Method and system for recognizing type iv secretion effect protein based on multi-modal information

Abstract

The application discloses a type IV secretory effector protein recognition method and system based on multi-modal information, and belongs to the technical field of bioinformatics and secretory protein recognition. The amino acid residue features of the protein are extracted by using a protein language model, and a spatial adjacency graph of the protein is constructed by using a three-dimensional structure predicted by a protein structure prediction model; then, sequence features and structure features of the protein are extracted by a deep sequence module and a hierarchical graph module respectively; meanwhile, a contrast learning module is introduced to realize cross-modal alignment of the same protein in a latent space; the sequence features and the structure features are bidirectionally interacted by using a cross attention module to obtain joint features, and finally, a classification result is output after a nonlinear transformation of a gate linear unit, and is used for prediction of the type IV secretory effector protein. The application realizes efficient cooperation of sequence and structure double modal information, and significantly improves the accuracy of type IV secretory effector protein recognition.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics and secretory protein recognition technology, and specifically relates to a method and system for recognizing type IV secretory effector proteins based on multimodal information. Background Technology

[0002] Type IV secretory effector proteins are key functional proteins that pathogens inject into host cells via the type IV secretion system, playing a crucial role in regulating host immunity, promoting bacterial invasion, and survival. Once inside host cells, type IV secretory effector proteins can interfere with signaling pathways, suppress immune responses, disrupt organelle function, and affect gene expression, thereby hijacking host cellular resources to support bacterial colonization and spread. Some pathogens, such as Helicobacter pylori, rely on type IV secretory effector proteins to invade host cells and cause disease. With the rapid growth of genomic data, accurate identification of type IV secretory effector proteins is crucial for understanding pathogen-host interaction mechanisms and developing anti-infection strategies. However, traditional experimental identification methods are time-consuming, costly, and have low throughput, making them unsuitable for large-scale screening. Therefore, computational prediction methods based on machine learning and deep learning have attracted widespread attention due to their efficiency and scalability.

[0003] In recent years, several tools for identifying type IV secretory effector proteins have been proposed. For example, DeepT3_4 (a joint classification model for type III and IV secretory effector proteins) combines recurrent neural networks and deep neural networks, utilizing amino acid character-level encoding and evolutionary conservation features to model sequence context dependencies and achieve joint classification of type III and IV effector proteins. T4Seeker (a hybrid model for identifying type IV secretory effector proteins) integrates multimodal features of distance residual (DR), evolutionary scale modeling (ESM), and long short-term memory networks (LSTM), which are significantly complementary. Using type III and IV effector proteins as negative controls, it constructs a highly specific and generalizable model for identifying type IV secretory effector proteins. Although the above methods have made some progress in identifying type IV effector proteins, they still generally rely on sequence information or shallow feature fusion and lack effective modeling of the three-dimensional spatial structure of proteins.

[0004] Given that protein function is highly dependent on its three-dimensional conformation, structural information plays a crucial role in accurately predicting protein function. With the rapid development of artificial intelligence and pre-trained protein language models, the emergence of high-precision structure prediction tools such as ESM-Fold and AlphaFold2 has made it possible to obtain protein three-dimensional structures on a large scale and at low cost, providing strong technical support for in-depth analysis of the mechanisms of action of pathogen effector proteins. In recent years, research teams both domestically and internationally have begun to combine structure prediction results with graph neural networks, effectively capturing long-range interactions within proteins by constructing spatial adjacency graphs or distance matrices between residues, significantly improving the accuracy of functional annotation.

[0005] However, most current methods for identifying type IV effector proteins still rely primarily on amino acid sequences or evolutionary information, failing to fully utilize high-confidence three-dimensional structural data. In particular, they lack effective modeling of cross-modal associations between sequence semantic features and spatial geometric topology, limiting the model's generalization ability and biological interpretability in low homology scenarios. Summary of the Invention

[0006] To overcome the limitations of existing methods for identifying type IV secretory effector proteins in terms of multimodal information fusion and feature expression capabilities, this invention provides a method and system for identifying type IV secretory effector proteins based on multimodal information. This invention acquires the amino acid residue features and spatial adjacency graph of proteins through a protein language model and a protein structure prediction model, respectively, constructs a learning framework for sequence and structure multimodal fusion, aligns modal representations using a contrastive learning module, and achieves deep fusion of sequence and structural features using a cross-attention module. Finally, the classification result is output after nonlinear transformation by a gated linear unit, significantly improving the accuracy and generalization ability of type IV secretory effector protein identification.

[0007] The present invention is achieved by the following technical solution.

[0008] A method for identifying type IV secretory effector proteins based on multimodal information includes the following steps:

[0009] Step S1: Input the amino acid sequence of the protein into the protein language model to obtain amino acid residue features; input the amino acid sequence of the protein into the protein structure prediction model to obtain a spatial adjacency graph.

[0010] Step S2: Input the amino acid residue features into the deep sequence module to extract the protein sequence features; input the amino acid residue features as node features and the spatial adjacency graph into the hierarchical graph module to obtain the protein structural features;

[0011] Step S3: Perform global average pooling on the sequence features and structural features respectively to obtain global sequence feature vector and global structural feature vector, and use the contrastive learning module to ensure that the sequence and structural representations of the same protein in the latent space are consistent, thereby achieving cross-modal alignment;

[0012] Step S4: Use the cross-attention module to perform bidirectional interaction between sequence features and structural features to obtain joint features;

[0013] Step S5: Perform global average pooling on the joint features and input them into a gated linear unit to determine whether they are type IV secretory effector proteins.

[0014] Furthermore, the protein language model adopts the esm2_t33_650M_UR50D version of the ESM-2 model; the protein structure prediction model is the ESM-Fold model, used to predict the three-dimensional structure of proteins from their amino acid sequences, and based on... - A spatial adjacency graph is constructed based on the criterion that the carbon atom spacing is less than 10 Å.

[0015] Furthermore, the deep sequence module consists of four identical sub-layers stacked together. Each sub-layer is connected sequentially by multi-head attention, layer normalization, feedforward network, and layer normalization; residual connections are introduced after each layer normalization.

[0016] Furthermore, the hierarchical graph module consists of two stacked graph convolutional networks, with each graph convolutional network followed by a residual connection, wherein the first layer graph convolutional network receives amino acid residue features and spatial adjacency graphs.

[0017] Furthermore, the contrastive learning module employs the InfoNCE loss function, treating sequence features and structural features of the same protein as positive sample pairs, and feature combinations of other samples as negative sample pairs, thereby aligning sequence features and structural features across modalities in the latent space.

[0018] Furthermore, the cross-attention module receives sequence features as queries and structural features as keys and values, enabling bidirectional interaction between structural and sequence information.

[0019] Furthermore, the gated linear unit includes a value branch linear transformation, a gated branch linear transformation, a Gaussian error linear unit activation function, a Sigmoid activation function, element-wise multiplication, and a linear layer, forming a nonlinear classification structure with a gating mechanism. The input features of the gated linear unit are divided into two branches. One branch is transformed by the value branch linear transformation to obtain a value vector, and the other branch is transformed by the gated branch linear transformation and activated by the Sigmoid function to obtain gate weights. After the value vector and the gate weights are multiplied element-wise, they are passed through the Gaussian error linear unit activation function and the linear layer, and the final output is a binary classification probability score.

[0020] This invention also provides a type IV secretory effector protein recognition system based on multimodal information, comprising:

[0021] The feature encoding module is used to receive the protein amino acid sequence, extract amino acid residue features using a pre-trained language model, and generate a three-dimensional structure and construct a spatial adjacency graph using a protein structure prediction model.

[0022] The sequence feature extraction module, connected to the feature encoding module, is composed of a deep sequence module and is used to extract sequence features from the amino acid residue features;

[0023] The structural feature extraction module, connected to the feature encoding module, is composed of a hierarchical graph module and is used to extract structural features from the amino acid residue features and the spatial adjacency graph.

[0024] A cross-modal alignment module, connected to the sequence feature extraction module and the structural feature extraction module, is used to align the sequence features and structural features of the same protein across modalities in the latent space through the contrastive learning module.

[0025] The feature fusion module, connected to the sequence feature extraction module and the structural feature extraction module, is used to obtain joint features by bidirectionally interacting sequence features and structural features using a cross-attention module.

[0026] The classification prediction module, connected to the feature fusion module, is used to perform nonlinear transformation and classification on the joint features and output a binary probability score to determine whether it is a type IV secretory effector protein.

[0027] The present invention also provides a non-volatile computer storage medium storing computer-executable instructions that execute the type IV secretory effector protein recognition method.

[0028] The present invention also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the type IV secretory effector protein recognition method.

[0029] The beneficial effects of this invention are:

[0030] Compared with existing type IV secretory effector protein identification technologies, this invention effectively avoids mutual interference between multimodal information and improves feature decoupling ability by designing a multimodal learning framework that fuses sequence and structure. A stacked graph convolutional network combined with residual connections is introduced into the hierarchical graph module to extract residue attributes and spatial topological relationships layer by layer, enhancing the depth and robustness of structural features. Furthermore, a contrastive learning module is introduced to align sequence and structural features in the latent space, significantly improving cross-modal consistency. A cross-attention module enables bidirectional interaction between sequence and structural features, fully mining complementary information between modalities to generate more discriminative joint features. Finally, a gated linear unit is used as a classifier to improve classification stability. Experimental results show that the overall prediction performance of this invention on the test set is significantly better than existing type IV secretory effector protein identification methods, demonstrating excellent recognition accuracy and promising application prospects. Attached Figure Description

[0031] Figure 1 This is a flowchart of a method for identifying type IV secretory effector proteins based on multimodal information;

[0032] Figure 2 This is a schematic diagram of the deep sequence module;

[0033] Figure 3 This is a structural diagram of the hierarchical graph module;

[0034] Figure 4 This is a schematic diagram of the gated linear unit. Detailed Implementation

[0035] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0036] like Figure 1 As shown, a method for identifying type IV secretory effector proteins based on multimodal information includes the following steps:

[0037] Step S1: Input the amino acid sequence of the protein into the protein language model to obtain amino acid residue features; input the amino acid sequence of the protein into the protein structure prediction model to obtain a spatial adjacency graph.

[0038] Step S2: Input the amino acid residue features into the deep sequence module to extract the protein sequence features; input the amino acid residue features as node features and the spatial adjacency graph into the hierarchical graph module to obtain the protein structural features;

[0039] Step S3: Perform global average pooling on the sequence features and structural features respectively to obtain global sequence feature vector and global structural feature vector, and use the contrastive learning module to ensure that the sequence and structural representations of the same protein in the latent space are consistent, thereby achieving cross-modal alignment;

[0040] Step S4: Input the sequence features and structural features into the cross-attention module to obtain joint features;

[0041] Step S5: Perform global average pooling on the joint features and input them into a gated linear unit to determine whether they are type IV secretory effector proteins.

[0042] In this embodiment, the protein language model uses the esm2_t33_650M_UR50D version of the ESM-2 model. esm2_t33_650M_UR50D is self-supervised pre-trained on large-scale protein sequence data and possesses powerful amino acid semantic representation capabilities. After inputting the target protein amino acid sequence, the protein language model outputs the features of each amino acid residue, with a dimension of 1280.

[0043] The protein structure prediction model ESM-Fold, based on the pre-trained ESM-1v model, can directly predict the three-dimensional structure of a protein sequence. In constructing the spatial adjacency graph, each amino acid residue is considered a node in the graph, and the spatial distance between the main chain atoms of each pair of amino acid residues is calculated. The shortest distance between any pair of residues is... - If the distance between carbon atoms is less than 10 Å, then an edge connection is established between the corresponding nodes to construct the spatial adjacency graph of the protein.

[0044] Amino acid residue features extracted using a protein language model are used for modeling the deep sequence module and the hierarchical graph module. These amino acid residue features are not only input into the subsequent deep sequence module but also serve as node features, combined with edge indices from a spatial adjacency graph constructed based on spatial distance, and input into the hierarchical graph module.

[0045] like Figure 2 As shown, the deep sequence module consists of four identical sub-layers stacked together. Each sub-layer is sequentially connected by multi-head attention, layer normalization, a feedforward network, and another layer normalization layer. Residual connections are introduced after each layer normalization. The deep sequence module is used to extract the contextual dependencies of protein amino acid sequences and output sequence features. Specifically, the amino acid residue features are represented as follows: Where R is the set of real numbers, N is the number of amino acid residues, d=1280, and d is the feature dimension. The amino acid residue feature X is input into the deep sequence module, and the calculation process is as follows:

[0046] For the Layer upon layer, Layer-by-layer output Context-aware features are computed using a multi-head attention mechanism:

[0047] ;

[0048] Then layer normalization is performed and residual connections are introduced:

[0049] ;

[0050] The output of the multi-head attention network, after being connected with residuals and normalized by layers, is fed into the feedforward network.

[0051] ;

[0052] After performing layer normalization and residual connection again, the output of the l-th sublayer is obtained:

[0053] ;

[0054] in, In the first Context-aware features in each sub-layer after multi-head attention computation For multi-head self-attention functions, In the first The output of each sublayer, after multi-head attention output, residual connection, and layer normalization, is the final output. In the first The output of the feedforward network in each sublayer is a feature obtained after further nonlinear transformation of the attention result. It is a feedforward network, consisting of two fully connected layers and one activation function. To modify the activation function of the linear unit, In the first The weights of the first fully connected layer of the feedforward network in each sub-layer In the first The bias of the first fully connected layer of the feedforward network in each sub-layer. In the first The weights of the second fully connected layer of the feedforward network in each sub-layer In the first The second fully connected layer of the feedforward network in each sub-layer is biased. The output of the l-th sub-layer Let l be the layer normalization function, where l represents the th layer in the deep sequence module. Sub-layers, S represents the total number of sub-layers in the deep sequence module. The final output features are the result of stacking S sub-layers, and the final output sequence features are... .

[0055] like Figure 3As shown, the hierarchical graph module consists of two stacked graph convolutional networks, each followed by a residual connection. The first-layer graph convolutional network receives amino acid residue features and a spatial adjacency graph. First, the amino acid residue features and spatial adjacency graph are input into the first-layer graph convolutional network. The output features of the first-layer network are combined with the original amino acid residue features through an addition operation to achieve a residual connection. Then, the fused features are input into the second-layer graph convolutional network, and graph convolution is performed again. The result of the second-layer network operation is then residually connected with the output features of the first-layer network to obtain the structural features. Specifically, the extracted amino acid residue features X are used as graph node features, combined with the spatial adjacency graph... Input is given to the hierarchical graph module, where V is the set of nodes and E is the set of edges; let the adjacency matrix of the spatial adjacency graph be... Where N is the total number of amino acid residues in the protein sequence, which is also equal to the number of nodes in the spatial adjacency graph, and the node characteristics are... .

[0056] First, define the node features. The adjacency matrix A is input into the first layer of the graph convolutional network:

[0057] ;

[0058] ;

[0059] The residual connection is achieved by adding the output features of the first-layer graph convolutional network to the original node features:

[0060] ;

[0061] Then, the features obtained by fusing the output features of the first layer of graph convolution with the original node features are input into the second layer of graph convolutional network:

[0062] ;

[0063] The output features of the second-layer graph convolutional network are added to the output of the first-layer graph convolutional network, and a residual connection is performed to obtain the structural features:

[0064] ;

[0065] in, To add a self-loop adjacency matrix, For the sigmoid function, The adjacency matrix and degree matrix for adding self-loops, where I is the identity matrix. The first layer of graph convolution is a learnable weight matrix. The learnable weight matrix for the second-layer graph convolution. The output features of the first layer graph convolution, This is the feature obtained by fusing the output features of the first layer graph convolution with the original node features. The output features of the second layer graph convolution, The structural features are obtained by adding the convolution of the first layer graph and the convolution of the second layer graph.

[0066] The process effectively aggregates the spatial relationships between amino acid residues through two-layer graph convolution and dual residual connections, thereby enhancing the ability to characterize structures.

[0067] After constructing sequence features With structural features After the joint representation, a contrastive learning module is introduced, aiming to maximize the similarity of the same protein across different modalities (sequence and structure) while minimizing the similarity between different samples. Specifically, this module first analyzes sequence features... and structural features Perform global average pooling separately to obtain global sequence feature vectors. and global structural feature vectors Where b is the training batch size and d is the feature dimension, the global sequence feature vector is then calculated. and global structural feature vectors Contrast loss between:

[0068] ;

[0069] ;

[0070] ;

[0071] in, This represents the loss calculated using sequence-enhanced view features as the query. k represents the loss calculated using structurally enhanced view features as the query. {0,…,b−1} represents the first traversal index of samples within the batch, m {0,…,b−1} represents the second traversal index of samples within a batch, where b represents the training batch size, used to enumerate all positive and negative sample candidates. This represents the original feature vector corresponding to the sequence augmentation view output by the encoder for the k-th sample. This represents the original feature vector corresponding to the structure-enhanced view output by the encoder for the k-th sample, where d is the feature dimension. This represents the unit feature vector after L2 normalization of the original features of the sequence. This represents the unit eigenvector after L2 normalization of the original structural features, where T is the matrix transpose, and τ>0 represents the temperature coefficient used to adjust the sharpness of the similarity distribution. and The terms represent the positive sample similarity index between the enhanced view of the same instance sequence and the enhanced view of the structure, respectively. Indicated by The sum of the similarity indices with all 2b features within the batch during the query constitutes the negative sample comparison set, and the other denominator term is... Indicated by The sum of similarity indices between the query and all 2b features within the batch forms the negative sample comparison set, and L represents the final symmetric loss. This entire contrastive learning module achieves cross-modal alignment of the same protein in the latent space, effectively improving the quality of feature interactions in the subsequent cross-attention module. This makes the fused representation more discriminative, thereby supporting performance improvements in classification tasks.

[0072] In the cross-attention module, sequence features are used as queries, and structural features are used as keys and values. Cross-modal attention weights are calculated to achieve bidirectional interaction between sequence and structural features. After weighted fusion, joint features with both semantic information and spatial context are generated. Specifically, let the sequence features... Structural features Then the cross-attention module can be defined as:

[0073] ;

[0074] ;

[0075] ;

[0076] ;

[0077] in, Let h be the query vector of the h-th attention head. This indicates that the projection result is divided into the h-th attention head along the feature dimension. Let h be the key vector of the h-th attention head. Let h be the value vector of the attention head. and These are the projection matrices for the keys and values, respectively. The joint features output by the cross-attention module. (N is the length of the feature sequence output by the cross-attention module, i.e., the number of residues). This indicates concatenation along the feature dimension, where H is the number of attention heads and h is the index of the current attention head. This indicates that similarity is transformed into a probability distribution, where T is the matrix transpose. To query the projection matrix, To output the projection matrix, This is the bias term for the output layer.

[0078] like Figure 4 As shown, the gated linear unit includes value branch linear transformation, gated branch linear transformation, Gaussian error linear unit activation function, Sigmoid activation function, element-wise multiplication, and linear layers, forming a nonlinear classification structure with a gated mechanism, used for discriminative weighting of joint features; the processing procedure of the gated linear unit can be defined as:

[0079] ;

[0080] ;

[0081] ;

[0082] Where N is the length of the feature sequence (number of amino acid residues) output by the cross-attention module. Let be the joint feature of the i-th amino acid residue, where i is an index variable representing the current processing position number, and c is the feature vector obtained by pooling the joint features obtained from the cross-attention module. Let v be a gated linear activation function, and g be the value vector and the gate signal, respectively. For the sigmoid function, This indicates element-wise multiplication. The activation function for the Gaussian error linear unit; To output a linear projection, This is a linear projection onto the residual path. Normalization along the category dimension yields the predicted probability distribution. This indicates that the category index with the highest probability is selected as the final prediction result. This is the final predicted category label.

[0083] The method of this invention outputs two probabilities using the Softmax function, and takes the probability of the type IV secreted effector protein category as the prediction score (ranging from 0 to 1). When the prediction score is greater than or equal to the threshold p (p is a preset scalar value used to divide the decision boundary between effector proteins and non-effector proteins, with a value range of (0,1), and a default value of 0.5), it is determined to be an effector protein; otherwise, it is a non-effector protein.

[0084] In summary, this invention provides a method for identifying type IV secretory effector proteins based on multimodal information. It acquires the amino acid residue features and spatial adjacency graph of proteins through a protein language model and a protein structure prediction model, respectively, constructs a learning framework that fuses sequence and structure multimodal features, aligns modal representations using a contrastive learning module, and achieves deep fusion of sequence and structural features using a cross-attention module. Finally, the classification result is output after nonlinear transformation by a gated linear unit. This method solves the problems of traditional models relying on handcrafted features and struggling to effectively model sequence-structure associations, significantly improving the accuracy of type IV secretory effector protein prediction.

[0085] To verify the effectiveness of the method of this invention, this embodiment uses a balanced dataset containing 1419 samples, divided into training, validation, and test sets in a ratio of 7:1.5:1.5. The method of this invention (T4BiPath) is trained on the training set, fine-tuned on the validation set to determine the optimal parameters, and finally evaluated on the test set based on T4BiPath. Table 1 shows the performance comparison of the method of this invention (T4BiPath) with other existing methods on independent test sets. Overall, T4BiPath outperforms other models. Specifically, T4BiPath outperforms other models in accuracy (2.2%~26.0%), F1 score (2.4%~20.9%), specificity (12.2%~44.9%), precision (0.2%~30.0%), recall (3.5%~14.3%), and MCC (Matthews correlation coefficient) by 4.5%~49.6%, respectively, significantly outperforming existing technologies. These results fully validate that the present invention, by fusing sequence and structural information and introducing a bimodal interaction mechanism, possesses stronger discriminative ability and robustness in the identification task of type IV secretory effector proteins. Specifically, Bastion4 is a multi-model ensemble voting prediction tool for type IV secretory effector proteins, T4SEpp is a multi-source feature fusion prediction model for type IV secretory effector proteins, DeepSecEbd is a deep learning feature embedding prediction tool for type IV secretory effector proteins, T4SEfinder is a key region focusing prediction tool for type IV secretory effector proteins, and T4Seeker is a multimodal hybrid identification model for type IV secretory effector proteins.

[0086] Table 1 compares the models with other models on the independent test set.

[0087]

[0088] This invention introduces a multimodal learning framework that integrates sequence and structure to capture the sequence semantics and spatial structure information of proteins, and combines a cross-attention module and a contrastive learning module to achieve alignment and deep fusion of sequence and structural features. To evaluate the contribution of each module to the recognition performance of type IV secretory effector proteins, an ablation experiment was designed: variant models were constructed with only the sequence module removed, only the structure module removed, the cross-attention module removed, and the contrastive learning module removed, and compared with the complete model T4BiPath on an independent test set. The experimental results are shown in Table 2. When only the sequence module or the structure module was removed, the model accuracy was 91.30% and 91.70%, respectively, indicating that sequence and structural information play a crucial and balanced role in the recognition of type IV secretory effectors. After removing the cross-attention module, the accuracy and F1 score decreased to 93.60% and 93.80%, respectively, indicating that this module effectively promoted the bidirectional complementarity of sequence and structural features. After removing the contrastive learning module, the MCC slightly decreased from 0.909 to 0.890, reflecting its stabilizing effect on cross-modal representation consistency. Overall, the complete model T4BiPath improves accuracy, F1 score, and MCC by 0.9%~4.1%, 1.0%~4.0%, and 1.9%~8.2% respectively compared to various ablation variants. This fully verifies the effectiveness and necessity of the multimodal architecture and its core components, namely the cross-attention module and the contrastive learning module, adopted in this invention in improving feature representation ability, robustness, and generalization performance.

[0089] Table 2 Ablation Experiment

[0090]

[0091] Another embodiment of the present invention provides a type IV secretory effector protein recognition system based on multimodal information, comprising:

[0092] The feature encoding module is used to receive the protein amino acid sequence, extract amino acid residue features using a pre-trained language model, and generate a three-dimensional structure and construct a spatial adjacency graph using a protein structure prediction model.

[0093] The sequence feature extraction module, connected to the feature encoding module, is composed of a deep sequence module and is used to extract sequence features from the amino acid residue features;

[0094] The structural feature extraction module, connected to the feature encoding module, is composed of a hierarchical graph module and is used to extract structural features from the amino acid residue features and the spatial adjacency graph.

[0095] A cross-modal alignment module, connected to the sequence feature extraction module and the structural feature extraction module, is used to align the sequence features and structural features of the same protein in the latent space through the contrastive learning module.

[0096] The feature fusion module, connected to the sequence feature extraction module and the structural feature extraction module, is used to achieve bidirectional interaction between sequence features and structural features through the cross-attention module to generate joint features;

[0097] The classification prediction module, connected to the feature fusion module, is used to perform nonlinear transformation and classification on the joint features and output a binary probability score to determine whether it is a type IV secretory effector protein.

[0098] Another embodiment of the present invention provides a non-volatile computer storage medium storing computer-executable instructions that perform the type IV secretory effector protein recognition method.

[0099] Another embodiment of the present invention provides an electronic device, including: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the type IV secretory effector protein recognition method.

[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any person skilled in the art can modify, substitute, or make equivalent transformations to the above embodiments without departing from the technical solution of the present invention, and the equivalent embodiments obtained shall still fall within the protection scope of the present invention.

Claims

1. A method for identifying type IV secretory effector proteins based on multimodal information, characterized in that, Includes the following steps: Step S1: Input the amino acid sequence of the protein into the protein language model ESM-2 to obtain amino acid residue features; input the amino acid sequence of the protein into the protein structure prediction model ESM-Fold to obtain a spatial adjacency graph constructed based on the criterion that the α-carbon atom spacing is less than 10 Å. Step S2: Input the amino acid residue features into a deep sequence module composed of multiple stacked Transformer blocks to extract the sequence features of the protein; input the amino acid residue features as node features and the spatial adjacency graph into a hierarchical graph module composed of two stacked graph convolutional networks, and introduce residual connections after each graph convolutional network to extract the structural features of the protein. Step S3: Perform global average pooling on the sequence features and structural features respectively to obtain global sequence feature vector and global structural feature vector. Then, by using the contrastive learning module with the InfoNCE loss function, the sequence features and structural features of the same protein are used as positive sample pairs, and the feature combinations of other proteins in the same batch are used as negative sample pairs. This ensures that the sequence and structural representation of the same protein in the latent space are consistent, and achieves cross-modal alignment. Step S4: Input sequence features and structural features into the cross-attention module, where sequence features are used as queries and structural features are used as keys and values, calculate cross-modal attention weights, realize bidirectional interaction between sequence and structural features, and obtain joint features; Step S5: Perform global average pooling on the joint features, and then input them into a gated linear unit. The gated linear unit includes a value branch linear transformation, a gated branch linear transformation, a sigmoid activation function, element-wise multiplication, a Gaussian error linear unit activation function, and a linear layer. The gate weights adaptively suppress noise in the non-effect region of the long sequence and output a binary classification probability score to determine whether it is a type IV secretory effector protein.

2. The method according to claim 1, characterized in that, The protein language model adopts the esm2_t33_650M_UR50D version of the ESM-2 model; the protein structure prediction model is the ESM-Fold model, which is used to predict the three-dimensional structure of proteins from the protein amino acid sequence and constructs a spatial adjacency graph based on the criterion that the α-carbon atom spacing is less than 10 Å.

3. The method according to claim 1, characterized in that, The deep sequence module consists of four identical sub-layers stacked together. Each sub-layer is composed of multi-head attention, layer normalization, feedforward network, and layer normalization connected in sequence. Residual connections are introduced after each layer normalization.

4. The method according to claim 1, characterized in that, The hierarchical graph module consists of two stacked graph convolutional networks. Each graph convolutional network is followed by a residual connection. The first graph convolutional network receives amino acid residue features and spatial adjacency graphs, while the second graph convolutional network receives the output of the first graph convolutional network and the residual connection results of the original amino acid residue features. Both graph convolutional networks use low-pass filtering aggregation to extract smooth structural domain topological features.

5. The method according to claim 1, characterized in that, The contrastive learning module uses the InfoNCE loss function, which treats the sequence features and structural features of the same protein as positive sample pairs, and the feature combinations of the remaining samples as negative sample pairs, so that the sequence features and structural features are aligned across modalities in the latent space. The loss function includes a loss anchored by the sequence features and a loss anchored by the structural features, and the two are added together to form a symmetric contrastive loss.

6. The method according to claim 1, characterized in that, The cross-attention module receives sequence features as queries and structural features as keys and values. It calculates cross-modal attention weights through a multi-head attention mechanism, and outputs joint features after linear transformation of the attention-weighted value vector. This enables structural information to guide sequence information and sequence information to supplement structural information.

7. The method according to claim 1, characterized in that, The gated linear unit includes a value branch linear transformation, a gated branch linear transformation, a Gaussian error linear unit activation function, a Sigmoid activation function, element-wise multiplication, and a linear layer, forming a nonlinear classification structure with a gating mechanism. The input features of the gated linear unit are divided into two branches. One branch is transformed by the value branch linear transformation to obtain a value vector, and the other branch is transformed by the gated branch linear transformation and activated by the Sigmoid function to obtain gate weights. After the value vector and gate weights are multiplied element-wise, they are passed through the Gaussian error linear unit activation function and the linear layer, and the final output is a binary classification probability score.

8. A system for recognizing type IV secretory effector proteins based on multimodal information, characterized in that, include: The feature encoding module is used to receive the protein amino acid sequence, extract amino acid residue features using the pre-trained language model ESM-2, and generate a three-dimensional structure using the protein structure prediction model ESM-Fold and construct a spatial adjacency graph based on the criterion that the α-carbon atom spacing is less than 10 Å. The sequence feature extraction module, connected to the feature encoding module, is composed of a deep sequence module and is used to extract sequence features from the amino acid residue features; The structural feature extraction module, connected to the feature encoding module, is composed of a hierarchical graph module and is used to extract structural features from the amino acid residue features and the spatial adjacency graph. The hierarchical graph module is composed of two stacked graph convolutional networks, with residual connections introduced after each graph convolutional network. A cross-modal alignment module, connected to the sequence feature extraction module and the structural feature extraction module, is used to align the sequence features and structural features of the same protein across modalities in the latent space through the contrastive learning module using the InfoNCE loss function. The feature fusion module, connected to the sequence feature extraction module and the structural feature extraction module, is used to perform bidirectional interaction between sequence features and structural features using the cross-attention module to obtain joint features, wherein the sequence features are used as queries and the structural features are used as keys and values. The classification prediction module, connected to the feature fusion module, is used to input the joint features into a gated linear unit after global average pooling for nonlinear transformation and classification, and output a binary classification probability score to determine whether it is a type IV secretory effector protein. The gated linear unit includes value branch linear transformation, gated branch linear transformation, sigmoid activation function, element-wise multiplication, Gaussian error linear unit activation function and linear layer.

9. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer can execute instructions to perform the method for recognizing type IV secretory effector proteins as described in any one of claims 1-7.

10. An electronic device, comprising: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, characterized in that the instructions are executed by the at least one processor to cause the at least one processor to perform the type IV secretory effector protein recognition method according to any one of claims 1-7.

Images