Feature selection method based on Pearson correlation and colinearity

Abstract

The invention provides a method based on Pearson's collinearity selection (PCS). The method is used for screening features closely related to a target phenotype from high-dimensional genome data. According to the PCS method, by calculating Pearson's correlation coefficients among the features, redundant features are effectively removed, the multi-collinearity problem is reduced, the quality of feature selection is remarkably improved, and the stability of model prediction is enhanced. The method is particularly suitable for the fields of genotype and phenotype correlation analysis and the like, and can optimize the data processing and analysis process. In addition, the PCS method has high cross-population adaptability. Feature selection and prediction verification are carried out on data sets of multiple populations, it is proved that multiple models can keep consistent good prediction performance among different data sets by using the PCS method, and the excellent generalization ability is shown. The method not only can effectively predict the phenotype of a single population, but also can be successfully applied to prediction tasks of other populations, and the application range of the method in the fields of precision medicine, crop breeding and the like is further widened. Compared with a traditional feature selection method, the PCS method has the advantages that feature redundancy is reduced, meanwhile, features highly related to the target phenotype can be better reserved, the prediction precision of the model is remarkably improved, and more efficient and reliable technical support is provided for genome data analysis.

Description

Technical Field

[0001] The invention belongs to the technical fields of genomics, feature selection and dimensionality reduction, and in particular to the application of a feature selection method based on Pearson correlation coefficient in genotype data analysis. Background Art

[0002] With the continuous development of genomic research, the scale and complexity of genotype data have increased significantly, especially in large-scale genome association studies, where the SNP (single nucleotide polymorphism) data involved are extremely high in dimensionality. The processing of these high-dimensional data brings huge challenges in computing and storage. Traditional feature selection methods, such as principal component analysis (PCA) and linear discriminant analysis (LDA), usually rely on the variability of global data or intra-class differences to screen features, but they may not be able to effectively identify key features related to the target phenotype and are often affected by noise and redundant features. In addition, there is usually a significant correlation between genotype features, which may cause multicollinearity problems, thereby affecting the stability and prediction accuracy of the model. Therefore, how to screen out features that contribute significantly to the phenotype in large-scale genotype data and reduce redundant information is the key to improving model performance. Summary of the invention

[0003] In order to efficiently screen features that are significantly correlated with the target phenotype from complex genomic data, solve the problems of multicollinearity and redundant features that are common in genotype data, and significantly improve the performance and stability of downstream prediction models.

[0004] The present invention provides an innovative feature selection method - Pearson collinearity selection (PCS), the core technology of the present invention includes the following key steps:

[0005] Step S1: Use a fixed-size sliding window to perform feature dimensionality reduction on the original single nucleotide polymorphism (SNP) data.

[0006] Step S2: Use the Pearson correlation coefficient to evaluate the features and target phenotypes, select important features and sort them.

[0007] Step S3: Remove redundant features by calculating the correlation between features and setting a threshold. Prioritize features that are highly correlated with the phenotype.

[0008] Step S4: Sort the features according to the Pearson correlation coefficient between the features and the target phenotype, and select the features with the highest ranking.

[0009] The present invention has the following advantages:

[0010] 1. The PCS method proposed in the present invention can screen features closely related to the target phenotype through the Pearson correlation coefficient, while removing redundant features, thereby effectively reducing the multicollinearity problem and improving the quality of feature selection and the stability of model prediction.

[0011] 2. The PCS method proposed in this paper has strong cross-population adaptability. By performing feature selection and prediction verification on data from different populations, it is proved that the PCS method can maintain good prediction performance across multiple data sets, has strong generalization ability, and is suitable for different populations and phenotype prediction tasks.

[0012] 3. Compared with traditional feature selection methods, the PCS method proposed in the present invention can not only reduce the redundancy between features, but also effectively retain features that are highly correlated with the target phenotype, thereby improving the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute a limitation of the present application. In the drawings:

[0014] Figure 1 The overall flow chart of the PCS method proposed by the present invention;

[0015] Figure 2 This is a flow chart of sliding window dimensionality reduction in the PCS method proposed in the present invention;

[0016] Figure 3 A flowchart for selecting and sorting important features in the PCS method proposed in the present invention;

[0017] Figure 4 The redundancy removal flow chart of the PCS method proposed by the present invention;

[0018] Figure 5 The present invention presents feature heatmaps of different feature selection methods;

[0019] Figure 6 The present invention verifies the influence of the PCS method on the prediction performance of the classical model;

[0020] Figure 7 This invention verifies the influence of the PCS method on the classical model in a cross-population dataset; DETAILED DESCRIPTION

[0021] The feature selection method proposed in the present invention, Pearson collinearity selection (PCS), effectively reduces data redundancy and improves the accuracy of phenotypic prediction by combining the steps of extracting genomic features through sliding windows, calculating the Pearson correlation coefficient, removing redundancy of correlation between features, and sorting.

[0022] The present invention is further described below through specific implementations.

[0023] Step S1: Use a fixed-size sliding window to perform feature dimensionality reduction on the original single nucleotide polymorphism (SNP) data.

[0024] Step S2: Use the Pearson correlation coefficient to evaluate the features and target phenotypes, select important features and sort them.

[0025] Step S3: Remove redundant features by calculating the correlation between features and setting a threshold. Prioritize features that are highly correlated with the phenotype.

[0026] Step S4: Sort the features according to the Pearson correlation coefficient between the features and the target phenotype, and select the features with the highest ranking.

[0027] 2. A method for selecting features based on Pearson correlation and collinearity according to claim 1, characterized in that the specific method in step S2 is as follows:

[0028] S21: Pearson correlation coefficient is used to calculate the linear correlation coefficient between each extracted feature and the target phenotype Its range is [-1, 1]. The closer r is to 0, the smaller the correlation is. The closer it is to 1, the positive correlation is positive, and the closer it is to -1, the negative correlation is negative. i and i are the observed values ​​of the eigenvalues ​​and target phenotypes, respectively. and is the mean of the trait and the target phenotype.

[0029] S22: After calculating the Pearson correlation coefficient value r for each feature and the target phenotype, its significance is further evaluated. To ensure that the selected features have practical predictive significance for the target phenotype, we use hypothesis testing (t-test) to calculate the significance value p-value of each correlation coefficient. In the statistical analysis, the null hypothesis set is "there is no significant correlation between the feature and the target phenotype", and the hypothesis is tested by calculating the corresponding statistic. The calculation formula of the statistic t is Where r is the Pearson correlation coefficient of the sample. n is the size of the sample. Use degrees of freedom df = n-2, and then calculate p-value = 2·(1-CDF(t,df)), where CDF is the cumulative distribution function. Since we usually perform a two-tailed test, we need to multiply by 2.

[0030] S23: Select features that are highly correlated with the target phenotype as candidate features. The specific method for selecting candidate features is: |r| is the absolute value of the Pearson correlation coefficient r between the feature and the target phenotype, is the average value of the absolute value of the linear correlation r between all features and the target feature, and is selected Features with p-value < 0.05 are candidate features. This criterion ensures that the selected features not only have strong relevance but also have statistical significance, ensuring that they have practical significance in phenotypic prediction.

[0031] 3. A method for selecting features based on Pearson correlation and collinearity according to claim 1, characterized in that the specific method in step S3 is as follows:

[0032] S31: Calculate the correlation between all candidate features and construct a feature correlation matrix. Redundant features are identified by calculating the correlation value of each pair of features.

[0033] S32: In order to effectively reduce the redundancy of data, a correlation threshold is set to remove redundant features that are highly correlated with other features. In this way, it is ensured that the information in the feature set is not repeated and the multicollinearity problem caused by the high correlation between features is avoided.

[0034] Embodiment 1

[0035] like Figure 2 As shown in the figure, we first input the original SNP features. SNP data is usually represented in the form of a two-dimensional matrix, where rows correspond to individuals and columns correspond to SNP sites. Due to the high-dimensional nature of genomic data, directly using the original data may lead to excessive computational complexity and storage pressure, so the sliding window method is used to reduce the dimension. Specifically, each time the window slides, a certain length of genomic sequence features are extracted. The window size is set to W and the step size is S, that is, each time the window slides S positions, W-dimensional features are extracted. In order to further reduce redundancy and improve the compactness of features, one or more key features in the window are selected for extraction according to needs.

[0036] Next, if Figure 3 As shown, the Pearson correlation coefficient and hypothesis test (such as t-test) are calculated based on the candidate features and target phenotypes in the previous step, and the features whose absolute value of the correlation coefficient |r| is greater than the average absolute value of r of all features and whose p-value is less than 0.05 are selected.

[0037] Then, if Figure 4 As shown, the correlation between all candidate features is calculated, a correlation matrix is ​​constructed, and redundant features are identified. A correlation threshold (such as 0.7) is set to remove redundant features that are highly correlated with other features and have low correlation. Finally, the remaining features are sorted according to their values ​​with the target phenotype, and features with strong correlation are retained first. Finally, the top ranked features are selected as input for use by downstream models.

[0038] In order to verify the effectiveness of the PCS method proposed in this invention, we did the following work:

[0039] 1) Select the data set and set the model evaluation indicators:

[0040] We used the RiceVarMap population rice SNP dataset from Huazhong Agricultural University, which contains 11 phenotypes. They are grain length (GL), grain thickness (GT), grain weight (GW), grain width (GWi), heading date (HD), grain length-width ratio (GWi), and

[0041] (LWR), effective ear number (NEP), ear number (NP), plant height (PH), spikelet length (SL), and yield (Y). The feature selection method was evaluated by combining the 10-fold cross-validation method with multiple classic model prediction methods. During the 10-fold cross-validation process, the feature subsets in the data set were randomly divided into ten parts, nine of which were used as training data sets and the remaining one as a test set for testing. Each experiment will produce the corresponding accuracy

[0042] (i.e., the correlation value between the predicted value and the observed value). In this experiment, all algorithms will obtain the average of ten results as an estimate of the model prediction accuracy. 2) Correlation analysis of the features extracted by LD and PCS methods:

[0043] like Figure 5 As shown, in order to verify the advantages of the Pearson collinearity selection (PCS) method proposed in the present invention in feature selection, this embodiment performs correlation analysis between features and correlation analysis between features and target phenotypes on the original data, the data processed by the LD method, and the data processed by the PCS method, revealing the differences between different feature selection methods in reducing redundancy and retaining valid feature information. We randomly selected 100 feature subsets for the original data, the data processed by the LD method, and the data processed by the PCS method, and calculated the Pearson correlation matrix between their features and the Pearson correlation matrix between the phenotype and the features, respectively. A heat map is used to visualize the correlation structure of each data set, where the darker the color, the stronger the correlation.

[0044] Figure 5 A, B, and C are heat maps of correlation between features (A: original data set, B: LD-processed data set, and C: PCS-processed data set), and D, E, and F are heat maps of correlation between features and phenotypes.

[0045] (D: original data set, E: LD processed data set, F: PCS processed data set). Figure 5 As can be seen from A, the collinearity between the features of the original data set is quite serious, such as Figure 5 B After LD treatment, the collinearity was significantly reduced. Figure 5E is obviously lighter than the original data graph C, indicating that the relationship between the feature and the phenotype has been weakened. Figure 5 As shown in C and F, it can effectively reduce collinearity and retain important features.

[0046] 3) Use PCA, LD and PCS methods to process data and compare the performance on different models:

[0047] like Figure 6 As shown in Figure 2, we compared the prediction accuracy of three feature extraction methods (principal component analysis [PCA], linkage disequilibrium [LD], and Pearson collinearity selection [PCS]) on seven different models and evaluated the impact of different feature selection methods on prediction accuracy. In this implementation, we selected DeepGS ( Figure 6 A) DNNGP( Figure 6 B) GenNet (Fig.

[0048] 6C) and LightGBM( Figure 6 D) rrBLUP( Figure 6 E) RF Figure 6 F) and SVM

[0049] ( Figure 6 G) Seven classic models. The feature data processed by different feature selection methods are input into each model for training, and the prediction accuracy of different methods on each model is evaluated. Obviously,

[0050] The PCS method has the best prediction accuracy among all models. Figure 6 As shown in FIG. 1 , this is the average effect of the predictions of the seven models, which can comprehensively present the performance of different characteristics. It can be clearly seen that the PCS feature selection method of the present invention has the best effect.

[0051] 4) The impact of using PCS method on classic models in cross-population datasets:

[0052] like Figure 7As shown, in order to verify the generalization ability of the PCS method proposed in the present invention in different populations, this embodiment compares and analyzes the prediction performance of the PCS method in different population data sets. Specifically, we selected the aspect ratio phenotype for cross-population feature selection and prediction. First, we extracted SNP features from the RiceVarMap (RVM) data set, and extracted SNP features at the corresponding positions from our own population data set (called OTHER) to perform cross-population prediction tasks. Subsequently, we merged the two data sets to form a mixed data set (CM), which included SNP features in RVM and OTHER data, and used DeepGS, DNNGP, LightGBM, rrBLUP, RF and SVM classic models for training and prediction. Excellent prediction performance was shown in any data set. Secondly, after training, the OTHER data set and CM data showed better prediction performance in most cases. This result verifies the generalization ability of the PCS method, indicating that the selected SNP features can not only effectively predict phenotypes in a specific population, but also be applicable to other populations.

[0053] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A feature selection method based on Pearson correlation and collinearity, characterized in that: include: Step S1: Use a fixed-size sliding window to perform feature dimensionality reduction on the original single nucleotide polymorphism (SNP) data. Step S2: Use the Pearson correlation coefficient to evaluate the features and target phenotypes, select important features and sort them. Step S3: Remove redundant features by calculating the correlation between features and setting a threshold. Prioritize features that are highly correlated with the phenotype. Step S4: Sort the features according to the Pearson correlation coefficient between the features and the target phenotype, and select the features with the highest ranking.

2. A method for feature selection based on Pearson correlation and collinearity according to claim 1, characterized in that: The specific method in step S2 is as follows: S21: Pearson correlation coefficient is used to calculate the linear correlation coefficient between each extracted feature and the target phenotype Its range is [-1, 1]. The closer r is to 0, the smaller the correlation is. The closer it is to 1, the positive correlation is positive, and the closer it is to -1, the negative correlation is negative. i and i are the observed values ​​of the eigenvalues ​​and target phenotypes, respectively. and is the mean of the trait and target phenotype. S22: After calculating the Pearson correlation coefficient value r for each feature and the target phenotype, its significance is further evaluated. To ensure that the selected features have actual predictive significance for the target phenotype, we use hypothesis testing (t-test) to calculate the significance value p-value of each correlation coefficient. In the statistical analysis, the null hypothesis set is "there is no significant correlation between the feature and the target phenotype", and the hypothesis is tested by calculating the corresponding statistic. The calculation formula of the statistic t is Where r is the Pearson correlation coefficient of the sample. n is the size of the sample. Use degrees of freedom df = n-2, and then calculate p-value = 2·(1-CDF(t,df)), where CDF is the cumulative distribution function. Since we usually perform a two-tailed test, we need to multiply by 2. S23: Select features that are highly correlated with the target phenotype as candidate features. The specific method for selecting candidate features is: |r| is the absolute value of the Pearson correlation coefficient r between the feature and the target phenotype, is the average value of the absolute value of the linear correlation r between all features and the target feature, and is selected Features with p-value < 0.05 are candidate features. This criterion ensures that the selected features not only have strong relevance but also have statistical significance, ensuring that they have practical significance in phenotypic prediction.

3. A method for feature selection based on Pearson correlation and collinearity according to claim 1, characterized in that: The specific method in step S3 is as follows: S31: Calculate the correlation between all candidate features and construct a feature correlation matrix. Redundant features are identified by calculating the correlation value of each pair of features. S32: In order to effectively reduce the redundancy of data, a correlation threshold is set to remove redundant features that are highly correlated with other features. In this way, it is ensured that the information in the feature set is not repeated and the multicollinearity problem caused by the high correlation between features is avoided.

Images