A method for constructing a machine learning model for auxiliary diagnosis of hepatocellular carcinoma based on a serum protein marker combination

Abstract

The application belongs to the technical field of biological medicine, and particularly relates to a method for constructing a machine learning model for auxiliary diagnosis of hepatocellular carcinoma based on a serum protein marker combination. The serum protein marker combination is composed of the following 7 serum protein markers: C8B, ADAM10, PRAP1, KIF27, HNRNPUL1, MYH7 and DSC3. The Gaussian Process classification model constructed in this way exhibits an AUC as high as 0.828 in the test set, and the sensitivity is as high as 0.85, which can very accurately separate SHHCC from conventional HCC before operation, solves the problem, and provides an effective tool for early identification and individualized treatment of SHHCC in the clinic.

Description

Technical Field

[0001] This invention belongs to the field of biomedical technology, specifically relating to a method for constructing a machine learning model for the auxiliary diagnosis of hepatocellular carcinoma based on a combination of serum protein biomarkers. Background Technology

[0002] Steatohepatitic hepatocellular carcinoma (SHHCC) is a specific pathological subtype of hepatocellular carcinoma (HCC). Compared to conventional hepatocellular carcinoma (CT-HCC), SHHCC is histopathologically characterized by significant steatosis, large bullous fat droplets, ballooning cell degeneration, and inflammatory cell infiltration. Due to its unique metabolic remodeling and immune microenvironment evolution, its malignancy and postoperative recurrence risk differ significantly from those of conventional liver cancer.

[0003] Currently, the clinical gold standard for SHHCC still heavily relies on histopathological diagnosis from surgically removed specimens or biopsies. However, preoperative biopsies are invasive and prone to missed diagnoses due to tumor heterogeneity. There are currently no serum proteins or other non-invasive diagnostic methods for SHHCC. Therefore, developing a combination of highly specific, non-invasive molecular diagnostic biomarkers based on peripheral serum and a machine learning model is of significant clinical importance in guiding the selection of personalized targeted / immunotherapy regimens. Summary of the Invention

[0004] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments.

[0005] As one aspect of the present invention, the present invention provides a reagent for detecting a combination of serum protein markers in the preparation of an auxiliary diagnostic kit for hepatocellular carcinoma of the steatohepatitis type, characterized in that the combination of serum protein markers consists of the following 7 serum protein markers: C8B, ADAM10, PRAP1, KIF27, HNRNPUL1, MYH7 and DSC3.

[0006] Preferably, the reagent includes antibodies, antibody fragments, or mass spectrometry-quantitative isotope internal standards that specifically bind to the seven serum differentially expressed proteins.

[0007] Preferably, the auxiliary diagnostic kit provides auxiliary diagnosis through a non-invasive diagnostic system, the system comprising: The data input module is used to receive the expression level data of the seven serum differentially expressed protein markers in the sample to be tested; The core diagnostic model module, built on machine learning algorithms, is used to calculate the input expression data and output diagnostic probability values. The result output module is used to indicate whether the sample to be tested is hepatocellular carcinoma of the steatohepatitis type based on the diagnostic probability value.

[0008] Preferably, the machine learning algorithm includes any one or more combinations of Gaussian process, logistic regression, stochastic gradient descent (SGD), quadratic discriminant analysis (QDA), and support vector machine (SVM).

[0009] Preferably, the underlying grid search optimization parameters of the Gaussian process are configured as follows: data preprocessing - Zscore, NRestart Optimizer: 0, Length Scale: 1.0; The underlying grid search optimization parameters for the logistic regression are configured as follows: Data preprocessing - Zscore, C: 1.0, Class Weight: balanced, Penalty: L2, Tol: 1e-05; The underlying grid search optimization parameters for the stochastic gradient descent are configured as follows: data preprocessing - Zscore, ClassWeight: None, Learning Rate: invscaling, Loss: square_loss, Penalty: L1; The underlying grid search optimization parameters for the secondary discriminant analysis are configured as follows: data preprocessing - Zscore, RegParam: 1.0, Tol: 1e-05; The underlying grid search optimization parameters of the support vector machine are configured as follows: data preprocessing - L2 norm regularization, C: 9.0, Class Weight: None, Gamma: 0.05, Kernel: rbf.

[0010] Preferably, the sample to be tested is human peripheral blood serum.

[0011] Preferably, the core diagnostic model module is constructed using the following method: 1) Obtain the original expression matrix of serum proteomics from hepatocellular carcinoma patients with known pathological phenotypes; 2) The original expression matrix was subjected to dual dimensionality reduction screening using LASSO regression and Boruta full feature selection algorithm, and the intersection was used to lock in the 7 core differential serum protein biomarkers; 3) Extract the expression levels of the seven core differentially expressed serum protein markers as feature variables, and input them into a machine learning algorithm in combination with known pathological phenotypes for training and cross-validation to obtain the core diagnostic model.

[0012] Preferably, in step 2), the LASSO regression uses 10-fold cross-validation and penalty parameter optimization to screen out expression features with non-zero coefficients; the Boruta full feature selection algorithm screens out important serum proteins with significant contributions through 500 iterations of random forest permutation.

[0013] Preferably, the reagent is a specific reagent for detecting the expression level of the serum protein marker composition in a serum sample.

[0014] The present invention also provides an in vitro diagnostic kit for identifying hepatocellular carcinoma of the steatohepatitis type, characterized in that it includes specific reagents for detecting the expression levels of the seven serum differentially expressed protein markers as described in claim 1.

[0015] The beneficial effects of this invention: This invention provides a highly specific, non-invasive combination of molecular diagnostic biomarkers for SHHCC and a machine learning model, solving the problem of the lack of specific serum proteins for SHHCC in existing technologies for non-invasive diagnosis. This invention uses a combination of LASSO regression and the Boruta algorithm to cross-screen a panel of 7 protein biomarkers—C8B, ADAM10, PRAP1, KIF27, HNRNPUL1, MYH7, and DSC3—from a high-dimensional serum proteome matrix. The Gaussian process classification model constructed based on this model showed an AUC as high as 0.828 and a sensitivity as high as 0.85 on the test set, enabling very accurate preoperative differentiation of SHHCC from conventional HCC, providing an effective tool for early identification and individualized treatment of SHHCC in clinical practice. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below, wherein: Figure 1 The graph shows the dual dimensionality reduction screening process for serum characteristic proteins (LASSO and Boruta intersection) and the differential expression statistics of 7 core differentially expressed serum proteins in the SHHCC group (ST group) and the conventional HCC group (CT group).

[0017] Figure 2 ROC curves and multidimensional performance index graphs of five machine learning classification models constructed based on seven serum differentially expressed proteins in the training set and internal test set. Detailed Implementation

[0018] To make the above-mentioned objectives, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to specific examples.

[0019] Example 1: 1. Clinical cohort and serum sample engineering This embodiment retrospectively collected and analyzed serum samples from independent hepatocellular carcinoma (HCC) patients based on liquid biopsy technology. High-dimensional serum proteomics expression matrices were obtained through standardized mass spectrometry quantitative detection of the discovery and validation sets, which were used to construct a highly specific non-invasive classifier.

[0020] Mass spectrometry experimental methods: Sample preparation: Protein enrichment: Dissolve the sample in an ice-water bath, centrifuge at 12000 rpm and 4℃ for 10 min, and take 120 μL of sample; Remove reagent H and disperse the magnetic beads evenly by blowing. Transfer 20 μL of reagent H to a new EP. Place the tube on a magnetic rack and let it stand for 2 minutes. Then, magnetically separate and discard the supernatant. Add 100 μL of reagent G, place on a constant temperature mixer, shake and wash the magnetic beads for 5 minutes (1300 rpm, room temperature), place on a magnetic rack and let stand for 2 minutes, then magnetically separate and discard the supernatant. Resuspend the magnetic beads in 100 μL of reagent G, add 100 μL of plasma, place on a constant temperature mixer, and incubate at 37°C with shaking for 1 h (1300 rpm). Place the beads on a magnetic rack and let them stand for 2 minutes. After magnetic separation, discard the supernatant, add 150 μL of reagent G, and wash with shaking for 5 minutes. Repeat the washing process three times. The low-abundance protein enriched on the magnetic beads will then be present.

[0021] Protein digestion: Add 40 μL of reagent A to the cleaned magnetic beads and mix thoroughly by pipetting. Incubate at 95℃ and 1000 rpm for 5 minutes; After heating is complete and the sample has returned to room temperature, add 15 μL of reagent B and mix well. Incubate at 37°C and 1300 rpm for 2 hours. After the enzymatic hydrolysis is complete, add 55 μL of reagent C, mix well, and terminate the enzymatic hydrolysis reaction.

[0022] Peptide desalting: Add all the sample to the desalting column and centrifuge at 700g for 1 min; Add 100 μL of reagent D to the desalting column, centrifuge at 700g for 1 min, and repeat once; Add 100 μL of reagent E to the desalting column, centrifuge at 700g for 1 min, and repeat once; Place the desalting column into a new centrifuge tube, add 100 μL of reagent F to the desalting column, centrifuge at 700g for 1 min, repeat once, and the final sample volume is 200 μL. The sample was concentrated using a vacuum refrigerated centrifuge, and the mobile phase A was reconstituted and then fed into the centrifuge.

[0023] Each sample (500g total peptides) was separated using a Vanquishneo nanoUPLC system and then coupled with a mass spectrometer (Astral) equipped with a nanoliter ion source for data acquisition. Chromatographic separation was performed using a 150μm ID×15cm reversed-phase column (EASYSpray™ HPLC (150μm×15cm), Thermo Scientific, USA). The mobile phase was an 80% acetonitrile-water-formic acid system, with mobile phase A being 0.1% formic acid aqueous solution and mobile phase B being 0.1% formic acid-80% acetonitrile solution. After equilibration of the column to 96% of phase A, the sample was directly loaded onto the column via an autosampler and then subjected to gradient separation for 8 min. Mass spectrometry analysis was performed using data-independent acquisition (DIA) mode, with a total analysis time of 8 min, and positive ion detection mode. The full scan range was 380-980 m / z, and the resolution was 240 kJ. The isolation window is 2 m / z, and the normalized collision energy (NCE) is 25%. The full scan range is 150-2000 m / z.

[0024] 2. Dual-engineering screening and trend description of serum characteristic proteins To accurately identify the gold standard for diagnosis of hepatocellular carcinoma (SHHCC) highly correlated with steatohepatitis type among thousands of serum proteins, this invention, in the R v4.2.3 environment, jointly invokes the glmnet (v4.1.10) and Boruta (v9.0.0) algorithms for feature dimensionality reduction (e.g.) Figure 1 (as shown) 1) Initial screening using LASSO regression: 10-fold cross-validation and penalty parameter optimization were used to initially screen 54 expression features with non-zero coefficients (including...). Figure 1 47 LASSO-specific features and intersection cores.

[0025] 2) Boruta Feature Selection: Through 500 iterations of random forest permutations, 30 important serum proteins (including) that significantly contribute to classification were identified. Figure 1 23 Boruta-specific features and their intersection core.

[0026] 3) Core Feature Identification: By taking the intersection of the two methods, seven core differentially expressed serum protein markers were ultimately precisely quantified (e.g., Figure 1 (As shown).

[0027] These 7 serum characteristic proteins in the SHHCC group ( Figure 1 The ST group and the conventional HCC group Figure 1 The specific expression trends and univariate statistical analysis among the CT group are detailed below: C8B (UniProt ID: P07358): showed a significant down-regulated trend in SHHCC tissue or peripheral serum (p = 0.0011).

[0028] ADAM10 (UniProt ID: O14672): showed a significant down-regulated trend in the SHHCC group (p = 0.022).

[0029] PRAP1 (UniProt ID: Q96NZ9): showed a significant down-regulated trend in the SHHCC group (p = 0.00069).

[0030] KIF27 (UniProt ID: Q86VH2): showed a significant down-regulated trend in the SHHCC group (p = 0.00056).

[0031] HNRNPUL1 (UniProt ID: Q9BUJ2): showed a significant up-regulated trend in the SHHCC group (p = 9.7 × 10^-05).

[0032] MYH7 (UniProt ID: P12883): showed a significant up-regulated trend in the SHHCC group (p = 0.0043).

[0033] DSC3 (UniProt ID: Q14574): showed a significant down-regulated trend in the SHHCC group (p = 4.8 × 10^-06).

[0034] Table 1. Seven proteins and their gene sequence numbers 3. Construction and Multidimensional Performance Evaluation of Serum Proteome Classification Model Based on the above seven serum differentially expressed protein biomarkers, this invention constructed five machine learning classification models: Gaussian Process, Logistic Regression, Stochastic Gradient Descent (SGD), Quadratic Discriminant Analysis (QDA), and Support Vector Machine (SVM). Figure 2 (As shown). The specific quantitative data for each model on the training set and the internal cross-validation test set are as follows: 3.1 Performance of the Model Training Set During the modeling and parameter fitting stages, all algorithms demonstrated stable classification capabilities: The Gaussian process performed best, with an area under the receiver operating characteristic curve (AUC) of 0.857. Among the auxiliary evaluation metrics, the accuracy was 0.79, the F1 score was 0.81, the precision was 0.77, the sensitivity was 0.86, and the specificity was 0.70.

[0035] Logistic Regression: AUC 0.845; accuracy 0.77, F1-score 0.80, precision 0.78, sensitivity 0.81, specificity 0.73.

[0036] SGD model: AUC 0.826; accuracy 0.79, F1-score 0.81, precision 0.78, sensitivity 0.85, specificity 0.72.

[0037] QDA model: AUC is 0.825; accuracy is 0.78, F1-score is 0.81, precision is 0.77, and sensitivity is 0.85.

[0038] SVM model: AUC is 0.820; although the sensitivity is slightly lower (0.61), its specificity is as high as 0.81, showing a very strong ability to exclude negative samples.

[0039] 3.2 Cross-validation performance on the model's internal test set In the internal cross-validation test set where no modeling was involved, the serum protein classifier still maintained extremely excellent robustness and generalization performance: Gaussian Process: Validation AUC reached 0.828, accuracy was 0.77, F1-score was 0.80, precision was 0.76, sensitivity was 0.85, and specificity was 0.67.

[0040] Logistic Regression: AUC was 0.822, accuracy was 0.76, F1-score was 0.77, precision was 0.79, and specificity reached 0.74.

[0041] The AUCs for SGD and QDA models were 0.820 and 0.818, respectively; their accuracy (0.79), F1-score (0.81), precision (0.80), sensitivity (0.85), and specificity (0.71) remained stable.

[0042] SVM model: AUC remains at 0.802; at the same time, its specificity remains stable at a high level of 0.80.

[0043] The training / test set is partitioned using five-fold cross-validation. The underlying grid search optimization parameters for the five serum protein machine learning models described in this embodiment are configured as follows: Gaussian Process: Data Preprocessing - Zscore, N Restart Optimizer: 0, LengthScale: 1.0 Logistic Regression: Data Preprocessing - Z-score, C: 1.0, Class Weight: balanced, Penalty: L2, Tol: 1e-05 SGD: Data Preprocessing - Z-score, Class Weight: None, Learning Rate: invscaling, Loss: square_loss, Penalty: L1 QDA: Data Preprocessing - Zscore, Reg Param: 1.0, Tol: 1e-05 SVM: Data Preprocessing - L2 Norm Regularization, C: 9.0, Class Weight: None, Gamma: 0.05, Kernel: rbf.

[0044] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. The application of a reagent for detecting a combination of serum protein markers in the preparation of an auxiliary diagnostic kit for hepatocellular carcinoma of the steatohepatitis type, characterized in that, The serum protein biomarker group consists of the following seven serum protein biomarkers: C8B, ADAM10, PRAP1, KIF27, HNRNPUL1, MYH7, and DSC3.

2. The application according to claim 1, characterized in that: The reagents include antibodies that specifically bind to the seven serum differentially expressed proteins, antibody fragments, or mass spectrometry-quantitative isotope internal standards.

3. The application according to claim 1 or 2, characterized in that: The auxiliary diagnostic kit provides auxiliary diagnosis through a non-invasive diagnostic system, which includes: The data input module is used to receive the expression level data of the seven serum differentially expressed protein markers in the sample to be tested; The core diagnostic model module, built on machine learning algorithms, is used to calculate the input expression data and output diagnostic probability values. The result output module is used to indicate whether the sample to be tested is hepatocellular carcinoma of the steatohepatitis type based on the diagnostic probability value.

4. The application according to claim 3, characterized in that: The machine learning algorithm includes any one or more combinations of Gaussian Process, Logistic Regression, Stochastic Gradient Descent (SGD), Quadratic Discriminant Analysis (QDA), and Support Vector Machine (SVM).

5. The application according to claim 4, characterized in that: The underlying grid search optimization parameters for the Gaussian process are configured as follows: Data preprocessing - Zscore, N Restart Optimizer: 0, Length Scale: 1.0; The underlying grid search optimization parameters for the logistic regression are configured as follows: Data preprocessing - Zscore, C: 1.0, ClassWeight: balanced, Penalty: L2, Tol: 1e-05; The underlying grid search optimization parameters for the stochastic gradient descent are configured as follows: data preprocessing - Zscore, ClassWeight: None, Learning Rate: invscaling, Loss: square_loss, Penalty: L1; The underlying grid search optimization parameters for the secondary discriminant analysis are configured as follows: data preprocessing - Zscore, Reg Param: 1.0, Tol: 1e-05; The underlying grid search optimization parameters of the support vector machine are configured as follows: data preprocessing - L2 norm regularization, C: 9.0, Class Weight: None, Gamma: 0.05, Kernel: rbf.

6. The application according to claim 5, characterized in that: The sample to be tested is human peripheral blood serum.

7. The application according to claim 3, characterized in that, The core diagnostic model module is constructed using the following methods: 1) Obtain the original expression matrix of serum proteomics from hepatocellular carcinoma patients with known pathological phenotypes; 2) The original expression matrix was subjected to dual dimensionality reduction screening using LASSO regression and Boruta full feature selection algorithm, and the intersection was used to lock in the 7 core differential serum protein biomarkers; 3) Extract the expression levels of the seven core differentially expressed serum protein markers as feature variables, and input them into a machine learning algorithm in combination with known pathological phenotypes for training and cross-validation to obtain the core diagnostic model.

8. The application according to claim 7, characterized in that: In step 2), the LASSO regression uses 10-fold cross-validation and penalty parameter optimization to screen out expression features with non-zero coefficients; the Boruta full feature selection algorithm screens out important serum proteins with significant contributions through 500 iterations of random forest permutation.

9. The application according to claim 1, characterized in that: The reagent is a specific reagent used to detect the expression level of the serum protein biomarker composition in a serum sample.

10. An in vitro diagnostic kit for identifying hepatocellular carcinoma of the steatohepatitis type, characterized in that, Includes specific reagents for detecting the expression levels of the seven serum differentially expressed protein markers as described in claim 1.

Images