A method for detecting single nucleotide variants in functional noncoding regions based on ensemble learning

Abstract

The present invention proposes a method for detecting single nucleotide variants in functional noncoding regions based on ensemble learning, which primarily addresses the technical problem of low prediction accuracy in the prior art. The method comprises the following steps: (1) obtaining single nucleotide variant data in functional noncoding regions and constructing a training sample set and a test sample set; (2) constructing a single nucleotide variant prediction model for functional noncoding regions; (3) training the single nucleotide variant prediction model for functional noncoding regions; and (4) obtaining detection results for single nucleotide variants in functional noncoding regions. The present invention has a simple experimental process and high detection accuracy, and can be applied to detecting single nucleotide variants in functional noncoding regions.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and relates to a method for detecting single nucleotide variations in functional non-coding regions, and specifically relates to a method for detecting single nucleotide variations in functional non-coding regions based on improved Stacking ensemble learning, which can be applied to detecting single nucleotide variations in functional non-coding regions. Background Art

[0002] Functional noncoding region (SNV) variants (SNRVs) are mutations caused by a single nucleotide in the noncoding regions of the human genome. These SNVs are widely involved in important biological processes, acting as regulatory mechanisms, such as DNA transcription and translation. They often contribute to the development of numerous diseases and medical conditions. Therefore, detecting SNVs in SNRVs can improve our understanding of complex diseases.

[0003] Traditional methods for detecting single nucleotide variants in functional non-coding regions rely on the classification capabilities of a single machine learning algorithm, which has many limitations: for example, classification models trained on small data sets are prone to overfitting; classification models trained with a single learning algorithm, such as an artificial neural network, may fall into local optimal solutions; and some data may not be correctly classified by a single classifier. Ensemble learning, on the other hand, refers to a method for making decisions by combining the learning results of multiple sub-learners through a combination strategy in supervised tasks. Compared to a single learner, ensemble learning often has better predictive performance and stronger generalization and robustness. This is because the prediction results of ensemble learning are generated by multiple sub-learners. The prediction results of strong learners can compensate for the deviations caused by weak learners, and the influence of weak learners on the results can also be reduced by assigning weights.

[0004] For example, Liu et al. published an article titled "Predicting the functional effects of human non-coding variants based on stacking ensemble learning" at the 2022 IEEE International Conference on Bioinformatics and Biomedicine, disclosing a method for predicting functional non-coding region single nucleotide variations based on Stacking ensemble learning. This method uses 5-fold cross-validation to provide training subsets and validation subsets, uses the prediction results of the base learner for ncSNVs as the input of the meta-learner, and predicts functional ncSNVs through the meta-learner. This method improves the detection accuracy, but it has the following defects: 1. This method uses 5-fold cross-validation to process the training set, and does not take into account the sample imbalance problem in the dataset of functional non-coding region single nucleotides, so that the divided dataset cannot guarantee full coverage of the patterns contained in the original training samples, and the randomness and diversity of the divided dataset are poor, which affects the algorithm's ability to accurately identify functional ncSNVs. 2. This method concatenates the predictions of all base learners without processing. This can result in the information captured by base learners with good prediction performance being given the same weight to the meta-learner as that captured by base learners with poor prediction performance. This method fails to account for the differences in predictive performance between base learners. This reduces the model's focus on base learners with strong predictive performance, potentially compromising the model's predictive power. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and propose a method for detecting single nucleotide variations in functional non-coding regions based on ensemble learning to solve the technical problem of low prediction accuracy in the prior art.

[0006] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0007] (1) Construct training sample set and test sample set:

[0008] Obtain N non-coding region single nucleotide variation ncSNV data D ori And each non-coding region single nucleotide variation ncSNV data Tags n , and for each non-coding region single nucleotide variation ncSNV data The annotation features in are normalized, and then the normalized Z non-coding region single nucleotide variation ncSNV data and their corresponding labels form the training sample set D tThe remaining T normalized non-coding region single nucleotide variation ncSNV data and their corresponding labels form the test sample set D te , where N≥1000, T=NZ,

[0009] (2) Construction of functional ncSNV prediction model:

[0010] Constructing a functional ncSNV prediction model O comprising a sequentially connected sampling module, a base learner module, and a meta-learner logistic regression model, wherein the base learner module comprises a random forest model, a decision tree model, and a support vector machine model arranged in parallel;

[0011] (3) Training the functional ncSNV prediction model:

[0012] (3a) Initialize the number of iterations j, the maximum number of iterations is J, the current meta-learner logistic regression model LR j The weight parameter is And let j = 0;

[0013] (3b) The training sample set D t As the input of the functional ncSNV prediction model O, the sampling module uses the Bootstrap sampling method to select D t Perform R replay sampling to generate training subset D train and its corresponding training sub-tag set Tag train , and validation subset D val and its corresponding verification sub-tag set Tag val , where D train_r ={d train_r_h |1≤h≤H}, D val_r ={d train_r_u |1≤u≤U}, U=ZH,1≤r≤R;

[0014] (3c) The three models arranged in parallel in the base learner module train the subset D train After learning each training sample in the validation subset D val Make predictions for each validation sample in the dataset and get the prediction result P v1 、P v2 and P v3 ;

[0015] (3d) By predicting the result P v1 、P v2 and P v3 and its corresponding AUROC values ​​V1, V2 and V3 to calculate P v1 、P v2 and Pv3 The weighted value P v1w 、P v2w and P v3w , and P v1w 、P v2w and P v3w Perform vertical splicing to obtain the meta-learner training set D mt ;

[0016] (3e) Meta-learner Logistic Regression Model LR j The training set D of the meta-learner mt Perform logistic regression to obtain the meta-learner training set D mt The predicted value of

[0017] (3f) adopts the cross entropy loss function L and passes and verify the subtag set Tag val Calculate the model LR of the jth iteration j The cross entropy loss function L j , then adopt the back propagation method and pass L j Calculate LR j The parameter gradient Finally, the gradient descent algorithm is used to LR j The weight parameter Update and get O for this iteration j , where the cross entropy loss function

[0018] (3g) Determine whether j ≥ J is true. If so, obtain the trained functional ncSNV prediction model O * Otherwise, let j = j + 1, O j =0, and execute step (3f);

[0019] (4) Obtaining functional ncSNV detection results:

[0020] (4a) The test sample set D te As a functional ncSNV prediction model * The base learner module predicts each test sample and obtains D te The prediction result P t , through the prediction result P t1 、P t2 and P t3 And the base learner module on the validation subset D val The AUROC values ​​of V1, V2 and V3 are calculated as P v1 、P v2 and P v3 The weighted value Pv1w 、P v2w and P v3w , P v1w 、P v2w and P v3w Add and take the mean to get the meta-learner test set D mte ;

[0021] (4b) Meta-learner Logistic Regression Model LR * Test set D for the meta-learner mte Perform logistic regression to obtain a trained functional ncSNV prediction model O * For the test sample set D te The predicted value of

[0022] Compared with the prior art, the present invention has the following advantages:

[0023] (1) The functional ncSNV prediction model constructed in the present invention includes a sampling module. During the training process of the model, the sampling module uses the Bootstrap sampling method to perform multiple playback sampling on the training sample set, forming new training sets and validation sets that are more random and diverse, and the number of positive sets in the samples is increased, avoiding the defects of the existing technology using K-fold cross-validation and effectively improving the detection accuracy.

[0024] (2) In the process of training the functional ncSNV prediction model and obtaining the test results of the present invention, the three models in the base learner module learn each training sample in the training subset respectively, predict each verification sample in the verification subset, and calculate the weighted value of each prediction result through the three prediction results and their corresponding AUROC values. The three weighted values ​​are then vertically spliced, which enhances the model's attention to learners with strong predictive ability and reduces the influence of weak learners on the model's predictive ability, thereby avoiding the influence of the prior art of directly splicing the three prediction results on the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart for implementing the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0027] Reference Figure 1 , the present invention comprises the following steps:

[0028] Step 1: Obtain single nucleotide variation data of functional non-coding regions and construct training sample sets and test sample sets.

[0029] 1.1) Download the 2019_07 version of the dataset D from the public website ori , including 31,287 single nucleotide variation data of non-coding regions and labels of single nucleotide variation data of each non-coding region;

[0030] 1.2) Normalize the annotation features in each non-coding region single nucleotide variation data in 1.1), and then form the training sample set D with the normalized 20,000 non-coding region single nucleotide variation ncSNV data and their corresponding labels. t The remaining 11287 normalized non-coding region single nucleotide variation ncSNV data and their corresponding labels constitute the test sample set D te , the normalized formula is:

[0031] d n ={s nc |1≤c≤C}

[0032]

[0033] Among them, d n yes The normalized result, s nc 、 They are d n 、 The cth annotation feature of .

[0034] Step 2: Build a functional ncSNV prediction model:

[0035] 2.1) Construct a sampling module including a sampling layer that implements 20 samplings with playback, construct a base learner module including a support vector machine model, a decision tree model, and a random forest model arranged in parallel, and construct a meta-learner logistic regression model including a serially arranged linear combination layer and a sigmoid activation function layer, wherein the support vector machine model is based on the constraint condition w T Mathematical model of d+b≥1 where w r is the hyperplane direction parameter of the support vector machine model, b r is the hyperplane location parameter of the support vector machine model, d o Represents the training data set D train* For the o-th ncSNV sample data, the decision tree model includes a root node layer, three decision node layers, and a leaf node layer arranged in series. The random forest model includes a decision module composed of a sampling module and a decision tree model arranged in series in parallel. The function

[0036] 2.2) Sequentially connect the sampling module, base learner module and meta-learner logistic regression model constructed in 2.1) to obtain a functional ncSNV prediction model O.

[0037] Step 3: Train the functional ncSNV prediction model.

[0038] (3a) Initialize the number of iterations j, the maximum number of iterations is J, the current meta-learner logistic regression model LR j The weight parameter is And let j = 0;

[0039] (3b) The training sample set D t As the input of the functional ncSNV prediction model O, the sampling module t Perform 20 replay samplings to generate the training subset D train and its corresponding training sub-tag set Tag train , and validation subset D val and its corresponding verification sub-tag set Tag val , where D train_r ={d train_r_h |1≤h≤H}, D val_r ={d train_r_u |1≤u≤U}, H>10, U=ZH, 1≤r≤20;

[0040] The above sampling module is used to t Perform 20 replay samplings to generate the training subset D train and its corresponding training sub-tag set Tag train , and validation subset D val and its corresponding verification sub-tag set Tag val , which can make the training subset and validation subset more random and diverse, and increase the number of positive sets in the validation subset, avoiding the defects of the existing technology using K-fold cross-validation and effectively improving the detection accuracy.

[0041] (3c) The three models arranged in parallel in the base learner module train the subset D train After learning each training sample in the validation subset D val Make predictions for each validation sample in the dataset and get the prediction result P v1 、P v2 and P v3 :

[0042] (3c1) Decision tree model for training subset D train Divide and calculate its annotation feature set S train Each annotation feature pair D train The Gini coefficient is divided according to the Gini coefficient.train , and obtain the learned decision tree model DT; the sampling module in the random forest model is trained on the subset D train Sampling is performed to obtain D train_1 , D train_2 ,…,D train_k ,…,D train_10 , using the sampled data D train_1 , D train_2 ,…,D train_k ,…,D train_10 After dividing the decision tree models separately, they are fused through a simple voting method to obtain the learned random forest model RF, where 1≤k≤K; the construction constraint is w r T d z +b r Mathematical model for maximizing separation distance ≥1 The constructed mathematical model is solved using the minimum sequence optimization algorithm to obtain D train_r The training results of SVM;

[0043] (3c2) The learned decision tree model DT, the learned random forest model RF and the learned support vector machine model SVM are used to validate the subset D val Make predictions for each validation sample in the dataset and get the prediction result P v1 、P v2 and P v3 ;

[0044] (3d) By predicting the result P v1 、P v2 and P v3 and its corresponding AUROC values ​​V1, V2 and V3 to calculate P v1 、P v2 and P v3 The weighted value P v1w 、P v2w and P v3w , and P v1w 、P v2w and P v3w Perform vertical splicing to obtain the meta-learner training set D mt :

[0045] (3d1) by predicting the result P v1 、P v2 and P v3 and its corresponding verification sub-tag set Tag val Calculate P v1 、P v2 and P v3 The corresponding AUROC value sets V1, V2 and V3;

[0046] (3d2) The prediction results P are respectively analyzed by AUROC value sets V1, V2 and V3. v1 、P v2 and P v3 The prediction results of each validation subset in the weighted, get P v1w 、P v2w and P v3w , where the prediction results The formula for weighting the prediction results of each validation subset in is:

[0047]

[0048] Among them, * represents cross product,

[0049] The above prediction result P v1 、P v2 and P v3 and its corresponding AUROC values ​​V1, V2 and V3 to calculate P v1 、P v2 and P v3 The weighted value P v1w 、P v2w and P v3w , which enhances the model's attention to learners with strong predictive ability, while reducing the impact of weak learners on the model's predictive ability, and avoids the impact of the existing technology of directly splicing three prediction results on prediction accuracy.

[0050] (3e) Meta-learner Logistic Regression Model LR j The training set D of the meta-learner mt Perform logistic regression to obtain the meta-learner training set D mt The predicted value of

[0051] (3f) adopts the cross entropy loss function L and passes and verify the subtag set Tag val Calculate the model LR of the jth iteration j The cross entropy loss function L j , then adopt the back propagation method and pass L j Calculate LR j The parameter gradient Finally, the gradient descent algorithm is used to LR j The weight parameter Update and get O for this iteration j , where the cross entropy loss function

[0052] (3g) Determine whether j ≥ J is true. If so, obtain the trained functional ncSNV prediction model O * Otherwise, let j = j + 1, O j =O, and execute step (3f).

[0053] Step 4: Obtain the detection results of functional ncSNVs:

[0054] (4a) The test sample set D te As a functional ncSNV prediction model * Input, functional ncSNV prediction model O * It is necessary to make predictions for the complete test sample set, so the sampling module does not sample the test sample set, and the base learner module predicts each test sample to obtain D te The prediction result P t , through the prediction result P t1 、P t2 and P t3 And the base learner module on the validation subset D val The AUROC values ​​of V1, V2 and V3 are calculated as P v1 、P v2 and P v3 The weighted value P v1w 、P v2w and P v3w , P v1w 、P v2w and P v3w Add and take the mean to get the meta-learner test set D mte ;

[0055] (4b) Meta-learner Logistic Regression Model LR * Test set D for the meta-learner mte Perform logistic regression to obtain a trained functional ncSNV prediction model O * For the test sample set D te The predicted value of

Claims

1. A method for detecting single nucleotide variations in functional non-coding regions based on ensemble learning, characterized in that: The steps include: (1) Construct training sample set and test sample set: Obtain N non-coding region single nucleotide variation ncSNV data D ori And each non-coding region single nucleotide variation ncSNV data Tags n , and for each non-coding region single nucleotide variation ncSNV data The annotation features in are normalized, and then the normalized Z non-coding region single nucleotide variation ncSNV data and their corresponding labels form the training sample set D t The remaining T normalized non-coding region single nucleotide variation ncSNV data and their corresponding labels form the test sample set D te , where N≥1000, T=NZ, (2) Construction of functional ncSNV prediction model: Constructing a functional ncSNV prediction model O comprising a sequentially connected sampling module, a base learner module, and a meta-learner logistic regression model, wherein the base learner module comprises a random forest model, a decision tree model, and a support vector machine model arranged in parallel; (3) Training the functional ncSNV prediction model: (3a) Initialize the number of iterations j, the maximum number of iterations is J, the current meta-learner logistic regression model LR j The weight parameter is And let j = 0; (3b) The training sample set D t As the input of the functional ncSNV prediction model O, the sampling module uses the Bootstrap sampling method to select D t Perform R replay sampling to generate training subset D train and its corresponding training sub-tag set Tag train , and validation subset D val and its corresponding verification sub-tag set Tag val , where D train_r ={d train_r_h |1≤h≤H}, D val_r ={d train_r_u |1≤u≤U}, U=ZH,1≤r≤R; (3c) The three models arranged in parallel in the base learner module train the subset D train After learning each training sample in the validation subset D val Make predictions for each validation sample in the dataset and get the prediction result P v1 、P v2 and P v3 ; (3d) By predicting the result P v1 、P v2 and P v3 and its corresponding AUROC values ​​V1, V2 and V3 to calculate P v1 、P v2 and P v3 The weighted value P v1w 、P v2w and P v3w , and P v1w 、P v2w and P v3w Perform vertical splicing to obtain the meta-learner training set D mt ; (3e) Meta-learner Logistic Regression Model LR j The training set D of the meta-learner mt Perform logistic regression to obtain the meta-learner training set D mt The predicted value of (3f) adopts the cross entropy loss function L and passes and verify the subtag set Tag val Calculate the model LR of the jth iteration j The cross entropy loss function L j , then adopt the back propagation method and pass L j Calculate LR j The parameter gradient Finally, the gradient descent algorithm is used to LR j The weight parameter Update and get O for this iteration j , where the cross entropy loss function (3g) Determine whether j ≥ J is true. If so, obtain the trained functional ncSNV prediction model O * Otherwise, let j = j + 1, O j =0, and execute step (3f); (4) Obtaining functional ncSNV detection results: (4a) The test sample set D te As a functional ncSNV prediction model * The base learner module predicts each test sample and obtains D te The prediction result P t , through the prediction result P t1 、P t2 and P t3 And the base learner module on the validation subset D val The AUROC values ​​of V1, V2 and V3 are calculated as P v1 、P v2 and P v3 The weighted value P v1w 、P v2w and P v3w , P v1w 、P v2w and P v3w Add and take the mean to get the meta-learner test set D mte ; (4b) Meta-learner Logistic Regression Model LR * Test set D for the meta-learner mte Perform logistic regression to obtain a trained functional ncSNV prediction model O * For the test sample set D te The predicted value of 2. The method for detecting single nucleotide variations in functional non-coding regions based on stacking ensemble learning according to claim 1, characterized in that: The ncSNV data for each non-coding region described in step (1) Normalize, the normalization formula is: d n ={s nc |1≤c≤C} Among them, d n yes The normalized result, s nc 、 They are d n 、 The cth annotation feature of .

3. The method for detecting single nucleotide variations in functional non-coding regions based on stacking ensemble learning according to claim 1, characterized in that: The sampling module, base learner module and meta-learner logistic regression model described in step (2), wherein: The sampling module includes a sampling layer that implements R times of replay sampling; Support vector machine model: with constraints as w T Mathematical model of d+b≥1 where w r is the hyperplane direction parameter of the support vector machine model, b r is the hyperplane location parameter of the support vector machine model, d o Represents the training data set D train* The oth ncSNV sample data in; Decision tree model: consists of a root node layer, three decision node layers, and one leaf node layer arranged in series; Random forest model: A model that includes 10 parallel-arranged decision modules consisting of sampling modules and a serial arrangement of decision tree models; The meta-learner logistic regression model includes a serially arranged linear combination layer and a sigmoid activation function layer, where the function 4. The method for detecting single nucleotide variations in functional non-coding regions based on stacking ensemble learning according to claim 1, characterized in that: The three models arranged in parallel in the base learner module described in step (3c) are respectively trained on the training subset D train After learning each training sample in the validation subset D val For each validation sample in , the prediction is performed as follows: (3c1) Decision tree model for training subset D train Divide and calculate its annotation feature set S train Each annotation feature pair D train The Gini coefficient is divided according to the Gini coefficient. train , and obtain the learned decision tree model DT; the sampling module in the random forest model is trained on the subset D train Sampling is performed to obtain D train_1 , D train_2 ,…,D train_k ,…,D train_10 , using the sampled data D train_1 , D train_2 ,…,D train_k ,…,D train_10 After dividing the decision tree models separately, they are fused through a simple voting method to obtain the learned random forest model RF, where 1≤k≤K; the construction constraint is w r T d z +b r Mathematical model for maximizing separation distance ≥1 The constructed mathematical model is solved using the minimum sequence optimization algorithm to obtain D train_r The training results of SVM; (3c2) The learned decision tree model DT, the learned random forest model RF and the learned support vector machine model SVM are used to validate the subset D val Make predictions for each validation sample in the dataset and get the prediction result P v1 、P v2 and P v3 .

5. The method for detecting single nucleotide variations in functional non-coding regions based on stacking ensemble learning according to claim 1, characterized in that: The prediction result P in step (3d) v1 、P v2 and P v3 and its corresponding AUROC values ​​V1, V2 and V3 to calculate P v1 、P v2 and P v3 The weighted value P v1w 、P v2w and P v3w , the implementation steps are: (3d1) by predicting the result P v1 、P v2 and P v3 and its corresponding verification sub-tag set Tag val Calculate P v1 、P v2 and P v3 The corresponding AUROC value sets V1, V2 and V3; (3d2) The prediction results P are respectively analyzed by AUROC value sets V1, V2 and V3. v1 、P v2 and P v3 The prediction results of each validation subset in the weighted, get P v1w 、P v2w and P v3w , where the prediction results The formula for weighting the prediction results of each validation subset in is: Among them, * represents cross product,

Images