Machine learning-based RE analysis and FTICRMS joint data decoding and optimizing method

Abstract

The invention discloses an RE analysis and FTICRMS joint data decoding and optimizing method based on machine learning. Comprising the steps of obtaining and preprocessing a training data set, constructing a convolutional neural network and decoding data, training an RE data decoding model, training an MS data decoding model, evaluating and optimizing data set quality, constructing an MS-FTICRMS joint algorithm, and testing and evaluating the model. According to the method, the MS-RE decoding model and the MS-MS decoding model are constructed and combined, so that the models can process high-resolution data and low-resolution data at the same time, the expression ability of the models is improved through multi-task learning, and the precision and efficiency of mass spectrum data decoding are greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of protein mass spectrometry analysis, and in particular relates to a RE analysis and FTICRMS joint data decoding and optimization method based on machine learning. Background Art

[0002] Mass spectrometry (MS) technology is one of the core tools for detecting and identifying molecular mass. In particular, Fourier transform ion cyclotron resonance mass spectrometry (FTICRMS) is widely used in the analysis of complex biomolecules, proteins, and metabolites due to its high resolution and high precision. On the other hand, reaction site (RE) analysis is crucial for understanding the three-dimensional structure, function, and interaction of proteins. In the context of large-scale biological data, how to efficiently and accurately integrate and decode mass spectrometry data and reaction site information has become an important challenge in current data analysis technology. The development of machine learning, especially deep learning technology, is bringing new opportunities to this field, improving the accuracy and efficiency of data decoding through automation and intelligent methods.

[0003] In traditional mass spectrometry analysis, mass spectrometry data is usually processed and decoded by manual or semi-automatic methods. For example, the data output by the mass spectrometer needs to go through multiple steps of manual screening, impurity removal, peak identification, etc. These steps are time-consuming and prone to human errors. The analysis of reaction sites generally relies on structural biology methods, such as X-ray crystallography. Although these methods can provide accurate three-dimensional structural information, the data processing is cumbersome and often requires the help of expert experience to complete. With the continuous increase in the scale of mass spectrometry data and biological data, traditional manual processing methods have been unable to cope with the surge in data volume and the complexity of analysis requirements. In addition, traditional mass spectrometry decoding often ignores the integrated analysis of different data sources (such as reaction site data and mass spectrometry data), which limits the comprehensiveness and accuracy of data decoding.

[0004] Although traditional mass spectrometry and reaction site analysis techniques have achieved certain results in the past, they still have many shortcomings in efficiency and accuracy. First, manual decoding of mass spectrometry data is inefficient, and the data processing speed is difficult to match the high-throughput requirements of modern scientific research. Second, manual removal of impurities and noise in mass spectrometry data is not only time-consuming, but also prone to bias, resulting in reduced accuracy of decoding results. Finally, traditional methods often process mass spectrometry data and reaction site information separately, lacking a unified model to integrate the two, thus affecting the overall decoding effect. Summary of the invention

[0005] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and to provide a method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning. By constructing an MS-RE decoding model combined with an MS-MS decoding model, the model can process high-resolution and low-resolution data at the same time, and improve the performance of the model through multi-task learning, thereby greatly improving the accuracy and efficiency of mass spectrometry data decoding.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning, comprising the following steps:

[0008] Collecting protein crystal structure data and mass spectrometry data, preliminarily screening and preprocessing the data to obtain a training data set; the training data set includes reaction site data and a mass spectrometry subgraph data set;

[0009] Constructing a data decoding model, training the data decoding model using reaction site data, automatically extracting and learning key local features of the reaction site data, obtaining a first data decoding model, training the data decoding model using a mass spectrum subgraph data set, learning different protein molecular features, and obtaining a second data decoding model;

[0010] Evaluate the first data decoding model through the characteristic curve to obtain a first evaluation result; comprehensively evaluate the second data decoding model through the quality evaluation index to obtain a second evaluation result, wherein the quality evaluation index includes resolution and peak characteristics; optimize the model using the first evaluation result and the second evaluation result respectively to obtain the best first data decoding model and second data decoding model;

[0011] Combine the best second data decoding model with the best first data decoding model to obtain a first combined decoding model, combine multiple best second data decoding models to obtain a second combined decoding model, and construct a joint decoding model using the first combined decoding model as a backbone and the second combined decoding model as an additional module, wherein an output end of the second combined decoding model is connected to an output end of the first combined decoding model;

[0012] The joint decoding model is tested using a given independent test set, and the decoding performance is evaluated using actual mass spectrometry data and protein structure data. The best joint decoding model is obtained based on the evaluation results, and the mass spectrometry data of the protein is decoded using the best joint decoding model.

[0013] As a preferred technical solution, the preliminary screening and preprocessing of data includes:

[0014] Extract spatial position information from protein crystal structure data through X-ray crystallography technology to obtain protein reaction site data;

[0015] Preliminarily screening the mass spectrometry data by comparing the protein structure to obtain a mass spectrometry data set, including an ion-free and solvent-free molecule set, a non-bridging data set, and a bridging data set, wherein the ion-free and solvent-free molecule set is a pure protein structure, and the non-bridging data set and the bridging data set have impurities;

[0016] The mass spectrometry data set is cut into fixed-size sub-graphs using a window algorithm;

[0017] The mass-to-charge ratios of water molecules, solvent molecules and other interfering data in the non-bridging data set and the bridged data set are identified by analyzing the characteristic peaks, and the water molecules, solvent molecules and other interfering data are filtered out as follows:

[0018]

[0019] Here, let the mass spectrum data be M, and the mass-to-charge ratio of the i-th data point of the mass spectrum data be m / z i , M water is the mass of water molecules, ∈ is an allowable error range, and the other interfering data include lipids, salt ions, additives or other impurity molecules.

[0020] As a preferred technical solution, the data decoding model includes N convolution blocks and a fully connected layer, the output of the convolution block is connected to the input of the next convolution block, and the output of the Nth convolution block is connected to the fully connected layer; the convolution block includes a convolution layer and a pooling layer, and the output of the convolution layer is connected to the pooling layer.

[0021] As a preferred technical solution, the evaluating the first data decoding model by using the characteristic curve includes:

[0022] Each sample in the reaction site data is predicted by different decision thresholds. For each threshold, the corresponding true positive rate and false positive rate are calculated. According to the true positive rate and false positive rate of each threshold, the ROC curve is generated;

[0023] The AUC value is calculated based on the ROC curve to evaluate the robustness of the current model and generate a first evaluation result.

[0024] As a preferred technical solution, the comprehensive evaluation of the second data decoding model by quality evaluation indicators includes:

[0025] Calculate the resolution and signal-to-noise ratio of the data based on the mass spectrum sub-image data set;

[0026] Analyze the peak features of the mass spectrum sub-graph, obtain the sample peak features, calculate the matching degree between the sample peak features and the target peak features, and mark the data with low matching degree as low-quality data; the sample peak features include peak shape, peak position and peak intensity.

[0027] As a preferred technical solution, the calculation of the matching degree is as follows:

[0028]

[0029] Among them, I observed is the observed peak intensity, I expected is the expected peak intensity and N is the number of peaks.

[0030] As a preferred technical solution, the second evaluation results are used to optimize the models respectively, specifically:

[0031] Exclude low-quality data;

[0032] Add low-quality data to the training dataset.

[0033] As a preferred technical solution, the first combined decoding model is used as the backbone part, and the second combined decoding model is used as an additional module to construct a joint decoding model, specifically:

[0034] The sample data is initially decoded by a first combined decoding model, and the sample data is refined by a second combined decoding model. The initial decoding result and the refined decoding result are concatenated to obtain a joint data decoding model; the second combined decoding model has at least one.

[0035] As a preferred technical solution, the evaluation of decoding performance using actual mass spectrometry data and protein structure data is specifically as follows:

[0036] The test output results are compared with the real mass spectrometry data, the comparison results are evaluated through performance indicators, and the best joint decoding model is screened based on the evaluation results; the performance indicators ROC curve and its AUC value, accuracy, recall rate and F1 value.

[0037] As a preferred technical solution, it also includes optimizing the joint decoding model by using hyperparameter optimization and transfer learning; the hyperparameter optimization includes grid search and Bayesian optimization.

[0038] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0039] (1) The present invention decodes mass spectrometry data and reaction site data by constructing a data decoding model. The algorithm can effectively extract local features in the data and further enhance the analysis capability of complex data through a multi-layer structure. This method can greatly improve the accuracy of mass spectrometry data decoding, especially in the detection of complex protein molecules and connection points.

[0040] (2) The present invention ensures the quality of input data by preprocessing and screening mass spectrometry data, especially removing interfering factors such as water molecules and solvent molecules, so that the model can learn important features from pure data. This data quality optimization process lays the foundation for efficient decoding of the model and improves the accuracy and robustness of the model.

[0041] (3) The present invention combines the MS-RE decoding model with the MS-MS decoding model, so that the model can process high-resolution and low-resolution data at the same time, and improves the performance of the model through multi-task learning. This multi-task integration enables the algorithm to maintain high decoding performance and consistency on different types of data. At the same time, because the algorithm performs multi-task processing on mass spectrometry data through the joint decoding model, it can adapt to different types of mass spectrometry data decoding tasks. Therefore, the scheme is not only applicable to current mass spectrometry decoding, but can also be extended to a wider range of mass spectrometry analysis and decoding fields, and has high application flexibility and extensiveness.

[0042] (4) The present invention introduces noise data during data preprocessing and decoding, so that the model has stronger noise resistance. In practical application scenarios, mass spectrometry data is often subject to various interferences. The algorithm can effectively deal with these interferences and improve the robustness of the model. At the same time, the introduction of transfer learning enables the model to quickly adapt to new mass spectrometry data tasks, improving its generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0044] Figure 1 It is a flow chart of the RE analysis and FTICRMS joint data decoding and optimization method based on machine learning according to an embodiment of the present invention;

[0045] Figure 2 This is a data preprocessing flow chart of an embodiment of the present invention;

[0046] Figure 3This is a schematic diagram of the data decoding model structure using a CNN neural network in an embodiment of the present invention;

[0047] Figure 4 This is a flow chart of data set quality assessment according to an embodiment of the present invention;

[0048] Figure 5 This is a structural diagram of a joint decoding model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0050] Reference to "embodiments" in this application means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments.

[0051] Example 1

[0052] See also Figure 1 ,This embodiment provides a method for joint data decoding and optimization of RE analysis and FTICRMS based on machine learning, which includes the following stages: training data set acquisition and preprocessing stage, model building and training stage, data set quality assessment and optimization stage, joint model building stage, and model testing and evaluation stage.

[0053] S1: Training data set acquisition and preprocessing stage.

[0054] In the training data set acquisition and preprocessing stage, first, this embodiment extracts three-dimensional crystal structure data from the PDB protein database as the main source of reaction site (RE) data. The PDB database contains three-dimensional crystal structures of different proteins, which are obtained through techniques such as X-ray crystallography and can provide spatial position information of proteins. Secondly, this embodiment obtains mass spectrometry data from the Acpdb database, where the Acpdb database contains a variety of protein mass spectrometry data, which are obtained through a mass spectrometer and are a collection of information such as molecular mass and charge ratio.

[0055] After completing the data collection, the mass spectrometry data is preliminarily screened by comparing the mass spectrometry data with the protein structure to obtain mass spectrometry data sets, including ion-free and solvent-free molecule sets, non-bridging data sets, and bridged data sets. The ion-free and solvent-free data sets represent pure protein structures without other impurities. The non-bridging data sets refer to pure protein molecule data, which can provide clear mass spectrometry maps and help accurately train the model. The bridged data sets contain impurities or other molecules bound to proteins. Although they can increase the model's ability to handle complex samples, they may also introduce noise. The non-bridging and bridged data sets distinguish whether there are connections between different proteins. Such a distinction helps optimize the model training process and ensure the accuracy and reliability of the decoding results.

[0056] Then, if Figure 2 As shown, in order to make these data better suited to the training of the neural network model, the raw data needs to be converted into a standardized data form suitable for model input. The mass spectrometry data is preprocessed as follows:

[0057] S11. Use the window algorithm to cut the mass spectrometry data set into sub-graphs of fixed size.

[0058] Mass spectrometry data is usually a two-dimensional or three-dimensional matrix data. By cutting it into fixed-size subgraphs, it can be adapted to the input format of the convolutional neural network. The cutting method can be implemented through a window-based algorithm. For example, given a mass spectrometry matrix of size n×n, a sliding window can be used, starting from the upper left corner, moving a fixed step size, and gradually extracting overlapping or non-overlapping sub-matrices. Assuming the step size is s and the window size is k×k, the number of subgraphs generated can be determined by the following formula.

[0059] Assuming the size of the original matrix is ​​m×m, the number of subgraphs is:

[0060]

[0061] This formula indicates that when extracting subgraphs through a sliding window, the number of subgraphs generated in each row and column can be calculated based on the window size k and the step size s, and then the total number can be obtained. In actual operation, m represents the row and column dimensions of the mass spectrometry data, k represents the size of the subgraph, and s represents the distance of each movement. In order to improve the training efficiency of the model, the size of the subgraph should be consistent, and the size that matches the network structure is usually selected, such as 64×64 or 128×128.

[0062] S12. Identify the mass-to-charge ratios corresponding to the water molecules, solvent molecules and other interfering data in the non-bridging data set and the bridge data set by analyzing characteristic peaks, and filter the water molecules, solvent molecules and other interfering data.

[0063] In the process of removing water molecules and solvent molecules, the mass-to-charge ratios corresponding to water molecules and solvent molecules can be identified by analyzing the characteristic peaks in the mass spectrometry data. The peaks of the mass spectrometry data represent different fragments in the protein. By analyzing these peaks, specific mass-to-charge ratio intervals can be filtered out to remove water molecules or solvent molecules. Assuming that the mass of the water molecule is 18 Da and the mass of the solvent molecule is W Da, the peak corresponding to the mass-to-charge ratio can be removed from the mass spectrum by setting the mass-to-charge ratio threshold filter range. Let the mass spectrometry data be M, where the mass-to-charge ratio of the i-th data point is m / z i , you can filter by the following conditions:

[0064]

[0065] Among them, M water is the mass of water molecules, and ∈ is an allowable error range. Similarly, solvent molecules can be screened to generate mass spectrometry data containing only protein structures.

[0066] There are other interfering data including lipids, salt ions, additives or other impurity molecules. The process of removing these interfering data is usually filtered by setting the allowed mass-to-charge ratio range, combined with signal-to-noise ratio analysis to exclude low-quality signals. In addition, by comparing with a known mass spectrum database, unmatched spectral data is identified and removed, and finally feature analysis is used to confirm and exclude peaks unrelated to protein structure, thereby ensuring the purity of the data set and the accuracy of the analysis.

[0067] S2: Model building and training phase.

[0068] The present embodiment constructs a data decoding model by establishing a CNN neural network. Convolutional neural network (CNN) is used to process the mass spectrometry data image after pretreatment. The input of the network is the mass spectrometry data image cut after pretreatment, and the output is the result of data quality optimization. When constructing the CNN model, the key is to determine the appropriate network structure and select the number and size of suitable convolutional layers, pooling layers and fully connected layers. In addition, the selection of optimizers (such as Adam or RMSProp) and the selection of loss functions (such as mean square error or cross entropy) also play a vital role in the convergence and performance of the model. In this process, it is also necessary to adjust the hyperparameters of the model, such as batch size, learning rate and training iterations. Through these adjustments, the model can automatically extract features from mass spectrometry data and enhance the ability to decode data. To prevent the model from overfitting problems, L2 regularization or Dropout technology is usually introduced.

[0069] The structure of the data decoding model constructed in this embodiment includes N convolution blocks and a fully connected layer. The output of the convolution block is connected to the input of the next convolution block, and the output of the Nth convolution block is connected to the fully connected layer. The convolution block includes a convolution layer and a pooling layer, and the output of the convolution layer is connected to the pooling layer. In a more specific example, Figure 3 As shown, two convolution blocks are used here.

[0070] After completing the construction of the basic data decoding model, the basic model will be trained at two levels, namely RE data decoding model training and MS data decoding model training.

[0071] For RE data decoding model training, we first need to build an RE database to provide sufficient training samples. Secondly, based on the convolutional neural network (CNN), the model can automatically extract and learn key features from these RE data. Specifically, the convolution layer in the model extracts local features from the data through weighted convolution operations, while the pooling layer is used for downsampling to reduce the amount of calculation. Assuming the input data is x and the convolution kernel is w, the output y of the convolution operation can be expressed as:

[0072] y=f(x*w+b)

[0073] Among them, * represents the convolution operation, b is the bias term, and f is the activation function. During the training process of the convolutional neural network, the back propagation algorithm is used to optimize the model, and the weights and biases of the convolution kernel are continuously adjusted by minimizing the loss function.

[0074] Similar to the RE data decoding model training, the MS data decoding model training is completed by building an MS database. In the MS data preprocessing, the MS mass spectrometry data is first processed and converted into an image format suitable for CNN model training. This step ensures that the model can learn different molecular features from the mass spectrometry data and build an effective MS data decoding model on this basis. Similar to the RE model, the MS model is also evaluated through ROC curves and AUC values. In addition, the cross-validation method can be used to further verify the performance of the model according to different mass spectrometry data types to ensure the generalization ability of the model.

[0075] S3, dataset quality assessment and optimization stage.

[0076] After completing the training in step S2 above, quality evaluation and optimization of the RE model and MS model will be carried out to ensure the effectiveness and robustness of the model.

[0077] (1) Quality evaluation and optimization of RE model.

[0078] This embodiment evaluates the quality of the RE model through the ROC curve and AUC value. The ROC curve reflects the relationship between the true positive rate and the false positive rate of the model. The ROC curve is generated based on the performance of the model in classifying samples at different thresholds.

[0079] First, the model predicts each sample through different decision thresholds. For each threshold, the corresponding true positive rate (TPR) and false positive rate (FPR) are calculated. The true positive rate (TPR) is defined as the proportion of samples correctly classified as positive to the actual positive samples, while the false positive rate (FPR) is defined as the proportion of negative samples incorrectly classified as positive to the actual negative samples. The formula is as follows:

[0080] True Positive Rate (TPR):

[0081] False Positive Rate (FPR):

[0082] Among them, TP is true positive, FP is false positive, FN is false negative, and TN is true negative.

[0083] In this step, the reaction site data is predicted by setting different decision thresholds. For each threshold, the true positive rate TPR and false positive rate FPR of the model are calculated. These calculation results are used to draw the receiver operating characteristic curve (ROC curve), which shows the performance of the model at different thresholds and helps to intuitively evaluate the classification performance of the model. By analyzing the ROC curve, the optimal threshold can be determined, thereby optimizing the predictive ability of the model.

[0084] Secondly, the AUC value is calculated based on the obtained ROC curve, and the AUC value is used to determine whether the currently trained RE is effective and robust, and the best RE model is selected based on the judgment result. The AUC value is the area under the ROC curve, and its value range is [0,1]. The closer the AUC value is to 1, the better the performance of the model is, and it can correctly distinguish positive and negative samples; conversely, the closer the AUC value is to 0, the worse the model performance is. When the AUC value is 0.5, it means that the model has no classification ability, and its performance is no different from random guessing; when the AUC value is greater than 0.5 and less than 1, the performance of the model is better than random guessing, but the specific degree of excellence needs to be judged according to the size of the AUC value; when the AUC value is equal to 1, it means that the model has perfect classification ability, that is, all positive examples are correctly predicted as positive examples, and all negative examples are correctly predicted as negative examples.

[0085] (2) Quality evaluation and optimization of MS models.

[0086] like Figure 4As shown in the figure, data set quality assessment and optimization are key steps in the mass spectrometry data decoding algorithm, which is directly related to the accuracy and stability of the model. In the MS mass spectrometry data decoding process, the data volume is huge and complex, and a strict quality assessment procedure must be used to ensure the validity of the data. This assessment procedure mainly conducts a comprehensive analysis of the high-resolution data in the RE database and MS database based on multiple indicators such as the resolution, peak intensity, and signal-to-noise ratio of the mass spectrometry data.

[0087] First, the resolution of the data. The resolution of the data directly affects the accuracy of the decoding. The higher the resolution, the more details can be captured. Therefore, the evaluation program ensures the accuracy of high-resolution data by calculating the resolution of the data.

[0088] Second, peak characteristics. The evaluation of peak characteristics is to analyze the shape, position and intensity of each peak in the mass spectrum to determine whether it is consistent with the expected biomolecular structure. By calculating the matching degree of peak characteristics, the program can find abnormal data and mark it as low-quality data. The formula is as follows:

[0089]

[0090] Among them, I observed is the observed peak intensity, I expected is the expected peak intensity, and N is the number of peaks. Data with low matching scores will be marked as "dissimilar".

[0091] For these low-quality data, the evaluation program can take two approaches: first, exclude them from the training data set to improve the training quality of the model. Second, generate and add a certain amount of low-quality data to the training set so that the model can perform more robustly in the face of noisy data. This method enables the model to adapt to complex situations in practical applications and improve decoding accuracy and anti-interference ability. For example, in protein mass spectrometry data, adding a small amount of noise samples can help the model better deal with data anomalies in real scenarios.

[0092] S4: Joint model building phase.

[0093] First, in order to solve the backbone decoding task and the enhanced decoding task, this embodiment combines the best RE model and MS model. In order to achieve the backbone decoding task, the RE model and the MS model are combined. The output of the MS model can provide basic information, and the RE model can deduce or predict the target based on these basic information, thereby inferring the reaction active site of the molecule through the mass spectrum information; in order to achieve the enhanced decoding task, multiple MS models are combined, and the fragment relationship in the molecular structure is analyzed layer by layer through the mass-to-charge ratio information of the parent ion and the daughter ion. In this way, the MS-RE decoding model and the MS-MS decoding model are obtained respectively.

[0094] Next, start building a joint decoding model. The construction of the MS-FTICRMS joint algorithm is a key step in this solution. This embodiment uses the MS-RE decoding model as the backbone and multiple MS-MS decoding models as additional modules to construct a joint data decoding model. The joint decoding model can process both high-resolution and low-resolution mass spectrometry data, and use the MS quality assessment program to screen the decoding results. Specifically, the data is preliminarily decoded by the MS-RE decoding model, and then the results of the MS-MS decoding model are used for further refinement. "Refinement" refers to further fine-tuning the preliminary decoding results using the "results of the MS-MS decoding model", rather than simply integrating them. This process analyzes and optimizes the output of the "MS-MS decoding model" to improve the accuracy and details of the decoding. Therefore, Figure 5 The “refined decoding” in should be more clearly defined as the optimization of the preliminary decoding results, rather than the simple integration of the two results. Figure 5 The two "refined decoding" appearing in the text respectively mean: the first one is the optimization of the preliminary decoding results, and the second one is a deeper refinement on the integrated joint decoding model to ensure the accuracy and completeness of the final decoding.

[0095] By building a joint model, we can effectively transform low-resolution data to have performance capabilities close to that of high-resolution data. This process combines the dual tasks of data feature extraction and quality assessment to ensure the accuracy and consistency of the decoding results. During the training process, the joint model is trained using a multi-task learning framework. The backbone of the model is composed of the MS-RE decoding task, while the MS-MS task is an additional module that further enhances the performance of the model.

[0096] S5: Model testing and evaluation phase.

[0097] like Figure 5 As shown in the figure, during the model testing and evaluation stage, the constructed MS-FTICRMS joint decoding model was tested using an independent test set, and the decoding performance was evaluated using actual mass spectrometry data and protein structure data. By comparing the model output results with the actual mass spectrometry data, the accuracy and efficiency of the model in decoding mass spectrometry data can be obtained. In addition, by comparing with the existing benchmark model, the advantages and disadvantages of the MS-FTICRMS joint decoding model can be fully evaluated. For the performance evaluation of the model, in addition to the ROC curve and AUC value, other evaluation indicators such as accuracy, recall rate, F1 value, etc. can also be considered to obtain more comprehensive performance feedback.

[0098] In order to ensure the performance of the joint decoding model in practical applications, hyperparameter optimization and transfer learning can be further introduced. The core of hyperparameter optimization is to find the optimal combination of model parameters, which can be achieved through grid search or Bayesian optimization. Grid search is an exhaustive method that tries every possible parameter combination, calculates the performance of the model under each combination, and selects the optimal combination. Assuming there are two hyperparameters: learning rate and batch size, the grid search will traverse all combinations within different value ranges of these parameters and evaluate their performance. Bayesian optimization is to gradually approach the optimal parameters by building a probabilistic model, which is more efficient. The formula is as follows:

[0099] x next =argmax(μ(x)+kσ(x))

[0100] Among them, μ(x) is the mean of the objective function, σ(x) is the uncertainty of the objective function, and K is the parameter that balances exploration and utilization. Through Bayesian optimization, the optimal parameter combination can be found with less computation and the decoding accuracy of the model can be improved.

[0101] In addition, transfer learning technology can also be used to optimize the training process of the joint decoding model. Transfer learning transfers the learned features to new tasks by pre-training on other similar datasets. This not only speeds up the training process, but also maintains the generalization ability of the model when the amount of data is small. Specifically, a preliminary model is first trained on a large-scale mass spectrometry database, and then fine-tuned on a specific task dataset. For example, if the model has been pre-trained on certain protein mass spectrometry data, then when it is applied to a new task, the model can quickly adapt to different mass spectrometry features, significantly improve accuracy and reduce training time.

[0102] Example 2

[0103] Embodiment 2 is based on step S1 of embodiment 1 and provides a more specific example.

[0104] Assume that a set of protein mass spectrometry data is obtained from the Acpdb database. The mass spectrum of one protein is 256×256 in size, and there are a large number of water molecules and solvent molecules. Through the preprocessing script, we cut the mass spectrum into 64×64 sub-graphs, and the step length of each movement is 32. Then, according to the formula, we can calculate:

[0105]

[0106] Therefore, 81 sub-graphs can be extracted from the mass spectrometry data. Next, the mass interval corresponding to the water molecules is removed by filtering the mass-to-charge ratio, and the preprocessed mass spectrometry data is used for subsequent model training.

[0107] This step provides a systematic method to process mass spectra containing interfering components, by cutting them into small sub-graphs and filtering unnecessary water molecules and solvent molecules, thereby enhancing the quality and usability of the data. This process lays the foundation for subsequent model training, improves the model's ability to identify and decode target proteins, and reflects the importance of data preprocessing in mass spectrometry analysis.

[0108] In addition, in order to balance the computational complexity of decoding and the robustness of decoding, the joint decoding model in Embodiment 2 adopts an MS-RE decoding model and an MS-MS decoding model composed of two layers of MS models.

[0109] The other steps of Example 2 are similar to those of Example 1 and will not be described again here.

[0110] It should be noted that, for the sake of convenience, the aforementioned method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present invention is not limited to the described order of actions, because according to the present invention, certain steps can be performed in other orders or simultaneously.

[0111] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A method for joint data decoding and optimization of RE analysis and FTICRMS based on machine learning, characterized in that: The steps include: Collecting protein crystal structure data and mass spectrometry data, preliminarily screening and preprocessing the data to obtain a training data set; the training data set includes reaction site data and a mass spectrometry subgraph data set; Constructing a data decoding model, training the data decoding model using reaction site data, automatically extracting and learning key local features of the reaction site data, obtaining a first data decoding model, training the data decoding model using a mass spectrum subgraph data set, learning different protein molecular features, and obtaining a second data decoding model; Evaluate the first data decoding model through the characteristic curve to obtain a first evaluation result; comprehensively evaluate the second data decoding model through the quality evaluation index to obtain a second evaluation result, wherein the quality evaluation index includes resolution and peak characteristics; Optimize the models respectively using the first evaluation result and the second evaluation result to obtain the best first data decoding model and second data decoding model; Combine the best second data decoding model with the best first data decoding model to obtain a first combined decoding model, combine multiple best second data decoding models to obtain a second combined decoding model, and construct a joint decoding model using the first combined decoding model as a backbone and the second combined decoding model as an additional module, wherein an output end of the second combined decoding model is connected to an output end of the first combined decoding model; The joint decoding model is tested using a given independent test set, and the decoding performance is evaluated using actual mass spectrometry data and protein structure data. The best joint decoding model is obtained based on the evaluation results, and the mass spectrometry data of the protein is decoded using the best joint decoding model.

2. The method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning according to claim 1 is characterized in that: The preliminary screening and preprocessing of the data includes: Extract spatial position information from protein crystal structure data through X-ray crystallography to obtain protein reaction site data; Preliminarily screening the mass spectrometry data by comparing the protein structure to obtain a mass spectrometry data set, including an ion-free and solvent-free molecule set, a non-bridging data set, and a bridging data set, wherein the ion-free and solvent-free molecule set is a pure protein structure, and the non-bridging data set and the bridging data set have impurities; The mass spectrometry data set is cut into fixed-size sub-graphs using a window algorithm; The mass-to-charge ratios of water molecules, solvent molecules and other interfering data in the non-bridging data set and the bridged data set are identified by analyzing the characteristic peaks, and the water molecules, solvent molecules and other interfering data are filtered out as follows: Here, let the mass spectrum data be M, and the mass-to-charge ratio of the i-th data point of the mass spectrum data be m / z i , M water is the mass of water molecules, ∈ is an allowable error range, and the other interfering data include lipids, salt ions, additives or other impurity molecules.

3. The method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning according to claim 1 is characterized in that: The data decoding model includes N convolution blocks and a fully connected layer, the output of the convolution block is connected to the input of the next convolution block, and the output of the Nth convolution block is connected to the fully connected layer; the convolution block includes a convolution layer and a pooling layer, and the output of the convolution layer is connected to the pooling layer.

4. The method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning according to claim 1 is characterized in that: The evaluating the first data decoding model by using the characteristic curve includes: Each sample in the reaction site data is predicted by different decision thresholds. For each threshold, the corresponding true positive rate and false positive rate are calculated. According to the true positive rate and false positive rate of each threshold, the ROC curve is generated; The AUC value is calculated based on the ROC curve to evaluate the robustness of the current model and generate a first evaluation result.

5. The method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning according to claim 1 is characterized in that: The comprehensive evaluation of the second data decoding model by using the quality evaluation index includes: Calculate the resolution and signal-to-noise ratio of the data based on the mass spectrum sub-image data set; Analyze the peak features of the mass spectrum sub-graph, obtain the sample peak features, calculate the matching degree between the sample peak features and the target peak features, and mark the data with low matching degree as low-quality data; the sample peak features include peak shape, peak position and peak intensity.

6. The method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning according to claim 5 is characterized in that: The calculation of the matching degree is as follows: Among them, I observed is the observed peak intensity, I expected is the expected peak intensity and N is the number of peaks.

7. The method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning according to claim 5 is characterized in that: The method includes optimizing the models respectively using the second evaluation results, specifically: Exclude low-quality data; Add low-quality data to the training dataset.

8. The method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning according to claim 1 is characterized in that: The first combined decoding model is used as the backbone part, and the second combined decoding model is used as an additional module to construct a joint decoding model, specifically: The sample data is initially decoded by a first combined decoding model, and the sample data is refined by a second combined decoding model. The initial decoding result and the refined decoding result are concatenated to obtain a joint data decoding model; the second combined decoding model has at least one.

9. The method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning according to claim 1, characterized in that: The evaluation of decoding performance using actual mass spectrometry data and protein structure data is specifically as follows: The test output results are compared with the real mass spectrometry data, the comparison results are evaluated through performance indicators, and the best joint decoding model is screened based on the evaluation results; the performance indicators ROC curve and its AUC value, accuracy, recall rate and F1 value.

10. The method for RE analysis and FTICRMS joint data decoding and optimization based on machine learning according to claim 1, characterized in that: It also includes optimizing the joint decoding model using hyperparameter optimization and transfer learning; the hyperparameter optimization includes grid search and Bayesian optimization.

Images