Construction method of colorectal cancer intelligent prediction model based on mass spectrum serum proteomics

Abstract

The invention discloses a method for constructing an intelligent colorectal cancer prediction model based on mass spectrum serum proteomics, which comprises the following steps of: screening out candidate micropeptides with diagnostic potential by combining high-precision mass spectrum quantification with AI function prediction, and further constructing the model by adopting an ensemble learning algorithm. The method is rigorous in process, and by introducing an advanced normalization algorithm, integrating a feature selection strategy and targeted sample imbalance processing, the biomarker screening robustness and the prediction model accuracy are remarkably improved. Meanwhile, through model interpretive analysis, theoretical support is provided for clinical application of the marker.

Description

(I) Technical Field

[0001] This invention relates to a method for constructing an intelligent prediction model for colorectal cancer based on mass spectrometry serum proteomics. (II) Background Technology

[0002] China is currently in a critical period of rapid economic development and social transformation. With the acceleration of industrialization and urbanization, residents' lifestyles and dietary structures have undergone significant changes, with high-salt, high-fat diets and irregular work-rest schedules becoming increasingly common. At the same time, the combined effects of work pressure and environmental factors have led to a continuous rise in the risk of chronic diseases. Clinical data shows that the five-year survival rate for patients with early-stage gastric cancer (stage I) can reach over 90%, while for those with advanced-stage (stage IV) cancer, it plummets to less than 30%. Therefore, establishing an efficient, accurate, and scalable early gastric cancer screening system has become a critical issue that urgently needs to be addressed in China's public health field.

[0004] Currently, the diagnosis of early colorectal cancer faces two major technical bottlenecks: firstly, early lesions are mostly confined to the mucosal layer, making them difficult to identify with routine imaging examinations; secondly, existing screening methods each have their limitations: while endoscopy, the gold standard, is highly accurate, its coverage in population screening is only about 20% due to its invasiveness, high equipment requirements, and high cost; and while serum pepsinogen (PG) testing is simple to perform, its accuracy is limited. This contradictory situation—"highly accurate technologies are difficult to popularize, while easily popularized technologies have low accuracy"—severely restricts the promotion of early gastric cancer screening.

[0005] In recent years, liquid biopsy technology, characterized by its minimally invasive and convenient nature, has provided a new approach to overcoming this predicament. However, current mainstream methods for detecting circulating tumor DNA (ctDNA) methylation still face many challenges: extremely low ctDNA levels in peripheral blood, easy sample degradation, and significant background noise interference lead to persistently high false negative (approximately 30%) and false positive (approximately 15%) rates, limiting their clinical application value.

[0006] Against this backdrop, micropeptides have demonstrated unique advantages as a novel biomarker. These functional small-molecule polypeptides encoded by non-coding RNA have the following characteristics: (1) small molecular weight and stable structure, making them more stable in blood, less susceptible to contamination, easier to detect, and more resistant to degradation; (2) high tissue specificity, with their expression profile closely related to the occurrence and development of colorectal cancer; and (3) higher detection sensitivity in peripheral blood compared to traditional protein biomarkers. Therefore, developing a high-performance detection method based on micropeptides is expected to break through the current technical bottleneck in early gastric cancer screening and provide key technical support for establishing a hierarchical screening system suitable for China's national conditions. (III) Summary of the Invention

[0007] The purpose of this invention is to provide a method for constructing an intelligent predictive model for colorectal cancer based on mass spectrometry-based serum proteomics. This method combines high-precision mass spectrometry quantification with AI-based predictive capabilities to screen for candidate micropeptides with diagnostic potential, and further employs an ensemble learning algorithm to construct the model. The method utilizes an optimized sample pretreatment process, including ultrafiltration of serum, high-resolution classical protein immunoprecipitation for reverse enrichment, and trypsin digestion, significantly increasing the proportion of micropeptides in the sample, thereby ensuring the stability and repeatability of target micropeptide detection. The established model is characterized by its ease of operation, non-invasiveness, and high specificity, meeting the clinical application standards for early colorectal cancer screening. It effectively overcomes the problems of poor compliance with traditional endoscopy and high false-positive rates in existing liquid biopsy techniques, providing an innovative solution for early colorectal cancer screening. It is expected to break through the current technical bottlenecks in early gastric cancer screening and provide key technical support for the improvement of my country's hierarchical medical system.

[0008] The technical solution adopted in this invention is:

[0009] This invention provides a method for constructing an intelligent prediction model for colorectal cancer based on mass spectrometry serum proteomics, the method comprising the following steps:

[0010] (1) Acquisition of raw micropeptide omics data: Low-abundance micropeptides were extracted from each blood sample by ultrafiltration and reverse enrichment. Disulfide bonds in the micropeptides were removed by traditional mass spectrometry pretreatment and thiol groups were methylated and blocked. Then, a mixture of peptides derived from blood micropeptides was obtained by enzymatic hydrolysis. The raw mass spectrometry data of the peptide mixture were obtained by high-resolution liquid chromatography-mass spectrometry (LC-MS / MS).

[0011] (2) Data preprocessing and quality control: The raw mass spectrometry data were preprocessed to eliminate differences between technical repetitions and to remove the reverse sequence library (decoy database) matching results used to assess the false positive rate in the mass spectrometry database search. Then, the delayed normalization algorithm was used to correct the peptide precursor ion intensity, which can effectively remove systematic bias between samples. Subsequently, the precursor ion intensity information at the peptide level was summarized to the micropeptide level. A dual screening strategy was designed based on the missing value pattern to identify candidate micropeptides. The appropriate statistical test method (t test or Mann-Whitney U test) was automatically selected according to the data distribution characteristics. The FDR correction (Benjamini-Hochberg method) was applied to control the false positive rate of multiple tests and obtain the micropeptide-sample quantitative matrix.

[0012] (3) Feature screening: In order to screen out the most discriminative micropeptide biomarkers, the micropeptide-sample quantitative matrix adopts a stratified random sampling strategy, dividing it into training set and test set. The training set is preprocessed with missing value imputation (SimpleImputer method, for example, using feature mean to fill) and Z-score standardization (StandardScaler method) to learn the imputation values ​​and standardization parameters (mean and standard deviation). The learned parameters are applied to the transformation of the training set and the test set to ensure the fairness of the model generalization ability evaluation. The SMOTE (Synthetic Minority Oversampling) algorithm is applied to the training set to handle class imbalance. The low variance filtering method is applied to the balanced training set after SMOTE processing to remove micropeptides whose expression levels hardly change in all samples. Then, the recursive feature elimination (RFECV) method combined with cross-validation is used. The evaluation results of multiple machine learning algorithms (such as random forest, gradient boosting tree) and information theory methods (mutual information) are combined to comprehensively rank all micropeptides, thereby screening out a stable combination of micropeptide biomarkers with significant diagnostic accuracy and development potential.

[0013] (4) Model training and validation: Based on the selected combination of micropeptide biomarkers, a classification prediction model is constructed. During the model training stage, in particular, to address the common class imbalance problem in clinical samples (such as fewer cancer samples than healthy samples), SMOTE (Synthetic Minority Oversampling) is used on the training set to avoid the model being biased towards the majority class and to improve the ability to identify cancer samples. By comparing the performance of various machine learning models (such as logistic regression, support vector machine, etc.) on an independent validation set, the optimal model is selected. The optimal model is then validated on an independent test set to evaluate its generalization ability and diagnostic accuracy in real-world scenarios.

[0014] Furthermore, in step (1), the blood samples of the subjects include colorectal cancer samples and normal control samples. The inclusion criteria for the samples include: blood samples of patients with pathologically confirmed gastrointestinal tumors before treatment, samples of healthy controls matched for age / sex, and all samples have obtained written informed consent from the donors.

[0015] Furthermore, in step (1), each 300 μL blood sample is pretreated as follows before mass spectrometry detection:

[0016] a. Take 300 μL of serum from each sample into a 10 KD ultrafiltration tube and centrifuge at 10000g at 4℃ for 90 min;

[0017] b. Transfer the supernatant from step a to a new EP tube, add 200 μL of High Select Top14 high abundance protein removal resin (Thermo Scientific), incubate at room temperature for 30 min, centrifuge at 12000 rpm for 1 min, and transfer the supernatant to a new EP tube;

[0018] c. Protein quantification of the supernatant from step b was performed using the BCA kit (beyotime); 80 μg of total protein supernatant was taken, an equal volume of acetone was added, mixed well, and 4 times the volume of supernatant was added in pre-cooled acetone at -20℃. The mixture was then allowed to stand at -20℃ for 3 hours.

[0019] d. Centrifuge the suspension obtained in step c at 13000 rpm for 15 min;

[0020] e. After dissolving the precipitate in step d with 300 μL of enzyme digestion buffer, incubate overnight at 37°C; the enzyme digestion buffer is prepared at pH 8.0 using 50 mM N2O3. HC The enzyme Lys-C (MedChemExpress) and sequencing-grade trypsin were dissolved in a buffer solution; the ratio of the endopeptide to the total protein in step c was 100:1, and the ratio of the trypsin to the total protein in step c was 50:1.

[0021] f. Add 1.5 μL of 5 mM TCEP aqueous solution (tris(2-carboxyethyl)phosphonic acid hydrochloride) and 6 μL of 11 mM IAM aqueous solution (iodoacetamide) to the 300 μL reaction system in step e, and incubate at 37°C in the dark for 1 h. Then adjust the pH to <2 with 50% TFA aqueous solution (trifluoroacetic acid) to terminate the digestion.

[0022] g. The product after the digestion in step f was dried under vacuum at 37°C and 4000 rpm using an Alpha 2-4 LSCbasic freeze dryer (Martin ChristGefriertrocknungsanlagen GmbH), and then redissolved in approximately 30 μL of 0.1% TFA aqueous solution.

[0023] h. The volume after cutting is approximately 2cm. 3 The C18 solid-phase extraction membrane (Empore) was packed into the chromatography column, and 20 μL of acetonitrile (ACN) was added to fully wet the packing material. The column was then centrifuged at 600 × g for 1 minute.

[0024] i. In step h, add 20 μL of a 50% ACN solution containing 0.1% TFA, adjust the packing environment, and centrifuge at 800 × g for 1 minute; the 50% ACN solution containing 0.1% TFA is prepared by mixing 500 μL of a 10% TFA aqueous solution, 25 ml of 100% acetonitrile, and 19.5 ml of double-distilled water;

[0025] j. Add 20 μL of 0.1% TFA aqueous solution to step i, centrifuge at 1000 × g for 1 minute, and repeat twice to ensure that the packing material is in a suitable state for binding with the target analyte;

[0026] k. Centrifuge the dissolved sample from step g at 12000 rpm, take the supernatant and add it to the chromatography column from step j to bind the target analyte to the packing material, and centrifuge at 600 × g for 2 minutes.

[0027] l. In step k, add 0.1% TFA aqueous solution, centrifuge at 1000 × g for 1 minute, repeat 2-3 times to remove non-specifically adsorbed impurities;

[0028] m. In step l, add 50 μL of 40% ACN solution containing 0.1% TFA, centrifuge at 600 × g for 1 minute, elute the target analyte and collect the eluent;

[0029] n. The eluent in step m was vacuum dried using an Alpha 2-4 LSCbasic freeze dryer at 37°C and 4000 rpm, and then redissolved in 5-10 μL of 0.1% TFA aqueous solution to obtain a solution for mass spectrometry detection.

[0030] Furthermore, the conditions for obtaining mass spectrometry data in step (1) are as follows:

[0031] Chromatographic separation was performed using a C18 reversed-phase column (75 μm inner diameter, 360 μm outer diameter, 150 mm length, 2 μm packing particle size); each sample was injected twice; gradient elution was performed at a constant flow rate of 300 nL / min for 70 minutes; mobile phase A was water / 0.1% formic acid, and mobile phase B was 80% acetonitrile / 0.1% formic acid; the elution gradient was set as follows: the initial concentration of phase B was 2%, linearly increasing to 28% within 58 minutes, then increasing to 35% within 65 minutes, and finally increasing to 98% within 70 minutes;

[0032] Mass spectrometry was performed using a Thermo Q Exactive HF-X mass spectrometer, controlled by Xcalibur 4.1 software in data-dependent acquisition mode. The acquisition sequence began with a full-scan mass spectrum (scan range 350–1800 m / z, resolution 60,000), followed by 20 data-dependent MS / MS scans (collision energies 30%). The automatic gain control target value was set to 3e6, the maximum injection time was 50 ms, and the MS2 spectrum resolution was set to 15,000.

[0033] Furthermore, step (2) preprocessing involves preliminary analysis using Thermo Xcalibur Qual Browser and Proteome software. First, the filenames are pattern matched to merge the data from all technical duplicate files belonging to the same biological sample, and then they are uniformly numbered to form a data table.

[0034] Further, the correction method in step (2) is as follows: taking the original precursor ion intensity value ms1_int_sum of each peptide in each sample as input, selecting the optimize.minimize function in the SciPy library, and selecting the SLSQP or L-BFGS-B algorithm as the solver, a unique normalization factor is calculated for each sample; the original precursor ion intensity value is multiplied by the corresponding normalization factor to obtain the corrected precursor ion intensity value ms1_int_sum_apex_dn.

[0035] Further, step (2) summarizes the precursor ion intensity information at the peptide level to the micropeptide level as follows: group the data according to the micropeptide identifier (protein_group) and the sample identifier (sample_group), sum the corrected precursor ion intensity values ​​ms1_int_sum_apex_dn of all peptides belonging to the same micropeptide in the same sample, and obtain the final quantitative abundance of the micropeptide in the sample; in order to make the data distribution closer to the normal distribution and stabilize the variance, perform a Log2 logarithmic transformation on all quantitative abundance values; before the transformation, for zero values, a very small pseudo-count can be added or the missing value filling strategy in subsequent steps can be used to handle it.

[0036] Furthermore, the dual screening strategy in step (2) is as follows: 1) Select proteins with high detection rates in at least one group (missing rate <50%); 2) Identify proteins with significant differences in detection rates between groups, and fill missing values ​​with one-tenth of the group mean to maintain data distribution characteristics.

[0037] Furthermore, step (3) uses 70% of the training set and 30% of the validation set.

[0038] Furthermore, the method for filtering low variance features in step (3) is as follows: redundant micropeptide features with minimal changes in expression levels in all samples are removed, and the feature_selection.VarianceThreshold module in the Scikit-learn library is used for preliminary dimensionality reduction.

[0039] Furthermore, step (3) combines the recursive feature elimination method with cross-validation as follows: use the feature_selection.RFECV module in the Scikit-learn library, combine the ensemble.RandomForestClassifier module as the evaluator and the model_selection.StratifiedKFold module to perform hierarchical cross-validation, use AUC as the performance index, and determine the best feature subset by iteratively removing the least important features.

[0040] Furthermore, the method for balancing the training set samples in step (3) is as follows: only the over_sampling.SMOTE module in the Imbalanced-learn library is used on the training set, and the SMOTE (Synthetic Minority Oversampling) algorithm is applied to generate synthetic samples to balance the number of positive and negative classes.

[0041] Furthermore, the method for multi-model training and selection in step (4) is as follows: on the balanced training set, using the selected marker features, multiple classifiers are trained using the ensemble.RandomForestClassifier module, linear_model.LogisticRegression, svm.SVC module and XGBClassifier module in the XGBoost library.

[0042] This invention also provides an application of the intelligent prediction model for colorectal cancer constructed based on mass spectrometry serum proteomics in assisting the diagnosis of colorectal cancer.

[0043] Compared with existing methods, the beneficial effects of this invention are mainly reflected in:

[0044] The method of this invention is rigorous, and by introducing advanced normalization algorithms, integrated feature selection strategies, and targeted sample imbalance handling, it significantly improves the robustness of biomarker screening and the accuracy of the predictive model. Furthermore, through model interpretability analysis, it provides theoretical support for the clinical application of biomarkers. (iv) Description of the attached drawings

[0045] Figure 1 The flowchart of the model construction of this invention.

[0046] Figure 2 DEG diagram of the corresponding peptide.

[0047] Figure 3 , Probability confusion matrix of validation set for colorectal cancer samples.

[0048] Figure 4 A schematic diagram of the area under the ROC curve for the validation set of colorectal cancer samples.

[0049] Figure 5 The probability confusion matrix of the colorectal cancer sample test set.

[0050] Figure 6 A schematic diagram of the area under the ROC curve for a colorectal cancer sample test set.

[0051] Figure 7 Validating the probability confusion matrix of colorectal cancer using traditional peptide detection indicators. (V) Detailed Implementation Methods

[0052] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto:

[0053] All technical solutions disclosed in this invention were implemented under Windows / macOS / Linux operating systems. The core algorithms and data analysis processes were implemented using the Python (version 3.9) programming language. The development and execution environment used the Visual Studio Code integrated development environment and its Jupyter Notebook plugin. Data processing and modeling relied on a series of open-source scientific computing libraries, mainly including Pandas, NumPy, Scikit-learn, Imbalanced-learn, and XGBoost.

[0054] The room temperature mentioned in the embodiments of the present invention refers to 25-30℃.

[0055] Example 1: Construction of an intelligent prediction model for colorectal cancer based on mass spectrometry serum proteomics

[0056] Reference Figure 1 The process for building an intelligent prediction model for colorectal cancer involves the following steps:

[0057] 1. Acquisition of raw micropeptide omics data

[0058] 1.1 Description of Sample Source:

[0059] In this embodiment, blood samples were collected from 174 colorectal cancer (CRC) patients and 25 healthy controls. All biological samples were clinical blood samples collected from cooperating medical institutions and approved by the ethics committee. The sample collection process strictly followed the relevant provisions of the "Ethical Review Measures for Biomedical Research Involving Human Subjects".

[0060] The blood samples used must meet the following pretreatment conditions:

[0061] a. Collect 3-5 mL of peripheral venous blood using an EDTA anticoagulant tube; b. Centrifuge at 4°C (3000 rpm, 10 min) within 2 hours after collection; c. Immediately aliquot and store the separated serum at -80°C.

[0062] The inclusion criteria for samples included: a. Blood samples from patients with pathologically confirmed gastrointestinal tumors before treatment; b. Samples from age / sex-matched healthy controls; c. Written informed consent from the donors was obtained for all samples.

[0063] 1.2 Preparation of peptide enzymatic hydrolysis mixture

[0064] a. Take 300 μL of the serum collected in step 1.1 into a 10 KD ultrafiltration tube and centrifuge at 10000g at 4℃ for 90 min.

[0065] b. Transfer all supernatant from step a to a new EP tube, add 200 μL of High Select Top14 high-abundance protein removal resin (purchased from Thermo Scientific, catalog number A36372), incubate at room temperature for 30 min, centrifuge at 12000 rpm for 1 min, and transfer the supernatant to a new EP tube.

[0066] c. Take a small amount of the supernatant from step b and use the BCA kit (beyotime) to determine the protein concentration according to the manufacturer's instructions. Take a portion of the remaining supernatant containing 80 μg of total protein, add an equal volume of acetone, and mix well. Add 4 volumes of pre-chilled acetone (-20°C) and incubate at -20°C for 3 hours.

[0067] d. After step c, centrifuge at 13000 rpm for 15 min, discard the supernatant, add an appropriate amount of ddH2O, centrifuge again, and discard the supernatant.

[0068] e. Add 300 μL of enzyme digestion buffer to the entire precipitate from step d, and incubate overnight at 37°C; the enzyme digestion buffer is prepared at pH 8.0 using 50 mM N2O3. HC The enzyme Lys-C (MedChemExpress, CAS No. 72561-05-8) and sequencing-grade trypsin (Shanghai Yaxin Biotechnology Co., Ltd., catalog No. CS0402) were dissolved in buffer solution to prepare the enzyme. The ratio of the enzyme to the total protein in step c was 100:1, and the ratio of the trypsin to the total protein in step c was 50:1.

[0069] f. Add 1.5 μL of 5 mM TCEP (tris(2-carboxyethyl)phosphonium hydrochloride) aqueous solution and 6 μL of 11 mM IAM (iodoacetamide) aqueous solution to the 300 μL system incubated overnight in step e, and incubate at 37°C in the dark for 1 h. Then adjust the pH to <2 with 50% TFA (trifluoroacetic acid) aqueous solution to terminate digestion.

[0070] g. The product after the digestion in step f was dried under vacuum at 37°C and 4000 rpm using an Alpha 2-4 LSCbasic freeze dryer (Martin ChristGefriertrocknungsanlagen GmbH), and then redissolved in about 30 μL of 0.1% TFA aqueous solution to obtain a peptide enzymatic hydrolysis mixture sample.

[0071] 1.3 C18 solid phase extraction membrane pretreatment

[0072] h. The volume after cutting is approximately 2cm. 3 Add the C18 solid-phase extraction membrane (Empore) to the chromatography column, then add 20 μL of acetonitrile (ACN) to fully wet the membrane, and centrifuge at 600 × g for 1 minute.

[0073] i. In step h, add 20 μL of 50% ACN solution containing 0.1% TFA (prepared by mixing 500 μL of 10% TFA aqueous solution, 25 mL of acetonitrile, and 19.5 mL of double-distilled water), adjust the membrane environment, and centrifuge at 800 × g for 1 minute.

[0074] j. In step i, add 20 μL of 0.1% TFA aqueous solution, centrifuge at 1000 × g for 1 minute, and repeat twice to ensure that the membrane is in a suitable state for the binding of the target analyte.

[0075] 1.4 Peptide purification

[0076] k. Centrifuge the peptide enzymatic digestion mixture sample from step g at 12000 rpm for 1 min, take the supernatant and add it to the chromatography column from step j to bind the target analyte to the membrane, and centrifuge at 600 × g for 2 min.

[0077] l. In step k, add 0.1% TFA aqueous solution, centrifuge at 1000 × g for 1 minute, repeat 2-3 times to remove non-specifically adsorbed impurities.

[0078] m. In step 1, add 50 μL of 40% ACN solution containing 0.1% TFA (prepared by mixing 400 μL of 10% TFA aqueous solution, 25 mL of acetonitrile, and 19.6 mL of double-distilled water), centrifuge at 600 × g for 1 minute, elute the target analyte, and collect the eluent.

[0079] n. The elution buffer in step m was vacuum dried using an Alpha 2-4 LSCbasic freeze dryer at 37°C and 4000 rpm, and then redissolved in 5-10 μL of 0.1% TFA aqueous solution to obtain the purified peptide sample.

[0080] 1.5 Raw micropeptide omics data

[0081] The purified peptide samples obtained in step 1.4 were analyzed using high-resolution liquid chromatography-mass spectrometry (LC-MS / MS) for micropeptidomics. The raw mass spectrometry data were processed using the Alphapept database retrieval software to generate a peptide identification and quantification list containing information such as peptide sequence, associated protein / micropeptide group, and precursor ion intensity (e.g., ms1_int_sum), in .csv file format. Each sample may correspond to one or more technically replicated .csv files.

[0082] LC-MS / MS analysis conditions: Chromatographic separation was performed using a C18 reversed-phase column (75 μm inner diameter, 360 μm outer diameter, 150 mm length, 2 μm particle size). Each sample was injected twice. Gradient elution was performed at a constant flow rate of 300 nL / min for 70 minutes. Mobile phase A was water / 0.1% formic acid, and mobile phase B was 80% acetonitrile / 0.1% formic acid. The elution gradient was set as follows: the initial concentration of phase B was 2%, linearly increasing to 28% within 58 minutes, then to 35% within 65 minutes, and finally to 98% within 70 minutes.

[0083] Mass spectrometry was performed using a Thermo Q Exactive HF-X mass spectrometer, controlled via Xcalibur 4.1 software in data-dependent acquisition mode. The acquisition sequence began with a single full-scan mass spectrum (350–1800 m / z, 60,000 m / s resolution), followed by 20 data-dependent MS / MS scans (30% collision energy). The automatic gain control target was set to 3e6, and the maximum injection time was 50 ms. The MS2 spectrum resolution was set to 15,000 m / s.

[0084] 2. Data Preprocessing and Quality Control

[0085] 2.1 Technical duplication merging and sample mapping:

[0086] To eliminate differences between technical replicates, the raw data were initially analyzed using Thermo Xcalibur Qual Browser and Proteome software. First, the filenames were pattern matched to merge the data from all technical replicates belonging to the same biological sample. Then, the data were uniformly numbered to form a single data table for easier subsequent management and analysis.

[0087] 2.2 Peptide Filtration

[0088] To reduce the false positive rate, the matching results of the reverse sequence library (decoy database) in the mass spectrometry database search were removed, that is, the entries with "REV" prefix in the protein_group column were deleted.

[0089] 2.3 Cross-sample precursor ion intensity correction based on optimization algorithm:

[0090] To address the systematic biases between different mass spectrometry batches, this invention employs an advanced delayed normalization algorithm. Unlike simple total ion current (TIC) normalization, this algorithm's core idea is to find a set of normalization factors that minimize global error through an optimization process.

[0091] 2.3.1 Algorithm Principle and Objective Function: The goal of this algorithm is to find a set of normalization factors. (in (This is the sample index), so that all sample pairs Peptides detected together The global error of the intensity ratio is minimized. Objective function It can be represented as:

[0092]

[0093] in: It is a peptide segment In the sample The original strength in; The sample to be solved The normalization factor; It is a sample and samples The set of peptides detected together; This represents the logarithmic ratio of the normalized intensity.

[0094] 2.3.2 Implementation: Using the original precursor ion intensity value ms1_int_sum for each peptide in each sample as input, the optimize.minimize function from the SciPy library is selected, and either SLSQP or L-BFGS-B algorithm is used as the solver to calculate a unique normalization factor for each sample. The original precursor ion intensity value is multiplied by the corresponding normalization factor to obtain the corrected precursor ion intensity value ms1_int_sum_apex_dn.

[0095] 2.4 Micropeptide Quantification and Data Conversion:

[0096] The precursor ion intensity information at the peptide level is aggregated to the micropeptide level. Specifically, the data is grouped according to micropeptide identifiers (protein_group) and sample identifiers (sample_group). The corrected precursor ion intensity values ​​(ms1_int_sum_apex_dn) of all peptides belonging to the same micropeptide in the same sample are summed to obtain the final quantitative abundance of that micropeptide in that sample. To make the data distribution closer to a normal distribution and stabilize the variance, all quantitative abundance values ​​are log2 transformed. Before transformation, zero values ​​can be handled by adding a very small pseudo-count or by using a missing value imputation strategy in subsequent steps.

[0097] 2.5 Differentially expressed peptide analysis:

[0098] A dual screening strategy based on missing value patterns was designed to identify candidate peptides: (1) proteins with high detection rates in at least one group (missing rate <50%) were selected; (2) proteins with significant differences in detection rates between groups were identified, and missing values ​​were filled with one-tenth of the group mean to maintain data distribution characteristics. Appropriate statistical tests (t-test or Mann-Whitney U test) were automatically selected based on data distribution characteristics, and FDR correction (Benjamini-Hochberg method) was applied to control the false positive rate of multiple tests, thereby obtaining 37 significantly differentially expressed peptides ( Figure 2 The generated micropeptide-sample quantification matrix.

[0099] 3. Feature Filtering

[0100] 3.1 Dataset Partitioning:

[0101] The peptide-sample quantification matrix generated in step 2.5 is divided into training and testing sets using a stratified random sampling strategy. In this example, the division ratio is 70% training set and 30% validation set. Stratified sampling ensures that the proportion of case samples (e.g., CRC patients) to control samples (e.g., healthy individuals) in each subset of data is consistent with the original population distribution, laying the foundation for robust evaluation of the model in subsequent steps.

[0102] 3.2 Data Standardization:

[0103] We constructed a preprocessing pipeline that includes imputation of missing values ​​using the SimpleImputer method (e.g., filling with the feature mean) and Z-score standardization using the StandardScaler method. This pipeline fits the model to the training set to learn the imputed values ​​and standardized parameters (mean and standard deviation). The learned parameters are then applied to the transformation between the training and test sets. This process strictly adheres to machine learning standards, effectively preventing premature leakage of test set information and ensuring the fairness of the model's generalization ability evaluation.

[0104] 3.3 Imbalance in processing categories:

[0105] To address the common class imbalance problem in clinical data (such as a significantly smaller sample size in the control group compared to the case group), the `over_sampling.SMOTE` module from the `Imbalanced-learn` library is used only on the training set. The `SMOTE` (Synthetic Minority Oversampling) algorithm is applied to generate synthetic samples to balance the number of positive and negative classes. `SMOTE` synthesizes new minority class samples similar to the original samples by linear interpolation between the nearest neighbors of the minority class samples (in this case, the control group), until the number of samples in both classes is balanced. Simultaneously, the original distribution of the test set remains unchanged to simulate real-world application scenarios.

[0106] 3.4 Preliminary Dimensionality Reduction:

[0107] For the balanced training set after SMOTE processing, low variance filtering (VarianceThreshold, threshold = 0.01) is first applied to remove redundant micropeptide features with minimal expression changes among samples. The feature_selection.VarianceThreshold module in the Scikit-learn library is used for preliminary dimensionality reduction to reduce the complexity of subsequent calculations and to initially eliminate noise.

[0108] 3.5 Automated Feature Subset Optimization:

[0109] To overcome the limitations of traditional methods that rely on manually setting the number of features based on experience, this invention employs a wrapper-style feature selection method, namely recursive feature elimination (RFECV) combined with cross-validation. This method uses the `feature_selection.RFECV` module from the Scikit-learn library, combined with the `ensemble.RandomForestClassifier` module (e.g., random forest) as the evaluator, and the `model_selection.StratifiedKFold` module for k-fold cross-validation (preferably k=5). Using the area under the receiver operating characteristic (AUC) curve as the performance metric, it iteratively removes the least important features, automatically finding the optimal number and combination of features that maximize the model's cross-validation performance, thereby obtaining an optimal combination of micropeptide biomarkers with high generalization potential (Table 1). The amino acid sequences are shown in SEQ ID NO.1~SEQ ID NO.14.

[0110] Table 1 Information on Micropeptide Biomarker Combinations

[0111]

[0112] 4. Model Training and Validation

[0113] 4.1 Model Training:

[0114] Using the optimal combination of micropeptide biomarkers selected in step 3.5, multiple machine learning classifiers based on different principles were trained in parallel on a balanced training set processed by SMOTE. These included: Random Forest (RF), Gradient Boosting Decision Tree (XGBoost), Logistic Regression (LR), and Support Vector Machine (SVM). Specifically, the ensemble.RandomForestClassifier module, linear_model.LogisticRegression, and svm.SVC module from the Scikit-learn library, as well as the XGBClassifier module from the XGBoost library, were used to train the various classifiers.

[0115] 4.2 Model selection based on validation set:

[0116] All trained candidate models are applied to independent validation sets for performance evaluation. The AUC value of each model on the validation set is calculated, and key metrics such as sensitivity, specificity, and precision are compared. The model with the best overall performance on the validation set is selected as the final diagnostic model. This step ensures the objectivity of model selection and avoids "overfitting" on the test set. In this embodiment, logistic regression (RF model) performs best; see... Figure 3 and Figure 4 .

[0117] 5. Model Evaluation and Interpretation:

[0118] 5.1 Generalization ability assessment:

[0119] The optimal model selected in 4.2 is applied to an independent test set that has never been involved in any training or tuning process before.

[0120] 5.2 Performance Visualization and Interpretation:

[0121] By plotting a confusion matrix, the model's prediction accuracy for different categories of samples is visually displayed, and the final sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) are calculated accordingly. Simultaneously, ROC curves are plotted on the test set, and AUC values ​​are calculated to comprehensively evaluate the model's overall discriminative ability and generalization performance. In this embodiment, the final model achieved an AUC value of 0.968 on the test set, and also performed well in sensitivity, specificity, and F1 score, demonstrating the model's high potential for clinical application. Figure 5 and Figure 6 .

[0122] Example 2: Application of Intelligent Prediction Model for Colorectal Cancer

[0123] 1. Traditional tumor marker detection products

[0124] Test sample set: 90 serum samples were randomly selected from the test set and validation set samples of Example 1, including 55 serum samples from clinically diagnosed cancer patients (positive samples) and 35 serum samples from healthy individuals (negative samples).

[0125] 2. Experimental methods for traditional tumor marker detection kits:

[0126] We selected four commonly used clinical enzyme-linked immunosorbent assay (ELISA) kits for tumor markers as benchmarks for comparison, specifically: human alpha-fetoprotein (AFP) ELISA kit (catalog number: JN19608); human gastrointestinal cancer marker CA199 (CA199) ELISA kit (catalog number: JN22858); human carcinoembryonic antigen (CEA) ELISA kit (catalog number: JN19902); and human carbohydrate antigen 125 (CA125) ELISA kit (catalog number: JN22666).

[0127] All the above kits were purchased from Shanghai Jining Biotechnology Co., Ltd. The content of the corresponding markers in the serum was detected according to the accompanying instructions, and the presence or absence of tumors was classified according to the detection threshold specified in the instructions.

[0128] 3. Experimental methods for the intelligent prediction model for colorectal cancer constructed in Example 1:

[0129] The intelligent prediction model for colorectal cancer constructed in Example 1 was used to detect the sample set to be tested, such as... Figure 6 As shown, the area under the ROC-micro curve is 0.97. This result suggests that the protein prediction model has high sensitivity and specificity, and has potential application value in the clinical prediction of benign and malignant breast tumors.

[0130] 4. Results

[0131] The results of testing 90 serum samples showed that the overall detection accuracy of the four existing ELISA kits was at a low level. Figure 7 The accuracy rates of the AFP kit were 38.9%, the CA199 kit 45.6%, the CEA kit 38.9%, and the CA125 kit 34.4%. These data indicate that using a single traditional tumor marker for detection has very limited ability to differentiate between cancer patients and healthy individuals, and the accuracy rate is far from meeting acceptable standards for clinical application.

[0132] In comparison, the colorectal cancer intelligent prediction model of this invention achieved an accuracy of 96.8% when detecting the same batch of unclassified test sets (90 samples).

[0133] Experimental Conclusion: This comparative experiment provides conclusive data demonstrating that the intelligent prediction model for colorectal cancer involved in this invention significantly outperforms existing single tumor marker ELISA detection methods in terms of accuracy. Its 96.8% accuracy rate showcases outstanding clinical diagnostic potential, fully validating the advanced nature, inventiveness, and significant technical effects of the technical solution described in this invention, and providing a more reliable new solution for early cancer screening and diagnosis.

Claims

1. A method for constructing an intelligent prediction model for colorectal cancer based on mass spectrometry serum proteomics, characterized in that, The method includes the following steps: (1) Acquisition of raw micropeptide omics data: After extracting low-abundance micropeptides from each blood sample by ultrafiltration and reverse enrichment, the disulfide bonds in the micropeptides are removed by traditional mass spectrometry pretreatment, and the thiol groups are blocked by methylation. Then, the peptide mixture derived from blood micropeptides is obtained by enzymatic hydrolysis. The raw mass spectrometry data of the peptide mixture is obtained by high-resolution liquid chromatography-mass spectrometry. (2) Data preprocessing and quality control: The raw mass spectrometry data were preprocessed to eliminate differences between technical repetitions and to remove the reverse sequence library matching results used to assess the false positive rate in the mass spectrometry database search. Then, the delayed normalization algorithm was used to correct the peptide precursor ion intensity, which can effectively remove systematic bias between samples. Subsequently, the precursor ion intensity information at the peptide level was summarized to the micropeptide level. A dual screening strategy was designed based on the missing value pattern to identify candidate micropeptides. The appropriate statistical test method was automatically selected according to the data distribution characteristics. FDR correction was applied to control the false positive rate of multiple tests and obtain the micropeptide-sample quantitative matrix. (3) Feature screening: In order to screen out the most discriminative micropeptide biomarkers, the micropeptide-sample quantitative matrix adopts a stratified random sampling strategy, which divides it into training set and test set. The training set is preprocessed with missing value imputation and Z-score standardization to learn the imputation values ​​and standardized parameters. The learned parameters are applied to the conversion between the training set and the test set to ensure the fairness of the model generalization ability evaluation. The SMOTE algorithm was applied to the training set to handle class imbalance. The low variance filtering method was applied to the balanced training set after SMOTE to remove micropeptides whose expression levels hardly changed in all samples. Then, a recursive feature elimination method combined with cross-validation was used. The evaluation results of multiple machine learning algorithms and information theory methods were combined to sort all micropeptides in order to screen out a stable combination of micropeptide biomarkers with significant diagnostic accuracy and development potential. (4) Model training and validation: Based on the selected combination of micropeptide biomarkers, a classification prediction model is constructed. In the model training stage, in particular, to address the common class imbalance problem in clinical samples, SMOTE is used on the training set to avoid the model being biased towards the majority class and to improve the ability to identify cancer samples. By comparing the performance of multiple machine learning models on an independent validation set, the optimal model is selected. The optimal model is then validated on an independent test set to assess its generalization ability and diagnostic accuracy in real-world scenarios.

2. The method as described in claim 1, characterized in that, In step (1), the blood samples of the subjects include colorectal cancer samples and normal control samples. The inclusion criteria for the samples include: blood samples of patients with pathologically confirmed gastrointestinal tumors before treatment, samples of age / sex matched healthy controls, and all samples have obtained written informed consent from the donors.

3. The method as described in claim 1, characterized in that, Step (1) Each 300 μL blood sample is pretreated as follows before mass spectrometry detection: a. Take 300 μL of serum from each sample into a 10 KD ultrafiltration tube and centrifuge at 10000g at 4℃ for 90 min; b. Transfer the supernatant from step a to a new EP tube, add 200 μL of High Select Top14 high abundance protein removal resin, incubate at room temperature for 30 min, centrifuge at 12000 rpm for 1 min, and transfer the supernatant to a new EP tube; c. Protein quantification of the supernatant from step b was performed using the BCA kit; 80 μg of total protein supernatant was taken, an equal volume of acetone was added, mixed well, and 4 times the volume of supernatant was added in pre-cooled acetone at -20℃. The mixture was then allowed to stand at -20℃ for 3 hours. d. Centrifuge the suspension obtained in step c at 13000 rpm for 15 min; e. After dissolving the precipitate in step d with 300 μL of enzyme digestion buffer, incubate overnight at 37°C; the enzyme digestion buffer is prepared at pH 8.0 using 50 mM N2O3. HC The enzyme Lys-C and sequencing-grade trypsin were dissolved in a buffer solution to prepare the enzyme; the ratio of the endopeptide to the total protein in step c was 100:1, and the ratio of the trypsin to the total protein in step c was 50:

1. f. Add 1.5 μL of 5 mM TCEP aqueous solution and 6 μL of 11 mM IAM aqueous solution to the 300 μL reaction system in step e, incubate at 37°C in the dark for 1 h, and then adjust the pH to <2 with 50% TFA aqueous solution to terminate digestion. g. The product after the digestion in step f was dried under vacuum at 37°C and 4000 rpm using an Alpha 2-4 LSCbasic freeze dryer, and then redissolved in about 30 μL of 0.1% TFA aqueous solution. h. The volume after cutting is approximately 2cm. 3 The C18 solid-phase extraction membrane was packed into the chromatography column, and 20 μL of acetonitrile was added to fully wet the packing material. The column was then centrifuged at 600 × g for 1 minute. i. In step h, add 20 μL of a 50% ACN solution containing 0.1% TFA, adjust the packing environment, and centrifuge at 800 × g for 1 minute; the 50% ACN solution containing 0.1% TFA is prepared by mixing 500 μL of a 10% TFA aqueous solution, 25 ml of 100% acetonitrile, and 19.5 ml of double-distilled water; j. Add 20 μL of 0.1% TFA aqueous solution to step i, centrifuge at 1000 × g for 1 minute, and repeat twice to ensure that the packing material is in a suitable state for binding with the target analyte; k. Centrifuge the dissolved sample from step g at 12000 rpm, take the supernatant and add it to the chromatography column from step j to bind the target analyte to the packing material, and centrifuge at 600 × g for 2 minutes. l. In step k, add 0.1% TFA aqueous solution, centrifuge at 1000 × g for 1 minute, repeat 2-3 times to remove non-specifically adsorbed impurities; m. In step l, add 50 μL of 40% ACN solution containing 0.1% TFA, centrifuge at 600 × g for 1 minute, elute the target analyte and collect the eluent; n. The eluent in step m was vacuum dried using an Alpha 2-4 LSCbasic freeze dryer at 37°C and 4000 rpm, and then redissolved in 5-10 μL of 0.1% TFA aqueous solution to obtain a solution for mass spectrometry detection.

4. The method as described in claim 1, characterized in that, The conditions for obtaining mass spectrometry data in step (1) are as follows: Chromatographic separation was performed using a C18 reversed-phase column with an inner diameter of 75 μm, an outer diameter of 360 μm, a length of 150 mm, and a packing particle size of 2 μm. Each sample was injected twice. Gradient elution was performed at a constant flow rate of 300 nL / min for 70 minutes. Mobile phase A was water / 0.1% formic acid, and mobile phase B was 80% acetonitrile / 0.1% formic acid. The elution gradient was set as follows: the initial concentration of phase B was 2%, which linearly increased to 28% within 58 minutes, then increased to 35% within 65 minutes, and finally increased to 98% within 70 minutes. Mass spectrometry was performed using a Thermo Q Exactive HF-X mass spectrometer, controlled by Xcalibur 4.1 software in data-dependent acquisition mode. The acquisition sequence began with a full-scan mass spectrum, covering the range of 350–1800 m / z with a resolution of 60,000, followed by 20 data-dependent MS / MS scans with a collision energy of 30%. The automatic gain control target value was set to 3e6, the maximum injection time was 50 ms, and the resolution of the MS2 spectrum was set to 15,000.

5. The method as described in claim 1, characterized in that, Step (2) preprocessing involves preliminary analysis using Thermo XcaliburQual Browser and Proteome software. First, filename pattern matching is performed to merge data from all technically repeated files belonging to the same biological sample, and then these are uniformly numbered to form a single data table. The correction method is as follows: using the original precursor ion intensity value ms1_int_sum of each peptide in each sample as input, the optimize.minimize function from the SciPy library is selected, and the SLSQP or L-BFGS-B algorithm is used as the solver to calculate a unique normalization factor for each sample. The original precursor ion intensity value is multiplied by the corresponding normalization factor to obtain the corrected value. The precursor ion intensity value ms1_int_sum_apex_dn is used to summarize the precursor ion intensity information at the peptide level to the micropeptide level. The method is as follows: the data are grouped according to the micropeptide identifier and the sample identifier. The corrected precursor ion intensity values ​​ms1_int_sum_apex_dn of all peptides belonging to the same micropeptide in the same sample are summed to obtain the final quantitative abundance of the micropeptide in the sample. In order to make the data distribution closer to the normal distribution and stabilize the variance, all quantitative abundance values ​​are transformed by log2 logarithm. Before the transformation, for zero values, a very small pseudo-count can be added or the missing value filling strategy in subsequent steps can be used to handle them.

6. The method as described in claim 1, characterized in that, Step (2) The dual screening strategy is as follows: 1) Select proteins with high detection rates in at least one group and a missing rate of <50%; 2) Identify proteins with significant differences in detection rates between groups and fill missing values ​​with one-tenth of the group mean to maintain the data distribution characteristics.

7. The method as described in claim 1, characterized in that, Step (3) 70% training set and 30% validation set.

8. The method as described in claim 1, characterized in that, Step (3) The low variance feature filtering method is as follows: remove redundant micropeptide features with minimal expression changes in all samples, and use the feature_selection.VarianceThreshold module in the Scikit-learn library for initial dimensionality reduction; the recursive feature elimination method combined with cross-validation is as follows: use the feature_selection.RFECV module in the Scikit-learn library, combined with the ensemble.RandomForestClassifier module as the evaluator and the model_selection.StratifiedKFold module for hierarchical cross-validation, using AUC as the performance index, and iteratively remove the least important features to determine the best feature subset; the training set sample balancing method is as follows: use the over_sampling.SMOTE module in the Imbalanced-learn library only on the training set, apply the SMOTE algorithm, and generate synthetic samples to balance the number of positive and negative classes.

9. The method as described in claim 1, characterized in that, Step (4) The method for training and selecting multiple models is as follows: On the balanced training set, using the selected marker features, train multiple classifiers using the ensemble.RandomForestClassifier module, linear_model.LogisticRegression, svm.SVC module in the Scikit-learn library and the XGBClassifier module in the XGBoost library.

10. The application of the intelligent prediction model for colorectal cancer constructed by the method of claim 1 in assisting the diagnosis of colorectal cancer.

Images