Deep learning-based top-down mass spectrum deconvolution method and system

Abstract

The invention provides a top-down mass spectrum deconvolution method and system based on deep learning. The method comprises the following steps: acquiring protein sample mass spectrum data for model training; extracting multi-dimensional physical characteristics representing signal peak characteristics from candidate experimental envelopes of protein sample mass spectrum data and corresponding theoretical envelope models; constructing an isotope envelope similarity evaluation model based on a multi-scale convolutional neural network, and training the isotope envelope similarity evaluation model by using a label generated based on a theoretical fragment ion matching result until the isotope envelope similarity evaluation model meets a preset training termination condition, obtaining a trained isotope envelope similarity evaluation model; and screening real isotope envelope signals in the mass spectrum data to be processed by using the trained isotope envelope similarity evaluation model, and generating a single isotope mass list for protein identification. According to the invention, the accuracy of top-down mass spectrum deconvolution can be improved.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a top-down mass spectrometry deconvolution method and system based on deep learning. Background Technology

[0002] Proteomics research is central to elucidating the molecular mechanisms of life activities, and top-down mass spectrometry analysis, which can fully preserve protein variant information (such as mutations and post-translational modifications), has become a key technology for precision proteomics. One of the core challenges of this technology lies in "deconvolution," that is, accurately resolving the single isotopic mass corresponding to a single protein variant from complex experimental mass spectrometry signals.

[0003] Traditional deconvolution methods (such as similarity matching based on the Averagine model and dynamic programming algorithms based on graph models) often suffer from insufficient accuracy and high false positive rates when dealing with real-world samples with low signal-to-noise ratios, multiple overlapping envelopes, and complex mixed charge states. In recent years, deep learning-based approaches (such as EnvCNN and FREE methods) have improved the intelligence level of envelope determination by introducing handcrafted features and convolutional neural networks.

[0004] However, existing models still have limitations: a single-scale receptive field is difficult to capture both the fine relationships between local peaks and the overall shape of the macro-envelope at the same time; the ability to distinguish key features is insufficient; and when dealing with training data with highly imbalanced classes, the model tends to favor the majority class, affecting the recognition of low-abundance true envelopes. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a top-down mass spectrometry deconvolution method and system based on deep learning, so as to improve the accuracy of top-down mass spectrometry deconvolution.

[0006] In a first aspect, the present invention provides a top-down mass spectrometry deconvolution method based on deep learning, the method comprising the following steps: Obtain protein sample mass spectrometry data for model training; Multidimensional physical features characterizing signal peak properties are extracted from candidate experimental envelopes and corresponding theoretical envelope models of protein sample mass spectrometry data. These multidimensional physical features include peak intensity distribution, mass-to-charge ratio shift, and similarity measure. An isotope envelope similarity evaluation model based on a multi-scale convolutional neural network was constructed. The model was trained using labels generated from theoretical fragment ion matching results until it met a preset training termination condition. The trained model consisted of an input layer, a feature extraction layer, and an output layer connected sequentially. The input layer received a feature matrix constructed from multi-dimensional physical features. The feature extraction layer was composed of multi-scale convolutional units, a channel attention mechanism, and residual blocks. The multi-scale convolutional units extracted local inter-peak relationships and macroscopic envelope morphology features at different scales in parallel. The channel attention mechanism adjusted the responses of different feature channels by generating weight vectors, and the residual blocks maintained gradient stability. The output layer output the similarity prediction probability score between the candidate experimental envelope and the theoretical envelope model. The trained isotope envelope similarity assessment model is used to screen the real isotope envelope signals in the mass spectrometry data to be processed, and a list of single isotope masses is generated for protein identification.

[0007] Optionally, before extracting multidimensional physical features characterizing the signal peak properties from the candidate experimental envelope and corresponding theoretical envelope model of the protein sample mass spectrometry data, the method further includes: The protein sample mass spectrometry data is converted into a standard mass spectrometry format file, and the initial isotope clusters in the standard mass spectrometry format file are identified by the deconvolution algorithm to generate candidate experimental envelopes containing mass-to-charge ratio distribution and charge state information. Based on the unique identifier of the spectrum, the candidate experimental envelope is mapped and associated with the protein spectrum matching file that records the known protein fragment information, so as to determine the theoretical envelope model corresponding to each candidate experimental envelope.

[0008] Optionally, similarity measures in multidimensional physical features include at least one of cosine similarity, Euclidean distance, Manhattan distance, and Chebyshev distance.

[0009] Optionally, the multi-scale convolutional units in the feature extraction layer include a first convolutional kernel, a second convolutional kernel, a third convolutional kernel, and a fourth convolutional kernel in parallel; the kernel sizes of the first convolutional kernel, the second convolutional kernel, the third convolutional kernel, and the fourth convolutional kernel decrease sequentially; the first and second convolutional kernels are used to capture macroscopic envelope morphological features, and the third and fourth convolutional kernels are used to capture local inter-peak relationships.

[0010] Optionally, the computational expression for the channel attention mechanism is:

[0011] in, Represents the weight vector. express Activation function Indicates batch normalization. This represents a multilayer perceptron. Indicates global average pooling. Indicates global max pooling. This represents the feature map output by a multi-scale convolutional unit.

[0012] Optionally, the loss function of the isotope envelope similarity assessment model is a binary cross-entropy function with class weights, which is used to optimize the accuracy of envelope binary classification and handle sample imbalance. The expression for the loss function is:

[0013] in, Indicates the loss value. Indicates protein sample The true label is used to determine whether the corresponding envelope is a true isotopic envelope. This indicates that the isotope envelope similarity assessment model is applicable to protein samples. The similarity prediction probability score reflects the degree of physical matching of the envelope similarity. and These represent the class weights of positive and negative samples, respectively. They are calculated using a balanced mode to improve the accuracy of identifying the true envelope and mitigate the impact of dataset imbalance. Positive samples correspond to the true envelope signal, while negative samples correspond to noise or false envelopes.

[0014] Optionally, the trained isotope envelope similarity assessment model is used to filter the true isotope envelope signals in the mass spectrometry data to be processed, generating a list of single isotope masses, including: The trained isotope envelope similarity assessment model is used to infer the candidate experimental envelope and output the similarity prediction probability score. Based on the comparison results of the pre-set score threshold and the similarity prediction probability score, high-confidence true envelope signals are selected, and the single isotope mass values ​​are calculated and updated according to the high-confidence true envelope signals to obtain a list of single isotope masses.

[0015] Secondly, the present invention provides a top-down mass spectrometry deconvolution system based on deep learning, comprising: The data acquisition module is used to acquire mass spectrometry data of protein samples for model training. The feature extraction module is used to extract multidimensional physical features characterizing the signal peaks from the candidate experimental envelope and the corresponding theoretical envelope model of protein sample mass spectrometry data. The multidimensional physical features include peak intensity distribution, mass-to-charge ratio shift, and similarity measure. The deep learning module is used to construct an isotope envelope similarity evaluation model based on a multi-scale convolutional neural network. The isotope envelope similarity evaluation model includes an input layer, a feature extraction layer, and an output layer connected sequentially. The input layer receives a feature matrix constructed from multi-dimensional physical features. The feature extraction layer is based on multi-scale convolutional units, a channel attention mechanism, and residual blocks. The multi-scale convolutional units are used to extract local inter-peak relationships and macroscopic envelope morphology features at different scales in parallel. The channel attention mechanism is used to adjust the responses of different feature channels by generating weight vectors, and the residual blocks are used to maintain gradient stability. The output layer outputs the similarity prediction probability score between the candidate experimental envelope and the theoretical envelope model. The deconvolution module is used to filter the true isotopic envelope signals in the mass spectrometry data to be processed using an isotopic envelope similarity assessment model, and generate a list of single isotopic masses for protein identification.

[0016] Optionally, the multi-scale convolutional units in the feature extraction layer include a first convolutional kernel, a second convolutional kernel, a third convolutional kernel, and a fourth convolutional kernel in parallel; the kernel sizes of the first convolutional kernel, the second convolutional kernel, the third convolutional kernel, and the fourth convolutional kernel decrease sequentially; the first and second convolutional kernels are used to capture macroscopic envelope morphological features, and the third and fourth convolutional kernels are used to capture local inter-peak relationships.

[0017] Optionally, the deep learning module also includes a model training module; The model training module uses labels generated based on theoretical fragment ion matching results to train the isotope envelope similarity assessment model until the isotope envelope similarity assessment model meets the preset training termination condition. The loss function of the trained isotope envelope similarity assessment model is a binary cross-entropy function with class weights, which is used to optimize the envelope binary classification accuracy and handle sample imbalance. The expression for the loss function is:

[0018] in, Indicates the loss value. Indicates protein sample The true label is used to determine whether the corresponding envelope is a true isotopic envelope. This indicates that the isotope envelope similarity assessment model is applicable to protein samples. The similarity prediction probability score reflects the degree of physical matching of the envelope similarity. and These represent the class weights of positive and negative samples, respectively. They are calculated using a balanced mode to improve the accuracy of identifying the true envelope and mitigate the impact of dataset imbalance. Positive samples correspond to the true envelope signal, while negative samples correspond to noise or false envelopes.

[0019] The present invention has at least the following beneficial effects: The isotope envelope similarity evaluation model constructed in this invention employs a feature extraction layer combining multi-scale convolutional units, channel attention mechanisms, and residual blocks. Multi-scale convolutional units can extract features from different receptive fields in parallel, simultaneously capturing local inter-peak relationships and macroscopic envelope morphology, thus solving the inherent problem of limited receptive fields in single-scale models. The channel attention mechanism dynamically evaluates and strengthens feature channels that contribute significantly to classification, suppressing redundant or interfering channels, thereby enhancing the model's ability to focus on key discriminative information. The introduction of residual blocks ensures stable gradient propagation in deep networks, preventing overfitting and enabling the model to perform deeper and more effective feature learning. This composite structure allows the model to maintain high-precision discrimination capabilities even in complex mass spectrometry scenarios such as severely overlapping envelopes and high background noise, which is beneficial for improving the accuracy of top-down mass spectrometry deconvolution. Furthermore, this invention employs a binary cross-entropy loss function with class weights. This function assigns higher weights to the "true envelope" class, which has a smaller sample size, forcing the model to pay more attention to these difficult-to-identify but crucial minority class samples during training. This processing mechanism significantly improves the model's sensitivity and recognition rate to low-abundance protein signals, effectively alleviates the model bias problem caused by data class imbalance, and makes the deconvolution results more accurate. Attached Figure Description

[0020] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0021] Figure 1 This is a flowchart of a deep learning-based top-down mass spectrometry deconvolution method in one embodiment of this application; Figure 2 This is a model structure diagram of an isotope envelope similarity assessment model in one embodiment of this application; Figure 3 This is a comparative embodiment diagram of a deep learning-based top-down mass spectrometry deconvolution method in another implementation of this application, and existing methods; wherein, Figure 3 (a) A comparison of the model performance of the deep learning-based top-down mass spectrometry deconvolution method provided in this invention with existing methods on multiple datasets; Figure 3 (b) A comparison of the number of PrSMs between the deep learning-based top-down mass spectrometry deconvolution method provided by this invention and existing methods on multiple datasets; Figure 4 This is a structural diagram of a top-down mass spectrometry deconvolution system based on deep learning, according to one embodiment of this application. Detailed Implementation

[0022] The technical solution of the present invention will now be described in detail and completely with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0023] In the description of this invention, it should be noted that the terms "upper", "lower", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0024] Example 1 like Figure 1 As shown, the deep learning-based top-down mass spectrometry deconvolution method provided by the present invention includes steps 11 to 14.

[0025] Step 11: Obtain mass spectrometry data of protein samples for model training.

[0026] In one feasible implementation, the protein sample mass spectrometry data is a raw mass spectrometry file (e.g., .raw format) of nine replicate samples of human ovarian tumor cells (OT) obtained in advance. The file contains core information such as signal intensity, retention time and mass-to-charge ratio (m / z).

[0027] To facilitate processing, protein sample mass spectrometry data needs to be preprocessed after acquisition. Specifically, the preprocessing steps include steps 11.1 to 11.2.

[0028] Step 11.1: Convert the protein sample mass spectrometry data into a standard mass spectrometry format file, and use the deconvolution algorithm to identify the initial isotope clusters in the standard mass spectrometry format file to generate candidate experimental envelopes containing mass-to-charge ratio distribution and charge state information.

[0029] In one feasible implementation, the Msconvert tool is used to convert the raw mass spectrometry file into the open standard mzML format. This step eliminates the differences in data formats between instrument manufacturers, facilitating subsequent processing by general-purpose software tools.

[0030] Subsequently, the TopFD tool was used to deconvolve the mzML file. TopFD identifies isotope clusters in the spectrum and calculates their individual isotope masses, while outputting key files, including: a feature file (.msalign) containing information such as individual isotope masses, an msalign file recording the original spectrum information, and most importantly, an envelope file (.env) recording detailed data such as the m / z range, peak intensity, and charge state of each candidate isotope envelope. This step yielded the candidate experimental envelopes.

[0031] Step 11.2: Based on the unique identifier of the spectrum, the candidate experimental envelopes are mapped and associated with the protein spectrum matching files that record known protein fragment information, so as to determine the theoretical envelope model corresponding to each candidate experimental envelope.

[0032] In one feasible implementation, based on the unique ID of each spectrum, the envelope file (.env) generated by TopFD is associated with and renamed a protein spectrum matching file (PrSM) that contains known protein fragment matching information and is generated through other means (such as preliminary database search). This ensures that each experimental envelope corresponds to a theoretical reference (i.e., a list of possible fragment ions), laying the foundation for subsequent annotation.

[0033] In practice, the associated file is read, and the samples required for supervised learning are labeled. Using the theoretical fragment ion list in the PrSM file as a reference, a mass tolerance (e.g., 15 ppm) is set. For an experimental envelope, if its mass matches the mass of any theoretical fragment ion within the tolerance range, the envelope is marked as a positive sample (label y=1), representing a "real" isotopic envelope signal generated by the target protein; otherwise, it is marked as a negative sample (label y=0), representing noise, chemical noise, or interference signals from non-target substances. The theoretical fragment ion types cover 14 common ions, including N-terminal ions (b, a, c) and C-terminal ions (y, x, z) and their variants. These ion types are applicable to different fragmentation modes such as CID, HCD, and ETD.

[0034] Step 12: Extract multidimensional physical features characterizing signal peak properties from the candidate experimental envelope and corresponding theoretical envelope model of the protein sample mass spectrometry data.

[0035] In this embodiment of the invention, the multidimensional physical features include peak intensity distribution, mass-to-charge ratio shift, and similarity metrics. Similarity metrics include at least one of cosine similarity, Euclidean distance, Manhattan distance, and Chebyshev distance.

[0036] In practice, firstly, for each positive sample, a theoretical isotope envelope is generated using an isotope distribution simulator based on the chemical formula (or amino acid sequence) of its matching theoretical fragment ions, including the precise mass and relative abundance of each isotope peak.

[0037] In one feasible implementation, the multidimensional physical features include the candidate experimental envelope peak intensity (normalized to a vector), the candidate experimental envelope peak mass-to-charge ratio, the theoretical envelope peak intensity vector (normalized to a vector), cosine similarity, Manhattan distance, Euclidean distance, and Chebyshev distance. Specifically, the candidate experimental envelope peak intensity is obtained by scaling the intensities of each isotope peak in the candidate experimental envelope to a sum of 1; cosine similarity is obtained by calculating the cosine of the cosine angle between the candidate experimental envelope peak intensity and the theoretical envelope peak intensity, measuring the similarity of the distribution shape; Euclidean distance is obtained by calculating the straight-line distance between the candidate experimental envelope peak intensity and the theoretical envelope peak intensity in space, used to measure the absolute numerical difference; Manhattan distance is obtained by calculating the sum of the absolute values ​​of the differences in each dimension between the candidate experimental envelope peak intensity and the theoretical envelope peak intensity; and Chebyshev distance is obtained by calculating the maximum value of the differences in each dimension between the candidate experimental envelope peak intensity and the theoretical envelope peak intensity.

[0038] In one feasible implementation, for each envelope pair, a 300×9 feature matrix is ​​constructed, where 300 represents the m / z range. The matrix is ​​divided into 300 intervals, each 0.01 m / z wide. The number 9 represents the nine feature dimensions corresponding to each peak (including the normalized value of the experimental envelope peak intensity, the normalized value of the theoretical envelope peak intensity, the m / z similarity between the theoretical and experimental envelope peaks, the intensity difference between the theoretical and experimental envelope peaks, the logarithmic ratio of the theoretical highest peak intensity to the baseline intensity, cosine similarity, Manhattan distance, Euclidean distance, and Chebyshev distance). This matrix converts the envelope pairs into a vector sequence, facilitating model input and processing.

[0039] Step 13: Construct an isotope envelope similarity evaluation model based on a multi-scale convolutional neural network, and train the isotope envelope similarity evaluation model using labels generated based on theoretical fragment ion matching results until the isotope envelope similarity evaluation model meets the preset training termination condition, thus obtaining the trained isotope envelope similarity evaluation model.

[0040] like Figure 2As shown in this embodiment of the invention, the isotope envelope similarity evaluation model includes an input layer 21, a feature extraction layer 22, and an output layer 23 connected in sequence. The input layer 21 receives a feature matrix constructed from multi-dimensional physical features. The feature extraction layer 22 is based on multi-scale convolutional units, a channel attention mechanism, and residual blocks. The multi-scale convolutional units are used to extract local inter-peak relationships and macroscopic envelope morphology features at different scales in parallel. The channel attention mechanism is used to adjust the responses of different feature channels by generating weight vectors, and the residual blocks are used to maintain gradient stability. The output layer 23 outputs the similarity prediction probability score between the candidate experimental envelope and the theoretical envelope model.

[0041] Specifically, in the input layer, a one-dimensional convolutional layer with a large kernel (e.g., 7) is first used, along with batch normalization (BatchNorm) and LeakyReLU activation functions, to quickly extract the basic features of the feature matrix and perform a downsampling (stride=2) to reduce the sequence length from 300 to 150.

[0042] The multi-scale convolutional units in the feature extraction layer include parallel first, second, third, and fourth convolutional kernels; the kernel sizes of the first, second, third, and fourth convolutional kernels decrease sequentially (e.g., kernel sizes of 11, 5, 3, and 1 respectively); the first, second, third, and fourth convolutional kernels correspond to convolutional branches 1, 2, 3, and 4, respectively. The convolutional branches are connected in parallel through concatenation layers. Specifically, for the input feature map... The calculation process of multi-scale convolution is as follows: for a convolution kernel size of... Expansion rate The convolution operation outputs a feature map. The calculation formula is: ,in Indicates the input feature map, Indicates the convolution kernel weights. Indicates the bias term. This represents the location index of the feature map. Dilated convolution effectively captures a wider range of envelope features by increasing the receptive field without increasing the number of parameters. By computing convolutions at different scales in parallel, four feature maps are obtained, each corresponding to a different kernel size and dilation rate. Each feature map has the same number of channels. Using convolutions with a dilation rate greater than 1 expands the receptive field without increasing the number of parameters, enabling the model to capture a wider range of envelope patterns.

[0043] Subsequently, the multi-scale features are concatenated through feature fusion to obtain a fused multi-scale feature representation, the number of which is the sum of the number of channels in each branch. The fused features are then adjusted for the number of channels using a 1×1 convolution, combined with batch normalization and the ReLU activation function, and further enhanced using a channel attention mechanism.

[0044] The computational expression for the channel attention mechanism is:

[0045] in, Represents the weight vector. express Activation function Indicates batch normalization. This represents a multilayer perceptron. Indicates global average pooling. Indicates global max pooling. This represents the feature map output by a multi-scale convolutional unit.

[0046] The specific calculation process is as follows: First, the input feature map... Global average pooling and global max pooling are performed separately to obtain two pooled feature vectors, each with a dimension equal to the number of channels in the input feature map. These two vectors are then processed by a shared multilayer perceptron (MLP) network. The MLP contains two fully connected layers: the first layer compresses the number of channels with the ReLU activation function, and the second layer restores the original number of channels without an activation function. The outputs of the two MLPs are summed and then processed with a sigmoid activation function and batch normalization to obtain a weight vector W. Finally, the weight vector is multiplied channel-by-channel with the original feature map to dynamically enhance the response of important feature channels and suppress irrelevant features.

[0047] In this embodiment of the invention, residual blocks are used to maintain gradient stability and prevent overfitting. Internally, they employ the LeakyReLU activation function and a Batch Normalization layer for nonlinear mapping. The design of the residual blocks follows the principle of residual connections, and its calculation process is as follows: the residual connection path is selected based on whether the number of input and output channels are the same: if the number of input channels is the same as the number of output channels, the residual path directly uses the input; otherwise, the number of channels is adjusted through a 1×1 convolution. The output of the main path is added to the output of the residual path, and then passed through the LeakyReLU activation function and a Dropout layer again. The Dropout layer randomly sets the output of some neurons to zero with a certain probability, thereby preventing overfitting. Residual connections allow gradients to propagate directly backward, alleviating the gradient vanishing problem in deep networks, thus supporting deeper network structures. Compared to ReLU, the LeakyReLU activation function can alleviate the neuron death problem, the Batch Normalization layer can stabilize the training process and accelerate convergence, and the Dropout layer can prevent overfitting.

[0048] It should be noted that after the channel attention mechanism, the feature extraction layer performs multiple downsampling operations through convolutions with a stride greater than 1, gradually increasing the number of channels. After downsampling, feature extraction is performed through residual blocks, combined with batch normalization, activation functions, and Dropout layers to complete the feature extraction process.

[0049] In the output layer: First, global average pooling is performed on the feature map output from the last feature extraction block to obtain a fixed-length feature vector. Then, 1-2 fully connected (Dense) layers are used for further integration, with Dropout (rate=0.5) to prevent overfitting. Finally, a fully connected layer with 1 neuron is passed through, and the sigmoid activation function is used to output a scalar value between 0 and 1. . This indicates the probability that the model predicts the envelope of the input candidate experiment to be the "true envelope".

[0050] It should be noted that, in the embodiments of the present invention, the loss function of the isotope envelope similarity evaluation model is a binary cross-entropy function with class weights, which is used to optimize the accuracy of envelope binary classification and handle sample imbalance. The expression for the loss function is:

[0051] in, Indicates the loss value. Indicates protein sample The true label is used to determine whether the corresponding envelope is a true isotopic envelope. This indicates that the isotope envelope similarity assessment model is applicable to protein samples. The similarity prediction probability score reflects the degree of physical matching of the envelope similarity. and These represent the class weights of positive and negative samples, respectively. Calculated using a balanced mode, these weights are used to improve the accuracy of identifying the true envelope and mitigate the impact of dataset imbalance. Positive samples correspond to the true envelope signal, while negative samples correspond to noise or spurious envelopes. It should be noted that the balanced mode calculation compensates for the learning bias caused by the order-of-magnitude difference between positive and negative samples by assigning higher class weights to positive samples.

[0052] During training, the loss function is optimized using the backpropagation algorithm, parameters are updated using the Adam optimizer, and a learning rate scheduler and early stopping mechanism are used to further optimize the training process. The early stopping mechanism monitors the validation set loss; if the loss does not decrease for 10 consecutive epochs, training is stopped, and the model weights at the point of lowest validation loss are restored. The learning rate scheduler includes learning rate decay; when the validation loss stagnates, the learning rate is halved. The training termination condition can be set by reaching a preset maximum number of training iterations.

[0053] Step 14: Use the trained isotope envelope similarity assessment model to screen the real isotope envelope signals in the mass spectrometry data to be processed, and generate a list of single isotope masses for protein identification.

[0054] Specifically, the process of using the trained isotope envelope similarity assessment model to screen the real isotope envelope signals in the mass spectrometry data to be processed and generating a single isotope mass list includes steps 14.1 to 14.2.

[0055] Step 14.1: Use the trained isotope envelope similarity evaluation model to infer the candidate experimental envelope and output the similarity prediction probability score.

[0056] In one feasible implementation, multidimensional physical features of the mass spectrometry data to be processed are extracted, and then the multidimensional physical features are converted into a matrix and input into a trained isotope envelope similarity evaluation model. The trained isotope envelope similarity evaluation model outputs a similarity prediction probability score.

[0057] Step 14.2: Based on the comparison results of the pre-set score threshold and the similarity prediction probability score, select the true envelope signals with high confidence, and calculate and update the single isotope mass values ​​according to the high confidence true envelope signals to obtain the single isotope mass list.

[0058] In one feasible implementation, when the similarity prediction probability score is higher than a pre-set score threshold, the candidate experimental envelope is determined to be "true." All envelopes determined to be "true" are used to generate the final single isotope mass list (.msalign file). This list contains fewer false positives and more true positives (especially low-abundance signals) compared to the original TopFD output.

[0059] After obtaining the list of single isotope masses, the optimized .msalign file is input into the protein identification software TopPIC to perform a database search, identify proteins and their variants, and output a protein spectrum matching (PrSM) list.

[0060] Example 2 In this embodiment of the invention, the deep learning-based top-down mass spectrometry deconvolution method (MACE) provided by this invention was compared with existing methods (FREE, using the original TopFD) on multiple datasets (ZF, OT). The comparison results are as follows. Figure 3 (a) Figure 3 As shown in (b). Wherein, Figure 3 (a) is a performance comparison chart of the deep learning-based top-down mass spectrometry deconvolution method provided in this invention with existing methods on multiple datasets. Figure 3 (b) is a comparison of the number of PrSMs on multiple datasets between the deep learning-based top-down mass spectrometry deconvolution method provided in this invention and existing methods. ZF represents samples from the cerebellum and thalamus tectum regions of mature female zebrafish, divided into six sub-datasets: CB1, CB2, CB3, Teo1, Teo2, and Teo3. OT represents samples from human ovarian tumor cells. Figure 3 (a) It can be seen that the model performance of the deep learning-based top-down mass spectrometry deconvolution method provided in this invention outperforms existing methods on multiple datasets. Figure 3 (b) It can be seen that on multiple datasets, the deep learning-based top-down mass spectrometry deconvolution method provided by this invention can obtain a comparable or higher number of PrSMs, and after manual inspection, the newly added PrSMs are mostly reliable low-abundance variants. This directly proves the actual improvement of the accuracy of top-down mass spectrometry deconvolution by this invention.

[0061] Therefore, the isotope envelope similarity evaluation model constructed in this invention employs a feature extraction layer combining multi-scale convolutional units, channel attention mechanisms, and residual blocks. Multi-scale convolutional units can extract features from different receptive fields in parallel, simultaneously capturing local inter-peak relationships and macroscopic envelope morphology, thus solving the inherent problem of limited receptive fields in single-scale models. The channel attention mechanism dynamically evaluates and strengthens feature channels that contribute significantly to classification, suppressing redundant or interfering channels, thereby enhancing the model's ability to focus on key discriminative information. The introduction of residual blocks ensures stable gradient propagation in deep networks, preventing overfitting and enabling the model to perform deeper and more effective feature learning. This composite structure allows the model to maintain high-precision discrimination capabilities even in complex mass spectrometry scenarios such as severely overlapping envelopes and high background noise, which is beneficial for improving the accuracy of top-down mass spectrometry deconvolution. Furthermore, this invention employs a binary cross-entropy loss function with class weights. This function assigns higher weights to the "true envelope" class, which has a smaller sample size, forcing the model to pay more attention to these difficult-to-identify but crucial minority class samples during training. This processing mechanism significantly improves the model's sensitivity and recognition rate to low-abundance protein signals, effectively alleviates the model bias problem caused by data class imbalance, and makes the deconvolution results more accurate.

[0062] Example 3 In this embodiment of the invention, a top-down mass spectrometry deconvolution system based on deep learning is also disclosed, such as... Figure 4 As shown, the system includes: The data acquisition module 401 is used to acquire protein sample mass spectrometry data for model training; The feature extraction module 402 is used to extract multidimensional physical features characterizing the signal peak properties from the candidate experimental envelope and the corresponding theoretical envelope model of the protein sample mass spectrometry data; the multidimensional physical features include peak intensity distribution, mass-to-charge ratio shift and similarity measure. Deep learning module 403 is used to construct an isotope envelope similarity evaluation model based on a multi-scale convolutional neural network. The isotope envelope similarity evaluation model includes an input layer, a feature extraction layer, and an output layer connected in sequence. The input layer receives a feature matrix constructed from multi-dimensional physical features. The feature extraction layer is composed of multi-scale convolutional units, a channel attention mechanism, and residual blocks. The multi-scale convolutional units are used to extract local inter-peak relationships and macroscopic envelope morphology features at different scales in parallel. The channel attention mechanism is used to adjust the responses of different feature channels by generating weight vectors. The residual blocks are used to maintain gradient stability. The output layer outputs the similarity prediction probability score of the candidate experimental envelope and the theoretical envelope model. The deconvolution module 404 is used to screen the true isotope envelope signals in the mass spectrometry data to be processed using an isotope envelope similarity assessment model, and generate a list of single isotope masses for protein identification.

[0063] In one feasible implementation, the multi-scale convolutional unit in the feature extraction layer 402 includes a first convolutional kernel, a second convolutional kernel, a third convolutional kernel, and a fourth convolutional kernel in parallel; the kernel sizes of the first convolutional kernel, the second convolutional kernel, the third convolutional kernel, and the fourth convolutional kernel decrease sequentially; the first convolutional kernel and the second convolutional kernel are used to capture macroscopic envelope morphological features, and the third convolutional kernel and the fourth convolutional kernel are used to capture local inter-peak relationships.

[0064] In one feasible implementation, the deep learning module 403 further includes a model training module; the model training module uses labels generated based on theoretical fragment ion matching results to train the isotope envelope similarity assessment model until the isotope envelope similarity assessment model meets a preset training termination condition, and the loss function of the trained isotope envelope similarity assessment model is a binary cross-entropy function with class weights, used to optimize the envelope binary classification accuracy and handle sample imbalance; the expression of the loss function is:

[0065] in, Indicates the loss value. Indicates protein sample The true label is used to determine whether the corresponding envelope is a true isotopic envelope. This indicates that the isotope envelope similarity assessment model is applicable to protein samples. The similarity prediction probability score reflects the degree of physical matching of the envelope similarity. and These represent the class weights of positive and negative samples, respectively. They are calculated using a balanced mode to improve the accuracy of identifying the true envelope and mitigate the impact of dataset imbalance. Positive samples correspond to the true envelope signal, while negative samples correspond to noise or false envelopes.

[0066] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. Their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here. Those skilled in the art will understand that, for the sake of convenience and brevity, the division of the above-mentioned functional units and modules is only used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0067] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0068] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A top-down mass spectrometry deconvolution method based on deep learning, characterized in that, include: Obtain protein sample mass spectrometry data for model training; Multidimensional physical features characterizing signal peak properties are extracted from the candidate experimental envelope and corresponding theoretical envelope model of the protein sample mass spectrometry data. The multidimensional physical characteristics include peak intensity distribution, mass-to-charge ratio shift, and similarity metric. An isotope envelope similarity evaluation model based on a multi-scale convolutional neural network is constructed, and the isotope envelope similarity evaluation model is trained using labels generated based on theoretical fragment ion matching results until the isotope envelope similarity evaluation model meets the preset training termination condition, thus obtaining the trained isotope envelope similarity evaluation model. The isotope envelope similarity evaluation model comprises an input layer, a feature extraction layer, and an output layer connected in sequence. The input layer receives a feature matrix constructed from the multi-dimensional physical features. The feature extraction layer is based on multi-scale convolutional units, a channel attention mechanism, and residual blocks. The multi-scale convolutional units extract local inter-peak relationships and macroscopic envelope morphology features at different scales in parallel. The channel attention mechanism adjusts the responses of different feature channels by generating weight vectors. The residual blocks maintain gradient stability. The output layer outputs the similarity prediction probability score between the candidate experimental envelope and the theoretical envelope model. The trained isotope envelope similarity assessment model is used to screen the real isotope envelope signals in the mass spectrometry data to be processed, and a list of single isotope masses is generated for protein identification.

2. The deep learning-based top-down mass spectrometry deconvolution method according to claim 1, characterized in that, Before extracting multidimensional physical features characterizing signal peak properties from the candidate experimental envelope and corresponding theoretical envelope model of the protein sample mass spectrometry data, the method further includes: The protein sample mass spectrometry data is converted into a standard mass spectrometry format file, and the initial isotope clusters in the standard mass spectrometry format file are identified using a deconvolution algorithm to generate a candidate experimental envelope containing mass-to-charge ratio distribution and charge state information. Based on the unique identifier of the spectrum, the candidate experimental envelopes are mapped and associated with protein spectrum matching files that record known protein fragment information, so as to determine the theoretical envelope model corresponding to each candidate experimental envelope.

3. The deep learning-based top-down mass spectrometry deconvolution method according to claim 2, characterized in that, The similarity measure in the multidimensional physical features includes at least one of cosine similarity, Euclidean distance, Manhattan distance, and Chebyshev distance.

4. The deep learning-based top-down mass spectrometry deconvolution method according to claim 3, characterized in that, The multi-scale convolutional unit in the feature extraction layer includes a first convolutional kernel, a second convolutional kernel, a third convolutional kernel, and a fourth convolutional kernel in parallel; the kernel sizes of the first convolutional kernel, the second convolutional kernel, the third convolutional kernel, and the fourth convolutional kernel decrease sequentially; the first convolutional kernel and the second convolutional kernel are used to capture the macroscopic envelope morphological features, and the third convolutional kernel and the fourth convolutional kernel are used to capture the local inter-peak relationships.

5. The deep learning-based top-down mass spectrometry deconvolution method according to claim 4, characterized in that, The calculation expression for the channel attention mechanism is as follows: in, Represents the weight vector. express Activation function Indicates batch normalization. This represents a multilayer perceptron. Indicates global average pooling. Indicates global max pooling. This represents the feature map output by a multi-scale convolutional unit.

6. The deep learning-based top-down mass spectrometry deconvolution method according to claim 5, characterized in that, The loss function of the isotope envelope similarity assessment model is a binary cross-entropy function with class weights, which is used to optimize the accuracy of envelope binary classification and handle sample imbalance. The expression for the loss function is: in, Indicates the loss value. Indicates protein sample The true label is used to determine whether the corresponding envelope is a true isotopic envelope. This indicates that the isotope envelope similarity assessment model is applicable to protein samples. The similarity prediction probability score reflects the degree of physical matching of the envelope similarity. and These represent the class weights of positive and negative samples, respectively. They are calculated using a balanced mode to improve the accuracy of identifying the true envelope and mitigate the impact of dataset imbalance. The positive samples correspond to the true envelope signal, and the negative samples correspond to noise or false envelopes.

7. The deep learning-based top-down mass spectrometry deconvolution method according to claim 6, characterized in that, The process of using the trained isotope envelope similarity evaluation model to filter the true isotope envelope signals in the mass spectrometry data to be processed and generating a single isotope mass list includes: The trained isotope envelope similarity assessment model is used to infer the candidate experimental envelope and output the similarity prediction probability score. Based on the comparison results of the pre-set score threshold and the similarity prediction probability score, high-confidence true envelope signals are selected, and the single isotope mass values ​​are calculated and updated according to the high-confidence true envelope signals to obtain a single isotope mass list.

8. A top-down mass spectrometry deconvolution system based on deep learning, characterized in that, include: The data acquisition module is used to acquire mass spectrometry data of protein samples for model training. The feature extraction module is used to extract multidimensional physical features characterizing the signal peak properties from the candidate experimental envelope and the corresponding theoretical envelope model of the protein sample mass spectrometry data. The multidimensional physical characteristics include peak intensity distribution, mass-to-charge ratio shift, and similarity metric. A deep learning module is used to construct an isotope envelope similarity evaluation model based on a multi-scale convolutional neural network. The isotope envelope similarity evaluation model includes an input layer, a feature extraction layer, and an output layer connected in sequence. The input layer receives a feature matrix constructed from the multi-dimensional physical features. The feature extraction layer is composed of multi-scale convolutional units, a channel attention mechanism, and residual blocks. The multi-scale convolutional units are used to extract local inter-peak relationships and macroscopic envelope morphology features at different scales in parallel. The channel attention mechanism is used to adjust the responses of different feature channels by generating weight vectors. The residual blocks are used to maintain gradient stability. The output layer outputs the similarity prediction probability score between the candidate experimental envelope and the theoretical envelope model. The deconvolution module is used to filter the true isotope envelope signals in the mass spectrometry data to be processed using the isotope envelope similarity evaluation model, and generate a list of single isotope masses for protein identification.

9. The deep learning-based top-down mass spectrometry deconvolution system according to claim 8, characterized in that, The multi-scale convolutional unit in the feature extraction layer includes a first convolutional kernel, a second convolutional kernel, a third convolutional kernel, and a fourth convolutional kernel in parallel; the kernel sizes of the first convolutional kernel, the second convolutional kernel, the third convolutional kernel, and the fourth convolutional kernel decrease sequentially; the first convolutional kernel and the second convolutional kernel are used to capture the macroscopic envelope morphological features, and the third convolutional kernel and the fourth convolutional kernel are used to capture the local inter-peak relationships.

10. The deep learning-based top-down mass spectrometry deconvolution system according to claim 9, characterized in that, The deep learning module also includes a model training module; The model training module uses labels generated based on theoretical fragment ion matching results to train the isotope envelope similarity evaluation model until the isotope envelope similarity evaluation model meets the preset training termination condition, thus obtaining the trained isotope envelope similarity evaluation model. The loss function of the isotope envelope similarity evaluation model is a binary cross-entropy function with class weights, which is used to optimize the envelope binary classification accuracy and handle sample imbalance. The expression for the loss function is: in, Indicates the loss value. Indicates protein sample The true label is used to determine whether the corresponding envelope is a true isotopic envelope. This indicates that the isotope envelope similarity assessment model is applicable to protein samples. The similarity prediction probability score reflects the degree of physical matching of the envelope similarity. and These represent the class weights of positive and negative samples, respectively. They are calculated using a balanced mode to improve the accuracy of identifying the true envelope and mitigate the impact of dataset imbalance. The positive samples correspond to the true envelope signal, and the negative samples correspond to noise or false envelopes.

Images