Protein function identification method, system, terminal, and storage medium

Abstract

A protein function identification method, a system, a terminal, and a storage medium. The method comprises: acquiring a bulk protein sequence and a bulk protein function corresponding to the bulk protein sequence, inputting the bulk protein sequence into multiple deep learning pre-trained models, respectively, and outputting multiple multi-dimensional feature vectors; performing feature fusion processing on the multiple multi-dimensional feature vectors to obtain a fused feature vector, and training a bi-directional long short-term memory model on the basis of the fused feature vector and the bulk protein function to obtain a protein function identification model; and acquiring a protein sequence to be identified, inputting said protein sequence into the protein function identification model, and outputting a protein function identification result. Feature vectors of a protein sequence are extracted by using multiple different deep learning models, and a bi-directional long short-term memory model is trained to obtain a protein function identification model, such that rapid and accurate identification of protein functions can be realized.

Description

A protein function identification method, system, terminal and storage medium TECHNICAL FIELD

[0001] The present application relates to the technical field of protein function identification, and particularly relates to a protein function identification method, system, terminal and computer readable storage medium. BACKGROUND

[0002] Proteins are extremely important compounds in living organisms. They are not only key components of biological structures, but also play a variety of important roles in life activities. In life science and medical research, identifying and understanding the functions of proteins is crucial.

[0003] Protein function prediction is a complex task that faces multiple challenges. First, proteins themselves are highly complex, including their amino acid sequences, three-dimensional structures, and multiple functions. The interactions and dependencies between these characteristics increase the difficulty of prediction. Second, many proteins not only have one function, but also assume multiple different biological roles, making protein function prediction a multi-label, multi-classification problem. In addition, the function of a protein can be influenced by its location in the cell, interactions, and expression patterns under different conditions.

[0004] To solve the above problems, the existing technology generally uses deep learning to predict protein functions, but the accuracy of the prediction results is not high.

[0005] Therefore, the prior art still needs to be improved and developed.

[0006] SUMMARY

[0007] The main purpose of the present application is to provide a protein function identification method, system, terminal and computer readable storage medium, which aims to solve the problem of low accuracy of protein function prediction in the prior art by using deep learning.

[0008] To achieve the above purpose, the present application provides a protein function identification method, which comprises the following steps:

[0009] Obtain the ontology protein sequence and the ontology protein function corresponding to the ontology protein sequence, and input the ontology protein sequence into a plurality of deep learning pre-training models respectively, and output a plurality of multi-dimensional feature vectors;

[0010] Perform feature fusion processing on the plurality of multi-dimensional feature vectors to obtain a fusion feature vector, and train a bidirectional long short-term memory model according to the fusion feature vector and the ontology protein function to obtain a protein function identification model;

[0011] obtaining a protein sequence to be identified, and inputting the protein sequence to be identified into the protein function identification model to output a protein function identification result.

[0012] Optionally, the protein function identification method, wherein the ontology protein sequence and the ontology protein function corresponding to the ontology protein sequence are obtained, and the ontology protein sequence is input into a plurality of deep learning pre-training models respectively to output a plurality of multi-dimensional feature vectors, specifically including:

[0013] determining a gene ontology, and determining an ontology label according to the gene ontology;

[0014] performing retrieval and search processing in a preset open source database according to the ontology label to obtain an ontology protein sequence and an ontology protein function corresponding to the ontology protein sequence;

[0015] obtaining a plurality of deep learning pre-training models, and inputting the ontology protein sequence into a plurality of the deep learning pre-training models respectively to output a plurality of multi-dimensional feature vectors.

[0016] Optionally, the protein function identification method, wherein the deep learning pre-training model includes a protein evolution scale model, a protein language model, a protein vector representation model, and a one-hot encoding model; and the multi-dimensional feature vector includes a first multi-dimensional feature vector, a second multi-dimensional feature vector, a third multi-dimensional feature vector, and a fourth multi-dimensional feature vector.

[0017] The obtaining a plurality of deep learning pre-training models, and inputting the ontology protein sequence into a plurality of the deep learning pre-training models respectively to output a plurality of multi-dimensional feature vectors, specifically includes:

[0018] obtaining a protein evolution scale model, a protein language model, a protein vector representation model, and a one-hot encoding model;

[0019] inputting the ontology protein sequence into the protein evolution scale model to output a first multi-dimensional feature vector;

[0020] inputting the ontology protein sequence into the protein language model to output a second multi-dimensional feature vector;

[0021] inputting the ontology protein sequence into the protein vector representation model to output a third multi-dimensional feature vector;

[0022] inputting the ontology protein sequence into the one-hot encoding model to output a fourth multi-dimensional feature vector.

[0023] Optionally, the protein function recognition method, wherein the inputting the ontology protein sequence into the one-hot encoding model outputs a fourth multi-dimensional feature vector, specifically comprising:

[0024] The ontology protein sequence is input into the one-hot encoding model, and the ontology protein sequence is processed by the one-hot encoding model for binary vector conversion to obtain a multi-dimensional matrix.

[0025] According to the ontology protein function, a label matrix is constructed, and a correspondence between the multi-dimensional matrix and the label matrix is obtained.

[0026] According to the correspondence, the feature vectors in the multi-dimensional matrix are valued to obtain the fourth multi-dimensional feature vector.

[0027] Optionally, the protein function recognition method, wherein the plurality of multi-dimensional feature vectors are subjected to feature fusion processing to obtain a fusion feature vector, and a bidirectional long short-term memory model is trained according to the fusion feature vector and the ontology protein function to obtain a protein function recognition model, specifically comprising:

[0028] The first multi-dimensional feature vector, the second multi-dimensional feature vector, the third multi-dimensional feature vector, and the fourth multi-dimensional feature vector are subjected to feature fusion processing to obtain a fusion feature vector.

[0029] The fusion feature vector is input into a bidirectional long short-term memory model to obtain context information of the fusion feature vector, and the context information is converted into a vector matrix.

[0030] According to the vector matrix, a Key vector and a Value vector are constructed, and a Query vector is constructed according to a label matrix corresponding to the ontology protein function.

[0031] The Key vector, the Value vector, and the Query vector are input into an attention mechanism layer and a category dense layer of the bidirectional long short-term memory model, and the attention mechanism layer and the category dense layer are trained according to the Key vector, the Value vector, and the Query vector to obtain a protein function recognition model.

[0032] Optionally, the protein function recognition method, wherein the protein function recognition model is obtained by inputting the protein sequence into the protein function recognition model, and outputting a protein function recognition result.

[0033] A protein sequence to be identified is obtained, and the protein sequence to be identified is input into the protein function recognition model.

[0034] The protein function recognition model is used for protein function recognition processing on the to-be-recognized protein sequence, and a protein function recognition result corresponding to the to-be-recognized protein sequence is output.

[0035] Optionally, the protein function recognition method, wherein the to-be-recognized protein sequence is obtained, and the to-be-recognized protein sequence is input into the protein function recognition model to output a protein function recognition result, and then the method further comprises:

[0036] An activation function is obtained, a probability result that the to-be-recognized protein sequence has the protein function recognition result is calculated according to the activation function, and the protein function recognition model is fine-tuned according to the probability result.

[0037] In addition, to achieve the above object, the present application further provides a protein function recognition system, wherein the protein function recognition system comprises:

[0038] A multi-dimensional feature vector generation module is configured to obtain an ontology protein sequence and an ontology protein function corresponding to the ontology protein sequence, and input the ontology protein sequence into a plurality of deep learning pre-training models respectively to output a plurality of multi-dimensional feature vectors.

[0039] A model training module is configured to perform feature fusion processing on the plurality of multi-dimensional feature vectors to obtain a fusion feature vector, and train a bidirectional long short-term memory model according to the fusion feature vector and the ontology protein function to obtain a protein function recognition model.

[0040] A protein function recognition module is configured to obtain a to-be-recognized protein sequence, and input the to-be-recognized protein sequence into the protein function recognition model to output a protein function recognition result.

[0041] In addition, to achieve the above object, the present application further provides a terminal, wherein the terminal comprises a memory, a processor, and a protein function recognition program stored in the memory and executable on the processor, and the protein function recognition program implements the steps of the protein function recognition method when executed by the processor.

[0042] In addition, to achieve the above object, the present application further provides a computer readable storage medium, wherein the computer readable storage medium stores a protein function recognition program, and the protein function recognition program implements the steps of the protein function recognition method when executed by a processor.

[0043] In the present application, the ontology protein sequence and the ontology protein function corresponding to the ontology protein sequence are obtained, and the ontology protein sequence is respectively input into a plurality of deep learning pre-training models to output a plurality of multi-dimensional feature vectors; the plurality of multi-dimensional feature vectors are subjected to feature fusion processing to obtain a fusion feature vector, and a bidirectional long short-term memory model is trained according to the fusion feature vector and the ontology protein function to obtain a protein function recognition model; a to-be-recognized protein sequence is obtained, and the to-be-recognized protein sequence is input into the protein function recognition model to output a protein function recognition result. The trained protein language model is ingeniously integrated to extract deep features and recognize complex patterns embedded in the protein sequence. In its core architecture, a bidirectional long short-term memory network model is used to effectively capture the inherent long-term dependencies in sequence data. In order to further enhance the distinguishing ability of the model, a category attention mechanism is introduced to strengthen the model's attention to key information. In addition, a category dense layer is added, which helps to carefully process and classify potential protein functions, and can accurately obtain the relationship between protein sequences and functions, so as to quickly and accurately realize the recognition and prediction of protein functions. BRIEF DESCRIPTION OF DRAWINGS

[0044] Fig. 1 is a flowchart of the preferred embodiment of the protein function recognition method of the present application;

[0045] Fig. 2 is a complete flowchart of the preferred embodiment of the protein function recognition method of the present application;

[0046] Fig. 3 is a schematic diagram of the ESM2 model representation method of the preferred embodiment of the protein function recognition method of the present application;

[0047] Fig. 4 is a schematic diagram of the ProtBert model representation method of the preferred embodiment of the protein function recognition method of the present application;

[0048] Fig. 5 is a schematic diagram of the Protein2vec model representation method of the preferred embodiment of the protein function recognition method of the present application;

[0049] Fig. 6 is a schematic diagram of the onehot model representation method of the preferred embodiment of the protein function recognition method of the present application;

[0050] Fig. 7 is a comparative analysis schematic diagram of the single protein sequence embedding technology of the preferred embodiment of the protein function recognition method of the present application;

[0051] Fig. 8 is a comparative analysis schematic diagram of the protein sequence embedding combination of the preferred embodiment of the protein function recognition method of the present application;

[0052] Fig. 9 is a structure diagram of the preferred embodiment of the protein function recognition system of the present application;

[0053] FIG. 10 is a block diagram of a preferred embodiment of the terminal of the present application. DETAILED DESCRIPTION

[0054] In order to make the objectives, technical solutions, and advantages of the present application clearer and more explicit, the present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific examples described herein are merely intended to explain the present application and not to limit the present application.

[0055] Proteins are extremely important compounds in living organisms, they are not only the key components of biological structures, but also play a variety of important roles in life activities. In life science and medical research, identifying and understanding the functions of proteins is crucial, and protein function prediction is a complex task that faces multiple challenges. First, proteins themselves have high complexity, including their amino acid sequences, three-dimensional structures, and multiple functions, the interactions and dependencies between these characteristics increase the difficulty of prediction. Second, many proteins not only have one function, but also assume multiple different biological roles, which makes protein function prediction a multi-label, multi-classification problem, solving this problem requires the ability to identify and distinguish the multiple functions of proteins, while considering their roles in different biological processes and structures. In addition, the function of a protein can be affected by its location in the cell, interactions, and expression patterns under different conditions.

[0056] Protein function prediction methods cover a variety of technical paths, from early sequence alignment to the latest deep learning models. Sequence alignment methods, such as Smith-Waterman and BLAST algorithms, are the earliest prediction means, which infer the function of the protein to be predicted by aligning known functional protein sequences. Subsequently, prediction methods based on functional domains and motifs emerged, which predict functions by identifying functional domains and motifs in proteins, these methods use known functional domain databases such as PROSITE and Pfam. With the development of machine learning technology, algorithms such as HMM and SVM were applied to protein function prediction, through training models to improve accuracy. In recent years, deep learning methods such as convolutional neural networks, recurrent neural networks, and transformer technology have made significant progress, these methods can automatically learn the representation and features of proteins, for example, ProteinBERT, DeepGO, and DeepGATGO models, extract high-level features from raw data, and thus improve the accuracy and coverage of prediction. In summary, the continuous evolution and innovation of protein function prediction methods provide diverse solutions to this challenging problem.

[0057] At present, the protein function prediction is generally realized by a deep learning method, which trains multiple gene ontology class spaces and realizes enhanced learning through an approximate semantic analysis method to improve the accuracy of protein function prediction.

[0058] The current deep learning method has two main technical defects: first, the current deep learning method excessively relies on static sequence information, has great limitations, only uses protein sequence information, and does not fully consider the influence of protein structure on function, resulting in that the prediction result is not comprehensive and accurate. Secondly, in the field of brain development related protein function prediction, there is no specific method, brain development is a complex process, and proteins play an important role in it, but the current prediction method fails to fully consider the particularity of brain development related proteins, and lacks specific prediction strategies and tools.

[0059] To solve the above problems, the application realizes a protein function recognition method based on a deep learning model, which aims to input protein sequences and their corresponding functions into a deep learning model to obtain the relationship between protein sequences and functions, so as to predict the relationship between proteins and their functions. Specifically, the feature data of multiple protein sequences is obtained, and the structure information of the protein sequences and the correlation information between the protein sequences are obtained through the constructed deep learning model, and then the information is used to predict the protein function on the brain development related protein data (herein, the application is described for application, but is not limited to brain development related protein prediction, and is also applicable to other proteins). The brain development related protein function prediction can be more accurate.

[0060] The protein function recognition method of the preferred embodiment of the application is shown in FIGS. 1 and 2, and the protein function recognition method comprises the following steps:

[0061] Step S10, obtaining ontology protein sequences and corresponding ontology protein functions of the ontology protein sequences, and inputting the ontology protein sequences into multiple deep learning pre-training models respectively, and outputting multiple multi-dimensional feature vectors.

[0062] The deep learning pre-training model includes a protein evolution scale model, a protein language model, a protein vector representation model and a one-hot encoding model; the multi-dimensional feature vector includes a first multi-dimensional feature vector, a second multi-dimensional feature vector, a third multi-dimensional feature vector and a fourth multi-dimensional feature vector.

[0063] The application uses a deep learning method to predict protein function, and systematically processes each protein sequence through four different deep learning pre-training models to obtain comprehensive embedding representation (wherein embedding is also called embedding vector), the deep learning pre-training models include ESM2 (protein evolution scale model), ProtBert (protein language model), Protein2vec (protein vector representation model) and onehot (one-hot encoding model).

[0064] Specifically, the gene ontology is determined, and the ontology label is determined according to the gene ontology; the ontology protein sequence and the ontology protein function corresponding to the ontology protein sequence are obtained by searching and processing in the preset open source database according to the ontology label.

[0065] First, the ontology related to brain development is selected from the gene ontology. Then, the protein sequence with the ontology label is searched from the Uniprot database. Then, each protein sequence is processed by four different protein language models to obtain comprehensive embedding representation (i.e. multiple multi-dimensional feature vectors in the application).

[0066] The protein evolution scale model, the protein language model, the protein vector representation model and the one-hot encoding model are obtained; the ontology protein sequence is input into the protein evolution scale model to output the first multi-dimensional feature vector. The ontology protein sequence is input into the protein language model to output the second multi-dimensional feature vector; the ontology protein sequence is input into the protein vector representation model to output the third multi-dimensional feature vector; the ontology protein sequence is input into the one-hot encoding model, and the ontology protein sequence is converted into a binary vector by the one-hot encoding model to obtain a multi-dimensional matrix; a label matrix is constructed according to the ontology protein function, and the corresponding relationship between the multi-dimensional matrix and the label matrix is obtained; the feature vectors in the multi-dimensional matrix are valued according to the corresponding relationship to obtain the fourth multi-dimensional feature vector.

[0067] The process of obtaining multi-dimensional feature vectors includes:

[0068] 1、As shown in Figure 3, Figure 3 is a 36-layer transformer model ESM2 containing a 300 million parameter training data set. The present application first uses a 36-layer transformer model ESM2 containing a 300 million parameter training data set, which can generate a 2560-dimensional embedding (after the body protein sequence is input into the ESM deep learning model, the last layer of the model outputs a vector with a fixed dimension of 2560, that is, the first multi-dimensional feature vector is obtained).

[0069] 2、As shown in Figure 4, Figure 4 is a ProtBert model based on BERT architecture. The present application constructs a ProtBert model based on BERT architecture. The ProtBert model will first be pre-trained on a wide protein sequence corpus (pre-training is to input a large amount of protein data to ProtBert model to learn the features therein, which are saved in the form of parameters in the model, and the trained parameters of the model can be directly called to convert the protein into a feature vector). During the training process, each sequence is regarded as an independent document, and the next sentence prediction is not used (the embedding vector dimension is 1024, that is, the second multi-dimensional feature vector in the present application is obtained).

[0070] 3、As shown in Figure 5, Figure 5 is a Protein2vec model. For the Protein2vec model, a k-mer method (k=3) is used to divide the sequence into a 100-dimensional vector, producing an embedding with a dimension of (L-k+1, 100), which is suitable for proteins with L amino acids.

[0071] 4、As shown in Figure 6, Figure 6 is a onehot model. The present application also applies a simple but effective one-hot encoding technique to represent 20 common amino acids as binary vectors (there are currently 20 common amino acids, so a 20-dimensional vector is used to represent it, and a specific position in the vector is assigned to each amino acid.

[0072] For example: A [1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]; C [0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]; D [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]… and so on, each amino acid is represented by a capital English letter), then the protein containing L amino acids is converted into an (L, 20) matrix (wherein an amino acid is a (1, 20) vector representation, two amino acids are a (2, 20) matrix, and L and amino acids are an (L, 20) matrix). Then the protein function is constructed into a label matrix (assuming n functions, as long as the protein sequence has this function, it is assigned a value of 1, and if it does not have this function, it is assigned a value of 0. Wherein the label matrix is similar to the matrix represented by {function 1, function 2, function 3}:

[0073] If the corresponding protein sequence exhibits a specific function related to the label, each position in the vector is assigned a value of 1; conversely, if the function does not exist, it is assigned a value of 0.

[0074] Step S20, performing feature fusion processing on a plurality of the multi-dimensional feature vectors to obtain a fused feature vector, and training a bidirectional long short-term memory model according to the fused feature vector and the ontology protein function to obtain a protein function recognition model.

[0075] As shown in FIG. 2, FIG. 2 is a complete flowchart of the present application, the protein sequence is first extracted by four pre-training models of deep learning, then the fused data is input into the bidirectional LSTM model to obtain the association between the sequences, then the output vector of the bidirectional LSTM is used as the Key vector and the Value vector of the attention mechanism layer, the protein function is converted into a label matrix and input into the attention mechanism layer as the Query vector, the association between the protein and its function is obtained through the attention mechanism, and finally the function of the protein is recognized through the category dense layer.

[0076] In the present application, after the protein sequence representation (i.e. the fused feature vector obtained after the protein sequence passes through the four models) is input into the bidirectional LSTM model, the context information of the combined sequence is obtained, then the protein vector is combined with the protein function using the multi-head attention mechanism, and finally the function of the protein is recognized through a dense layer.

[0077] Specifically, the first multi-dimensional feature vector, the second multi-dimensional feature vector, the third multi-dimensional feature vector and the fourth multi-dimensional feature vector are subjected to feature fusion processing to obtain a fusion feature vector; the fusion feature vector is input into a bidirectional long short-term memory model to obtain context information of the fusion feature vector, and the context information is converted into a vector matrix; a Key vector and a Value vector are constructed according to the vector matrix, and a Query vector is constructed according to a label matrix corresponding to the ontology protein function; the Key vector, the Value vector and the Query vector are input into an attention mechanism layer and a category dense layer of the bidirectional long short-term memory model, and the attention mechanism layer and the category dense layer are trained according to the Key vector, the Value vector and the Query vector to obtain a protein function recognition model.

[0078] Further, the four features extracted by the models are fused (i.e., the embedding vectors extracted by each pre-trained model are directly spliced, and the vector dimension of one protein after splicing is (2560, 1024, 100, 20, corresponding to ESM2, ProtBert, Protein2vec and onehot respectively), and the dimension of the fused feature vector after splicing is 3704), and then the fused feature vector is input into a bidirectional LSTM model to obtain the context information of the combined sequence. The bidirectional LSTM model (Bi-directional Long Short-Term Memory, Bi-LSTM, bidirectional long short-term memory network) is a special type of recurrent neural network, which is composed of two independent LSTM layers, one processing sequence from front to back (forward), and the other from back to front (reverse), and the outputs of the two layers are usually combined at each time step. The forward LSTM: processes the sequence step by step from time step 1 to T, receives input and updates its memory state at each time step, and then outputs a vector representing the context information from the beginning of the sequence to the current time step. The reverse LSTM layer: processes the sequence in reverse order from time step T to 1 in a similar manner to the forward layer, but it provides context information from the end of the sequence to the current time step. The outputs of the forward and reverse layers are usually combined at each time step, so that the output of each time step contains information from both directions.

[0079] As shown in FIG. 2, after the context information is converted into a vector matrix, the vector matrix is used as a Key vector and a Value vector of a category attention mechanism, and the label matrix is used as a Query vector. Finally, the Key vector, the Value vector and the Query vector are passed through a category dense layer to identify the protein function (after the protein sequence is converted into a vector by the deep learning model, the protein sequence is finally calculated by using the activation function softmax to calculate the probability that the protein may have the function through a fully connected layer in the category dense layer).

[0080] The protein sequence data and the human chromosome related data adopted in the present application do not need to be obtained through biological experiments, but can be directly downloaded from a public database. Meanwhile, the present application can extract sequence similarity information and structure information in the protein sequence through four different pre-training models of deep learning, and then explore the relationship between the protein and its function by using the constructed deep learning model RecGOBD (i.e., the protein function identification model in the present application, which can be called the gene ontology related brain development protein function identification model for the study of brain proteins).

[0081] In step S30, a to-be-identified protein sequence is obtained, and the to-be-identified protein sequence is input into the protein function identification model to output a protein function identification result.

[0082] Specifically, a to-be-identified protein sequence is obtained, and the to-be-identified protein sequence is input into the protein function identification model. The to-be-identified protein sequence is subjected to protein function identification processing by the protein function identification model, and a protein function identification result corresponding to the to-be-identified protein sequence is output.

[0083] After the protein function identification model is constructed in the present application, the protein sequence can be subjected to function identification. After the to-be-identified protein sequence is input into the protein function identification model, the to-be-identified protein sequence is analyzed by the protein function identification model, so as to quickly and accurately obtain the protein function corresponding to the to-be-identified protein sequence.

[0084] Further, an activation function is obtained, and a probability result that the to-be-identified protein sequence has the protein function identification result is calculated according to the activation function. The protein function identification model is subjected to model fine-tuning according to the probability result.

[0085] In addition, the present application adopts a comprehensive method to explore genetic mutations on the human genome and their effects on protein function. Specifically, first, a large amount of data is downloaded from the 22 autosomes of the human genome, and then 10838 amino acid sequences representing the chromosome state before and after mutation are carefully selected and extracted. By analyzing each amino acid sequence, the exact location of the mutation can be accurately determined. Then, the amino acid sequences before and after the mutation are input into the previously trained prediction model (i.e. the protein function recognition model in the present application), which integrates four different embeddings. This method can compare the functional differences of the same protein before and after mutation, and this comparison focuses on the changes in the GO (Gene Ontology) annotations of protein function, providing valuable information about the biological processes, cellular components, and molecular functions that the protein may be involved in. Finally, among the 10838 amino acid sequences, 2484 amino acid sequences were found to have GO changes due to mutation.

[0086] The present application can also be applied to Autism Spectrum Disorder (ASD, Autism Spectrum Disorder is a group of complex neurodevelopmental disorders characterized by impaired social interaction and communication, as well as restrictive and repetitive behavior patterns. This condition usually begins to manifest in early childhood and continues to affect individuals throughout their lives, and it is believed that the cause of this disorder may be the mutation of amino acids on the chromosome leading to changes in brain development-related functions), screening amino acid sequences from the 22 autosomes of the human genome, and accurately determining the exact location of the mutations by analyzing each sequence. Subsequently, the amino acid sequences before and after the mutation are input into the previously trained prediction model to compare the functional differences of the same protein before and after mutation (equivalent to predicting changes in protein function, which can be verified by artificial experiments later).

[0087] The present application is verified through the following experiment: the present application initially selects ontologies related to brain development from gene ontology (Gene Ontology, GO for short, is a widely used bioinformatics tool aimed at providing a consistent method of describing the properties of genes and gene products. It is a system containing three main ontologies for describing molecular function (Molecular Function), cellular component (Cellular Component), and biological process (Biological Process)). Then, the protein sequences with these ontology labels are retrieved from the Uniprot database, and then the feature vectors of these protein sequences are extracted through four different deep learning pre-training models. The feature vectors generated by different pre-training models are compared and fused, and then put into the deep learning model constructed by the present application for training and testing. In the present application, AUPR and AUROC are used as evaluation indicators, as shown in Figure 7, which is the AUROC and AUPR curves obtained by inputting the feature vectors obtained by comparing four different pre-training models processing protein sequences into RecGOBD. When only a single pre-training model is used for protein sequence embedding, it can be concluded that the ESM2 model provides better prediction results because it can extract the inherent features of the sequence itself and the closer relationship between sequences can be seen. The results show that the ESM model is the best for feature extraction.

[0088] As shown in Figure 8, which is the AUROC and AUPR curves obtained by inputting the fused features extracted by the pre-training model into the RecGOBD model. The present application has carried out detailed statistical analysis on the data, calculated the mean and variance of various embeddings, and used graphical representation for intuitive visualization. In particular, when combining four embedding techniques (one-hot encoding, protein2vec, ProtBert and ESM2), it is observed that this configuration performs well in performance indicators, achieving optimal mean values and showing extremely low variability. Specifically, this model configuration achieves an average AUROC of 0.917 (with a variance of ±0.003) and an average AUPR of 0.694 (with a variance of ±0.008). These results show that the model combining embedding techniques not only performs well in performance, but also shows significant stability and consistency. As can be seen from the results, the results obtained by fusing the features extracted by ESM2, ProtBert, Protein2vec and onehot model are the best.

[0089] In summary, the present application develops a protein function recognition model based on a deep learning model, which is specially used for predicting protein function, named RecGOBD (also known as gene ontology related brain development protein function recognition). The model ingeniously integrates the trained protein language model to extract deep features and recognize the complex patterns embedded in the protein sequence. In its core architecture, a bidirectional long short-term memory (Bi-LSTM) network is adopted to strategically deploy to effectively capture the inherent long-term dependencies in the sequence data. In order to further enhance the model's ability to distinguish protein functions, the present application also sets up a category attention mechanism to strengthen the model's focus on key information. In addition, a category dense layer is added, which helps to carefully process and classify the potential protein functions.

[0090] Further, as shown in FIG. 9, based on the above-mentioned protein function recognition method, the present application also correspondingly provides a protein function recognition system, wherein the protein function recognition system comprises:

[0091] A multi-dimensional feature vector generation module 51 is configured to obtain an ontology protein sequence and an ontology protein function corresponding to the ontology protein sequence, and input the ontology protein sequence into a plurality of deep learning pre-training models respectively, and output a plurality of multi-dimensional feature vectors;

[0092] A model training module 52 is configured to perform feature fusion processing on the plurality of multi-dimensional feature vectors to obtain a fusion feature vector, and train a bidirectional long short-term memory model according to the fusion feature vector and the ontology protein function to obtain a protein function recognition model;

[0093] A protein function recognition module 53 is configured to obtain a to-be-recognized protein sequence, and input the to-be-recognized protein sequence into the protein function recognition model to output a protein function recognition result.

[0094] Further, as shown in FIG. 10, based on the above-mentioned protein function recognition method and system, the present application also correspondingly provides a terminal, which comprises a processor 10, a memory 20 and a display 30. FIG. 10 only shows part of the components of the terminal, but it should be understood that it is not required to implement all the shown components, and more or fewer components can be alternatively implemented.

[0095] The memory 20 can be an internal storage unit of the terminal in some embodiments, such as a hard disk or a memory of the terminal. The memory 20 can also be an external storage device of the terminal in other embodiments, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal. Further, the memory 20 can include both the internal storage unit and the external storage device of the terminal. The memory 20 is used to store application software and various data installed on the terminal, such as program codes of the terminal, etc. The memory 20 can also be used to temporarily store data that has been output or will be output. In an embodiment, the memory 20 stores a protein function recognition program 40, which can be executed by the processor 10 to implement the protein function recognition method.

[0096] The processor 10 can be a central processing unit (CPU), a microprocessor or other data processing chip in some embodiments, which is used to run program codes or process data stored in the memory 20, such as to execute the protein function recognition method, etc.

[0097] The display 30 can be an LED display, a liquid crystal display, a touch liquid crystal display, an OLED (Organic Light-Emitting Diode) touch, etc. in some embodiments. The display 30 is used to display information of the terminal and to display a visual user interface. The components 10-30 of the terminal communicate with each other through a system bus.

[0098] In an embodiment, the following steps are implemented when the processor 10 executes the protein function recognition program 40 in the memory 20:

[0099] Obtaining a body protein sequence and a body protein function corresponding to the body protein sequence, and inputting the body protein sequence into a plurality of deep learning pre-training models respectively to output a plurality of multi-dimensional feature vectors;

[0100] Performing feature fusion processing on the plurality of multi-dimensional feature vectors to obtain a fusion feature vector, and training a bidirectional long short-term memory model according to the fusion feature vector and the body protein function to obtain a protein function recognition model;

[0101] Obtaining a to-be-recognized protein sequence, and inputting the to-be-recognized protein sequence into the protein function recognition model to output a protein function recognition result.

[0102] wherein the obtaining the ontology protein sequence and the ontology protein function corresponding to the ontology protein sequence, and inputting the ontology protein sequence into a plurality of deep learning pre-training models respectively, outputting a plurality of multi-dimensional feature vectors, specifically comprising:

[0103] determining a gene ontology, and determining an ontology label according to the gene ontology;

[0104] According to the ontology label, searching and finding in the preset open source database, obtaining the ontology protein sequence and the ontology protein function corresponding to the ontology protein sequence;

[0105] obtaining a plurality of deep learning pre-training models, and inputting the ontology protein sequence into a plurality of deep learning pre-training models respectively, outputting a plurality of multi-dimensional feature vectors.

[0106] wherein the deep learning pre-training model includes a protein evolution scale model, a protein language model, a protein vector representation model and a one-hot encoding model; the multi-dimensional feature vector includes a first multi-dimensional feature vector, a second multi-dimensional feature vector, a third multi-dimensional feature vector and a fourth multi-dimensional feature vector;

[0107] The obtaining a plurality of deep learning pre-training models, and inputting the ontology protein sequence into a plurality of deep learning pre-training models respectively, outputting a plurality of multi-dimensional feature vectors, specifically comprising:

[0108] obtaining a protein evolution scale model, a protein language model, a protein vector representation model and a one-hot encoding model;

[0109] inputting the ontology protein sequence into the protein evolution scale model, outputting a first multi-dimensional feature vector;

[0110] inputting the ontology protein sequence into the protein language model, outputting a second multi-dimensional feature vector;

[0111] inputting the ontology protein sequence into the protein vector representation model, outputting a third multi-dimensional feature vector;

[0112] inputting the ontology protein sequence into the one-hot encoding model, outputting a fourth multi-dimensional feature vector.

[0113] wherein the inputting the ontology protein sequence into the one-hot encoding model, outputting a fourth multi-dimensional feature vector, specifically comprising:

[0114] inputting the ontology protein sequence into the one-hot encoding model, and converting the ontology protein sequence into a binary vector through the one-hot encoding model to obtain a multi-dimensional matrix;

[0115] constructing a tag matrix according to the ontology protein function, and obtaining a correspondence between the multi-dimensional matrix and the tag matrix;

[0116] According to the correspondence, the feature vectors in the multi-dimensional matrix are valued and processed to obtain the fourth multi-dimensional feature vector.

[0117] Among them, the multi-dimensional feature vectors are fused to obtain a fusion feature vector, and a bidirectional long short-term memory model is trained according to the fusion feature vector and the ontology protein function to obtain a protein function recognition model, specifically including:

[0118] The first multi-dimensional feature vector, the second multi-dimensional feature vector, the third multi-dimensional feature vector and the fourth multi-dimensional feature vector are fused to obtain a fusion feature vector;

[0119] The fusion feature vector is input into a bidirectional long short-term memory model to obtain context information of the fusion feature vector, and the context information is converted into a vector matrix;

[0120] According to the vector matrix, a Key vector and a Value vector are constructed, and a Query vector is constructed according to the tag matrix corresponding to the ontology protein function;

[0121] The Key vector, the Value vector and the Query vector are input into the attention mechanism layer and the category dense layer of the bidirectional long short-term memory model, and the attention mechanism layer and the category dense layer are trained according to the Key vector, the Value vector and the Query vector to obtain a protein function recognition model.

[0122] Among them, the protein sequence to be identified is obtained, and the protein sequence to be identified is input into the protein function recognition model to output a protein function recognition result, specifically including:

[0123] The protein sequence to be identified is obtained, and the protein sequence to be identified is input into the protein function recognition model;

[0124] The protein function recognition model is used to perform protein function recognition processing on the protein sequence to be identified, and a protein function recognition result corresponding to the protein sequence to be identified is output.

[0125] Among them, the protein sequence to be identified is obtained, and the protein sequence to be identified is input into the protein function recognition model to output a protein function recognition result, and then further including:

[0126] An activation function is obtained, and a probability result of the to-be-identified protein sequence having the protein function recognition result is calculated according to the activation function, and the protein function recognition model is fine-tuned according to the probability result.

[0127] The application further provides a computer readable storage medium, wherein the computer readable storage medium stores a protein function recognition program, and the protein function recognition program, when executed by a processor, implements the steps of the protein function recognition method.

[0128] In summary, the application provides a protein function recognition method and related equipment, the method comprising: obtaining a body protein sequence and a body protein function corresponding to the body protein sequence, and inputting the body protein sequence into a plurality of deep learning pre-training models respectively to output a plurality of multi-dimensional feature vectors; performing feature fusion processing on the plurality of multi-dimensional feature vectors to obtain a fusion feature vector, and training a bidirectional long short-term memory model according to the fusion feature vector and the body protein function to obtain a protein function recognition model; obtaining a to-be-identified protein sequence and inputting the to-be-identified protein sequence into the protein function recognition model to output a protein function recognition result. The application ingeniously integrates a trained protein language model to extract deep features and recognize complex patterns embedded in a protein sequence. In its core architecture, a bidirectional long short-term memory network model is used to effectively capture the inherent long-term dependencies in sequence data. In order to further enhance the distinguishing ability of the model, a category attention mechanism is introduced to strengthen the model's attention to key information. In addition, a category dense layer is added, which helps to carefully process and classify potential protein functions, and can accurately obtain the relationship between protein sequences and functions, so as to quickly and accurately recognize and predict protein functions.

[0129] It should be noted that in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that a process, method, article or terminal including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such a process, method, article or terminal. Without more limitations, the element defined by the statement "including a" does not exclude the presence of another identical element in the process, method, article or terminal including the element.

[0130] Of course, 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 relevant hardware (such as a processor, a controller, etc.) through a computer program. The program can be stored in a computer-readable computer-readable storage medium, and the program can include the processes of the above-mentioned method embodiments when executed. The computer-readable storage medium can be a memory, a disk, an optical disk, etc.

[0131] It should be understood that the application is not limited to the above examples, and those skilled in the art can make improvements or changes according to the above description, and all these improvements and changes shall belong to the protection scope of the appended claims of the application.

Claims

1. A method for identifying protein function, characterized in that, The method for identifying the function of the protein includes: Obtain the ontological protein sequence and the ontological protein function corresponding to the ontological protein sequence, and input the ontological protein sequence into multiple deep learning pre-trained models to output multiple multidimensional feature vectors; The multidimensional feature vectors are fused to obtain a fused feature vector, and the bidirectional long short-term memory model is trained based on the fused feature vector and the ontological protein function to obtain a protein function recognition model. The protein sequence to be identified is obtained and input into the protein function identification model, and the protein function identification result is output.

2. The method for identifying protein function according to claim 1, characterized in that, The process of acquiring the ontology protein sequence and the corresponding ontology protein function, and inputting the ontology protein sequence into multiple deep learning pre-trained models to output multiple multidimensional feature vectors, specifically includes: Determine the gene ontology, and determine the ontology tag based on the gene ontology; Based on the ontology tag, a search and processing is performed in a preset open-source database to obtain the ontology protein sequence and the ontology protein function corresponding to the ontology protein sequence. Multiple deep learning pre-trained models are obtained, and the ontology protein sequence is input into each of the multiple deep learning pre-trained models to output multiple multidimensional feature vectors.

3. The method for identifying protein function according to claim 2, characterized in that, The deep learning pre-trained models include protein evolution scaling models, protein language models, protein vector representation models, and one-hot encoding models. The multidimensional feature vector includes a first multidimensional feature vector, a second multidimensional feature vector, a third multidimensional feature vector, and a fourth multidimensional feature vector; The process of acquiring multiple deep learning pre-trained models and inputting the ontology protein sequence into each of the multiple deep learning pre-trained models to output multiple multidimensional feature vectors specifically includes: Obtain protein evolution scaling models, protein language models, protein vector representation models, and one-hot coding models; The ontological protein sequence is input into the protein evolution scaling model, and the first multidimensional feature vector is output. The ontology protein sequence is input into the protein language model, and a second multidimensional feature vector is output. The ontological protein sequence is input into the protein vector representation model, and a third multidimensional feature vector is output. The ontological protein sequence is input into the one-hot coding model, and a fourth multidimensional feature vector is output.

4. The method for identifying protein function according to claim 3, characterized in that, The step of inputting the ontological protein sequence into the one-hot coding model and outputting the fourth multidimensional feature vector specifically includes: The ontological protein sequence is input into the one-hot coding model, and the ontological protein sequence is subjected to binary vector conversion processing through the one-hot coding model to obtain a multidimensional matrix; A tag matrix is ​​constructed based on the ontological protein function, and the correspondence between the multidimensional matrix and the tag matrix is ​​obtained; The eigenvectors in the multidimensional matrix are assigned values ​​according to the correspondence to obtain the fourth multidimensional eigenvector.

5. The method for identifying protein function according to claim 4, characterized in that, The step of performing feature fusion processing on multiple multidimensional feature vectors to obtain a fused feature vector, and training a bidirectional long short-term memory model based on the fused feature vector and the ontological protein function to obtain a protein function recognition model, specifically includes: The first multidimensional feature vector, the second multidimensional feature vector, the third multidimensional feature vector, and the fourth multidimensional feature vector are subjected to feature fusion processing to obtain a fused feature vector. The fused feature vector is input into a bidirectional long short-term memory model to obtain the context information of the fused feature vector, and the context information is converted into a vector matrix; Construct a Key vector and a Value vector based on the vector matrix, and construct a Query vector based on the tag matrix corresponding to the ontology protein function; The Key vector, Value vector, and Query vector are input into the attention mechanism layer and category-dense layer of the bidirectional long short-term memory model, and the attention mechanism layer and category-dense layer are trained based on the Key vector, Value vector, and Query vector to obtain a protein function recognition model.

6. The method for identifying protein function according to claim 1, characterized in that, The process of obtaining the protein sequence to be identified, inputting the protein sequence into the protein function identification model, and outputting the protein function identification result specifically includes: Obtain the protein sequence to be identified and input the protein sequence to be identified into the protein function identification model; The protein function identification model is used to perform protein function identification processing on the protein sequence to be identified, and the protein function identification result corresponding to the protein sequence to be identified is output.

7. The method for identifying protein function according to claim 1 or 6, characterized in that, The process of obtaining the protein sequence to be identified, inputting the protein sequence to be identified into the protein function identification model, and outputting the protein function identification result further includes: An activation function is obtained, and the probability of the protein sequence to be identified possessing the protein function identification result is calculated based on the activation function. The protein function identification model is then fine-tuned based on the probability result.

8. A protein function recognition system, characterized in that, The protein function identification system includes: A multidimensional feature vector generation module is used to obtain the ontological protein sequence and the ontological protein function corresponding to the ontological protein sequence, and input the ontological protein sequence into multiple deep learning pre-trained models to output multiple multidimensional feature vectors. The model training module is used to perform feature fusion processing on multiple multidimensional feature vectors to obtain fused feature vectors, and to train the bidirectional long short-term memory model based on the fused feature vectors and the ontological protein function to obtain a protein function recognition model. The protein function identification module is used to acquire the protein sequence to be identified, input the protein sequence to be identified into the protein function identification model, and output the protein function identification result.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a protein function identification program stored in the memory and executable on the processor, wherein when the protein function identification program is executed by the processor, it implements the steps of the protein function identification method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a protein function identification program, which, when executed by a processor, implements the steps of the protein function identification method as described in any one of claims 1-7.

Images