Glutarylation site prediction method and device based on dual-generator WGAN-GP

Abstract

The application provides a glutarylation site prediction method and device based on a double-generator WGAN-GP, which comprises the following steps: acquiring glutarylated protein data, taking a lysine residue as a window center, setting a fixed window size to cut a sequence fragment, taking glutarylated samples as positive samples, and taking non-glutarylated samples as unlabeled samples; establishing and training a reliable non-glutarylated site selection model, the reliable non-glutarylated site selection model adopts WGAN-GP, and comprises two generators and one discriminator; selecting reliable negative samples from the unlabeled samples by using the discriminator in the reliable non-glutarylated site selection model; and predicting glutarylated sites based on glutarylated site and reliable non-glutarylated site information to obtain the prediction results of the glutarylated sites and the non-glutarylated sites. The application can improve the purity of the non-glutarylated sites used for model training, and improve the accuracy and reliability of the glutarylated site prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning glutamylation site prediction, in particular to a glutamylation site prediction method and device based on dual generator WGAN-GP. BACKGROUND

[0002] Glutamylation site prediction is of great significance for exploring its biological mechanisms and the treatment and prevention of related diseases. However, traditional biological experimental methods consume a large amount of time and financial resources. Using computer technology to develop new prediction methods can quickly and efficiently identify potential modification sites in protein sequences.

[0003] Most existing glutamylation site prediction methods use classification methods in machine learning to make predictions, which are sensitive to data balance. However, glutamylation site data is extremely imbalanced, meaning that glutamylation sites have been experimentally verified, but non-glutamylation post-translational modification data has not been experimentally verified. The number of positive samples is much smaller than the number of unlabeled samples. The inequality of the number of samples of different categories leads to serious bias towards the majority class samples, thus greatly affecting the accuracy and reliability of the prediction. Existing computational methods mainly use oversampling and undersampling methods. For example, the latest solution is a machine learning method called BiPepGlut to identify glutamylation sites. Its prediction model uses an evolutionary method based on dipeptides for feature extraction and an Extra-Trees (ET) classifier for classification. In the case of data imbalance, BiPepGlut oversamples positive samples (glutamylation).

[0004] However, the undersampling method cannot guarantee the reliability of the selected negative samples, and the oversampling method is prone to introducing noise, affecting the accuracy and reliability of the prediction. SUMMARY

[0005] Therefore, the present application provides a glutamylation site prediction method and device based on dual generator WGAN-GP to improve the purity of non-glutamylation sites in the training set and improve the accuracy and reliability of glutamylation site prediction.

[0006] To this end, the present application provides the following technical solutions:

[0007] The present application provides a glutamylation site prediction method based on dual generator WGAN-GP, comprising:

[0008] Glutamylation protein data is obtained, with lysine residues as the window center and a fixed window size to extract sequence fragments. Glutamylation samples are positive samples, and non-glutamylation samples are unlabeled samples.

[0009] The reliable non-pentanoylation site selection model is established and trained, and the reliable non-pentanoylation site selection model adopts a bulldozer distance principle-based dual-generator generated adversarial network WGAN-GP with gradient penalty, including two generators and one discriminator;

[0010] The discriminator module in the reliable non-pentanoylation site selection model is used to select reliable non-pentanoylation sites from unlabeled samples.

[0011] Based on the pentanoylation site data and the reliable non-pentanoylation site sample set, a pentanoylation site prediction model is constructed to obtain the prediction results of the pentanoylation site and the non-pentanoylation site.

[0012] Further, the reliable non-pentanoylation site selection model is trained, including:

[0013] In the discriminator training stage, the pentanoylation post-translational modification data set, the pentanoylation post-translational modification data set generated by the first generator G1, and the unlabeled sample generated by the second generator G2 are input into the discriminator D for training and learning, wherein the purpose of the pentanoylation post-translational modification data set and the pentanoylation post-translational modification data set generated by G1 is to improve the discrimination ability of D, and the unlabeled sample generated by G2 is used as noise data to improve the generalization ability of D; the existing data labels are used to train and learn the discriminator, so that the discriminator obtains better initial network parameters, and the generator does not participate in the training in this stage;

[0014] In the generator training stage, a large amount of unlabeled data set is used to generate an adversarial training of G1, G2 and D, the unlabeled data is input into the generator, the generator generates new data, the generated new data is input into D, and the distribution difference between the generated new data and the input data is calculated according to the loss function, and the parameters of the generator are optimized;

[0015] Through the above two stages, the trained generator and discriminator are obtained.

[0016] Further, based on the pentanoylation site and the reliable non-pentanoylation site sample, the pentanoylation site prediction is performed, including:

[0017] The pentanoylation site and the reliable non-pentanoylation site sample are deeply extracted;

[0018] The key information extracted through the attention layer is output through the Sigmoid activation function by adopting a three-layer fully connected layer combined with the Dropout layer and the ReLU activation function, the binary cross-entropy loss function is used to measure the difference between the prediction and the actual value, and the prediction results of the pentanoylation site and the non-pentanoylation site are obtained.

[0019] Further, the glutarylated post-translational modification dataset and the reliable non-glutarylated post-translational modification dataset are deeply extracted, including:

[0020] The deep feature extraction layer includes a convolutional neural network layer, a Bi-LSTM layer and an attention layer, the convolutional neural network layer is used to extract hidden information of glutarylated translation, the Bi-LSTM is used to learn sequence information of the protein and extract dependency relationship of amino acids, and the attention layer is used to identify key information of feature information.

[0021] Further, the protein sequence is further converted into a numerical vector through a multi-view feature encoding mode, wherein the multi-view feature encoding includes CKSAAP composed of K-interval amino acid pairs, AAC composed of amino acids, AAF composed of amino acid indexes and BLOSUM62.

[0022] Further, the model variants removing the convolutional neural network, removing the Bi-LSTM or removing the attention layer are developed, and the importance of different modules on the overall result of the model is tested.

[0023] The application further provides a glutarylated site prediction device based on a dual-generator WGAN-GP, including:

[0024] The sample acquisition unit is used for acquiring glutarylated protein data, taking lysine residues as window centers, setting a fixed window size to cut sequence fragments, and taking glutarylated samples as positive samples and non-glutarylated samples as unlabeled samples.

[0025] The negative sample selection model unit is used for establishing and training a non-glutarylated site sample selection model, and the reliable non-glutarylated site sample selection model adopts a dual-generator generative adversarial network WGAN-GP based on a bulldozer distance principle with a gradient penalty term, including two generators and one discriminator.

[0026] The negative sample selection unit is used for selecting reliable non-glutarylated site samples from the unlabeled samples by using the discriminator in the reliable non-glutarylated site sample selection model.

[0027] The prediction unit is used for performing glutarylated site prediction based on the positive samples and the reliable non-glutarylated site samples, and obtaining prediction results of glutarylated sites and non-glutarylated sites.

[0028] The advantages and positive effects of the present application are: in the present application, first, the lysine sequence is encoded by multi-view features, and the sequence information is converted into a digital vector, and then a double generator WGAN-GP algorithm is proposed to select reliable non-pentanoylation site samples, solving the most important challenge in pentanoylation site prediction-data imbalance. Compared with common machine learning methods, the present application constructs a deep feature extraction network, mainly including a convolutional neural network layer, a Bi-LSTM layer and an attention layer, wherein the convolutional neural network layer is used to extract the hidden information of the sequence, the Bi-LSTM extracts the dependency relationship of the feature sequence, and the attention layer is used to identify the key information of the feature information, and finally a fully connected neural network is used for pentanoylation site prediction. The method of the present application is helpful to improve the prediction accuracy of pentanoylation, and has certain value for actual disease diagnosis and treatment and subsequent drug research and development. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0030] Figure 1 The flow chart of the pentanoylation site prediction method based on the double generator WGAN-GP in the embodiments of the present application is shown.

[0031] Figure 2 The schematic diagram of selecting reliable non-pentanoylation post-translational modification samples in the embodiments of the present application is shown. DETAILED DESCRIPTION

[0032] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0033] It should be noted that the terms "first", "second", and the like in the description and in the claims of the present application and above-described accompanying drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0034] As Figure 1 indicated, the glutarylation site prediction method of the dual generator WGAN-GP provided in the embodiments of the application mainly includes three parts of data preparation, model building and model testing.

[0035] 1. Data preparation:

[0036] Step 1: Obtain glutarylated protein data from the database, set a fixed window size to intercept sequence fragments with lysine residues as the window center, and glutarylated samples as positive samples, and non-glutarylated samples as unlabeled samples.

[0037] Step 2: Extract attribute features from protein sequences of positive and negative samples through multi-view feature encoding mode, so as to achieve the purpose of all-around characterization of glutarylated samples and non-glutarylated samples, wherein the multi-view feature encoding includes K-interval amino acid pair composition (CKSAAP), amino acid composition (AAC), amino acid index (AAF), and BLOSUM62.

[0038] 2. Model building:

[0039] Step 1: Select reliable non-glutarylated site samples.

[0040] In glutarylation site prediction, unlabeled samples are large, and it is crucial to select reliable negative samples from unlabeled sample data using existing positive sample data, therefore, based on the generator-adversarial network (WGAN-GP) with gradient penalty based on bulldozer distance principle, a reliable negative sample selection model is proposed, and the dual generator WGAN-GP is composed of two generators G1 and G2 and a discriminator D, as Figure 2 indicated.

[0041] 1) Discriminator training phase. First, the glutaroylated post-translational modification dataset, the glutaroylated post-translational modification dataset in the generated data of the first generator G1 and the unlabeled sample in the generated data of the second generator G2 are input into the discriminator D for training learning, wherein the purpose of the glutaroylated post-translational modification dataset and the glutaroylated post-translational modification dataset in the generated data of the first generator G1 is to improve the ability of the discriminator D to identify positive samples, and the unlabeled sample in the generated data of the second generator G2 is used as noise data to improve the generalization ability of the discriminator D. The existing data labels are used to train the discriminator, so that the discriminator obtains better initial network parameters, and the generator does not participate in training in this stage.

[0042] 2) Generator training phase. A large amount of unlabeled data set is used to perform generative adversarial training on the first generator G1, the second generator G2 and the discriminator D. The unlabeled data is input into the generator, the generator generates new data, and the generated new data is input into the discriminator. According to the loss function, the distribution difference between the generated new data and the input data is calculated, and the parameters of the generator are optimized.

[0043] Through the iterative cycle training of the above two stages, the generator and the discriminator are mutually opposed, and the generalization ability of the discriminator is continuously improved. Finally, the trained generator and discriminator are obtained, and only the discriminator is used to select reliable non-glutaroylated post-translational modification samples from the unlabeled samples.

[0044] Step 2: Glutaroylation site prediction.

[0045] In order to better capture the protein sequence information, the features after reliable negative sample selection are deeply extracted. The deep feature extraction layer mainly includes a convolutional neural network layer, a Bi-LSTM layer and an attention layer. The convolutional neural network layer is used to extract the hidden information of the feature sequence, the Bi-LSTM is used to learn the sequence information of the feature, and the attention layer is used to identify the key information of the feature information.

[0046] For glutaroylation site prediction, three fully connected layers combined with Dropout layers and ReLU activation functions are used. The key information extracted by the attention layer is output through a Sigmoid activation function, and a binary cross-entropy loss function is used to measure the difference between the prediction and the actual value.

[0047] 3、Model testing:

[0048] Step 1: The protein sequence is converted into a numerical vector through multi-view feature coding, and reliable non-pentanoylated post-translational modification samples are selected through the dual generator WGAN-GP to solve the problem of data imbalance, and then deep feature extraction is performed through the convolutional neural network, Bi-LSTM and attention mechanism, and finally the extracted features are put into the full connection layer for prediction. The loss value between the true sample label value and the model prediction value is recorded, and the index value for evaluating the classification ability of the model is recorded.

[0049] Step 2: Test various parameters and important modules that affect the performance of the model. Observe the influence of the convolutional neural network, Bi-LSTM, attention layer and other modules in the model on the test results of the model. Develop model variants that remove the convolutional neural network, remove the Bi-LSTM, remove the attention layer, etc. Test the importance of different modules to the overall result of the model.

[0050] Step 3: Test the ability of the model to identify potential pentanoylation sites. For several common drugs used to treat diseases in real society, the site information of the protein is predicted and observed.

[0051] In the above embodiment, first, the lysine sequence is subjected to multi-view feature coding to convert the sequence information into a digital vector, and then the dual generator WGAN-GP algorithm is used to select reliable non-pentanoylated post-translational modification samples to solve the most important challenge in pentanoylation site prediction - the problem of data imbalance. Compared with common machine learning methods, the present application constructs a deep feature extraction network, mainly including a convolutional neural network layer, a Bi-LSTM layer and an attention layer, wherein the convolutional neural network layer is used to extract hidden information of the sequence, the Bi-LSTM extracts the dependency relationship of the feature sequence, the attention layer is used to identify key information of the feature information, and finally the full connection neural network is used for pentanoylation site prediction. The method of the present application is helpful to improve the accuracy of pentanoylation prediction, and has certain value for actual disease diagnosis and treatment and subsequent drug research and development.

[0052] Corresponding to the pentanoylation site prediction method of the dual generator WGAN-GP in the present application, the present application also provides a pentanoylation site prediction device of the dual generator WGAN-GP, comprising:

[0053] A sample acquisition unit is configured to acquire pentanoylated protein data, set a fixed window size to cut sequence fragments with lysine residues as the window center, and pentanoylated samples as positive samples and non-pentanoylated samples as unlabeled samples.

[0054] The negative sample selection model unit is configured to establish and train a reliable non-pentanoylation site sample selection model, and the reliable non-pentanoylation site sample selection model adopts an improved generative adversarial network WGAN-GP based on the bulldozer distance principle with gradient penalty, and includes two generators and one discriminator.

[0055] The negative sample selection unit is configured to select reliable non-pentanoylation site samples from the unlabeled samples by using the generator in the reliable non-pentanoylation site sample selection model.

[0056] The prediction unit is configured to perform pentanoylation site prediction based on the pentanoylation site samples and the reliable non-pentanoylation site samples, to obtain the prediction results of the pentanoylation sites and the non-pentanoylation sites.

[0057] For the improved WGAN-GP pentanoylation site prediction device of the embodiment of the present application, since it corresponds to the double-generator WGAN-GP pentanoylation site prediction method in the above embodiment, the description is relatively simple, and the relevant similarities can be referred to the description of the double-generator WGAN-GP pentanoylation site prediction method in the above embodiment, which will not be described in detail here.

[0058] In the several embodiments of the present application, it should be understood that the disclosed technology can be implemented in other ways. Of course, the unit embodiment described above is only illustrative, and for example, the division of the units can be a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interfaces, units or modules, and can be electrical or other forms.

[0059] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on multiple units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment scheme.

[0060] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0061] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application, essentially or in other words, the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0062] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting glutarylation sites based on a dual generator WGAN-GP, characterized in that: include: To obtain glutarylated protein data, a fixed window size was set with lysine residues as the window center to intercept sequence fragments. The glutarylated samples were considered positive samples, and the non-glutarylated samples were considered unlabeled samples. Establish and train a reliable non-glutarylation site selection model, which uses a double-generator generative adversarial network based on the bulldozer distance principle with gradient penalty WGAN-GP, consisting of two generators and one discriminator; selecting reliable non-glutarylated sites from the unlabeled sample using a discriminator module in the reliable non-glutarylated site selection model; constructing a glutarylation site prediction model based on the glutarylation site data and the reliable non-glutarylation site sample set to obtain prediction results of glutarylation sites and non-glutarylation sites; Training a reliable non-glutarylation site selection model, including: During the discriminator training phase, the glutarylation PTM dataset, the data generated by the first generator G1, and the data generated by the second generator G2, which are unlabeled samples, are input into the discriminator D for training. The purpose of the glutarylation PTM dataset and the data generated by G1 is to improve the discrimination ability of D, while the data generated by G2, which are unlabeled samples, is used as noise data to improve the generalization ability of D. The discriminator is trained and learned using the existing data labels to obtain better initial network parameters. The generator does not participate in training at this stage. During the generator training phase, a large number of unlabeled datasets are used to perform generative adversarial training on G1, G2, and D. Unlabeled data is fed into the generator, which then generates new data. This new data is then fed into D. The loss function is used to calculate the distribution gap between the generated new data and the input data, and the generator parameters are optimized. Through the above two stages, training is carried out in sequence, and finally the trained generator and discriminator are obtained.

2. The method for predicting glutarylation sites based on the dual generator WGAN-GP according to claim 1, characterized in that: Glutarylation site prediction is performed based on the glutarylation site and the reliable non-glutarylation site 2 point samples, including: performing deep extraction on the glutarylation site and the reliable non-glutarylation site samples; A three-layer fully connected layer combined with a Dropout layer and a ReLU activation function was used. The key information extracted by the attention layer was output through a Sigmoid activation function. The binary cross-entropy loss function was used to measure the gap between the prediction and the actual situation, and the prediction results of glutarylation sites and non-glutarylation sites were obtained.

3. The method for predicting glutarylation sites based on the dual generator WGAN-GP according to claim 2, characterized in that: Deep extraction is performed on the glutarylation post-translational modification dataset and the reliable non-glutarylation post-translational modification dataset, including: The deep feature extraction layer includes a convolutional neural network layer, a Bi-LSTM layer, and an attention layer. The convolutional neural network layer is used to extract hidden information of glutarylation translation, the Bi-LSTM layer is used to learn protein sequence information and extract amino acid dependencies, and the attention layer is used to identify key information of feature information.

4. The method for predicting glutarylation sites based on the dual generator WGAN-GP according to claim 1, characterized in that: Also includes: The protein sequence is converted into a numerical vector through multi-view feature encoding, where the multi-view feature encoding includes K-spaced amino acid pair composition CKSAAP, amino acid composition AAC, amino acid index AAF and BLOSUM62.

5. The method for predicting glutarylation sites based on the dual generator WGAN-GP according to claim 2, characterized in that: Also includes: Develop model variants that remove convolutional neural networks, remove Bi-LSTMs, or remove attention layers, and test the importance of different modules to the overall model results.

6. A glutarylation site prediction device based on a dual generator WGAN-GP, characterized in that: include: The sample acquisition unit is used to obtain glutarylated protein data, with lysine residues as the window center, and a fixed window size is set to intercept sequence fragments. The glutarylated sample is the positive sample, and the non-glutarylated sample is the unlabeled sample; A negative sample selection model unit is used to establish and train a reliable non-glutarylation site sample selection model. The reliable non-glutarylation site sample selection model adopts a dual-generator generative adversarial network (WGAN-GP) based on the bulldozer distance principle with a gradient penalty term, including two generators and one discriminator. a negative sample selection unit, configured to select reliable non-glutarylation site samples from the unlabeled samples using the discriminator in the reliable non-glutarylation site sample selection model; a prediction unit, configured to perform glutarylation site prediction based on the positive sample and the reliable non-glutarylation site sample, and obtain prediction results of glutarylation sites and non-glutarylation sites; Training a reliable non-glutarylation site selection model, including: During the discriminator training phase, the glutarylation PTM dataset, the data generated by the first generator G1, and the data generated by the second generator G2, which are unlabeled samples, are input into the discriminator D for training. The purpose of the glutarylation PTM dataset and the data generated by G1 is to improve the discrimination ability of D, while the data generated by G2, which are unlabeled samples, is used as noise data to improve the generalization ability of D. The discriminator is trained and learned using the existing data labels to obtain better initial network parameters. The generator does not participate in training at this stage. During the generator training phase, a large number of unlabeled datasets are used to perform generative adversarial training on G1, G2, and D. Unlabeled data is fed into the generator, which then generates new data. This new data is then fed into D. The loss function is used to calculate the distribution gap between the generated new data and the input data, and the generator parameters are optimized. Through the above two stages, training is carried out in sequence, and finally the trained generator and discriminator are obtained.

Images