Peptide sequencing method, apparatus, device, storage medium, and program product

Abstract

The application provides a peptide sequencing method, device, equipment, storage medium and program product, and relates to the technical field of artificial intelligence. The method comprises the following steps: acquiring mass spectrum data of a peptide to be sequenced; inputting the mass spectrum data into a peptide sequencing model to obtain a peptide sequencing result output by the peptide sequencing model; wherein the peptide sequencing model is obtained by training an initial model based on TSARseqNovo algorithm based on sample mass spectrum data and sample peptide sequencing results corresponding to the sample mass spectrum data. Through the above method, the analysis accuracy of mass spectrum information can be significantly improved, the reasoning efficiency of the model can be improved, and the quality and accuracy of the peptide sequencing result can be effectively ensured.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a peptide sequencing method, apparatus, device, storage medium, and program product. Background Technology

[0002] In mass spectrometry-based proteomics, peptide identification plays a crucial role in analyzing the protein composition of complex biological samples. The processing steps for peptide identification include acquiring mass spectrometry images of the peptide, performing detailed analysis, and pairing them based on key data (such as mass number, charge, and retention time) to infer their constituent amino acid sequences. Currently, the main techniques for peptide identification include library search techniques, which compare and match observed mass spectrometry information with pre-stored peptide information in existing databases to identify the peptide. The limitation of this method is that it is not effective in identifying antibodies, neoantigens in human leukocyte antigens, or novel proteins and peptides not listed in current genomics databases.

[0003] Another mainstream approach is de novo peptide sequencing, which involves obtaining the amino acid sequence of a peptide through tandem mass spectrometry without relying on sequence databases. Early methods primarily used non-deep learning approaches, including heuristic search and dynamic programming, to evaluate the matching degree between the peptide and the corresponding spectral observations. Subsequently, the Novo algorithm incorporated decision trees into its scoring function, enhancing its peptide recognition capabilities. With the explosive development of deep learning methods, researchers proposed the DeepNovo model, which combines convolutional neural networks with long short-term memory networks, significantly improving its de novo peptide sequencing capabilities. Building on this, the PointNovo model, by employing a sequence-invariant network architecture, enhances its feature extraction capabilities from high-resolution mass spectrometry data, further improving de novo peptide sequencing performance. The CasaNovo model, based on the Transformer framework, directly maps peaks observed from mass spectrometry data to amino acid sequences in the feature space, significantly improving peptide recognition and sequencing performance.

[0004] However, the CasaNovo model, as the most commonly used model, also has shortcomings: In terms of the inference object, the CasaNovo model mainly relies on the historical information of peptides for inference, ignoring the information of individual amino acids. Although this strategy improves its prediction performance for peptides, the accuracy decreases when analyzing information of individual amino acids. In addition, when analyzing mass-to-charge ratio and intensity information, the CasaNovo model does not make full use of the spatial arrangement correlation of spectral peaks. These shortcomings significantly affect the CasaNovo model's ability to extract global and local features of mass spectrometry information. In terms of inference speed, the CasaNovo model can only process the prediction analysis of one amino acid in a single inference process, which significantly slows down the inference process.

[0005] Therefore, the accuracy of existing peptide sequencing methods is not high. Summary of the Invention

[0006] This invention provides a peptide sequencing method, apparatus, device, storage medium, and program product to address the shortcomings of low accuracy in existing peptide sequencing methods.

[0007] This invention provides a peptide sequencing method, comprising: acquiring mass spectrometry data of the peptide to be sequenced; inputting the mass spectrometry data into a peptide sequencing model to obtain peptide sequencing results output by the peptide sequencing model; wherein, the peptide sequencing model is obtained by training an initial model based on the TSARseqNovo algorithm based on the sample mass spectrometry data and the sample peptide sequencing results corresponding to the sample mass spectrometry data.

[0008] According to a peptide sequencing method provided by the present invention, the peptide sequencing model includes a mass spectrometry peak information fusion encoding module, a semi-autoregressive decoding module, and a mask optimization decoding module. The mass spectrometry peak information fusion encoding module is used to extract features from mass spectrometry data to obtain mass spectrometry peak fusion embedding features. The semi-autoregressive decoding module is used to perform initial prediction based on the mass spectrometry peak fusion embedding features to obtain amino acid prediction results for the peptide to be sequenced. The mask optimization decoding module is used to perform optimized prediction based on the amino acid prediction results to obtain peptide sequencing results.

[0009] According to a peptide sequencing method provided by the present invention, a mass spectrometry peak information fusion encoding module includes a mass spectrometry peak encoding module, a local and global mass spectrometry feature encoding module, a mass spectrometry peak fusion encoding module, and multiple encoders. The mass spectrometry peak encoding module is used to perform feature encoding on the mass spectrometry data to obtain mass spectrometry peak embedding encoding. The local and global mass spectrometry feature encoding module is used to perform feature encoding on the mass spectrometry data to obtain local and global spectral feature embedding encoding. The mass spectrometry peak fusion encoding module is used to fuse the mass spectrometry peak embedding encoding and the local and global spectral feature embedding encoding to obtain mass spectrometry peak fusion encoding. All encoders are used to generate mass spectrometry peak fusion embedding features based on the mass spectrometry peak fusion encoding.

[0010] According to a peptide sequencing method provided by the present invention, an initial prediction is performed based on mass spectrometry peak fusion embedding features to obtain amino acid prediction results of the peptide to be sequenced, including: generating amino coding information of the peptide to be sequenced; and performing an initial prediction based on the amino coding information and mass spectrometry peak fusion embedding features to obtain amino acid prediction results of the peptide to be sequenced.

[0011] According to a peptide sequencing method provided by the present invention, the amino acid prediction results include multiple amino identifiers of the peptide to be sequenced and the prediction confidence of each amino identifier; the peptide sequencing results are obtained by optimizing the prediction based on the amino acid prediction results, including: determining the target amino identifier based on the amino acid prediction results; the target amino identifier is an amino identifier whose prediction confidence is less than a preset threshold; and optimizing the prediction based on the target amino identifier and the mass spectrometry peak fusion embedding features to generate peptide sequencing results.

[0012] According to a peptide sequencing method provided by the present invention, the peptide sequencing model is trained using a preset loss function; wherein, the preset loss function is determined based on a first loss function and a second loss function, the first loss function being the cross-entropy loss function of the semi-autoregressive decoding module, and the second loss function being the cross-entropy loss function of the mask optimization decoding module.

[0013] The present invention also provides a peptide sequencing device, comprising: an acquisition module for acquiring mass spectrometry data of a peptide to be sequenced; and a sequencing module for inputting the mass spectrometry data into a peptide sequencing model to obtain peptide sequencing results output by the peptide sequencing model; wherein the peptide sequencing model is obtained by training an initial model based on the TSARseqNovo algorithm based on the sample mass spectrometry data and the sample peptide sequencing results corresponding to the sample mass spectrometry data.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any of the peptide sequencing methods described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the peptide sequencing methods described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements any of the peptide sequencing methods described above.

[0017] The peptide sequencing method, apparatus, device, storage medium, and program products provided by this invention use a peptide sequencing model based on the TSARseqNovo algorithm to analyze and predict the mass spectrometry data of the peptide to be sequenced, and obtain peptide sequencing results. Compared with the existing CasaNovo model, the peptide sequencing model based on the TSARseqNovo algorithm has better feature encoding and inference capabilities, which can significantly improve the accuracy of mass spectrometry information analysis, improve the inference efficiency of the model, and effectively ensure the quality and accuracy of peptide sequencing results. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts of the peptide sequencing method provided by the present invention.

[0020] Figure 2 This is the second schematic diagram of the peptide sequencing method provided by the present invention.

[0021] Figure 3 This is one of the schematic diagrams of parallel inference of the semi-autoregressive decoding module provided by the present invention.

[0022] Figure 4 This is the second schematic diagram of parallel inference of the semi-autoregressive decoding module provided by the present invention.

[0023] Figure 5 This is a schematic diagram of the processing flow of the semi-autoregressive decoding module provided by the present invention.

[0024] Figure 6 This is one of the schematic diagrams showing the evaluation results of different models provided by this invention on nine species datasets.

[0025] Figure 7 This is the second illustration of the evaluation results of different models provided by this invention on nine species datasets.

[0026] Figure 8 This is the third illustration of the evaluation results of different models provided by this invention on nine species datasets.

[0027] Figure 9 This is one of the schematic diagrams illustrating the evaluation results of different models provided by this invention on a merged high-resolution dataset.

[0028] Figure 10This is the second illustration of the evaluation results of different models provided by this invention on a merged high-resolution dataset.

[0029] Figure 11 This is the third illustration of the evaluation results of different models provided by this invention on a merged high-resolution dataset.

[0030] Figure 12 This is a schematic diagram of the peptide sequencing device provided by the present invention.

[0031] Figure 13 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0033] Please see Figure 1 , Figure 1 This is one of the flowcharts illustrating the peptide sequencing method provided by the present invention. In this embodiment, the peptide sequencing method includes steps S110 to S120, each step of which is detailed below: S110: Obtain mass spectrometry data of the peptide to be sequenced.

[0034] S120: Input mass spectrometry data into the peptide sequencing model to obtain the peptide sequencing results output by the peptide sequencing model.

[0035] The peptide sequencing model is obtained by training an initial model based on the TSARseqNovo algorithm, which is based on sample mass spectrometry data and the corresponding sample peptide sequencing results.

[0036] The peptide sequencing method provided in this embodiment uses a peptide sequencing model based on the TSARseqNovo algorithm to analyze and predict the mass spectrometry data of the peptide to be sequenced, and obtain peptide sequencing results. Compared with the existing CasaNovo model, the peptide sequencing model based on the TSARseqNovo algorithm has better feature encoding and reasoning capabilities, which can significantly improve the accuracy of mass spectrometry information analysis, improve the model's reasoning efficiency, and effectively ensure the quality and accuracy of peptide sequencing results.

[0037] In some embodiments, the peptide sequencing model includes a mass spectrometry peak information fusion encoding module, a semi-autoregressive decoding module, and a mask optimization decoding module. The mass spectrometry peak information fusion encoding module is used to extract features from mass spectrometry data to obtain mass spectrometry peak fusion embedding features. The semi-autoregressive decoding module is used to perform initial prediction based on the mass spectrometry peak fusion embedding features to obtain amino acid prediction results for the peptide to be sequenced. The mask optimization decoding module is used to perform optimized prediction based on the amino acid prediction results to obtain peptide sequencing results.

[0038] Please see Figure 2 , Figure 2 This is the second schematic diagram of the peptide sequencing method provided by the present invention.

[0039] like Figure 2 As shown, this embodiment uses a peptide sequencing model based on the TSARseqNovo algorithm. This model includes a mass spectrometry peak information fusion encoding (MS PeakFusion Encoder) module, a semi-autoregressive (SAR) decoding module, and a masking refinement (MR) decoding module.

[0040] Among them, the mass spectrometry peak information fusion encoding module is used to extract features from mass spectrometry data and obtain mass spectrometry peak fusion embedding features.

[0041] Specifically, the mass spectrometry peak information fusion encoding module is mainly used to calculate the mass spectrometry peak information fusion embedding features: In this module, the mass spectrometry data is first encoded, taking into account the mass spectrometry peak encoding and the local and global mass spectrometry feature encoding; then the feature encoding is fused to obtain the mass spectrometry peak fusion encoding; then the mass spectrometry peak fusion encoding is input into the standard Transformer encoder to obtain the mass spectrometry peak fusion embedding features.

[0042] The semi-autoregressive decoding module is used to perform initial predictions based on mass spectrometry peak fusion embedding features to obtain amino acid prediction results for the peptide to be sequenced.

[0043] Specifically, the semi-autoregressive decoding module encodes amino acid information. It performs a positional addition operation on the positional and mass information of the amino acids in the peptide to be sequenced, obtaining the amino acid encoding information corresponding to the peptide. The initial identifier for each peptide is set to [b], and the end identifier is set to [e]. The length of each peptide is 100, meaning the longest encoding length for an amino acid is 100. After obtaining the amino acid encoding information, the semi-autoregressive decoding module uses the amino acid encoding information and mass spectrometry peak fusion embedding features as input. The module calculates a new set of amino acid identifiers and their confidence levels. The new amino acid prediction results replace the current amino acid identifiers, and the encoding is re-encoded, repeating the above process. This process iterates until a predetermined stop identifier [e] is encountered or the number of amino acids reaches 100. During this process, for each set of prediction results, amino acid identifiers with prediction confidence levels less than a preset threshold are selected and marked as [m], thus obtaining the amino acid prediction results for the peptide to be sequenced.

[0044] The mask optimization and decoding module is used to optimize predictions based on amino acid prediction results to obtain peptide sequencing results.

[0045] Specifically, the mask optimization decoding module can perform secondary prediction based on the amino identifier corresponding to the [m] position and the mass spectrometry peak fusion embedding feature until the prediction confidence of all predicted amino identifiers is greater than or equal to the preset threshold, thereby generating the final peptide sequencing result and optimizing amino acids to improve the overall peptide prediction accuracy.

[0046] In some embodiments, the mass spectrometry peak information fusion encoding module includes a mass spectrometry peak encoding module, a local and global mass spectrometry feature encoding module, a mass spectrometry peak fusion encoding module, and multiple encoders. The mass spectrometry peak encoding module is used to perform feature encoding on the mass spectrometry data to obtain mass spectrometry peak embedding encoding. The local and global mass spectrometry feature encoding module is used to perform feature encoding on the mass spectrometry data to obtain local and global spectral feature embedding encoding. The mass spectrometry peak fusion encoding module is used to perform feature fusion on the mass spectrometry peak embedding encoding and the local and global spectral feature embedding encoding to obtain mass spectrometry peak fusion encoding. All encoders are used to generate mass spectrometry peak fusion embedding features based on the mass spectrometry peak fusion encoding.

[0047] In the Transformer-based model architecture, data embedding and encoding can reflect subtle changes between different spectral peaks and differences between different amino acids, thus playing a crucial role in subsequent peptide prediction and analysis. The CasaNovo model employs sine-cosine encoding to convert the mass-to-charge ratio and intensity of mass spectrometry data into high-dimensional vector embeddings. While this embedding method can reflect the differences between different spectral peaks, it fails to consider the correlations between peaks. Therefore, the peptide sequencing model based on the TSARseqNovo algorithm introduces a novel embedding and encoding method, which relies on three main components: a mass spectrometry peak encoding module, a local and global mass spectrometry feature encoding module (i.e., a local-global mass spectrometry feature encoding module), and a mass spectrometry peak fusion encoding module.

[0048] Specifically, the mass spectrometry peak information fusion coding module includes a mass spectrometry peak coding module, a local and global mass spectrometry feature coding module, a mass spectrometry peak fusion coding module, and multiple encoders.

[0049] The mass spectrometry peak encoding module uses the sine-cosine encoding technique in the CasaNovo algorithm to encode the features of the mass spectrometry data, converting the peak information of the mass spectrometry data into a high-dimensional vector to obtain the mass spectrometry peak embedding code.

[0050] The embedding formula for each peak is as follows: In the above formula, This represents the mass-to-charge ratio of the i-th peak. Represents the dimensionality of the entire spectral peak embedding. This represents the minimum change in wavelength between the sine and cosine waves. This represents the maximum value of the change in wavelength between the sine and cosine waves.

[0051] This embedding coding method can convert the mass-to-charge ratio and intensity in mass spectrometry data into vector information in a high-dimensional space.

[0052] The local-global mass spectrometry feature encoding module is used to encode the features of mass spectrometry data to obtain local and global spectral feature embedding codes (i.e., local-global spectral feature embedding codes).

[0053] In this embodiment, the local-global mass spectrometry feature encoding module comprehensively considers the spatial relationships between different spectral peaks in the mass spectrometry data, integrating spatial and local information features. The local-global spectral feature embedding encoding includes global intensity ranking ratio, global half-intensity ranking ratio, local intensity ranking ratio, local half-intensity ranking ratio, global relative intensity magnitude, and local relative magnitude, as detailed below: (1) Global strength ranking ratio ( ): The ranking is the result of sorting the intensity of each spectral peak from largest to smallest. The total number of spectral peaks.

[0054] (2) Global half-strength ranking ratio ( ): The ranking is the result of sorting the spectral peaks from largest to smallest after halving their intensity. The total number of spectral peaks.

[0055] (3) Local intensity ranking ratio ( The intensity of a spectral peak within a specified window size is defined as local information. In this embodiment, the window size is set to... ,in This represents the ranking of local spectral peak intensities from largest to smallest. This represents the total number of local spectral peaks.

[0056] (4) Local half-intensity ranking ratio ( ): This is the ranking of the local spectral peaks after their intensities have been halved, sorted from largest to smallest. This represents the total number of local spectral peaks.

[0057] (5) Global relative intensity ( ): The size of a certain spectral peak It is the peak value with the highest intensity among all spectral peaks.

[0058] (6) Local relative size ( ): The size of a certain spectral peak It is the peak value with the highest intensity among the local spectral peaks.

[0059] The local-global mass spectrometry feature encoding module can construct a single feature from the above six types of features. The matrix is ​​denoted as The matrix is ​​then passed through a transformation network branch. Mapping to a high-dimensional space enables pattern encoding independent of spectral peak order, resulting in the final local-global spectral feature embedding encoding.

[0060] The mass spectrometry peak fusion coding module is used to perform feature fusion on mass spectrometry peak embedding coding and local and global spectral feature embedding coding to obtain mass spectrometry peak fusion coding.

[0061] The fusion formula for each peak is as follows: ; in, , , The embedding encoding result represents the i-th peak; The mass-to-charge ratio represents mass spectrometry information; The intensity of mass spectrometry information; This represents the local-global spectral feature embedding encoding of the transformation network branch processing.

[0062] All encoders are used to generate mass spectrum peak fusion embedding features based on mass spectrum peak fusion encoding.

[0063] Specifically, in the calculation of mass spectrum peak fusion embedding features, six standard Transformer architecture encoders are used. Each encoder consists of a multi-head self-attention module and a feedforward network to analyze mass spectrum peak fusion encoding and generate mass spectrum peak fusion embedding features.

[0064] The advanced semantic features extracted by the TSARseqNovo model can effectively help the subsequent semi-autoregressive decoding module and mask optimization decoding module improve their feature extraction capabilities, so as to more effectively identify and interpret complex features in mass spectrometry data.

[0065] The basic operation of the encoder can be expressed by the following formula: ; in, , and These represent the three matrix representations learned from the linear mapping component in the standard Transformer architecture encoder. Therefore, the output of the encoder part can be denoted as: ; in, Represents the encoders from the first to the kth. Encoding for mass spectrometry peak fusion.

[0066] The peptide sequencing method provided in this embodiment adopts a new mass spectrometry peak information fusion encoding module and proposes a new encoding strategy. It integrates mass spectrometry peak encoding and local-global mass spectrometry feature encoding of mass spectrometry data, so that the final mass spectrometry peak fusion embedding features are not limited to mass-to-charge ratio and intensity information. This method can significantly enhance the performance of the model.

[0067] In some embodiments, an initial prediction is performed based on mass spectrometry peak fusion embedding features to obtain the amino acid prediction results of the peptide to be sequenced, including: generating amino coding information of the peptide to be sequenced; and performing an initial prediction based on the amino coding information and mass spectrometry peak fusion embedding features to obtain the amino acid prediction results of the peptide to be sequenced.

[0068] In the TSARseqNovo model, in order to ensure that the model can effectively consider the information of individual amino acids, this embodiment introduces a dynamic amino acid encoding method in the semi-autoregressive decoding module, which takes into account the mass, position and specific characteristics of each amino acid. This allows the corresponding encoding to effectively capture the relationships between peptide segments and the characteristics of amino acids, such as their mass and relative position, when the model training ends.

[0069] Specifically, the semi-autoregressive decoding module encodes amino acid information to generate amino-encoded information for the peptide to be sequenced. This amino-encoded information includes the amino acid identifier of the peptide. The predicted sequence of the amino acid identifier uses [b] as a start marker. This marker, along with the mass spectrometry peak fusion embedding feature, serves as the input to the semi-autoregressive decoding module (also known as the SAR decoder) to calculate its hidden state vector and new amino identifiers. The predicted amino identifier is then re-encoded, and new amino acid identifiers are iteratively generated repeatedly until a stop marker [e] is encountered or the maximum length of 100 is reached. The amino acid encoding process of the semi-autoregressive decoding module can be expressed by the following formula: ; ; In the above formula, This represents the hidden unit vector of the last layer of the SAR decoder; This represents the sequence result predicted by the SAR decoder; This represents the information encoding corresponding to the amino acid output by the SAR decoder. It should be noted that when the current peptide sequence length reaches 100 or when the end symbol [e] is encountered, it needs to be reset to [b].

[0070] The process involves iterative prediction until a predetermined stop identifier [e] is encountered or the number of amino acids reaches 100. During this process, for each set of prediction results, amino acid identifiers with prediction confidence less than a preset threshold are selected and marked as [m] to obtain the amino acid prediction results of the peptide to be sequenced.

[0071] It should be noted that in the semi-autoregressive decoding module, the input data consists of mass spectrometry peak fusion embedding features and amino coding information. The output amino acid prediction results include multiple amino identifiers of the peptide to be sequenced and the prediction confidence of each amino identifier. The semi-autoregressive decoding module introduces the idea of ​​semi-autoregression. By incorporating this idea, the semi-autoregressive decoding module can process k amino coding information in parallel. Compared to the traditional method of the CasaNovo model, which can only process one amino coding information at a time, the method in this embodiment can effectively improve the inference efficiency of the model.

[0072] Please see Figures 3 to 5 , Figure 3This is one of the schematic diagrams of parallel inference in the semi-autoregressive decoding module provided by the present invention. Figure 4 This is the second schematic diagram of parallel inference of the semi-autoregressive decoding module provided by the present invention. Figure 5 This is a schematic diagram of the processing flow of the semi-autoregressive decoding module provided by the present invention.

[0073] like Figure 3 and Figure 5 As shown, for a peptide of length M to be sequenced, it can be divided into... Multiple groups process simultaneously; such as Figure 4 As shown, each group may include k amino code information.

[0074] Accordingly, in order to achieve parallel processing, the SAR decoder has also improved its corresponding mask matrix, as shown in the following formula: in, ,and This represents a floating-point operation. This operation allows the attention mechanism to focus on at most k subsequent positions from a given mark, thus enabling parallel processing. Its attention mechanism uses the following formula: ; in, , , These represent the Query matrix, Key matrix, and Value matrix in the attention mechanism, respectively. This corresponds to the vector dimension. This modification of the traditional attention mechanism enables parallel group prediction, enhancing the model's computational efficiency for relevant peptide features.

[0075] Unlike the CasaNovo model, which employs a stepwise prediction strategy based on the standard Transformer model, the peptide sequencing method presented in this embodiment introduces a semi-autoregressive approach. While enabling parallel prediction analysis of multiple amino groups, it comprehensively considers historical amino group information from the previous generation of results in the analysis results. This allows the final generated results to incorporate more features, effectively improving the accuracy of the model's prediction analysis results and significantly increasing the model's inference speed. Compared to the baseline model, its inference efficiency is improved by 2-3 times.

[0076] In some embodiments, the amino acid prediction results include multiple amino identifiers of the peptide to be sequenced and the prediction confidence of each amino identifier; optimizing the prediction based on the amino acid prediction results to obtain peptide sequencing results includes: determining the target amino identifier based on the amino acid prediction results; the target amino identifier is an amino identifier whose prediction confidence is less than a preset threshold; optimizing the prediction based on the target amino identifier and mass spectrometry peak fusion embedding features to generate peptide sequencing results.

[0077] Both the mask optimization decoding module (also known as the MR decoder) and the semi-autoregressive decoding module employ a simplified version of the self-attention mechanism instead of the multi-head attention mechanism. The main function of the mask optimization decoding module is to identify amino identifiers in the overall peptide with a predicted confidence level lower than a preset threshold. Based on the amino identifiers with a predicted confidence level lower than the preset threshold and the mass spectrometry peak fusion embedding features, the module performs prediction again to regenerate amino identifiers with a predicted confidence level greater than or equal to the preset threshold. This process continues until the predicted confidence level of all amino identifiers is greater than or equal to the preset threshold, at which point the final peptide sequencing result is generated, aiming to improve the accuracy of the overall sequence.

[0078] It should be noted that in the initial stage of model training, the model's prediction results have significant uncertainty, therefore its confidence score is set relatively low. As the model is gradually optimized, its prediction accuracy gradually improves. Therefore, during training, the confidence threshold can be gradually increased to ensure the reliability of the model results. During this process, amino identifiers with prediction confidence scores below the preset threshold are also marked as [m] for further analysis. The processing flow of the mask optimization decoding module can be expressed by the following formula: ; ; in, This indicates the hidden state of the last layer of the MR decoder; It is an amino acid identifier whose prediction confidence is less than a preset threshold; The optimized peptide results are generated by the MR decoder. The optimized output of the MR decoder is considered as the final result prediction in the TSARseqNovo model.

[0079] The peptide sequencing method provided in this embodiment optimizes the mask structure of the existing CasaNovo model by introducing a secondary optimization module, namely the mask optimization decoding module, into the model. For amino identifiers whose prediction confidence is less than a preset threshold in the semi-autoregressive decoding module prediction results, a secondary prediction is performed, which realizes fine optimization and adjustment of the model prediction and can improve the overall prediction accuracy of the model.

[0080] In some embodiments, the peptide sequencing model is trained using a preset loss function; wherein the preset loss function is determined based on a first loss function and a second loss function, the first loss function being the cross-entropy loss function of the semi-autoregressive decoding module, and the second loss function being the cross-entropy loss function of the mask-optimized decoding module.

[0081] Specifically, the peptide sequencing model is trained using a preset loss function, which considers the cross-entropy (CE) loss of the semi-autoregressive decoding module and the cross-entropy loss of the mask optimization decoding module. The semi-autoregressive decoding module calculates the loss between the predicted peptide and the real peptide, while the mask optimization decoding module focuses on calculating the loss between the optimized final peptide and the real peptide.

[0082] Preset loss function The calculation formula is as follows: ; in, For the cross-entropy loss of the semi-autoregressive decoding module, The first loss function; To optimize the cross-entropy loss of the decoding module using a mask, This is the second loss function.

[0083] To evaluate the performance of the TSARseqNovo model in this embodiment, it was compared with other established peptide sequence analysis tools, such as the PEAKs model, Novor model, DeepNovo model, PointNovo model, and CasaNovo model. The focus was on metrics such as peptide-level prediction accuracy, amino acid labeling recall, and amino acid precision, which met the previous evaluation criteria for the DeepNovo, PointNovo, and CasaNovo models.

[0084] Two main datasets were used for model evaluation: a nine-species benchmark dataset and a merged high-resolution dataset. The nine-species benchmark dataset, containing approximately 1.5 million MS spectra from nine different experiments, was used to evaluate the DeepNovo, PointNovo, and CasaNovo models. For the ground truth values ​​in these datasets, peptides with a false discovery rate of less than or equal to 1% from CasaNovo model-related papers were used as ground truth values ​​for both model training and evaluation. The merged high-resolution dataset, containing approximately 700,000 MS spectra from seven high-resolution proteomics datasets (PXD008808, PXD011246, PXD012645, PXD012979, PXD008844, and PXD010559), had ground truth values ​​derived from the π-HelixNovo study.

[0085] For validation, leave-one-out cross-validation, previously used by the DeepNovo and CasaNovo models, was employed. This method uses eight of the nine species datasets for training, reserving the ninth for testing. For the merged dataset, following the strategy of the PointNovo model, PXD008808, PXD011246, PXD012645, and PXD012979 were used as the training set. The remaining datasets, ABRF, PXD008844, and PXD010559, were used for testing, thus comprehensively evaluating the model's performance under different MS conditions.

[0086] Please see Figures 6 to 8 , Figure 6 This is one of the schematic diagrams illustrating the evaluation results of different models provided by this invention on nine species datasets. Figure 7 This is the second illustration of the evaluation results of different models provided by this invention on nine species datasets. Figure 8 This is the third illustration of the evaluation results of different models provided by this invention on nine species datasets.

[0087] like Figures 6 to 8 As shown, the TSARseqNovo model outperforms many existing models in terms of performance.

[0088] Depend on Figure 6 The results show that, compared to the traditional PEAKs and Novor models, the TSARseqNovo model achieves an overall performance improvement of 18.0% over the PEAKs model and 34.8% over the Novor model. Compared to existing deep learning-based DeepNovo, PointNovo, and CasaNovo models, the TSARseqNovo model achieves 12.6% higher peptide-level accuracy than the DeepNovo model, 11.4% higher than the PointNovo model, and 11.4% higher than the CasaNovo model (4.9%). It also shows improvements of 1.9%, 3.9%, 4.1%, 5.1%, 8.3%, 3.8%, and 6.8% in Human, Yeast, M. mazei, Honeybee, Tomato, Bacilus, and Clam bacteria, respectively.

[0089] Depend on Figure 7 As can be seen, in terms of amino acid accuracy, the TSARseqNovo model in this embodiment has improvements of 6.1%, 23.4%, 5.1%, 3.7% and 4.7% compared with the PEAKs model, Novor model, DeepNovo model, PointNovo model and CasaNovo model, respectively.

[0090] Depend on Figure 8 As can be seen, in terms of recall, the TSARseqNovo model in this embodiment has an improvement of 6.8%, 23.8%, 5.5%, 3.8%, and 5.0% compared to the PEAKs model, Novor model, DeepNovo model, PointNovo model, and CasaNovo model, respectively.

[0091] Please see Figures 9 to 11 , Figure 9 This is one of the schematic diagrams illustrating the evaluation results of different models provided by this invention on a merged high-resolution dataset. Figure 10 This is the second illustration of the evaluation results of different models provided by this invention on a merged high-resolution dataset. Figure 11 This is the third illustration of the evaluation results of different models provided by this invention on a merged high-resolution dataset.

[0092] In the merged high-resolution dataset, the TSARseqNovo model of this embodiment is trained on the sub-datasets PXD008808, PXD011246, PXD012645 and PXD012979, and inference validated on the sub-datasets ABRF, PXD008844 and PXD010559.

[0093] Depend on Figure 9 As can be seen, the TSARseqNovo model significantly outperforms the CasaNovo and PointNovo models, achieving performance improvements of 7.0% and 3.3%, respectively.

[0094] However, the TSARseqNovo model shows some decline in both the accuracy and recall of amino acid predictions.

[0095] Depend on Figure 10 and Figure 11 The results show that the TSARseqNovo model performs well on the ABRF dataset, improving accuracy by 3.2% and recall by 3.6% compared to the PointNovo model. In contrast, the TSARseqNovo model shows a 1.6% decrease in accuracy and a 2.2% decrease in recall on the PXD008844 dataset; a 1.6% decrease in accuracy and a 2.2% decrease in recall on the PXD010559 dataset; and a 3.8% and 3.1% decrease in accuracy and recall, respectively, on the PXD01 dataset.

[0096] To further explore the performance of the TSARseqNovo model in terms of inference speed and accuracy, the Yeast dataset was used for further evaluation, and the results are shown in Table 1.

[0097] Table 1

[0098] Table 1 shows that the TSARseqNovo model investigated the impact of parallel decoding lengths for decoding two and three amino acids on inference speed and accuracy. The results indicate that increasing the decoding length of the TSARseqNovo model significantly improves its inference speed. Specifically, when the inference length is set to 3, the processing speed reaches 234 mass spectrometry data points per second, twice that of the CasaNovo model. At this speed, the TSARseqNovo model's accuracy (52.1%) surpasses that of the CasaNovo model (50.5%). Although the accuracy of the TSARseqNovo model is slightly lower than that of the PointNovo model, its inference speed is approximately 10 times faster. Furthermore, when the parallel decoding length is set to 3, the TSARseqNovo model's predicted amino acid accuracy decreases from 76.1% to 74.8%, amino acid recall decreases from 75.3% to 74.6%, and peptide accuracy decreases from 54.4% to 52.1%. This indicates that while longer decoding lengths improve inference speed, they directly impact the accuracy of the TSARseqNovo model.

[0099] Overall, the TSARseqNovo model shows a significant improvement in inference speed while maintaining high accuracy. Therefore, it has great application value as a powerful tool for peptide sequence analysis.

[0100] This invention also provides a peptide sequencing device. Please refer to [link / reference]. Figure 12 , Figure 12 This is a schematic diagram of the peptide sequencing device provided by the present invention. In this embodiment, the peptide sequencing device includes an acquisition module 1210 and a sequencing module 1220.

[0101] The acquisition module 1210 is used to acquire mass spectrometry data of the peptide to be sequenced.

[0102] The sequencing module 1220 is used to input the mass spectrometry data into the peptide sequencing model to obtain the peptide sequencing results output by the peptide sequencing model.

[0103] The peptide sequencing model is obtained by training an initial model based on the TSARseqNovo algorithm, using sample mass spectrometry data and the corresponding sample peptide sequencing results.

[0104] In some embodiments, the peptide sequencing model includes a mass spectrometry peak information fusion encoding module, a semi-autoregressive decoding module, and a mask optimization decoding module. The mass spectrometry peak information fusion encoding module is used to extract features from mass spectrometry data to obtain mass spectrometry peak fusion embedding features. The semi-autoregressive decoding module is used to perform initial prediction based on the mass spectrometry peak fusion embedding features to obtain amino acid prediction results for the peptide to be sequenced. The mask optimization decoding module is used to perform optimized prediction based on the amino acid prediction results to obtain peptide sequencing results.

[0105] In some embodiments, the mass spectrometry peak information fusion encoding module includes a mass spectrometry peak encoding module, a local and global mass spectrometry feature encoding module, a mass spectrometry peak fusion encoding module, and multiple encoders. The mass spectrometry peak encoding module is used to perform feature encoding on the mass spectrometry data to obtain mass spectrometry peak embedding encoding. The local and global mass spectrometry feature encoding module is used to perform feature encoding on the mass spectrometry data to obtain local and global spectral feature embedding encoding. The mass spectrometry peak fusion encoding module is used to perform feature fusion on the mass spectrometry peak embedding encoding and the local and global spectral feature embedding encoding to obtain mass spectrometry peak fusion encoding. All encoders are used to generate mass spectrometry peak fusion embedding features based on the mass spectrometry peak fusion encoding.

[0106] In some embodiments, an initial prediction is performed based on mass spectrometry peak fusion embedding features to obtain the amino acid prediction results of the peptide to be sequenced, including: generating amino coding information of the peptide to be sequenced; and performing an initial prediction based on the amino coding information and mass spectrometry peak fusion embedding features to obtain the amino acid prediction results of the peptide to be sequenced.

[0107] In some embodiments, the amino acid prediction results include multiple amino identifiers of the peptide to be sequenced and the prediction confidence of each amino identifier; optimizing the prediction based on the amino acid prediction results to obtain peptide sequencing results includes: determining the target amino identifier based on the amino acid prediction results; the target amino identifier is an amino identifier whose prediction confidence is less than a preset threshold; optimizing the prediction based on the target amino identifier and mass spectrometry peak fusion embedding features to generate peptide sequencing results.

[0108] In some embodiments, the peptide sequencing model is trained using a preset loss function; wherein the preset loss function is determined based on a first loss function and a second loss function, the first loss function being the cross-entropy loss function of the semi-autoregressive decoding module, and the second loss function being the cross-entropy loss function of the mask-optimized decoding module.

[0109] The present invention also provides an electronic device. Figure 13 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 13As shown, the electronic device may include a processor 1310, a communications interface 1320, a memory 1330, and a communication bus 1340, wherein the processor 1310, the communications interface 1320, and the memory 1330 communicate with each other via the communication bus 1340. The processor 1310 can call logical instructions in the memory 1330 to execute peptide sequencing methods.

[0110] Furthermore, the logical instructions in the aforementioned memory 1330 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0111] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the peptide sequencing methods provided by the methods described above.

[0112] The present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to execute the peptide sequencing methods provided by the above methods.

[0113] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0114] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A peptide sequencing method, characterized in that, include: Obtain mass spectrometry data of the peptide to be sequenced; The mass spectrometry data is input into the peptide sequencing model to obtain the peptide sequencing results output by the peptide sequencing model. The peptide sequencing model is obtained by training an initial model based on the TSARseqNovo algorithm based on sample mass spectrometry data and the corresponding sample peptide sequencing results. The peptide sequencing model includes a mass spectrometry peak information fusion encoding module, a semi-autoregressive decoding module, and a mask optimization decoding module. The mass spectrometry peak information fusion encoding module is used to extract features from the mass spectrometry data to obtain mass spectrometry peak fusion embedding features; The semi-autoregressive decoding module is used to perform initial prediction based on the mass spectrometry peak fusion embedding features to obtain the amino acid prediction results of the peptide to be sequenced. The mask optimization and decoding module is used to optimize the prediction based on the amino acid prediction results to obtain the peptide sequencing results. The mass spectrometry peak information fusion coding module includes a mass spectrometry peak coding module, a local and global mass spectrometry feature coding module, a mass spectrometry peak fusion coding module, and multiple encoders; The mass spectrometry peak encoding module is used to perform feature encoding on the mass spectrometry data to obtain mass spectrometry peak embedding encoding; The embedding formula for each peak is as follows: ； in, The mass-to-charge ratio represents the i-th peak; The dimension of the spectral peak embedding is represented; This represents the minimum change in wavelength between the sine and cosine waves. This represents the maximum value of the change between the sine and cosine wavelengths; The local and global mass spectrometry feature encoding module is used to perform feature encoding on the mass spectrometry data to obtain local and global spectral feature embedding codes; The mass spectrum peak fusion coding module is used to perform feature fusion on the mass spectrum peak embedding coding and the local and global spectral feature embedding coding to obtain the mass spectrum peak fusion coding; All of the encoders are used to generate the mass spectrum peak fusion embedding features based on the mass spectrum peak fusion encoding; The amino acid encoding process of the semi-autoregressive decoding module is expressed by the following formula: ； ； in, This represents the hidden unit vector of the last layer of the semi-autoregressive decoding module; This represents the sequence result predicted by the semi-autoregressive decoding module; This represents the information encoding corresponding to the amino acid output by the semi-autoregressive decoding module; The processing flow of the mask optimization decoding module is expressed by the following formula: ； ； in, It is an amino acid identifier whose prediction confidence is less than a preset threshold.

2. The peptide sequencing method according to claim 1, characterized in that, The initial prediction based on the mass spectrometry peak fusion embedding features to obtain the amino acid prediction results of the peptide to be sequenced includes: Generate the amino coding information of the peptide to be sequenced; Initial predictions are made based on the amino-encoding information and the mass spectrometry peak fusion embedding features to obtain the amino acid prediction results of the peptide to be sequenced.

3. The peptide sequencing method according to claim 1, characterized in that, The amino acid prediction results include multiple amino identifiers of the peptide to be sequenced and the prediction confidence of each amino identifier; The optimization prediction based on the amino acid prediction results to obtain the peptide sequencing results includes: Based on the amino acid prediction results, a target amino acid identifier is determined; the target amino acid identifier is an amino acid identifier whose prediction confidence is less than a preset threshold. The peptide sequencing results are generated by optimizing the prediction based on the target amino identifier and the mass spectrometry peak fusion embedding features.

4. The peptide sequencing method according to claim 1, characterized in that, The peptide sequencing model is trained using a preset loss function; The preset loss function is determined based on a first loss function and a second loss function. The first loss function is the cross-entropy loss function of the semi-autoregressive decoding module, and the second loss function is the cross-entropy loss function of the mask-optimized decoding module.

5. A peptide sequencing device, characterized in that, include: The acquisition module is used to acquire mass spectrometry data of the peptide to be sequenced; The sequencing module is used to input the mass spectrometry data into the peptide sequencing model to obtain the peptide sequencing results output by the peptide sequencing model. The peptide sequencing model is obtained by training an initial model based on the TSARseqNovo algorithm based on sample mass spectrometry data and the corresponding sample peptide sequencing results. The peptide sequencing model includes a mass spectrometry peak information fusion encoding module, a semi-autoregressive decoding module, and a mask optimization decoding module. The mass spectrometry peak information fusion encoding module is used to extract features from the mass spectrometry data to obtain mass spectrometry peak fusion embedding features; The semi-autoregressive decoding module is used to perform initial prediction based on the mass spectrometry peak fusion embedding features to obtain the amino acid prediction results of the peptide to be sequenced. The mask optimization and decoding module is used to optimize the prediction based on the amino acid prediction results to obtain the peptide sequencing results. The mass spectrometry peak information fusion coding module includes a mass spectrometry peak coding module, a local and global mass spectrometry feature coding module, a mass spectrometry peak fusion coding module, and multiple encoders; The mass spectrometry peak encoding module is used to perform feature encoding on the mass spectrometry data to obtain mass spectrometry peak embedding encoding; The embedding formula for each peak is as follows: ； in, The mass-to-charge ratio represents the i-th peak; The dimension of the spectral peak embedding is represented; This represents the minimum change in wavelength between the sine and cosine waves. This represents the maximum value of the change between the sine and cosine wavelengths; The local and global mass spectrometry feature encoding module is used to perform feature encoding on the mass spectrometry data to obtain local and global spectral feature embedding codes; The mass spectrum peak fusion coding module is used to perform feature fusion on the mass spectrum peak embedding coding and the local and global spectral feature embedding coding to obtain the mass spectrum peak fusion coding; All of the encoders are used to generate the mass spectrum peak fusion embedding features based on the mass spectrum peak fusion encoding; The amino acid encoding process of the semi-autoregressive decoding module is expressed by the following formula: ； ； in, This represents the hidden unit vector of the last layer of the semi-autoregressive decoding module; This represents the sequence result predicted by the semi-autoregressive decoding module; This represents the information encoding corresponding to the amino acid output by the semi-autoregressive decoding module; The processing flow of the mask optimization decoding module is expressed by the following formula: ； ； in, It is an amino acid identifier whose prediction confidence is less than a preset threshold.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the peptide sequencing method as described in any one of claims 1 to 4.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the peptide sequencing method as described in any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the peptide sequencing method as described in any one of claims 1 to 4.

Images