Method for improving multi-cancer detection accuracy based on single-stranded sequencing data
Through a deep neural network model based on single-strand sequencing data, identifying DNA differential sites, combining DNA methylation and semi-methylation characteristics, the problem of insufficient accuracy of multiple cancer detection in the prior art is solved, and higher detection is achieved Accuracy and a wider range of applications.
Patent Information
- Application Number
- CN202410842494.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-27
- Publication Date
- 2024-11-19
AI Technical Summary
In the prior art, methylation detection based on double-stranded DNA is difficult to effectively utilize the semi-methylation phenomenon, resulting in insufficient accuracy of multiple cancer detection.
Using a deep neural network model based on single-strand sequencing data, we identify DNA differential methylation sites (DMR) and DNA differential hemimethylation sites (DHMR). We train multiple models and perform secondary calibration to combine DNA A Cytogenic and semimethylation characteristics were performed for cancer detection.
It significantly improves the accuracy of multi-cancer detection, which is better than the method of using single methylated or semimethylated sites, improves cancer prediction accuracy, and reduces the impact of population heterogeneity.
Smart Images

Figure CN118983004A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data, and belongs to the technical field of inspection and detection. Background Art
[0002] Extracellular DNA (cfDNA) in the blood contains DNA fragments released into the blood after cell apoptosis, including a large number of DNA fragments from normal blood cells and a small number of DNA fragments released into the blood circulation by cancer cells. The methylation information on DNA is tissue-specific, so it can be used as a feature for multi-cancer detection. However, there is not only complete methylation on DNA, but also hemimethylation. Hemimethylation is often caused by the failure of complete methylation or active demethylation during the occurrence of cancer, so it is also an important feature of cancer detection. In the existing technology, the abundance signal of DNA methylation is mainly detected by building a library of double-stranded DNA.
[0003] Therefore, a method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data is needed to solve the above problems. Summary of the invention
[0004] Purpose of the invention: In view of the problems existing in the prior art, the present invention provides a method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data.
[0005] A method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data, comprising the following steps:
[0006] Step 1: Use blood samples to obtain a training set, which contains blood from multiple types of cancer patients and normal people, and perform single-stranded DNA methylation sequencing on the blood samples;
[0007] Step 2: Use the training set to identify DNA differential methylation sites (DMRs) and DNA differential hemimethylation sites (DHMRs) specific to each type of cancer;
[0008] Step 3: Divide the training set into N equal parts to obtain N groups of data for training N models;
[0009] Step 4: Use the N groups of data to train the DMR model and the DHMR model using a deep neural network DNN model to obtain N DMR models and N DHMR models;
[0010] Step 5: Fuse the N DMR sub-models and the N DHMR sub-models, perform secondary calibration, obtain a unified model, and use the unified model to detect the blood sample.
[0011] Furthermore, in step 3, each of the N groups of data contains (N-1) / N of the training set. In order to reduce the impact of population heterogeneity on the model, the training set is divided into N equal parts for training N models, and each group of data contains (N-1) / N of the training set samples.
[0012] Furthermore, in step 2, the training set is used to identify DNA differential methylation site DMRs specific to various types of cancer, including the following steps: first, the abundance of DNA methylation at each site is calculated; second, the negative binomial distribution is used to test the DNA differential methylation site DMR specific to each group of samples in the training set; finally, the test is repeated for each group of samples to obtain all specific DNA differential methylation site DMRs in the training set.
[0013] Furthermore, in step 2, the training set is used to identify the DNA differential hemimethylation sites DHMR specific to various types of cancer, including the following steps: first, the DNA hemimethylation level of each site is calculated; second, the T test is used to identify the DNA differential hemimethylation site DHMR specific to each group of samples in the training set; finally, the test is repeated for each group of samples to obtain all specific DNA differential hemimethylation sites DHMR in the training set.
[0014] Furthermore, the DNA hemimethylation level of each site was calculated using the following formula:
[0015]
[0016] Among them, Watson represents the value of DNA methylation level on the Watson chain, and Crick represents the value of DNA methylation level on the Crick chain. The asymmetry of methylation levels on the Watson chain and the Crick chain was evaluated.
[0017] Furthermore, the DMR model in step 4 is a pair of other three-classification models, including a first input layer, a first output layer of three nodes, a first hidden layer including 64 nodes, a second hidden layer including 32 nodes and a third hidden layer including 16 nodes, the first hidden layer, the second hidden layer and the third hidden layer all use linear activation functions, and the first output layer is a softmax multi-classification function.
[0018] Furthermore, the DHMR model in step 4 is a pair of other three-classification models, including a second input layer, a three-node second output layer, a fourth hidden layer including 64 nodes, a fifth hidden layer including 32 nodes and a sixth hidden layer including 16 nodes, the fourth hidden layer, the fifth hidden layer and the sixth hidden layer all use linear activation functions, and the second output layer is a softmax multi-classification function.
[0019] Furthermore, in step 5, the N DMR sub-models and the N DHMR sub-models are fused and recalibrated to obtain a unified model, which includes the following steps:
[0020] Step 5.1: Use N DMR sub-models and N DHMR sub-models to predict the samples in the training set respectively, and obtain N DMR prediction results and N DHMR prediction results;
[0021] Step 5.2: Combine the N DMR prediction results and the N DHMR prediction results as the features of the quadratic calibration model, input the features of the 2N quadratic calibration models into the general linear model GLMnet regularized by elastic net, and train to obtain a unified model.
[0022] Principle of the invention: The present invention is based on sequencing data of single-stranded DNA, which can not only detect the abundance of methylation, but also the degree of hemimethylation of DNA, thereby increasing the dimension of the data. Deep learning models are good at efficiently processing data of complex structural signals, good at discovering hidden relationships and patterns, and are suitable for modeling complex biological signals; and are particularly good at processing large sample data, so the predictive classification and prediction accuracy of the model can be effectively expanded by increasing the sample size. The present invention can greatly improve the detection accuracy of multiple cancers by identifying cancer-specific DNA methylation sites and DNA hemimethylation sites, combined with machine learning methods, and its accuracy is higher than that of methods that use only methylation sites or hemimethylation sites.
[0023] Beneficial effects: The method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data of the present invention simultaneously exploits two features, DNA methylation and DNA hemimethylation. DNA hemimethylation is an important phenomenon in the process of cancer occurrence. The present invention uses this feature for the first time to improve the accuracy of cancer detection. The integrated model of DNA methylation and hemimethylation features proposed by the present invention significantly improves the accuracy of cancer prediction, which is better than the prediction model of any single feature. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 A flow chart of a method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data of the present invention;
[0025] Figure 2 Comparison of the detection performance of the present invention in the normal group of patients, including the final model, the DMR model and the DHMR model;
[0026] Figure 3 Comparison of the detection performance of the present invention in patients with liver cancer, including the final model, the DMR model, and the DHMR model;
[0027] Figure 4 The detection performance comparison of the present invention in brain cancer patients includes the final model, the DMR model and the DHMR model. DETAILED DESCRIPTION
[0028] The preferred embodiments of the present invention will be described below in conjunction with the accompanying drawings to more clearly and completely illustrate the technical solutions of the present invention.
[0029] See also Figure 1 As shown, the purpose of the present invention is to simultaneously utilize the characteristics of DNA methylation and DNA hemimethylation, and combine with a deep learning model to develop a method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data, including the following steps:
[0030] Step 1: Divide the blood samples into training set and validation set. The training set should contain blood from multiple types of cancer patients and normal people. Perform single-stranded DNA methylation sequencing on the samples.
[0031] Step 2: Use the training set samples to identify the DNA differential methylation sites DMR and DNA differential hemimethylation sites DHMR specific to each type of cancer:
[0032] Step 2.1: DMR can be identified using existing methods. First, the abundance of DNA methylation at each site is calculated, and then a group of patients is tested for specific significant sites using negative binomial distribution. Repeat the test for each group of patients to obtain all DMR sites of each group of cancer patients;
[0033] Step 2.2: Identify DHMR. First, calculate the DNA hemimethylation level of each site, using the abundance information of single-strand sequencing, hemimethylation level = log2 (Watson / Crick), that is, evaluate the asymmetry of methylation levels on the Watson chain and the Crick chain. Then use the T test to identify the significant sites specific to each group of patients, repeat this step for each group of patients, and obtain all DHMR sites of each group of cancer patients.
[0034] Step 3: In order to reduce the impact of population heterogeneity on the model, the training set is divided into N equal parts for training N models. Each set of data contains (N-1) / N of the training set samples.
[0035] Step 4: For each of the N equal parts of data, use the deep neural network DNN model to train the DMR model and the DHMR model respectively:
[0036] Step 4.1: Using the DMR sites identified in step 2 as features, the signals of N training set samples at the DMR sites, i.e., methylation abundance, are input into the deep neural network model to train and obtain N DMR models;
[0037] Step 4.2: Using the DHMR sites identified in step 2 as features, the signals of N training set samples at the DHMR sites, i.e., the hemimethylation levels, are input into the deep neural network model to train and obtain N DMR models;
[0038] Step 4.3: Each DMR or DHMR model is a “one-vs-other” multi-cancer classification model.
[0039] Step 5: Fuse the DMR sub-model and the DHMR sub-model and perform secondary calibration to obtain the final unified model:
[0040] Step 5.1: Use the N DMR models and N DHMR models trained in step 4 to predict all training set samples respectively;
[0041] Step 5.2: Each sample in the training set is given N DMR prediction probabilities and N DHMR prediction probabilities, and the 2N prediction results are combined as features of the secondary calibration model. The 2N features of all samples are input into the general linear model (GLMnet) regularized with elastic nets, and the final unified model is trained;
[0042] Step 5.3: Evaluate the performance of the trained model using ROC curve.
[0043] Step 6: Evaluate the performance of the model in the validation set. The results show that the performance of the unified model is better than that of the DMR single model or the DHMR single model, indicating that the DHMR feature is an important indicator of the cancer process.
[0044] Step 6.1: Evaluate the performance of the DMR model on the validation set.
[0045] Step 6.2: Evaluate the performance of the DHMR model on the validation set.
[0046] Step 6.3: Evaluate the performance of the final unified model on the validation set.
[0047] The present invention is a method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data, which simultaneously mines two features, DNA methylation and DNA hemimethylation. The main advantages are:
[0048] 1. The cfDNA single-strand sequencing method based on the novel method is highly sensitive, requiring only about 200-300 μl of blood to obtain data with an accuracy comparable to that obtained with about 80 ml of blood using the existing method;
[0049] 2. DNA hemimethylation is an important phenomenon in the process of cancer occurrence. This invention is the first to use this feature for cancer detection;
[0050] 3. The integrated model of DNA methylation and hemimethylation features proposed in this invention significantly improves the accuracy of cancer prediction, which is better than the prediction model of any single feature;
[0051] 4. The present invention discloses a multi-cancer detection model based on deep learning, which has a wider range of applications compared to the prediction model of a single cancer. Deep learning models are good at processing large samples and big data with complex signal structures, so they have greater potential to expand the types of cancer detection.
[0052] Example 1
[0053] Step 1: Collect blood samples from 221 patients, including normal group, liver cancer group and brain cancer group. Divide them into a training set of 175 samples and a validation set of 46 samples. Perform single-stranded DNA methylation sequencing on the samples.
[0054] Step 2: Use the training set samples to identify specific DMR sites and DHMR sites in each group of samples.
[0055] Step 2.1: For the identification of DMR sites, the threshold p<0.01 and the methylation Log fold change >1 were required. A total of 100 DMR sites were identified.
[0056] Step 2.1: For the identification of DHMR sites, the threshold p<0.01 was required, that is, the difference in hemimethylation was significant. A total of 30,784 DHMR sites were identified.
[0057] Step 3: Divide the training set into 10 equal parts for training 10 models. Each group of data contains 90% of the training set samples.
[0058] Step 4: For each of the 10 equal parts of data, use the deep neural network DNN model to train the DMR model and the DHMR model respectively, see Figure 1 .
[0059] Step 4.1: Using the DMR sites identified in step 2 as features, the signals of the 10 training set samples at the DMR sites are trained to obtain 10 DMR models.
[0060] Step 4.2: Using the DHMR sites identified in step 2 as features, the signals of the 10 training set samples at the DHMR sites are trained to obtain 10 DMR models.
[0061] Step 4.3: Each DMR or DHMR model is a "one-vs-other" three-class model. It includes an input layer, a three-node output layer, and three hidden layers with 64 nodes, 32 nodes, and 16 nodes respectively. The hidden layer uses a linear activation function, and the output layer is a "softmax" multi-classification function. The L2 regularization coefficient is 0.1.
[0062] Step 5: Fuse the DMR sub-model and the DHMR sub-model, perform secondary calibration, and obtain the final unified model.
[0063] Step 5.1: Use the 10 DMR models and 10 DHMR models trained in step 4 to predict all training set samples.
[0064] Step 5.2: For each sample in the training set, 10 DMR prediction probabilities and 10 DHMR prediction probabilities were obtained, and the 20 prediction results were combined as features of the secondary calibration model. The 20 features of all samples were used to input the general linear model (GLMnet) regularized with elastic net. The α and λ coefficients in the GLMnet model were adjusted in the interval of 0-1 at intervals of 0.1, and the parameters with the best training effect were used as the final model parameters.
[0065] Step 6: Evaluate the performance of the model in 46 validation sets. For each sample group, we found that the AUC of the final model was greater than that of the DMR model and the DHMR model, see Figure 2 , Figure 3 and Figure 4 .
[0066] Step 6.1: Evaluate the performance of the DMR model in the validation set. The AUCs for the normal group, liver cancer group, and brain cancer group were 0.88, 0.90, and 0.94, respectively.
[0067] Step 6.2: Evaluate the performance of the DHMR model in the validation set. The AUCs for the normal group, liver cancer group, and brain cancer group were 0.91, 0.98, and 0.88, respectively.
[0068] Step 6.3: Evaluate the performance of the final unified model in the validation set. The AUCs for the normal group, liver cancer group, and brain cancer group were 0.93, 0.97, and 0.96, respectively.
Claims
1. A method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data, characterized in that: The following steps are involved: Step 1: Use blood samples to obtain a training set, which contains blood from multiple types of cancer patients and normal people, and perform single-stranded DNA methylation sequencing on the blood samples; Step 2: Use the training set to identify DNA differential methylation sites (DMRs) and DNA differential hemimethylation sites (DHMRs) specific to each type of cancer; Step 3: Divide the training set into N equal parts to obtain N groups of data for training N models; Step 4: Use the N groups of data to train the DMR model and the DHMR model using a deep neural network DNN model to obtain N DMR models and N DHMR models; Step 5: Fuse the N DMR sub-models and the N DHMR sub-models, perform secondary calibration, obtain a unified model, and use the unified model to detect the blood sample.
2. The method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data according to claim 1, characterized in that: In step 3, the N groups of data all contain (N-1) / N of the training set.
3. The method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data according to claim 1, characterized in that: In step 2, the training set is used to identify DNA differential methylation site DMRs specific to various types of cancer, including the following steps: first, the abundance of DNA methylation at each site is calculated; second, the negative binomial distribution is used to test the DNA differential methylation site DMR specific to each group of samples in the training set; finally, the test is repeated for each group of samples to obtain all specific DNA differential methylation site DMRs in the training set.
4. The method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data according to claim 1, characterized in that: In step 2, the training set is used to identify the DNA differential hemimethylation sites DHMR specific to various types of cancer, including the following steps: first, the DNA hemimethylation level of each site is calculated; second, the T test is used to identify the DNA differential hemimethylation site DHMR specific to each group of samples in the training set; finally, the test is repeated for each group of samples to obtain all specific DNA differential hemimethylation sites DHMR in the training set.
5. The method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data according to claim 4, characterized in that: The DNA hemimethylation level of each site was calculated using the following formula: , Among them, Watson represents the value of DNA methylation level on the Watson chain, and Crick represents the value of DNA methylation level on the Crick chain.
6. The method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data according to claim 1, characterized in that: The DMR model in step 4 is a pair of other three-classification models, including a first input layer, a first output layer of three nodes, a first hidden layer including 64 nodes, a second hidden layer including 32 nodes and a third hidden layer including 16 nodes. The first hidden layer, the second hidden layer and the third hidden layer all use linear activation functions, and the first output layer is a softmax multi-classification function.
7. The method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data according to claim 1, characterized in that: The DHMR model in step 4 is a pair of other three-classification models, including a second input layer, a three-node second output layer, a fourth hidden layer including 64 nodes, a fifth hidden layer including 32 nodes and a sixth hidden layer including 16 nodes. The fourth hidden layer, the fifth hidden layer and the sixth hidden layer all use linear activation functions, and the second output layer is a softmax multi-classification function.
8. The method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data according to claim 1, characterized in that: In step 5, the N DMR sub-models and the N DHMR sub-models are fused and recalibrated to obtain a unified model, which includes the following steps: Step 5.1: Use N DMR sub-models and N DHMR sub-models to predict the samples in the training set respectively, and obtain N DMR prediction results and N DHMR prediction results; Step 5.2: Combine the N DMR prediction results and the N DHMR prediction results as the features of the quadratic calibration model, input the features of the 2N quadratic calibration models into the general linear model GLMnet regularized by elastic net, and train to obtain a unified model.
9. The method for improving the accuracy of multi-cancer detection based on single-stranded sequencing data according to claim 1, characterized in that: In step 1, the blood of multiple types of cancer patients includes blood of liver cancer patients and blood of brain cancer patients.
