Mass spectrum data processing and identification method based on full-connection variational neural network

Abstract

The invention relates to a mass spectrum data processing and identification method based on a full-connection variational neural network, and the method comprises the steps: S1, obtaining mass spectrum original data and pathological annotation information, and carrying out the data preprocessing; s2, inputting the preprocessed data into an unsupervised learning model based on variational self-coding, learning low-dimensional feature representation of mass spectrum data, and realizing data visualization; s3, associating the low-dimensional feature representation learned by the model with features in the original data to realize peak value learning, and obtaining a potential feature ion list related to a tumor cluster or anatomy; and S4, inputting the low-dimensional feature representation learned by the model and the potential feature ions into a Gaussian mixture model for clustering analysis, and manually selecting an optimal clustering model in combination with pathological annotation information to realize co-localization of the cluster of the region of interest and the feature ions. Compared with the prior art, the method has the advantages that efficient memory use and efficient peak value recognition are realized, and the clustering speed is higher while the accuracy is ensured.

Description

Technical Field

[0001] The present invention relates to the field of computer-aided medicine, and in particular to a mass spectrometry data processing and recognition method based on a fully connected variational neural network. Background Art

[0002] The study of biomolecular recognition and its molecular mechanisms is an important area of ​​life science research. It involves the study of molecular structure, mechanics and dynamics. These factors interact with each other to enable molecules to achieve efficient and highly selective binding and recognition. By studying the molecular mechanisms of biomolecular recognition, we can gain a deeper understanding of various biological processes in organisms, such as gene expression regulation and signal transduction. In addition, this research can also help develop new treatments and diagnostic tools, providing a scientific basis for the treatment and prevention of diseases.

[0003] Raw data Biomarkers are biochemical indicators that can mark changes or possible changes in the structure or function of systems, organs, tissues, cells and subcellular structures. These biomarkers are mostly derived from human tissues or body fluids and can reflect physiological or pathological processes, as well as the biological effects of exposure or therapeutic interventions.

[0004] Mass spectrometry imaging is an emerging technology with the potential to improve biomarker discovery, metabolomics research, drug application, and clinical diagnostics. With the advent of high-resolution mass spectrometry imaging, the volume and high-dimensional nature of the data still pose computational and storage complexities for existing data mining, clustering, and visualization methods, hindering the accurate identification of biologically relevant molecular patterns. However, the complexity of the data slows down processing time and poses computational and memory challenges that hinder the analysis of multiple samples required to extract biologically relevant patterns.

[0005] Peak picking is a common preprocessing step for analyzing raw data biomarkers. However, the selection of peak picking parameters is subjective and may affect subsequent clustering analysis and biological interpretation. In addition, after peak picking, the data still has high dimensional complexity, and existing data dimensionality reduction algorithms are still limited by computing memory and computing speed. Summary of the invention

[0006] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a mass spectrometry data processing and recognition method based on a fully connected variational neural network, avoiding the traditional time-consuming preprocessing, adopting a small batch processing strategy, realizing efficient use of memory, and adopting a peak learning method based on a neural network weight matrix to realize high-efficiency peak recognition. Cluster analysis is performed on low-dimensional feature representation based on a Gaussian mixture model, which ensures faster clustering speed while ensuring clustering accuracy.

[0007] The purpose of the present invention can be achieved by the following technical solutions:

[0008] A mass spectrometry data processing method based on a fully connected variational neural network comprises the following steps:

[0009] Step S1, obtaining mass spectrometry raw data and pathological annotation information, and performing data preprocessing;

[0010] Step S2, inputting the preprocessed data into an unsupervised learning model based on variational autoencoder to learn the low-dimensional feature representation of mass spectrometry data and realize data visualization;

[0011] Step S3, associating the low-dimensional feature representation learned by the model with the features in the original data to achieve peak learning and obtain a list of potential m / z ions related to tumor clusters or anatomy;

[0012] Step S4, inputting the low-dimensional feature representation and potential m / z ions learned by the model into a Gaussian mixture model (GMM) for cluster analysis, manually selecting the best clustering model in combination with pathological annotation information, and achieving co-localization of region of interest (ROI) clusters and characteristic ions;

[0013] Wherein, the unsupervised learning model is a fully connected variational neural network model.

[0014] Furthermore, in step S1, the obtained mass spectrometry raw data and pathological annotation information are in Imzml format, and data preprocessing includes:

[0015] Perform TIC normalization on the data;

[0016] Convert data in Imzml format to data in HDF5 format;

[0017] The intensity normalization operation is used to scale the pixel values ​​of the image block to [0, 1].

[0018] Furthermore, in step S2, the computing framework of the unsupervised learning model is based on a fully connected variational neural network with no less than five layers, and the neural network includes an input layer, a hidden layer and an output layer.

[0019] Furthermore, the feature learning of the unsupervised learning model includes:

[0020] The input of each layer is batch normalized before any neuron is activated to correct the covariate shift;

[0021] After batch normalization, the neurons in the output layer are activated using the sigmoid function, and the neurons in the remaining layers are activated using the linear rectification function (ReLU);

[0022] After the neuron is activated, the model learns the underlying nonlinear spectral structure of the input data through manifold learning, and optimizes the Kullback-Leibler divergence and variational autoencoder network loss function to minimize the error between the input data and the reconstructed data, thereby capturing accurate low-dimensional feature representations and correcting covariate shifts through batch regularization to improve the stability and convergence of model learning.

[0023] During the learning process, reconstruction loss, regularization loss and mean square error are used as evaluation indicators to optimize parameters.

[0024] Furthermore, the Kullback-Leibler divergence is always non-negative, and the formula of the Kullback-Leibler divergence is as follows:

[0025]

[0026] Among them, z is the latent variable, x is the input data, Φ is the parameter of the recognition model, θ is the parameter of the generation model, KL is the Kullback-Leibler divergence, q Φ (z|x) is the approximate posterior distribution of the latent variable z defined by the encoder parameters Φ, p θ (z|x) is the generative distribution of data x defined by the generative model parameters θ, KL(q Φ (z|x))||p θ (z|x) is q Φ (z|x) and p θ The Kullback-Leibler divergence between (z|x);

[0027] The formula of the variational autoencoder network loss function is as follows:

[0028]

[0029] in, is the variational lower bound, x (i) For a single data point, p θ (z) is the prior distribution of latent variables.

[0030] Furthermore, the reconstruction loss is used to evaluate the matching degree between the decoded sample and the input, and the regularization loss is used to evaluate the overfitting degree of the latent space to the training data. The smaller the mean square error value, the more accurate the model. The mean square error value formula is:

[0031]

[0032] Where MSE is the mean square error, n is the number of data points, and X i is the i-th original data, Xi′′ is the i-th predicted data.

[0033] Furthermore, in step S3, when implementing peak learning and identifying the potential m / z ion list, the low-dimensional feature representation learned by the model is associated with the features in the original data, and identification is performed based on the threshold analysis of back propagation of the neural network weight hyperparameters.

[0034] Furthermore, the threshold analysis includes:

[0035] Weight parameter threshold analysis, for each low-dimensional feature representation represented by the i-th neuron of layer h2, first identify the j-th neuron of the previous hidden layer h1 with the largest scaled weight value Then, using the weight vector Calculate the threshold T, which is a one-dimensional vector that holds the weights between all d neurons in the input layer and the jth neuron identified at the hidden layer; finally, a set of P neurons in the input layer with weights greater than T(w dj (1) ≥T), are retrieved and each represents an m / z variable, the formula is as follows,

[0036] T=mean(w dj (1) )+β*std(w dj (1) ),β∈[1,2.5];

[0037] Among them, w dj (1) represents the weights between all d neurons in the input layer and the jth neuron identified at the hidden layer. β is an adjustable parameter that gives the same threshold weight to each low-dimensional feature representation. Good performance is observed between [1, 2.5].

[0038] Local maximum identification, assigns the retrieved observed variable representing m / z bin to its associated m / z molecule, and identifies the feature m / z on the average spectrum as the local maximum closest to the m / z bin.

[0039] Furthermore, in step S4, an iterative method is used to perform cluster analysis, which enables the Gaussian mixture model clustering process to be applied to different numbers of clusters k (k-clusters) within an expected range, wherein the number of clusters k is a user-adjustable parameter.

[0040] Furthermore, in step S4, the co-localization between the region of interest clusters and the characteristic ions is measured by using a Pearson correlation coefficient, wherein the Pearson correlation coefficient is between 0 and 1, and the higher the coefficient, the higher the degree of co-localization between the region of interest clusters and the characteristic ions.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. In the design of the network model structure, the present invention proposes a neural network model for mass spectrometry image data analysis. This is a robust and universal probabilistic generation model based on a fully connected variational autoencoder, which effectively learns and visualizes the potential nonlinear spectral manifold of the original mass spectrometry image, avoiding the traditional time-consuming preprocessing. A small batch processing strategy is adopted to achieve efficient use of memory. In addition, the Kullback-Leibler divergence and the variational autoencoder cost function are optimized simultaneously to minimize the error between the original data and the reconstructed data, thereby capturing accurate low-dimensional feature representations, and batch regularization is used to correct covariate shifts to improve the stability and convergence of model learning;

[0043] 2. Based on the neural network weight matrix method, the low-dimensional feature representation learned by the model is associated with the original features to achieve peak learning and obtain a list of potential m / z ions related to tumor clusters or anatomy. This method can achieve high-efficiency peak identification;

[0044] 3. Perform cluster analysis on low-dimensional feature representation based on Gaussian mixture model. An iterative method is used to apply the Gaussian mixture model clustering process to different numbers of k-clusters within the expected range; the best clustering model is manually selected in combination with the pathological annotation results. The clusters of the region of interest are associated with a list of potential m / z ions with biological relevance to identify m / z ions co-localized with the region of interest, and the tumor-specific related metabolic characteristics are explored in combination with the pathological molecular identification results. In terms of time performance, the efficiency is improved by about 50% compared with clustering that requires peak picking, the clustering speed is faster, and the clustering results are highly consistent. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the process of the present invention;

[0046] Figure 2 This is a schematic diagram of the basic network framework structure in the embodiment;

[0047] Figure 3 Schematic diagram of an unsupervised learning framework structure of mass spectrometry image data based on variational autoencoder in an embodiment;

[0048] Figure 4 A schematic diagram of a threshold analysis structure based on back propagation of a neural network weight matrix according to the present invention;

[0049] Figure 5 It is a schematic diagram of a threshold analysis structure based on back propagation of a neural network weight matrix in an embodiment. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0051] A mass spectrometry data analysis method based on a fully connected variational neural network according to an embodiment of the present invention is as follows: Figure 1 As shown, the following steps are included:

[0052] Step 1, obtaining mass spectrometry raw data and pathological annotation information in the data format of Imzml, and performing data preprocessing;

[0053] In this example, four types of public data sets were used, three of which were 2D slice mass spectrometry image data sets, and one was a 3D slice data set containing four 2D slices, from human and mouse models. Pathological information included H&E staining annotation results and HMDB database molecular matching results or secondary mass spectrometry standard molecular identification results.

[0054] Preferably, the preprocessing method is as follows:

[0055] (1) The original data were TIC standardized using the Cardinal package in the R language;

[0056] (2) Use the python h5py library to convert the data in the Imzml standard format into the hdf5 format and save the .h5 file;

[0057] (3) Intensity normalization is used to scale the pixel values ​​of the image blocks to [0, 1] to facilitate network processing and analysis.

[0058] Step 2: Input the preprocessed data into an unsupervised learning model based on variational autoencoder to learn the low-dimensional feature representation of mass spectrometry data and realize data visualization;

[0059] Figure 2 Schematic diagram of the basic network framework structure of the embodiment of the present invention. The probability encoder is used as the recognition model q Φ (z|x) to infer the true but intractable distribution p of the k-dimensional latent variable (z) θ Approximate estimate of (z|x) (low-dimensional feature representation). Mathematically, the spectrum of mass spectrometry image data can be represented as a set of high-dimensional vectors, and the calculation formula 1 is:

[0060]

[0061] Among them, the data point x (i) is the conditional distribution p θ (z|x) is generated. The true posterior distribution p θ (z|x) will provide a compressed representation of the observed high-dimensional data, i.e., a low-dimensional feature representation. (i) The high dimensional complexity of the posterior distribution makes it computationally intractable. Therefore, we introduce q Φ (z|x) identification model to approximate the true intractable posterior p θ (z|x). First, assume that the recognition model is derived from μ Φ (x) and σ Φ (x) is sampled from a normal distribution parameterized by the latent variable (z). Inference of the latent variable (z) will make the marginal likelihood estimator p θ (z|x) reconstructs data point x (i) The calculation formula 2 is:

[0062]

[0063] Among them, the variational parameter Φ needs to be estimated because Φ makes q Φ (z|x) is as close as possible to the true posterior p θ (z|x). A reparameterization technique is used to decode the sample points and finally calculate the reconstruction loss, which is back-propagated through the network. First, an auxiliary variable is merged With the input data point x (i) Form a continuous function g Φ (ε,x (i) ), which can then be used to approximate the posterior z (i) ~q Φ (z,|,x (i) ) to sample the latent variables. The calculation formula 3 is:

[0064] z (i) =g Φ (ε,x (i) )=μ Φ (x (i) )+diag(σ Φ (x (i) )).ε

[0065] During the training process, reconstruction loss and regularization loss are often used for parameter optimization. Reconstruction loss evaluates whether the decoded samples match the input, and regularization loss evaluates whether the latent space is overfitting the training data.

[0066] Figure 3This is a schematic diagram of the framework structure of unsupervised learning of mass spectrometry image data based on variational autoencoder according to an embodiment of the present invention. In this embodiment, a fully connected neural network consisting of an input layer, three hidden layers (h1, h2 and h3) and an output layer is used. The number of neurons in the input layer and output layer of the model is the number of m / z bins. The number of neurons in h1, h2 and h3 is set to 512, 5, 512 respectively. The low-dimensional feature representation compressed in the 5-dimensional space of the mass spectrometry image data is captured in the hidden layer h2 and then generated in the output layer by the generative model p. θ (z|x) is used to reconstruct data given only low-dimensional feature representations. Batch normalization is applied to the input of each layer before any neuron activation to correct for covariate shifts that degrade the learning process and improve the stability and convergence of model learning. After batch normalization, except for the neurons in the output layer, which are activated by the sigmoid function, the linear rectification function (ReLU) is used to activate neurons in other layers, while minimizing the reconstruction loss between the original data and the reconstructed data by optimizing the Kullback-Leibler divergence and the VAE cost function, thereby capturing accurate low-dimensional feature representations.

[0067] For continuous data, Gaussian distribution is used and MSE is used as the loss function. MSE is the mean of the sum of squares of the errors between the predicted data and the original data, which can be used to evaluate the quality of data fitting. The smaller the MSE value, the better the predicted data (X i″ ) more accurately describes the original data (X i ). Calculation formula 4 is:

[0068]

[0069] Before model training, the cost function is optimized based on minibatch processing, the batch size is set to 64, the adaptive stochastic gradient descent method optimized by Adam is used, the default learning rate is 0.001, and the epoch is set to 100. During the training process, reconstruction loss and regularization loss are often used for parameter optimization. Reconstruction loss is used to evaluate whether the decoded samples match the input, and regularization loss is used to evaluate whether the latent space is overfitting the training data.

[0070] The initial training epoch is 200. When the model converges within 100 training times and the learning rate is the default value of 0.001, the effect is better.

[0071] The preliminary results of the pre-experiment on low-dimensional feature representation learning based on variational autoencoder neural network are analyzed, and the heavy parameters of the learning model are determined. The pre-experimental results of four types of data sets are shown in Table 1.

[0072] Table 1 Preliminary experimental results of low-dimensional feature representation learning model on four types of datasets

[0073]

[0074] It can be seen from Table 1 that when the number of neurons in h1, h2 and h3 of the network are 512, 5, 512 respectively, the effect is better.

[0075] Step 3: Associating the low-dimensional feature representation learned by the model with the features in the original data to achieve peak learning and obtain a list of potential feature ions related to tumor clusters or anatomy;

[0076] Step 31, weight parameter threshold analysis. For each low-dimensional feature representation represented by the i-th neuron of layer h2, the weight parameter W of the input layer and output layer of the neural network is (L) Perform threshold analysis, such as Figure 4 First, identify the jth neuron of the previous hidden layer h1 with the largest scaled weight value See also Figure 5 The red line in ; then, using the weight vector Calculate the threshold T, which is a one-dimensional vector that holds the weights between all d neurons in the input layer and the jth neuron identified at the hidden layer; finally, a set of P neurons in the input layer with weights greater than T(w dj (1) ≥T), are retrieved and each represents an m / z variable, see Figure 5 The blue line in the equation 4 is:

[0077] T=mean(w dj (1) )+β*std(w dj (1) ),β∈[1,2.5]

[0078] Among them, the parameter β shows that the same threshold weight is given to each low-dimensional feature representation. Good performance is observed between [1, 2.5]. In this example, β is set to 2.5 for peak learning.

[0079] Step 32, local maximum identification. Since the raw data is analyzed without prior peak picking preprocessing, the retrieved observation variables represent m / z bins, which need to be assigned to their associated m / z molecules, and the characteristic m / z on the average spectrum is identified as the local maximum closest to the m / z bin.

[0080] In step 4, the low-dimensional feature representation and potential characteristic ions are input into the Gaussian mixture model for cluster analysis, and the best clustering model is manually selected in combination with the pathological annotation information to achieve co-localization of the region of interest clusters and characteristic ions.

[0081] In summary, the embodiment of the present invention automatically extracts relevant features of the original mass spectrometry image data through a variational autoencoder neural network to achieve the purpose of exploring tumor-specific metabolic features. Compared with the method that requires peak picking, this method avoids time-consuming traditional preprocessing steps that may affect subsequent analysis results, and the recognition and analysis speed is faster. Compared with the manual annotation method, this method avoids subjective bias and has a higher efficiency in biomarker identification. At the same time, this example proves the robustness of this method when applied to different tissue types through experimental results of data sets of different types and resolutions.

[0082] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for mass spectrometry data processing and recognition based on a fully connected variational neural network, characterized in that it comprises the following steps: Step S1, obtaining mass spectrometry raw data and pathological annotation information, and performing data preprocessing; Step S2, inputting the preprocessed data into an unsupervised learning model based on variational autoencoder to learn the low-dimensional feature representation of mass spectrometry data and realize data visualization; Step S3, associating the low-dimensional feature representation learned by the model with the features in the original data to achieve peak learning and obtain a list of potential m / z ions related to tumor clusters or anatomy; Step S4, inputting the low-dimensional feature representation and potential m / z ions learned by the model into a Gaussian mixture model for cluster analysis, and manually selecting the best clustering model in combination with the pathological annotation information to achieve co-localization of the region of interest clusters and characteristic ions; Wherein, the unsupervised learning model is a fully connected variational neural network model.

2. The method for mass spectrometry data processing and recognition based on a fully connected variational neural network according to claim 1, characterized in that: In step S1, the obtained mass spectrometry raw data and pathological annotation information are in Imzml format, and data preprocessing includes: Perform TIC normalization on the data; Convert data in Imzml format to data in HDF5 format; The intensity normalization operation is used to scale the pixel values ​​of the image block to [0, 1].

3. The method for mass spectrometry data processing and recognition based on a fully connected variational neural network according to claim 1, characterized in that: In step S2, the computing framework of the unsupervised learning model is based on a fully connected variational neural network with no less than five layers, and the neural network includes an input layer, a hidden layer and an output layer.

4. The method for mass spectrum data processing and recognition based on a fully connected variational neural network according to claim 3, characterized in that: The feature learning of the unsupervised learning model includes: The input of each layer is batch normalized before any neuron is activated to correct the covariate shift; After batch normalization, the neurons in the output layer are activated using the sigmoid function, and the neurons in the remaining layers are activated using the linear rectification function; After the neuron is activated, the model learns the underlying nonlinear spectral structure of the input data through the manifold, while optimizing the Kullback-Leibler divergence and the variational autoencoder network loss function to minimize the error between the input data and the reconstructed data, and finally captures accurate low-dimensional feature representation; During the learning process, reconstruction loss, regularization loss and mean square error are used as evaluation indicators to optimize parameters.

5. The method for mass spectrometry data processing and recognition based on a fully connected variational neural network according to claim 4, characterized in that: The Kullback-Leible r The divergence is always non-negative, Kullback-Leible r The formula for divergence is as follows: Among them, z is the latent variable, x is the input data, Φ is the parameter of the recognition model, θ is the parameter of the generation model, KL is the Kullback-Leibler divergence, q Φ (z|x) is the approximate posterior distribution of the latent variable z defined by the encoder parameters Φ, p θ (z|x) is the generative distribution of data x defined by the generative model parameters θ, KL(q Φ (z|x))||p θ (z|x) is q Φ (z|x) and p θ The Kullback-Leibler divergence between (z|x); The formula of the variational autoencoder network loss function is as follows: in, is the variational lower bound, x(i) is a single data point, and pθ(z) is the prior distribution of the latent variable.

6. The method for mass spectrum data processing and recognition based on a fully connected variational neural network according to claim 4, characterized in that: The reconstruction loss is used to evaluate the matching degree between the decoded sample and the input, and the regularization loss is used to evaluate the overfitting degree of the latent space to the training data. The smaller the mean square error value, the more accurate the model. The formula for the mean square error value is: Where MSE is the mean square error, n is the number of data points, and X i is the i-th original data, X l′′ is the i-th predicted data.

7. The method for mass spectrometry data processing and recognition based on a fully connected variational neural network according to claim 1, characterized in that: In step S3, when implementing peak learning and identifying the potential m / z ion list, the low-dimensional feature representation learned by the model is associated with the features in the original data, and identification is performed based on the threshold analysis of back propagation of the neural network weight hyperparameters.

8. The method for mass spectrum data processing and recognition based on a fully connected variational neural network according to claim 7, characterized in that: The threshold analysis includes: Weight parameter threshold analysis, the formula is as follows, T=mean(w dj (1) +β*std(w dj (1) ,β∈[1,2.5]; Among them, w dj (1) represents the weights between all d neurons in the input layer and the jth neuron identified at the hidden layer, and β is an adjustable parameter that gives the same threshold weight to each low-dimensional feature representation; Local maximum identification, assigns the retrieved observation variables representing m / z bins to their associated m / z molecules, and identifies the feature m / z on the average spectrum as the local maximum closest to the m / z bin.

9. The method for mass spectrometry data processing and recognition based on a fully connected variational neural network according to claim 1, characterized in that: In step S4, cluster analysis is performed using an iterative method that enables the Gaussian mixture model clustering process to be applied to different numbers of cluster numbers k within an expected range, wherein the number of clusters k is a user-adjustable parameter.

10. The method for mass spectrometry data processing and recognition based on a fully connected variational neural network according to claim 1, characterized in that: In step S4, the co-localization between the region of interest clusters and the characteristic ions is measured by the Pearson correlation coefficient, wherein the Pearson correlation coefficient is between 0 and 1, and the higher the coefficient is, the higher the degree of co-localization between the region of interest clusters and the characteristic ions is.

Images