Liver cancer screening model and method and apparatus for constructing the same

Abstract

The application discloses a hepatocellular carcinoma screening model and a construction method and device thereof. The construction method utilizes a fusion model based on machine learning, and specifically comprises the following steps: step one) collecting DNA methylation data of patients diagnosed with hepatocellular carcinoma and non-hepatocellular carcinoma; step two) analyzing the DNA methylation data of the patients diagnosed with hepatocellular carcinoma and the non-hepatocellular carcinoma to obtain methylation information data corresponding to each sequencing sequence site meeting the requirements; and step three) utilizing the methylation information data of the previous step to construct, optimize and test the fusion model. The application can efficiently distinguish hepatocellular carcinoma samples from non-hepatocellular carcinoma samples, and the method has high sensitivity, high specificity and non-invasiveness.

Description

Technical Field

[0001] This invention belongs to the fields of genomics and molecular biology, and particularly relates to cancer screening models and their construction methods and devices, especially to a hepatocellular carcinoma screening model and its construction methods and devices. Background Technology

[0002] Hepatocellular carcinoma (HCC) is a common malignant tumor, with incidence and mortality rates ranking among the highest in the world in my country. Currently, most HCC patients in China are diagnosed at an advanced or even terminal stage, making radical surgery impossible and severely impacting prognosis. Improving early detection rates and implementing personalized treatment can significantly reduce HCC mortality. However, early-stage HCC patients often present with mild or no symptoms. While liver ultrasound, serum alpha-fetoprotein (AFP) testing, and PIVKA-II testing are commonly used screening and diagnostic techniques for HCC, their sensitivity for early-stage HCC is relatively low. Therefore, there is an urgent need to develop a sensitive, reliable, and minimally invasive detection method to identify early-stage HCC for timely intervention.

[0003] Cell-free DNA (cfDNA) in plasma refers to degraded DNA fragments released into the blood plasma after cell death or apoptosis. In cancer patients, a portion of cfDNA originates from tumor cells and is also known as circulating tumor DNA (ctDNA). Next-generation sequencing (NGS) can analyze various tumor-specific signals carried by ctDNA, especially DNA methylation modifications, a valuable biomarker reflecting early tumor changes.

[0004] In recent years, the use of ctDNA-related biomarkers for clinical translational research, combined with non-invasive methods such as liquid biopsy (e.g., blood sampling) for early tumor diagnosis, has become a hot topic in early cancer screening research. After obtaining tumor-specific signals carried by ctDNA, machine learning methods are typically combined to build early cancer screening models using these signals as features. Fusion models belong to the ensemble learning model category in machine learning. This algorithm constructs and combines multiple learners to complete classification tasks. This method often achieves significantly better generalization performance than a single learner, resulting in a better-performing and more robust classifier, which helps solve key computational problems and has been applied in fields such as facial recognition and speech recognition. However, there are no reports of applying this technology to early hepatocellular carcinoma screening. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for constructing a hepatocellular carcinoma screening model, which can efficiently distinguish between hepatocellular carcinoma samples and non-hepatocellular carcinoma samples. This method has high sensitivity, high specificity, and is non-invasive.

[0006] Therefore, the technical solution of the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for constructing a screening model for hepatocellular carcinoma, the method utilizing a fusion model based on machine learning.

[0008] Preferably, the construction method includes at least the following steps:

[0009] Step 1: Collect DNA methylation data from patients diagnosed with hepatocellular carcinoma and from non-hepatocellular carcinoma patients;

[0010] Step 2: Analyze the DNA methylation data of patients diagnosed with hepatocellular carcinoma and non-hepatocellular carcinoma data to obtain methylation information data corresponding to each sequencing site that meets the requirements;

[0011] Step 3: Using the methylation information data from the previous step, construct, optimize, and test the fusion model.

[0012] Preferably, the non-hepatocellular carcinoma patients mentioned in step one include DNA methylation data from normal individuals, individuals with cirrhosis, and individuals with hepatitis;

[0013] Preferably, step one includes at least the following steps:

[0014] 101) Peripheral blood was drawn from the subject, centrifuged, plasma was separated, and cell-free DNA was extracted from the plasma; the subject was either a patient diagnosed with hepatocellular carcinoma or a non-hepatocellular carcinoma patient;

[0015] 102) Perform DNA methylation sequencing on cell-free DNA in plasma to construct a methylation sequencing library;

[0016] 103) The methylation sequencing library was hybridized and enriched with selected hepatocellular carcinoma biomarker probes to obtain the final library for sequencing and analysis of the subject's DNA methylation data.

[0017] Existing experimental methods for constructing DNA methylation sequencing libraries include, but are not limited to, bisulfite sequencing, enzymatic methylation conversion sequencing (EM-seq), and TAPS, all of which can be used in this invention.

[0018] Preferably, step two includes at least the following steps:

[0019] 201) After the DNA methylation data is downloaded, use data processing software to perform quality control, alignment, sorting and deduplication of the DNA methylation data described in step one;

[0020] 202) Set appropriate conditions to filter out sequencing sequences that do not meet the requirements (filtering conditions include but are not limited to: sequencing quality, sequence length, etc.) and obtain methylation information data corresponding to each sequencing sequence site that meets the requirements.

[0021] Specifically, the original FASTQ file was processed using fastp (0.21.0), with default parameters used to remove Illumina-specific adapters. Low-quality sequences were removed using the parameters -u 20 and -q 20. The processed clean fastq data was then methylated and aligned using bismark (0.23.0). Deduplication was performed using deduplicate_bismark. The alignment file generated by bismark includes methylation information for each base of each read, counts the number of CHG and CHH sequences in each read, and removes reads with a sum of CHG and CHH greater than or equal to 3. Finally, methylation annotation information was obtained using bismark_methylation_extractor (bismark).

[0022] Preferably, step three includes at least the following steps:

[0023] 301) Classify the methylation information data obtained in step two into data of patients diagnosed with hepatocellular carcinoma and data of non-hepatocellular carcinoma, and add them to the label column respectively, labeled as hepatocellular carcinoma and non-hepatocellular carcinoma;

[0024] 302) Divide the categorized data into training and test sets according to a ratio of 70% and 30% respectively;

[0025] 303) Construct a fusion model by inputting the feature data and labels of the training set to train and optimize the fusion model to obtain the final model;

[0026] 304) Input the test set into the final model and use the classification results to evaluate the performance of the final model.

[0027] Preferably, the fusion model described in step 303) is divided into two layers: the first layer consists of two models, and the second layer consists of a multinomial logistic regression model.

[0028] The first layer includes a hepatocellular carcinoma vs. normal model, trained using normal human data and hepatocellular carcinoma data in the training set (HCC vs. Normal, hereinafter referred to as the "HN model") and a hepatocellular carcinoma vs. liver disease model, trained using liver disease data and hepatocellular carcinoma data in the training set (HCC vs. Liver-disease, hereinafter referred to as the "HL model").

[0029] Preferably, both models in the first layer are trained using GDBT (Gradient Boosting Decision Tree), i.e., the gradient boosting iterative decision tree algorithm;

[0030] Preferably, the second-layer multinomial logistic regression model is constructed as follows: the two models of the first layer are fused using logistic regression in a stack manner, and the result of the fused model is used as a prediction of hepatocellular carcinoma and non-hepatocellular carcinoma of the sample; wherein, the process of constructing the second-layer model can be implemented using R language and / or Python language, preferably Python (v.3.6.7).

[0031] The first layer of the model uses the GBDT algorithm. The specific process of the GBDT algorithm is as follows:

[0032] 1-1) First, initialize the weak classifier:

[0033]

[0034] In formula (1), N is the total number of training set samples, y i Let x be the label of the training set sample, and let x be the feature data of the training set.

[0035] 1-2) Calculate the residual r of m trees using formula (2). m,i The iterations are performed sequentially, where m = 1, 2, ..., M, M is the maximum number of iterations; i = 1, 2, ..., N.

[0036]

[0037] 1-3) For the leaf node region R of the m-th regression tree mj Where j = 1, 2, ..., J m J m Let J be the number of tree nodes in the m-th regression tree. m For each leaf node region, the optimal residual fitting value C is calculated using formula (3). mj :

[0038]

[0039] 1-4) Update the classifier until a final strong classifier F is obtained. M (x), the calculation formulas in the process are shown in formulas (4) and (5):

[0040]

[0041]

[0042] The specific construction method of the first layer of the model can be described as follows: Using data from normal individuals and hepatocellular carcinoma (HCC) patients, as well as data from liver disease (cirrhosis + hepatitis) and HCC patients in the training set, as training data, two models are trained according to steps 1-1) to 1-4) above, named HN (HCC + normal) model and HL (HCC + liver disease) model, respectively. The feature engineering method is set to select features with weights greater than 0 based on the model's scoring of feature weights. The model parameters are optimized using grid search combined with cross-validation. After optimization, the first layer training is complete.

[0043] The second layer of the model is trained using a model constructed using the logistic regression algorithm. Specifically, the logistic regression algorithm can be described as follows:

[0044] 2-1) First, construct a prediction function:

[0045]

[0046] 2-2) Furthermore, in order to avoid overfitting, L2 regularization is added to the logistic regression algorithm. The specific process can be described as: taking the partial derivative of the loss function J and iterating until the loss value reaches the minimum and the model converges.

[0047] Specifically, as shown in formulas (7) and (8):

[0048]

[0049]

[0050] The specific construction method of the second layer of the model can be described as follows: the outputs of the two models in the first layer are used as features and input in matrix form. The model is then trained using the logistic regression algorithm according to steps 2-1) to 2-2) above to obtain the final model. The parameters of the second layer model are also optimized using grid search combined with cross-validation. After optimization, the second layer is trained and the final model is formed.

[0051] Specifically, the effect evaluation is as follows:

[0052] The optimized final model is input using the test set obtained from step three, and its performance is evaluated using the classification results. Based on the classification results, a feature classification performance graph of the final model is plotted; the sensitivity, specificity, and accuracy of the final model are calculated; the ROC curve and AUC value of the final model are plotted; and comparisons are made between the final model and existing detection methods AFP and PIVKA-II.

[0053] Secondly, the present invention provides an apparatus for constructing a hepatocellular carcinoma screening model, the apparatus comprising:

[0054] Module 1: Used to collect DNA methylation data from patients diagnosed with hepatocellular carcinoma and from non-hepatocellular carcinoma patients;

[0055] Module 2: Used to analyze DNA methylation data from patients diagnosed with hepatocellular carcinoma and non-hepatocellular carcinoma data to obtain methylation information data corresponding to each sequencing site that meets the requirements;

[0056] Module 3: Used to build, optimize, and test fusion models using the methylation information data from Module 2.

[0057] Thirdly, the present invention provides a hepatocellular carcinoma screening model, which is constructed by the method described in the first aspect of the present invention.

[0058] Fourthly, the present invention provides a hepatocellular carcinoma screening device, wherein the hepatocellular carcinoma screening device includes a hepatocellular carcinoma screening model constructed by the method described in the first aspect of the present invention.

[0059] Compared with existing technologies, the hepatocellular carcinoma screening model, its construction method, and apparatus provided by this invention have at least the following advantages:

[0060] This invention utilizes next-generation sequencing (NGS) to obtain marker DNA methylation data and constructs a machine learning fusion model to predict hepatocellular carcinoma. The constructed model exhibits good stability and robustness, and can efficiently distinguish between hepatocellular carcinoma and non-hepatocellular carcinoma samples, providing a highly sensitive, specific, and non-invasive approach to hepatocellular carcinoma detection. Attached Figure Description

[0061] Figure 1 This is the overall technical roadmap for the model construction method of this invention;

[0062] Figure 2 This is a technical roadmap for constructing the fusion model in Example 1;

[0063] Figure 3 This is a heatmap showing the feature classification effect of Example 1;

[0064] Figure 4 The ROC curve for evaluating the fusion model in Example 1 is shown.

[0065] Figure 5 This is a comparison chart of the results of the fusion model in Example 1 and existing detection methods. Detailed Implementation

[0066] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0067] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0068] Example 1

[0069] The roadmap for the fusion model construction technology in this embodiment is as follows: Figure 2 As shown.

[0070] Step 1: Collect DNA methylation data

[0071] This study recruited 596 volunteers, including 268 with hepatocellular carcinoma, 76 with liver disease, and 274 healthy individuals. All samples were processed using the following methods:

[0072] Peripheral blood was collected from volunteers into cfDNA preservation tubes (Xiamen Zhixuan). Plasma was separated by centrifugation, and cfDNA was extracted from the plasma. The cfDNA was ligated with methylated adapters using the VAHTS Universal DNA library Prep Kit for Illumina V3 (Novazia), and the ligation products were then enzymatically transformed. The Enzymatic Methyl-seq Conversion Module (NEB) was used. After amplification and purification of the transformation product, it was captured using the target probe. The captured library was obtained by PCR amplification and finally sequenced.

[0073] Step Two: Data Processing Methods

[0074] After the DNA methylation sequence sequencing data from step one was processed, the raw FASTQ file was processed using fastp (0.21.0). Illumina-specific adapters were removed using default parameters, and low-quality sequences were removed using the parameters -u 20 and -q 20. The processed clean fastq data was then aligned to methylation using bismark (0.23.0). Deduplication was performed using deduplicate_bismark. The alignment file generated by bismark includes methylation information for each base of each read, counts the number of CHG and CHH in each read, and removes reads with a sum of CHG and CHH greater than or equal to 3. Finally, methylation annotation information was obtained using bismark_methylation_extractor (bismark). The parameters were set as follows: comprehensive --bedGraph --counts --cytosine_report --CX --buffer_size 20G --parallel 16.

[0075] Step 3: Fusion Model Construction and Optimization

[0076] The data from the 596 volunteers obtained in step two were divided into two groups based on clinical information: one group included healthy individuals and patients with liver disease (non-hepatocellular carcinoma), and the other group included hepatocellular carcinoma. The two groups were then randomly divided: 70% was allocated to the training set (n=417, healthy individuals=195, liver disease=54, hepatocellular carcinoma=168) for model training, and 30% was allocated to the test set (n=179, healthy individuals=84, liver disease=23, hepatocellular carcinoma=72) for model testing. The process was single-blind. Model construction and optimization were preferably implemented using Python (v.3.6.7). Specifically, the GradientBoostingClassifier package from scikit-learn (v1.1.0) was used as input, along with data from healthy individuals, hepatocellular carcinoma patients, and patients with liver disease and hepatocellular carcinoma, respectively. The GBDT algorithm was then used to train two models: the HN model and the HL model. The first layer of each model was then trained. Then, the LogisticRegression package from scikit-learn (v1.1.0) is called to construct the second layer of the model using the outputs of the HN and HL models as features. L2 regularization is used to prevent overfitting, and the final model is trained using a stack-based logistic regression algorithm. The model parameters are optimized using grid search combined with cross-validation. Adjustable parameters include solver, max_iter, and C-value.

[0077] Step 4: Model Result Testing

[0078] 1) Input the test set data obtained in step three into the model to obtain the model prediction results for each sample in the test set. Then, calculate and graphically display the model performance. This part includes:

[0079] Based on the values ​​of hepatocellular carcinoma-related methylation site features used in the model, feature data heatmaps were plotted in the training and test sets for the non-hepatocellular carcinoma and hepatocellular carcinoma patient groups. Figure 3 Based on the selected features, the differences between hepatocellular carcinoma (HCC) and non-HCC populations can be clearly observed. Therefore, the combination of HCC-related methylation sites, in which the methylation signals at these sites in the cell-free DNA of plasma from HCC patients and normal controls show significant differences, can be used as liquid biopsy markers for detecting HCC.

[0080] 2) Compare the actual results on the test set, calculate the specific sensitivity and specificity, plot the ROC curves of the model on the test set and training set, and calculate the AUC and confidence interval. This part is implemented using R (v.3.6.0) language and calculated by the pROC (v.1.18.0) package. The test set compiles the model prediction results of hepatocellular carcinoma patients with clinical information including AFP and PIVKA-II test results, and compares the accuracy of these two clinically used protein tests with the model's test results.

[0081] The model results constructed in this embodiment are calculated as follows:

[0082] The model performed well on the training set with 5-fold cross-validation, achieving a sensitivity of 92% and a specificity of 94% (AUC = 0.980, 95% CI 0.968–0.991). On the test set, the sensitivity was 90% and the specificity was 97% (AUC = 0.957, 95% CI 0.925–0.989). See below for detailed results. Figure 4 The model's sensitivity in the hepatocellular carcinoma (HCC) test set increased with cancer stage, reaching 85% (35 / 41) for stage I patients, 89% (8 / 9) for stage II patients, and 100% (23 / 23) for stage III-IV patients. The model also demonstrated high specificity in the non-HCC test set, with 91% specificity in patients with liver disease and 99% specificity in the general population. Specific results are shown in Table 1.

[0083] In the test set of hepatocellular carcinoma patients, the model's prediction results were compared with those of commonly used existing techniques for serum AFP testing (threshold 20 ng / mL). It can be seen that the model's accuracy is higher than that of AFP in both early-stage (n=50) and late-stage (n=22) hepatocellular carcinoma patients. Similarly, for PIVKA-II testing (threshold 40 ng / mL), the model exhibits higher accuracy in both early-stage (n=15) and late-stage (n=11) hepatocellular carcinoma patients. See below for detailed results. Figure 5 .

[0084] Table 1 Classification effect of the fusion model test set in Example 1

[0085]

[0086] In summary, this invention utilizes next-generation sequencing (NGS) to obtain marker DNA methylation data and constructs a machine learning fusion model to predict hepatocellular carcinoma. The constructed model exhibits good stability and robustness. The model can efficiently distinguish between hepatocellular carcinoma samples and non-hepatocellular carcinoma samples, providing a highly sensitive, specific, and non-invasive method for detecting hepatocellular carcinoma.

[0087] It should be noted and understood that various modifications and improvements can be made to the invention described in the detailed description above without departing from the spirit and scope of the invention as claimed in the appended claims. Therefore, the scope of the claimed technical solutions is not limited to any specific exemplary teachings given.

[0088] The applicant declares that the above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention and should not be construed as limiting the specific implementation of the invention to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for constructing a screening model for hepatocellular carcinoma, characterized in that, The method utilizes a machine learning-based fusion model and includes the following steps: Step 1: Collect DNA methylation data from patients diagnosed with hepatocellular carcinoma and from non-hepatocellular carcinoma patients; Step 2: Analyze the DNA methylation data of patients diagnosed with hepatocellular carcinoma and non-hepatocellular carcinoma data to obtain methylation information data corresponding to each sequencing site that meets the requirements; Step 3: Using the methylation information data from the previous step, construct, optimize, and test the fusion model; Step three includes the following steps: 301) Classify the methylation information data obtained in step two into data of patients diagnosed with hepatocellular carcinoma and data of non-hepatocellular carcinoma, and add them to the label column respectively, labeled as hepatocellular carcinoma and non-hepatocellular carcinoma; 302) Divide the categorized data into training and test sets according to a ratio of 70% and 30% respectively; 303) Construct a fusion model by inputting the training set data and labels to train and optimize the fusion model to obtain the final model; 304) Input the test set into the final model and use the classification results to evaluate the performance of the final model; Among them, the fusion model described in step 303) is divided into two layers: the first layer consists of two models, and the second layer consists of a multinomial logistic regression model. The first layer includes a hepatocellular carcinoma and normal model and a hepatocellular carcinoma and liver disease model. Both models in the first layer are trained using the GDBT algorithm, and the specific process is as follows: 1-1) First, initialize the weak classifier: In formula (1), N is the total number of training set samples, y i Let x be the label of the training set sample, and let x be the feature data of the training set. 1-2) Calculate the residual r of m trees using formula (2). m,i The iterations are performed sequentially, where m = 1, 2, ..., M, M is the maximum number of iterations; i = 1, 2, ..., N. 1-3) For the leaf node region R of the m-th regression tree mj Where j = 1, 2, ..., J m J m Let J be the number of tree nodes in the m-th regression tree. m For each leaf node region, the optimal residual fitting value C is calculated using formula (3). mj : 1-4) Update the classifier until a final strong classifier F is obtained. M (x), the calculation formulas in the process are shown in formulas (4) and (5): The second-level multinomial logistic regression model is constructed as follows: The two models in the first layer are fused using logistic regression in a stacked manner, and the result of the fused model is used as a prediction of hepatocellular carcinoma and non-hepatocellular carcinoma in the samples; the process of building the second layer model is implemented using R and / or Python languages; The second layer of the model is trained using a model constructed using the logistic regression algorithm. Specifically, the logistic regression algorithm is as follows: 2-1) First, construct a prediction function: 2-2) Furthermore, to avoid overfitting, L2 regularization is added to the logistic regression algorithm. The specific process can be described as: taking the partial derivative of the loss function J and iterating until the loss value reaches its minimum and the model converges. Specifically, as shown in formulas (7) and (8): The specific construction method of the second layer of the model can be described as follows: the outputs of the two models in the first layer of the model are used as features and input in the form of a matrix. The model is trained using the logistic regression algorithm according to steps 2-1) to 2-2) above to obtain the final model. The parameters of the second layer model are also optimized using grid search combined with cross-validation. After optimization, the second layer is trained and the final model is formed. The effect evaluation in step 304) is as follows: The optimized final model is input using the test set obtained from step three. The final model is evaluated using the classification results. The feature classification effect diagram used by the final model is plotted based on the classification results. The sensitivity, specificity, and accuracy of the final model are calculated. The ROC curve and AUC value of the final model are plotted. The final model is compared with the detection results of existing detection methods AFP and PIVKA-II.

2. The construction method according to claim 1, characterized in that, The non-hepatocellular carcinoma individuals mentioned in Step 1 include DNA methylation data from normal individuals, individuals with cirrhosis, and individuals with hepatitis; Step one includes the following steps: 101) Peripheral blood was drawn from the subject, centrifuged, plasma was separated, and cell-free DNA was extracted from the plasma; the subject was either a patient diagnosed with hepatocellular carcinoma or a non-hepatocellular carcinoma patient; 102) Perform DNA methylation sequencing on cell-free DNA in plasma to construct a methylation sequencing library; 103) The methylation sequencing library was hybridized and enriched with selected hepatocellular carcinoma biomarker probes to obtain the final library for sequencing and analysis of the subject's DNA methylation data.

3. The construction method according to claim 1, characterized in that, Step two includes the following steps: 201) Use data processing software to perform quality control, alignment, sorting, and deduplication of the DNA methylation data described in step one; 202) Set appropriate conditions to filter out sequencing sequences that do not meet the requirements, and obtain methylation information data corresponding to each sequencing sequence site that meets the requirements.

4. A hepatocellular carcinoma screening device, characterized in that, The hepatocellular carcinoma screening device includes a hepatocellular carcinoma screening model constructed by the method of any one of claims 1-3.

Images