Method for predicting interaction between transmembrane proteins, and device and medium

Abstract

A method for predicting an interaction between transmembrane proteins, and a computer device and a storage medium. The method comprises: inputting an amino acid sequence of a first protein and an amino acid sequence of a second protein into a transmembrane prediction neural network, and when it is identified from output results that both the first protein and the second protein are transmembrane proteins, performing screening on the output results of the transmembrane prediction neural network to obtain an extracellular amino acid fragment sequence of the first protein and an extracellular amino acid fragment sequence of the second protein; and inputting the extracellular amino acid fragment sequence of the first protein and the extracellular amino acid fragment sequence of the second protein into a trained attention mechanism model, in order to obtain an interaction score, wherein the interaction score is used for representing the likelihood of an interaction between the first protein and the second protein.

Description

Methods, equipment, and media for predicting transmembrane protein-transmembrane protein interactions.

[0001] This application claims priority to Chinese patent application No. 202510478762.0, filed on April 16, 2025, the entire contents of which are incorporated herein by reference. Technical Field

[0002] This application relates to the field of bioinformatics, and in particular to a method, computer device, and storage medium for predicting the interaction between transmembrane proteins. Background Technology

[0003] Accurate prediction of membrane protein interactions plays a crucial role in biology, drug design, and molecular biology. With the deepening of protein interaction research, predicting and understanding interactions between membrane proteins has become a current research hotspot. Research on protein interactions also has significant value in drug design. Understanding the binding and interaction patterns between proteins can help design more precise and effective drugs. Research in these areas has profound implications for biological research, drug development, and personalized medicine.

[0004] Discussions of protein-protein interactions span fields such as biology, medicine, and drug design. Protein-protein interactions are fundamental to cellular function, signal transduction, and many physiological processes within organisms. In medicine, a precise understanding of protein-protein interactions is crucial for elucidating disease mechanisms. For example, the development of certain diseases is associated with dysregulation or abnormal interactions between proteins. This research is significant not only in basic science but also has a profound impact on clinical medicine and new drug development.

[0005] The medical field has an urgent need for accurate prediction of protein-protein interactions. However, current methods primarily rely on biological experiments to test these interactions. This is not only extremely resource-intensive but also inefficient. Technical issues

[0006] Therefore, it is necessary to provide a method, computer device, and storage medium for predicting the interaction between transmembrane proteins, addressing the aforementioned technical problems. This method automates the determination of the interaction between transmembrane proteins using a computer, without requiring excessive human and material resources, and also improves the efficiency of determining the interaction between transmembrane proteins. Technical solutions

[0007] This application provides a method for predicting the interaction between transmembrane proteins, comprising: obtaining the amino acid sequences of a first protein and a second protein to be predicted; inputting the amino acid sequences of the first protein and the second protein into a transmembrane prediction neural network respectively; when the first protein and the second protein are identified as transmembrane proteins based on the output of the transmembrane prediction neural network, selecting the extracellular amino acid fragment sequence of the first protein from multiple amino acid fragment sequences of the first protein output by the transmembrane prediction neural network, and selecting the extracellular amino acid fragment sequence of the second protein from multiple amino acid fragment sequences of the second protein output by the transmembrane prediction neural network; inputting the extracellular amino acid fragment sequences of the first protein and the second protein into a trained attention mechanism model to obtain an interaction score output by the trained attention mechanism model, wherein the interaction score is used to characterize the probability of interaction between the first protein and the second protein, and the trained attention mechanism model is trained by the extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between each sample transmembrane protein.

[0008] In one embodiment, the model training of the attention mechanism model includes: obtaining the extracellular amino acid fragment sequence of a first sample single transmembrane protein and the extracellular amino acid fragment sequence of a second sample single transmembrane protein; obtaining the interaction score between the first sample single transmembrane protein and the second sample single transmembrane protein, wherein the interaction score is determined by one or more biological research methods or information mining methods; and inputting the extracellular amino acid fragment sequence of the first sample single transmembrane protein and the extracellular amino acid fragment sequence of the second sample single transmembrane protein, as well as the interaction score, into the self-attention mechanism model to perform model training on the self-attention mechanism model.

[0009] In one embodiment, when the first protein and the second protein are identified as single transmembrane proteins based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein is 1, and the number of extracellular amino acid fragment sequences of the second protein is 1; the interaction score between the first protein and the second protein is obtained by inputting 1 extracellular amino acid fragment sequence of the first protein and 1 extracellular amino acid fragment sequence of the second protein into a trained attention mechanism model.

[0010] In one embodiment, when the first protein and the second protein are identified as multi-transmembrane proteins based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein and the number of extracellular amino acid fragment sequences of the second protein are both multiple. A method for predicting the interaction between transmembrane proteins further includes: pairwise combining each extracellular amino acid fragment sequence of the first protein with each extracellular amino acid fragment sequence of the second protein to obtain a combined dataset. The combined dataset contains multiple combination elements, each of which contains one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein. The extracellular amino acid fragment sequences of the first and second proteins are input into a trained attention mechanism model to obtain an interaction score output by the trained attention mechanism model. This includes: inputting each combination element into the trained attention mechanism model to obtain the interaction score of each combination element output by the trained attention mechanism model; and outputting the highest score among the interaction scores of each combination element, where the highest score characterizes the probability of interaction between the first and second proteins.

[0011] In one embodiment, a method for predicting the interaction between transmembrane proteins further includes: identifying multiple target combination elements whose interaction scores are greater than a set threshold; obtaining the position of the extracellular amino acid fragment sequence of the first protein in the first protein and the position of the extracellular amino acid fragment sequence of the second protein in the second protein; and displaying the interaction scores of each target combination element and the positions of the extracellular amino acid fragment sequences of the first protein and the second protein in the second protein.

[0012] In one embodiment, a method for predicting the interaction between transmembrane proteins further includes: when the highest score is less than a score threshold, outputting information indicating that the interaction between the first protein and the second protein is weak.

[0013] In one embodiment, the attention mechanism model includes a word embedding layer, a position encoding layer, a multi-head attention layer, and a multi-layer perceptron layer connected in sequence; wherein, the multi-head attention layer contains an encoding layer composed of multiple connected encoders, and the self-attention representation of the extracellular amino acid fragment sequence generated by the encoding layer is input to the multi-layer perceptron layer after feature fusion.

[0014] In one embodiment, the hyperparameters of the word embedding layer, position encoding layer, multi-head attention layer, and multilayer perceptron layer in the attention mechanism model are determined by: configuring the attention mechanism model to be trained using preset hyperparameters; obtaining extracellular amino acid fragment sequences of multiple sample transmembrane proteins and interaction scores between each sample transmembrane protein; training the attention mechanism model to be trained using the extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between each sample transmembrane protein, adjusting the hyperparameters of the attention mechanism model and the length of the extracellular amino acid fragment sequences of the input sample transmembrane proteins during the model training process; and determining the hyperparameters of the word embedding layer, position encoding layer, multi-head attention layer, and multilayer perceptron layer in the attention mechanism model and the length threshold of the extracellular amino acid fragment sequences input to the attention mechanism model when the memory overflow of the device running the model training is detected.

[0015] In one embodiment, when the first protein is identified as a single-transmembrane protein and the second protein as a multi-transmembrane protein based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein is one, and the number of extracellular amino acid fragment sequences of the second protein is multiple. The method further includes: pairwise combining one extracellular amino acid fragment sequence of the first protein with each extracellular amino acid fragment sequence of the second protein to obtain a combined dataset, wherein the combined dataset contains multiple combination elements, and each combination element contains one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein; inputting the extracellular amino acid fragment sequences of the first protein and the extracellular amino acid fragment sequences of the second protein into a trained attention mechanism model to obtain an interaction score output by the trained attention mechanism model, including: inputting each combination element into the trained attention mechanism model to obtain an interaction score of each combination element output by the trained attention mechanism model; outputting the highest score among the interaction scores of each combination element, wherein the highest score is used to characterize the probability of interaction between the first protein and the second protein.

[0016] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0017] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above. Beneficial effects

[0018] This application obtains the amino acid sequences of a first protein and a second protein to be predicted. The amino acid sequences of the first and second proteins are then input into a transmembrane prediction neural network. When the output of the transmembrane prediction neural network identifies both the first and second proteins as transmembrane proteins, the extracellular amino acid fragment sequence of the first protein is selected from multiple amino acid fragment sequences output by the network, and the extracellular amino acid fragment sequence of the second protein is selected from multiple amino acid fragment sequences output by the network. The extracellular amino acid fragment sequences of the first and second proteins are then input into a trained attention mechanism model to obtain an interaction score output by the model. This interaction score characterizes the probability of interaction between the first and second proteins. The trained attention mechanism model is obtained through training on the extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between these samples. Therefore, transmembrane protein interactions can be automatically determined by computer, without requiring excessive human and material resources, and the efficiency of determining transmembrane protein interactions is improved. Attached Figure Description

[0019] Figure 1 is a flowchart illustrating a method for predicting the interaction between transmembrane proteins according to an embodiment of this application.

[0020] Figure 2 is a flowchart illustrating the model training process of the attention mechanism model in the embodiments of this application;

[0021] Figure 3 is a schematic diagram of the process of predicting the interaction score of extracellular amino acid fragment sequences using a trained attention mechanism model in an embodiment of this application.

[0022] Figure 4 is an internal structure diagram of the computer device in an embodiment of this application. Embodiments of the present invention

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] This application provides a method for predicting the interactions between transmembrane proteins. As shown in Figure 1, in one embodiment, a method for predicting the interactions between transmembrane proteins includes the following steps:

[0025] S102, Obtain the amino acid sequences of the first protein and the second protein to be predicted.

[0026] In this embodiment, the first protein and the second protein are two proteins whose interaction is to be tested. The amino acid sequences of the first protein and the second protein are extracted. The amino acid sequence of a protein refers to the information recording the order of all amino acid residues constituting a specific protein from the N-terminus (amino terminus) to the C-terminus (carboxyl terminus). This ordered amino acid sequence determines the primary structure of the protein and plays a decisive role in its three-dimensional structure and function.

[0027] S104, the amino acid sequences of the first protein and the second protein are respectively input into the transmembrane prediction neural network. When the first protein and the second protein are identified as transmembrane proteins based on the output of the transmembrane prediction neural network, the extracellular amino acid fragment sequence of the first protein is selected from the multiple amino acid fragment sequences of the first protein output by the transmembrane prediction neural network, and the extracellular amino acid fragment sequence of the second protein is selected from the multiple amino acid fragment sequences of the second protein output by the transmembrane prediction neural network.

[0028] In this embodiment, the transmembrane prediction neural network is an open-source model, such as TMHMM software like DeepTMHMM. The amino acid sequence of the first protein is input into the transmembrane prediction neural network, which outputs the transmembrane prediction result of the first protein and one or more amino acid fragment sequences based on their relative positions on the cell membrane. The amino acid sequence of the second protein is input into the transmembrane prediction neural network, which outputs the transmembrane prediction result of the second protein and one or more amino acid fragment sequences based on their relative positions on the cell membrane.

[0029] The transmembrane prediction result is either a transmembrane protein or a non-transmembrane protein. One or more amino acid fragment sequences based on their relative position on the cell membrane include extracellular amino acid fragment sequences, intracellular amino acid fragment sequences, and amino acid fragment sequences on the cell membrane.

[0030] S106, the extracellular amino acid fragment sequences of the first and second proteins are input into the trained attention mechanism model to obtain the interaction score output by the trained attention mechanism model. The interaction score characterizes the probability of interaction between the first and second proteins. The trained attention mechanism model is obtained through training on the extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between the transmembrane proteins of each sample.

[0031] In this embodiment, when step S104 determines that both the first and second proteins are transmembrane proteins, the extracellular amino acid fragment sequences of the first and second proteins are selected and input into the trained attention mechanism model. The trained attention mechanism model outputs an interaction score between the first and second proteins, and the interaction between them is determined based on the interaction score. The trained attention mechanism model is trained using extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between these proteins. Each sample transmembrane protein used to train the attention mechanism model is a single transmembrane protein. Since a single transmembrane protein contains only one extracellular amino acid fragment sequence, the interaction between the extracellular amino acid fragment sequences of two single transmembrane proteins can be macroscopically represented as an interaction between the two single transmembrane proteins. Based on this, the attention mechanism model can be trained using multiple sample single transmembrane proteins, enabling the trained attention mechanism model to identify the interaction between two extracellular amino acid fragment sequences.

[0032] In one embodiment, model training of the attention mechanism model includes: obtaining the extracellular amino acid fragment sequence of a first sample single transmembrane protein and the extracellular amino acid fragment sequence of a second sample single transmembrane protein; obtaining the interaction score between the first sample single transmembrane protein and the second sample single transmembrane protein, wherein the interaction score is determined by one or more biological research methods or protein information inference methods; and inputting the extracellular amino acid fragment sequence of the first sample single transmembrane protein and the extracellular amino acid fragment sequence of the second sample single transmembrane protein, as well as the interaction score, into the self-attention mechanism model to perform model training on the self-attention mechanism model.

[0033] In this embodiment, the interaction score between the first sample single-transmembrane protein and the second sample single-transmembrane protein is determined using one or more biological research methods or information mining methods. For example, biological research methods include gene co-expression, gene adjacency, gene fusion, and gene co-occurrence. Information mining methods include text mining. By setting standards, the interaction scores determined by various biological research methods or information mining methods are comprehensively calculated to finally obtain the interaction score between the two sample single-transmembrane proteins. For example, if the interaction score between two single-transmembrane proteins determined by gene co-expression is 0.8 and the interaction score determined by gene co-occurrence is 0.7, then the interaction score between the two single-transmembrane proteins can be determined to be 0.75.

[0034] Specifically, the training process of the trained attention mechanism model is as follows:

[0035] First, the amino acid sequences of multiple proteins and their experimentally demonstrated interaction scores are obtained from a protein database. Then, a deep neural network processes the amino acid sequences, breaking them down into different fragments based on their location on the cell membrane and outputting the corresponding sequences. Additionally, the network outputs a predicted probability of whether the protein is a transmembrane protein. For example, it outputs that the protein is a transmembrane protein, and provides the extracellular, intracellular, and intracellular amino acid fragment sequences of the protein.

[0036] During the training phase of the attention mechanism model, transmembrane proteins and their extracellular amino acid fragment sequences were selected through the above screening process. Furthermore, experimentally proven interaction scores for each transmembrane protein were collected. This yielded extracellular amino acid fragment sequences of multiple sample transmembrane proteins for the attention mechanism model, as well as interaction scores between these proteins. Subsequently, the attention mechanism model was trained using these extracellular amino acid fragment sequences and interaction scores.

[0037] In one example of this embodiment, the attention mechanism model includes a word embedding layer, a position encoding layer, a multi-head attention layer, and a multi-layer perceptron layer connected in sequence; wherein, the multi-head attention layer contains an encoding layer composed of multiple connected encoders, and the self-attention representation of the extracellular amino acid fragment sequence generated by the encoding layer is input to the multi-layer perceptron layer after feature fusion.

[0038] Specifically, the word embedding layer processes the two input amino acid fragment sequences separately to obtain a masked representation of the amino acid fragment sequence. The positional encoding layer processes the masked representation of the amino acid fragment sequence to obtain a representation of the amino acid fragment sequence containing the amino acid position order. The encoding layer, composed of multiple connected encoders in the multi-head attention layer, masks the amino acid fragment sequence again to obtain a self-attention representation of the amino acid fragment sequence. The self-attention representations of the two amino acid fragment sequences are concatenated to obtain a feature fusion representation. The multilayer perceptron layer of the attention mechanism model processes the feature fusion representation to generate a score predicting the likelihood of interactions.

[0039] It is important to emphasize that, compared with traditional attention mechanism models, the attention mechanism model in this application includes a coding layer but no decoding layer in the multi-head attention layer. That is, the self-attention representation of the extracellular amino acid fragment sequence generated by the coding layer is directly used for feature fusion, without the need for a decoding layer to perform decoding before feature fusion. Since amino acid fragment sequences do not have the complex grammatical structure of natural language, decoding is unnecessary. Direct feature fusion of the self-attention representation of the extracellular amino acid fragment sequence generated by the coding layer can be perceived and recognized by the multilayer perceptron layer, having minimal impact on the final interaction score. Based on this, setting a coding layer but not a decoding layer in the multi-head attention layer reduces the model complexity of the attention mechanism model without affecting the final prediction result. When using the attention mechanism model to predict the interaction score between the first and second proteins, it can reduce the computational load of the model and improve the prediction efficiency of the interaction score between the two proteins.

[0040] In one example of this embodiment, the determination of hyperparameters for the word embedding layer, position encoding layer, multi-head attention layer, and multilayer perceptron layer in the attention mechanism model includes: configuring the attention mechanism model to be trained using preset hyperparameters; obtaining extracellular amino acid fragment sequences of multiple sample transmembrane proteins and interaction scores between each sample transmembrane protein; training the attention mechanism model to be trained using the extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between each sample transmembrane protein, adjusting the hyperparameters of the attention mechanism model and the length of the extracellular amino acid fragment sequences of the input sample transmembrane proteins during the model training process; and determining the hyperparameters of the word embedding layer, position encoding layer, multi-head attention layer, and multilayer perceptron layer in the attention mechanism model, as well as the length threshold of the extracellular amino acid fragment sequences input to the attention mechanism model, when a memory overflow is detected in the device running the model training.

[0041] Specifically, an attention mechanism model was configured with a set of hyperparameters. The model was trained using extracellular amino acid fragment sequences of transmembrane proteins from multiple samples and the interaction scores between these proteins. During training, the hyperparameters within the attention mechanism model and the length of the input extracellular amino acid fragment sequences were adjusted. Using the GPU memory of the device running the model as a constraint, the hyperparameters and the length of the extracellular amino acid fragment sequences were repeatedly adjusted. When GPU memory overflow occurred while the prediction performance was acceptable, the hyperparameters of the word embedding layer, position encoding layer, multi-head attention layer, and multilayer perceptron layer in the attention mechanism model, as well as the length threshold of the input extracellular amino acid fragment sequences, were determined. In a specific training result, the length threshold of the input extracellular amino acid fragment sequence was determined to be 300 amino acids. The hyperparameters of the attention mechanism model in the debugging results are as follows:

[0042] The final values ​​of the hyperparameters are: word embedding layer embedding dimension 64, number of heads in multi-head attention 4, number of encoders in the encoding layer 6, and dropout rate 0.1.

[0043] The hyperparameters of the above attention mechanism model are designed for three reasons: first, proteins do not have the complex syntax of natural language; second, they can reduce GPU memory requirements; and third, they can prevent overfitting.

[0044] Based on the attention mechanism model given in the above embodiments, the specific training and application processes are described below:

[0045] As shown in Figure 2, the word embedding layer performs word embedding operations on amino acid fragment sequences. Word embedding operations can include mapping amino acid fragment sequences to a high-dimensional vector space. For example, if the amino acid fragment sequence data is a one-dimensional sequence, where each data element is an amino acid symbol, and there are 20 standard amino acid symbols, the word embedding operation can map each amino acid symbol to a high-dimensional vector of a specified dimension.

[0046] The positional encoding layer performs positional encoding on the amino acid sequence vector representation output by the word embedding layer. Positional encoding involves embedding the original positional information of the sequence into the input amino acid sequence vector representation. Positional encoding ensures that the model considers the positional information of elements when processing sequence data, which is crucial for understanding amino acids that are far apart in a protein sequence but may have important interactions. For example, in some embodiments, the maximum processable amino acid sequence length is configured to 10,000, thereby generating 10,000 unknown positional codes, which are then embedded into the vector representation of the amino acid sequence based on different input data. All modules are applied sequentially to the amino acid sequence vector representation.

[0047] Multi-head self-attention is used to learn amino acid fragment sequences, allowing the model to simultaneously focus on information from different subspaces or "heads" of the input sequence. Each attention head independently learns and weights the correlations between amino acids at different positions. Finally, the results of all attention heads are spliced ​​or merged to obtain a more comprehensive and detailed understanding of the entire amino acid fragment sequence. Therefore, encoders configured to use multi-head self-attention mechanisms can efficiently uncover deep patterns and complex interactions in amino acid sequences, contributing to the prediction of protein structure and function.

[0048] A multilayer perceptron (MLP) can consist of at least three layers: an input layer, a hidden layer, and an output layer. Because the MLP is fully connected, every unit in each layer is connected to all units in the next layer via specific weights. Except for the input layer units, each unit is essentially a neuron component using a non-linear activation function. The MLP employs a supervised learning algorithm called backpropagation for model training. It is precisely because of its multilayer structure and non-linear activation mechanism that the MLP is clearly distinguishable from a linear perceptron and exhibits effective discrimination capabilities on non-linearly separable datasets.

[0049] If we assume that all neurons in a Multilayer Perceptron (MLP) use only linear activation functions—that is, the weighted input is mapped to the output of each neuron through a linear function—linear algebra principles state that, regardless of the number of layers, such a structure is equivalent to a simple two-layer input-output model. However, in actual MLPs, some neurons use nonlinear activation functions, designed to mimic the behavior of biological neurons in generating action potentials or adjusting firing frequencies. During training, the model's learning process is achieved by comparing the error between the actual output and the expected output for each training data point, and adjusting the connection weights between neurons accordingly. This is an example of supervised learning, where the learning mechanism is based on the backpropagation principle, which is actually an extension of the minimum mean square error algorithm in linear perceptrons. In this application, as shown in Figure 2, the output of the MLP is the predicted score of the interaction score between two amino acid fragment sequences. During model training, the loss of its output is adjusted using the true label. The true label refers to a probability value between 0 and 1.

[0050] The specific model training process is as follows:

[0051] The training set of samples was compiled from amino acid sequences of proteins collected from open-source databases, along with protein-protein interaction scores. The collected data was filtered using a deep neural network to select single-transmembrane extracellular amino acid fragment sequences and their associated protein-protein interaction scores, containing amino acid sequences of approximately 1100 proteins and approximately 49,000 interaction scores. Multi-transmembrane extracellular amino acid fragment sequences could also be selected for training during other stages of model training. The model was trained using the type of extracellular amino acid fragment sequence used for training, and the resulting model was then used to test the interaction scores of that type of extracellular amino acid fragment sequence. For the training set, 80% of the data was used for model training, 10% was used as a validation set to verify the model's effectiveness during training, and the remaining 10% was used as a test set to evaluate the final model's performance, as detailed in the table below.

[0052] MSE RMSER2MAEEVSMAPEMax Error0.01450.1210.6700.0920.68134.187%0.72

[0053] MSE (Mean Square Error) is the average of the sum of squared errors;

[0054] RMSE (Root Mean Square Error) is the square root of MSE;

[0055] R² (ratio of determination) – the larger the value, the better the fit.

[0056] MAE (Mean Absolute Error) is the average of the absolute values ​​of the error.

[0057] EVS (Explained Variance Score);

[0058] MAPE (Mean Absolute Percentage Error).

[0059] In the application phase of the trained attention mechanism model, the amino acid sequences of each pair of proteins are used as inputs and passed through the transmembrane prediction neural network. When the output prediction result is that both are transmembrane proteins, the pairwise combination datasets of the two fragments are constructed by using at least two extracellular amino acid fragment sequences output by the transmembrane prediction neural network. Each extracellular amino acid fragment sequence is then input into the trained attention mechanism model one by one.

[0060] As shown in Figure 3, the trained attention mechanism model receives at least two extracellular amino acid fragment sequences and generates a predicted score for the interaction between the two extracellular amino acid fragment sequences. The predicted interaction score is a probability value between 0 and 1.

[0061] For the application stage of the trained attention mechanism model, the following are some examples:

[0062] In one embodiment, when the first protein and the second protein are identified as single transmembrane proteins based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein is 1, and the number of extracellular amino acid fragment sequences of the second protein is 1; the interaction score between the first protein and the second protein is obtained by inputting 1 extracellular amino acid fragment sequence of the first protein and 1 extracellular amino acid fragment sequence of the second protein into a trained attention mechanism model.

[0063] In this embodiment, both the first and second proteins are single-transmembrane proteins. In this case, the output of the transmembrane prediction neural network contains only one extracellular amino acid fragment sequence. The interaction score between the first and second proteins is directly determined by one extracellular amino acid fragment sequence of the first and second proteins. Therefore, after inputting one extracellular amino acid fragment sequence of the first and second proteins into the trained attention mechanism model, the interaction between the first and second proteins is determined directly by referring to the interaction score output by the trained attention mechanism model. Specifically, the interaction between the first and second proteins can be directly evaluated using the output interaction score.

[0064] In another embodiment, when the first protein and the second protein are identified as multi-transmembrane proteins based on the output of the transmembrane prediction neural network, the first protein has multiple extracellular amino acid fragment sequences, and the second protein has multiple extracellular amino acid fragment sequences. At this point, before the step of inputting the extracellular amino acid fragment sequences of the first protein and the second protein into the trained attention mechanism model to obtain the interaction score output by the trained attention mechanism model, the method further includes: pairwise combining each extracellular amino acid fragment sequence of the first protein with each extracellular amino acid fragment sequence of the second protein to obtain a combined dataset. The combined dataset contains multiple combination elements, and each combination element contains one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein. Inputting the extracellular amino acid fragment sequences of the first protein and the second protein into the trained attention mechanism model to obtain the interaction score output by the trained attention mechanism model includes: inputting each combination element into the trained attention mechanism model to obtain the interaction score of each combination element output by the trained attention mechanism model; and outputting the highest score among the interaction scores of each combination element, where the highest score is used to characterize the probability of interaction between the first protein and the second protein.

[0065] In this embodiment, both the first and second proteins are multi-transmembrane proteins. In this case, the output of the transmembrane prediction neural network contains multiple extracellular amino acid fragment sequences for both the first and second proteins. These multiple extracellular amino acid fragment sequences of the first and second proteins are paired to obtain a combined dataset. Each pair of these combined extracellular amino acid fragment sequences is then input into a trained attention mechanism model to obtain an interaction score for each pair. After obtaining the interaction scores for all pairwise combinations, the highest score is used to predict the interaction between the first and second proteins. Specifically, the interaction score output by the attention mechanism model each time characterizes the probability of interaction between the two extracellular amino acid fragment sequences of the two proteins in that input. For multi-transmembrane proteins, this process is repeated until all extracellular amino acid fragment sequences of the first and second proteins have their interaction scores obtained, ultimately outputting a list of interaction score results for the first and second proteins. The highest interaction score in the interaction score list characterizes the probability of interaction between the first and second proteins.

[0066] In another embodiment, when the first protein is identified as a single transmembrane protein and the second protein as a multi-transmembrane protein based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein is 1, and the number of extracellular amino acid fragment sequences of the second protein is multiple. At this point, before the step of inputting the extracellular amino acid fragment sequences of the first protein and the second protein into the trained attention mechanism model to obtain the interaction score output by the trained attention mechanism model, the method further includes: pairwise combining one extracellular amino acid fragment sequence of the first protein with each extracellular amino acid fragment sequence of the second protein to obtain a combined dataset. The combined dataset contains multiple combination elements, and each combination element contains one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein. Inputting the extracellular amino acid fragment sequences of the first protein and the second protein into the trained attention mechanism model to obtain the interaction score output by the trained attention mechanism model includes: inputting each combination element into the trained attention mechanism model to obtain the interaction score of each combination element output by the trained attention mechanism model; and outputting the highest score among the interaction scores of each combination element, where the highest score is used to characterize the probability of interaction between the first protein and the second protein.

[0067] In this embodiment, the first protein is a single-transmembrane protein and the second protein is a multi-transmembrane protein. In this case, the output of the transmembrane prediction neural network contains one extracellular amino acid fragment sequence for the first protein and multiple extracellular amino acid fragment sequences for the second protein. The multiple extracellular amino acid fragment sequences of the first and second proteins are paired to obtain a combined dataset. Each pair of combined extracellular amino acid fragment sequences is then input into a trained attention mechanism model to obtain an interaction score for each pair. After obtaining the interaction scores for all pairwise combined extracellular amino acid fragment sequences, the highest score is used to predict the interaction between the first and second proteins. Specifically, the interaction score output by the attention mechanism model each time is used to characterize the probability of interaction between the two extracellular amino acid fragment sequences of the two proteins in that input. For single-transmembrane and multi-transmembrane proteins, this process is repeated until all extracellular amino acid fragment sequences of the first and second proteins have their interaction scores obtained, and finally, a list of interaction score results for the first and second proteins is output. The highest interaction score in the interaction score result list is used to characterize the probability of interaction between the first and second proteins.

[0068] In one example of this embodiment, after the step of determining the interaction between the first protein and the second protein based on the highest score in the interaction scores of each combination element, the method further includes: determining a plurality of target combination elements whose interaction scores are greater than a set threshold; obtaining the position of the extracellular amino acid fragment sequence of the first protein in the first protein and the position of the extracellular amino acid fragment sequence of the second protein in the second protein; and displaying the interaction scores of each target combination element and the positions of the extracellular amino acid fragment sequences of the first protein and the second protein in the second protein.

[0069] In another example, after determining the interaction between the first protein and the second protein based on the highest score among the interaction scores of each combination element, the method further includes: when the highest score is less than a score threshold, outputting information indicating that the interaction between the first protein and the second protein is weak.

[0070] Specifically, a threshold for the interaction score is preset. This threshold is used to indicate that the interaction strength between two extracellular amino acid fragment sequences reaches a certain level. When the interaction score of any combination element exceeds the preset threshold, the information of that combination element is output. The information of the combination element includes its interaction score, the position of the extracellular amino acid fragment sequence of the first protein within the first protein, and the position of the extracellular amino acid fragment sequence of the second protein within the second protein. For example, as shown in the table below:

[0071]

[0072] Therefore, this technology can provide researchers with visualizations of the interactions between multiple extracellular amino acid fragment sequences in the first and second proteins, offering analytical assistance for protein interactions. For example, if the interaction of multiple extracellular amino acid fragment sequences exceeds a set threshold, it indicates that there are multiple amino acid fragments with significant interactions within the extracellular amino acid sequences of the two proteins. Researchers can use this information to conduct subsequent biological experiments on the two proteins, such as developing novel drugs targeting these proteins. Alternatively, if the first and second proteins have numerous transmembrane regions, and the number of extracellular amino acid fragment sequences interacting with a threshold is small, such as only one or two interacting segments, researchers can use this information to conduct subsequent biological experiments on the two proteins.

[0073] Alternatively, a scoring threshold can be set for the output score of the attention mechanism model. Using this threshold as a reference, if the highest score is lower than the threshold, it indicates a weak interaction between the first and second proteins. In this case, information representing the weak interaction between the first and second proteins should be output to directly visually remind researchers of the weak interaction. As analyzed above, the score output by the attention mechanism model is a value between 0 and 1, and the scoring threshold can be selected as 0.5.

[0074] In the medical field, an accurate understanding of protein-protein interactions is crucial for elucidating disease mechanisms. For example, the development of certain diseases is associated with dysregulation or abnormal interactions between proteins. These studies are not only significant in basic science but also have a profound impact on clinical medicine and new drug development. Therefore, the pharmaceutical field urgently needs to accurately predict protein-protein interactions. This application presents a method for predicting transmembrane protein-transmembrane protein interactions, which can predict protein interactions using computer science, laying a technological foundation for protein-interaction-based drug design and development, and improving the efficiency of drug design.

[0075] It should be understood that although the steps in the flowchart are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0076] In one embodiment, a computer device is provided, which can be a server running a transmembrane protein-to-transmembrane protein interaction prediction method according to this application. Its internal structure is shown in Figure 4. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface receives the amino acid sequences of a first protein and a second protein to be predicted via a network. When the computer program is executed by the processor, it implements the aforementioned transmembrane protein-to-transmembrane protein interaction prediction method.

[0077] Those skilled in the art will understand that the structure shown in Figure 4 is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device on which the present application is intended to be applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0078] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it performs the following steps: acquiring the amino acid sequences of a first protein and a second protein to be predicted; inputting the amino acid sequences of the first protein and the second protein into a transmembrane prediction neural network; when the first protein and the second protein are identified as transmembrane proteins based on the output of the transmembrane prediction neural network, selecting extracellular amino acid fragment sequences of the first protein from multiple amino acid fragment sequences of the first protein output by the transmembrane prediction neural network, and selecting extracellular amino acid fragment sequences of the second protein from multiple amino acid fragment sequences of the second protein output by the transmembrane prediction neural network; inputting the extracellular amino acid fragment sequences of the first protein and the second protein into a trained attention mechanism model to obtain an interaction score output by the trained attention mechanism model, wherein the interaction score is used to characterize the probability of interaction between the first protein and the second protein, and the trained attention mechanism model is trained using extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between each sample transmembrane protein.

[0079] In one embodiment, the model training of the attention mechanism model includes: obtaining the extracellular amino acid fragment sequence of a first sample single transmembrane protein and the extracellular amino acid fragment sequence of a second sample single transmembrane protein; obtaining the interaction score between the first sample single transmembrane protein and the second sample single transmembrane protein, wherein the interaction score is determined by one or more biological research methods or information mining methods; and inputting the extracellular amino acid fragment sequence of the first sample single transmembrane protein and the extracellular amino acid fragment sequence of the second sample single transmembrane protein, as well as the interaction score, into the self-attention mechanism model to perform model training on the self-attention mechanism model.

[0080] In one embodiment, when the first protein and the second protein are identified as single transmembrane proteins based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein is 1, and the number of extracellular amino acid fragment sequences of the second protein is 1; the interaction score between the first protein and the second protein is obtained by inputting 1 extracellular amino acid fragment sequence of the first protein and 1 extracellular amino acid fragment sequence of the second protein into a trained attention mechanism model.

[0081] In one embodiment, when the first protein and the second protein are identified as multi-transmembrane proteins based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein and the second protein are both multiple. When the processor executes the computer program, it further implements the following steps: pairwise combining each extracellular amino acid fragment sequence of the first protein with each extracellular amino acid fragment sequence of the second protein to obtain a combined dataset. The combined dataset contains multiple combination elements, each containing one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein. When the processor executes the computer program to input the extracellular amino acid fragment sequences of the first and second proteins into a trained attention mechanism model to obtain the interaction score output by the trained attention mechanism model, it specifically implements the following steps: inputting each combination element into the trained attention mechanism model to obtain the interaction score of each combination element output by the trained attention mechanism model; outputting the highest score among the interaction scores of each combination element, where the highest score characterizes the probability of interaction between the first and second proteins.

[0082] In one embodiment, when the processor executes the computer program, it further performs the following steps: determining multiple target combination elements whose interaction scores are greater than a set threshold; obtaining the position of the extracellular amino acid fragment sequence of the first protein in the first protein and the position of the extracellular amino acid fragment sequence of the second protein in the second protein; displaying the interaction scores of each target combination element and the positions of the extracellular amino acid fragment sequences of the first protein and the second protein in the second protein.

[0083] In one embodiment, when the processor executes the computer program, it also performs the following steps: when the highest score is less than a score threshold, it outputs information indicating that the interaction between the first protein and the second protein is weak.

[0084] In one embodiment, the attention mechanism model includes a word embedding layer, a position encoding layer, a multi-head attention layer, and a multi-layer perceptron layer connected in sequence; wherein, the multi-head attention layer contains an encoding layer composed of multiple connected encoders, and the self-attention representation of the extracellular amino acid fragment sequence generated by the encoding layer is input to the multi-layer perceptron layer after feature fusion.

[0085] In one embodiment, the hyperparameters of the word embedding layer, position encoding layer, multi-head attention layer, and multilayer perceptron layer in the attention mechanism model are determined by: configuring the attention mechanism model to be trained using preset hyperparameters; obtaining extracellular amino acid fragment sequences of multiple sample transmembrane proteins and interaction scores between each sample transmembrane protein; training the attention mechanism model to be trained using the extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between each sample transmembrane protein, adjusting the hyperparameters of the attention mechanism model and the length of the extracellular amino acid fragment sequences of the input sample transmembrane proteins during the model training process; and determining the hyperparameters of the word embedding layer, position encoding layer, multi-head attention layer, and multilayer perceptron layer in the attention mechanism model and the length threshold of the extracellular amino acid fragment sequences input to the attention mechanism model when the memory overflow of the device running the model training is detected.

[0086] In one embodiment, when the first protein is identified as a single-transmembrane protein and the second protein as a multi-transmembrane protein based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein is one, and the number of extracellular amino acid fragment sequences of the second protein is multiple. When the processor executes the computer program, it further implements the following steps: pairwise combining one extracellular amino acid fragment sequence of the first protein with each extracellular amino acid fragment sequence of the second protein to obtain a combined dataset. The combined dataset contains multiple combination elements, each of which contains one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein. When the processor executes the computer program to input the extracellular amino acid fragment sequences of the first and second proteins into a trained attention mechanism model to obtain the interaction score output by the trained attention mechanism model, it specifically implements the following steps: inputting each combination element into the trained attention mechanism model to obtain the interaction score of each combination element output by the trained attention mechanism model, and outputting the highest score among the interaction scores of each combination element. The highest score is used to characterize the probability of interaction between the first and second proteins.

[0087] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a processor, the computer program performs the following steps: obtaining the amino acid sequence of a first protein to be predicted and the amino acid sequence of a second protein; inputting the amino acid sequences of the first protein and the second protein into a transmembrane prediction neural network, respectively; when the first protein and the second protein are identified as transmembrane proteins based on the output of the transmembrane prediction neural network, selecting the extracellular amino acid fragment sequence of the first protein from multiple amino acid fragment sequences of the first protein output by the transmembrane prediction neural network, and selecting the extracellular amino acid fragment sequence of the second protein from multiple amino acid fragment sequences of the second protein output by the transmembrane prediction neural network; inputting the extracellular amino acid fragment sequence of the first protein and the extracellular amino acid fragment sequence of the second protein into a trained attention mechanism model to obtain an interaction score output by the trained attention mechanism model, wherein the interaction score is used to characterize the probability of interaction between the first protein and the second protein, and the trained attention mechanism model is trained by the extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between each sample transmembrane protein.

[0088] In one embodiment, the model training of the attention mechanism model includes: obtaining the extracellular amino acid fragment sequence of a first sample single transmembrane protein and the extracellular amino acid fragment sequence of a second sample single transmembrane protein; obtaining the interaction score between the first sample single transmembrane protein and the second sample single transmembrane protein, wherein the interaction score is determined by one or more biological research methods or information mining methods; and inputting the extracellular amino acid fragment sequence of the first sample single transmembrane protein and the extracellular amino acid fragment sequence of the second sample single transmembrane protein, as well as the interaction score, into the self-attention mechanism model to perform model training on the self-attention mechanism model.

[0089] In one embodiment, when the first protein and the second protein are identified as single transmembrane proteins based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein is 1, and the number of extracellular amino acid fragment sequences of the second protein is 1; the interaction score between the first protein and the second protein is obtained by inputting 1 extracellular amino acid fragment sequence of the first protein and 1 extracellular amino acid fragment sequence of the second protein into a trained attention mechanism model.

[0090] In one embodiment, when the first protein and the second protein are both identified as multi-transmembrane proteins based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein and the number of extracellular amino acid fragment sequences of the second protein are both multiple. When the computer program is executed by the processor, it further performs the following steps: pairwise combining each extracellular amino acid fragment sequence of the first protein with each extracellular amino acid fragment sequence of the second protein to obtain a combined dataset. The combined dataset contains multiple combination elements, and each combination element contains one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein. When the program is executed by the processor to input the extracellular amino acid fragment sequences of the first and second proteins into a trained attention mechanism model, and to obtain the interaction score output by the trained attention mechanism model, the specific steps for determining the interaction between the first and second proteins based on the interaction score output by the trained attention mechanism model are as follows: input each combination element into the trained attention mechanism model to obtain the interaction score of each combination element output by the trained attention mechanism model; output the highest score among the interaction scores of each combination element, and the highest score is used to characterize the probability of the interaction between the first and second proteins.

[0091] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: identifying multiple target combination elements whose interaction scores are greater than a set threshold; obtaining the position of the extracellular amino acid fragment sequence of the first protein in the first protein and the position of the extracellular amino acid fragment sequence of the second protein in the second protein; displaying the interaction scores of each target combination element and the positions of the extracellular amino acid fragment sequences of the first protein and the second protein in the second protein.

[0092] In one embodiment, when the computer program is executed by the processor, it further performs the following steps: when the highest score is less than a score threshold, outputting information indicating that the interaction between the first protein and the second protein is weak.

[0093] In one embodiment, the attention mechanism model includes a word embedding layer, a position encoding layer, a multi-head attention layer, and a multi-layer perceptron layer connected in sequence; wherein, the multi-head attention layer contains an encoding layer composed of multiple connected encoders, and the self-attention representation of the extracellular amino acid fragment sequence generated by the encoding layer is input to the multi-layer perceptron layer after feature fusion.

[0094] In one embodiment, the hyperparameters of the word embedding layer, position encoding layer, multi-head attention layer, and multilayer perceptron layer in the attention mechanism model are determined by: configuring the attention mechanism model to be trained using preset hyperparameters; obtaining extracellular amino acid fragment sequences of multiple sample transmembrane proteins and interaction scores between each sample transmembrane protein; training the attention mechanism model to be trained using the extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between each sample transmembrane protein, adjusting the hyperparameters of the attention mechanism model and the length of the extracellular amino acid fragment sequences of the input sample transmembrane proteins during the model training process; and determining the hyperparameters of the word embedding layer, position encoding layer, multi-head attention layer, and multilayer perceptron layer in the attention mechanism model and the length threshold of the extracellular amino acid fragment sequences input to the attention mechanism model when the memory overflow of the device running the model training is detected.

[0095] In one embodiment, when the first protein is identified as a single-transmembrane protein and the second protein as a multi-transmembrane protein based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein is one, and the number of extracellular amino acid fragment sequences of the second protein is multiple. When the processor executes the computer program, it further implements the following steps: pairwise combining one extracellular amino acid fragment sequence of the first protein with each extracellular amino acid fragment sequence of the second protein to obtain a combined dataset. The combined dataset contains multiple combination elements, each of which contains one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein. When the processor executes the computer program to input the extracellular amino acid fragment sequences of the first and second proteins into a trained attention mechanism model to obtain the interaction score output by the trained attention mechanism model, it specifically implements the following steps: inputting each combination element into the trained attention mechanism model to obtain the interaction score of each combination element output by the trained attention mechanism model, and outputting the highest score among the interaction scores of each combination element. The highest score is used to characterize the probability of interaction between the first and second proteins.

[0096] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0097] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0098] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for predicting the interaction of a transmembrane protein with a transmembrane protein, wherein, include: Obtain the amino acid sequences of the first protein and the second protein to be predicted; The amino acid sequences of the first protein and the second protein are respectively input into a transmembrane prediction neural network. When the first protein and the second protein are identified as transmembrane proteins based on the output of the transmembrane prediction neural network, the extracellular amino acid fragment sequence of the first protein is selected from multiple amino acid fragment sequences of the first protein output by the transmembrane prediction neural network, and the extracellular amino acid fragment sequence of the second protein is selected from multiple amino acid fragment sequences of the second protein output by the transmembrane prediction neural network. The extracellular amino acid fragment sequences of the first protein and the second protein are input into a trained attention mechanism model to obtain an interaction score output by the trained attention mechanism model. The interaction score is used to characterize the probability of interaction between the first protein and the second protein. The trained attention mechanism model is trained using extracellular amino acid fragment sequences of multiple sample transmembrane proteins and the interaction scores between the transmembrane proteins of each sample.

2. The method of claim 1, wherein, The training of the attention mechanism model includes: Obtain the extracellular amino acid fragment sequence of the first sample of a single transmembrane protein and the extracellular amino acid fragment sequence of the second sample of a single transmembrane protein; The interaction scores of the first sample single transmembrane protein and the second sample single transmembrane protein are obtained, and the interaction scores of the first sample single transmembrane protein and the second sample single transmembrane protein are determined by one or more biological research methods or information mining methods. The extracellular amino acid fragment sequences of the first sample single transmembrane protein and the second sample single transmembrane protein, as well as the interaction scores of the first sample single transmembrane protein and the second sample single transmembrane protein, are input into the self-attention mechanism model to perform model training on the self-attention mechanism model.

3. The method of claim 1, wherein, When the first protein and the second protein are identified as single transmembrane proteins based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein is 1, and the number of extracellular amino acid fragment sequences of the second protein is 1. The interaction score between the first protein and the second protein is obtained by inputting one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein into a trained attention mechanism model.

4. The method of claim 1, wherein, When the first protein and the second protein are identified as multi-transmembrane proteins based on the output of the transmembrane prediction neural network, the first protein has multiple extracellular amino acid fragment sequences, and the second protein has multiple extracellular amino acid fragment sequences; the method further includes: Each extracellular amino acid fragment sequence of the first protein and each extracellular amino acid fragment sequence of the second protein are combined in pairs to obtain a combined dataset. The combined dataset contains multiple combination elements, and each combination element contains one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein. The extracellular amino acid fragment sequences of the first protein and the second protein are input into a trained attention mechanism model to obtain the interaction score output by the trained attention mechanism model, including: Each combination element is input into a trained attention mechanism model to obtain the interaction score of each combination element output by the trained attention mechanism model. Output the highest score among the interaction scores of each combination element, which is used to characterize the likelihood of interaction between the first protein and the second protein.

5. The method of claim 4, wherein, The method further includes: Identify multiple target combination elements whose interaction scores are greater than a set threshold in the interaction score of each combination element; Obtain the position of the extracellular amino acid fragment sequence of the first protein in the first protein and the position of the extracellular amino acid fragment sequence of the second protein in the second protein in each combination element; The interaction scores of each target combination element are displayed, as well as the positions of the extracellular amino acid fragment sequences of the first protein and the second protein in the first protein.

6. The method of claim 5, wherein, The method further includes: When the highest score is less than the score threshold, information indicating that the interaction between the first protein and the second protein is weak is output.

7. The method of claim 1, wherein, The attention mechanism model includes a word embedding layer, a position encoding layer, a multi-head attention layer, and a multi-layer perceptron layer connected in sequence. The multi-head attention layer includes a coding layer composed of multiple connected encoders, and the self-attention representation of the extracellular amino acid fragment sequence generated by the coding layer is input to the multilayer perceptron layer after feature fusion.

8. The method of claim 7, wherein, The hyperparameters of the word embedding layer, positional encoding layer, multi-head attention layer, and multi-layer perceptron layer in the attention mechanism model are determined in the following ways: The attention mechanism model to be trained is configured by pre-defined hyperparameters; Obtain extracellular amino acid fragment sequences of transmembrane proteins from multiple samples and the interaction scores between transmembrane proteins from each sample; The attention mechanism model to be trained is obtained by scoring the extracellular amino acid fragment sequences of transmembrane proteins from multiple samples and the interactions between transmembrane proteins from each sample. During the model training process, the hyperparameters of the attention mechanism model and the length of the extracellular amino acid fragment sequences of the input transmembrane proteins are adjusted. When a memory overflow is detected in the device running the model training, the hyperparameters of the word embedding layer, position encoding layer, multi-head attention layer, and multi-layer perceptron layer in the attention mechanism model, as well as the length threshold of the extracellular amino acid fragment sequence input to the attention mechanism model, are determined.

9. The method according to claim 1, wherein, When the first protein is identified as a single transmembrane protein and the second protein as a multi-transmembrane protein based on the output of the transmembrane prediction neural network, the number of extracellular amino acid fragment sequences of the first protein is 1, and the number of extracellular amino acid fragment sequences of the second protein is multiple. The method further includes: The extracellular amino acid fragment sequence of the first protein is combined with each extracellular amino acid fragment sequence of the second protein in pairs to obtain a combined dataset. The combined dataset contains multiple combination elements, and each combination element contains one extracellular amino acid fragment sequence of the first protein and one extracellular amino acid fragment sequence of the second protein. The extracellular amino acid fragment sequences of the first and second proteins are input into a trained attention mechanism model to obtain the interaction score output by the trained attention mechanism model, including: Each combination element is input into a trained attention mechanism model to obtain the interaction score of each combination element output by the trained attention mechanism model. Output the highest score among the interaction scores of each combination element, which is used to characterize the likelihood of interaction between the first protein and the second protein.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

11. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images