Method and device for predicting protein-protein interaction, electronic equipment and storage medium

Abstract

The application provides a protein interaction prediction method and device, electronic equipment and storage medium, comprising: obtaining a first amino acid sequence corresponding to a first protein and a second amino acid sequence corresponding to a second protein; performing protein prediction based on the first amino acid sequence and the second amino acid sequence through a pre-trained target prediction network to obtain a protein prediction result; wherein the target prediction network comprises an amino acid embedding subnetwork, a vector encoding subnetwork and a prediction subnetwork, and the protein prediction result is used to represent the probability of interaction between the first protein and the second protein. The application can efficiently and accurately predict whether the proteins will interact, and can also realize cross-species protein interaction prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of protein interaction prediction, and in particular to a protein interaction prediction method and device, an electronic device and a storage medium. BACKGROUND

[0002] Predicting the interaction between proteins helps to reveal the life activity process of cells, and is an important means of mining and verifying functional genes and other markers from massive data. Traditional protein prediction techniques include wet experiment methods and computational methods. Among them, yeast two-hybrid, immunoprecipitation, fluorescence resonance energy transfer and the like are typical wet experiment methods, but wet experiment methods usually require a large amount of sample, highly purified protein, and are time-consuming and costly; in addition, the development of computational biology and bioinformatics technology provides a new method for exploring protein interaction, but the existing computational methods usually have OOD (Out of distribution) problem, making it difficult to make accurate predictions for unknown species of proteins. SUMMARY

[0003] Therefore, the purpose of the present application is to provide a protein interaction prediction method, device, electronic device and storage medium, which can efficiently and accurately predict whether proteins will interact with each other, and can also realize cross-species protein interaction prediction.

[0004] In a first aspect, an embodiment of the present application provides a protein interaction prediction method, comprising: obtaining a first amino acid sequence corresponding to a first protein and a second amino acid sequence corresponding to a second protein; performing protein prediction based on the first amino acid sequence and the second amino acid sequence through a pre-trained target prediction network to obtain a protein prediction result; wherein the target prediction network comprises an amino acid embedding subnetwork, a vector encoding subnetwork and a prediction subnetwork, and the protein prediction result is used to represent the probability of interaction between the first protein and the second protein.

[0005] In an implementation, the vector encoding sub-network includes a first vector encoding sub-network and a second vector encoding sub-network, and the output of the amino acid embedding sub-network is connected to the input of the first vector encoding sub-network and the second vector encoding sub-network respectively, and the output of the first vector encoding sub-network and the second vector encoding sub-network is connected to the prediction sub-network; the protein prediction is performed based on the first amino acid sequence and the second amino acid sequence by the pre-trained target prediction network to obtain a protein prediction result, including: embedding the first amino acid sequence into a vector space by the amino acid embedding sub-network to obtain a first embedding vector corresponding to the first protein, and embedding the second amino acid sequence into the vector space to obtain a second embedding vector corresponding to the second protein; encoding the first embedding vector by the first vector encoding sub-network to obtain a first encoding result corresponding to the first protein; and encoding the second embedding vector by the second vector encoding sub-network to obtain a second encoding result corresponding to the second protein; performing protein prediction based on the first encoding result and the second encoding result by the prediction sub-network to obtain a protein prediction result.

[0006] In an implementation, the first vector encoding sub-network includes a projection unit and a plurality of encoding units; the first embedding vector is encoded by the first vector encoding sub-network to obtain a first encoding result corresponding to the first protein, including: compressing the first embedding vector from a current dimension to a specified dimension by the projection unit; performing self-attention calculation and feedforward neural calculation on the first embedding vector of the specified dimension or the output vector of a previous encoding unit by the encoding unit to obtain an output vector of the encoding unit; and determining the output vector of the encoding unit at the end as the first encoding result corresponding to the first protein.

[0007] In an implementation, the encoding unit includes a multi-head attention layer and a feedforward neural layer; the output vector of the previous encoding unit is calculated by the encoding unit to obtain an output vector of the encoding unit, including: performing multi-head attention operation on the output vector of the previous encoding unit by the multi-head attention layer, and performing normalization operation on the sum of the output vector of the previous encoding unit and the result of the multi-head attention operation to obtain an intermediate vector; the output vector of the encoding unit is determined according to the intermediate vector by the feedforward neural layer according to the following formula: wherein, the output vector of the encoding unit is represented as y, the intermediate vector is represented as h, the output vector of the previous encoding unit is represented as x, the weight matrix of the multi-head attention layer is represented as W, the weight matrix of the feedforward neural layer is represented as V, and the bias vector of the feedforward neural layer is represented as b. ​wherein LN represents the normalization operation, and W1, W2, b1, b2 are network parameters of the feed-forward neural layer.

[0008] In an embodiment, the first vector encoding subnetwork and the second vector encoding subnetwork adopt a twin network architecture, and the first vector encoding subnetwork and the second vector encoding subnetwork share network parameters.

[0009] In an embodiment, the protein prediction is performed based on the first encoding result and the second encoding result to obtain a protein prediction result, including: performing average pooling operation on the first encoding result to obtain a first average pooling result, and performing average pooling operation on the second encoding result to obtain a second average pooling result; determining a Hadamard product of the first average pooling result and the second average pooling result; determining a protein prediction result based on the Hadamard product and network parameters of the prediction subnetwork by using a Softmax function; wherein if a probability represented by the protein prediction result is greater than a preset threshold, it is determined that interaction occurs between the first protein and the second protein, and if the probability represented by the protein prediction result is less than the preset threshold, it is determined that no interaction occurs between the first protein and the second protein.

[0010] In an embodiment, the method further includes: training network parameters of the amino acid embedding subnetwork by using a first training data set; wherein the amino acid embedding subnetwork adopts a PortT5 model; freezing the network parameters of the amino acid embedding subnetwork, and training network parameters of the vector encoding subnetwork and the prediction subnetwork by using a second training set to obtain a target prediction network.

[0011] In a second aspect, an embodiment of the present application further provides a protein interaction prediction device, including: a sequence acquisition module configured to acquire a first amino acid sequence corresponding to a first protein and a second amino acid sequence corresponding to a second protein; and a prediction module configured to perform protein prediction based on the first amino acid sequence and the second amino acid sequence by using a pre-trained target prediction network to obtain a protein prediction result; wherein the target prediction network includes an amino acid embedding subnetwork, a vector encoding subnetwork and a prediction subnetwork, and the protein prediction result is used to represent a probability of interaction occurring between the first protein and the second protein.

[0012] In a third aspect, an embodiment of the present application further provides an electronic device including a processor and a memory, wherein the memory stores computer executable instructions capable of being executed by the processor, and the processor executes the computer executable instructions to implement the method according to any one of the first aspect.

[0013] In a fourth aspect, the embodiments of the present application further provide a computer readable storage medium, which stores computer executable instructions, and the computer executable instructions, when invoked and executed by a processor, cause the processor to implement the method according to any one of the first aspect.

[0014] The method, device, electronic device and storage medium for predicting protein interaction provided by the embodiments of the present application first obtain a first amino acid sequence corresponding to a first protein and a second amino acid sequence corresponding to a second protein, and then perform protein prediction based on the first amino acid sequence and the second amino acid sequence through a pre-trained target prediction network to obtain a protein prediction result. The target prediction network includes an amino acid embedding sub-network, a vector encoding sub-network and a prediction sub-network, and the protein prediction result is used to represent the probability of interaction between the first protein and the second protein. The above method provides a target prediction network including an amino acid embedding sub-network, a vector encoding sub-network and a prediction sub-network. Using the target prediction network, whether the proteins will interact can be efficiently and accurately predicted according to the amino acid sequences of the proteins, and cross-species protein interaction prediction can also be realized.

[0015] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application will be realized and achieved by the structure particularly pointed out in the description, claims and drawings.

[0016] In order to make the above objectives, characteristics and advantages of the present application more apparent and easy to understand, the following preferred embodiments are specifically described below, and the accompanying drawings are referred to for detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0018] Figure 1 A flowchart of a protein interaction prediction method provided by the embodiments of the present application;

[0019] Figure 2 A structure diagram of a target prediction network provided by the embodiments of the present application;

[0020] Figure 3Another structure schematic diagram of a target prediction network provided by an embodiment of the present application is provided.

[0021] Figure 4 A flowchart of another protein interaction prediction method provided by an embodiment of the present application is provided.

[0022] Figure 5 A model prediction effect provided by an embodiment of the present application is provided.

[0023] Figure 6 A structure schematic diagram of a protein interaction prediction device provided by an embodiment of the present application is provided.

[0024] Figure 7 A structure schematic diagram of an electronic device provided by an embodiment of the present application is provided. DETAILED DESCRIPTION

[0025] To make the objectives, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described below in connection with the embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0026] Proteins are important macromolecules in living organisms, formed by folding long chains of amino acids in space into three-dimensional structures. Proteins usually do not act alone, but form "molecular machines" by interacting with other proteins to perform biological functions at the cellular and tissue levels. Protein interaction, also known as protein interaction, refers to the physical contact between proteins through electrostatic attraction, hydrogen bonds, hydrophobic forces, etc. non-random collision, docking at specific sites to form a complex. Therefore, whether protein interaction depends on the geometry and physicochemical properties of the protein surface, and it is a complex process. Protein interaction is one of the basic units of function execution at the molecular level, and it runs through various links of life processes such as gene expression, signal transduction, and immune regulation. For example, the spike protein of the new coronavirus interacts with and binds to the host cell target ACE2 receptor protein, realizes fusion with the host cell membrane, and thus completes the infection of the host cell.

[0027] Traditional protein prediction techniques include wet experimental methods and computational methods. Yeast two-hybrid, co-immunoprecipitation, fluorescence resonance energy transfer, etc. are typical wet experimental methods, but wet experimental methods usually require a large amount of sample, highly purified protein, and are time-consuming and costly. In recent years, the development of computational biology and bioinformatics technology has provided a new method for exploring protein interactions, such as computer simulation and molecular dynamics simulation based on protein sequence and structure, which can predict protein interactions and provide information on the mechanism and dynamics of protein interactions. The development of deep learning technology provides a new method for predicting protein interactions. Deep learning technology has the ability to process large-scale data, which can learn features from massive protein sequence and structure data and make predictions of protein interactions. However, deep learning models usually have OOD (Out of distribution) problems. When there is a difference in the distribution of the test set and the training set, the prediction accuracy of the deep learning prediction model will decrease significantly. In the task of protein interaction prediction, the OOD problem is that the interaction prediction model is difficult to make accurate predictions for proteins of unknown species.

[0028] Based on this, the present embodiment provides a protein interaction prediction method, device, electronic equipment and storage medium, which can efficiently and accurately predict whether proteins will interact, and also realize cross-species protein interaction prediction.

[0029] In order to facilitate the understanding of the present embodiment, first, a protein interaction prediction method disclosed by the present embodiment is introduced in detail, referring to the flowchart of the protein interaction prediction method shown in Figure 1 The method mainly includes the following steps S102 to S104:

[0030] Step S102, obtaining a first amino acid sequence corresponding to a first protein and a second amino acid sequence corresponding to a second protein. The first protein and the second protein can be proteins of the same species or proteins across species.

[0031] Step S104, performing protein prediction based on the first amino acid sequence and the second amino acid sequence through a pre-trained target prediction network to obtain a protein prediction result. The protein prediction result is used to represent the probability of interaction between the first protein and the second protein. Optionally, when the probability is greater than a preset threshold, it indicates that the first protein and the second protein can interact, otherwise when the probability is less than the preset threshold, it indicates that the first protein and the second protein will not interact.

[0032] In an implementation, the target prediction network can be referred to as PPITrans, which includes an amino acid embedding sub-network, a vector encoding sub-network, and a prediction sub-network. The amino acid embedding sub-network can adopt a PortT5 model, which is used to embed the amino acid sequences of proteins into a vector space to obtain embedding vectors of the proteins (which can also be referred to as embedding representations). The input of the amino acid embedding sub-network is the first amino acid sequence of the first protein and the second amino acid sequence of the second protein, and the output is the first embedding vector of the first protein and the second embedding vector of the second protein. The vector encoding sub-network is used to encode the embedding vectors to obtain encoding results. The input of the vector encoding sub-network is the first embedding vector and the second embedding vector, and the output is the first encoding result of the first protein and the second encoding result of the second protein. The prediction sub-network is used to perform average pooling operation on the encoding results of the proteins, and predict the probability of interaction between the two proteins based on the Hadamard product of the average pooling results. The input of the prediction sub-network is the first encoding result and the second encoding result, and the output is the protein prediction result.

[0033] The protein interaction prediction method provided by the embodiments of the present application provides a target prediction network, which includes an amino acid embedding sub-network, a vector encoding sub-network, and a prediction sub-network. The target prediction network can be used to efficiently and accurately predict whether proteins will interact with each other according to the amino acid sequences of the proteins, and can also realize cross-species protein interaction prediction.

[0034] To facilitate the understanding of the foregoing embodiments, the embodiments of the present application provide a target prediction network. Referring to Figure 2 , a structural schematic diagram of a target prediction network is shown, Figure 2 which shows that the target prediction network includes an amino acid embedding sub-network, a first vector encoding sub-network, a second vector encoding sub-network, and a prediction sub-network. The outputs of the amino acid embedding sub-network are connected to the inputs of the first vector encoding sub-network and the second vector encoding sub-network, respectively. The outputs of the first vector encoding sub-network and the second vector encoding sub-network are connected to the prediction sub-network.

[0035] In an implementation, in the network construction stage, the network parameters of the amino acid embedding sub-network can be trained by using a first training data set, and then the network parameters of the amino acid embedding sub-network are frozen, and the network parameters of the vector encoding sub-network and the prediction sub-network are trained by using a second training set to obtain the target prediction network. To facilitate understanding, the embodiments of the present application provide a specific structure of the target prediction network. Referring to Figure 3 , a structural schematic diagram of another target prediction network is shown, and specifically,

[0036] (1) For the amino acid embedding subnetwork: the embodiment of the present application adopts the ProtT5 model to dynamically embed the amino acid sequence, the ProtT5 model contains 3 billion parameters, which is first pre-trained on the BFD (Big Fantastic Database, protein sequence database) dataset, and then fine-tuned on the UniRef50 dataset. Among them, the UniRef50 dataset contains 45 million and 2.122 billion proteins, covering a wide range of biological species. After embedding is completed, each residue in the protein is embedded into a 1024-dimensional vector, which contains the position and context information of the residue.

[0037] (2) For the first vector encoding subnetwork and the second vector encoding subnetwork: the first vector encoding subnetwork and the second vector encoding subnetwork adopt a twin network architecture, and the first vector encoding subnetwork and the second vector encoding subnetwork share network parameters. Taking the first vector encoding subnetwork as an example, the first vector encoding subnetwork includes a projection unit and a plurality of encoding units, and each encoding unit includes a multi-head attention layer and a feed-forward neural layer. For example, the first vector encoding subunit and the second vector subunit can each include 6 encoding units. In actual application, since the ProtT5 model requires a large amount of computing resources during fine-tuning, the PPITrans freezes its parameters and adds encoding units (which can use Transformer encoders) on top of it to further encode the protein sequence. In the pre-trained amino acid embedding subnetwork, the embedding vector size has reached 1024, resulting in the need for more computing resources for the encoding unit, but the performance of the encoding unit cannot be improved, therefore, the PPITrans provided by the embodiment of the present application adds a projection unit before the encoding unit to convert the size of the embedding vector to the hidden size 256 of the encoding unit. Since each input sample contains two proteins, the embodiment of the present application adopts a twin network architecture for protein encoding, that is, the vector encoding subnetwork actually includes the first vector encoding subnetwork and the second vector encoding subnetwork, and the two vector encoding subnetworks share network parameters.

[0038] (3) For the prediction subnetwork: the prediction subnetwork performs average pooling operations on the first encoding result and the second encoding result respectively, and predicts the probability of interaction between the two proteins according to the Hadamard product of the two average pooling results.

[0039] On the basis of the above target prediction network, the embodiment of the present application provides an implementation of obtaining a protein prediction result by a pre-trained target prediction network based on a first amino acid sequence and a second amino acid sequence, see the following steps 1 to 3:

[0040] Step 1, embedding the first amino acid sequence into a vector space by an amino acid embedding subnetwork to obtain a first embedding vector corresponding to the first protein, and embedding the second amino acid sequence into the vector space to obtain a second embedding vector corresponding to the second protein. In an optional embodiment, first, the first amino acid sequence is input into the amino acid embedding subnetwork (PortT5 model), and the first amino acid sequence is embedded into the vector space by the amino acid embedding subnetwork to obtain the first embedding vector; then the second amino acid sequence is input into the amino acid embedding subnetwork, and the second amino acid sequence is embedded into the vector space by the amino acid embedding subnetwork to obtain the second embedding vector. Wherein each amino acid of the protein is embedded as a vector with a length of 1024 The embedding vector of the protein is represented as

[0041] Step 2, encoding the first embedding vector by a first vector encoding subnetwork to obtain a first encoding result corresponding to the first protein; and encoding the second embedding vector by a second vector encoding subnetwork to obtain a second encoding result corresponding to the second protein. For ease of understanding, the present embodiment takes the first vector encoding subnetwork as an example to provide an embodiment of encoding the first embedding vector by the first vector encoding subnetwork to obtain the first encoding result corresponding to the first protein, which is specifically described as follows:

[0042] Step 2.1, compressing the first embedding vector from the current dimension to the specified dimension by a projection unit. In an embodiment, the projection unit compresses the first embedding vector from 1024 dimensions to 265 dimensions:

[0043]

[0044] wherein, is the parameter matrix of the projection unit, LN represents the layer normalization operation, represents the first embedding vector of 265 dimensions, represents the first embedding vector of 1024 dimensions.

[0045] Step 2.2, by the encoding unit, performing self-attention calculation and feedforward neural calculation on the first embedding vector of the specified dimension or the output vector of the previous encoding unit to obtain the output vector of the encoding unit. In an embodiment, the protein is further encoded using a Transformer encoder, each layer of the Transformer encoder including two main steps: (1) self-attention calculation; and (2) feedforward neural layer calculation. Specifically, for the encoding unit at the head end, the 265-dimensional first embedding vector is sequentially subjected to self-attention calculation and feedforward neural calculation by the encoding unit, and for other encoding units, the output vector of the previous encoding unit is sequentially subjected to self-attention calculation and feedforward neural calculation by the encoding unit.

[0046] Step 2.3, determining the output vector of the encoding unit at the tail end as the first encoding result corresponding to the first protein.

[0047] For the convenience of understanding step 2.2, the embodiment of the present application further provides an implementation of obtaining the output vector of the encoding unit by performing self-attention calculation and feedforward neural calculation on the output vector of the previous encoding unit, see steps a to b as follows:

[0048] Step a, performing multi-head attention operation on the output vector of the previous encoding unit by the multi-head attention layer, and performing normalization operation on the sum of the output vector of the previous encoding unit and the result of the multi-head attention operation to obtain an intermediate vector. Specifically, the intermediate vector can be determined according to the following formula:

[0049]

[0050] wherein, represents the intermediate vector of the encoding unit, represents the output vector of the l-1th encoding unit, i.e., the output vector of the previous encoding unit, MultiHead represents the multi-head attention operation.

[0051] Step b, determining the output vector of the encoding unit according to the intermediate vector by the feedforward neural layer according to the following formula:

[0052]

[0053] wherein, represents the output vector, represents the intermediate vector, LN represents the normalization operation, W1, W2, b1, b2 are network parameters of the feedforward neural layer, b1∈R 4d and b2∈Rd represents the learnable parameters. Assuming the lengths of the two proteins are m and n respectively, after encoding, the first encoding result is The second encoding result is

[0054] Step 3, through the predictor network, the protein prediction is performed based on the first encoding result and the second encoding result, and the protein prediction result is obtained. For details, see steps 3.1 to 3.3 as follows:

[0055] Step 3.1, the first encoding result is subjected to an average pooling operation to obtain a first average pooling result, and the second encoding result is subjected to an average pooling operation to obtain a second average pooling result. In an embodiment, the encoding of the protein is subjected to an average pooling operation:

[0056]

[0057]

[0058] wherein a represents the first protein, b represents the second protein, that is, the first average pooling result, that is, the second average pooling result.

[0059] Step 3.2, the Hadamard product of the first average pooling result and the second average pooling result is determined.

[0060] Step 3.3, the protein prediction result is determined based on the Hadamard product and the network parameters of the predictor network by using a Softmax function. In an embodiment, a multi-layer perceptron can be used to predict according to the average pooling result, and for details, see the following formula:

[0061]

[0062] wherein, represents the Hadamard product, W c and b c represent the parameters of the prediction module, represents the probability of model prediction. In an embodiment, if the probability represented by the protein prediction result is greater than a preset threshold, it is determined that the interaction occurs between the first protein and the second protein, and if the probability represented by the protein prediction result is less than the preset threshold, it is determined that the interaction does not occur between the first protein and the second protein. For example, when , it is considered that the two proteins will interact, otherwise, they will not interact.

[0063] For the convenience of understanding the foregoing embodiments, a specific embodiment of a protein interaction prediction method is provided in the embodiments of the present application, which is described in detail with reference toFigure 4 Another flowchart of the prediction method of protein interaction is shown, which mainly comprises the following steps S402 to S408:

[0064] In step S402, an embedding representation of the protein is obtained using an amino acid embedding subnetwork. The embedding representation is the first embedding vector and the second embedding vector.

[0065] In step S404, the embedding representation of the protein is compressed from 1024 dimensions to 256 dimensions using a projection unit in the vector encoding subnetwork.

[0066] In step S406, the embedding representation of the protein is encoded using a Transformer encoder in the vector encoding subnetwork.

[0067] In step S408, the prediction subnetwork is used to make a prediction based on the encoded result of the protein to obtain a protein prediction result.

[0068] The embodiment of the present application compares PPITrans with existing deep learning-based protein interaction prediction models based on standard data sets, wherein the positive samples of the training set and the test set are from the STRING database, and the negative samples are constructed by random pairing, and the scale is ten times that of the positive samples. See Figure 5 The prediction effect of the model is shown, Figure 5 The prediction effect of all models on six model organism data is shown, and the evaluation indicators include F1 score and AUPR score, because both are more discriminant in non-uniform binary classification evaluation. The experimental results show that the classification performance of PPITrans greatly exceeds that of traditional interaction prediction models on the test sets of six model organisms. The AUPR scores on the test sets of human (Human), mouse (Mouse), and fruit fly (Fly) are all more than 0.9, and the F1 scores on the test sets of model organisms except yeast (Yeast) and E. coli (E. coli) are all more than 0.8, which shows that PPITrans can correctly predict most of the test samples. More importantly, on the test sets except human, PPITrans has greater advantages than traditional methods. On the fruit fly test set, PPITrans compared with the best model PIPR+DSCRIPT, the AUPR score is more than 0.34, and the F1 score is more than 0.31. The above results mean that PPITrans has stronger generalization ability.

[0069] In summary, the present application proposes and completes a prediction model based on protein sequence, called PPITrans. PPITrans includes an amino acid embedding subnetwork, a vector encoding subnetwork, and a prediction subnetwork. The amino acid embedding subnetwork adopts a pre-trained ProtT5 model to convert the amino acid sequence into a vector representation. The vector encoding subnetwork adopts a twin network structure based on Transformer, which is used to further encode the embedding vectors of the input two proteins. The prediction subnetwork adopts a multilayer perceptron to determine whether the interaction occurs according to the Hadamard product of the protein coding. The present application embodiment proposes and completes a protein interaction prediction model based on the Transformer network structure. The model can efficiently determine whether the proteins will interact according to the amino acid sequence of the proteins, and can predict the protein interaction across species.

[0070] For the protein interaction prediction method provided by the foregoing embodiment, the present application embodiment provides a protein interaction prediction device, which refers to the structural schematic diagram of a protein interaction prediction device shown in Figure 6 The device mainly includes the following parts:

[0071] The sequence acquisition module 602 is configured to acquire a first amino acid sequence corresponding to a first protein and a second amino acid sequence corresponding to a second protein.

[0072] The prediction module 604 is configured to perform protein prediction based on the first amino acid sequence and the second amino acid sequence by using a pre-trained target prediction network, to obtain a protein prediction result. The target prediction network includes an amino acid embedding subnetwork, a vector encoding subnetwork, and a prediction subnetwork. The protein prediction result is used to represent the probability of interaction between the first protein and the second protein.

[0073] The protein interaction prediction device provided by the present application embodiment provides a target prediction network, which includes an amino acid embedding subnetwork, a vector encoding subnetwork, and a prediction subnetwork. The target prediction network can be used to efficiently and accurately predict whether the proteins will interact according to the amino acid sequence of the proteins, and can also realize the prediction of protein interaction across species.

[0074] In one embodiment, the vector encoding subnetwork includes a first vector encoding subnetwork and a second vector encoding subnetwork. The output of the amino acid embedding subnetwork is connected to the inputs of the first vector encoding subnetwork and the second vector encoding subnetwork, respectively. The outputs of both the first and second vector encoding subnetworks are connected to the prediction subnetwork. The prediction module 604 is further configured to: embed a first amino acid sequence into a vector space through the amino acid embedding subnetwork to obtain a first embedding vector corresponding to a first protein, and embed a second amino acid sequence into a vector space to obtain a second embedding vector corresponding to a second protein; encode the first embedding vector through the first vector encoding subnetwork to obtain a first encoding result corresponding to the first protein; and encode the second embedding vector through the second vector encoding subnetwork to obtain a second encoding result corresponding to the second protein; and predict the protein based on the first and second encoding results through the prediction subnetwork to obtain a protein prediction result.

[0075] In one embodiment, the first vector encoding subnetwork includes a projection unit and multiple encoding units; the prediction module 604 is further configured to: compress the first embedding vector from the current dimension to a specified dimension through the projection unit; perform self-attention calculation and feedforward neural calculation on the first embedding vector of the specified dimension or the output vector of the previous encoding unit through the encoding unit to obtain the output vector of the encoding unit; and determine the output vector of the encoding unit located at the end as the first encoding result corresponding to the first protein.

[0076] In one embodiment, the encoding unit includes a multi-head attention layer and a feedforward neural layer; the prediction module 604 is further configured to: perform a multi-head attention operation on the output vector of the previous encoding unit through the multi-head attention layer, and normalize the sum of the output vector of the previous encoding unit and the result of the head attention operation to obtain an intermediate vector; and determine the output vector of the encoding unit based on the intermediate vector through the feedforward neural layer according to the following formula: Among them, the The output vector is represented by the... Let LN represent the intermediate vector, LN represent the normalization operation, and W1, W2, b1, and b2 are all network parameters of the feedforward neural layer.

[0077] In one implementation, the first vector coding subnetwork and the second vector coding subnetwork adopt a twin network architecture, and the first vector coding subnetwork and the second vector coding subnetwork share network parameters.

[0078] In an implementation, the prediction module 604 is further configured to: perform an average pooling operation on the first encoding result to obtain a first average pooling result, and perform an average pooling operation on the second encoding result to obtain a second average pooling result; determine a Hadamard product of the first average pooling result and the second average pooling result; determine a protein prediction result based on the Hadamard product and network parameters of the prediction subnetwork by using a Softmax function; and if a probability represented by the protein prediction result is greater than a preset threshold, determine that the first protein and the second protein interact with each other, and if the probability represented by the protein prediction result is less than the preset threshold, determine that the first protein and the second protein do not interact with each other.

[0079] In an implementation, the apparatus further includes a training module configured to: train network parameters of the amino acid embedding subnetwork by using a first training data set; wherein the amino acid embedding subnetwork adopts a PortT6 model; freeze the network parameters of the amino acid embedding subnetwork, and train network parameters of the vector encoding subnetwork and the prediction subnetwork by using a second training set to obtain a target prediction network.

[0080] The apparatus provided in the embodiments of the present application has the same implementation principle and technical effects as the foregoing method embodiments. For brevity, the apparatus embodiment part is not mentioned in the foregoing method embodiment part, and the corresponding content can be referred to in the foregoing method embodiment part.

[0081] The electronic device provided in the embodiments of the present application includes a processor and a storage device. The storage device stores a computer program. When the computer program is executed by the processor, the computer program performs the method according to any one of the foregoing embodiments.

[0082] Figure 7 A structural schematic diagram of an electronic device provided in the embodiments of the present application is shown in FIG. 1. The electronic device 100 includes a processor 70, a memory 71, a bus 72, and a communication interface 73. The processor 70, the communication interface 73, and the memory 71 are connected through the bus 72. The processor 70 is configured to execute executable modules stored in the memory 71, such as a computer program.

[0083] The memory 71 can include a high-speed random access memory (RAM) and can also include a non-volatile memory, such as at least one disk memory. The communication interface 73 can be wired or wireless, and can be used to implement communication connection between the system network element and at least one other network element, and can use the Internet, a wide area network, a local area network, a metropolitan area network, etc.

[0084] The bus 72 can be an ISA bus, a PCI bus, an EISA bus, or the like. The bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, Figure 7 Only one bidirectional arrow is used to represent the bus in the figure, but it does not mean that there is only one bus or only one type of bus.

[0085] The memory 71 is configured to store a program, and the processor 70 executes the program after receiving an execution instruction. The method executed by the device for defining the flow process according to any of the foregoing embodiments of the application can be applied to the processor 70 or implemented by the processor 70.

[0086] The processor 70 can be an integrated circuit chip with a processing capability of signals. In the implementation process, each step of the foregoing method can be completed by an integrated logic circuit of hardware in the processor 70 or an instruction in the form of software. The processor 70 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), and the like; or can be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. Each method, step, and logic block diagram disclosed in the embodiments of the application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, or other mature storage media in the art. The storage medium is located in the memory 71, and the processor 70 reads information in the memory 71 and combines the hardware to complete the steps of the foregoing method.

[0087] The computer program product of the readable storage medium provided by the embodiments of the application includes a computer readable storage medium storing program codes, and the program codes include instructions for executing the method described in the foregoing method embodiments. For specific implementation, reference can be made to the foregoing method embodiments, which will not be described here.

[0088] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the present application that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of software products. The computer software product is stored in a storage medium and includes instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0089] Finally, it should be noted that: the above-described embodiments are only specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit them. The protection scope of the present application is not limited thereto. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can make modifications or easily think of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed by the present application, or make equivalent replacements to some technical features. The modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for predicting protein-protein interaction, characterized by, The method comprises the following steps: obtaining a first amino acid sequence corresponding to a first protein and a second amino acid sequence corresponding to a second protein; performing protein prediction based on the first amino acid sequence and the second amino acid sequence through a pre-trained target prediction network to obtain a protein prediction result; wherein the target prediction network comprises an amino acid embedding subnetwork, a vector encoding subnetwork and a prediction subnetwork, the vector encoding subnetwork comprises a first vector encoding subnetwork and a second vector encoding subnetwork, the output of the amino acid embedding subnetwork is connected to the input of the first vector encoding subnetwork and the second vector encoding subnetwork respectively, the output of the first vector encoding subnetwork and the second vector encoding subnetwork is connected to the prediction subnetwork, the first vector encoding subnetwork comprises a projection unit and a plurality of encoding units, and the protein prediction result is used to represent the probability of interaction between the first protein and the second protein; performing protein prediction based on the first amino acid sequence and the second amino acid sequence through a pre-trained target prediction network to obtain a protein prediction result, comprising: embedding the first amino acid sequence into a vector space through the amino acid embedding subnetwork to obtain a first embedding vector corresponding to the first protein, and embedding the second amino acid sequence into the vector space to obtain a second embedding vector corresponding to the second protein; encoding the first embedding vector through the first vector encoding subnetwork to obtain a first encoding result corresponding to the first protein, and encoding the second embedding vector through the second vector encoding subnetwork to obtain a second encoding result corresponding to the second protein; performing protein prediction based on the first encoding result and the second encoding result through the prediction subnetwork to obtain a protein prediction result; encoding the first embedding vector through the first vector encoding subnetwork to obtain a first encoding result corresponding to the first protein, comprising: compressing the first embedding vector from a current dimension to a specified dimension through the projection unit; performing self-attention calculation and feedforward neural calculation on the first embedding vector of the specified dimension or the output vector of the previous encoding unit through the encoding unit to obtain the output vector of the encoding unit; determining the output vector of the encoding unit at the end as the first encoding result corresponding to the first protein.

2. The method of predicting protein-protein interaction according to claim 1, wherein, The encoding unit comprises a multi-head attention layer and a feedforward neural layer; performing self-attention calculation and feedforward neural calculation on the output vector of the previous encoding unit through the encoding unit to obtain the output vector of the encoding unit, comprising: performing multi-head attention operation on the output vector of the previous encoding unit through the multi-head attention layer, and performing normalization operation on the sum of the output vector of the previous encoding unit and the result of the multi-head attention operation to obtain an intermediate vector; determining the output vector of the encoding unit according to the intermediate vector through the feedforward neural layer according to the following formula: ； wherein the denotes the output vector, the denotes the intermediate vector, LN denotes the normalization operation, , , , are network parameters of the feedforward neural layer.

3. The method of predicting protein-protein interaction according to claim 1, wherein The first vector encoding subnetwork and the second vector encoding subnetwork adopt a twin network architecture, and the first vector encoding subnetwork and the second vector encoding subnetwork share network parameters.

4. The method of predicting protein-protein interaction according to claim 1, wherein Based on the first encoding result and the second encoding result, protein prediction is performed to obtain a protein prediction result, including: performing average pooling operation on the first encoding result to obtain a first average pooling result, and performing average pooling operation on the second encoding result to obtain a second average pooling result; determining the Hadamard product of the first average pooling result and the second average pooling result; using a Softmax function, determining a protein prediction result based on the Hadamard product and network parameters of the prediction subnetwork; wherein, if the probability represented by the protein prediction result is greater than a preset threshold, it is determined that interaction occurs between the first protein and the second protein, and if the probability represented by the protein prediction result is less than the preset threshold, it is determined that no interaction occurs between the first protein and the second protein.

5. The method of predicting protein-protein interactions according to any one of claims 1 to 4, characterized in that, The method further includes: training network parameters of the amino acid embedding subnetwork using a first training data set; wherein the amino acid embedding subnetwork adopts a PortT5 model; freezing the network parameters of the amino acid embedding subnetwork, and training network parameters of the vector encoding subnetwork and the prediction subnetwork using a second training set to obtain a target prediction network.

6. A device for predicting protein-protein interactions, characterized in that including: a sequence acquisition module for acquiring a first amino acid sequence corresponding to a first protein and a second amino acid sequence corresponding to a second protein; a prediction module for performing protein prediction based on the first amino acid sequence and the second amino acid sequence through a pre-trained target prediction network to obtain a protein prediction result; wherein the target prediction network includes an amino acid embedding subnetwork, a vector encoding subnetwork and a prediction subnetwork, the vector encoding subnetwork includes a first vector encoding subnetwork and a second vector encoding subnetwork, the output of the amino acid embedding subnetwork is connected to the input of the first vector encoding subnetwork and the second vector encoding subnetwork respectively, the output of the first vector encoding subnetwork and the second vector encoding subnetwork is connected to the prediction subnetwork, the first vector encoding subnetwork includes a projection unit and a plurality of encoding units, and the protein prediction result is used to represent the probability of interaction between the first protein and the second protein. The prediction module is specifically configured to: embed the first amino acid sequence into a vector space through the amino acid embedding subnetwork to obtain a first embedding vector corresponding to the first protein, and embed the second amino acid sequence into the vector space to obtain a second embedding vector corresponding to the second protein; encode the first embedding vector through the first vector encoding subnetwork to obtain a first encoding result corresponding to the first protein, and encode the second embedding vector through the second vector encoding subnetwork to obtain a second encoding result corresponding to the second protein; The predictor network is used to predict proteins based on the first encoding result and the second encoding result, and a protein prediction result is obtained. The prediction module is specifically configured to: The projection unit is configured to compress the first embedding vector from a current dimension to a specified dimension; The encoding unit is configured to perform self-attention calculation and feedforward neural calculation on the first embedding vector of the specified dimension or an output vector of a previous encoding unit, and obtain an output vector of the encoding unit; The output vector of the encoding unit at the end is determined as the first encoding result corresponding to the first protein.

7. An electronic device, comprising: The processor executes the computer executable instructions to implement the method in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores computer executable instructions, and the computer executable instructions, when called and executed by the processor, cause the processor to implement the method in any one of claims 1 to 5. The computer readable storage medium stores computer executable instructions, and the computer executable instructions, when called and executed by the processor, cause the processor to implement the method in any one of claims 1 to 5.

Images