CN107301323B - A method for constructing a classification model related to psoriasis - Google Patents

Abstract

The invention relates to the technical field of medical detection, in particular to a method for constructing a classification model related to psoriasis, which comprises the following steps: (1) selecting psoriasis susceptible sites; (2) converting the susceptible loci into input data according to different types of susceptible loci; (3) and classifying the data by using an Adaboost-SVM model. At present, relevant technologies are lacked to classify and predict psoriasis data, and only the existence of judgment sites is remained to infer the diseased situation. The invention utilizes the effective machine learning classifier SVM to classify, integrates the SVM by the adaboost frame, and improves the accuracy of the classifier. The model can integrate SNP, amino acid and type data for classification, comprehensively considers the information of each dimension, and improves the accuracy of the classification result.

Description

A method for constructing a classification model related to psoriasis

technical field

The invention relates to the technical field of medical detection, in particular to a method for constructing a classification model related to psoriasis.

Background technique

Psoriasis, also known as psoriasis, is a common and complex disease. It has been reported that the occurrence of psoriasis is related to genetic factors, especially the human leukocyte antigen region (HLA), but the real relevant site is unknown.

With the development of sequencing technology and the deepening of genome research, high-depth sequencing and precise variant detection of Chinese MHC regions were reported in Nature Genetics last year, and several psoriasis prone spots were located in its genome association analysis. sense site. However, there is still a lack of classification and prediction models of susceptibility loci based on HLA regions. Therefore, there is an urgent need to develop relevant classification prediction tools to use HLA region susceptible loci to classify and predict data.

Psoriasis is most significantly associated with HLA, but current technologies lack the ability to target HLA regions. Recently, a breakthrough has been made in accurate mutation detection in the HLA region, and the psoriasis-related susceptibility loci on the HLA have been accurately located. The present invention encodes these susceptibility sites and then uses the machine learning model Adaboost to classify them, and can integrate the susceptibility site information found by using the HLA region. The machine learning model is used to comprehensively analyze the data, improve the classification accuracy, and provide a basis for the prevention and screening of psoriasis.

SUMMARY OF THE INVENTION

The purpose of the present invention is to solve the above-mentioned deficiencies of the prior art, find the biomarkers related to psoriasis based on the full coverage of the MHC region, and use SVM-Adaboost to construct a psoriasis-related susceptibility site based on the independent related susceptibility sites of the HLA region. The classification model provides a method for constructing a classification model related to psoriasis, and provides a basis for the prevention and screening of psoriasis.

The present invention is achieved through the following technical solutions:

1 Data processing and transformation

Coding the variation of each sample. Variation information, including HLA type (C*06:02, C*07:04, DPB1*05:01), single nucleotide polymorphism sites (SNP sites) and amino acids, was obtained from high-throughput sequencing data (snp31443520, B:Y33Y, B:Y91C, B:Y140S, snp32472030).

Then, for each sample, according to the susceptibility locus, it is converted into the input data required by the present invention. Edit distance scoring was used for HLA types, and 0/1 scoring was used for SNPs and amino acids. The specific methods are as follows: ① For the susceptible HLA type, calculate the edit distance between the type and the susceptible type of each individual and score; ② For the SNP site, if the mutation exists, it is recorded as 1, and if there is no mutation, it is recorded as 0; ③ For amino acid mutations, the presence of the mutation is scored as 1, and the absence of the mutation is scored as 0.

After the scoring is completed, the data is randomly split into a test set and a training set. Note that the test set and training set data do not overlap. When the number of samples is small, the data can be divided into 5 parts (10 parts) according to the 5-fold crossover method (or 10-fold crossover method), and each time 1 is taken as the test set, and the rest are used as the training set.

2 Classification of data using adaboost-SVM model

The invention utilizes the adaboost method to integrate the support vector machine (SVM) classifier, integrates and utilizes all susceptible site information, and improves the accuracy of data classification.

2.1 About the construction of the classification model

2.1.1 Subclassification model SVM

Support Vector Machine Model SVM is a classic machine learning classification software, which belongs to supervised learning. The Gaussian kernel function (Equation 1) first utilized by the present invention projects the data into a high-dimensional space.

Among them, x is any point in the space, y is the center of the selected space, σ is the width parameter, and K(x, y) is the spatial distance from x to y.

Afterwards, the SVM model is used to construct the separation plane in the high-dimensional space. The construction of the separation plane is mainly determined by the points closest to the separation plane (as shown in Figure 1, point A is one of the closest points), and the connection between the nearest point and the separation plane is called a support vector. The plane when the vector is maximized is set as the separation plane, that is, the data is maximally separated by the separation plane. The present invention adopts the SVM model based on python 2 (refer to the website https://www.manning.com/books/machine-learning-in-action).

2.1.2 Classification Model Integration Algorithm Adaboost

Adaboost is an ensemble method that improves the performance of classifiers based on errors. Through multiple training of each sample, the classifier is repeatedly corrected by the error rate and finally integrated to obtain the integrated results. Specific method: First, assign the same weight to the samples. The SVM is then trained on the training dataset and the error rate (ε, Equation 2) of this classifier is calculated.

Error rate ε=Number of correct classifications/Number of total samples (Formula 2)

Then adjust the Gaussian kernel function σ, and then perform SVM again on the same dataset. In the second training of the classifier, the weight of each sample will be re-adjusted (here the weight is a multi-dimensional vector), in which the next classification weight of the correctly classified sample will be reduced, and the lower classification weight of the wrongly classified sample will be reduced. The secondary weight will increase. That is to say, in the end, the weight of the correct classification will be larger than the weight of the wrong classification. The specific method is to calculate the weight α of each classifier according to the error rate.

The weights can be updated after α is calculated.

Correct classification:

Misclassification:

α is the weight of the basic classifier in the final classifier, ε is the error rate of the classifier; (t) represents the order, t represents this time, and t+1 represents the next time; D _i is the weight of the ith training sample.

After calculating the weight D, start to enter the next round of iteration. The process of training and adjusting the weights is repeated continuously until the training error rate is 0 or the number of weak classifiers reaches a specified value. The present invention adopts the adaboost integration framework based on python2 (refer to the website https://www.manning.com/books/machine-learning-in-action)

3 Classify and evaluate data

After the input training set and test set are constructed, they are substituted into the constructed adaboost-SVM model for classification. The results of the classification model are compared with the actual disease or not. The results were evaluated by calculating the accuracy and plotting the ROC curve.

The ROC curve is the method used to select the best signal model. Usually, the area under the ROC curve (AUC) can be calculated to judge whether the classification model is good or bad. For details, refer to Table 1.

Table 1

The beneficial effects of the present invention are:

At present, there is a lack of relevant technologies to classify and predict psoriasis data, and it only stops at judging the presence or absence of loci to infer the disease situation. The invention uses the effective machine learning classifier SVM for classification, and integrates the SVM through the adaboost framework, so as to improve the accuracy of the classifier. The model can integrate SNP, amino acid and type data for classification, comprehensively consider the information of each dimension, and improve the accuracy of the classification results.

Description of drawings

Fig. 1 is a schematic diagram of constructing a separation plane with an SVM model in a high-dimensional space;

Fig. 2 is the ROC curve of training set classification result of the present invention;

Fig. 3 is the ROC curve of the classification result of the test set of the present invention.

Detailed ways

In order to better understand the present invention, the present invention will be further described below with reference to the embodiments and the accompanying drawings. The following embodiments are only to illustrate the present invention and not to limit it.

Example 1

A total of 5168 patients with psoriasis under the age of 30 were selected for the study. The adaboost-SVM model based on the python2 language was used to construct a model for the classification of susceptible loci.

1 Data processing and transformation

In this implementation case, the variation information ped and map files of the sample are first obtained through variation detection. Afterwards, HLA region variation information was extracted according to the susceptible sites (Table 2). Types (1, 2, 7) are scored according to edit distance (see Table 3 for scoring matrix), and amino acid sites and SNP sites (3, 4, 5, 6, 8) are scored according to their presence or absence, The presence is scored as 1, the absence is scored as 0.

Table 2 Susceptibility sites

Table 3 Edit distance scoring matrix

The data list is obtained. Due to the data volume of 5168 cases, 2000 cases are selected as the training set in this case, and the remaining samples are used as the test set.

2 Substitute the model

Substitute the processed data into the adaboost-SVM model constructed by the present invention for calculation. In this case, 9 SVM classifiers are set, and the value of σ ranges from 30 to 3, decreasing successively from large to small.

3 get the result

As shown in Figures 2 and 3, the classification error rate of this case is 23.9%, the training set AUC (area under the ROC curve) is 0.833, and the test set AUC is 0.868, indicating that the present invention achieves good results in this embodiment.

The above-mentioned embodiments are only to describe the preferred embodiments of the present invention, and do not limit the scope of the present invention. On the premise of not departing from the design spirit of the present invention, various modifications made by those of ordinary skill in the art to the technical solutions of the present invention and improvements, all should fall within the protection scope determined by the claims of the present invention.

Claims (3)

1. a construction method of a classification model relevant to psoriasis, is characterized in that, comprises the following steps: (1) Select psoriasis susceptibility sites; (2) According to different types of susceptibility loci, convert it into input data; (3) Use the Adaboost-SVM model to classify data; The psoriasis susceptibility site of step (1) includes at least one of HLA type, SNP site and amino acid; The susceptibility site of the HLA type includes at least one of C*06:02, C*07:04, and DPB1*05:01; The SNP site and the susceptibility site of the amino acid include at least one of snp31443520, B:Y33Y, B:Y91C, B:Y140S, and snp32472030; The transformation method described in step (2) is: using edit distance scoring for HLA type, and 0/1 scoring for SNP and amino acid; the specific method is as follows: 1. For susceptible HLA type, calculate the difference between the type and the susceptible HLA type of each individual; 2) For SNP sites, if the mutation exists, it is recorded as 1, and if there is no mutation, it is recorded as 0; (3) for amino acid mutations, if the mutation exists, it is recorded as 1, and if there is no mutation, it is recorded as 0; The classification of step (3) includes the following steps: (31) using Gaussian kernel function to project the data into high-dimensional space, and then use SVM model to construct the separation plane in the high-dimensional space; (32) Give the same equal weight to the sample, then train the SVM on the training data set data and calculate the error rate of the classifier to train the weak classifier, and then combine the weak classifiers obtained by each training into a strong classifier; (33) Classify and evaluate data. 2. the construction method of a kind of classification model relevant to psoriasis according to claim 1, is characterized in that, the formula of the described Gaussian kernel function of step (31) is:

Among them, x is any point in the space, y is the center of the selected space, σ is the width parameter, and K(x, y) is the spatial distance from x to y. 3. The method for constructing a classification model related to psoriasis according to claim 1, wherein the evaluation method in step (33) is to calculate the area under the ROC curve.

Images