Selenium peptide de novo sequencing method based on full convolution attention mechanism neural network

Abstract

The invention provides a selenium peptide de novo sequencing method based on a full convolution attention mechanism neural network, and belongs to the field of selenium-containing natural polypeptide identification and detection.Specific seleno-amino acid corresponding to each characteristic fragment in all selenium peptides is used as initial amino acid, then the next symbol in the N / C end direction is sequentially predicted, and the sequence of the next symbol in the N / C end direction is determined; taking the symbol with the highest probability as the symbol at the corresponding position; learning selenium peptide N and C terminal amino acid features in a known selenium peptide sequence by using a full convolution attention mechanism neural network, then carrying out feature fusion, and standardizing to 0-1; taking all identified selenium peptide sequences as an input layer, performing data feature separation by using an activation function of a neural network hidden layer, calculating each separation feature, obtaining the probability that the corresponding selenium peptide sequence is a natural selenium peptide, removing unmatched selenium peptide sequences, and outputting the identified natural selenium peptide sequence by an output layer; and optimizing the weight and bias of the model, inputting all identified selenium peptide sequences into the optimization model, obtaining a sequencing model when the error between the selenium peptide sequences and a real sequence does not change any more, and inputting the selenium peptide sequences correspondingly identified by all selenium peptides of a to-be-detected sample into the sequencing model to obtain the natural selenium peptide sequence identified by the to-be-detected sample.

Description

Technical Field

[0001] The present invention belongs to the field of identification and detection of natural selenium-containing peptides, and specifically relates to a seleno peptide de novo sequencing method based on a fully convolutional attention mechanism neural network. Background Art

[0002] Selenium is an essential trace element for human growth and metabolism. It has been proven to have physiological functions such as anti-oxidation, anti-inflammatory, and improving the body's immunity. As a biologically active form of selenium, natural selenopeptides are an important source of organic selenium. The difference in amino acid composition and arrangement order will affect the biological activity of selenopeptides. Accurate determination of the amino acid sequence of selenopeptides is of great significance for exploring the structure-activity relationship of selenopeptides, evaluating the content and types of selenopeptides in food, rational diet, and preventing nutritional deficiencies.

[0003] The existing peptide identification methods are mainly based on amino acid sequence databases and spectral library search and identification. Due to the specificity of selenopeptides and their relative scarcity in databases, the existing databases lack complete selenopeptide sequence information. The selenium element in selenopeptides usually exists in the form of selenocysteine ​​(SeC) or selenomethionine (SeM). The introduction of selenoamino acids will change the mass spectrometry characteristics of peptides. The existing amino acid sequence spectral library and search algorithm do not fully consider the particularity of selenoamino acids, and cannot accurately identify selenopeptide spectra during the matching process.

[0004] In summary, due to the incompleteness of the existing amino acid sequence database, the diversity of selenopeptide sequences, the complexity of selenopeptide modifications, the lag in updating the amino acid spectrum library and the inadequacy of the search engine scoring algorithm, the current analytical methods are not sufficient to achieve the effective identification of natural selenopeptides. Summary of the invention

[0005] In response to the problems existing in the prior art, the present invention provides a method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network, which realizes the accurate sequencing of natural selenopeptides by utilizing the fully convolutional attention mechanism neural network.

[0006] The present invention is achieved through the following technical solutions: A method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network comprises the following steps: S1, taking the specific selenoamino acid corresponding to each characteristic fragment in all selenopeptides as the starting amino acid, and then predicting the next possible symbol in the N / C terminal direction of the selenopeptide sequence in turn, taking the symbol with the highest probability as the symbol of the corresponding position, and obtaining the corresponding identified selenopeptide sequence; S2, using a fully convolutional attention mechanism neural network to learn the N- and C-terminal amino acid features of selenopeptides in known selenopeptide sequences, then fusion the obtained features, and then normalize all features to 0 to 1 to obtain a standardized data set; All identified selenopeptide sequences are used as the input layer, and the activation function of the hidden layer of the neural network is used to separate the data features to obtain the corresponding separation features, and each separation feature is calculated using the standardized data set to obtain the probability that the corresponding selenopeptide sequence is a natural selenopeptide, and the unmatched selenopeptide sequences are removed, and the output layer outputs the identified natural selenopeptide sequence; The input layer, hidden layer and output layer form an initial model, optimize the weights and biases in the initial model to obtain an optimized model, input all identified selenopeptide sequences into the optimized model to obtain an optimized selenopeptide sequence, and obtain a sequencing model when the error between the optimized selenopeptide sequence and the true sequence no longer changes; S3, all the selenopeptides of the sample to be tested are subjected to the process of S1 to obtain the corresponding identified selenopeptide sequences, which are then input into the sequencing model to obtain the natural selenopeptide sequences identified by the sample to be tested.

[0007] Preferably, the characteristic fragment described in S1 is obtained as follows: The secondary spectra of all selenopeptides were converted into matrices containing molecular formula, mass-to-charge ratio and abundance, and then several common fragments were captured. If there was a mass spectrum peak cluster consistent with the selenium isotope peak cluster, the corresponding common fragment was the characteristic fragment of the selenopeptide.

[0008] Preferably, S1 inputs a matrix containing molecular formula, mass-to-charge ratio and abundance into a residual dense attention network for capture to obtain several common fragments.

[0009] Preferably, S1 combines fragment ions a, a 2+ , a-H2O, a-NH3, b, b 2+ , b-H2O, b-NH3, y, y 2+ , y-H2O and y-NH3 information, and use the deep learning-based target detection algorithm Ion-CNN to predict the next possible symbol in the N / C-terminal direction of the selenopeptide sequence in turn.

[0010] Preferably, S1 calculates the probability of the symbol score at each position in all peptide sequences using the LogSoftmax function, and selects the symbol with the highest probability as the symbol at the corresponding position.

[0011] Preferably, S2 inputs the obtained features into the fully connected layer of the fully convolutional attention mechanism neural network for fusion.

[0012] Preferably, S2 applies the Softmax function to each feature to normalize all features to between 0 and 1, thereby obtaining a standardized data set.

[0013] Preferably, S2 uses the ReLu function as the activation function of the hidden layer of the neural network to separate the data features in all identified selenopeptide sequences.

[0014] Preferably, the unmatched selenopeptide sequence in S2 is a selenopeptide sequence with a probability of less than 60%.

[0015] Preferably, S2 uses a cross entropy loss function to calculate the error between the optimized selenopeptide sequence and the true sequence.

[0016] Compared with the prior art, the present invention has the following beneficial technical effects: The present invention discloses a method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network. The specific selenoamino acid corresponding to each characteristic fragment in all selenopeptides is used as the starting amino acid for selenopeptide identification, which facilitates bidirectional simultaneous sequencing of the N / C terminals, and then the symbol with the highest probability can be the symbol of the corresponding position to obtain the corresponding identified selenopeptide sequence; the known selenopeptide sequence is learned by a fully convolutional attention mechanism neural network to obtain the N- and C-terminal amino acid features of the selenopeptide, and then feature fusion and standardization are performed, so that the activation function is used to separate the data features of all identified selenopeptide sequences, and then the identified mismatched selenopeptide sequences are screened out, and selenium-containing peptides in food are screened to obtain a list of natural selenopeptides. By optimizing this initial model, the error between the optimized selenopeptide sequence and the true sequence can no longer change, and finally the final sequencing model is used to sequence the selenopeptide sequences identified by all selenopeptides of the sample to be tested, and the natural selenopeptide sequence identified by the sample to be tested can be obtained, which can achieve accurate identification of selenopeptides. The natural selenopeptide identification method established by the present invention does not rely on existing databases, and can directly use secondary spectra and fragment ion information to infer peptide sequences, which is conducive to the discovery of new selenopeptides and the accurate analysis of the amino acid sequence of selenopeptides, helps to reveal the metabolic pathway of selenium in organisms, and provides a scientific basis for in-depth research on the biological effects of selenium. DETAILED DESCRIPTION

[0017] The principles and advantages of the present invention are further explained and illustrated in detail below in conjunction with specific embodiments so that those skilled in the art can better understand the present invention. The following description is only exemplary and does not limit its content.

[0018] The present invention provides a method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network, which specifically comprises the following steps: Step 1: Search in the public database to obtain the secondary spectra of all selenopeptides, and divide 88-92% into the training set, and the rest into the validation set and the test set in equal proportion; Step 2: Convert the secondary spectra of all selenopeptides in the training set into matrices containing molecular formula, mass-to-charge ratio and abundance, and then input them into the residual dense attention network to capture several common fragments. If there is a mass spectrum peak cluster that is consistent with the selenium isotope peak cluster, the corresponding common fragment is regarded as the characteristic fragment of the selenopeptide.

[0019] The amount of selenium contained in each selenopeptide is determined based on the abundance of the isotope peak cluster of the characteristic fragment of the selenopeptide. Step 3: According to the characteristic fragment of each selenopeptide, the specific type of selenoamino acid corresponding to it is determined, and the selenoamino acid is used as the starting amino acid for selenopeptide identification, and the fragment ions a, a 2+ , a-H2O, a-NH3, b, b 2+ , b-H2O, b-NH3, y, y 2+ , y-H2O and y-NH3 information, and use the deep learning-based target detection algorithm Ion-CNN to predict the next possible symbol in the N / C-terminal direction of the selenopeptide sequence in turn. The symbol is one of 23 symbols, including 20 standard amino acids, selenocysteine, selenomethionine and terminator.

[0020] The LogSoftmax function is used to calculate the probability of the sequence symbol score at each position in all peptide sequences, and the sequence symbol with the highest probability is selected as the sequence symbol at that position.

[0021] Step 4: Using the known selenopeptide sequences in the training set as input, the fully convolutional attention mechanism neural network is used to learn the N- and C-terminal amino acid features of selenopeptides. The obtained features are then input into the fully connected layer of the neural network for feature fusion. The Softmax function is applied to each feature, and all features are standardized to 0 to 1, indicating the importance of each input, which is recorded as the standardized data set.

[0022] All the selenopeptide sequences identified in step 3 are used as the input layer, and the ReLu function is used as the activation function of the hidden layer to separate the data features in the identified selenopeptide sequences, and the corresponding separation features are obtained. The separation features are calculated using the standardized data set to obtain the probability that the selenopeptide sequences identified in step 3 are natural selenopeptides. The selenopeptide sequences with a probability of less than 60% are removed, and the output layer outputs the identified natural selenopeptide sequences, which are the theoretically predicted sequences of selenopeptides. The input layer, hidden layer, and output layer form the initial model.

[0023] Step 5: Use the validation set to optimize the established initial model. The specific process is as follows: The cross entropy loss function is applied to calculate the error between the theoretically predicted sequence of selenopeptides and the real sequence in the validation set. Then, the weight parameters and bias parameters in the initial model are optimized by the sgd optimizer in an iterative manner to obtain the first optimized model. All the identified selenopeptide sequences in step 3 are input into the first optimized model to obtain the first optimized selenopeptide sequence. The error between it and the real sequence is calculated again. If the error changes compared with the previous value, it is optimized again, and then the new error is calculated using the same method. When the error no longer changes, the iteration is stopped. The model at this time is the sequencing model. Step 6, the secondary maps of the selenopeptides in the test set were converted into selenopeptide sequences using the methods in steps 2-3 and then input into the sequencing model obtained in step 5, and the model prediction sequences were compared with the actual sequences in the test set (including the number of sequences that were consistent, and the number and proportion of prediction errors), and the model performance was evaluated using accuracy, sensitivity and positive predictive value (PPV) as evaluation indicators. The results showed that the accuracy was greater than 93%, the sensitivity was greater than 84%, and the PPV was greater than 89%, so the established sequencing model can be used for the accurate prediction of natural selenopeptides.

[0024] Obtain the secondary spectrum of the sample to be tested in a public database, convert it into a selenopeptide sequence using the method of steps 2 to 3, and input it into the model optimized in step 5 to obtain the identified natural selenopeptide sequence.

[0025] Example 1 Step 1: Search the public database to obtain the secondary spectra of all selenopeptides in samples of different foods (selenium-enriched soybeans, selenium-enriched rice, selenium-enriched oysters, selenium-enriched tea, and selenium-enriched milk), and divide them into training set (1836), validation set (102), and test set (102) at a ratio of 90%, 5%, and 5%; Step 2: The secondary spectra of the selenopeptides in the training set were converted into a matrix containing molecular formula, mass-to-charge ratio and abundance, and then input into the residual dense attention network to capture several common fragments. The fragments with mass spectrum peak clusters consistent with the selenium isotope peak cluster were screened, namely M, M+1.0007 Da, M+1.99809 Da, M+3.99730 Da, M+5.99748 Da, and the characteristic fragments of selenopeptides were finally obtained: C3H7Se + m / z 122.97075, C4H 10 NSe + m / z 151.99730, C5H7OSe + m / z 162.96566, C5H8NOSe + m / z 177.97656, C7H 11 N2O2Se + m / z 234.99803.

[0026] The number of selenium in each selenopeptide is determined based on the abundance of the isotope peak cluster of the characteristic fragment. M / I M+1 / I M+2 / I M+4 / I M+6 = 18.887:15.380:47.914:100:17.597, then the number of selenium is 1; if IM / I M+1 / I M+2 / I M+4 / I M+6 = 37.775:30.760:95.827:100:35.195, then the number of selenium is 2; if I M / I M+1 / I M+2 / I M+4 / I M+6 =37.775:30.760:47.914:100:69.569, then the number of selenium is 3; if I M / I M+1 / I M+2 / I M+4 / I M+6 = 56.662:46.140:71.871:100:52.792, then the number of selenium is 4; if I M / I M+1 / I M+2 / I M+4 / I M+6 = 75.549:61.520:95.827:100:41.742, the number of selenium is 5. Taking selenopeptide ES(Se)MP as an example, according to the secondary spectrum m / z Mass spectrum peak cluster I at 510.12234 M / I M+1 / I M+2 / I M+4 / I M+6 = 18.887:15.380:47.914:100:17.597, and it is determined that the number of selenium it contains is 1.

[0027] Step 3: Determine the corresponding selenoamino acid type based on the characteristic fragment of each selenopeptide. If it contains the fragment ion C2H3O2 + m / z 59.0127, C3HO2 + m / z 68.99711, C3H6NO + m / z 72.04439, CH3OSe + m / z 110.93436, it is selenocysteine; if it contains fragment ion C4H6NO + m / z 84.04439, C4H5O2 + m / z 85.02841, C4H8O + m / z86.06004, C4H8NO2 + m / z 102.05495, C3H7Se + m / z 122.97075, which is selenomethionine. This selenoamino acid is used as the starting amino acid for selenopeptide identification, combined with fragment ions a, a 2+ , a-H2O, a-NH3, b, b 2+ , b-H2O, b-NH3, y, y 2 + , y-H2O and y-NH3 information, and use Ion-CNN to predict the next possible symbol in the N / C-terminal direction of the selenopeptide sequence in turn. The symbol is one of 23 symbols, including 20 standard amino acids, selenocysteine, selenomethionine and terminator.

[0028] According to the secondary spectrum of ES(Se)MP, m / z 84.04408, m / z 102.05469 fragment ion, confirming that it contains selenomethionine. Based on y2 ion m / z 293.03989 information, using Ion-CNN to predict the amino acid sequence symbol of O1, using the LogSoftmax function to calculate the probability of the sequence symbol score at each position, and inferring that the amino acid sequence symbol of O1 is P. Based on y3 ion m / z 380.07192 information, it is inferred that the amino acid sequence symbol of N1 is S. Based on b3 ion m / z 396.06683 information, the inferred amino acid sequence symbol of N2 is E, and the highest score is obtained when it is combined with ES(Se)MP. The final identified selenopeptide sequence is ES(Se)MP; Table 1 Sequence symbol score probability table for each position of selenopeptide ES(Se)MP

[0029] Step 4: Using the known selenopeptide sequences in the training set as input, the fully convolutional attention mechanism neural network was used to learn the N- and C-terminal amino acid features of selenopeptides (Table 2). The obtained features were then input into the fully connected layer for feature fusion. The Softmax function was applied to each feature, and all features were standardized to 0 to 1 to indicate the importance of each input, which was recorded as the standardized feature dataset.

[0030] The selenopeptide sequence identified in step 3 is used as the input layer, and the ReLu function is used as the activation function of the hidden layer to separate the data features in the identified selenopeptide sequence, and the separated data features are obtained. The separated data features are calculated using the standardized feature data set to obtain the probability that the selenopeptide sequence identified in step 3 is a natural selenopeptide, and the selenopeptide sequences with a probability of less than 60% are removed. The output layer outputs the identified natural selenopeptide sequence, which is the theoretical prediction sequence of selenopeptides. The input layer, hidden layer, and output layer form the initial model.

[0031] Table 2 Common amino acids at the N and C termini of selenopeptides

[0032] Step 5: Use the validation set to optimize the established model. The specific process is as follows: The cross entropy loss function is applied to calculate the error between the theoretically predicted sequence of selenopeptides and the true sequence. The weight parameters and bias parameters in the initial model are optimized by the sgd optimizer. When the error no longer changes, the iteration is stopped and the model at this time is the sequencing model. In step 6, the secondary spectra of the selenopeptides in the test set were converted into selenopeptide sequences using the methods of steps 2 and 3, and then input into the sequencing model obtained in step 5. The model performance was evaluated using accuracy, sensitivity, and positive predictive value as evaluation indicators (Table 3). The results showed that the accuracy was greater than 93%, the sensitivity was greater than 84%, and the PPV was greater than 89%, indicating that the established sequencing model can be used to accurately predict natural selenopeptides in five types of selenium-rich food matrices: selenium-enriched soybeans, selenium-enriched rice, selenium-enriched tea, selenium-enriched milk, and selenium-enriched oysters.

[0033] Table 3 Model performance evaluation results

[0034] Using public databases, the secondary spectra of all selenopeptides in selenium-enriched soybeans, selenium-enriched rice, selenium-enriched oysters, selenium-enriched tea and selenium-enriched milk were obtained, converted into selenopeptide sequences using the methods of steps 2 to 3, and input into the optimized model in step 5 to obtain the identified natural selenopeptide sequences.

[0035] The natural selenopeptide sequences identified in selenium-enriched soybeans are: TYG(Se)ME, EIL(Se)CKQ, EPF(Se)M, ES(Se)MP, ERM(Se)CE, E(Se)CKQP; The natural selenopeptide sequences identified in selenium-enriched rice are: LNGKLTG(Se)M, VIDLIRH(Se)M, LGAPDVGHP(Se)M, TVGGAPAGRIV(Se)ME, VQRD(Se)MKLVPY, T(Se)MMM, (Se)MDPGQQ; The natural selenopeptide sequences identified in selenium-enriched oysters are: E(Se)MS(Se)MMF, LLV(Se)MY, LVGAT(Se)MF, LTV(Se)MDF, LVVK(Se)CF, VVVS(Se)MY, LAEAGV(Se)MY, VVV(Se)CPAEE, EG(Se)MFF, LLT(Se)MY; The natural selenopeptide sequences identified in selenium-enriched tea are: LGVPIV(Se)MHDY, LP(Se)MF, VKRL(Se)MMY, and EKFKR(Se)M.

[0036] The natural selenopeptide sequences identified in selenium-enriched milk are: TPVEVL(Se)CKGF, TDDI(Se)M.

[0037] The above implementation modes are only used to illustrate the present invention and are not used to limit the present invention. Any modifications, equivalent substitutions and improvements made on the basis of the technical essence of the present invention within the spirit and principles of the present invention shall be included in the scope of the present invention. The patent protection scope of the present invention is defined by the claims.

Claims

1. A method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network, characterized in that: The steps include: S1, taking the specific selenoamino acid corresponding to each characteristic fragment in all selenopeptides as the starting amino acid, and then predicting the next possible symbol in the N / C terminal direction of the selenopeptide sequence in turn, taking the symbol with the highest probability as the symbol of the corresponding position, and obtaining the corresponding identified selenopeptide sequence; S2, using a fully convolutional attention mechanism neural network to learn the N- and C-terminal amino acid features of selenopeptides in known selenopeptide sequences, then fusion the obtained features, and then normalize all features to 0 to 1 to obtain a standardized data set; All identified selenopeptide sequences are used as the input layer, and the activation function of the hidden layer of the neural network is used to separate the data features to obtain the corresponding separation features, and each separation feature is calculated using the standardized data set to obtain the probability that the corresponding selenopeptide sequence is a natural selenopeptide, and the unmatched selenopeptide sequences are removed, and the output layer outputs the identified natural selenopeptide sequence; The input layer, hidden layer and output layer form an initial model, optimize the weights and biases in the initial model to obtain an optimized model, input all identified selenopeptide sequences into the optimized model to obtain an optimized selenopeptide sequence, and obtain a sequencing model when the error between the optimized selenopeptide sequence and the true sequence no longer changes; S3, all the selenopeptides of the sample to be tested are subjected to the process of S1 to obtain the corresponding identified selenopeptide sequences, which are then input into the sequencing model to obtain the natural selenopeptide sequences identified by the sample to be tested.

2. The method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network according to claim 1, characterized in that: The characteristic fragment described in S1 is obtained as follows: The secondary spectra of all selenopeptides were converted into matrices containing molecular formula, mass-to-charge ratio and abundance, and then several common fragments were captured. If there was a mass spectrum peak cluster consistent with the selenium isotope peak cluster, the corresponding common fragment was the characteristic fragment of the selenopeptide.

3. The method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network according to claim 2, characterized in that: S1 inputs the matrix containing molecular formula, mass-to-charge ratio and abundance into the residual dense attention network for capture, and obtains several common fragments.

4. The method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network according to claim 1, characterized in that: S1 binds to fragment ions a, a 2+ , a-H2O, a-NH3, b, b 2+ , b-H2O, b-NH3, y, y 2+ , y-H2O and y-NH3 information, and use the deep learning-based target detection algorithm Ion-CNN to predict the next possible symbol in the N / C-terminal direction of the selenopeptide sequence in turn.

5. The method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network according to claim 4, characterized in that: S1 uses the LogSoftmax function to calculate the symbol score probability of each position in all peptide sequences, and selects the symbol with the highest probability as the symbol of the corresponding position.

6. The method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network according to claim 1, characterized in that: S2 inputs the obtained features into the fully connected layer of the fully convolutional attention mechanism neural network for fusion.

7. The method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network according to claim 6, characterized in that: S2 applies the Softmax function to each feature, normalizing all features to a range of 0 to 1 and obtaining a standardized data set.

8. The method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network according to claim 1, characterized in that: S2 uses the ReLu function as the activation function of the hidden layer of the neural network to separate the data features in all identified selenopeptide sequences.

9. The method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network according to claim 1, characterized in that: The unmatched selenopeptide sequences described in S2 are selenopeptide sequences with a probability of less than 60%.

10. The method for de novo sequencing of selenopeptides based on a fully convolutional attention mechanism neural network according to claim 1, characterized in that: S2 uses the cross entropy loss function to calculate the error between the optimized selenopeptide sequence and the true sequence.

Images