A protein interaction prediction method, system, computer device and storage medium

Abstract

The present invention discloses a protein interaction prediction method, system, computer device and storage medium, which relates to the field of biotechnology, including: an input layer receives protein feature coding and reshapes it into a two-dimensional vector; a hidden layer extracts features through convolution, pooling, residual blocks and fully connected layers, and introduces nonlinear transformations through ReLU functions; an output layer outputs interaction probabilities through Sigmoid functions, and optimizes through binary cross entropy loss functions; and a protein interaction prediction model is trained through a stochastic gradient descent optimizer. The protein interaction prediction method provided by the present invention uses a UniPPI model with different combinations of traditional feature extraction. Compared with traditional machine learning models and deep learning models, the UniPPI model has higher accuracy, precision and F1 score.

Description

Technical Field

[0001] The present invention relates to the field of biotechnology, and in particular to a protein interaction prediction method, system, computer equipment and storage medium. Background Art

[0002] Protein-protein interaction is one of the important ways of interaction between molecules in cells. It plays a very important role in the life activities and metabolic regulation of cells. The study of protein-protein interactions is crucial to explore and clarify the biological processes of related proteins in cells. At present, the prediction methods of protein-protein interactions can generally be divided into methods based on biological experiments and methods based on calculations.

[0003] The protein interaction research methods based on biological realization mainly include the following:

[0004] 1. Yeast two-hybrid technology: This method connects proteins A and B through the DNA binding domain (DBD) and activation domain (AD) respectively. When A and B interact, a fully functional transcription factor is formed, which can activate the expression of the reporter gene. The expression can be used to determine whether the proteins interact. However, this method is not applicable to some membrane proteins and eukaryotic cell proteins and has some limitations.

[0005] 2. Immunoprecipitation technology: This method uses antibodies to precipitate the target protein or protein complex, so that their interaction relationship can be further studied. This technology uses specific antibodies to bind to the target protein to form an antigen-antibody complex, and then separates the complex through immunoprecipitation. Finally, the target protein is obtained through the elution step to detect whether there are other proteins that interact with it. This method has low efficiency for protein interaction detection with relatively low affinity, and the sensitivity of detection is limited by the concentration of the target protein. At the same time, the antibody equipment is relatively cumbersome and expensive.

[0006] 3. Bimolecular fluorescence complementation technology: This method marks two complementary fluorescent protein fragments on the protein of interest. When these fragments bind to each other, the fluorescent proteins will reassemble and produce fluorescent signals, thus marking the interaction of proteins. However, this technology has some limitations, such as the formation of fluorescent groups requires a certain amount of time, which may lead to a lag in the results, and the non-specific binding of fluorescent protein fragments may cause false positive results, affecting the accuracy of the results.

[0007] Computational protein interaction research methods usually include molecular docking methods based on genomic information, evolutionary information, protein interaction network, and physical model. These methods have their own advantages and limitations. Methods based on genomic information rely on a large amount of prior knowledge of the genome, and may be difficult to accurately predict in the absence of complete genetic information; methods based on evolutionary information are based on conservative evolutionary assumptions and are sensitive to the quality and quantity of evolutionary data; methods based on protein interaction networks rely entirely on known interaction networks and may miss potential interaction relationships; molecular docking methods based on physical models usually require three-dimensional structural information of proteins, and obtaining structural information is time-consuming and computationally complex, which limits its practical application range. Summary of the invention

[0008] In view of the above-mentioned problems, the present invention is proposed.

[0009] Therefore, the technical problems solved by the present invention are: 1. Biological realization is time-consuming, costly, and requires a lot of manpower and material resources.

[0010] Second, computational protein interaction prediction methods require complex computational processes and a large amount of biological prior knowledge.

[0011] To solve the above technical problems, the present invention provides the following technical solutions: a protein interaction prediction model, comprising: an input layer receives protein feature codes and reshapes them into two-dimensional vectors; a hidden layer extracts features through convolution, pooling, residual blocks and fully connected layers, and introduces nonlinear transformations through ReLU functions; an output layer outputs interaction probabilities through Sigmoid functions and optimizes them through binary cross entropy loss functions; and the protein interaction prediction model is trained through a stochastic gradient descent optimizer.

[0012] As a preferred embodiment of the protein interaction prediction model described in the present invention, the protein feature encoding data is a 1900-dimensional vector encoded by the protein sequence feature extraction model, and protein A and protein B are respectively input into the network; the one-dimensional vector with a dimension of 1900 is reshaped, and the one-dimensional vector with a dimension of 1900 * 1 is reshaped into a two-dimensional vector data with a dimension of 38 * 50; and a convolution operation is performed on the reshaped two-dimensional vector data.

[0013] As a preferred solution of the protein interaction prediction model described in the present invention, the feature extraction includes performing a one-dimensional convolution operation on the reshaped two-dimensional vector, using a convolution kernel size of 7, a step size of 2, and a number of channels of 64; the feature map after the convolution operation is processed by a maximum pooling layer, with a pooling window size of 2 and a step size of 2; the pooled feature map enters a residual block, and the residual block includes a trunk part and a jump connection part; the trunk part is composed of three convolution layers, with convolution kernel sizes of 1, 3, and 1, respectively, and the number of channels of 64, 64, and 256, respectively; the jump connection part adds the input and output; the feature map processed by the residual block is flattened into a one-dimensional vector and input into a fully connected layer; the number of fully connected layers is four, and the number of neurons is 256, 128, 32, and 8, respectively; a Dropout layer is introduced to prevent overfitting of the model.

[0014] As a preferred embodiment of the protein interaction prediction model of the present invention, the binary cross entropy loss function is expressed as:

[0015] ,

[0016] in, Represents the true classification label; represents the predicted value of the model; N is the number of samples.

[0017] As a preferred embodiment of the protein interaction prediction model of the present invention, the training by stochastic gradient descent optimizer includes using stochastic gradient descent optimizer to train the model, setting the initial learning rate to 0.005, the decay coefficient to 0.001, and the learning rate update method to be expressed as:

[0018] ,

[0019] Among them, lrate represents the learning rate, initial_lrate represents the value of the initial learning rate, decay is the decay coefficient, and iteration is the number of iterations.

[0020] Another problem to be solved by the present invention is how to provide a protein sequence feature extraction model. To solve the above technical problems, the present invention provides the following technical solutions: a protein sequence feature extraction model, comprising: an input layer receives protein sequence data and transmits it to a hidden layer; the hidden layer processes the protein sequence data and extracts features through an mLSTM model; and the output layer extracts a 1900-dimensional vector from the hidden layer as a feature representation of the protein sequence.

[0021] Another problem to be solved by the present invention is how to provide a sequence-based protein interaction prediction method. To solve the above technical problems, the present invention provides the following technical solutions: a protein sequence feature extraction model, including collecting protein interaction data sets; extracting protein sequence features through a protein sequence feature extraction model; predicting protein interaction relationships through a protein interaction prediction model; and calculating evaluation indicators to evaluate the protein interaction prediction model.

[0022] Another object of the present invention is to provide a protein interaction prediction system. To solve the above technical problems, the present invention provides the following technical solutions: a protein interaction prediction system, comprising: an acquisition module, an extraction module, a prediction module and an evaluation module; the acquisition module is used to collect protein interaction data sets; the extraction module is used to extract protein sequence features through a protein sequence feature extraction model; the prediction module is used to predict protein interaction relationships through a protein interaction prediction model; the evaluation module is used to calculate an evaluation index to evaluate the protein interaction prediction model.

[0023] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the above-mentioned protein interaction prediction method when executing the computer program.

[0024] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the protein interaction prediction method described above.

[0025] Beneficial effects of the present invention: The UniPPI model used in the protein interaction prediction method provided by the present invention is compared with different combinations of traditional feature extraction, traditional machine learning models and deep learning models. UniPPI completely surpasses other methods on the brewer's yeast dataset, with an accuracy of 95.56%, a precision of 97.61%, an F1 score of 95.46%, and a Matthews correlation coefficient of 91.2%; the accuracy on the Helicobacter pylori dataset is 88.54%, the precision is 87.48%, the F1 score is 88.88%, and the Matthews correlation coefficient is 77.13%; the accuracy on the human dataset is 98.59%, the precision is 99.3%, the F1 score is 98.6%, and the Matthews correlation coefficient is 97.19%. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0027] Figure 1 A network structure diagram of a protein interaction prediction model provided by one embodiment of the present invention.

[0028] Figure 2 A residual block structure diagram of a protein interaction prediction model provided by one embodiment of the present invention.

[0029] Figure 3 This is a structural diagram of an mLSTM model of a protein sequence feature extraction model provided by one embodiment of the present invention.

[0030] Figure 4 A protein feature extraction process diagram of a protein sequence feature extraction model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0032] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0033] Example 1

[0034] Reference Figure 1-Figure 2 , as an embodiment of the present invention, provides a protein interaction prediction model, comprising:

[0035] The protein interaction prediction model UniPPI is built using the TensorFlow framework and is divided into three parts: input layer, hidden layer, and output layer. The input layer is the starting point of the neural network and is responsible for receiving protein feature encoding data, that is, the protein feature encoding extracted using Unirep in the previous step. Subsequently, the data will undergo a series of nonlinear transformations and feature extractions through the hidden layer. The neurons in the hidden layer will perform a weighted combination of the input encoding features and introduce nonlinear transformations through activation functions, allowing the network to learn more complex patterns and rules. The final output is a classification of protein interactions, predicting whether proteins interact with each other. The UniPPI network results are as follows: Figure 1 shown.

[0036] The input data of the network are two 1900-dimensional vectors of protein A and protein B, which are encoded by Unirep respectively. The vector is first convolved, and the one-dimensional vector with a dimension of 1900 needs to be reshaped before that. The purpose of this step is to reorganize the original one-dimensional vector (dimension is 1900 * 1) into two-dimensional vector data (dimension is 38*50) suitable for network training without changing the data content. Through this reshaping operation, the structure and characteristics of the network can be better utilized to improve the expressiveness and training effect of the data in the network. The convolution operation includes convolution layer and maximum pooling. A one-dimensional convolution operation is used, the convolution kernel is set to 7, the step size is set to 2, and the number of channels is 64. Then it enters the maximum pooling layer with a pooling window of 2 and a step size of 2. Then connect a residual block. The residual block structure is as follows Figure 2 shown.

[0037] Protein A and protein B refer to two different protein sequences.

[0038] The residual block backbone consists of three convolutional layers, with kernel sizes of 1, 3, and 1, and channels of 64, 64, and 256. The jump connection combines the input with the output of the backbone through a layer of convolution, and then the data is averaged and pooled. The two proteins are connected at the output of the convolution module and flattened into a one-dimensional vector, followed by four fully connected layers with 256, 128, 32, and 8 neurons, and finally output the prediction results. A Dropout layer is added to the network to prevent overfitting of the model.

[0039] The UniPPI protein interaction prediction model uses binary cross-entropy as the loss function, which is usually used for binary classification problems, especially in binary classification tasks in neural networks. The mathematical expression of BinaryCross-Entropy is as follows:

[0040] ,

[0041] in, Represents the true classification label, with a value of 0 or 1; is the predicted value of the model, indicating the probability that the i-th sample belongs to the positive class; N is the number of samples.

[0042] The activation function used in the model is the ReLU function, which is a simple and efficient activation function with the advantages of simple calculation, sparsity, suppression of gradient disappearance, nonlinearity and stable gradient update. The calculation formula of the ReLU function is as follows:

[0043] ,

[0044] The output layer uses the Sigmoid activation function, which maps the input real number to a continuous output between 0 and 1. Its calculation formula is as follows:

[0045] ,

[0046] The learning rate is an important parameter used to adjust the parameter update speed when training the model. The Stochastic Gradient Descent (SGD) optimizer is used to train the model. The experiment sets the initial learning rate to 0.005 and the decay coefficient to 0.001. The learning rate update method is as follows:

[0047] ,

[0048] Where lrate represents the learning rate, initial_lrate represents the value of the initial learning rate, decay is the decay coefficient, and iteration is the number of iterations.

[0049] Through this learning rate decay mechanism, a larger learning rate can be used in the initial stage to quickly explore the parameter space and accelerate convergence. As training progresses, the learning rate will gradually decrease, making the model more stable and accurate when approaching the optimal solution.

[0050] The structure of the present invention and different network structures were experimented on a Saccharomyces cerevisiae dataset, and the experimental results are shown in Table 1.

[0051] Table 1 Comparison of different network structures on the Saccharomyces cerevisiae dataset

[0052] ,

[0053] Table 1 shows the five-fold cross-validation performance of different network structures on the Saccharomyces cerevisiae protein interaction dataset. The results show that the CNN+ResNet+Multilayer Perceptron (MLP) network structure performs best in all five evaluation indicators, proving that the model has excellent performance.

[0054] Example 2

[0055] Reference Figure 3-4 , as an embodiment of the present invention, provides a protein sequence feature extraction model system, comprising:

[0056] This method uses Unirep to extract protein sequence features. Unlike traditional feature extraction methods based on amino acid sequences, Unirep does not require manual feature design, but learns the representation of protein sequences through an end-to-end neural network, which is derived from the reasoning information of the neural network and has better adaptability and generalization capabilities.

[0057] First, 24 million amino acid sequences were obtained from the UniRef50 database, and then the protein sequences were trained using a multiplicative long short-term memory (mLSTM) RNNs model with 1900 hidden units. The model structure is as follows: Figure 3 shown.

[0058] The model includes input layer, hidden layer and output layer. The input layer receives a vector as input at each moment and passes this data to the next layer. The main function of the hidden layer is to extract the features of the input data through the complex structure of the neural network and convert it into a more advanced representation so as to better express and understand the inherent structure and rules of the data. The output layer is responsible for outputting the data results processed by the hidden layer.

[0059] The model input is a protein sequence, and the model is trained by predicting the probability of the next amino acid. In this process, the model continuously learns how to represent proteins, updates parameters, and finally takes a 1900-dimensional vector from the hidden layer of the model as the feature of the protein sequence, such as Figure 4 shown.

[0060] Example 3

[0061] In one embodiment of the present invention, a sequence-based protein interaction prediction method is provided, comprising:

[0062] Collect protein interaction datasets.

[0063] This method first collects three datasets for evaluating the performance of the model, namely, the Saccharomyces cerevisiae protein dataset, the human protein dataset, and the Helicobacter pylori protein dataset. The specific information of the datasets is shown in Table 2.

[0064] Table 2 Information of three evaluation datasets

[0065] ,

[0066] By conducting experimental studies on these three protein interaction datasets from different sources, the performance and applicability of the proposed protein interaction prediction model can be more comprehensively evaluated, thereby providing important experimental verification and reference for the progress in the field of protein interaction research.

[0067] Build a macro dataset.

[0068] The construction of the macro dataset is used to explore the general rules of protein interaction, provide a large dataset for building a protein interaction prediction model, and then build a protein interaction prediction platform with this prediction model as the core. The positive data in the macro dataset are obtained from three commonly used protein interaction databases: DIP (Database of Interacting Proteins), BioGRID (Biological General Repository for Interaction Datasets), and STRING (SearchTool for the Retrieval of Interacting Genes / Proteins). Since the probability of random protein pair interactions in nature is about one thousandth, a negative dataset is constructed by random protein combination, excluding confirmed positive data. The macro dataset has a total of about 1.6 million entries.

[0069] Protein sequence features are extracted through a protein sequence feature extraction model.

[0070] Predict protein-protein interaction relationships using protein-protein interaction prediction models.

[0071] Calculate evaluation indicators to evaluate protein interaction prediction models.

[0072] According to the four values ​​in the confusion matrix, we can calculate the evaluation index to evaluate the model. The confusion matrix is ​​shown in Table 3.

[0073] Table 3 Confusion matrix table for binary classification problem

[0074]

[0075] Accuracy: Accuracy is one of the most commonly used model evaluation indicators. It measures the proportion of correct predictions made by the model on all samples, that is, the ratio of the number of correctly classified samples to the total number of samples. The calculation formula is as follows:

[0076] ,

[0077] Precision: Precision measures the proportion of samples predicted as positive by the model that are actually positive, that is, the proportion of samples predicted as positive to samples that are actually positive. The calculation formula is as follows:

[0078] ,

[0079] F1 score (F-Score): F1 score is a comprehensive indicator that takes into account both precision and recall, and can more comprehensively evaluate the performance of a classifier or model. The calculation formula is as follows:

[0080] ,

[0081] Matthews Correlation Coefficient (MCC): MCC is a metric used to evaluate the performance of binary classification problems. It combines the elements of the confusion matrix and remains stable even when the categories are unbalanced. The calculation formula is as follows:

[0082] ,

[0083] AUC (Area Under the ROC Curve): AUC is an evaluation indicator commonly used in binary classification problems. It is the area between the ROC curves drawn with the True Positive Rate (TPR) as the vertical axis and the False Positive Rate (FPR) as the horizontal axis. AUC can be used to measure the model's ability to distinguish between positive and negative samples. The formulas for TPR and FTR are as follows:

[0084] .

[0085] Example 4

[0086] As one embodiment of the present invention, a sequence-based protein interaction prediction system is provided, comprising:

[0087] Acquisition module, extraction module, prediction module and evaluation module;

[0088] The acquisition module is used to collect protein interaction datasets;

[0089] The extraction module is used to extract protein sequence features through a protein sequence feature extraction model;

[0090] The prediction module is used to predict protein interaction relationships through protein interaction prediction models;

[0091] The evaluation module is used to calculate evaluation indicators to evaluate protein interaction prediction models.

[0092] Example 5

[0093] An embodiment of the present invention is different from the first four embodiments in that:

[0094] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, RandomAccess Memory), disk or optical disk, etc., which can store program codes.

[0095] The logic and / or steps represented in the block diagram or otherwise described herein, for example, may be considered as an ordered list of executable instructions for implementing logical functions, and may be embodied in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, apparatus, or device and execute instructions), or in conjunction with such instruction execution system, apparatus, or device. For purposes of this specification, "computer-readable medium" may be any device that can contain, store, communicate, propagate, or transmit a program for use by an instruction execution system, apparatus, or device, or in conjunction with such instruction execution system, apparatus, or device.

[0096] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0097] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0098] Example 6

[0099] As an embodiment of the present invention, a protein interaction prediction method is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.

[0100] The present invention selects five commonly used traditional methods for calculating features, including autocovariance AC, autocorrelation descriptor AD, joint triplet CT, pseudo amino acid composition PseAAC, and multivariate mutual information MMI, as input features of the model, and compares them with the feature extraction method of UniPPI, and evaluates their performance compared with the UniPPI method on the brewer's yeast dataset.

[0101] The present invention selected different combination methods of three features, four features and five features for experiments. There are 10 combinations of three features, 5 combinations of four features and 1 combination of all five features, totaling 16 feature combination methods. In order to accurately evaluate the performance of the model, the five-fold cross-validation method was adopted, and each feature combination method was comprehensively compared with the UniPPI method on six evaluation indicators. The experimental results are shown in Table 4.

[0102] Table 4 Evaluation performance comparison table

[0103] ,

[0104] From Table 4, we can see that there are significant differences in the model performance of different traditional feature combinations, and the more features the better is not necessarily true. Information redundancy may occur, which will have a negative effect on model training.

[0105] Secondly, the UniPPI model performed best in the five evaluation indicators, accuracy, precision, F1 score, MCC and AUC, reaching 0.9556, 0.9761, 0.9546, 0.9120 and 0.9856 respectively. This shows that the UniPPI model has high accuracy and reliability in predicting protein interactions, can effectively distinguish positive and negative samples, and has high predictive ability.

[0106] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A protein interaction prediction model, characterized in that: include: The input layer receives protein feature encoding and reshapes it into a two-dimensional vector; The hidden layer extracts features through convolution, pooling, residual blocks and fully connected layers, and introduces nonlinear transformations through the ReLU function; The output layer outputs the interaction probability through the Sigmoid function and is optimized through the binary cross entropy loss function; The protein interaction prediction model was trained by stochastic gradient descent optimizer; The protein feature encoding data is a 1900-dimensional vector encoded by Unirep after protein sequence feature extraction, and protein A and protein B are input into the network respectively; Reshape the one-dimensional vector with a dimension of 1900, and reshape the one-dimensional vector with a dimension of 1900*1 into a two-dimensional vector data with a dimension of 38*50; Perform convolution operation on the two-dimensional vector data after data reshaping; The feature extraction includes performing a one-dimensional convolution operation on the reshaped two-dimensional vector, using a convolution kernel size of 7, a step size of 2, and a channel number of 64; The feature map after the convolution operation is processed by the maximum pooling layer with a pooling window size of 2 and a stride of 2; The pooled feature map enters the residual block, which includes a trunk part and a jump connection part; The backbone consists of three convolutional layers, with convolution kernel sizes of 1, 3, and 1, and channel numbers of 64, 64, and 256, respectively; The skip connection part adds the input and the output; The feature map processed by the residual block is flattened into a one-dimensional vector and input into the fully connected layer; The number of fully connected layers is four, and the number of neurons is 256, 128, 32, and 8 respectively; The Dropout layer is introduced to prevent the model from overfitting.

2. The protein interaction prediction model according to claim 1, characterized in that: The binary cross entropy loss function is expressed as, Among them, y i Represents the true classification label; represents the predicted value of the model; N is the number of samples.

3. The protein interaction prediction model according to claim 2, characterized in that: The training by stochastic gradient descent optimizer includes using stochastic gradient descent optimizer to train the model, setting the initial learning rate to 0.005, the decay coefficient to 0.001, and the learning rate update method to be expressed as: Among them, lrate represents the learning rate, initial_lrate represents the value of the initial learning rate, decay is the decay coefficient, and iteration is the number of iterations.

4. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the protein interaction prediction model described in claim 1 are implemented.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the protein interaction prediction model described in claim 1 are implemented.

Images