A gene expression feature selection method

Abstract

The present invention provides a gene expression feature selection method, which relates to the fields of bioinformatics and machine learning. The method uses whether the joint mutual information between an original gene expression dataset and a gene expression dataset corresponding to a candidate feature subset and a category vector is equal as a judgment condition for maximum correlation, uses the maximum conditional mutual information formula as an objective function, iteratively selects gene features that meet the objective function to obtain a candidate feature subset with maximum correlation, and then deletes redundant features in the candidate feature subset to screen out a minimum feature subset with both maximum correlation and minimum redundancy.

Description

Technical Field

[0001] The present invention relates to the technical fields of bioinformatics and machine learning, and in particular to a gene expression feature selection method based on maximum conditional mutual information. Background Art

[0002] In recent years, cancer has become one of the most important factors affecting human health. The occurrence and development of cancer are closely related to genes. Using gene expression data for early diagnosis of cancer is of great significance for the discovery and treatment of cancer.

[0003] Gene expression data typically contains thousands or even tens of thousands of features, most of which exhibit severe multicollinearity, hindering the performance and accuracy of machine learning algorithms. This phenomenon is also known as the "curse of dimensionality." Feature selection is an effective means of addressing data dimensionality reduction. However, existing feature selection methods lack accurate correlation and redundancy evaluation formulas, making it difficult to ensure that the selected feature subset has both maximum correlation and minimum redundancy with the target classification. Furthermore, for high-dimensional datasets, existing search methods, such as exhaustive search, complete search, heuristic search, and random search, suffer from high computational complexity and unsatisfactory search results.

[0004] Therefore, a gene expression feature selection method is urgently needed to select the minimum feature subset that has the maximum correlation and minimum redundancy characteristics with the target classification and has the least number of gene features. Summary of the Invention

[0005] The purpose of the present invention is to provide a gene expression feature selection method to select a minimum gene feature subset that has maximum relevance and minimum redundancy with the target classification.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A gene expression feature selection method, comprising:

[0008] Calculate the first joint mutual information between the original gene expression dataset and the tumor category vector;

[0009] Selecting a candidate feature subset such that a second joint mutual information between the gene expression dataset corresponding to the candidate feature subset and the category vector is equal to the first joint mutual information, and the candidate feature subset is a minimum feature subset of the original gene feature set corresponding to the original gene expression dataset;

[0010] Determine whether each gene feature in the candidate feature subset is a redundant feature according to a redundant feature judgment formula;

[0011] All redundant gene features are removed from the candidate feature subset to obtain the optimal feature subset.

[0012] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0013] The present invention provides a gene expression feature selection method, which uses the equality of the joint mutual information between the original gene expression dataset and the gene expression dataset corresponding to the candidate feature subset and the category vector as the judgment condition for maximum correlation, takes the maximum conditional mutual information formula as the objective function, iteratively selects gene features that meet the objective function to obtain the candidate feature subset with maximum correlation, and then deletes redundant features in the candidate feature subset, thereby screening out the minimum feature subset with both maximum correlation and minimum redundancy. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The following is a brief introduction to the drawings required for use in the embodiments.

[0015] Figure 1 A flow chart of a gene expression feature selection method provided in Example 1 of the present invention;

[0016] Figure 2 A comparison chart of the number of feature selections corresponding to the method of the present invention and the prior art in Example 1 of the present invention;

[0017] Figure 3 1 is a comparison chart of the detection accuracy of the method of the present invention and the prior art respectively in Example 1 of the present invention;

[0018] Figure 4 1 is a comparison chart of the running time of the method of the present invention and the prior art in Example 1 of the present invention. DETAILED DESCRIPTION

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] Example 1:

[0021] To select the smallest gene signature subset with the greatest correlation and least redundancy for tumor classification based on gene expression datasets, see Figure 1 This embodiment provides a gene expression feature selection method, including:

[0022] S1: Determine the original gene expression dataset D F Each gene feature f i (1≤i≤n,i∈N) corresponding gene expression vector Is it a discrete value? If not, use a supervised discretization algorithm to discretize the gene expression vector;

[0023] S2: Calculate the first joint mutual information between the original gene expression dataset and the tumor category vector C. The expression of the first joint mutual information is:

[0024]

[0025] Among them, MI(D F ; C) represents the first joint mutual information; D F is an m×n matrix, representing the original gene expression dataset, m is the number of samples (patients), n is the number of gene features, Represents gene feature f i The gene expression vector, Indicates the gene feature f corresponding to the kth (1≤k≤m,k∈N) sample i The gene expression level; F = {f1,f2,…,f n}, indicating the difference between the original gene expression dataset D F The corresponding original gene feature set; C = (c1, c2, ..., c m ) T , represents the tumor category of m samples, that is, the category vector, such as benign tumor, malignant tumor grade 1, grade 2, grade 3, etc.; p() represents the probability density function.

[0026] S3: Select a candidate feature subset so that the second joint mutual information between the gene expression dataset corresponding to the candidate feature subset and the category vector is equal to the first joint mutual information, and the candidate feature subset is the minimum feature subset of the original gene feature set corresponding to the original gene expression dataset.

[0027] To quickly and accurately search for a subset of candidate features that meets the above conditions, S3 specifically includes the following iterative feature selection process:

[0028] (1) setting the initial value of the candidate feature subset to an empty set;

[0029] (2) Selecting a gene feature with maximum conditional mutual information from the original gene feature set minus the current candidate feature subset according to the maximum conditional mutual information formula. The maximum conditional mutual information formula is expressed as:

[0030]

[0031] Among them, f max represents the gene signature with the maximum conditional mutual information; Represents the current candidate feature subset F S Corresponding gene expression dataset; F S ={f s1 ,f s2 ,…,f sn}, represents the current candidate feature subset; f i It is the original gene feature set F minus the current candidate feature subset F S Gene features in the subsequent collection; Represents gene feature f i The corresponding gene expression level;

[0032] (3) placing the gene feature with the maximum conditional mutual information into the current candidate feature subset;

[0033] (4) Determine whether the second joint mutual information between the gene expression dataset corresponding to the current candidate feature subset and the category vector is equal to the first joint mutual information. The determination formula is:

[0034]

[0035] Among them, MI(D F ; C) represents the first joint mutual information; represents the second joint mutual information; Represents the current candidate feature subset F S Corresponding gene expression dataset; F S represents the current candidate feature subset; F S is a subset of F; C is a category vector;

[0036] (5) If yes, the current candidate feature subset is used as the final candidate feature subset;

[0037] (6) If not, return to step (2).

[0038] The maximum conditional mutual information formula ensures that F S The mutual information between the two classes C grows fastest, thus ensuring that F S Contains the least number of genetic features.

[0039] When the second joint mutual information is equal to the first joint mutual information, the candidate feature subset has the greatest correlation with the class vector. However, the maximum correlation criterion does not guarantee that there are no redundant features in the candidate feature subset. Redundant features are usually defined as features that are irrelevant to the class vector under the conditions of other features in the feature set. This also includes:

[0040] S4: According to the redundant feature judgment formula, each gene feature in the candidate feature subset is judged one by one in the reverse order of the gene feature selection order to determine whether it is a redundant feature. The expression of the redundant feature judgment formula is:

[0041]

[0042] in, Represents the candidate feature subset F S Corresponding gene expression dataset; It's F s The gene feature currently being judged in si The corresponding expression vector, when the above expression is established, the gene feature f si It is judged to be a redundant feature;

[0043] S5: Eliminate all redundant gene features from the candidate feature subset to obtain the optimal feature subset.

[0044] The method described in this example is primarily used for dimensionality reduction and feature selection in high-dimensional gene expression datasets. It can select a minimal subset of gene features with the greatest correlation and minimal redundancy with the target classification. Compared to existing feature selection methods based on correlation and redundancy analysis, the method in this example has lower computational complexity, a more accurate objective function, better dimensionality reduction, and higher classification accuracy for the feature subset.

[0045] To verify the feasibility of the above method, this example uses the Brca_2016_Expression dataset, a gene expression dataset for breast cancer patients, as input. This dataset contains 1,980 breast cancer samples, each containing 24,368 gene expression values, all of which are continuous variables. Furthermore, this dataset includes five cancer categories: LuminalA (718 samples), Luminal B (488 samples), HER2 (240 samples), Basal-like (329 samples), and Normal-like (199 samples). This dataset is a typical high-dimensional, small-sample dataset and is publicly available on the website cbioportal.org.

[0046] The 8 existing feature selection algorithms (Inforgain, GainRatio, SU, ChiSquare, Fisher, CFS, FCBF and mRMR) and the method of this embodiment are used to select features of the dataset Brca_2016_Expression. The experimental results are as follows: Figure 2 、 3 and 4. Figure 2The feature selection results of 8 feature selection methods including the embodiment method are shown. Among them, the number of gene features obtained by the embodiment method is the least, which fully demonstrates that the embodiment method has the best effect on data dimensionality reduction. Figure 3 The detection accuracy of the feature subsets obtained by different feature selection methods for three different machine learning models (Naive Bayes, C4.5 and RandomForest) is shown. The results show that for the three different machine learning models, the feature subsets obtained by the method of this embodiment all achieved the best detection accuracy. Figure 4 The execution time of various feature selection methods, including the method of the present invention, is shown. The execution time of the method of this embodiment is slightly higher than that of individual evaluation methods such as InfoGain, Fisher, and SU, but significantly lower than that of subset evaluation methods such as FCBF, CFS, and mRMR. This is primarily due to the fact that the method of this embodiment eliminates conditionally irrelevant features during each iteration, thereby improving the algorithm's operational efficiency.

[0047] Based on the above experimental results, it is fully proved that the method of this embodiment is superior to the existing feature selection method in terms of dimensionality reduction effect and classification accuracy. In addition, in terms of operating efficiency, the method of this embodiment is significantly better than the existing subset evaluation method and is close to the individual evaluation method. Therefore, the method of this embodiment has good operating efficiency.

Claims

1. A method for selecting gene expression feature, characterized in that: include: Calculate the first joint mutual information between the original gene expression dataset and the tumor category vector; Selecting a candidate feature subset such that a second joint mutual information between the gene expression dataset corresponding to the candidate feature subset and the category vector is equal to the first joint mutual information, and the candidate feature subset is a minimum feature subset of the original gene feature set corresponding to the original gene expression dataset; According to the redundant feature judgment formula, each gene feature in the candidate feature subset is judged one by one in the reverse order of the gene feature selection order to determine whether it is a redundant feature; Eliminating all redundant gene features from the candidate feature subset to obtain the optimal feature subset; The selection process of the candidate feature subset specifically includes: Setting the initial value of the candidate feature subset to an empty set; Selecting a gene feature with maximum conditional mutual information from the original gene feature set minus the current candidate feature subset according to the maximum conditional mutual information formula; Putting the gene feature with the maximum conditional mutual information into the current candidate feature subset; Determining whether a second joint mutual information between the gene expression dataset corresponding to the current candidate feature subset and the category vector is equal to the first joint mutual information; If yes, the current candidate feature subset is used as the final candidate feature subset; If not, return to the step of "selecting the gene feature with the maximum conditional mutual information from the set obtained by subtracting the current candidate feature subset from the original gene feature set according to the maximum conditional mutual information formula"; The judgment formula of the redundant feature is: in, Represents the candidate feature subset F S Corresponding gene expression dataset; It's F s The gene feature currently being judged in si The corresponding expression vector; C represents the category vector.

2. The gene expression feature selection method according to claim 1, wherein Before calculating the first joint mutual information between the original gene expression dataset and the tumor category vector, the method further includes: discretizing the gene expression vectors of non-discrete values ​​in the original gene expression dataset; and using the discretized original gene expression dataset as a new original gene expression dataset.

3. The gene expression feature selection method according to claim 1, wherein The expression of the maximum conditional mutual information formula is: Among them, f max represents the gene feature with the maximum conditional mutual information; F = {f1,f2,…,f n }, indicating the difference between the original gene expression dataset D F The corresponding original gene feature set; F S ={f s1 ,f s2 ,…,f sn }, represents the candidate feature subset; Represents the candidate feature subset F S Corresponding gene expression dataset; f i is the original gene feature set F minus the candidate feature subset F S Gene features in the subsequent collection; Represents gene feature f i The corresponding gene expression level; C represents the category vector.

Images