CN111896609A - A method for analyzing mass spectrometry data based on artificial intelligence - Google Patents

Abstract

A method for analyzing mass spectrometry data based on artificial intelligence, the method comprises: using a laser-assisted desorption/ionization mass spectrometer to collect metabolite small molecule fingerprints for each of the samples; extracting the absolute intensity of the fingerprints ; Input the processed data into the multi-layer neural network for sample grouping processing. A method for calculating the importance of contribution to sample discrimination, converting fingerprint data into a two-dimensional image; using a saliency feature analysis method to calculate the data in the metabolite screening picture library, and sorting all the features to screen out Substances that contribute the most to sample discrimination. The beneficial effects of the present invention are that samples are quickly grouped for mass spectrometry data, and the interpretability of the classification model is greatly improved.

Description

A method for analyzing mass spectrometry data based on artificial intelligence

technical field

The invention belongs to the field of artificial intelligence-assisted mass spectrometry data mining, and specifically designs metabolite fingerprints obtained based on mass spectrometry and artificial intelligence analysis technology for constructing a sample grouping model and calculating the importance of grouping.

Background technique

Mass spectrometry detection methods show the advantages of high-throughput analysis and multi-metabolite detection, and are the main method to detect untargeted metabolites. However, the application of mass spectrometry detection methods also faces problems such as lengthy preprocessing, including high complexity and low metabolite abundance in biological samples. With the development of nanotechnology, the recently developed nano-assisted laser desorption/ionization mass spectrometry (LDI MS), due to its high analytical throughput (~300 samples/hour) and precise metabolic identification (mass error <50ppm), Become the most practical tool for metabolic analysis.

Deep learning is mainly used in the auxiliary analysis of large and high-latitude datasets. It can accept input of various data types, and has become a cutting-edge technology in the analysis of various medical data. As a frontier field of machine learning, deep learning has become a major analysis tool. Because it has the characteristics of optimizing the loss function as much as possible to learn the relevant data rules and mining the potential features of the data as much as possible, it is widely used in various field. It has been widely used in the biomedical field. However, traditional machine learning method analysis cannot be well applied to mass spectrometry data analysis and mining, because the mass spectrometry data itself has huge sample characteristics, and there will be problems such as overfitting and underfitting that reduce the accuracy. In addition, deep learning is a black box, and it is difficult to select important features from the points to explain the mechanism of the diagnosis principle.

The main method of common feature selection in deep learning is the saliency region map, which is mainly used in the image field. It is extended to solve complex scene understanding problems in various fields such as neuroscience, psychology and medical diagnosis. However, this method of significance analysis has not yet been applied to mass spectrometry data analysis.

SUMMARY OF THE INVENTION

Aiming at the problems of long time consumption, high data dimension and complex combination in mass spectrometry data analysis, the present invention provides a fast, accurate and efficient classification model based on an improved multi-layer neural network to realize rapid grouping of samples and calculate the classification contribution The importance of the method greatly improves the interpretability of the classification model.

A method for analyzing mass spectrometry data based on artificial intelligence, using a multi-layer neural network to analyze and process mass spectrometry data to realize grouping of samples;

Methods include:

Step 1: Pipette the sample onto the mass spectrometer target plate, and use it as a thin layer for subsequent mass spectrometry analysis after drying;

Step 2: Collect small molecule fingerprints of metabolites between 100 and 1000 in positive ion mode for each analyzed sample using a laser-assisted desorption/ionization mass spectrometer, without any smoothing procedure;

Step 3: Extract the absolute intensity of the original metabolic fingerprint, and perform centralized preprocessing on the data extracted from all samples;

Step 4: Input the data of Step 3 into the neural network for sample grouping processing.

Further, in step 2, at least 2 independent experiments were performed on each sample to eliminate intra-individual bias and improve the reproducibility and stability of the analysis.

Further, the multi-layer neural network includes: network input, network main body, and network output; the network main body includes a feature extraction part, a nonlinear feature interaction layer, and a classification layer; the network input is first processed by the feature extraction part, and the output of the feature extraction part It is processed by the nonlinear feature interaction layer, the output of the nonlinear feature interaction layer is processed by the classification layer, and the output of the classification layer is the network output;

The principle formula from the network input all the way to the classification layer is:

x_input=concatenate(x_spectral,x_ext) (1)

x_fs=feature_extract(x_input) (2)

x_nl=feature_interaction(x_fs) (3)

y_pred=softmax(x_nl) (4)

Further, the network input is 1-1024-dimensional multimodal features (x_input), including the original mass spectral data input (x_spectral), and other parts are filled with 0s. Based on the finiteness of the samples, all multimodal features are simply scaled and centered.

Further, the feature extraction part (feature_extract) is composed of four layers of local connection (LocallyConnected1D) layers. Each LocallyConnected1D layer divides all features into 32 intervals for fully connected feature extraction (32 fully connected with their own parameters). Compared with the four-layer fully connected architecture, it can reduce the network width and parameter scale while finely modeling the characteristics of mass spectrometry data. The extraction process also indirectly reduces overfitting.

Further, the principle formula of four-layer LocallyConnected1D layer stacking:

Further, the nonlinear feature interaction layer (feature_interaction) learns its nonlinear relationship for the 96 latent features obtained in the feature extraction part. In the feature interaction part, each layer can simultaneously extract the discrete Relu activation features, and can also extract the approximate quadratic relationship of the linear combination of features, which can be extracted as fusion features through residuals or splicing and merging, which can better extract nonlinear features. Dropout regularization can further alleviate overfitting and enhance generalization performance. The nonlinear feature interaction layer can be regarded as a new self-attention mechanism suitable for the fully connected layer, and has the fusion ability of discrete and quadratic features. Compared with the multi-layer fully connected architecture, the network width is reduced. The nonlinearity is enhanced along with the parameter scale, which is beneficial to reduce overfitting under limited samples and improve the final classification performance.

Further, the principle formula of the nonlinear feature interaction layer:

Further, non-target detection of metabolic fingerprints after sample preprocessing was performed to obtain relevant metabolite databases, to construct the mapping relationship between grouping information and metabolic profiles, and to divide training set samples and blind test set samples.

Further, the neural network is trained using the sample data, and 3/4 of the training set data is randomly used as the training group and 1/4 as the test group. 10-fold cross-validation (10-fold) training is performed on the training group samples based on the multi-layer neural network, and classification is achieved by counting the accurate average value of the final model.

A method for calculating the importance of sample discrimination contribution, including the following steps:

Step 11: Convert mass spectrometry data into two-dimensional images, and build a metabolite screening image library;

Step 12: Calculate the data in the metabolite screening image library by using the saliency maps, and sort all the features to screen out the substances that contribute the most to the sample discrimination.

The present invention has the following technical effects: rapid grouping of samples for mass spectrometry data is realized, and the interpretability of the classification model is greatly improved.

Description of drawings

FIG. 1 is a schematic diagram of a neural network structure in an embodiment of the present invention.

Detailed ways

The preferred embodiments of the present application will be described below with reference to the accompanying drawings, so as to make its technical content clearer and easier to understand. The present application can be embodied in many different forms of embodiments, and the protection scope of the present application is not limited to the embodiments mentioned herein.

The concept, specific structure and technical effects of the present invention will be further described below to fully understand the purpose, features and effects of the present invention, but the protection of the present invention is not limited to this.

In an embodiment of the present invention, the data of the sample to be tested is first prepared, and the steps are as follows:

Step 1: Pipette the sample onto the mass spectrometer target plate, and use it as a thin layer for subsequent mass spectrometry analysis after drying;

Step 2: Collect metabolite small molecule fingerprints between 100-1000 in positive ion mode for each analyzed sample using a laser-assisted desorption/ionization mass spectrometer, without any smoothing procedure, and perform five Independent experiments to remove intra-individual bias and improve the reproducibility and stability of diagnostic results;

Step 3: Extract the absolute intensity of the original metabolic fingerprint (between 100-1000 m/z mass-to-charge ratio), and perform centralized preprocessing on the data extracted from all samples for further machine learning;

Non-target detection of metabolic fingerprints after sample preprocessing was performed to obtain relevant metabolite databases, to construct the mapping relationship between grouping information and metabolic profiles, and to divide training set samples and blind test set samples.

In this embodiment, the neural network structure for processing the data extracted from the sample is as follows:

The input to this network is a multimodal feature (x_input) of 1-1024 dimensions, which includes the raw mass spectral data input (x_spectral), and the other parts are padded with 0s. Based on the finiteness of the samples, all features are simply scaled and centered (-1,1). The main body of the network is divided into two parts, followed by the input is the feature extraction part (feature_extract), the feature extraction layer is followed by the nonlinear feature interaction layer (feature_interaction), and finally the reorganized 96 features are input to the Softmax classification layer for classification probability output .

The principle formula from the 1024-dimensional input to the Softmax layer:

x_input=concatenate(x_spectral,x_ext) (1)

x_fs=feature_extract(x_input) (2)

x_nl=feature_interaction(x_fs) (3)

y_pred=softmax(x_nl) (4)

The feature extraction part (feature_extract) is composed of four LocallyConnected1D layers stacked. Each LocallyConnected1D layer divides all features into 32 intervals for fully connected feature extraction (32 fully connected layers with their own parameters), so as to reflect mass spectrometry data. Compared with the four-layer fully connected architecture, the network width and parameter scale can be reduced, and the feature extraction process of mass spectrometry data can be finely modeled, and the feature extraction process of mass spectrometry data can also be indirectly reduced. overfitting.

The principle formula of four-layer LocallyConnected1D layer stacking:

The nonlinear feature interaction layer (feature_interaction) learns the nonlinear relationship of the 96 latent features obtained in the feature extraction part. In the feature interaction part, each layer can simultaneously extract the discrete Relu activation features, and can also extract the approximate quadratic relationship of the linear combination of features, which can be extracted as fusion features through residuals or splicing and merging, which can better extract nonlinear features. Dropout regularization can further alleviate overfitting and enhance generalization performance. The nonlinear feature interaction layer can be regarded as a new self-attention mechanism suitable for the fully connected layer, and has the fusion ability of discrete and quadratic features. Compared with the multi-layer fully connected architecture, the network width is reduced. The nonlinearity is enhanced along with the parameter scale, which is beneficial to reduce overfitting under limited samples and improve the final classification performance.

The principle formula of the nonlinear feature interaction layer:

Use the sample data to train the above neural network, randomly select 3/4 of the training set data as the training group and 1/4 as the test group. 10-fold cross-validation (10-fold) training is performed on the training group samples based on the multi-layer neural network, and the classification is realized by counting the accurate average value of the final model;

The trained network is used for blind test set sample analysis, and by analyzing the accuracy of prediction, it is verified that the multi-layer neural network-based grouping model can achieve accurate classification;

Further analysis of the contribution of the classification model includes the following steps:

Step 11: Convert the mass spectrometry data into a two-dimensional image, and construct a metabolite screening image library;

Step 12: Calculate the data in the metabolite screening image library by using the saliency maps, and sort all the features to screen out the substances that contribute the most to the sample discrimination.

The preferred specific embodiments of the present application are described in detail above. It should be understood that many modifications and changes can be made in accordance with the concept of the present application without creative efforts by those skilled in the art. Therefore, any technical solutions that can be obtained by those skilled in the art through logical analysis, reasoning or limited experiments on the basis of the prior art according to the concept of the present application shall fall within the protection scope determined by the claims.

Claims (10)

1. A method for analyzing mass spectrum data based on artificial intelligence is characterized in that the mass spectrum data is analyzed and processed by using a multilayer neural network, and grouping of samples is realized; the method comprises the following steps:

step 1: absorbing and moving the sample to a mass spectrum target plate, drying the sample, and performing subsequent mass spectrum analysis as a thin layer;

step 2: collecting the metabolite small molecule fingerprint spectrogram between the positive ion mode 100-1000 for each sample by adopting a laser-assisted desorption/ionization mass spectrometer without any smoothing program;

and step 3: extracting absolute intensity of the fingerprint spectrum, and performing centralized preprocessing on the extracted data;

and 4, step 4: and (4) inputting the data processed in the step (3) into the multilayer neural network, and performing sample grouping processing.

2. The artificial intelligence based method of analyzing mass spectrometry data of claim 1, wherein in step 2, at least 2 independent experiments are performed on each of the samples.

3. The artificial intelligence based method of analyzing mass spectrometry data of claim 1, wherein the multi-layer neural network comprises: network input, network subject, network output; the network main body comprises a feature extraction part, a nonlinear feature interaction layer and a classification layer; the network input is firstly processed by the feature extraction part, the output of the feature extraction part is processed by the nonlinear feature interaction layer, the output of the nonlinear feature interaction layer is processed by the classification layer, and the output of the classification layer is the network output;

the principle formula from the network input up to the classification level is:

x_input＝concatenate(x_spectral,x_ext) (1)

x_fs＝feature_extract(x_input) (2)

x_nl＝feature_interaction(x_fs) (3)

y_pred＝softmax(x_nl) (4)。

4. the method for artificial intelligence based analysis of mass spectrometry data of claim 3, wherein the network input is a multi-modal 1-1024 dimensional signature including raw mass spectrometry data input, and the rest is filled with 0 s.

5. The method for analyzing mass spectrometry data based on artificial intelligence as claimed in claim 3, wherein the feature extraction part is formed by stacking four local connection layers, each of the local connection layers divides all features into 32 intervals for full connection feature extraction.

6. The method for artificial intelligence based analysis of mass spectrometry data of claim 3, wherein the nonlinear feature interaction layer learns the nonlinear relationship of the 96 latent features obtained by the feature extraction portion.

7. The artificial intelligence based method of analyzing mass spectral data of claim 6, wherein the principle formula of the nonlinear feature interaction layer is:

8. the method for analyzing mass spectrometry data based on artificial intelligence of any one of claims 1 or 2, wherein the fingerprint is subjected to non-target detection, a related metabolite database is obtained, a mapping relation between grouping information and a metabolic profile is constructed, and a training set sample and a blind measurement set sample are divided.

9. The method for artificial intelligence based analysis of mass spectrometry data of claim 8, wherein 3/4 of the data of the training set samples are used as a training set and 1/4 is used as a test set. And performing 10-fold cross validation (10-fold) training on the training group samples based on the multilayer neural network, and realizing classification by counting the accurate average value of the final model.

10. A method for calculating the significance of a sample discrimination contribution, comprising the steps of:

step 11: converting the fingerprint data as set forth in claim 8 into two-dimensional images to construct a metabolite screening picture library;

step 12: and (3) calculating the data in the metabolite screening picture library by using a significant feature analysis method, sequencing all features, and screening out the substances which have the greatest contribution to sample discrimination.

Images