Methamphetamine false positive early warning method based on convolutional neural network assisted similarity retrieval

Abstract

The application provides a methamphetamine false positive early warning method based on convolutional neural network assisted similarity retrieval, comprising the following steps: (A1) data preprocessing; (A2) mass spectrum data decomposition; (A3) similarity matching; (A4) similarity retrieval result output; (A5) mass spectrum data tensorization processing; (A6) methamphetamine quantitative prediction based on a convolutional neural network; (A7) qualitative prediction based on extracted characteristic signal retrieval first, and then quantitative prediction based on similarity and difference between mass spectrum data by using a convolutional neural network deep learning, so that a support sample with the highest matching degree is selected as a retrieval result to infer a query sample, qualitative and quantitative identification of the query sample is realized by outputting three-dimensional results of similarity, quantitative value and probability value, and a methamphetamine false positive early warning is given to a quantitative prediction result of the query sample. The application has the advantages of high identification accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to mass spectrometry technology, in particular to a methamphetamine false positive early warning method based on convolutional neural network assisted similarity retrieval. BACKGROUND

[0002] Traditional laboratory mass spectrometry technology has problems such as high detection cost, harsh operating environment, high professional operation requirement, complex and long time sample pretreatment, and is not suitable for on-site rapid screening of drug users.

[0003] Facing the urgent need of on-site rapid detection of biological samples, with the rapid development of direct ionization ionization technology and small portable mass spectrometry, direct ionization mass spectrometry (Ambient Ionization Mass Spectrometry, AIMS) fully plays the advantages of accurate mass spectrometry analysis, wide spectrum and direct ionization technology without or only simple sample pretreatment, and has the technical advantages of real-time, in-situ, rapid and low cost, etc. The application disadvantages of traditional mass spectrometry detection cannot be applied to on-site rapid detection, and on-site detection of drugs in complex matrix under open environment is realized. Due to the influence of multiple factors such as open on-site detection environment and direct sampling, AIMS data usually presents the characteristics of low repeatability, relatively complex background signal and difficulty in unifying multiple feature matching. Therefore, in the process of on-site rapid detection of trace drugs of biological samples, it may lead to false positive error identification of drug user screening, especially when the concentration of drug detection limit of drug user is close. In order to reduce the influence of high abundance background signal in AIMS data on the identification accuracy of low abundance signal of trace drugs, improve the identification accuracy of positive samples, and explore the potential value of complex AIMS spectrum data.

[0004] Mass spectrometry library similarity retrieval is an effective method for identifying target substances. However, under the interference of complex background signal and noise, the existing spectrum library retrieval method often has false positive identification problem. In order to improve the similarity and accuracy of mass spectrometry data identification, the intelligent application method of data mining is studied, the potential relationship between data characteristics in a large amount of complex data is clarified, and the mass spectrometry data processing and classification identification model is successfully established. SUMMARY

[0005] In order to solve the above problems in the prior art, the present application provides a methamphetamine false positive early warning method based on convolutional neural network assisted similarity retrieval.

[0006] The purpose of the present application is realized by the following technical scheme:

[0007] The methamphetamine false positive early warning method based on convolutional neural network assisted similarity retrieval comprises the following steps:

[0008] (A1) Data preprocessing: Extract ion intensity and normalize abundance from the original AIMS data respectively;

[0009] (A2) Mass spectrum data decomposition: Construct a mass spectrum data matrix as the background signal source of the sample mass spectrum data to be decomposed;

[0010] (A3) Similarity matching;

[0011] (A4) Similarity search result output: Sort the similarity values obtained by similarity matching from high to low, select the top n matching support samples with higher similarity values as the result, output n support samples and their corresponding similarity, and identify the query sample as positive or negative according to the search result;

[0012] (A5) Mass spectrum data tensorization processing: For the similarity search result, 1 query sample and n support samples form n sample pairs, and the original mass spectrum data of the query sample and the support sample in the sample pair are converted into a two-dimensional matrix after normalization;

[0013] (A6) Methamphetamine quantitative prediction based on convolutional neural network: Use convolutional neural network to establish a mass spectrum data classification model to evaluate the probability of the query sample and the support sample having the same methamphetamine content in the input two-dimensional matrix data;

[0014] According to the probability value, reorder the n sample pairs, select the methamphetamine content of the support sample with the highest probability, its reference probability and similarity value as the prediction result, and realize the quantitative prediction of the query sample.

[0015] (A7) By first searching the qualitative characteristics based on the extracted features, and then using convolutional neural network to learn the similarity and difference between mass spectrum data for quantitative prediction, the support sample with the highest matching degree is selected as the search result to predict the query sample. The three-dimensional results of similarity, quantitative value and probability value are output to realize the qualitative and quantitative identification of the query sample;

[0016] When the quantitative prediction of the query sample is not less than 0.1 ng / mg, and the probability value meets the judgment condition, if the similarity between the query sample and the support sample with the highest probability is less than 0.6, it is considered that there is a qualitative conflict between the similarity search result and the quantitative prediction result for the query sample. The quantitative prediction result of the query sample is given a methamphetamine false positive warning.

[0017] The technical concept of the present application is that in the actual application of hair sample trace drug detection, the direct ionization mass spectrum data has the characteristics of low repeatability and complex background signal under the influence of multiple factors such as open detection environment, sample introduction and complex matrix, which leads to the difficulty of directly and universally simple application of spectrum data analysis. Around the data characteristics of hair methamphetamine direct ionization mass spectrum detection, aiming at the false positive problem, a false positive warning method of methamphetamine based on convolutional neural network assisted similarity retrieval is proposed. Taking blank samples and spiked samples as support samples, firstly, a similarity retrieval model integrating generalized weighted robust principal component analysis algorithm and improved Pearson correlation coefficient is constructed, which can accurately remove the background and extract the characteristic signal of the query sample and the support sample, and then the similarity matching between the characteristic signals of the query sample and the support sample is carried out to give the qualitative retrieval result, then a quantitative prediction model based on convolutional neural network is established, the retrieval result is reordered and the support sample with the highest probability is selected as the matching object to output the similarity value, the quantitative prediction value and the reference probability, and finally a false positive warning strategy of methamphetamine based on the similarity value and the probability value is constructed to realize the false positive warning of methamphetamine.

[0018] Compared with the prior art, the present application has the beneficial effects of:

[0019] 1. Around the spectrum data characteristics of hair methamphetamine direct ionization mass spectrum detection, aiming at the false detection problem of methamphetamine false positive caused by low repeatability and complexity of background signal under the influence of complex matrix and open field environment, the potential feature information value of complex mass spectrum data is mined, a rapid discrimination model of hair sample methamphetamine positive and methamphetamine negative and a methamphetamine false positive accurate discrimination warning strategy are constructed, and the identification accuracy of methamphetamine positive sample is improved.

[0020] 2. The present application proposes a new background removal and feature extraction method based on m / z domain information, which decomposes the low rank structure part from the mass spectrum data by generalized weighted robust principal component analysis algorithm to realize the deduction of complex background signal, overcomes the limitation that the conventional mass spectrum data processing method is not suitable for background signal deduction and feature extraction under the characteristics of low repeatability and complexity of direct ionization mass spectrum data, uses convolutional neural network to learn the potential features of mass spectrum data, constructs a quantitative prediction model of convolutional neural network assisted similarity retrieval and forms a methamphetamine false positive warning strategy, improves the identification accuracy of methamphetamine positive, and better solves the problem that the traditional hair methamphetamine screening method relying on target ion intensity is not accurate due to the low repeatability and complexity of mass spectrum data. BRIEF DESCRIPTION OF DRAWINGS

[0021] The disclosure of the present application will become more apparent with reference to the accompanying drawings. It is easy for those skilled in the art to understand that these drawings are only used to illustrate the technical solutions of the present application, and are not intended to limit the protection scope of the present application. In the drawings:

[0022] Figure 1 is a flow chart of the methamphetamine false positive early warning method based on convolutional neural network assisted similarity retrieval according to an embodiment of the present application;

[0023] Figure 2 is a flow chart of mass spectrometry data decomposition based on a generalized weighted robust principal component analysis model;

[0024] Figure 3 is a schematic diagram of sample pair mass spectrometry data tensorization processing;

[0025] Figure 4 is a schematic diagram of mass spectrometry data decomposition results of two randomly selected support samples;

[0026] Figure 5 is a schematic diagram of mass spectrometry data decomposition results of three randomly selected query samples;

[0027] Figure 6 is a schematic diagram of model output probability values of 95 positive generated data of false methamphetamine quantitative prediction;

[0028] Figure 7 is a schematic diagram of quantitative prediction results and their probability values of real hair samples as a query sample set. DETAILED DESCRIPTION

[0029] Figures 1-7 The optional specific embodiments of the present application described in the following description and the accompanying drawings are presented for the purpose of teaching and understanding the present application. Some conventional aspects have been simplified or omitted for the sake of clarity. Those skilled in the art should understand that variations or modifications of the specific embodiments described herein could be made without departing from the scope of the present application. Those skilled in the art should understand that the features described below could be combined in various ways to form multiple variations of the present application. Thus, the present application is not limited to the optional specific embodiments described below, but only by the claims and their equivalents. Example 1

[0030] The methamphetamine false positive early warning method based on convolutional neural network assisted similarity retrieval of the present application embodiment 1, as shown in Figure 1 , includes the following steps:

[0031] (A1) Data preprocessing. Extract ion intensity and normalize abundance for original AIMS data respectively. Extracting ion intensity refers to the sum of ion signal intensity within a certain mass-to-charge ratio range [1]. When extracting ion intensity for all target ions, the range of unit m / z should consider the actual resolution of the mass spectrometry system. Normalize abundance using total ion intensity to convert the absolute intensity of mass spectrometry data into relative intensity, which can reduce the influence of low repeatability of AIMS data. Output preprocessed mass spectrometry data .

[0032] (A2) Mass spectrum data decomposition. Import the original AIMS data of m blank samples, construct the mass spectrum data matrix D after data preprocessing , which is the source of background signal of the sample mass spectrum data to be decomposed. The formula is as follows:

[0033] After data preprocessing, the sample s to be decomposed is added to the matrix D, and the matrix changes to ; then the data matrix D is decomposed based on the generalized weighted robust principal component analysis model , to obtain the background signal matrix A , the characteristic signal matrix E and the dense small noise matrix H , where A s , E s and H s are the background signal, characteristic signal and noise signal of the sample s, respectively, and the matrices , and record the background, characteristic and noise signals of the m blank samples. The structure of the mass spectrum data decomposition model is shown in Figure 2 , which mainly consists of two parts, namely the cyclic update of matrix and parameter.

[0034] The cyclic update matrix , and . In each cycle, a random sequence similar to (1, 2, 3) is generated , and then the sequence is traversed. For the i-th element in the sequence: if , then , and then singular value vectors and are obtained by singular decomposition of the matrices and , respectively. Then the singular value vectors of are updated combined with the weight vector W, and finally the matrix is updated, and the calculation formula is as follows:

[0035] If , update the matrix . The calculation formula is as follows:

[0036] In formula (3), is a soft threshold function. If , update the matrix . The calculation formula is as follows: The parameters are cyclically updated, mainly updating the matrix Y and the parameter μ, which are used to update the matrix in the next cycle. The calculation formula is as follows:

[0037] (A3) Similarity matching. Spectral library searching is to infer the structural properties of the query sample by similarity matching with the identified mass spectra in the spectral library. Spectral similarity scoring is widely used in spectral library searching, and its general process includes three steps: feature extraction, ion peak intensity conversion and core scoring calculation function [3]. The Pearson correlation coefficient has better similarity scoring application effect than cosine distance, Spearman correlation coefficient and mean square error (MSE) [4]. Therefore, a matching model is constructed based on the Pearson correlation coefficient, and the similarity value is obtained by matching calculation between the feature signals of the query sample and the support sample. The calculation formula is as follows: In the formula, X represents the feature signal of the query sample, and Y represents the feature signal of the support sample. Generally, the similarity value obtained by directly calculating the above formula will return null. This is because the feature signal decomposed from the sample s may appear . To solve this problem, the above formula is improved in this method, and low-intensity Gaussian noise is added to the feature signal, that is, , to avoid returning null similarity value.

[0038] (A4) Similarity search result output. The similarity values obtained by similarity matching are sorted from high to low, and the top n support samples with higher similarity values are selected as the results, and n support samples and their corresponding similarity values are output. The query sample can be qualitatively identified as positive or negative for methamphetamine according to the search results.

[0039] (A5) Mass spectral data tensorization processing. For the similarity search results, 1 query sample and n support samples form n sample pairs, and the original mass spectral data of the query sample and the support sample in the sample pair are converted into a two-dimensional matrix after normalization abundance, and the specific process is shown in Figure 3 . The original mass spectral data is divided into b segments after normalization abundance, each segment contains a relative intensity value, and each segment of mass spectral data of the query sample and the support sample is arranged in order, and finally converted into a two-dimensional matrix data of a x 2b.

[0040] (A6) Methamphetamine quantitative prediction based on convolutional neural network. A mass spectral data classification model is established using convolutional neural network to evaluate the probability that the query sample and the support sample in the input two-dimensional matrix data have the same methamphetamine content. According to the probability value, the n sample pairs are reordered, and the methamphetamine content of the support sample with the highest probability, its reference probability and similarity value are selected as the prediction results, and the methamphetamine quantitative prediction of the query sample is realized.

[0041] (A7) False positive warning strategy analysis of methamphetamine; By searching the qualitative characteristics of the extracted characteristic signals first, and then using convolutional neural network deep learning to quantitatively predict the similarities and differences between mass spectrometry data, the highest matching support sample is selected as the search result to speculate the query sample, and the three-dimensional results of similarity, quantitative value and probability value are output to realize the qualitative and quantitative identification of the query sample. In the similarity search qualitative, when the Pearson correlation coefficient value is greater than or equal to 0.6, it is considered that there is a strong correlation between the vectors, so when the quantitative value gives the positive identification result of the query sample of methamphetamine, if the similarity is less than 0.6, the positive qualitative result is not established. Therefore, based on the similarity and quantitative value and its reference probability, the false positive warning strategy of the method for methamphetamine can be proposed. When the quantitative prediction of the query sample is greater than or equal to 0.1 ng / mg and the probability value meets the judgment condition, if the similarity between the query sample and the support sample with the highest probability is less than 0.6, it is considered that there is a qualitative conflict between the similarity search result and the quantitative prediction result for the query sample. The quantitative prediction result of the query sample is given a false positive warning of methamphetamine. Example 2

[0042] Application example of the false positive warning method for methamphetamine based on convolutional neural network assisted similarity search according to the application example 1.

[0043] The target drug is methamphetamine, which is relatively more popular among users. The precursor ion m / z 150 is mainly fragmented into two fragment ions m / z 91 and m / z 119. The support sample set contains 300 blank hair samples and 100 hair samples with MS / MS mass spectrum data of methamphetamine detection. The hair methamphetamine content of the added sample ranges from 0.1 ng / mg to 1 ng / mg (interval 0.1), and there are 10 added samples for each different methamphetamine content. Since the real hair samples of methamphetamine positive and false positive are not easy to obtain, the query sample set contains 10 positive samples and 13 false positive samples. Considering that the number of negative samples has no important influence on the feature mining and identification of positive and false positive sample data, 10 negative samples are added to the query sample set to further prove the accuracy of the method in the qualitative identification of methamphetamine positive or negative.

[0044] The detection data is obtained in different environments, the same instrument, different operators and other actual field application scenarios.

[0045] Figure 4The mass spectrum data decomposition results of two randomly selected support samples are shown. (a) represents the mass spectrum data after data preprocessing of the original secondary spectrum of the sample; (b) represents the background signal obtained after signal decomposition of the sample mass spectrum data; (c) represents the characteristic signal obtained after signal decomposition of the sample mass spectrum data; (d) represents the noise signal obtained after signal decomposition of the sample mass spectrum data. As can be seen from the figure, the obtained characteristic signal retains the fragment ion information with obvious relative intensity fluctuation, including the fragment ion characteristics of the target substance (i.e. methamphetamine); this method can effectively remove the background signal, reduce the data complexity, and promote the identification of target ion characteristics.

[0046] Figure 5 The mass spectrum data decomposition results of three randomly selected query samples are shown, including one positive sample, one negative sample and one false positive sample. (a) represents the mass spectrum data after data preprocessing of the original secondary spectrum of the sample; (b) represents the background signal obtained after signal decomposition of the sample mass spectrum data; (c) represents the characteristic signal obtained after signal decomposition of the sample mass spectrum data; (d) represents the noise signal obtained after signal decomposition of the sample mass spectrum data. As can be seen from the figure, compared with the characteristic signals of the three samples, the false positive sample detection data is affected by environmental noise or ions of non-target substances in complex matrix, which greatly increases the ion intensity of m / z91 and m / z119, directly exceeding the intensity threshold for determining methamphetamine use, resulting in false positive detection.

[0047] The positive change of ion relative intensity in the complex matrix sample is mainly obtained The similarity values obtained by similarity matching are sorted from high to low, and the top 10 matching support samples with higher similarity values are selected as the retrieval results. Considering the sparsity of mass spectrum data and the size of mass spectrum data converted into a two-dimensional matrix, a network model suitable for current mass spectrum data processing and classification is established based on convolutional neural network, and the specific network structure configuration is shown in Table 1. Due to the sparsity of mass spectrum data, the first convolution kernel size is set to 15x15, and the larger the convolution kernel size, the larger the receptive field, which can extract more effective potential information from the data. The sizes of the second and third convolution kernels are set to 5x5 and 3x3 respectively to reduce the amount of calculation. For the last layer output layer, [y0, y1] is the output result, y0 represents the negative sample score, and y1 represents the positive sample score. The value of y1-y0 is taken as the prediction probability value, with a value range of [-1, +1], and the higher the probability value, the higher the possibility that the query sample and the corresponding support sample have the same methamphetamine content.

[0048] Table 1: Quantitative prediction network structure

[0049] .

[0050] Based on the support sample set, 100 examples of different methamphetamine content data were generated using a one-dimensional signal data enhancement method, and the generated data set was used as a query sample to fully train the mass spectrometry data classification network. The accuracy of rank was used as an evaluation index to analyze the output results of the model to evaluate the performance of the proposed method. As an evaluation standard for retrieval and identification, the accuracy of rank represents that the top k support samples with higher matching with the query sample in the retrieval list are selected as the retrieval results. When k = 1, it represents selecting the support sample with the highest matching degree, and if the support sample is the same category as the query sample, it is considered correct identification. When k = 2, it represents selecting the top 2 support samples with higher matching, and if there is a sample in the two support samples that is the same category as the query sample, it is also considered correct identification. When k = 3, 4, …, it is the same. The calculation formula is as follows. The difference is that the query sample and the retrieved support sample with the same methamphetamine content are considered correct identification.

[0051] For the identification of 1100 examples of generated data of methamphetamine content, the recognition rates of similarity retrieval and convolutional neural network assisted similarity retrieval are shown in Table 2. The results show that the convolutional neural network in the method can reorder the similarity retrieval results, and effectively assist the similarity retrieval to realize quantitative prediction and identification by selecting the methamphetamine content of the support sample with the highest prediction probability value as the quantitative result. Under the condition of selecting the top 10 matching samples in similarity retrieval and k = 1, the recognition rate can reach 91.4%, and the positive samples and negative samples of methamphetamine can be accurately distinguished by 100%. For the samples that cannot be correctly identified, considering that the similarity retrieval result does not contain the support sample corresponding to the query sample, the selection range of the retrieval result can be expanded to improve the recognition rate, but at the same time, the calculation cost will also increase.

[0052] Figure 6 The probability values of 95 examples of false methamphetamine quantitative prediction are shown. For sample pairs with different methamphetamine content, the model will output low probability values tending to -1. Combined with the probability values of the correct quantitative prediction results of the model, the judgment condition of the quantitative prediction result can be given: when the output probability value is greater than or equal to 0.6, the quantitative prediction of the query sample can be considered true; when the output probability value is less than 0.6, the prediction result is identified as false quantitative identification, and the query sample can be identified qualitatively according to the similarity retrieval result.

[0053] Table 2: Comparison of recognition rates of similarity retrieval and convolutional neural network assisted similarity retrieval in methamphetamine quantitative prediction

[0054]

[0055] The second spectrum data of 33 real hair samples for methamphetamine detection were used as the query sample set to verify the performance of the quantitative prediction model after training. Figure 7 The quantitative prediction results and their probability values of all samples were shown, (a) showed that the query sample was predicted to be 0.0 ng / mg, (b) showed that the sample was predicted to be greater than or equal to 0.1 ng / mg; the output probability value of all samples was greater than 0.6. According to the determination condition of the quantitative prediction result, the quantitative prediction result of all query samples was considered to be true, and the model accurately identified 10 negative samples and 11 false positive samples; the quantitative prediction results of 10 positive samples and the remaining 2 false positive samples were shown in Table 3, the quantitative prediction results of the positive samples were consistent with the expected values, and the 2 false positive samples were incorrectly identified. However, the similarity values of the 2 false positive samples were less than 0.6, and according to the false positive warning strategy of the method, there was a qualitative conflict between the similarity search result and the quantitative prediction result of the query sample, and the false positive warning of methamphetamine was given. Therefore, under this strategy, the quantitative prediction accuracy of the model reached 93.9% in the 33 real sample set, the false positive rate was reduced to 8.7%, and the false positive warning accuracy reached 100%.

[0056] Table 3: Quantitative prediction results of 10 positive samples and 2 false positive samples

[0057] .

Claims

1. A methamphetamine false positive early warning method based on convolutional neural network assisted similarity retrieval, the methamphetamine false positive early warning method comprising the following steps: (A1) Data preprocessing: extracting total ion intensity and normalized abundance from the original AIMS data, converting the absolute intensity of mass spectrometry data into relative intensity, and outputting the preprocessed mass spectrometry data ; (A2) Mass spectrum data decomposition: construct a mass spectrum data matrix as the background signal source of the sample mass spectrum data to be decomposed; (A3) Similarity matching; (A4) Similarity retrieval result output: sort the similarity values obtained from the similarity matching from high to low, select the top n matching support samples with higher similarity values as the results, output the n support samples and their corresponding similarity, and identify the query sample as methamphetamine positive or negative according to the retrieval results; (A5) Mass spectrum data tensorization processing: for the similarity retrieval results, 1 query sample and n support samples form n sample pairs, and the original mass spectrum data of the query sample and the support sample in the sample pair are converted into a two-dimensional matrix after normalization abundance; (A6) Methamphetamine quantitative prediction based on convolutional neural network: a mass spectrum data classification model is established using a convolutional neural network to evaluate the probability that the query sample and the support sample have the same methamphetamine content in the input two-dimensional matrix data; reorder the n sample pairs according to the probability value, select the methamphetamine content of the support sample with the highest probability, its reference probability and similarity value as the prediction result, and realize the quantitative prediction of the query sample; (A7) By first retrieving the extracted characteristic signal to determine the qualitative, and then using the convolutional neural network to learn the similarity and difference between the mass spectrum data to quantitatively predict, the support sample with the highest matching degree is selected as the retrieval result to speculate the query sample, and the three-dimensional results of similarity, quantitative value and probability value are output to realize the qualitative and quantitative identification of the query sample; When the quantitative prediction of the query sample is not less than 0.1 ng / mg, and the probability value meets the judgment condition, if the similarity between the query sample and the support sample with the highest probability is less than 0.6, it is considered that there is a qualitative conflict between the similarity retrieval result and the quantitative prediction result for the query sample, and the methamphetamine false positive warning is given for the quantitative prediction result of the query sample.

2. The methamphetamine false positive early warning method based on convolutional neural network assisted similarity retrieval according to claim 1, characterized in that, In step (A3), a matching model is constructed based on Pearson correlation coefficient, and a similarity value is obtained by matching calculation between the characteristic signals of the query sample and the support sample: ; X represents the characteristic signal of the query sample, and Y represents the characteristic signal of the support sample. 3.The methed of false positive warning of methamphetamine based on convolutional neural network aided similarity retrieval according to claim 1, characterized in that, In step (A5), the original mass spectrum data is divided into b segments after normalization abundance, each segment contains a relative intensity value, and the mass spectrum data of the query sample and the support sample are arranged in order, finally converted into a two-dimensional matrix data of a x 2b.

4. The methamphetamine false positive early warning method based on convolutional neural network assisted similarity retrieval according to claim 1, characterized in that, In step (A7), in the similarity retrieval qualitative, when the Pearson correlation coefficient value is not less than 0.6, it is considered that there is a strong correlation between the vectors, so when the quantitative value gives the positive identification result of the query sample, if the similarity is less than 0.6, the positive qualitative result is not established.

Images