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0066] The technical solution of the present invention: 1. Sample cohort collection, involving healthy individuals, benign lesions, precancerous lesions, and cancer; 2. Sample pretreatment: extraction of peptides and metabolites; 3. Mass spectrometry detection; 4. Data preprocessing; 5. Modeling and analysis, stacking of dual-omics models; analysis of multiple models, and comparative analysis with biomarkers; 6. Test set verification and risk stratification.

[0067] The specific technical solution is as follows ( Figure 8 ): (1) Serum samples were collected from the subjects and divided into 5 subgroups according to their diseases, including esophageal cancer, esophageal inflammatory / functional diseases, esophageal proliferative diseases / benign tumors, esophageal precancerous lesions, and healthy controls. The final study cohort was determined according to the pre-defined inclusion and exclusion criteria. Clinical epidemiological information related to patients was collected, including age, sex, pathological stage, type, tumor markers (CA199, CEA, AFP, etc.), smoking history, alcohol consumption history, dietary habits, and family history of esophageal cancer.

[0068] (2) The BiopSi silicon-based micro / nanoparticle kit and the Met-Si Array mass spectrometry chip kit were used to enrich peptides and metabolic molecules in serum samples, respectively. Peptide molecules: Serum peptide extraction was performed in Bio-pSi® (Hangzhou Huijian Technology Co., Ltd.) 96-well plates pre-embedded with silicon-based micro / nanoparticles. Serum sample and ultrapure water were added to each well of the 96-well plate, vortexed and incubated, centrifuged to remove the supernatant and washed once. Then, buffer was added to each well of the plate to release the peptides from the silicon-based micro / nanoparticle channels. The peptide sample was mixed with the matrix solution and then spotted onto the mass spectrometry target plate. Metabolic molecules: Serum metabolic extraction was performed using liquid-liquid extraction. All pipetting operations were completed using the ClinSpot® I automated pipetting workstation (Hangzhou Huijian Technology Co., Ltd.). Methanol was added to the 96-well PCR plate containing the sample and vortexed and incubated. Then, methyl tert-butyl ether (MTBE) was added and vortexed and incubated. Finally, ultrapure water was added to each well and centrifuged. Transfer the supernatant to a new 96-well PCR plate, vacuum dry it, resuspend it in lipid rehydration solution, and spot the sample onto a silicon-based micro / nano chip.

[0069] (3) Mass spectrometry detection.

[0070] Peptide omics detection: Samples were detected using a ClinMS-Plat® I MALDI-TOF mass spectrometer (Hangzhou Huijian Technology Co., Ltd.). The detection mode was linear positive ion mode, and the m / z range was set to 1000~20000.

[0071] The detection mode is linear positive ion mode, with an m / z range of 600~20000, a delay of 250ns, and a laser energy range of 20~40%. Each detection point accumulates 50... 16 = 800 laser bombardments; rigorous mass spectrometry mass axis calibration is performed before data acquisition.

[0072] Metabolomics detection: Mass spectrometry acquisition was performed using an autoflex maX MALDI TOF / TOF instrument (Bruker Daltonics). Measurement modes were negative and positive reflection ion modes. The relative laser pulse energy was optimized to 55%–63%, the relative laser energy was set to 61% of the total energy, the laser beam parameter was set to "4_large", and the laser spot diameter (spot size) was controlled at 80–100 μm. Instrument parameters were set as follows: pulsed ion extraction time 100–120 nanoseconds, primary ion source voltage 19.18–20.00 kV, secondary ion source voltage 16.92–17.75 kV; lens and reflector voltages were finely tuned to obtain optimal mass spectrometry resolution. The mass-to-charge ratio (m / z) scan range was 50–1000 Daltons, with 1250 laser bombardments accumulated at each detection point; rigorous mass axis calibration was performed before acquisition.

[0073] (4) Data preprocessing and biomarker screening: Baseline correction, normalization, peak identification and alignment were performed on the acquired serum peptide and metabolite spectrum data, and peptide mass spectrometry peaks and metabolite mass spectrometry peaks were annotated. Biomarker screening: Based on the training set, significant metabolites and peptides (fold difference, FDR < 0.05) obtained from differential analysis were screened using Boruta features, and then the Sequential Forward Floating Selection (SFFS) algorithm was used to determine the combination of feature biomarkers.

[0074] (5) Modeling, analysis, and validation: Modeling was performed based on the selected biomarkers. The training set was used as the model data, and 10-fold cross-validation was employed for model training and parameter optimization. After model training, validation was performed on both the training set (cross-validation prediction) and the test set. The Youden index was calculated based on the cross-validation ROC curve of the training set to determine the optimal classification threshold, which was then uniformly applied to the test set. Finally, the best-performing basic model among the various omics models was selected as the first-level learner, and its predicted probability was used as input. Logistic regression was used to construct a Stacking ensemble model, and its diagnostic performance was evaluated on the test set and prospective cohort.

[0075] (6) Risk value classification: low, medium and high risk thresholds are determined by cross-validation of the training set and the score distribution of esophageal cancer, esophageal inflammatory / functional diseases, esophageal proliferative diseases / benign tumors, esophageal precancerous lesions and healthy controls in the test set.

[0076] Advantages of this invention: 1. Richer detection information dimensions, resulting in higher diagnostic performance. By combining metabolite and peptide omics, complementary disease information can be obtained. Metabolites reflect global metabolic pathway alterations caused by tumor growth, while peptides reflect the expression and degradation products of tumor-specific proteins. This combination overcomes the limitations of single-omics in terms of information dimensions, significantly improving the detection performance of early esophageal cancer.

[0077] 2. The risk stratification output model is more adapted to clinical needs.

[0078] Compared to the traditional binary output of "cancer / non-cancer," this system establishes three risk levels: low, medium, and high, providing clear stratified management for clinical practice: high-risk individuals are promptly recommended for endoscopic examination and pathological biopsy; medium-risk individuals are advised to have regular follow-up examinations or further confirmation using other auxiliary testing methods; and low-risk individuals can maintain routine health management. This design precisely meets the clinical needs for risk stratification and precise prevention and control among high-risk groups for esophageal cancer.

[0079] 3. High throughput, low cost, and easy operation

[0080] MALDI-TOF MS offers high-throughput detection capabilities, requiring only a small amount of serum sample for testing. It is non-invasive or minimally invasive, making it suitable for large-scale testing of high-risk populations.

[0081] The amino acid sequence of the polypeptide group: SEQ ID NO:1 (fragment of Caveolae-associated protein 2): ALTSEEAERSD SEQ ID NO:2 (fragment of Immunoglobulin heavy constant gamma 2): GSFFLYSKLTV SEQ ID NO:3 (fragment of Immunoglobulin lambda variable 3-10): YELTQPPSVSVSPGQ SEQ ID NO:4 (Fragment of Endoplasmic reticulum aminopeptidase 2): SQFSVPSSYHFTED SEQ ID NO:5 (fragment of Apolipoprotein AI): HLSTLSEKAKPALED SEQ ID NO:6 (fragment of Immunoglobulin kappa variable 2-28): DIVMTQSPLSLPVTPG SEQ ID NO:7 (fragment of Apolipoprotein B-100): ASSSWGPQMSASVHLD SEQ ID NO:8 (fragment of Large ribosomal subunit protein uL22): GLDVDSLVIEHIQVNK SEQ ID NO:9 (fragment of Prothrombin): TATSEYQTFFNPRTFGSG SEQ ID NO:10 (Fragment of Phosphoglycerate kinase 1): QIVWNGPVGVFEWEAFAR SEQ ID NO:11 (fragment of Complement C1q subcomponent subunit A): HIYQGSEADSVFSGFLIFPSA SEQ ID NO:12 (fragment of Fibrinogen beta chain): GWTVIQNRQDGSVDFGRKWD SEQ ID NO:13 (fragment of Alpha-2-macroglobulin): PKGNRIAQWQSFQLEGGLKQFSFPL SEQ ID NO:14 (Fragment of Fructose-bisphosphate aldolase A): TVPPAVTGITFLSGGQSEEEASINLNAINK SEQ ID NO:15 (fragment of Apolipoprotein AI): DEPPQSPWDRVKDLATVYVDVLKDSGRDY SEQ ID NO:16 (fragment of Heterogeneous nuclear ribonucleoprotein F): MRPGAYSTGYGGYEEYSGLSDGYGFTTDLFGR SEQ ID NO:17 (fragment of Hemoglobin subunit alpha): VLSPADKTNVKAAWGKVGAHAGEYGAEALERMFL SEQ ID NO:18 (fragment of Alpha-1-antitrypsin): LEAIPMSIPPEVKFNKPFVFLMIEQNTKSPLFMGKVVNPTQK The numerical ranges and parameters involved in this invention have been presented as precisely as possible in the specific embodiments. However, any numerical value inevitably contains standard deviations due to individual test methods. Therefore, unless otherwise expressly stated, it should be understood that all numerical ranges or specific data used in this disclosure may have a reasonable deviation within a certain range, such as ±10%, ±5%, ±1%, or ±0.5%.

[0082] The test materials used in this invention are all common commercial products and can be purchased on the market.

[0083] The present invention will be further illustrated below with reference to the embodiments: like Figure 1 As shown, the screening method for marker combinations provided in this application may include: S20. Extract peptides and / or metabolites from the subject's serum as analytical samples.

[0084] Optionally, the subjects include healthy controls, patients with esophageal cancer, esophageal precancerous lesions, esophageal inflammatory / functional disorders, and esophageal proliferative diseases / benign tumors. Healthy controls are individuals confirmed by clinical endoscopy to be free of esophageal diseases, serving as a normal baseline reference. Esophageal cancer patients are individuals pathologically diagnosed with esophageal malignancies. Esophageal precancerous lesions are esophageal epithelial lesions with malignant transformation potential, including low-grade intraepithelial neoplasia and high-grade intraepithelial neoplasia. Esophageal inflammatory / functional disorders are diseases involving inflammation of the esophageal mucosa or abnormal esophageal function, including reflux esophagitis and eosinophilic esophagitis. Esophageal proliferative diseases / benign tumors are lesions involving benign proliferation of esophageal mucosal epithelial or stromal cells, including esophageal polyps and leiomyomas. All cohort samples in this invention have had their clinical results confirmed endoscopically or pathologically.

[0085] The serum samples used were collected in accordance with the hospital's strict requirements and standards, and have already eliminated the influence of factors such as diet and exercise, thus objectively reflecting the metabolic level of the individuals sampled. Among them, metabolites reflect global metabolic pathway changes caused by tumor growth, and peptides reflect the expression and degradation products of tumor-specific proteins.

[0086] S40. Obtain peptide omics data and / or metabolomics data of the analyzed sample by mass spectrometry detection.

[0087] Optionally, the mass spectrometry platform is a MALDI-TOF MS mass spectrometry platform; The mass spectrometry detection mode for the metabolomics is a reflectance positive / negative ion mode, with a mass-to-charge ratio scanning range of 50–1000 Daltons and a laser energy setting range of 55%–63%. The mass spectrometry detection mode for the peptide omics is linear positive ion mode, the mass-to-charge ratio scanning range is 600~20000 Daltons, and the laser energy setting range is 20%~40%.

[0088] Specifically, for metabolic spectrum detection: data acquisition was performed using a reflective positive / negative ion mode. The laser relative pulse energy was optimized to 55%–63%, and the laser spot diameter was controlled at 80–100 micrometers. Instrument parameters were set as follows: pulsed ion extraction time 100–120 nanoseconds, primary ion source voltage 19.18–20.00 kV, and secondary ion source voltage 16.92–17.75 kV; lens and reflector voltages were finely tuned to obtain optimal mass spectrometry resolution. The mass-to-charge ratio (m / z) scan range was 50–1000 Daltons, with 1250 laser bombardments accumulated at each detection point; rigorous mass axis calibration was performed before acquisition.

[0089] Specifically, for peptide profiling: the detection mode is linear positive ion mode, the m / z range is set to 600~20000, the delay is 250ns, the laser energy is set to 20~40%, and the cumulative energy at each detection point is 50. 16 = 800 laser bombardments; rigorous mass spectrometry mass axis calibration is performed before data acquisition.

[0090] S60. Preprocess and annotate the peptide omics data and / or metabolomics data.

[0091] The preprocessing includes baseline correction, normalization, peak identification, or peak alignment.

[0092] Baseline correction is the process of removing non-signal baseline drift in a mass spectrum caused by factors such as instrument background and chemical noise. In practice, the baseline typically appears as a slowly varying low-frequency signal, covering the entire mass spectrum and affecting the accurate calculation of peak intensity and area. Baseline correction estimates and subtracts the baseline signal using algorithms, allowing the mass spectral peaks to stand out against a flat background. Various algorithms can be used for baseline correction, such as nonlinear least squares fitting, wavelet transform, morphological filtering, and local minimum estimation. It can also be implemented using software packages such as MALDIquant, PROcess, and baseline in R, or the SciPy library in Python.

[0093] Normalization refers to the standardization of mass spectrometry data from different samples to eliminate systematic differences caused by variations in sample loading, instrument response fluctuations, and other factors. In MALDI-TOF MS detection, even samples from the same source may exhibit differences in total ion intensity when detected at different batches or different targets. Normalization scales the spectral data of each sample according to certain rules, making the data from different samples comparable. Commonly used normalization methods include total ion current normalization (dividing the intensity of each peak by the total ion intensity of the sample), median normalization (dividing the intensity of each peak by the median of the intensities of all peaks in the sample), quantile normalization, and internal standard normalization.

[0094] Peak identification refers to the process of detecting and locating true mass spectrometry peaks from baseline-corrected and normalized mass spectrometry data. Raw mass spectrometry data contains a large amount of signal, requiring the differentiation of true analyte signal peaks from background noise. Peak identification is achieved by using algorithms to detect local maxima of signal intensity and determine whether they exceed a set signal-to-noise ratio threshold. Various peak identification algorithms can be employed, such as peak detection based on continuous wavelet transform, methods based on local maximum search, and methods based on Gaussian mixture models. Commonly used software tools include MALDIquant, MassSpecWavelet, and XCMS.

[0095] Peak alignment refers to the process of matching and corresponding peaks with the same mass-to-charge ratio in different sample spectra. Due to instrument measurement errors and differences between samples, the mass-to-charge ratio of the same substance may drift slightly in different sample spectra. Peak alignment uses algorithms to cluster and match peaks in each sample spectrum according to their mass-to-charge ratio, ensuring the comparability of the same substance across different samples. Various peak alignment methods can be used, such as nearest neighbor-based matching, density-based clustering matching, and dynamic time warping methods.

[0096] Peak annotation refers to the process of identifying molecular identities by mapping detected mass spectral peaks to known peptides or metabolites. Peak annotation is achieved by matching experimentally obtained mass-to-charge ratios with theoretical values ​​in a database. Peptide annotation can be performed using peptide databases obtained through LC-MS mass spectrometry detection, which contain information such as the peptide's mass-to-charge ratio, retention time, and amino acid sequence. This database was established by detecting the primary and secondary mass spectrometry data of peptides from the modeling cohort samples mentioned in this embodiment of the invention on an LC-MS mass spectrometry platform. Metabolite annotation can be performed using metabolic databases, including HMDB (Human Metabolomics Database) and a local LC-MS metabolic database built by the inventive team. Through peak annotation, mass spectral peaks can be correlated with specific molecules, providing a molecular basis for subsequent biomarker identification.

[0097] S80. After screening the significantly differentially expressed peptides and metabolites identified by differential analysis using Boruta features, the biomarker combination is determined by the sequence floating selection algorithm.

[0098] Differential analysis involves statistically comparing the expression levels of various molecules under different disease states to identify molecules with significant differences in expression between different groups. Differential analysis compares the expression levels of various peptides / metabolites between disease subgroups and healthy controls, calculating the fold change (FC) and statistical significance. The fold change reflects the magnitude of change in molecular expression levels, while the false discovery rate (FDR) is used to control for false positives in multiple comparisons.

[0099] Optionally, a screening criterion can be set to meet preset thresholds: FC > 1.5 and FDR < 0.05. For example, differential analysis can be performed using methods such as t-tests, Mann-Whitney U tests, ANOVA, and Kruskal-Wallis tests, with multiple comparison correction using the Benjamini-Hochberg method to control for FDR. The purpose of differential analysis is to initially screen for potentially disease-related candidate molecules from a large number of detected molecules.

[0100] Boruta feature selection is a feature selection algorithm based on random forests. Its core objective is to identify all features that are truly relevant to the target variable. The Boruta algorithm creates randomly shuffled shadow features as references and iteratively compares the importance of the true features with those of the shadow features to determine whether each true feature has real predictive power. The advantages of Boruta feature selection lie in its robustness, ability to handle collinearity among features, and the fact that it does not require a pre-defined number of features. It is suitable for high-dimensional, small-sample data, especially in omics datasets where the number of features is much greater than the number of samples. The Boruta algorithm can be implemented using the `Boruta` package in R or the `boruta_py` library in Python.

[0101] Sequential Forward Floating Selection (SFFS) is a feature selection method whose core idea is to find the optimal feature combination by dynamically adjusting the selected feature subset. The basic process of SFFS is as follows: starting from an empty set, a forward search strategy is used to add an optimal feature at a time, maximizing the evaluation function value. After each addition, a backward floating step is performed, which checks if a certain feature exists in the current feature subset that, if removed, would further improve the evaluation function value; if so, it is removed. The addition and removal steps are repeated until a stopping condition is met. The evaluation function of SFFS typically uses performance metrics of classification models, such as AUC, accuracy, and F1 score. The SFFS algorithm can be implemented custom-based or using feature selection modules in machine learning libraries or tools such as mRMR.

[0102] The biomarker combinations obtained using the above screening methods include peptide groups and / or metabolite groups.

[0103] The polypeptide group includes at least one of the following: a polypeptide with a mass-to-charge ratio of 1207.1678, a polypeptide with a mass-to-charge ratio of 1261.7113, a polypeptide with a mass-to-charge ratio of 1588.3147, a polypeptide with a mass-to-charge ratio of 1630.5943, a polypeptide with a mass-to-charge ratio of 1639.0169, a polypeptide with a mass-to-charge ratio of 1655.2586, a polypeptide with a mass-to-charge ratio of 1660.2189, a polypeptide with a mass-to-charge ratio of 1779.0734, and a polypeptide with a mass-to-charge ratio of 2010.6. The peptides with a mass-to-charge ratio of 646, 2105.0986, 2272.2004, 2363.8626, 2876.9920, 3057.8379, 3364.4412, 3474.8535, 3586.6758, or 4788.7358 are listed.

[0104] The metabolite group includes at least one of the following: LPA 16:0, LPA 20:0, LPE 18:2, LPA 22:0, LPE 20:4, SM 33:1; O2, PE P-38:3, SHexCer 38:2; O3, SHexCer 40:0; O3, SHexCer 41:1; O3, SHexCer 42:1; O3, SHexCer 42:2; O4, FA 22:4, LPC 18:0, CE 16:1, PA 40:9, PG 36:4, Taurine, Thiodiacetic acid, p-Cresol sulfate, O-methoxycatechol-O-sulphate, Pyrogallol-1-O-sulphate, Heptadecanoic acid (FA 17:0), Hawkinsin, 5,8,11-Eicosatrienoic acid (FA 20:3), Docosahexaenoic acid (FA 22:6) or Adrenic acid (FA 22:4).

[0105] like Figure 2 As shown, the method for constructing the esophageal cancer risk stratification model provided in this application may include: S100. Based on the peptidomics and metabolomics data of the analyzed samples, and the combination of biomarkers as described above, construct peptide models and metabolic models.

[0106] The analyzed samples refer to serum polypeptide spectrum data and metabolic spectrum data that have been preprocessed and annotated with peaks according to the method described in Example 1, which include mass spectrometry peak intensity information corresponding to each biomarker.

[0107] A peptide model is a machine learning classification model built based on peptide omics data. Its input is the mass spectrometry peak intensity value or the standardized relative quantitative value of the peptide biomarker in the sample, and its output is the predicted probability value that the sample belongs to esophageal cancer. A metabolic model is a machine learning classification model built based on metabolomics data. Its input is the mass spectrometry peak intensity value or concentration value of the metabolite biomarker in the sample, and its output is the predicted probability value that the sample belongs to esophageal cancer.

[0108] Optionally, the peptide model and metabolic model are models constructed based on random forest, LASSO linear regression, or XGBoost algorithms.

[0109] S150. Obtain the esophageal cancer prediction probability value of the polypeptide model and the esophageal cancer prediction probability value of the metabolic model corresponding to the analysis sample.

[0110] The esophageal cancer prediction probability value is a number between 0 and 1, representing the model's confidence level in the analyzed sample being esophageal cancer. A higher value indicates a higher probability that the sample has esophageal cancer. The esophageal cancer prediction probability value of the peptide model reflects the discrimination result of the combination of peptide biomarkers for esophageal cancer, while the esophageal cancer prediction probability value of the metabolic model reflects the discrimination result of the combination of metabolite biomarkers for esophageal cancer.

[0111] Optionally, the peptide and metabolic models trained on the training set can be used to predict each analyzed sample separately, obtaining two independent predicted probability values. These two probability values ​​discriminate the sample from two different dimensions: peptide and metabolism, providing complementary information.

[0112] S200. The esophageal cancer prediction probability value of the polypeptide model and the esophageal cancer prediction probability value of the metabolic model are fused and calculated to obtain an esophageal cancer risk stratification model, and the esophageal cancer risk stratification model outputs an esophageal cancer risk score.

[0113] Fusion computing refers to integrating the prediction results of multiple individual models according to a certain strategy to obtain better prediction performance. Optionally, fusion computing is implemented through the stacking ensemble strategy. Stacking is a hierarchical model ensemble method. Its principle is to use the prediction outputs of multiple first-level learners (also known as base learners or basic models) as input features to train a second-level learner (also known as a meta-learner or ensemble model) for the final prediction.

[0114] In this invention, the esophageal cancer risk stratification model is a Stacking ensemble model. Its secondary learner uses a logistic regression algorithm, taking the esophageal cancer prediction probability values ​​of the peptide model and the esophageal cancer prediction probability values ​​of the metabolic model as input features. The weight coefficients of each input feature are determined through training, thereby constructing the Stacking ensemble model.

[0115] This application also provides an esophageal cancer risk stratification model, including: The data acquisition module is used to acquire peptide omics data and metabolomics data of the test subject samples, including omics data corresponding to the biomarker combinations as described above. The calculation module includes a peptide model and a metabolic model, which are used to obtain the esophageal cancer prediction probability value of the peptide model and the esophageal cancer prediction probability value of the metabolic model corresponding to the sample of the subject to be tested. The output module is used to fuse the esophageal cancer prediction probability value of the polypeptide model and the esophageal cancer prediction probability value of the metabolic model to obtain an esophageal cancer risk score and output the esophageal cancer discrimination result.

[0116] Optionally, the calculation formula for the esophageal cancer risk stratification model is as follows: Logit(P) = w0 + w1 × esophageal cancer prediction probability value of metabolic model + w2 × esophageal cancer prediction probability value of peptide model (Equation I). Where P represents the predicted risk probability of esophageal cancer by the model; w0 is -7.5 to -6.0, representing the intercept term of the model; w1 is 8.0 to 9.5, representing the weight coefficient corresponding to the predicted probability value of esophageal cancer by the metabolic model; and w2 is 7.0 to 8.5, representing the weight coefficient corresponding to the predicted probability value of esophageal cancer by the peptide model. Risk Score = 1 / [1+exp(-Logit(P))] (Equation II), where Risk Score represents the risk score for esophageal cancer.

[0117] Specifically, the calculation formula for the esophageal cancer risk stratification model includes: Logit(P) = -6.774772 + 8.701288 × esophageal cancer prediction probability value of the metabolic model + 7.640766 × esophageal cancer prediction probability value of the peptide model (Equation I); Risk Score = 1 / [1+exp(-(-6.774772 + 8.701288×esophageal cancer prediction probability value of metabolic model + 7.640766×esophageal cancer prediction probability value of peptide model)] (Equation II).

[0118] Optionally, the boundary threshold of the model can be determined based on the ROC curve, the maximum Yoden index principle, and a preset positive risk probability threshold. The boundary threshold includes the boundary threshold between low risk and medium risk, and the boundary threshold between medium risk and high risk.

[0119] The ROC curve (Receiver Operating Characteristic curve) is a curve plotted with the false positive rate on the x-axis and the true positive rate on the y-axis, used to evaluate the diagnostic performance of a binary classification model. The area under the curve (AUC) reflects the model's overall discriminative ability; a higher AUC value indicates better model performance. The Youden index is a comprehensive indicator of the model's diagnostic efficacy, calculated as sensitivity + specificity - 1. The threshold corresponding to the maximum Youden index is the optimal classification threshold. The principle of maximizing the Youden index means selecting the point on the ROC curve that maximizes the Youden index as the optimal classification threshold, at which the model's sensitivity and specificity achieve an optimal balance. The preset positive risk probability threshold refers to a pre-set positive discrimination probability value, set to 70% in this invention.

[0120] Specifically, 0.2365 is set as the dividing threshold between low and intermediate risk, and 0.6597 is set as the dividing threshold between intermediate and high risk. An esophageal cancer risk score below 0.2365 is output as low risk; an esophageal cancer risk score between 0.2365 and 0.6597 is output as intermediate risk; and an esophageal cancer risk score above 0.6597 is output as high risk.

[0121] This application also provides a computer program product, including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement the biomarker combination screening method as described above and output the biomarker combination screening results, or implement the esophageal cancer risk stratification model construction method as described above and output an esophageal cancer risk score.

[0122] This application embodiment also provides an electronic device, including at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the biomarker combination screening method as described above, and output the biomarker combination screening results, or to implement the esophageal cancer risk stratification model construction method as described above, and output an esophageal cancer risk score. (Reference) Figure 3 The diagram illustrates the structure of an electronic device suitable for implementing the method for constructing an esophageal cancer risk stratification model as described in the embodiments of this application, or for implementing the method for screening biomarker combinations provided in the embodiments of this application. The electronic devices in the embodiments of this application may include, but are not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0123] like Figure 3 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0124] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, memory cards, hard drives, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0125] This application embodiment also provides a computer storage medium carrying one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement the biomarker combination screening method as described above and output the biomarker combination screening results, or implement the esophageal cancer risk stratification model construction method as described above and output an esophageal cancer risk score.

[0126] Application Cases

[0127] 7.1 Biomarker Screening

[0128] 1) A total of 830 serum samples were enrolled, including 189 cases of esophageal cancer, 38 cases of precancerous esophageal lesions, 163 cases of inflammatory / functional esophageal diseases, 31 cases of esophageal proliferative diseases / benign tumors, and 409 healthy controls. Statistical analysis showed no significant differences in age, sex, and BMI among the cohorts. The cohorts were then randomly divided into training and test sets at a ratio of 7 / 3 according to their disease subgroups.

[0129] 2) The BiopSi silicon-based micro / nano particle kit and the Met-Si Array mass spectrometry chip kit were used to enrich peptides and metabolic molecules in serum samples, respectively. Extracts were then analyzed for peptide and metabolic profiles using a ClinMS-Plat® I MALDI-TOF mass spectrometer (Hangzhou Huijian Technology Co., Ltd.) and an autoflex maX MALDI TOF / TOF instrument (Bruker Daltonics).

[0130] 3) Baseline correction, normalization, peak identification, and alignment were performed on the acquired serum peptide and metabolic spectrum data. Peptide mass spectrum peaks and metabolic mass spectrum peaks were annotated using peptide and metabolic databases (obtained via LC-MS mass spectrometry). Biomarker screening: Based on the training set, significant metabolites and peptides (fold change > 1.5, FDR < 0.05) obtained from differential analysis were subjected to Boruta feature screening, and then the Sequential Forward Floating Selection (SFFS) algorithm was used to determine the combination of feature biomarkers. Specific peptides and metabolic biomarkers are shown in Tables 1 and 2 below: Table 1. Peptides

[0131] Table 2 Metabolites

[0132] 7.2 Model Building and Validation 1

[0133] Based on the biomarkers discovered in section 7.1, the Random Forest (RF) algorithm was used to model the model using the training set and validate it using the test set. Peptide (Pep) and metabolism (Met) models were established separately. A random forest classification model was built based on the training set, and the model parameters were optimized using 10-fold cross-validation. ROC-AUC was used as the evaluation metric, and the optimal combination of mtry, splittrule, and min.node.size was determined through parameter search. The model performance was then evaluated on the test set.

[0134] ① Peptide omics model parameters: The peptide model employs the Ranger algorithm to construct a random forest model. The forest size (number of trees) is set to the default value of 500 trees, and the maximum tree depth is set to the default value (max.depth = 0, meaning no manual limit is imposed on tree depth; the algorithm automatically grows the trees based on node splitting conditions). Node splitting parameters are optimized using 10-fold cross-validation, with the optimal parameter combination being: mtry = 9, splitle = extratrees, min.node.size = 1. The model's random seed is set to 888,888 to ensure repeatable model training.

[0135] ②Metabolic model parameters: The metabolomics model employs the Ranger algorithm to construct a random forest model. The forest size (number of trees) is set to the default value of 500 trees, and the maximum tree depth is set to the default value (max.depth = 0, meaning no manual limit is imposed on tree depth; the algorithm automatically grows the trees based on node splitting conditions). Node splitting parameters are optimized using 10-fold cross-validation, with the optimal parameter combination being: mtry = 3, split = extratrees, min.node.size = 1. The model's random seed is set to 888,888 to ensure repeatable model training.

[0136] The performance testing and verification results of the model built by the RF (Random Forest) algorithm are shown in Tables 3 and 4.

[0137] Table 3 Cross-validation set

[0138] Table 4 Test Set

[0139] 7.3 Model Building and Validation 2

[0140] Based on the biomarkers discovered in section 7.1, the Lasso (linear regression) algorithm was used to model the model using the training set and validate it using the test set. Peptide (Pep) and metabolism (Met) models were established separately. First, the features in the training set are centered and standardized. Then, the penalty parameter λ (lambda) is optimized on the training set based on 10-fold cross-validation, and the optimal parameter combination is selected using ROC-AUC as the model performance evaluation metric. LASSO uses L1 regularization to filter key features and builds the final model based on the optimal parameters.

[0141] ① Peptide omics model parameters: The final model was fitted using the optimal penalty parameters (λ = 0.00615, α = 1), resulting in 14 non-zero features, and a LASSO Logistic Regression model was constructed. The model intercept was -1.797743, and the final classification threshold was determined to be 0.3534 using the Youden exponent from the ROC curve of the training set. The model's predicted probability was calculated using the Logistic function. The formula for this model is as follows: Logit(P) = -1.797743 + 0.070853 ×Pep_1207.1677972373 - 0.239864 ×Pep_1261.71129942741 - 0.385892 × Pep_1630.59433947327 + 0.193592 × Pep_1639.01686830117 - 0.911961 × Pep_1779.07336118498 - 0.035909 × Pep_2010.66461734472 - 0.663244 × Pep_2272.20044892586 - 0.720545 × Pep_2363.86262483669 + 0.579528 × Pep_2876.99198994687 - 0.403037 × Pep_3057.8379163008 + 0.064375 ×Pep_3364.44124664455 + 0.482181 ×Pep_3474.85350592591 + 0.355377 The final output risk score: Risk Score = 1 / (1+exp(-(-1.797743 + 0.070853 × Pep_1207.1677972373 -0.239864 × Pep_1261.71129942741 – 0.385892 +0.385694 × Pep_1630.59433947327 + 0.193592 × Pep_1639.01686830117 -0.911961 × Pep_1779.07336118498 - 0.035909 × Pep_2010.66461734472 -0.663244 × Pep_2272.20044892586 - 0.720545 × Pep_2363.86262483669 +0.579528 × Pep_2876.99198994687 - 0.403037 × Pep_3057.8379163008 + 0.064375× Pep_3364.44124664455 + 0.482181 × Pep_3474.85350592591 + 0.355377 × Pep_4788.7357737536))) ②Metabolic model parameters: The final model was fitted using the optimal penalty parameters (λ = 0.00515, α = 1), resulting in 25 non-zero features, and a LASSO Logistic regression model was constructed. The model intercept was -1.696885, and the final classification threshold was determined to be 0.3028 using the Youden exponent from the ROC curve of the training set. The model's predicted probability was calculated using the Logistic function. The calculation formula for this model is as follows: Logit(P) = -1.696885 - 0.136485 × LPA 16:0+ 0.37206 × LPA 20:0 +0.228412 × LPE 18:2 + 0.331868 × LPA 22:0+ 0.078373 × LPE 20:4 - 0.540068× SM 33:1;O2 + 0.084174 × PE P-38:3 - 0.141604 × SHexCer 38:2;O3 -0.225074 × SHexCer 40:0;O3 - 0.241221 × SHexCer 41:1;O3- 0.043708 × SHexCer42:2;O4 - 0.520298 × FA 22:4 - 0.595782 × LPC 18:0 + 0.567543 × CE 16:1 -0.024947 × PA 40:9 + 0.551934 × Taurine - 0.135035 × Thiodiacetic acid +0.50745 × p-Cresol sulfate - 0.694942 × O-methoxycatechol-O-sulphate -0.180838 × Pyrogallol-1-O-sulphate - 0.381052 × Heptadecanoic acid (FA 17:0) + 0.64515 × Hawkinsin - 0.686768 × 5,8,11-Eicosatrienoic acid (FA 20:3)- 0.350542 × Docosahexaenoic acid (FA 22:6) + 0.256421 × Docosahexaenoicacid (FA 22:6) . Risk Score (Risk Score): Risk Score = 1 / (1+exp(-(-1.696885 - 0.136485 × LPA 16:0+ 0.37206 × LPA 20:0 + 0.228412 × LPE 18:2 + 0.331868 × LPA 22:0+ 0.078373 × LPE 20:4- 0.540068 × SM 33:1;O2 + 0.084174 × PE P-38:3 - 0.141604 × SHexCer 38:2;O3 - 0.225074 ×SHexCer 40:0;O3 - 0.241221 ×SHexCer 41:1;O3- 0.043708 ×SHexCer 42:2;O4 - 0.520298 × FA 22:4 - 0.595782 × LPC 18:0 + 0.567543 × CE16:1 - 0.024947 × PA 40:9 + 0.551934 × Taurine - 0.135035 × Thiodiacetic acid + 0.50745 × p-Cresol sulfate - 0.694942 × O-methoxycatechol-O-sulphate- 0.180838 × Pyrogallol-1-O-sulphate - 0.381052 × Heptadecanoic acid (FA17:0) + 0.64515 × Hawkinsin - 0.686768 × 5,8,11-Eicosatrienoic acid (FA 20:3) - 0.350542 × Docosahexaenoic acid (FA 22:6) + 0.256421 × Docosahexaenoic acid (FA 22:6))) The performance testing and validation results of the model built using the Lasso (linear regression) algorithm are shown in Tables 5 and 6.

[0142] Table 5 Cross-validation set

[0143] Table 6 Test Set

[0144] 7.4 Model Building and Validation 3

[0145] Based on the biomarkers discovered in section 7.1, the XGBoost algorithm was used to model the model using the training set and validate it using the test set. Peptide (Pep) and metabolism (Met) models were established separately: An Extreme Gradient Boosting (XGBoost) classification model was built based on the training set, and hyperparameter optimization was performed using 10-fold cross-validation combined with grid search, with ROC-AUC as the evaluation metric. Optimized parameters included nrounds, max_depth, eta, gamma, colsample_bytree, subsample, and min_child_weight, and the parameter combination with the highest average AUC was selected to build the final model. The model adopted a binary logistic regression objective function (objective = binary:logistic), and the evaluation metric was set to AUC (eval_metric = auc). The final model parameters were determined by cross-validation, with the number of boost rounds (nrounds) determined based on the optimal number of iterations. The model random seed was set to 888,888 to ensure the reproducibility of the model training process. The final model was evaluated on the test set, and the optimal classification threshold was determined by combining the predicted probabilities from the training set cross-validation with the Youden index.

[0146] ① Peptide omics model parameters: The final optimized parameters of the XGBoost model for peptidomics are: number of boost rounds (nrounds) = 50, maximum tree depth (max_depth) = 4, learning rate (eta) = 0.1, node split regularization parameter (gamma) = 0, feature sampling ratio (colsample_bytree) = 1, minimum leaf node weight (min_child_weight) = 1, sample sampling ratio (subsample) = 0.7, and the best number of iterations (best_nrounds) is determined to be 46 through 10-fold cross-validation.

[0147] ②Metabolic model parameters: The final optimized parameters of the metabolomics XGBoost model are: number of boost rounds (nrounds) = 100, maximum tree depth (max_depth) = 4, learning rate (eta) = 0.1, node split regularization parameter (gamma) = 0, feature sampling ratio (colsample_bytree) = 0.7, minimum leaf node weight (min_child_weight) = 3, sample sampling ratio (subsample) = 0.7, and the best number of iterations (best_nrounds) was determined to be 86 after 10-fold cross-validation.

[0148] The performance testing and verification results of the model built using the XGBoost algorithm are shown in Tables 7 and 8.

[0149] Table 7 Cross-validation set

[0150] Table 8 Test Set

[0151] 7.5 Establishment and Validation of a Two-Omics Esophageal Cancer Risk Stratification Model

[0152] The models built using different algorithms described in sections 7.2-7.4 all demonstrated good diagnostic efficacy, indicating that the discovered biomarker combinations have good stability. Comparing the AUC and accuracy results of the cross-validation sets for the three models, the RF model was deemed optimal. Therefore, the RF peptide and RF metabolism models were selected to construct the esophageal cancer risk stratification model. The RF model calculated the Youden index based on the ROC curves of the training set cross-validation to determine the optimal classification threshold, which was then uniformly applied to the test set at 0.2365. A Stacking ensemble model was constructed using Logistic regression, and its diagnostic performance was evaluated using the test set and tumor markers.

[0153] First, Random Forest (RF) classification models were built based on metabolomics and peptideomics data, respectively, to obtain the predicted probability of esophageal cancer for each sample through training. Then, the predicted probabilities output by the two independent models were used as input variables for a secondary model, and a Logistic Regression model was used as a meta-classifier to further build a multi-omics ensemble prediction model. The calculation formula for this model is as follows: Logit(P)= -6.774772 + 8.701288×Metabolite Model + 7.640766×PeptideModel; Where P represents the esophageal cancer risk probability predicted by the ensemble model, -6.774772 represents the model intercept term, and 8.701288 and 7.640766 represent the weight coefficients corresponding to the metabolomics prediction probability and the peptide probabilities, respectively; Metabolite Model represents the esophageal cancer prediction probability value of the metabolic model, and Peptide Model represents the esophageal cancer prediction probability value of the peptide model.

[0154] The final output risk score is obtained by transforming it using the sigmoid function, that is: Risk Score = 1 / [1+exp(-(-6.774772 + 8.701288×Metabolite Model +7.640766×Peptide Model))].

[0155] Based on the receiver operating characteristic (ROC) curve analysis results in the training set, the optimal classification threshold was determined using the maximum Youden index principle. When the risk score output by the model is higher than this threshold, the sample is classified as having a positive risk of esophageal cancer; otherwise, it is classified as having a non-esophageal cancer risk.

[0156] The primary classifiers employed Random Forest (RF) models, obtaining the out-of-fold (OOF) prediction probability for each training sample through 10-fold cross-validation. Subsequently, the OOF prediction probabilities generated by the metabolomics and peptidomics RF models were used as input variables for the secondary logistic regression model, constructing a Stacking ensemble model. The secondary logistic regression model further utilized 10-fold cross-validation to generate OOF prediction results, which were used for model performance evaluation and determination of the optimal classification threshold. Finally, the secondary logistic regression model was refitted based on the OOF prediction probabilities of the primary model across all training samples, and this final Stacking model was applied to the test set for independent prediction.

[0157] Note: Logistic regression coefficients reflect the linear influence of a variable in the meta-classifier, while SHAP values ​​reflect the actual contribution of a variable in the prediction process across all samples. Therefore, although metabolomics prediction scores have higher regression coefficients, peptideomics results may still show a higher SHAP contribution due to their greater predictive variability and stronger classification ability across samples.

[0158] The performance testing and validation results of the esophageal cancer risk stratification model are shown in Tables 9 and 10.

[0159] Table 9 Cross-validation set

[0160] Table 10 Performance comparison of the esophageal cancer risk stratification model and the tumor marker model in the test.

[0161] After establishing the esophageal cancer risk stratification model, the score distribution of healthy (HC), esophagitis (BED), benign esophageal hyperplasia (BHL), esophageal precancerous lesions (EPL), and esophageal cancer (EC) samples in the modeling cohort was analyzed (Figure A: scatter plot). Linear fitting was then used to fit these scores into two curves: EC (EC & EPL) and NC (HC, BED, BHL). Figure 4 A score of 0.2365 was set as the low- and medium-risk thresholds, while a score of 0.6597, representing a 70% probability of a positive result, was set as the medium- and high-risk thresholds. The results of each subgroup and stage of the test set are shown in Table 11.

[0162] Table 11 Results of each subgroup and stage of the test set

[0163] 7.6 Establishment and Application of Risk Stratification System

[0164] 1) Based on the esophageal cancer risk stratification model established in 7.5, the esophageal cancer risk stratification system was established by classifying the subgroups of esophageal health, esophageal inflammation, benign esophageal hyperplasia, esophageal precancerous lesions, and esophageal cancer into low-risk, intermediate-risk, and high-risk groups according to the model score distribution.

[0165] Table 12 Establishment of Esophageal Cancer Risk Stratification System

[0166] 2) This project established a brand-new risk stratification system for esophageal cancer. In the real-world research phase, the project compared its performance with that of mainstream clinical methods: esophagoscopy or gastroscopy.

[0167] First, a questionnaire survey was conducted among the intended participants to screen for high-risk esophageal cancer cases, totaling 931 individuals. These participants underwent both gastroscopy and the esophageal cancer risk stratification assessment (serum peptide and metabolomics testing) as part of this project. A double-blind method was used to compare the results of the two methods, using the results of gastroscopy or resection histopathology as the gold standard to evaluate the advantages of this project's esophageal cancer risk stratification.

[0168] Gastroscopy results: Of 931 samples, 11 cases of esophageal cancer, 5 cases of precancerous lesions, 39 cases of benign esophageal hyperplasia, 178 cases of esophagitis, and 698 healthy individuals were detected. The positive rate was only 5.91% (55 / 931). Therefore, based on questionnaires and gastroscopy, the positive detection rate of esophageal cancer is low, which will significantly increase the pressure on clinical resources. Due to my country's large population, the questionnaire and gastroscopy screening method resulted in nearly 95% or more of high-risk individuals undergoing unnecessary endoscopic examinations, leading to a waste of endoscopic screening resources. Furthermore, endoscopy is an invasive procedure with low compliance, high economic costs, and high technical requirements for operators. Even if comprehensive endoscopic screening is only promoted in high-incidence areas, these difficulties still exist.

[0169] The results of the esophageal cancer risk stratification assessment (Table 13) show that the NPV of low-risk, intermediate-risk, and high-risk individuals identified by the esophageal cancer risk stratification model was 99.77%, the PPV (intermediate-risk) was 6.12%, and the PPV (high-risk) was 42.31%, which fully demonstrates the effectiveness of the esophageal cancer risk stratification system in the real world.

[0170] Table 13 Results of Esophageal Cancer Risk Stratification Assessment

[0171] Based on the esophageal cancer risk stratification assessment established in this project, the positive rate reached 10.20% in the intermediate-risk population and as high as 50% in the high-risk population, significantly improving screening efficiency. Therefore, this project established a new, objective, convenient, and universally applicable risk stratification method to further accurately identify high-risk populations, thereby optimizing existing endoscopic screening strategies and improving their effectiveness and adoption.

[0172] Comparison of Models in Example 2

[0173] A) A comparison was made between single peptide and metabolic models and models jointly established using clinical tumor markers CEA, CA199, and AFP (Tables 14 and 15). Regardless of the modeling algorithm, the performance of tumor marker models was weaker than that of peptide or metabolic models.

[0174] Table 14 Performance comparison of single peptide, metabolic model and clinical tumor marker model in cross-validation dataset.

[0175] Table 15 Performance comparison of single peptides, metabolic models, and clinical tumor marker models in the test set.

[0176] like Figure 5As shown, Figures A and D are the ROC curves of the cross-validation set and test set for the three peptide models (RF, LASSO, and XGBoost), respectively; Figures B and E are the ROC curves of the cross-validation set and test set for the three metabolic models (RF, LASSO, and XGBoost), respectively; and Figures C and F are the ROC curves of the cross-validation set and test set for the three tumor marker models (RF, LASSO, and XGBoost), respectively. B) The esophageal cancer risk stratification model was compared with the model jointly established by clinical tumor markers CEA, CA199, and AFP (Table 16).

[0177] Table 16 Performance Comparison of Esophageal Cancer Risk Stratification Model and Clinical Tumor Marker Model

[0178] like Figure 6 As shown in Figures A and B, the performance of the esophageal cancer risk stratification model and tumor markers is compared using the cross-validation set; Figures C and D, the performance of the esophageal cancer risk stratification model and tumor markers is compared using the test set.

[0179] C) Performance comparison in real-world studies: Table 17 shows a performance comparison between the esophageal cancer risk stratification model and the clinical tumor marker model in real-world studies.

[0180] Table 17 Performance Comparison of Esophageal Cancer Risk Stratification Models and Clinical Tumor Marker Models in Real-World Studies

[0181] The project also compared its performance with tumor markers, plotting DCA curves to demonstrate its performance advantages, such as... Figure 7 As shown.

[0182] In summary, the technical contribution of this invention is the first-ever identification of a combination of combined biomarkers consisting of specific serum endogenous polypeptide sequences and specific metabolic ion characteristics in real-world screenings encompassing esophageal inflammatory / functional diseases, benign proliferative diseases, precancerous lesions, esophageal cancer, and healthy individuals. The polypeptide biomarkers are defined by their exact amino acid sequences rather than their corresponding complete precursor proteins, and the metabolic biomarkers are defined by specific molecular species and mass-to-charge ratios.

[0183] This invention further establishes peptide and metabolic basis models for the different distribution characteristics of two types of mass spectrometry data, and performs two-stage fusion of their cross-validation prediction probabilities to form low, medium, and high risk outputs through pre-determined thresholds. This technical solution achieved an AUC of 0.955 in an independent test set and 0.971 in a high-risk cohort of 931 patients from five centers with gastroscopy / pathology controls; the positive rate reached 10.20% in the medium-risk population and as high as 50% in the high-risk population. This effect cannot be directly predicted by existing single metabolic markers, single peptide markers, or general stacking algorithms, demonstrating that the specific molecular combinations and their fusion relationships have functional mutual support and synergistic technical effects.

[0184] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0185] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0186] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), memory, read-only memory (ROM), electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, removable disks, CD-ROMs (Compact Disc-Read-Only Memory), or any other form of storage medium known in the art.

[0187] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0188] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A combination of markers, characterized in that, Including the peptide group and / or metabolite group; The polypeptide group includes at least one of the following: a polypeptide with a mass-to-charge ratio of 1207.1678, a polypeptide with a mass-to-charge ratio of 1261.7113, a polypeptide with a mass-to-charge ratio of 1588.3147, a polypeptide with a mass-to-charge ratio of 1630.5943, a polypeptide with a mass-to-charge ratio of 1639.0169, a polypeptide with a mass-to-charge ratio of 1655.2586, a polypeptide with a mass-to-charge ratio of 1660.2189, a polypeptide with a mass-to-charge ratio of 1779.0734, and a polypeptide with a mass-to-charge ratio of 2010.

6. The following peptides have mass-to-charge ratios: 646, 2105.0986, 2272.2004, 2363.8626, 2876.9920, 3057.8379, 3364.4412, 3474.8535, 3586.6758, or 4788.7358. The metabolite group includes at least one of the following: LPA 16:0, LPA 20:0, LPE 18:2, LPA 22:0, LPE 20:4, SM 33:1; O2, PE P-38:3, SHexCer 38:2; O3, SHexCer 40:0; O3, SHexCer 41:1; O3, SHexCer 42:1; O3, SHexCer 42:2; O4, FA 22:4, LPC 18:0, CE 16:1, PA 40:9, PG 36:4, Taurine, Thiodiacetic acid, p-Cresol sulfate, O-methoxycatechol-O-sulphate, Pyrogallol-1-O-sulphate, Heptadecanoic acid (FA 17:0), Hawkinsin, 5,8,11-Eicosatrienoic acid (FA 20:3), Docosahexaenoic acid (FA 22:6) or Adrenic acid (FA 22:4).

2. The marker combination according to claim 1, characterized in that, The polypeptide with a mass-to-charge ratio of 1207.1678 is a fragment of small pit-associated protein 2, and its amino acid sequence is shown in SEQ ID NO: 1; The polypeptide with a mass-to-charge ratio of 1261.7113 is a fragment of the γ2 constant region of the immunoglobulin heavy chain, and its amino acid sequence is shown in SEQ ID NO: 2; The polypeptide with a mass-to-charge ratio of 1588.3147 is a fragment of the variable region 3-10 of immunoglobulin λ, and its amino acid sequence is shown in SEQ ID NO: 3; The polypeptide with a mass-to-charge ratio of 1630.5943 is a fragment of endoplasmic reticulum aminopeptidase 2, and its amino acid sequence is shown in SEQ ID NO: 4; The polypeptide with a mass-to-charge ratio of 1639.0169 is a fragment of apolipoprotein AI, and its amino acid sequence is shown in SEQ ID NO: 5; The polypeptide with a mass-to-charge ratio of 1655.2586 is a fragment of immunoglobulin κ variable region 2-28, and its amino acid sequence is shown in SEQ ID NO: 6; The polypeptide with a mass-to-charge ratio of 1660.2189 is a fragment of apolipoprotein B-100, and its amino acid sequence is shown in SEQ ID NO:

7. The polypeptide with a mass-to-charge ratio of 1779.0734 is a fragment of the large ribosomal subunit protein uL22, and its amino acid sequence is shown in SEQ ID NO: 8; The polypeptide with a mass-to-charge ratio of 2010.6646 is a fragment of prothrombin, and its amino acid sequence is shown in SEQ ID NO: 9; The polypeptide with a mass-to-charge ratio of 2105.0986 is a fragment of phosphoglycerate kinase 1, and its amino acid sequence is shown in SEQ ID NO:

10. The polypeptide with a mass-to-charge ratio of 2272.2004 is a fragment of the complement C1q subunit A chain, and its amino acid sequence is shown in SEQ ID NO: 11; The polypeptide with a mass-to-charge ratio of 2363.8626 is a fragment of the β chain of fibrinogen, and its amino acid sequence is shown in SEQ ID NO: 12; The polypeptide with a mass-to-charge ratio of 2876.9920 is a fragment of α-2-macroglobulin, and its amino acid sequence is shown in SEQ ID NO: 13; The polypeptide with a mass-to-charge ratio of 3057.8379 is a fragment of fructose-1,5-bisphosphate aldolase A, and its amino acid sequence is shown in SEQ ID NO:

14. The polypeptide with a mass-to-charge ratio of 3364.4412 is a fragment of apolipoprotein AI, and its amino acid sequence is shown in SEQ ID NO:

15. The polypeptide with a mass-to-charge ratio of 3474.8535 is a fragment of heterologous ribonucleoprotein F, and its amino acid sequence is shown in SEQ ID NO:

16. The polypeptide with a mass-to-charge ratio of 3586.6758 is a fragment of the hemoglobin α subunit, and its amino acid sequence is shown in SEQ ID NO:

17. The polypeptide with a mass-to-charge ratio of 4788.7358 is a fragment of α-1-antitrypsin, and its amino acid sequence is shown in SEQ ID NO:

18.

3. The screening method for marker combinations according to claim 1 or 2, characterized in that, include: Peptides and / or metabolites were extracted from the serum of the subjects as analytical samples; the subjects included healthy controls, patients with esophageal cancer, esophageal precancerous lesions, esophageal inflammatory / functional diseases, and esophageal proliferative diseases / benign tumors; Mass spectrometry was used to obtain peptide omics data and / or metabolomics data of the analyzed samples. The peptide omics data and / or metabolomics data are preprocessed and peak annotation is performed; the preprocessing includes baseline correction, normalization, peak identification or peak alignment. After Boruta feature screening of significantly differentially expressed peptides and metabolites identified by differential analysis, biomarker combinations were determined using a sequence floating selection algorithm.

4. The screening method according to claim 3, characterized in that, The conditions for the mass spectrometry detection include: The mass spectrometry platform is a MALDI-TOF MS mass spectrometry platform; The mass spectrometry detection mode for the metabolomics is a reflectance positive / negative ion mode, with a mass-to-charge ratio scanning range of 50–1000 Daltons and a laser energy setting range of 55%–63%. The mass spectrometry detection mode for the peptide omics is linear positive ion mode, the mass-to-charge ratio scanning range is 600~20000 Daltons, and the laser energy setting range is 20%~40%. The criteria for differential analysis screening include: the difference fold meets a preset threshold and the false detection rate is less than 0.

05.

5. The use of the biomarker combination described in claim 1 or 2 in the preparation of a risk assessment reagent for esophageal cancer-related diseases, or in the construction of an esophageal cancer risk stratification model.

6. A method for constructing an esophageal cancer risk stratification model, characterized in that, include: Based on the peptidomics and metabolomics data of the analyzed samples, and the combination of biomarkers as described in claim 1 or 2, a peptide model and a metabolic model are constructed; the peptide model and metabolic model are models constructed based on random forest, LASSO linear regression or XGBoost algorithm; Obtain the esophageal cancer prediction probability value of the polypeptide model and the esophageal cancer prediction probability value of the metabolic model corresponding to the analyzed sample; The esophageal cancer prediction probability values ​​of the polypeptide model and the metabolic model are fused together to obtain an esophageal cancer risk stratification model, which outputs an esophageal cancer risk score.

7. The method according to claim 6, characterized in that, The esophageal cancer risk stratification model is a Stacking ensemble model. Its secondary learner uses a logistic regression algorithm to fuse the esophageal cancer prediction probability values ​​of the peptide model and the metabolic model as input features.

8. An esophageal cancer risk stratification model, characterized in that, include: The data acquisition module is used to acquire peptide omics data and metabolomics data of the subject sample to be tested, including omics data corresponding to the biomarker combination as described in claim 1 or 2. The calculation module includes a peptide model and a metabolic model, which are used to obtain the esophageal cancer prediction probability value of the peptide model and the esophageal cancer prediction probability value of the metabolic model corresponding to the sample of the subject to be tested. The output module is used to fuse the esophageal cancer prediction probability values ​​of the polypeptide model and the esophageal cancer prediction probability values ​​of the metabolic model to obtain an esophageal cancer risk score and output the esophageal cancer discrimination result.

9. The model according to claim 8, characterized in that, The calculation formula for the esophageal cancer risk stratification model includes: Logit(P) = w0 + w1 × esophageal cancer prediction probability value of metabolic model + w2 × esophageal cancer prediction probability value of peptide model (Equation I). Where P represents the predicted risk probability of esophageal cancer by the model; w0 is -7.5 to -6.0, representing the intercept term of the model; w1 is 8.0 to 9.5, representing the weight coefficient corresponding to the predicted probability value of esophageal cancer by the metabolic model; and w2 is 7.0 to 8.5, representing the weight coefficient corresponding to the predicted probability value of esophageal cancer by the peptide model. Risk Score = 1 / [1+exp(-Logit(P))] (Equation II), where Risk Score represents the risk score for esophageal cancer.

10. The model according to claim 9, characterized in that, The boundary thresholds of the model are determined based on the ROC curve, the maximum Yoden index principle, and the preset positive risk probability threshold. The boundary thresholds include the boundary threshold between low risk and medium risk, and the boundary threshold between medium risk and high risk. Among them, an esophageal cancer risk score below the low-risk and intermediate-risk threshold is output as low risk; an esophageal cancer risk score between the two thresholds is output as intermediate risk; and an esophageal cancer risk score above the intermediate-risk and high-risk threshold is output as high risk.

Images