Protein-polypeptide binding site detection method and related equipment

Abstract

The invention provides a protein-polypeptide binding site detection method and related equipment. The method comprises the following steps: acquiring to-be-detected data, inputting the to-be-detected data into a trained binding site detection model, and outputting a two-tuple set as binding site detection data; the binding site detection data is used for indicating a binding site region and a type of a protein-polypeptide sequence corresponding to the to-be-detected data, the binding site region is a binding region taking continuous binding sites as a whole, and the to-be-detected data is a protein sequence; wherein the binding site detection model is composed of an encoder based on ProteinBERT and a non-autoregression decoder based on Transform, the encoder is used for obtaining context perception information representation of each amino acid in a protein sequence, and the decoder is used for directly generating a two-tuple set. According to the application, the regionalized binding site information is introduced by adopting a detection thought to realize prediction of the protein-polypeptide binding site.

Description

Technical Field

[0001] The present application relates to the technical field of bioinformatics, and in particular to a method for detecting protein-peptide binding sites and related equipment. Background Art

[0002] Proteins are the core executors of life activities. They play a vital role in regulating key biological processes such as cell metabolism, signal transduction, immune defense, and cell cycle by interacting with molecules such as polypeptides, DNA, RNA, and metal ions. Among them, the binding of proteins to polypeptides plays an important role in physiological processes such as immune response, transcriptional regulation, cell migration, and repair. In the immune system, many proteins identify and eliminate foreign pathogens by binding to polypeptides to protect the body from infection; in gene expression regulation, polypeptides, as part of transcription factors, regulate the transcription level of specific genes, thereby affecting cell growth, differentiation, and stress response. In addition, the interaction between proteins and polypeptides also has important functions in cell signal transduction, molecular recognition, and cell cycle control. In-depth research on the interaction between proteins and polypeptides and their mechanisms is of great significance for understanding protein functions and developing new therapeutic targets and drugs.

[0003] In the study of protein-peptide interactions, traditional methods usually rely on a series of complex experimental techniques, such as X-ray crystallography, nuclear magnetic resonance (NMR), cryo-electron microscopy (cryo-EM) and molecular docking, which simulate and locate potential binding sites by analyzing the three-dimensional structure of proteins. However, these methods have significant limitations, including high costs, time-consuming experimental processes, structural analysis challenges caused by small peptide size and peptide flexibility. In addition, these methods usually rely on known three-dimensional protein structures and are difficult to apply to proteins that have not yet been resolved or have highly dynamic properties. Therefore, accurately identifying protein-peptide binding sites remains a difficult problem in the fields of structural biology and bioinformatics. Especially in the context of time-consuming and costly experimental methods, the market urgently needs to develop an effective computational prediction method.

[0004] Based on this, the present application provides a method and related equipment for detecting protein-polypeptide binding sites. Summary of the invention

[0005] In view of the problems existing in the prior art, the present application provides a method and related equipment for detecting protein-peptide binding sites, which adopts the concept of detection to introduce regionalized binding site information to realize the prediction of protein-peptide binding sites.

[0006] The purpose of this application is achieved by the following technical solutions:

[0007] In a first aspect, the present application provides a method for detecting a protein-polypeptide binding site, the method comprising:

[0008] Acquire data to be detected, input the data to be detected into a trained binding site detection model, and output a set of two-tuples as binding site detection data; the binding site detection data is used to indicate the binding site region and category of the protein-polypeptide sequence corresponding to the data to be detected, the binding site region is a binding region that regards continuous binding sites as a whole, and the data to be detected is a protein sequence;

[0009] The binding site detection model consists of a ProteinBERT-based encoder and a Transformer-based non-autoregressive decoder, wherein the encoder is used to obtain the context-aware information representation of each amino acid in the protein sequence, and the decoder is used to directly generate a set of bigrams, which includes region information and category information.

[0010] In some possible implementations, the encoder is a pre-trained BERT model, which is used to obtain a protein sequence and encode it through ProteinBERT, learn the context information of amino acids in the sequence through masked language technology, and map each amino acid into a high-dimensional embedding vector, so that the protein sequence is converted into an embedding matrix.

[0011] In some possible implementations, the decoder is a non-autoregressive decoder, which is used to identify bigrams from a given embedding matrix and form a bigram set; the bigram set T = {(r1, c1), ..., (r n ,c n )} satisfies the conditional probability of formula (2):

[0012]

[0013] Among them, T i =(r i start ,r i end ,c i ), r i start Indicates the starting position of the binding region, r i end Indicates the end position of the binding region, c i represents the type of ligand, S represents the protein sequence, and p L (n|S) is used to model the size of the set of two-tuples, and p(T i |S,T j≠i ; μ) represents the binary group T iThere is a correlation between the generation of and the protein sequence S and other pairs in the set.

[0014] In some possible implementations, the non-autoregressive decoder is composed of N layers of the same Transformer architecture, and the non-autoregressive decoder recognizes a binary from a given embedding matrix in the following manner:

[0015] Each layer of Transformer is equipped with a multi-head self-attention mechanism through a multi-head self-attention module, which receives target query information and is used to capture the relationship between the combined regions of the binary; the target query information includes m embedding vectors, where m is a positive integer greater than 35;

[0016] The input embedding matrix and the target query information are integrated through the multi-head cross attention mechanism of the multi-head cross attention module, so that the target query information is converted into m output embedding vectors, denoted as P d ∈R m×d (d=d k ), where each output embedding vector is independently decoded into a combined region and category through a feed-forward neural network, thereby generating m bigrams;

[0017] Among them, unmasked self-attention is used in the multi-head self-attention mechanism, and the information of all positions is considered when processing the sequence to achieve parallel processing of the entire sequence.

[0018] In some possible implementations, the methods for obtaining the start position and the end position of the combined area include:

[0019] Through formula (4) and formula (5), two classifiers are used to predict the starting position and the ending position of the binding area respectively:

[0020]

[0021] Among them, A c ∈R z×d , A i ∈R d×d (i=1,2,3,4) and a i ∈R d (i=1,2) are all learnable parameters, z represents the total number of categories (peptides and none), l is the length of the protein sequence, P e is the output of the encoder.

[0022] In some possible implementations, the BERT model is obtained by pre-training on a BFD database.

[0023] In some possible implementations, the loss function of the binding site detection model is a bipartite graph matching loss function, and the process of bipartite graph matching loss includes:

[0024] Find the optimal matching step, match the predicted bigrams and the real bigrams according to the similarity between them, evaluate the matching degree between each predicted bigram and all the real bigrams, and find the optimal matching pair;

[0025] In the step of calculating the loss function, the matching costs between all true binary pairs and predicted binary pairs are summed to obtain the final loss value.

[0026] In a second aspect, the present application further provides an electronic device, comprising a memory and at least one processor, wherein the memory stores a computer program, and the at least one processor is configured to implement a method as described in any one of the first aspects when executing the computer program.

[0027] In a third aspect, the present application further provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded by a processor of a terminal device and executing any of the methods described in the first aspect.

[0028] In a fourth aspect, the present application further provides a computer program product, wherein the computer program product comprises a computer program, and when the computer program is executed by at least one processor, the steps of any one of the methods described in the first aspect are implemented.

[0029] In combination with the above-mentioned technical solutions and the technical problems solved, the technical solutions to be protected by the present invention and the positive effects thereof are as follows: on the one hand, the continuous amino acid binding sites are regarded as a whole binding region for prediction, and the dependency between the binding region and the protein sequence is effectively captured by accurately identifying the starting position and the ending position of the binding region, so as to achieve integrity and spatial dependence. On the other hand, only the more universal protein sequence is used as the model input to enhance the generalization ability of the model. On this basis, special emphasis is placed on enhancing the mutual influence relationship between the binding region and the binding region, so as to improve the modeling ability of the model for complex biological relationships. On the other hand, the binding region is predicted as a whole, the imbalance ratio of the data set is reduced and more holistic, and a bipartite graph matching loss function is used to directly optimize the optimal match, improve the problem of data imbalance, and avoid the limitations of the relevant strategies. This novel method of regional level detection not only provides new ideas for binding site prediction, but also enables the model to better adapt to complex biological backgrounds, and has significant biological significance and practical application value. On the other hand, traditional experimental techniques are not only complicated to operate, but also require expensive equipment and consumables. This technical solution replaces some experimental operations with computational prediction methods, reducing dependence on expensive experimental techniques. At the same time, by using an encoder based on ProteinBERT and a non-autoregressive decoder based on Transformer, binding sites can be predicted directly from protein sequences without the need for actual physical experiments, greatly improving prediction efficiency and effectively solving the problem of time-consuming and costly experimental methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The present application is further described below in conjunction with the accompanying drawings and implementation methods.

[0031] Figure 1 It is a schematic diagram of the framework of a binding site detection model provided in an embodiment of the present application.

[0032] Figure 2 Schematic diagram of the predicted and experimental binding sites provided in the examples of this application.

[0033] Figure 3 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0034] Below, in conjunction with the accompanying drawings and specific implementation methods, the present application is further described. It should be noted that, under the premise of no conflict, the various embodiments described below or the various technical features can be arbitrarily combined to form a new embodiment. The implementation procedures of the present application will be described below with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. The present application can also be implemented or applied through other different specific implementation procedures, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be understood that the preferred embodiments are only for illustrating the present application, not for limiting the scope of protection of the present application.

[0035] The following briefly describes the relevant terms and application fields of the embodiments of the present application to facilitate understanding by those skilled in the art.

[0036] In the study of protein-peptide interactions, traditional methods usually rely on a series of complex experimental techniques, which are time-consuming and costly. Based on this, with the rapid development of machine learning, deep learning and large models, the prediction of protein-peptide binding sites has achieved remarkable success, but these methods also have certain limitations.

[0037] Specifically, these methods can generally be divided into three categories according to the type of protein information utilized: methods based on protein sequence information, methods based on protein structure information, and methods that simultaneously integrate protein sequence and structure information.

[0038] Regarding methods based on protein structure information, although AlphaFold has made remarkable achievements in the field of protein structure prediction, not all protein sequences can obtain known real three-dimensional structures, and adding predicted structural information may introduce errors or uncertainties.

[0039] The fusion method uses the sequence and structural features of proteins, which can not only capture the potential information in the protein sequence, but also use the structural data to provide a more comprehensive background, thereby improving the accuracy and reliability of the prediction. However, it will also be affected by the predicted structural errors and the alignment of the features between the two modalities. Therefore, in this paper, we only use protein sequences as the research object.

[0040] Regarding the methods based on protein sequence information, the current methods for predicting binding sites based on protein sequences also have certain limitations. First, the existing methods all regard each amino acid in the protein sequence as an independent prediction unit and perform a binary prediction on whether each amino acid is a binding site. This ignores the continuity and spatial interdependence of protein-peptide binding sites. In fact, protein-peptide binding sites may exist in regions of several consecutive amino acids. Such regionalized binding sites can better reflect the relationship between peptide binding and are more consistent with biological mechanisms. Secondly, the existing methods extract sequence context feature information and then independently classify the amino acids point by point. During the training process, they fail to fully capture the relationship between sites and sites and between sites and the entire protein sequence in the binding process. The lack of contextual information in the sequence limits the generalization ability of the model when dealing with complex biological relationships. Finally, the ratio of the number of binding sites to non-binding sites in protein sequences is seriously unbalanced. Existing methods usually use strategies such as undersampling, weighted or contrastive learning to solve the data imbalance problem. However, undersampling may lead to the loss of negative sample information, limiting the learning ability of the model; weighted methods rely too much on weight settings and lack universality and robustness; although contrastive learning helps to learn the relative relationship between samples, inappropriate selection of negative samples may introduce noise, resulting in a decrease in the generalization ability of the model.

[0041] The present application is different from the related art in that it aims to use the idea of ​​detection to introduce regionalized binding site information to achieve prediction of protein-polypeptide binding sites. The following first describes the method embodiment, and then describes the device, etc.

[0042] Method embodiment.

[0043] Figure 1 Schematic diagram of a binding site detection model provided in an embodiment of the present application. The method for detecting a binding site of a protein-peptide provided in this embodiment includes:

[0044] Acquire data to be detected, input the data to be detected into a trained binding site detection model, and output a set of two-tuples as binding site detection data; the binding site detection data is used to indicate the binding site region and category of the protein-polypeptide sequence corresponding to the data to be detected, the binding site region is a binding region that regards continuous binding sites as a whole, and the data to be detected is a protein sequence;

[0045] The binding site detection model consists of a ProteinBERT-based encoder and a Transformer-based non-autoregressive decoder, wherein the encoder is used to obtain the context-aware information representation of each amino acid in the protein sequence, and the decoder is used to directly generate a set of bigrams, which includes region information and category information.

[0046] The technical solution provided in this embodiment first obtains the protein sequence data to be detected, and inputs the protein sequence into a trained binding site detection model. The binding site detection model consists of an encoder based on ProteinBERT and a non-autoregressive decoder based on Transformer. Among them, the role of the encoder is to obtain the context-aware information representation of each amino acid in the protein sequence. ProteinBERT is a pre-trained BERT model specifically used to process protein sequences. It captures the long-range dependencies between amino acids through a self-attention mechanism, and maps each amino acid into a high-dimensional embedding vector, thereby obtaining the context information of each amino acid in the sequence.

[0047] The task of the decoder (based on Transformer) is to directly generate a set of two-tuples, which contain multiple sets of region information and category information. Among them, the region information refers to the starting and ending positions of the protein-peptide binding site, and the category information refers to the type of binding site (such as peptide). The non-autoregressive decoder does not depend on the order of sequence generation and can process the entire sequence in parallel, which improves the prediction efficiency. The set of two-tuples output by the decoder indicates the binding site region and category of the protein-peptide sequence corresponding to the data to be detected. The binding site region treats the continuous binding sites as a whole binding region, which can more accurately capture the dependency between the binding region and the protein sequence.

[0048] Therefore, by predicting binding sites in a regionalized manner, the model can more effectively learn the interdependencies within the binding region, between regions, and between regions and proteins, thereby improving the accuracy of the prediction. Compared with existing methods that usually treat each amino acid as an independent prediction unit and ignore the continuity and spatial interdependence of binding sites, this method overcomes this limitation through regionalized prediction. The model only takes the protein sequence as input, which enhances the generalization ability of the model and enables it to adapt to different protein sequences and complex biological backgrounds. The non-autoregressive decoder can process the entire sequence in parallel instead of building the output step by step, which improves computational efficiency.

[0049] It is worth noting that the binding site detection model mentioned in this application is the PPRD (Protein-Peptide Regions Detection) model, which is stated as the PPRD model in some descriptions.

[0050] In an exemplary embodiment, the encoder is a pre-trained BERT model, which is used to obtain a protein sequence and encode it through ProteinBERT, learn the context information of amino acids in the sequence through masked language technology, and map each amino acid into a high-dimensional embedding vector, so that the protein sequence is converted into an embedding matrix.

[0051] The protein sequence is input into the ProteinBERT encoder. Each amino acid is regarded as a "word" in the sequence and is mapped to a specific numerical representation. ProteinBERT uses the self-attention mechanism to capture the dependencies between amino acids in the sequence, so that the representation of each amino acid contains not only its own information, but also the contextual information of the entire sequence. Each amino acid is mapped to a high-dimensional embedding vector that can represent the complex biological properties and contextual information of the amino acid. After the entire protein sequence is encoded by ProteinBERT, it is converted into an embedding matrix, in which each row corresponds to the embedding vector of an amino acid, thereby providing rich input features for the subsequent decoder.

[0052] Therefore, through the self-attention mechanism, the model can capture the long-range dependencies between amino acids. At the same time, since the embedding vector of each amino acid contains rich contextual information, it helps to improve the accuracy of subsequent binding site prediction.

[0053] In an exemplary embodiment, the BERT model is obtained by pre-training on the BFD database. The BFD database refers to the Big Fantastic Database, a protein sequence database built by Martin Steinegger, which contains a large number of protein families and sequences.

[0054] See also Figure 1 , as an example, a ProteinBERT-based encoder is provided.

[0055] The encoder is designed to obtain context-aware information representation of each amino acid in the protein sequence. Drawing on the successful application of deep Transformer models in natural language processing, the pre-trained BERT model ProtBert-Big Fantastic Database is used. By pre-training on large-scale protein sequence data from BFD, the self-attention mechanism is used to effectively capture the long-range dependencies between amino acids. It is worth noting that protein sequences do not have reasonable semantic logical relationships like natural languages. Therefore, it is different from the original BERT. ProteinBERT converts each amino acid into a high-dimensional embedding vector and only learns the contextual information in the sequence through the masked language technique (MLM), without using the conventional next sequence prediction (NSP) technology.

[0056] The protein sequence is first split into tokens suitable for encoder input. Each amino acid in the sequence is mapped to a specific numeric value as if it were a word in a sentence. For example, alanine A is mapped to 7, tyrosine Y is mapped to 21, and arginine R is mapped to 14. It can be assumed that the length of each protein sequence is different. The shortest is only 31 amino acids, and the longest can reach 1485 amino acids. In order not to add too much redundant information, the protein sequences are not padded to the same length. Therefore, the protein sequence S = (s1, s2, s3, ..., s l ) is input to the encoder, where s i represents the i-th amino acid, and l represents the length of the protein sequence. After ProteinBERT encoding, each amino acid s i is mapped into a high-dimensional embedding vector Therefore, the entire protein sequence S is converted into an embedding matrix, represented as where l is the length of the protein sequence, including the start and end markers [CLS] and [SEP], and d k is the dimension of the embedding vector for each amino acid, Represents the embedding vector of the ith amino acid in the sequence.

[0057] In an exemplary embodiment, the decoder is a non-autoregressive decoder, which is used to identify two-tuples from a given embedding matrix and form a two-tuple set; the two-tuple set T = {(r1, c1), ..., (r n ,c n )} satisfies the conditional probability of formula (2):

[0058]

[0059] in, Indicates the starting position of the binding region, Indicates the end position of the binding region, c i represents the type of ligand, S represents the protein sequence, and p L (n|S) is used to model the size of the set of two-tuples, and p(T i |S,T j≠i ; μ) represents the binary group T i There is a correlation between the generation of and the protein sequence S and other pairs in the set.

[0060] The decoder is a non-autoregressive decoder, which is used to identify bigrams from a given embedding matrix and form a bigram set. The conditional probability representation of the bigram set satisfies formula (2). The non-autoregressive decoder does not depend on the order of sequence generation and can process the entire sequence in parallel, which improves the prediction efficiency.

[0061] Thus, by considering the interrelationships between pairs, the model is able to predict binding regions more accurately. By considering the interdependencies between binding regions, the model is able to identify true binding sites more accurately. The model does not rely on a specific sequence length or structure and can adapt to protein sequences of different lengths and complexities.

[0062] In an exemplary embodiment, the non-autoregressive decoder is composed of N layers of the same Transformer architecture, and the non-autoregressive decoder recognizes a bigram from a given embedding matrix in the following manner:

[0063] Each layer of Transformer is equipped with a multi-head self-attention mechanism through a multi-head self-attention module, which receives target query information and is used to capture the relationship between the combined regions of the binary; the target query information includes m embedding vectors, where m is a positive integer greater than 35;

[0064] The input embedding matrix and the target query information are integrated through the multi-head cross attention mechanism of the multi-head cross attention module, so that the target query information is converted into m output embedding vectors, denoted as P d ∈R m×d (d=d k ), where each output embedding vector is independently decoded into a combined region and category through a feed-forward neural network (FNN), thereby generating m bigrams;

[0065] Among them, unmasked self-attention is used in the multi-head self-attention mechanism, and the information of all positions is considered when processing the sequence to achieve parallel processing of the entire sequence.

[0066] The non-autoregressive decoder consists of N layers of identical Transformer architectures, each of which contains a multi-head self-attention module and a multi-head cross-attention module. Each layer of Transformer receives the target query information through the multi-head self-attention mechanism, which includes m embedding vectors, where m is a positive integer greater than 35. The multi-head self-attention mechanism can capture the mutual relationship between the binding regions of the bigrams and enhance the model's understanding of the internal dependencies of the sequence. It can be considered that the target query information is a set of learnable embedding vectors used to represent the predicted bigram set. The multi-head cross-attention mechanism is used to integrate the input embedding matrix (contextual information of the protein sequence) and the target query information. This integration enables the target query information to absorb the contextual information of the protein sequence, thereby more accurately predicting the binding region and category.

[0067] Each output embedding vector is independently decoded into a binding region and a class by FFN, generating m bigrams. The bigrams indicate the potential binding site region and its class in the protein sequence.

[0068] As a result, multi-head self-attention and multi-head cross-attention mechanisms enhance the model's understanding of complex sequence dependencies and improve the model's expressiveness. Unmasked self-attention allows the model to process the entire sequence in parallel instead of building the output step by step, significantly improving computational efficiency.

[0069] In an exemplary embodiment, the method for obtaining the start position and the end position of the combined area includes:

[0070] Through formula (4) and formula (5), two classifiers are used to predict the starting position and the ending position of the binding area respectively:

[0071]

[0072] Among them, A c ∈R z×d , A i ∈R d×d (i=1,2,3,4) and a i ∈R d (i=1,2) are all learnable parameters, z represents the total number of categories (peptides and none), l is the length of the protein sequence, P e is the output of the encoder.

[0073] In this embodiment, the starting position and the ending position of the combined region are predicted by two independent classifiers, corresponding to formula (4) and formula (5) respectively. Both formulas use the softmax function, which can convert the output into a probability distribution.

[0074] Thus, by independently predicting the start and end positions of the binding region, the model is able to more accurately identify specific regions of protein-peptide interaction.

[0075] In specific applications, see Figure 1 , input protein sequence, the sequence is embedded into the encoder, the encoder includes multi-head attention (module) and feedforward neural network, the decoder includes multi-head attention (module), multi-head cross attention (module) and multi-layer perceptron (MLP, a feedforward artificial neural network), multi-head attention is used to receive target query information, multi-head cross attention is connected to the output end of the encoder, and the multi-layer perceptron is connected to the output end of the multi-head cross attention module. The output end of the decoder is connected to two independent classifiers (not shown in the figure), each classifier independently evaluates the possibility of each position in the sequence becoming the starting or ending position of the binding region. There are two feedforward neural networks (FNN) and each corresponds to a classifier, so that the output embedding vector after the classifier is independently decoded into the binding region and category through the feedforward neural network. The final output result is the regional information and category of the binding site and the blank area (i.e., no area).

[0076] See also Figure 1 As an example, a Transformer-based non-autoregressive decoder is provided.

[0077] In this example, the prediction problem after binding site regionalization is modeled as a set prediction problem, and a set of two-tuples is directly generated through a transformer-based non-autoregressive decoder, which contains region information and category information. In the related art, the two-tuple is converted into a sequence, and an autoregressive decoder is used to generate each two-tuple one by one in sequence. For this method, the conditional probability of the target two-tuple set can be expressed by formula (1):

[0078]

[0079] In contrast, the non-autoregressive decoder used in this example does not need to learn the generation order of bigrams, compared with the sequence-to-sequence autoregressive method, avoiding the problem of sequence dependency. In addition, it can simultaneously utilize bidirectional context information, rather than being limited to a unidirectional generation process from left to right, thereby extracting bigram sets more efficiently.

[0080] Therefore, the non-autoregressive decoder can directly model the target bigrams based on formula (2). Specifically, the core goal of the task is to identify all possible bigrams from a given protein sequence. Formally, given an input sequence S, the target bigram set T = {(r1,c1),…,(r n ,c n )} can be expressed as:

[0081]

[0082] Among them, T i =(r i start ,r i end ,c i ), r i start Indicates the starting position of the binding region, r i end Indicates the end position, c i Indicates the type of ligand (for example, the ligand in protein-ligand binding can include ATP, ADP, DNA, RNA, peptides, metal ions, etc.). In order to make the model adaptable to multi-category ligand recognition, we designed a binary representation, but in this scenario, we only consider one type of ligand-peptide. L (n|S) is used to model the size of the set of two-tuples, and p(T i |S,T j≠i ; μ) represents the binary group T i The generation of depends not only on the protein sequence S, but also on other pairs in the set, i.e. other binding regions T j≠i There is a correlation between them.

[0083] Before the decoding process begins, m learnable embedding vectors (called target query information) are used for initialization, which can be shared, that is, all input sequences are initialized with the same binary target query. Therefore, the decoder must first know the size of the target binary set, that is, the number of binding regions corresponding to each protein sequence in the dataset. In other words, in this example, we first need to calculate p in formula (2) L (n|X) and simplifies the modeling process by requiring the non-autoregressive decoder to generate a fixed number of m target predictions for each input protein sequence, where the value of m is set to a value significantly larger than the maximum number of bigrams in a protein sequence.

[0084] The non-autoregressive decoder consists of N layers of identical Transformer architectures. Each layer of Transformer is equipped with a multi-head self-attention mechanism to capture the correlation between the regions that bind to the bigram and integrate the information of the input sequence through a multi-head cross-attention mechanism. Unlike the autoregressive decoder, the non-autoregressive decoder is not restricted by the output autoregressive factorization, which means that during the decoding process, early steps do not need to block access to information from later steps. Therefore, instead of using random masks in the multi-head self-attention mechanism, unmasked self-attention is used, allowing the model to fully consider information at all positions when processing the sequence. Because it can process the entire sequence in parallel instead of building the output step by step, this design makes the non-autoregressive decoder more efficient in generating bigrams.

[0085] Finally, the m target queries are converted into m output embedding vectors, denoted as P d ∈R m×d (d=d k ). Then, P d Each output embedding vector of is independently decoded into a combined region and category by a feed-forward neural network, thereby generating m final predicted bigrams. Specifically, given P d The output embedding p in d ∈R d , the prediction of the ligand binding category (here there is only one category, namely peptide) is achieved by the following formula (3):

[0086] P c =softmax(A c p d ) (3)

[0087] At the same time, the regional information of the protein binding site is decoded by using two classifiers to predict the starting position and the ending position respectively:

[0088]

[0089] Among them, A c ∈R z×d , A i ∈R d×d (i=1,2,3,4) and a i ∈R d (i=1,2) are all learnable parameters, z represents the total number of categories (peptides and none), l is the length of the protein sequence, P e is the output of proteinBERT encoder.

[0090] In an exemplary embodiment, the loss function of the binding site detection model is a bipartite graph matching loss function, and the process of bipartite graph matching loss includes:

[0091] Find the optimal matching step, match the predicted bigrams and the real bigrams according to the similarity between them, evaluate the matching degree between each predicted bigram and all the real bigrams, and find the optimal matching pair;

[0092] In the step of calculating the loss function, the matching costs between all true binary pairs and predicted binary pairs are summed to obtain the final loss value.

[0093] Among them, the bipartite graph matching loss function is a method used to optimize model predictions. It calculates the loss by optimally matching the predicted results with the true labels, thereby guiding model training. In the step of finding the best match, the similarity or match between each predicted bigram and all true bigrams is first evaluated, which can be achieved by calculating the similarity measure between the two. Based on the similarity score, an algorithm (such as a greedy algorithm or a Hungarian algorithm) is used to find the best matching pair. The best match means finding a matching scheme among all possible matches so that the overall matching cost between the predicted bigram and the true bigram is minimized.

[0094] In the step of calculating the loss function, the loss function is calculated after the optimal matching pair is found, and the matching costs between all true bigrams and predicted bigrams are summed. The matching cost can be the difference between the predicted bigram and the true bigram in terms of position, category, etc. By summing the costs, the final loss value is obtained, which can be used for gradient descent optimization in model training.

[0095] In specific applications, the technical solution provided in this embodiment does not use the loss function used by traditional target detection methods, but selects a bipartite graph matching loss based on the characteristics of combining regions to form a set of binary group predictions. This loss function avoids excessive reliance on candidate boxes. Through optimal matching, the model can better align the prediction and the real target during the training process.

[0096] The process of bipartite graph matching loss is mainly divided into two steps: finding the best match and calculating the loss function. In the step of finding the best match, we first need to match the predicted bigrams and the real bigrams based on their similarity. Usually, we use cosine similarity, Euclidean distance or other correlation measurement methods to evaluate the matching degree between each predicted bigram and all real bigrams. Based on this, we can find the best matching pair through greedy algorithm or Hungarian algorithm.

[0097] In application, use Denote a set of m predicted pairs, T = {T1, T2, ..., T n} represents a real set of two-tuples, where T i represents the i-th true bigram Here m must be greater than n. Therefore, in order to match the real bigrams with the predicted bigrams, the real bigrams are replaced by the empty set. Fill in to get T = {T1, T2, ..., T m}. At the same time, a permutation optimization method is used to calculate the matching cost of each pair of binary pairs and minimize the total matching cost to obtain the optimal match. The matching cost is optimized to find the optimal one. For the specific goal, see formula (6):

[0098]

[0099] Among them, θ∈∏(m) represents the permutation space of all lengths m, that is, all possible matching combination regions; Represents the real pair T i and predicted bigrams The matching cost between .

[0100] Matching Cost It is necessary to consider combining category matching and area span matching:

[0101]

[0102] in, Indicates the matching of categories. Indicates the matching of the region span, Indicates that category matching is considered only when the true category is not empty.

[0103] The matching cost of a region span needs to take into account the matching costs of the region's start and end positions:

[0104]

[0105] In this step, the optimization algorithm is solved by the Hungarian algorithm, which efficiently finds the minimum cost match by calculating the matching cost between each predicted bigram and the real bigram. Compared with brute force exhaustive permutation, the Hungarian algorithm has a significant advantage in computational complexity. It optimizes through a one-to-one matching method without considering the order, so it can determine an optimal bipartite matching solution by calculating the matching degree of each pair of bigrams.

[0106] The second step is to calculate the loss function. Specifically, it is to sum the matching costs between all true binary pairs and predicted binary pairs to get the final loss value. In this application, the loss function defined is similar to the loss function in common target detection methods:

[0107]

[0108] The present application provides a technical solution that is different from the existing method for binary classification of amino acids one by one. The positive effects of the technical solution to be protected by the present application at least include:

[0109] On the one hand, the continuous amino acid binding sites are regarded as a whole binding region for prediction. By accurately identifying the starting and ending positions of the binding region, the dependency between the binding region and the protein sequence is effectively captured, achieving integrity and spatial dependence.

[0110] On the other hand, only more universal protein sequences are used as model input to enhance the generalization ability of the model. On this basis, special emphasis is placed on enhancing the mutual influence relationship between binding regions, thereby improving the model's ability to model complex biological relationships.

[0111] On the other hand, the binding region is predicted as a whole, the imbalance ratio of the data set is reduced and more holistic, and a bipartite graph matching loss function is used to directly optimize the optimal match, which improves the problem of data imbalance and avoids the limitations of related strategies. This novel region-level detection method not only provides new ideas for binding site prediction, but also enables the model to better adapt to complex biological backgrounds, which has significant biological significance and practical application value.

[0112] On the other hand, traditional experimental techniques are not only complicated to operate, but also require expensive equipment and consumables. This technical solution replaces some experimental operations with computational prediction methods, reducing the reliance on expensive experimental techniques; at the same time, using the ProteinBERT-based encoder and the Transformer-based non-autoregressive decoder, binding sites can be predicted directly from protein sequences without the need for actual physical experiments, greatly improving prediction efficiency and effectively solving the problem of time-consuming and costly experimental methods.

[0113] The proposed PPRD model is compared with the existing methods on the same dataset. The proposed PPRD model is trained on GPU and uses python3.8.10 and Pytorch framework. The initial learning rate of ProteinBERT used in the experiment is set to 0.00001, and the initial learning rate of the non-autoregressive decoder is set to 0.00002. During the training process, the learning rate is decayed at a rate of 0.01 to avoid overfitting. Ten layers of bidirectional Transformer are stacked in the non-autoregressive decoder and the number of generated bigrams is set to 40. The AdamW optimizer is selected, which can effectively avoid the gradient vanishing problem and optimize the training process through weight decay. At the same time, the gradient clipping technique is used to prevent gradient explosion. In addition, the strategy of training and testing is adopted, that is, a test evaluation is performed immediately after each round of training to monitor the performance of the model in real time and prevent overfitting. A total of 100 rounds of training are conducted.

[0114] In order to fairly evaluate and compare the performance of the method protected by this application with existing methods, a standard data set consistent with previous studies was selected. The data sets were all from the BioLip database and were widely used in the training and testing of protein-peptide binding site prediction. In order to ensure the wide applicability of the data, the same comparison process was also followed, and the BLAST tool was used to remove redundant protein sequences with sequence similarity higher than 30%. When defining a binding site, if the distance between any heavy atom of a residue in the protein sequence and any heavy atom in the polypeptide is less than Finally, a total of 1279 polypeptide-bound protein sequences were obtained, including 16749 binding residues and 290943 non-binding residues, which corresponded to 9010 binding regions in the method of the present application, and the rest were background non-binding regions, which were randomly divided into training sets and independent test sets, accounting for 90% and 10% of the total data, respectively.

[0115] In order to comprehensively evaluate the performance of the model proposed in this application and compare it with existing methods, the same evaluation indicators as the existing methods were selected for evaluation. However, since we used the first 100 position information after prediction sorting to combine into a binding region in the process of regionalized prediction of binding sites, the predicted classification probability corresponding to each residue could not be obtained after regionalization, so it was impossible to directly compare with the existing methods to draw PR curves and calculate AUC. For this reason, other evaluation indicators commonly used in classification tasks were used, including recall, specificity, precision and Matthews correlation coefficient (MCC) to comprehensively measure the performance of the model in the task of identifying protein-peptide interaction sites. Recall and precision measure the performance of the model in identifying positive and negative examples, and specificity and MCC can measure the impact of false negatives and false positives. The results of existing methods are directly taken from their work, including: Pepsite, Peptimap, SPRINT-Seq, SPRINT-Str, PepBind, Visual, PepNN-Seq, PepBCL, PepCNN.

[0116] According to the results in Table 1 below, it can be seen that the method provided by the present application has reached a level comparable to that of existing methods based on sequence information in terms of evaluation indicators. In particular, it is superior to the current methods in terms of both specificity and precision. This shows that the model proposed in the present application can accurately identify and distinguish binding sites from non-binding sites, thereby improving the accuracy of prediction.

[0117] Table 1 Comparison with previous method results

[0118]

[0119] Note: The methods marked with * indicate that the protein structure information is added.

[0120] In order to further verify the effectiveness of the method of this application, two representative protein sequences will be selected for in-depth analysis. The binding sites of these two protein sequences are predicted using the method of this application. In order to intuitively demonstrate the accuracy of the prediction results, the three-dimensional structures of these proteins will be visualized using PyMol software, as shown in the figure below. Figure 2 As shown, Figure 2 This is a schematic diagram of the predicted and experimental binding sites provided in the examples of this application, which intuitively compares the predicted binding sites with the binding sites actually obtained in the experiment. Through this visual comparison, not only can the performance of the model on different protein sequences be evaluated, but also its applicability and robustness in complex biological data can be verified.

[0121] See also Figure 2, PyMol software was used to visualize the 3D structures of the two proteins (Figures A and B are protein 1uj0, and Figures C and D are protein 2aka). Gray represents the non-binding area, blue represents the binding area obtained by the experiment, and red represents the binding area predicted by the method of the present application.

[0122] Obviously, this application proposes a new method for predicting protein-peptide binding sites, which adopts an encoder-decoder architecture to predict binding sites in a regionalized manner, rather than the existing method of binary classification of single amino acids. In this way, the model can more effectively learn the interdependencies within the binding region, between regions, and between regions and proteins, thereby overcoming the existing methods that ignore the conformational changes during the binding process of polypeptide proteins. The experimental results show the effectiveness of this method. Not only are the evaluation results comparable to those of existing methods, but it also provides new computational ideas and tools for further protein ligand identification and drug design.

[0123] Device Embodiments.

[0124] This embodiment provides an electronic device, including a processor, a memory, and a computer program stored in the memory and capable of running on the processor, and when the computer program is executed by the processor, the steps of the method described in any one of the method embodiments are implemented.

[0125] See also Figure 3 , Figure 3 It is a structural schematic diagram of an electronic device provided in an embodiment of the present application.

[0126] The electronic device 10 may include, for example, at least one memory 11 , at least one processor 12 , and a bus 13 connecting different platform systems.

[0127] The memory 11 may include a (computer) readable medium in the form of a volatile memory, such as a random access memory (RAM) 111 and / or a cache memory 112 , and may further include a read-only memory (ROM) 113 .

[0128] The memory 11 also stores a computer program, which can be executed by the processor 12 so that the processor 12 implements the steps of any of the above methods.

[0129] The memory 11 may also include a utility 114 having at least one program module 115, such program module 115 including but not limited to: an operating system, one or more application programs, other program modules and program data, each of which or some combination may include the implementation of a network environment.

[0130] Accordingly, the processor 12 may execute the above-mentioned computer program, and may execute the utility 114 .

[0131] The processor 12 may adopt one or more application specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field programmable gate arrays (FPGAs), or other electronic components.

[0132] The bus 13 may be a local bus representing one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor, or any of a variety of bus architectures.

[0133] The electronic device 10 may also communicate with one or more external devices such as a keyboard, a pointing device, a Bluetooth device, etc., and may also communicate with one or more devices that can interact with the electronic device 10, and / or communicate with any device (such as a router, a modem, etc.) that enables the electronic device 10 to communicate with one or more other computing devices. Such communication may be performed through an input / output interface 14. In addition, the electronic device 10 may also communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN) and / or a public network, such as the Internet) through a network adapter 15. The network adapter 15 may communicate with other modules of the electronic device 10 through the bus 13. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 10 in actual applications, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0134] Storage Medium Embodiments.

[0135] This embodiment provides a computer-readable storage medium, which stores instructions, and the instructions are suitable for being loaded by a processor of a terminal device and executing any method described in the method embodiment. Its specific embodiment is consistent with the embodiment recorded in the above method embodiment and the technical effects achieved, and some contents are not repeated here.

[0136] Program Product Embodiments.

[0137] The embodiment of the present application also provides a computer program product, the specific embodiments of which are consistent with the embodiments recorded in the above method embodiments and the technical effects achieved, and some contents will not be repeated here.

[0138] The computer program product includes a computer program, and when the computer program is executed by at least one processor, the steps of any of the above methods are implemented. The computer program product can be a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a terminal device, such as a personal computer. However, the computer program product of the present invention is not limited to this, and the computer program product can be any combination of one or more computer-readable media.

[0139] It should be noted that, in the embodiments of the present application, "at least one" refers to one or more, and "plurality" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can represent: a, b, c, a and b, a and c, b and c or a and b and c, where a, b and c can be single or multiple. It is worth noting that "at least one" can also be interpreted as "one or more items".

[0140] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are configured to distinguish similar objects, and do not have to be configured to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0141] This application is explained from the perspectives of purpose of use, effectiveness, progress and novelty, and has met the functional enhancement and usage requirements emphasized by the Patent Law. The above description and drawings of this application are only the preferred embodiments of this application, and are not intended to limit this application. Therefore, all structures, devices, features, etc. that are similar or identical to this application, that is, all equivalent replacements or modifications made in accordance with the scope of the patent application of this application, should fall within the scope of protection of the patent application of this application.

Claims

1. A method for detecting protein-polypeptide binding sites, characterized in that: The method comprises: Acquire data to be detected, input the data to be detected into a trained binding site detection model, and output a set of two-tuples as binding site detection data; the binding site detection data is used to indicate the binding site region and category of the protein-polypeptide sequence corresponding to the data to be detected, the binding site region is a binding region that regards continuous binding sites as a whole, and the data to be detected is a protein sequence; The binding site detection model consists of a ProteinBERT-based encoder and a Transformer-based non-autoregressive decoder, wherein the encoder is used to obtain the context-aware information representation of each amino acid in the protein sequence, and the decoder is used to directly generate a set of bigrams, which includes region information and category information.

2. The binding site detection method according to claim 1, characterized in that: The encoder is a pre-trained BERT model, which is used to obtain a protein sequence and encode it through ProteinBERT, learn the context information of amino acids in the sequence through masked language technology, and map each amino acid into a high-dimensional embedding vector, so that the protein sequence is converted into an embedding matrix.

3. The binding site detection method according to claim 2, characterized in that: The decoder is a non-autoregressive decoder, which is used to identify two tuples from a given embedding matrix and form a two-tuple set; the two-tuple set T = {(r1, c1), ..., (r n ,c n )} satisfies the conditional probability of formula (2): in, Indicates the starting position of the binding region, Indicates the end position of the binding region, c i represents the type of ligand, S represents the protein sequence, and p L (n|S) is used to model the size of the set of two-tuples, and p(T i |S,T j≠i ; μ) represents the binary group T i There is a correlation between the generation of and the protein sequence S and other pairs in the set.

4. The binding site detection method according to claim 3, characterized in that: The non-autoregressive decoder is composed of N layers of the same Transformer architecture. The non-autoregressive decoder recognizes bigrams from a given embedding matrix in the following ways: Each layer of Transformer is equipped with a multi-head self-attention mechanism through a multi-head self-attention module, which receives target query information and is used to capture the relationship between the combined regions of the binary; the target query information includes m embedding vectors, where m is a positive integer greater than 35; The input embedding matrix and the target query information are integrated through the multi-head cross attention mechanism of the multi-head cross attention module, so that the target query information is converted into m output embedding vectors, denoted as P d ∈R m×d (d=d k ), where each output embedding vector is independently decoded into a combined region and category through a feed-forward neural network, thereby generating m bigrams; Among them, unmasked self-attention is used in the multi-head self-attention mechanism, and the information of all positions is considered when processing the sequence to achieve parallel processing of the entire sequence.

5. The binding site detection method according to claim 4, characterized in that: Methods for obtaining the start and end positions of the combined area include: Through formula (4) and formula (5), two classifiers are used to predict the starting position and the ending position of the binding area respectively: Among them, A c ∈R z×d , A i ∈R d×d (i=1,2,3,4) and a i ∈R d (i=1,2) are all learnable parameters, z represents the total number of categories (peptides and none), l is the length of the protein sequence, P e is the output of the encoder.

6. The binding site detection method according to claim 2, characterized in that: The BERT model is obtained by pre-training on the BFD database.

7. The binding site detection method according to claim 1, characterized in that: The loss function of the binding site detection model is a bipartite graph matching loss function. The process of bipartite graph matching loss includes: Find the optimal matching step, match the predicted bigrams and the real bigrams according to the similarity between them, evaluate the matching degree between each predicted bigram and all the real bigrams, and find the optimal matching pair; In the step of calculating the loss function, the matching costs between all true binary pairs and predicted binary pairs are summed to obtain the final loss value.

8. An electronic device, characterized in that: The electronic device comprises a memory and at least one processor, wherein the memory stores a computer program, and the at least one processor is configured to implement the method according to any one of claims 1 to 7 when executing the computer program.

9. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by at least one processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images