Plant non-histone lysine crotonylation site identification method and system

Abstract

The invention relates to a plant non-histone lysine crotonylation site identification method and a plant non-histone lysine crotonylation site identification system. The method comprises the following steps: extracting non-histone lysine crotonylation Kcr sites from a protein sequence to form a data set; extracting a first feature vector by using a self-attention mechanism and a smoothing function through a protein language model ESM2; capturing a long-distance dependency relationship and context information in the protein sequence through a bidirectional long-short-term memory network, and extracting a second feature vector; and carrying out feature fusion processing and then carrying out classification prediction to obtain a Kcr site identification result. By constructing a protein sequence data set, ESM2 utilizes a self-attention mechanism and a smoothing function, and BiLSTM captures a long-distance dependency relationship and context information in the protein sequence, efficient feature extraction of the protein sequence is realized, and more comprehensive information can be provided through fusion processing, so that the detection precision of Kcr site detection and recognition is improved.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method and system for identifying plant non-histone lysine crotonylation sites. Background Art

[0002] In the field of bioinformatics, post-translational modification has the ability to change certain specific amino acid residues through a variety of modification methods. Lysine crotonylation (Kcr) is an important post-translational modification that is widely present in histones and non-histones. This site has been widely detected in animals and plants. In plants, global identification and functional analysis of this site have been carried out in species such as papaya, rice, wheat, tobacco powder, and peanuts. It was found that the modification of this site is involved in photosynthesis, ribosomes, and antioxidant systems, and can actively regulate the cold tolerance of wheat; in addition, it is also involved in functions such as gene transcription regulation. Therefore, accurate prediction of this site is very important for understanding the regulation of proteins in human biology.

[0003] Currently, the commonly used experimental methods for detecting Kcr sites include high-performance liquid chromatography separation, stable isotope labeling of amino acids in cell culture, immunoaffinity enrichment, and high-resolution liquid chromatography-tandem mass spectrometry. With the rapid development of artificial intelligence technology, a large number of machine learning-based methods have been proposed, including predictors based on traditional machine learning and deep learning. However, most existing predictors are only used for predictions of a certain species, and there is no model specifically for predicting Kcr sites in plants.

[0004] However, methods based on biological experiments are usually time-consuming, labor-intensive and costly, and are not suitable for large-scale data analysis; methods based on machine learning rely heavily on traditional feature encoding methods, require complex feature engineering steps, are not comprehensive in feature extraction, have poor prediction accuracy, and lack analytical understanding of model interpretability. Therefore, traditional Kcr site detection and recognition methods have the problem of low recognition accuracy. Summary of the invention

[0005] Based on this, in order to solve the above technical problems, a method and system for identifying plant non-histone lysine crotonylation sites are provided, which can improve the detection accuracy of Kcr site detection and identification.

[0006] A method for identifying plant non-histone lysine crotonylation sites, the method comprising:

[0007] Retrieving protein sequences corresponding to various species in a database, extracting non-histone lysine crotonylation Kcr sites in the protein sequences to form a data set;

[0008] Inputting the data set into a protein language model ESM2, and extracting a first feature vector in the protein sequence by using a self-attention mechanism and a smoothing function through the protein language model ESM2;

[0009] Inputting the data set into a bidirectional long short-term memory network BiLSTM, capturing the long-distance dependency and context information in the protein sequence through the bidirectional long short-term memory network, and extracting the second feature vector in the protein sequence;

[0010] Performing feature fusion processing on the first feature vector and the second feature vector to obtain feature representation, performing classification prediction based on the feature representation, and outputting prediction probabilities of each category;

[0011] The Kcr site identification results are obtained based on the predicted probabilities of each category.

[0012] In one embodiment, non-histone lysine crotonylation Kcr sites are extracted from the protein sequence to form a data set, including:

[0013] Extracting peptides of target length from each of the protein sequences, respectively, deleting redundant data of each of the peptides of target length, and obtaining non-histone lysine crotonylation Kcr sites as sample data;

[0014] The sample data is divided into a training data set and a test data set.

[0015] In one embodiment, the protein language model ESM2 uses a self-attention mechanism and a smoothing function to extract a first feature vector in the protein sequence, including:

[0016] The protein sequence is divided into various amino acid fragments by using the protein language model ESM2, and each of the amino acid fragments is encoded and converted into a vector representation to generate an input vector;

[0017] The input vector is transmitted to a deep encoder in the protein language model ESM2, and a first feature vector is determined from the protein sequence by the deep encoder based on a self-attention mechanism;

[0018] Among them, the protein language model ESM2 uses the smoothing function to assist the self-attention mechanism in feature extraction.

[0019] In one embodiment, the method further comprises:

[0020] extracting an attention score from the protein sequence through the self-attention mechanism;

[0021] Using STREME to identify and analyze the protein sequence to obtain motif analysis results;

[0022] The attention score is compared with the motif analysis result, and the protein language model ESM2 is adjusted according to the comparison result.

[0023] In one embodiment, the method further comprises:

[0024] The amino acid substitution matrix BLOSUM62 is used to encode the protein sequence, and the substitution frequency of the amino acids in the protein sequence is calculated to convert the amino acid sequence into a numerical vector.

[0025] In one embodiment, capturing the long-distance dependency and context information in the protein sequence by the bidirectional long short-term memory network and extracting the second feature vector in the protein sequence includes:

[0026] Inputting the numerical vector into the bidirectional long short-term memory network;

[0027] The bidirectional long short-term memory network learns context information of the protein sequence based on the numerical vector;

[0028] capturing dependencies in the protein sequence based on the context information;

[0029] A second feature vector in the protein sequence is identified according to the context information and the dependency relationship.

[0030] In one embodiment, performing feature fusion processing on the first feature vector and the second feature vector to obtain feature representation includes:

[0031] The first feature vector is output through a first channel, and the second feature vector is output through a second channel;

[0032] The first feature vector and the second feature vector output by the two channels are concatenated to complete feature fusion and obtain feature representation.

[0033] In one embodiment, classification prediction is performed based on the feature representation, and the prediction probability of each category is output, including:

[0034] Inputting the feature representation into a fully connected layer, performing a linear transformation on the feature representation according to weights and bias parameters through the fully connected layer, and obtaining a linearly transformed feature representation;

[0035] The activation function is used to introduce nonlinear features into the feature representation after the linear transformation, and then the softmax function is used for normalization processing to convert the prediction values ​​of each category into probability values.

[0036] In one embodiment, the method further comprises:

[0037] Obtaining each probability value of each category corresponding to each of the feature representations;

[0038] The highest probability value is found from each of the probability values, and the highest probability value is output as the predicted probability.

[0039] A plant non-histone lysine crotonylation site recognition system, the system comprising:

[0040] A data set construction module is used to retrieve protein sequences corresponding to various species in the database, extract non-histone lysine crotonylation Kcr sites in the protein sequences, and construct a data set;

[0041] An ESM2 feature extraction module is used to input the data set into a protein language model ESM2, and extract the first feature vector in the protein sequence by using a self-attention mechanism and a smoothing function through the protein language model ESM2;

[0042] A BiLSTM feature extraction module, used for inputting the data set into a bidirectional long short-term memory network BiLSTM, capturing the long-distance dependency and context information in the protein sequence through the bidirectional long short-term memory network, and extracting a second feature vector in the protein sequence;

[0043] A classification prediction module, used for performing feature fusion processing on the first feature vector and the second feature vector to obtain a feature representation, performing classification prediction based on the feature representation, and outputting a prediction probability of each category;

[0044] The site identification module is used to obtain the Kcr site identification result based on the predicted probability of each category.

[0045] The above-mentioned plant non-histone lysine crotonylation site identification method and system can facilitate the subsequent full mining of protein sequence information by constructing a protein sequence data set; through ESM2, the self-attention mechanism and smoothing function are used to realize efficient feature extraction of protein sequences to capture rich structural and functional information in protein sequences; the features obtained by capturing long-distance dependencies and contextual information in protein sequences through BiLSTM not only contain local information of amino acids, but also integrate the global context of the sequence, providing strong support for subsequent classification and prediction tasks; fusing the extracted features can provide more comprehensive information, thereby improving the detection accuracy of Kcr site detection and recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A diagram showing the application environment of a method for identifying plant non-histone lysine crotonylation sites in one embodiment;

[0047] Figure 2 A schematic diagram of a process for identifying plant non-histone lysine crotonylation sites in one embodiment;

[0048] Figure 3 Schematic diagram of the structure of the DeepNhKcr model in one embodiment;

[0049] Figure 4 A schematic diagram showing a comparison of the performance of two models on different species in one embodiment;

[0050] Figure 5 Schematic diagram of performance index comparison of DeepNhKcr model in one embodiment;

[0051] Figure 6 Attention score heat map and motif analysis graph;

[0052] Figure 7 Schematic diagram of the DeepNhKcr model's predictive ability for unknown data;

[0053] Figure 8 A structural diagram of a plant non-histone lysine crotonylation site recognition system in one embodiment;

[0054] Fig. 9 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0056] It is understood that the terms "first", "second", etc. used in this application may be used herein to describe feature vectors, but these feature vectors are not limited by these terms. These terms are only used to distinguish a first feature vector from another feature vector. For example, without departing from the scope of this application, a first feature vector may be referred to as a second feature vector, and similarly, a second feature vector may be referred to as a first feature vector. Both the first feature vector and the second feature vector are feature vectors, but they are not the same feature vector.

[0057] The plant non-histone lysine crotonylation site identification method provided in the embodiments of the present application can be applied to Figure 1 In the application environment shown. Figure 1As shown, the application environment includes a computer device 110. The computer device 110 can retrieve protein sequences corresponding to various species in the database, extract non-histone lysine crotonylation Kcr sites in the protein sequence, and form a data set; the computer device 110 can input the data set into the protein language model ESM2, and extract the first feature vector in the protein sequence by using the self-attention mechanism and smoothing function of the protein language model ESM2; the computer device 110 can input the data set into the bidirectional long short-term memory network BiLSTM, and capture the long-distance dependency and context information in the protein sequence through the bidirectional long short-term memory network, and extract the second feature vector in the protein sequence; the computer device 110 can perform feature fusion processing on the first feature vector and the second feature vector to obtain feature representation, perform classification prediction based on the feature representation, and output the prediction probability of each category; the computer device 110 can obtain the Kcr site recognition result based on the prediction probability of each category. Among them, the computer device 110 can be, but not limited to, various personal computers, laptops, smart phones, robots, unmanned aerial vehicles, tablet computers and other devices.

[0058] In one embodiment, Figure 2 As shown, a method for identifying plant non-histone lysine crotonylation sites is provided, comprising the following steps:

[0059] Step 202, searching the database for protein sequences corresponding to each species, extracting non-histone lysine crotonylation Kcr sites from the protein sequences, and forming a data set.

[0060] The computer equipment can retrieve the corresponding protein sequence of each species from the UniPort and NVBI databases, which can include papaya, rice, wheat, tobacco, peanut and other species.

[0061] In one embodiment, a method for identifying plant non-histone lysine crotonylation sites is provided, which may also include a process for constructing a data set, the specific process including: extracting peptides of target length from each protein sequence, deleting redundant data of each peptide of target length, and obtaining non-histone lysine crotonylation Kcr site as sample data; dividing the sample data to obtain a training data set and a test data set.

[0062] Among them, the computer equipment can extract peptides of length 29 in the protein sequence, in which the K (lysine) residue is located in the center, and when the 14 residues located upstream and downstream are insufficient, the missing residues are replaced by X. Redundant samples were deleted through the CD-HIT package (sequence identity threshold is 70%), and non-histone Kcr sites of five plant species were obtained; finally, all samples were divided into a training data set and an independent test set for the model in a ratio of 7:3.

[0063] In this embodiment, the non-histone Kcr sites of the five plant species obtained may include Papaya (2386 samples), Rice (644 samples), Wheat (2710 samples), Tabacum (833 samples) and Peanut (2724 samples), totaling 9297 positive samples and 36313 negative samples. Then, all samples can be divided into training data sets and independent test sets in a ratio of 7:3. In order to evaluate the performance of the model on each species, it is necessary to allocate according to the species. Specifically, the distribution ratio of positive and negative is retained for each species. Finally, 6508 positive samples and 25418 negative samples are used as training data sets, totaling 31926 samples; 2789 positive samples and 10895 negative samples are used as independent test sets, totaling 13684 samples.

[0064] Step 204: input the data set into the protein language model ESM2, and use the self-attention mechanism and smoothing function of the protein language model ESM2 to extract the first feature vector in the protein sequence.

[0065] Among them, ESM2 (Evolutionary Scale Modeling 2) is an advanced protein language model that aims to achieve efficient feature extraction of protein sequences through deep learning technology. The protein language model adopts an evolutionary transformation neural network architecture and makes full use of the self-attention mechanism and smoothing function to capture the rich structural and functional information in protein sequences.

[0066] In one embodiment, a provided method for identifying plant non-histone lysine crotonylation sites may also include a process of extracting features by ESM2, the specific process including: dividing the protein sequence into individual amino acid fragments through the protein language model ESM2, and encoding each amino acid fragment and converting it into a vector representation to generate an input vector; transmitting the input vector to a deep encoder in the protein language model ESM2, and determining a first feature vector from the protein sequence through the deep encoder based on a self-attention mechanism; wherein the protein language model ESM2 uses a smoothing function to assist the self-attention mechanism for feature extraction.

[0067] In one embodiment, when performing ESM2 feature extraction, it can also include: extracting an attention score from a protein sequence through a self-attention mechanism; using STREME to identify and analyze the protein sequence to obtain a motif analysis result; comparing the attention score with the motif analysis result, and adjusting the protein language model ESM2 according to the comparison result.

[0068] In the initial stage of feature extraction, the protein sequence is first processed by tokenization, which mainly divides the long sequence into several amino acid fragments (tokens). Each amino acid fragment will be encoded and converted into a vector representation. This process not only retains the basic information of the amino acids, but also provides structured data input for subsequent modeling. After encoding and embedding, a set of model input vectors are generated, which can effectively represent each amino acid in the sequence and its context. Next, these input vectors are fed into the deep encoder for further processing: the encoder is designed based on the self-attention mechanism, which enables the model to dynamically determine which parts of the information are most critical at different positions in the sequence; this mechanism takes into account long-range dependencies and allows the model to focus on multiple regions of the sequence at the same time, thereby effectively capturing complex patterns and structural features. In addition, the model also uses a smoothing function to improve the stability of the feature extraction process and reduce potential noise interference. Through the processing of the ESM2 module, the model can not only identify the local features of amino acids, but also mine the global patterns in long sequences, laying a solid foundation for subsequent classification tasks. Ultimately, after this series of complex processing, the model is able to generate high-dimensional feature vectors that contain rich information about protein sequences and have extremely high expressive power, providing important support for the identification of lysine crotonylation sites.

[0069] Among them, the self-attention mechanism of ESM2 is described as follows:

[0070] Given an input vector X (represented as query, key, and value matrices), the single-head attention mechanism is calculated as follows:

[0071]

[0072] Where Q = XW Q is the query vector obtained by linear transformation, W Q is the query weight matrix; K = XW K is the key vector obtained by linear transformation, W K is the key weight matrix; V = XW V is the value vector obtained by linear transformation, W V is the value weight matrix; d k is the dimension of the key vector (usually the same as the dimension of the query and value vectors).

[0073] The multi-head attention mechanism divides the input into h heads, each head calculates the attention value independently, and then concatenates the results together:

[0074] MultiHead(Q,K,V)=Concat(head1,head2,…,head h )W O

[0075] The calculation for each head is as follows:

[0076]

[0077] in is the query, key, and value weight matrix corresponding to the i-th head; Concat is the operation of concatenating the outputs of all heads together to form a longer vector; W o is the final weight matrix used to transform the connection result back to the original dimension. The self-attention mechanism enables the model to focus on different parts of information in different subspaces at the same time, enhancing the model's ability to represent and capture complex patterns.

[0078] By introducing the attention mechanism into the protein language model ESM2, attention scores can be extracted from the query sequence. This process enables ESM2 to automatically focus on the most important part of the sequence for the prediction result, thereby identifying key features. STREME is then used to identify and analyze potential motifs or conserved sequences in biological sequences, and statistical methods are used to calculate the significance of motif occurrence. By generating many randomly arranged sequences, the algorithm can calculate the probability of observing a specific motif under a random distribution. Here, the P value needs to be calculated. The smaller the P value, the lower the probability of the motif appearing in the input sequence and the stronger its biological functionality.

[0079] The extracted attention scores are compared with the results of motif analysis to further enhance the understanding of the model's decision-making process. Specifically, the attention scores reflect the degree of attention paid by the model to different regions when processing the sequence, while motif analysis provides an analysis of known biological patterns and sequence features. By combining the two, it is not only possible to confirm whether the regions focused on by the model are consistent with biologically important motifs, but also to reveal the potential biological mechanisms that the model may adopt in its decision-making process. This comparative analysis provides a transparent way for researchers to interpret the behavior of the model and verify the reliability of its predictions, thereby enhancing the trust and explanatory power of the model's application. Ultimately, this method not only improves the interpretability of the model, but also provides important guidance for subsequent biological research and applications.

[0080] Specifically, the attention mechanism is used to extract attention scores from the query sequence, and compared with the motif analysis of the motif to make an interpretable analysis of the model. The attention score can visualize the model's attention to each input part when processing the sequence. In the attention mechanism, the multi-head attention scores are obtained, and the multi-head attention matrix is ​​summed up. Then, this matrix is ​​visualized as a heat map, which can intuitively show the model's attention pattern when processing the sequence.

[0081] STREME is a method for identifying potential motifs or conserved sequences in biological sequences. Its core principle is to analyze the frequency of occurrence of short sequences in a given sequence to infer the possible biological functions of these sequences. After inputting the original sequence, STREME identifies candidate motifs by calculating the frequency of occurrence of each short sequence, and uses statistical methods to evaluate the significance of these motifs, and calculates the P value to determine their biological relevance in the input data. Finally, the identified motifs, their frequency of occurrence and statistical significance in the sequence, and the drawn sequence diagram are obtained.

[0082] For a given short sequence (motif) s, the frequency of occurrence in the input sequence set can be expressed as:

[0083]

[0084] Among them, N(s) is the number of times the short sequence s appears in all input sequences, Ntota l is the total length of all sequences.

[0085] To assess the significance of a motif, STREME calculates a P value, typically using a hypergeometric distribution or a binomial distribution to assess the probability of observing the motif in a random sequence. The P value can be expressed as:

[0086] P(s)=P(X≥k|n,N,M)

[0087] Where X is the number of occurrences of motif s observed in the random sequence, k is the number of occurrences of the motif observed in the input sequence, n is the length of the random sequence, N is the total possible number of occurrences of the motif, and M is the total number of sequences.

[0088] STREME may calculate a significance score S for each identified motif, typically based on a combination of frequency and P value: S(s) = -log(P(s)) × f(s); the score can help rank and select the most biologically significant motifs.

[0089] Through these calculations, STREME can effectively identify sequence patterns that are important in a specific biological context. The extracted attention scores are compared with the results of motif analysis to confirm whether the regions that the model focuses on are consistent with biologically important motifs, ultimately achieving the purpose of interpretability.

[0090] Step 206, input the data set into a bidirectional long short-term memory network BiLSTM, capture the long-distance dependency and context information in the protein sequence through the bidirectional long short-term memory network, and extract the second feature vector in the protein sequence.

[0091] In one embodiment, a method for identifying plant non-histone lysine crotonylation sites is provided, which may also include a process for encoding a protein sequence, and the specific process includes: encoding the protein sequence using the amino acid substitution matrix BLOSUM62, and calculating the substitution frequency of amino acids in the protein sequence, and converting the amino acid sequence into a numerical vector.

[0092] First, the computer device can encode the input protein sequence using BLOSUM62 (BLOcks SUbstitution Matrix 62). BLOSUM62 is a widely used amino acid substitution matrix that assigns a numerical score to each amino acid by analyzing the substitution frequency of amino acids during evolution. This encoding process converts the original amino acid sequence into a numerical vector, allowing the model to process and understand the sequence data. After BLOSUM62 encoding, the generated numerical vector is used as the input feature of BiLSTM.

[0093] In this embodiment, a BiLSTM bidirectional long short-term memory network is used to extract features from protein sequences to capture long-range dependencies and contextual information in protein sequences.

[0094] In one embodiment, a provided method for identifying plant non-histone lysine crotonylation sites may also include a process of extracting features using a BiLSTM bidirectional long short-term memory network, the specific process including: inputting a numerical vector into the bidirectional long short-term memory network; the bidirectional long short-term memory network learns the context information of the protein sequence based on the numerical vector; capturing the dependency relationship in the protein sequence based on the context information; and identifying a second feature vector in the protein sequence based on the context information and the dependency relationship.

[0095] Among them, the uniqueness of BiLSTM lies in its bidirectional structure, which contains hidden layer states in both forward and backward directions. This design enables the model to simultaneously consider the contextual information before and after each position in the sequence, thereby more comprehensively capturing the dependencies in the sequence. In the forward propagation process, BiLSTM starts from the beginning of the sequence, gradually processes the input of each time step backward, and updates its hidden state. This process enables the model to learn the contextual information from the beginning of the sequence to the current position; while in the backward propagation process, BiLSTM starts from the end of the sequence, processes the input forward, and captures the contextual information from the current position to the end of the sequence. By combining information from these two directions, BiLSTM can effectively capture long-distance dependencies and identify potential patterns and structural features in the sequence. In addition, BiLSTM also has the characteristics of a memory unit, which can effectively retain important information when processing long sequences, avoiding the gradient vanishing problem of traditional RNN in long sequence processing, which enables BiLSTM to maintain high accuracy and stability when processing complex biological sequence data. Finally, after BiLSTM processing, the model generates a set of high-dimensional feature vectors, which not only contain the local information of amino acids, but also integrate the global context of the sequence, providing strong support for subsequent classification and prediction tasks.

[0096] BiLSTM runs two LSTMs simultaneously on the time series, one for forward processing and the other for backward processing. A key part of LSTM is the cell state and various gating controls, including input gate, forget gate, and output gate. The relevant calculation method formula can be expressed as:

[0097] i t =σ(W i ·[h t-1 , X t ]+b i )

[0098] f t =σ(W f ·[h t-1 , X t ]+b f )

[0099]

[0100] o t =σ(W o ·[h t-1 , X t ]+b o )

[0101] h t =o t *tanh(C t)

[0102]

[0103] Where t represents the current step, t-1 represents the previous step, h represents the hidden value, W, C and b are the weight, candidate unit value and bias matrix respectively, i is the i-th amino acid, It is an element-by-element operation, and σ represents the sigmoid function, which is used to control the information flow of the gating mechanism.

[0104] BiLSTM computes a bidirectional hidden state by running LSTM layers forward and backward along the time axis. The forward LSTM computes sequentially from the first element of the sequence to the last element, while the backward LSTM does the opposite. The two hidden states are combined to form the final bidirectional hidden state. The output of the bidirectional LSTM captures the contextual information before and after each time step in the sequence, providing a more comprehensive feature representation.

[0105] In this embodiment, Focal loss is a loss function based on binary cross entropy loss. By introducing a dynamic scaling factor, it dynamically reduces the weight of samples that are easy to distinguish during training, and quickly focuses on samples that are difficult to distinguish. Among them, the number of negative samples in the data set is significantly greater than that of positive samples, which may cause the model to be biased towards learning negative samples during training. In order to address this problem, an improved focal loss is used as the loss function, and the calculation formula can be expressed as:

[0106]

[0107] Where y represents the true label (positive sample is 1, negative sample is 0); ′ represents the predicted probability given by the model; is the balance weight factor used to solve the class imbalance problem; (1-y′) γ is the modulation factor that distinguishes easy samples from difficult samples, and γ is a parameter.

[0108] Step 208: perform feature fusion processing on the first feature vector and the second feature vector to obtain feature representation, perform classification prediction based on the feature representation, and output the prediction probability of each category.

[0109] In the feature extraction stage, ESM2 and BiLSTM respectively perform in-depth feature extraction on the protein sequence and generate high-dimensional feature vectors. Next, the classification prediction stage will effectively fuse the obtained features to provide more comprehensive information for subsequent classification tasks. The feature fusion process is to splice the feature vectors from ESM2 and BiLSTM to form a unified feature representation.

[0110] In one embodiment, a provided method for identifying plant non-histone lysine crotonylation sites may also include a process of performing feature fusion processing, the specific process including: outputting a first vector feature through a first channel, and outputting a second feature vector through a second channel; splicing the first feature vector and the second feature vector output by the two channels to complete feature fusion and obtain feature representation.

[0111] In one embodiment, a provided method for identifying plant non-histone lysine crotonylation sites may also include a process of converting predicted values ​​into probability values, the specific process including: inputting the feature representation into a fully connected layer, performing a linear transformation on the feature representation according to weights and bias parameters through the fully connected layer to obtain a linearly transformed feature representation; using an activation function to introduce nonlinear features into the linearly transformed feature representation, and then using a softmax function to perform normalization processing to convert the predicted values ​​of each category into probability values.

[0112] In one embodiment, the computer device may also obtain each probability value of each category corresponding to each feature representation; search for the highest probability value from each probability value, and output the highest probability value as the predicted probability.

[0113] Among them, the feature fusion process aims to combine the advantages of ESM2 and BiLSTM, making full use of the ability of ESM2 in capturing local patterns and structural features, and the advantages of BiLSTM in capturing long-distance dependencies and contextual information. In this way, richer feature information can be obtained, thereby improving the accuracy of classification. The fused feature vector is input into the fully connected layer for further processing. The function of the fully connected layer is to map the input feature vector to the probability distribution of the target category. The fully connected layer linearly transforms the input features through a series of weights and bias parameters, and applies an activation function to introduce nonlinear characteristics. This process enables the model to learn the complex relationship between features and categories. At the output of the fully connected layer, a probability distribution is also generated to represent the predicted probability of each category. In order to obtain the final prediction result, the softmax function is used to normalize the output and convert the predicted values ​​of all categories into probability values. Finally, the category with the highest probability is selected as the prediction result. In addition, in order to improve the robustness and generalization ability of the model, regularization technology is also introduced in the classification prediction stage to prevent overfitting. It can effectively enhance the performance of the model on unseen data and ensure its reliability in practical applications. Therefore, by fusing the features extracted by the ESM2 and BiLSTM modules and inputting them into the fully connected layer for prediction, efficient and accurate lysine crotonylation site identification can be achieved.

[0114] Step 210, obtaining the Kcr site identification result based on the predicted probability of each category.

[0115] In one embodiment, a method for identifying plant non-histone lysine crotonylation sites is provided, which can be implemented by a plant non-histone lysine crotonylation site recognition model, namely, the DeepNhKcr model, such as Figure 3 As shown, the DeepNhKcr model can include a dataset creation module, an ESM2 module, a BiLSTM module, and a classification prediction module. The dataset module is Figure 3 Part A of the ESM2 module Figure 3 Part B in the figure can realize the function of protein language model ESM2; the BiLSTM module is Figure 3 The C and D parts in the figure can realize the function of the bidirectional long short-term memory network BiLSTM; the classification prediction module is Figure 3 Part E of the .

[0116] The processed original protein sequence is used as the input sequence of the ESM2 encoder module to extract the high-dimensional latent feature representation. In addition, the original protein sequence is encoded with BLOSUM62 as the input of the BiLSTM encoder module to extract the output features of BiLSTM. At the same time, the output features of the two channels are fused to generate the final evolutionary output features. The site recognition model uses a fully connected layer to predict whether the input sequence is a Kcr site, and uses an improved Focal loss function to balance the imbalance of positive and negative samples.

[0117] In this embodiment, the site recognition model can also be evaluated by evaluating the performance of the site recognition model through five-fold cross validation and independent testing, and quantitative analysis is performed using indicators such as accuracy (ACC), F1-score, Matthews correlation coefficient (MCC), sensitivity (SN) and specificity (SP). Ablation experiments are performed to verify the contribution of each module to the performance of the site recognition model and ensure the effectiveness and robustness of the site recognition model.

[0118] Specifically, in this embodiment, five commonly used evaluation indicators are used: sensitivity (SN), specificity (SP), accuracy (ACC), Matthews correlation coefficient (MCC), and F1 score. The formulas of these indicators are described as follows:

[0119]

[0120] Among them, TP, FN, TN, and FP represent the number of true positive, false negative, true negative, and false positive samples respectively, SN represents the proportion of true Kcr, and SP represents the proportion of non-kcr samples correctly identified by the model.

[0121] The f1 score comprehensively evaluates the effectiveness of the model in detecting true Kcr samples by balancing the number of true positives, false positives, and false negatives; MCC and ACC are both used to measure the overall predictive performance of the model; these indicators are used to quantitatively evaluate the overall effectiveness of the prediction model; for each of the aforementioned indicators, the higher the value, the better the performance of the model.

[0122] In one embodiment, a method for identifying plant non-histone lysine crotonylation sites may be implemented using the DeepNhKcr model to conduct experiments. Specifically, in order to evaluate the performance of DeepNhKcr, a five-fold cross validation and independent test experiment were performed, as shown in the following table:

[0123]

[0124] Among them, the experimental results are given in the first two rows of the above table; in order to demonstrate the superior performance of the DeepNhKcr model, it is compared with the existing models, and the results are given in the third row of the above table. It can be seen that the DeepNhKcr model surpasses the previous models in ACC (accuracy), F1-score (F1 score), MCC (Matthews correlation coefficient) and SP (specificity) evaluation indicators; by comparing the currently popular pre-training models: ESM2, BERT and ESM1b, the results are shown in the second, fourth and fifth rows of the above table. The ESM2 model is used in the second row, which can effectively prove the prediction ability of ESM2 in this model; in addition, eight commonly used protein sequence encoding methods are compared as input features, including BE, AAC, AAI, ZSC, CKSAAGP, GDPC, GTPC and BLOSUM62 encoding methods, and the results are shown in the following table:

[0125]

[0126]

[0127] It can be seen that the BLOSUM62 encoding method achieved the best prediction ability; in summary, these results demonstrate the effectiveness of the proposed DeepNhKcr model in identifying Kcr sites.

[0128] Next, in order to further verify the reliability of the model, an ablation experiment was conducted on the DeepNhKcr model. Specifically, the ESM2 module was removed from the model, the BiLSTM module was removed, the BLOSUM62 encoding module was replaced with binary encoding, and the focal loss module was replaced with cross entropy loss. The impact of each component in the model on its prediction performance was evaluated in independent tests. The results are shown in the following table. Each module helps to improve the prediction performance of our model.

[0129]

[0130] Considering the effects of species specificity, it is crucial to evaluate to what extent the predictive power of the DeepNhKcr model can be generalized to different types of plants. Therefore, the performance of the DeepNhKcr model was evaluated by training on individual species in papaya, rice, wheat, tobacco, and peanut, and the results were compared with previous models. Figure 4 As shown in the figure, the DeepNhKcr model outperforms the previous models on different species, and the prediction capabilities on different species are slightly different. Compared with its general model (mixed data training), the general model shows obvious advantages and demonstrates good generalization ability.

[0131] Next, the DeepNhKcr model can be analyzed for interpretability: First, to analyze the conservation of the Kcr site, a probability-based motif analysis was performed on the data, such as Figure 5 As shown in A, the Kcr site is conserved, and it is very necessary to develop a model to identify the Kcr site. Secondly, the discriminative ability of the model is visualized, and the input layer and linear layer data are reduced in dimension and visualized. Figure 5 In B, we can see that the positive samples are clustered to one side, highlighting the effective classification ability of the model. In order to further study the robustness of the model, a mutation experiment was conducted. Specifically, random mutations were performed on 1-3 sites of the Kcr sample. Figure 5 As shown in C, as the mutation rate increases, the model's discriminative ability gradually decreases; however, even when there are mutations at three sites, the model still maintains a certain degree of predictive ability, indicating that the model has strong generalization and robustness. In addition, by comparing the sequence learned by the model with the motif patterns obtained by traditional tools, it is shown that the DeepNhKcr model has learned the key information in the sequence. The visualization of the attention score heat map and the motif analysis diagram is shown in Figure 2. Figure 6 shown.

[0132] Next, in order to evaluate the prediction ability of the DeepNhKcr model for unknown data, a cross-species validation was performed: training on one species and performing prediction evaluation on another species. Figure 7 As shown in A, in the heat map based on ACC, MCC, SN and SP indicators, it can be seen that rice varieties are slightly different from other species; for further explanation, motif analysis was performed on the five species as shown in Figure 7 B in the figure can well explain this finding; overall, it is relatively stable, indicating that the DeepNhKcr model can also show strong predictive ability even when the species has not been seen, highlighting its practicality.

[0133] It should be understood that, although the various steps in the above-mentioned flow chart are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the above-mentioned flow chart may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0134] In one embodiment, Figure 8 As shown, a plant non-histone lysine crotonylation site identification system is provided, comprising: a data set construction module 810, an ESM2 feature extraction module 820, a BiLSTM feature extraction module 830, a classification prediction module 840 and a site identification module 850, wherein:

[0135] The data set construction module 810 is used to search the protein sequences corresponding to each species in the database, extract the non-histone lysine crotonylation Kcr sites in the protein sequences, and construct a data set;

[0136] ESM2 feature extraction module 820, used to input the data set into the protein language model ESM2, and extract the first feature vector in the protein sequence by using the self-attention mechanism and smoothing function of the protein language model ESM2;

[0137] The BiLSTM feature extraction module 830 is used to input the data set into the bidirectional long short-term memory network BiLSTM, capture the long-distance dependency and context information in the protein sequence through the bidirectional long short-term memory network, and extract the second feature vector in the protein sequence;

[0138] The classification prediction module 840 is used to perform feature fusion processing on the first feature vector and the second feature vector to obtain feature representation, perform classification prediction based on the feature representation, and output the prediction probability of each category;

[0139] The site identification module 850 is used to obtain the Kcr site identification result based on the prediction probability of each category.

[0140] In one embodiment, the data set construction module 810 is also used to extract peptides of target length from each protein sequence, delete redundant data of each target length peptide, and obtain non-histone lysine crotonylation Kcr site as sample data; divide the sample data to obtain a training data set and a test data set.

[0141] In one embodiment, the ESM2 feature extraction module 820 is also used to divide the protein sequence into individual amino acid fragments through the protein language model ESM2, and convert each amino acid fragment into a vector representation after encoding to generate an input vector; the input vector is transmitted to the deep encoder in the protein language model ESM2, and the first feature vector is determined from the protein sequence based on the self-attention mechanism through the deep encoder; wherein the protein language model ESM2 uses a smoothing function to assist the self-attention mechanism for feature extraction.

[0142] In one embodiment, the ESM2 feature extraction module 820 is also used to extract an attention score from a protein sequence through a self-attention mechanism; use STREME to identify and analyze the protein sequence to obtain a motif analysis result; compare the attention score with the motif analysis result, and adjust the protein language model ESM2 according to the comparison result.

[0143] In one embodiment, the BiLSTM feature extraction module 830 is also used to encode the protein sequence using the amino acid substitution matrix BLOSUM62, and calculate the substitution frequency of amino acids in the protein sequence, and convert the amino acid sequence into a numerical vector.

[0144] In one embodiment, the BiLSTM feature extraction module 830 is also used to input the numerical vector into a bidirectional long short-term memory network; the bidirectional long short-term memory network learns the context information of the protein sequence based on the numerical vector; captures the dependency relationship in the protein sequence based on the context information; and identifies the second feature vector in the protein sequence based on the context information and the dependency relationship.

[0145] In one embodiment, the classification prediction module 840 is also used to output the first feature vector through the first channel and the second feature vector through the second channel; the first feature vector and the second feature vector output from the two channels are concatenated to complete feature fusion and obtain feature representation.

[0146] In one embodiment, the classification prediction module 840 is also used to input the feature representation into the fully connected layer, perform linear transformation on the feature representation according to the weight and bias parameters through the fully connected layer, and obtain the feature representation after the linear transformation; use the activation function to introduce nonlinear features into the feature representation after the linear transformation, and then use the softmax function to perform normalization processing to convert the prediction value of each category into a probability value.

[0147] In one embodiment, the classification prediction module 840 is further used to obtain each probability value of each category corresponding to each feature representation; find the highest probability value from each probability value, and output the highest probability value as the predicted probability.

[0148] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Fig. 9 As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a method for identifying plant non-histone lysine crotonylation sites is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a key, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0149] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0150] In one embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the steps of a method for identifying plant non-histone lysine crotonylation sites are implemented.

[0151] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of a method for identifying plant non-histone lysine crotonylation sites are implemented.

[0152] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0153] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0154] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A method for identifying plant non-histone lysine crotonylation sites, characterized in that: The method comprises: Retrieving protein sequences corresponding to various species in a database, extracting non-histone lysine crotonylation Kcr sites in the protein sequences to form a data set; Inputting the data set into a protein language model ESM2, and extracting a first feature vector in the protein sequence by using a self-attention mechanism and a smoothing function through the protein language model ESM2; Inputting the data set into a bidirectional long short-term memory network BiLSTM, capturing the long-distance dependency and context information in the protein sequence through the bidirectional long short-term memory network, and extracting the second feature vector in the protein sequence; Performing feature fusion processing on the first feature vector and the second feature vector to obtain feature representation, performing classification prediction based on the feature representation, and outputting prediction probabilities of each category; The Kcr site identification results are obtained based on the predicted probabilities of each category.

2. The method for identifying plant non-histone lysine crotonylation sites according to claim 1, characterized in that: Non-histone lysine crotonylation Kcr sites are extracted from the protein sequence to form a data set, including: Extracting peptides of target length from each of the protein sequences, respectively, deleting redundant data of each of the peptides of target length, and obtaining non-histone lysine crotonylation Kcr sites as sample data; The sample data is divided into a training data set and a test data set.

3. The method for identifying plant non-histone lysine crotonylation sites according to claim 1, characterized in that: The protein language model ESM2 uses a self-attention mechanism and a smoothing function to extract a first feature vector in the protein sequence, including: The protein sequence is divided into various amino acid fragments by using the protein language model ESM2, and each of the amino acid fragments is encoded and converted into a vector representation to generate an input vector; The input vector is transmitted to a deep encoder in the protein language model ESM2, and a first feature vector is determined from the protein sequence by the deep encoder based on a self-attention mechanism; Among them, the protein language model ESM2 uses the smoothing function to assist the self-attention mechanism in feature extraction.

4. The method for identifying plant non-histone lysine crotonylation sites according to claim 3, characterized in that: The method further comprises: extracting an attention score from the protein sequence through the self-attention mechanism; Using STREME to identify and analyze the protein sequence to obtain motif analysis results; The attention score is compared with the motif analysis result, and the protein language model ESM2 is adjusted according to the comparison result.

5. The method for identifying plant non-histone lysine crotonylation sites according to claim 1, characterized in that: The method further comprises: The amino acid substitution matrix BLOSUM62 is used to encode the protein sequence, and the substitution frequency of the amino acids in the protein sequence is calculated to convert the amino acid sequence into a numerical vector.

6. The method for identifying plant non-histone lysine crotonylation sites according to claim 5, characterized in that: The method captures the long-distance dependency and context information in the protein sequence by using the bidirectional long short-term memory network, and extracts the second feature vector in the protein sequence, including: Inputting the numerical vector into the bidirectional long short-term memory network; The bidirectional long short-term memory network learns context information of the protein sequence based on the numerical vector; capturing dependencies in the protein sequence based on the context information; A second feature vector in the protein sequence is identified according to the context information and the dependency relationship.

7. The method for identifying plant non-histone lysine crotonylation sites according to claim 1, characterized in that: Performing feature fusion processing on the first feature vector and the second feature vector to obtain feature representation includes: The first feature vector is output through a first channel, and the second feature vector is output through a second channel; The first feature vector and the second feature vector output by the two channels are concatenated to complete feature fusion and obtain feature representation.

8. The method for identifying plant non-histone lysine crotonylation sites according to claim 1, characterized in that: Based on the feature representation, classification prediction is performed and the prediction probability of each category is output, including: Inputting the feature representation into a fully connected layer, performing a linear transformation on the feature representation according to weights and bias parameters through the fully connected layer, and obtaining a linearly transformed feature representation; The activation function is used to introduce nonlinear features into the feature representation after the linear transformation, and then the softmax function is used for normalization processing to convert the prediction values ​​of each category into probability values.

9. The method for identifying plant non-histone lysine crotonylation sites according to claim 8, characterized in that: The method further comprises: Obtaining each probability value of each category corresponding to each of the feature representations; The highest probability value is found from each of the probability values, and the highest probability value is output as the predicted probability.

10. A plant non-histone lysine crotonylation site recognition system, characterized in that: The system comprises: A data set construction module is used to retrieve protein sequences corresponding to various species in the database, extract non-histone lysine crotonylation Kcr sites in the protein sequences, and construct a data set; An ESM2 feature extraction module is used to input the data set into a protein language model ESM2, and extract the first feature vector in the protein sequence by using a self-attention mechanism and a smoothing function through the protein language model ESM2; A BiLSTM feature extraction module, used for inputting the data set into a bidirectional long short-term memory network BiLSTM, capturing the long-distance dependency and context information in the protein sequence through the bidirectional long short-term memory network, and extracting a second feature vector in the protein sequence; A classification prediction module, used for performing feature fusion processing on the first feature vector and the second feature vector to obtain a feature representation, performing classification prediction based on the feature representation, and outputting a prediction probability of each category; The site identification module is used to obtain the Kcr site identification result based on the predicted probability of each category.

Images