Dual-branch protein sequence feature fusion interactive prediction method and system

Abstract

This invention relates to the field of bioinformatics, and particularly to a method and system for predicting protein interaction using a two-branch protein sequence feature fusion approach. The method includes: acquiring a first protein sequence and a second protein sequence and vectorizing them respectively to generate corresponding primary features; performing dynamic channel attention enhancement processing on the first protein's primary features to generate enhanced features for the first protein; performing semantic embedding encoding processing on the second protein's primary features to generate semantic features for the second protein; fusing the enhanced features and semantic features to obtain a two-branch interaction feature set; constructing nodes of a graph isomorphic network using the elements of the set; generating protein pair structure embedding representations based on feature weighting aggregation of neighboring nodes; inputting these representations into a fully connected classifier; and outputting protein interaction prediction results. This invention effectively solves the problem that existing technologies cannot fully capture the lack of collaborative representation of multi-scale biological semantics in protein interaction prediction.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a method and system for interactive prediction of bibranched protein sequence features. Background Technology

[0002] Protein-protein interactions (PPIs) serve as the molecular basis for cell signaling, metabolic regulation, and immune responses in life activities. Accurate prediction of PPIs is of significant scientific value for elucidating disease mechanisms and identifying drug targets. Traditional experimental methods are limited by false positives and time-consuming processes in high-throughput screening, driving the development of computational prediction techniques.

[0003] Existing computational methods are mainly based on three technical approaches. First, sequence similarity-driven methods (such as BLAST) rely on homologous sequence alignment to infer interactions, but cannot capture functional associations between non-homologous proteins. Second, shallow machine learning models (such as SVM / RF) achieve classification prediction through manually designed features, but due to the single dimension of feature representation, it is difficult to integrate the local physical properties and global topological constraints of proteins. Third, deep neural network architectures (such as CNN-LSTM hybrid models), although they can extract local sequence patterns and model temporal dependencies, still have receptive field limitations when dealing with long-distance residue interactions and ignore the semantic information implicit in the evolutionary conservation of proteins. Of particular note is that PPI prediction methods based on graph neural networks (GNNs), although modeling protein-protein interaction relationships through graph structures, still suffer from the fragmentation of local and global features, isolation of semantic representation, and coarse-grained modeling of node features.

[0004] The aforementioned shortcomings make it difficult for existing models to capture complex biological phenomena such as allosteric effect transmission and conformational change-induced binding site remodeling, which severely limits the accuracy and generalization ability of PPI prediction.

[0005] The information disclosed in this background section is intended only to enhance the understanding of the general background of this disclosure and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] This invention provides a dual-branch protein sequence feature fusion interactive prediction method and system, which can effectively solve the problems in the background art.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A dual-branch protein sequence feature fusion interactive prediction method, the method comprising: Acquire at least one pair of target protein sequence data to generate a first protein sequence and a second protein sequence; The first protein sequence and the second protein sequence are vectorized according to the convolutional neural network embedding layer to generate primary features of the first protein and primary features of the second protein. Dynamic channel attention enhancement processing is performed on the first protein primary features to generate the first protein enhanced features; The primary features of the second protein are semantically embedded and encoded to generate semantic features of the second protein. The first protein enhancement feature and the second protein semantic feature are fused together to obtain a dual-branch interactive feature set. The elements of the bi-branch interaction feature set are constructed as nodes of a graph isomorphic network, and protein pair structure embedding representations are generated based on the feature weighted aggregation of neighborhood nodes. The protein pair structure embedding representation is input into a fully connected classifier, which outputs protein interaction prediction results.

[0008] Furthermore, dynamic channel attention enhancement processing is performed on the primary features of the first protein, including: The local correlation features of the primary features of the first protein are obtained based on a one-dimensional convolutional layer, and the local correlation features are normalized by BatchNorm1D normalization to generate normalized features. The normalized features are input into the SEBlock1D module, the sequence dimension is compressed to 1 through global average pooling, and channel weight coefficients are generated through two fully connected layers. The channel weight coefficients are then multiplied element-wise with the normalized features to generate weighted features. MaxPool1D pooling is performed on the weighted features to generate pooled features, and the pooled features are input into GRU for time series modeling to obtain GRU features; The attention score of the GRU feature is calculated and normalized to a weight coefficient using softmax; the first weighted feature is obtained by weighting and summing the weight coefficients. The first weighted feature is input into the fully connected layer to generate the first protein enhancement feature.

[0009] Furthermore, semantic embedding encoding processing includes: The primary features of the second protein are input into the residual layer of the ESM2 model; Obtain the output vectors of multiple hidden layers in the residual layer; The output vectors of each hidden layer are weighted and summed based on the gating vectors with learnable parameters to generate the second protein semantic features.

[0010] Furthermore, feature fusion processing includes: The first protein enhancement feature and the second protein semantic feature are concatenated along the channel dimension; The concatenated features are input into a one-dimensional convolutional layer with a kernel size of 3, a stride of 1, and padding of 1, and the output feature dimension is adjusted to 8. MaxPool1D pooling with a window size of 2, a stride of 2, and padding of 0 is performed on the convolution output to flatten the pooled features into a one-dimensional vector to generate flattened features; A Dropout operation with a dropout ratio of 0.3 is applied to the flattened features, and the Dropout output is fed into a fully connected layer to generate the dual-branch interactive feature set.

[0011] Furthermore, feature-weighted aggregation based on neighborhood nodes includes: In the graph isomorphic network, the features of the target node are added to the features of its neighboring nodes. The contribution ratio between the target node and its neighboring nodes is controlled by weighting coefficients and trainable parameters. The target node features are then updated by nonlinear transformation through a neural network. Perform LayerNorm normalization on the updated target node features; The normalized target node features are then residually concatenated with the original input target node features. ReLU activation and BatchNorm1d normalization are performed sequentially on the residual connection results to generate the protein pair structure embedding representation.

[0012] Furthermore, the self-attention feature extraction process performed when generating the first protein enhancement feature and the second protein semantic feature includes: The input features are mapped to a query matrix, a key matrix, and a value matrix based on linear transformation; the input features are either the first protein enhancement feature or the second protein semantic feature. Calculate the scaled dot product attention score of the query matrix and the key matrix, and normalize the attention score into attention weights using the softmax function; The value matrix is ​​weighted and summed using the attention weights to generate weighted features and perform layer normalization. The normalized weighted features are residually concatenated with the input features; The residual connection result is input into a feedforward neural network consisting of two fully connected layers, wherein the output of the first fully connected layer is activated by ReLU and then input into the second fully connected layer. The output of the feedforward neural network is normalized and residually connected with the input features to generate the final first protein enhancement feature or the second protein semantic feature.

[0013] Furthermore, the graph isomorphic network uses amino acid residues in protein pairs as nodes, and when the spatial Euclidean distance between any two amino acid residues in a protein pair is less than or equal to 8 angstroms, a connection edge is established between the nodes corresponding to the amino acid residues.

[0014] Furthermore, protein-protein interaction prediction results are generated, including: The protein pair structure is embedded in a fully connected layer representing the input cascade; Nonlinear feature transformation and feature compression operations are performed through the cascaded fully connected layers to generate structural fusion features; The structural fusion features are input into a multilayer perceptron, and the multilayer perceptron outputs an interaction probability value mapped to the range of 0-1 as the interaction prediction result.

[0015] A dual-branch protein sequence feature fusion interactive prediction system, the system comprising: The data acquisition module acquires at least one pair of target protein sequence data and generates a first protein sequence and a second protein sequence. The feature vector module vectorizes the first protein sequence and the second protein sequence according to the embedding layer of the convolutional neural network, respectively, to generate primary features of the first protein and primary features of the second protein. The enhancement processing module performs dynamic channel attention enhancement processing on the primary features of the first protein to generate enhanced features of the first protein. The semantic encoding module performs semantic embedding encoding on the primary features of the second protein to generate semantic features of the second protein. The feature fusion module performs feature fusion processing on the first protein enhancement feature and the second protein semantic feature to obtain a dual-branch interactive feature set. The embedding representation module constructs nodes of a graph isomorphic network from the elements of the bi-branch interaction feature set, and generates a protein pair structure embedding representation based on the weighted aggregation of features from neighboring nodes. The results prediction module takes the protein pair structure embedding representation as input to the fully connected classifier and outputs the protein interaction prediction results.

[0016] Furthermore, the result prediction unit includes: The data input unit embeds protein pair structures into a fully connected layer representing the input cascade; The transformation and compression unit performs nonlinear feature transformation and feature compression operations through cascaded fully connected layers to generate structural fusion features; The perceptual mapping unit inputs the structural fusion features into the multilayer perceptron, and the multilayer perceptron outputs interaction probability values ​​mapped to the range of 0-1 as the interaction prediction results.

[0017] The technical solution of this invention can achieve the following technical effects: By combining a dual-branch fusion architecture that integrates basic protein sequence features with deep features of the ESM2 language model, and utilizing one-dimensional convolution and channel attention mechanisms to enhance local feature extraction, a self-attention mechanism to model global dependencies, and graph isomorphic networks to optimize node feature weighted aggregation and residual connections, this approach successfully addresses the problems of existing technologies failing to fully capture multi-level features and long-term temporal dependencies of protein sequences and neglecting semantic information of protein language models. Specifically, it addresses the lack of collaborative representation of multi-scale biological semantics in protein interaction prediction, thereby significantly improving the accuracy and generalization ability of protein-protein interaction prediction.

[0018] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart illustrating the interactive prediction method based on the fusion of two-branch protein sequence features; Figure 2 A flowchart illustrating the dynamic channel attention enhancement process; Figure 3 This is a flowchart illustrating the feature fusion process. Figure 4 This is a flowchart illustrating the feature-weighted aggregation process based on neighborhood nodes. Figure 5 A schematic diagram of the process for performing self-attention feature extraction. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0022] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0023] Example 1; like Figure 1 As shown, this application provides a two-branch protein sequence feature fusion interactive prediction method, the method including: S10: Acquire at least one pair of target protein sequence data and generate a first protein sequence and a second protein sequence; S20: Vectorize the first protein sequence and the second protein sequence according to the convolutional neural network embedding layer to generate primary features of the first protein and primary features of the second protein. S30: Perform dynamic channel attention enhancement processing on the primary features of the first protein to generate enhanced features of the first protein; S40: Perform semantic embedding encoding on the primary features of the second protein to generate semantic features of the second protein; S50: The first protein enhancement feature and the second protein semantic feature are fused to obtain a dual-branch interactive feature set; S60: Construct the elements of the bi-branch interactive feature set into nodes of a graph isomorphic network, and generate protein pair structure embedding representations based on the weighted aggregation of features from neighboring nodes; S70: Input the protein pair structure embedding representation into the fully connected classifier and output the protein interaction prediction results.

[0024] Specifically, firstly, at least one pair of target protein sequence data is acquired and input into the system as the first and second protein sequences, respectively. To facilitate subsequent processing, the amino acid letter sequences can be standardized and encoded, for example, using 20-dimensional or 23-dimensional one-hot encoding of amino acids or pre-trained amino acid embedding vectors as initial symbol representations. The sequence lengths are also standardized; longer sequences can be segmented into fixed windows, while shorter sequences can be padded to a fixed length, for example, a maximum of 1000 residues. Subsequently, the first and second protein sequences are vectorized using convolutional neural network embedding layers to generate primary features. In a preferred embodiment, the convolutional embedding employs a strategy of parallel processing of multi-size convolutional kernels. To capture local structural information at different scales, after convolution, normalization and nonlinear activation are used to obtain a vector representation of each position. Finally, different convolution channels are concatenated or projected to obtain a first protein primary feature with dimension 128 and a second protein primary feature with dimension 128. For the first protein primary feature, this embodiment uses dynamic channel attention enhancement to strengthen the response to key channels, thereby improving the model's sensitivity to local features that can indicate interaction sites. Specifically, global pooling of the channel dimensions can be performed first to obtain channel statistics, and then a lightweight dynamic mapping network is used to generate channel-level weights. In the preferred embodiment, this mapping network is a two-layer fully connected structure with a middle layer... A smaller latent dimension, such as 16 or 32, is used, and a dynamic gating mechanism is introduced so that the weights not only depend on the channel statistics of the current input, but can also be adjusted based on sequence length or other contextual information. Finally, the channel weights are multiplied channel by channel to the first protein primary feature, thereby obtaining the first protein enhanced feature. The improvement of this dynamic channel attention over fixed attention is that it can adaptively adjust the channel importance to adapt to the diversity of different proteins. For the second protein primary feature, this embodiment uses semantic embedding encoding to capture the global semantic and contextual dependency information of the sequence. Preferably, a bidirectional recurrent encoder or an attention-based encoder is used, whose latent dimension matches the convolutional embedding dimension. The first protein enhancement feature and the second protein semantic feature are combined and positional encoding is introduced during the encoding process to preserve sequence order information. After encoding, the second protein semantic feature is obtained by positional projection and normalization. This semantic encoding is able to represent the local long-range dependency relationship with convolution, thereby improving the ability to identify long-distance interaction sites. Then, the first protein enhancement feature and the second protein semantic feature are fused to construct a two-branch interaction feature set. The preferred fusion strategy is a combination of multiple operations to make full use of the two-branch information. For example, a combination of concatenation, element-wise multiplication and difference operations is used, followed by dimensionality transformation and fusion through a convolution or linear mapping layer. At the same time, residual connections and batch normalization can be introduced after fusion to maintain numerical stability.In a specific example, the two branches of features are first aligned by position, or the nodes after segmentation and aggregation are aligned. Then, the concatenation vector, element-wise product vector, and absolute difference vector are calculated. After concatenating these three, a fusion vector for each position or segment is obtained through a linear mapping used for interactive feature compression, forming a two-branch interactive feature set. Subsequently, the elements of this interactive feature set are used as nodes in a graph neural network to construct a graph isomorphic network to obtain structured protein pair representations. In a preferred embodiment, each graph node corresponds to a residue position or a fixed-length fragment. The construction of edges between nodes can be based on a hybrid strategy of sequence proximity and feature similarity, preserving sequential neighborhood connections to reflect peptide chain continuity while using k-nearest neighbors, for example, k=5. Alternatively, threshold similarity can be used to connect non-ordered neighbors to reflect the potential for long-distance interactions. Graph isomorphic networks employ a discriminative neighborhood feature weighted aggregation mechanism. In each round of neighborhood aggregation, learnable neighborhood weights are used to weight and sum the features of neighboring nodes, and the node representation is updated through a non-linear transformation. In specific implementations, a three-layer graph isomorphic network structure is preferred to gradually expand the receptive domain. The hidden dimension of each layer can be set to 256. The neighborhood weights can be learned either by a small attention subnetwork or by using a differentiable weight function based on feature similarity, thereby achieving feature weighted aggregation based on neighboring nodes to generate more discriminative node representations. The readout strategy of the graph isomorphic network can use a combination of global averaging and max pooling, concatenated to obtain the protein. The overall structural embedding representation of the protein pair is obtained, which integrates local interaction signals and global topological information to better describe the interaction potential of the protein pair. Finally, the obtained protein pair structural embedding representation is input into a set of fully connected classification layers to output protein interaction prediction results. In a preferred embodiment, the classifier consists of two to three fully connected layers, each followed by an activation function and dropout regularization, for example, a dropout ratio of 0.3 to 0.5. The final output is a single real value mapped to an interaction probability through a sigmoid function, which is then used for binary classification decisions. During training, a binary cross-entropy loss is used, and an adaptive optimizer, such as Adam, is employed. The initial learning rate can be set to 1e-4, and end-to-end training is performed. To prevent overfitting, strategies such as weight decay and early stopping are employed. Batch sizes of 32 to 128 can be used, and the number of training epochs depends on the data size, for example, 50 to 200 epochs. To verify and enhance generalization ability, data augmentation and balancing can be performed during training. For example, long sequences can be randomly pruned or some residues can be masked to improve robustness to incomplete sequences. Similarly, negative sample screening and resampling strategies can be used in the training set to address sample imbalance. The hyperparameters in the above embodiments, such as the number of convolutional kernels, embedding dimension, number of layers, node partitioning strategy, aggregation weight generation method, and training hyperparameters, are all preferred configurations. In practical applications, these parameters can be reasonably adjusted according to the specific dataset size, sequence length distribution, and computational resources.

[0025] The technical solution of this invention combines a dual-branch fusion architecture of basic protein sequence features and deep features of the ESM2 language model. It utilizes one-dimensional convolution and channel attention mechanisms to enhance local feature extraction, self-attention mechanisms to model global dependencies, and graph isomorphic networks to optimize node feature weighted aggregation and residual connections. This successfully solves the problems of existing technologies being unable to fully capture multi-level features and long-term temporal dependencies of protein sequences and ignoring semantic information of protein language models, namely the lack of collaborative representation of multi-scale biological semantics in protein interaction prediction. As a result, it significantly improves the accuracy and generalization ability of protein-protein interaction prediction.

[0026] Furthermore, such as Figure 2 As shown, dynamic channel attention enhancement processing is performed on the primary features of the first protein, including: The local correlation features of the primary features of the first protein are obtained based on the one-dimensional convolutional layer, and the local correlation features are normalized by BatchNorm1D normalization to generate normalized features. The normalized features are input into the SEBlock1D module, the sequence dimension is compressed to 1 through global average pooling, and channel weight coefficients are generated through two fully connected layers. The channel weight coefficients are then multiplied element-wise with the normalized features to generate weighted features. MaxPool1D pooling is performed on the weighted features to generate pooled features, and the pooled features are input into GRU for time series modeling to obtain GRU features; Calculate the attention score of the GRU feature, normalize it to a weight coefficient using softmax, and sum the weights according to the weight coefficients to obtain the first weighted feature; The first weighted feature is input into the fully connected layer to generate the first protein enhancement feature.

[0027] As a preferred embodiment of the above, a one-dimensional convolutional layer is first used to obtain the local correlation features of the primary features of the first protein. This layer uses convolutional kernels of different sizes to capture multi-scale local structural information. The design of the convolutional kernels can be arranged from small to large, for example, with sizes of 3, 5, and 9. The output after convolution is normalized by BatchNorm1D. This process ensures that the values ​​remain stable in subsequent processing and helps to speed up convergence. In this embodiment, the normalized features are then input into SEBlock1D, which generates channel weights based on an adaptive gating mechanism. First, global average pooling is used to compress the sequence dimension to a single scalar. This allows the module to obtain global summary information for each channel. Then, through two fully connected layers, the module effectively generates weight coefficients for each channel. The first layer reduces dimensionality to maintain adjustability, while the second layer boosts the channel count back to its original value. After non-linear activation, such as ReLU, the final weights are output. These weight coefficients are applied element-wise to the normalized features to obtain weighted features, allowing information from each channel to be redistributed according to importance. Subsequently, the weighted features are subjected to MaxPool1D pooling to extract more representative features. The pooling operation aims to retain the highest response values ​​to obtain simplified yet information-rich features. These features are then fed into a gated recurrent unit (GRU) to perform temporal modeling. GRU is a preferred choice for processing sequence information, retaining only key information and ignoring redundancy. Through the transformation of the GRU, the features are modeled as information that changes over time in the sequence. This example preferably introduces a single-layer GRU to simplify computation. The features output by the GRU are used to calculate attention scores through an attention mechanism, thereby evaluating weights. These scores are normalized using a softmax function to ensure that the sum of all weight coefficients is 1, thus achieving the final weighted summation to obtain the comprehensive first weighted feature. Finally, this first weighted feature is input into a fully connected layer to generate the first... Protein enhancement features: This fully connected layer not only further compresses or expands the feature dimension, but also increases the model's expressive power by introducing nonlinear activation functions, such as sigmoid or tanh. In this embodiment, by combining local association, channel adaptation, and temporal dynamic modeling, a multi-level attention mechanism is used to enhance the representational power of primary features, providing more accurate information support for subsequent protein interaction prediction. The overall design not only makes full use of traditional convolution and modern attention methods, but also creates a fusion feature representation suitable for complex protein structures through sequence temporal modeling and adaptive recompression processes to address the challenge of extremely diverse protein sequence information.

[0028] Furthermore, semantic embedding encoding processing includes: The primary features of the second protein are input into the residual layer of the ESM2 model; Obtain the output vectors of multiple hidden layers in the residual layer; The output vectors of each hidden layer are weighted and summed based on the gating vectors with learnable parameters to generate the second protein semantic features.

[0029] As a preferred embodiment of the above, the enhanced features from the first branch and the semantic features from the second branch are first concatenated along the channel dimension. Before concatenation, the two branches are made consistent in the sequence length dimension to ensure alignment of corresponding positions. For example, when the sequence lengths of the two branches are inconsistent, the longer sequence can be segmented into fixed windows or the shorter sequence can be zero-padding or resampling can be used to make the lengths consistent. Alternatively, linear projection can be performed on the two branches before concatenation to map the channel dimension to similar dimensions so that subsequent convolution processing is more stable. On the feature map obtained after channel concatenation, a one-dimensional convolution operation is preferably used for local feature integration, and the size of the one-dimensional convolution kernel is 3. The specific settings of stride 1 and padding 1 ensure that the convolution simultaneously considers the information of its left and right neighbors at each position, and that the output and input maintain the same length in the long dimension of the sequence. Setting the number of output channels after the convolution operation to 8 significantly reduces the feature dimension and compresses the amount of information, facilitating subsequent computation and improving generalization ability. Batch normalization and non-linear activation (e.g., ReLU) are typically applied after convolution to enhance training stability and expressive power. Subsequently, max pooling with a window size of 2, stride of 2, and padding of 0 is performed on the convolution output. This pooling setting reduces the sequence length to half its original length with each execution, thereby preserving high-response features while reducing the load on subsequent flattening and fully connected layers. The input dimension is determined by flattening the multi-channel features after pooling to transform the two-dimensional channel-length representation into a one-dimensional vector representation. The order of channel priority or length priority during flattening should be clearly defined and maintained consistent during training and inference. For example, in a specific example, if the concatenated output after convolution has 8 channels and the sequence length is 200, then after pooling the length becomes 100, and the flattened vector dimension is 8 multiplied by 100, equaling 800. This example is only for illustrating the approach to dimension changes and is not a limitation. Next, a random deactivation operation with a dropout ratio of 0.3 is applied to the flattened features to suppress overfitting and improve model robustness. Finally, the Dropout-processed vector is input into a fully connected layer. A mapping layer is used to generate a set of bi-branch interactive features. This fully connected layer can be set to one or more layers according to the task requirements. The output dimension can be adjusted according to the node feature dimension required by the downstream graph construction or classifier input, for example, the output is 128 or 256 dimensions. An activation function and normalization are added after the fully connected layer to obtain stable and easy-to-train interactive features. To improve the applicability and performance of the implementation, this embodiment also preferably introduces residual connections or inter-layer normalization before and after convolution. In terms of pooling strategy, a combination of max pooling and average pooling can be used to balance responsiveness and smoothness. In the training stage, batch normalization, learning rate scheduling and early stopping techniques are used to obtain better generalization effect.

[0030] Furthermore, such as Figure 3 As shown, the feature fusion process includes: The first protein enhancement feature and the second protein semantic feature are spliced ​​along the channel dimension; The concatenated features are input into a one-dimensional convolutional layer with a kernel size of 3, a stride of 1, and padding of 1, and the output feature dimension is adjusted to 8. MaxPool1D pooling with a window size of 2, a stride of 2, and padding of 0 is performed on the convolution output to flatten the pooled features into a one-dimensional vector to generate flattened features; A Dropout operation with a dropout ratio of 0.3 is applied to the flattened features, and the Dropout output is fed into the fully connected layer to generate a set of bi-branch interactive features.

[0031] As a preferred embodiment of the above, firstly, the aforementioned first protein enhancement feature and second protein semantic feature are directly concatenated along the channel dimension. Assuming that the channel dimension of each of the first enhancement feature and the second semantic feature is 128 and the sequence length is L, the resulting feature is a tensor with a channel dimension of 256 and a sequence dimension of L. Here, the concatenation is performed channel-by-channel along the channel axis to preserve the information of each branch and provide a joint representation for subsequent interaction modeling. Then, the concatenated feature is input into a one-dimensional convolutional operator. The one-dimensional convolution uses a kernel size of 3, a stride of 1, and padding of 1 to ensure that the length of the convolutioned sequence is the same as the length of the input sequence. Simultaneously, the sliding operation of the convolution kernel across adjacent positions is used to capture the two branches. The convolutional layer adjusts the output feature dimension to 8, reducing the number of channels to 8. In implementation, batch normalization and a non-linear activation function (e.g., ReLU) are preferably applied immediately after the convolution to accelerate training and enhance expressiveness. Max pooling is then performed on the convolution output with a pooling window size of 2, a stride of 2, and no padding. This step downsamples the sequence length to approximately half, with an exact value of L / 2 for even-length sequences, to highlight the strongest response and reduce subsequent computational complexity. The tensor shape after pooling is 8×(L / 2). Next, the pooled features are flattened into a one-dimensional vector in a channel-priority manner. Specifically, all channels and the time dimension are unfolded into continuous vectors, which serve as the input to the fully connected layer. Before the fully connected layer, a dropout operation is applied to the flattened features to improve the model's generalization ability; the preferred dropout ratio is 0.3. This dropout operation randomly zeros 30% of the units during training and disables it during inference. The output of Dropout is then fed into the fully connected layer. Preferably, this fully connected layer first projects the high-dimensional flattened vectors onto the desired interactive feature space, for example, projecting them into a set of vectors of length 512 or 128. Specifically, this can be achieved by mapping the flattened vectors into several feature vectors and reshaping them into a set of elements to serve as input for subsequent graph structure nodes. In implementation, the fully connected layer can be designed as a two-layer structure: the first layer is used for dimensionality reduction and combined with non-linear activation to enhance representation ability, and the second layer is used to reduce the final dimension. Mapping and normalization are performed to generate a stable set of bi-branch interactive features. To adapt to protein inputs of different lengths, length normalization and variable shape processing strategies are preferably added to the above process. For example, short sequences are padded or long sequences are segmented before flattening, and the segmented results are concatenated or pooled after fully connected mapping to obtain an interactive feature set of uniform size. In addition, residual connections or channel attention can be added between convolution output and pooling to further retain important signals. For the selection of hyperparameters, preferred values ​​are 8 convolution output channels, 2 pooling windows, 0.3 dropout ratio, and 128 / 512 fully connected output dimensions. In practical applications, these values ​​can be adjusted according to the data scale and computing resources.

[0032] Furthermore, such as Figure 4 As shown, the feature-weighted aggregation based on neighborhood nodes includes: In graph isomorphic networks, the features of the target node are added to the features of its neighboring nodes. The contribution ratio between the target node and its neighboring nodes is controlled by weighting coefficients and trainable parameters, and the target node features are updated by nonlinear transformation through a neural network. Perform LayerNorm normalization on the updated target node features; The normalized target node features are residually concatenated with the original input target node features. ReLU activation and BatchNorm1d normalization are performed sequentially on the residual connection results to generate protein pair structure embedding representations.

[0033] As a preferred embodiment of the above, for each target node in the graph, its predefined set of neighboring nodes is first collected. The neighborhood can be determined by sequential adjacency or by k-nearest neighbors based on feature similarity. Preferably, k is 4 or 5 to balance information content and computational complexity. Then, the features of the target node and the features of each of its neighboring nodes are added together by element dimension to form a candidate merging vector. Next, a set of learnable weighting coefficients is used to control the contribution ratio of the target node and its neighboring nodes in the merging process. The generation of these weighting coefficients preferably involves scoring each pair of target nodes and neighboring nodes by inputting a small trainable attention network. This attention network is a two-layer fully connected structure. The first layer connects... The vector is mapped to a lower dimension, for example, a hidden dimension of 64. After nonlinear activation, the second layer outputs a scalar score. The scores of all neighborhoods are then normalized using softmax to obtain the final neighborhood weights. In addition, an independent, learnable self-loop weight is set to measure the relative importance of the target node itself in the merging process, so that neighborhood information can be adaptively amplified or suppressed by adjusting the ratio of neighborhood weights to self-loop weights. The weighted neighborhood contribution and the target self-loop contribution are summed according to the normalized weights to obtain a weighted aggregation vector. Subsequently, this weighted aggregation vector is input into a trainable nonlinear transformation unit to update the target node features. This transformation preferably adopts a first... The structure involves a linear mapping followed by nonlinear activation. For example, a linear mapping layer projects the input to a hidden dimension consistent with the node representation (e.g., 256 in some cases). Then, GELU or ReLU activations are used to enhance representational power. Optionally, a second linear mapping layer is introduced after this first to increase representational complexity. Dropout is optionally added during the mapping process to suppress overfitting. After the transformation, the updated features are then normalized using LayerNorm to eliminate differences in numerical distribution between different nodes and stabilize training. Normalization is preferably performed on the channel dimension, with an epsilon value of approximately 1e-5 to prevent numerical instability. The normalized features are then compared with the... The original input features of the target node are subjected to residual connections, which are performed element-wise. A learnable scaling factor is allowed to be introduced at the residual to adjust the fusion weight of the original input and the updated features. This residual mechanism preserves the original node information and introduces new information after neighborhood fusion, which helps to alleviate the gradient vanishing problem during deep propagation. Nonlinear activation is applied to the results of the residual connections in sequence, with ReLU being preferred to ensure sparse activation and computational efficiency. Then, BatchNorm1d normalization is performed to normalize the feature distribution within the batch. The momentum and epsilon of BatchNorm can use common default values, such as momentum 0.1. The node representation after sequence processing is the final node-level output. This update process can be repeated for all nodes on the graph for several iterations, preferably 2 to 3 iterations to gradually expand the receptive domain and information fusion depth. After updating all nodes, the structural embedding representation of the protein pair can be obtained by performing a readout operation on the node representation. The preferred readout strategy is to concatenate the results of global average pooling and global max pooling and then normalize them to form the final representation, thus balancing the overall trend with significant local signals. The hyperparameters in the above process, such as neighborhood size, attention network hidden dimension, node representation dimension, number of iterations, and normalization details, are preferred values ​​and can be adjusted appropriately based on the specific dataset size, sequence length distribution, and computational resources.

[0034] Furthermore, such as Figure 5 As shown, the self-attention feature extraction process performed when generating the first protein enhancement feature and the second protein semantic feature includes: The input features are mapped to a query matrix, a key matrix, and a value matrix based on linear transformation; the input features are either the first protein enhancement feature or the second protein semantic feature. Calculate the attention score of the scaled dot product of the query matrix and the key matrix, and normalize the attention score into attention weights using the softmax function; The value matrix is ​​weighted and summed using attention weights to generate weighted features, and then layer normalization is performed. The normalized weighted features are residually connected to the input features; The residual connection result is input into a feedforward neural network consisting of two fully connected layers, wherein the output of the first fully connected layer is activated by ReLU and then input into the second fully connected layer. The output of the feedforward neural network is normalized and residually connected with the input features to generate the final first protein enhancement feature or second protein semantic feature.

[0035] As a preferred embodiment of the above, the input feature vector, which can be a first protein enhancement feature or a second protein semantic feature, is first mapped to a query matrix, a key matrix, and a value matrix through three sets of learnable linear transformations, assuming its channel dimension is D and its sequence length is L. Specifically, each set of linear transformations is converted into its own weight matrix and bias term, and the query, key, and value after the linear transformation can form a tensor of dimension L×D. To enhance the model's ability to model long-range dependencies, the single-head attention is preferably extended to a multi-head form, that is, D is divided into several sub-dimensions according to the number of heads H, and each head independently performs subsequent calculations on its sub-dimensions. Then, the inputs of each head are... The original dimensions are concatenated back; for example, when H=8 heads and total dimension D=256, the dimension of each head is 32. Next, the dot product similarity score of the query matrix and the key matrix is ​​calculated in each attention head, and this score is scaled for stable training, that is, scaled by the square root of the key vector dimension to suppress the numerical amplification caused by the increase of dimension. Then, the scaled similarity score is normalized to attention weights along the dimension of the key using the softmax function, and then the corresponding value matrix is ​​weighted and summed using the attention weights to generate weighted features. In actual implementation, in order to prevent overfitting and enhance the robustness of the model, the weights or the weighted summation result can be adjusted. A small dropout is introduced, preferably with a dropout ratio of 0.1. The weighted outputs from each head are concatenated head-to-head and then mapped back to the original dimensions using a set of linear projections. Layer normalization is then performed on the weighted features. The normalized weighted features are then linked element-wise with the input features to preserve original information and support deep gradient flow. After the residual connections, the residual results are input into a feedforward neural network consisting of two fully connected layers for position-wise nonlinear transformation. The first fully connected layer maps the features to an intermediate dimension larger than the input to expand expressive power; preferably, the intermediate dimension is two to four times the original number of channels. For example, if D is... The intermediate layers can be set to 512 or 1024. After the output of the first layer, ReLU activation is applied to introduce nonlinearity, and dropout can be optionally applied to suppress overfitting. The activated output is then input into the second fully connected layer to map back to the original dimension D. The output of the feedforward network is then normalized to stabilize the numerical distribution, and a second residual connection is performed with the input of the feedforward network, i.e., the aforementioned residual connection result or earlier input features. Finally, the residual connection results are output as the final features in the preferred order of this scheme. In this embodiment, it is recommended to retain an appropriate amount of regularization between the two linear layers of the feedforward network and at the input and output, such as a dropout ratio of 0.1. Weight initialization strategies, such as common Xavier or Kaiming initialization, are used to facilitate training convergence. To further enhance sequence position sensitivity, positional encoding or relative positional bias can be added before the input mapping, such as optional sine / cosine positional encoding or learnable positional embeddings. GELU can be used as a replacement activation to obtain a smoother nonlinear response. This dual residual and layer normalization design can significantly improve the training stability and expressiveness of deep attention structures. Multi-head attention and feedforward extension enable the model to capture long-range dependencies and perform high-order combinations of local patterns, thereby generating high-quality first protein enhancement features or second protein semantic features for downstream interaction prediction.

[0036] Furthermore, the graph isomorphic network uses amino acid residues in protein pairs as nodes, and when the spatial Euclidean distance between any two amino acid residues in a protein pair is less than or equal to 8 angstroms, a connection edge is established between the nodes corresponding to the amino acid residues.

[0037] As a preferred embodiment of the above, firstly, for a given protein pair, the three-dimensional coordinate information of each protein is obtained, preferably using the experimentally resolved structure file. When an experimental structure is unavailable, a high-precision predicted structure can be used as a substitute. To ensure coordinate reliability, low-confidence residues can be removed or downweighted based on the prediction confidence level. For example, a confidence threshold of 70 can be set to remove highly unreliable residues. Subsequently, the residues in the protein sequence are matched one-to-one with the residues in the structure file, considering insertions, deletions, and multi-strand markers. A representative point is selected for each residue for distance calculation. If side-chain interactions are of interest, the side-chain centroid or a specific atom can be selected as a representative point. If the representative atom is missing in the structure... If a residue is missing, other available atomic coordinates are used as a substitute, or no edge is created for that residue. For all residue representative points, the spatial distance between any two residue representative points is calculated based on Euclidean geometry, with 8 angstroms as the default threshold. When the distance between two residues is less than or equal to 8 angstroms, an undirected connection edge is created between the corresponding nodes. The value of 8 angstroms is preferred because it can cover the positional relationships of direct contact and short-range weak interactions, while also taking into account the slight offset caused by structural noise. In practical applications, the threshold can be adjusted within the range of approximately 6 to 12 angstroms to adapt to different sensitivity requirements according to task needs. To improve computational efficiency and scalability, this embodiment suggests constructing a spatial index before distance calculation and performing accurate distance calculation only for spatial nearest neighbors. This avoids the overhead of full-contrast computation for large and complex structures. During graph construction, additional edge information is recorded as edge features, such as precise distance scalars, whether binary contact markers are ≤8 Å, whether inter-chain or intra-chain markers cross protein chains, sequence spacing (sequence position difference), and simple geometric orientation or relative position encoding. These edge features can be used for subsequent graph network weighting or as input to attention weights. When processing multi-body complexes or structures containing missing fragments, it is preferable to segment the graph or prune and sample subgraphs with excessive nodes, while preserving key cross-segment connections to ensure the integrity of interaction patterns. For edge weight design, a distance-driven learnable mapping method is recommended. For example, the original distance can be mapped to a vectorized distance description through a set of basis functions and the relative importance can be learned within the network, or the distance decay function can be directly used as the initial weights and subsequent network adjustments can be allowed. In addition, it is preferable to retain self-loops in the graph to enhance the preservation of node self-information, and to perform threshold filtering or normalization on redundant or weakly correlated edges when constructing edges to suppress the influence of noise on the representation. To ensure the robustness of this embodiment, it is recommended to perform simple preprocessing on the structure before constructing the graph, including coordinate deduplication, abnormal atom removal, and smoothing out isolated noise points using the median of neighboring residues. In the actual training and prediction process, ablation experiments can be conducted on different thresholds and representative point strategies to determine the optimal configuration.

[0038] Furthermore, the predicted results of protein-protein interactions include: Embed protein pair structures into fully connected layers that represent input cascades; Nonlinear feature transformation and feature compression operations are performed through cascaded fully connected layers to generate structural fusion features; The structural fusion features are input into a multilayer perceptron, and the multilayer perceptron outputs interaction probability values ​​mapped to the range of 0-1 as the interaction prediction results.

[0039] In a preferred embodiment, the protein pair structure embedding representation obtained through graph structuring and readout is first introduced as an input vector into a cascaded fully connected transformation unit. In a preferred implementation, the dimension of this input vector can be 256 or 512, etc., to adapt to the dimension of the upstream output. The cascaded fully connected units perform several nonlinear transformations and dimension compression operations to extract more compact and discriminative structural fusion features. Specifically, in a preferred configuration, the input dimension is first mapped to an intermediate dimension, for example, from 512 to 256, through a first fully connected layer, and then a nonlinear activation function (preferably ReLU or GELU) is applied. The process involves normalization (using either BatchNorm or LayerNorm for stable training) and dropout (preferably a dropout ratio of 0.2–0.4 to prevent overfitting), followed by a second fully connected layer to further compress the dimensions (e.g., mapping 256 to 128). This is followed by activation, normalization, and dropout. If necessary, another layer can be added to further compress the features to the target fusion dimension, such as 64 or 128. This cascaded structure design not only progressively abstracts higher-order features but also maintains numerical stability and generalization ability at each step through normalization and regularization. The output is the fused structure. Features; then the fused features of this structure are input into a multilayer perceptron for final prediction mapping. The multilayer perceptron preferably consists of two to three fully connected layers, each followed by an activation function and a suitable amount of dropout. An example configuration is as follows: the first layer maps 128 to 64 and passes through ReLU with a dropout of 0.2; the second layer maps 64 to 16 and passes through ReLU; the third layer is a single neuron outputting a single value, which is then mapped to the 0-1 range using a sigmoid function to obtain the interaction probability value. During the training phase, binary cross-entropy loss is used as the main loss function, and it can be combined with weight decay, such as L2 regularization, and category addition. To address the imbalanced sample problem, a weighted or focus loss method can be used. An adaptive optimizer (e.g., AdamW) is preferred for end-to-end training. The initial learning rate can be set to 1e-4 and used in conjunction with a learning rate scheduler, such as warm-up with cosine decay or performance-based dynamic reduction and early stopping mechanisms to prevent overfitting. The training batch size can be selected from 32 to 128, and the number of training epochs depends on the amount of data, for example, 50 to 200 epochs. To improve the reliability and interpretability of the predicted probabilities, the output probabilities can be calibrated after training, for example, by temperature scaling based on the validation set or equal probability regression calibration. The default threshold for binary decision-making is 0.5. However, it is preferable to optimize specific metrics by selecting the most suitable threshold on the validation set using ROC curves or precision-recall curves, such as maximizing F1 or ensuring a certain precision under high recall. During the inference stage, model ensemble can be used to average or weight the outputs of multiple models and estimate uncertainty using dropout-based Monte Carlo sampling to obtain a more robust probability estimate. Furthermore, to facilitate downstream analysis, intermediate representations before the final layer can be used as interpretable local features for visualization or importance attribution, such as gradient weighting or ensemble gradient methods, to locate key structural regions affecting prediction. The aforementioned hyperparameters, such as the number of cascaded layers, dimension per layer, activation method, dropout ratio, optimization and regularization strategies, calibration and threshold selection methods, etc., are preferred settings in this invention. In practical applications, they can be adjusted according to data scale, positive and negative sample distribution, and computational resources.

[0040] Example 2; Based on the same inventive concept as the dual-branch protein sequence feature fusion interactive prediction method in the foregoing embodiments, the present invention also provides a dual-branch protein sequence feature fusion interactive prediction system, the system comprising: The data acquisition module acquires at least one pair of target protein sequence data and generates a first protein sequence and a second protein sequence. The feature vector module vectorizes the first protein sequence and the second protein sequence according to the embedding layer of the convolutional neural network, respectively, to generate primary features of the first protein and primary features of the second protein. The enhancement processing module performs dynamic channel attention enhancement processing on the primary features of the first protein to generate enhanced features of the first protein. The semantic encoding module performs semantic embedding encoding on the primary features of the second protein to generate semantic features of the second protein. The feature fusion module performs feature fusion processing on the first protein enhancement feature and the second protein semantic feature to obtain a dual-branch interactive feature set. The embedding representation module constructs nodes of a graph isomorphic network from the elements of the bi-branch interaction feature set, and generates a protein pair structure embedding representation based on the weighted aggregation of features from neighboring nodes. The results prediction module takes the protein pair structure embedding representation as input to the fully connected classifier and outputs the protein interaction prediction results.

[0041] The adjustment system described above in this invention can effectively realize the dual-branch protein sequence feature fusion interactive prediction method, and the technical effects it can achieve are as described in the above embodiments, and will not be repeated here.

[0042] Furthermore, the outcome prediction unit includes: The data input unit embeds protein pair structures into a fully connected layer representing the input cascade; The transformation and compression unit performs nonlinear feature transformation and feature compression operations through cascaded fully connected layers to generate structural fusion features; The perceptual mapping unit inputs the structural fusion features into the multilayer perceptron, and the multilayer perceptron outputs interaction probability values ​​mapped to the range of 0-1 as the interaction prediction results.

[0043] Similarly, the above-mentioned optimization schemes for the system can also achieve the optimization effects corresponding to the methods in Embodiment 1, which will not be repeated here.

[0044] Although this application has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made thereto without departing from the spirit and scope of this application. Accordingly, this specification and drawings are merely exemplary illustrations of the application as defined herein, and are to be considered as covering any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Thus, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A dual-branch protein sequence feature fusion interactive prediction method, characterized in that, The method includes: Acquire at least one pair of target protein sequence data to generate a first protein sequence and a second protein sequence; The first protein sequence and the second protein sequence are vectorized according to the convolutional neural network embedding layer to generate primary features of the first protein and primary features of the second protein. Dynamic channel attention enhancement processing is performed on the first protein primary features to generate the first protein enhanced features; The primary features of the second protein are semantically embedded and encoded to generate semantic features of the second protein. The first protein enhancement feature and the second protein semantic feature are fused together to obtain a dual-branch interactive feature set. The elements of the bi-branch interaction feature set are constructed as nodes of a graph isomorphic network, and protein pair structure embedding representations are generated based on the feature weighted aggregation of neighborhood nodes. The protein pair structure embedding representation is input into a fully connected classifier, which outputs protein interaction prediction results.

2. The dual-branch protein sequence feature fusion interactive prediction method according to claim 1, characterized in that, Dynamic channel attention enhancement processing is performed on the primary features of the first protein, including: The local correlation features of the primary features of the first protein are obtained based on a one-dimensional convolutional layer, and the local correlation features are normalized by BatchNorm1D normalization to generate normalized features. The normalized features are input into the SEBlock1D module, the sequence dimension is compressed to 1 through global average pooling, and channel weight coefficients are generated through two fully connected layers. The channel weight coefficients are then multiplied element-wise with the normalized features to generate weighted features. MaxPool1D pooling is performed on the weighted features to generate pooled features, and the pooled features are input into GRU for time series modeling to obtain GRU features; The attention score of the GRU feature is calculated and normalized to a weight coefficient using softmax; the first weighted feature is obtained by weighting and summing the weight coefficients. The first weighted feature is input into the fully connected layer to generate the first protein enhancement feature.

3. The dual-branch protein sequence feature fusion interactive prediction method according to claim 1, characterized in that, Semantic embedding encoding processing includes: The primary features of the second protein are input into the residual layer of the ESM2 model; Obtain the output vectors of multiple hidden layers in the residual layer; The output vectors of each hidden layer are weighted and summed based on the gating vectors with learnable parameters to generate the second protein semantic features.

4. The dual-branch protein sequence feature fusion interactive prediction method according to claim 1, characterized in that, Feature fusion processing includes: The first protein enhancement feature and the second protein semantic feature are concatenated along the channel dimension; The concatenated features are input into a one-dimensional convolutional layer with a kernel size of 3, a stride of 1, and padding of 1, and the output feature dimension is adjusted to 8. MaxPool1D pooling with a window size of 2, a stride of 2, and padding of 0 is performed on the convolution output to flatten the pooled features into a one-dimensional vector to generate flattened features; A Dropout operation with a dropout ratio of 0.3 is applied to the flattened features, and the Dropout output is fed into a fully connected layer to generate the dual-branch interactive feature set.

5. The dual-branch protein sequence feature fusion interactive prediction method according to claim 1, characterized in that, Feature-weighted aggregation based on neighborhood nodes includes: In the graph isomorphic network, the features of the target node are added to the features of its neighboring nodes. The contribution ratio between the target node and its neighboring nodes is controlled by weighting coefficients and trainable parameters. The target node features are then updated by nonlinear transformation through a neural network. Perform LayerNorm normalization on the updated target node features; The normalized target node features are then residually concatenated with the original input target node features. ReLU activation and BatchNorm1d normalization are performed sequentially on the residual connection results to generate the protein pair structure embedding representation.

6. The dual-branch protein sequence feature fusion interactive prediction method according to claim 1, characterized in that, The self-attention feature extraction process performed during the generation of the first protein enhancement feature and the second protein semantic feature includes: The input features are mapped to a query matrix, a key matrix, and a value matrix based on linear transformation; the input features are either the first protein enhancement feature or the second protein semantic feature. Calculate the scaled dot product attention score of the query matrix and the key matrix, and normalize the attention score into attention weights using the softmax function; The value matrix is ​​weighted and summed using the attention weights to generate weighted features and perform layer normalization. The normalized weighted features are residually concatenated with the input features; The residual connection result is input into a feedforward neural network consisting of two fully connected layers, wherein the output of the first fully connected layer is activated by ReLU and then input into the second fully connected layer. The output of the feedforward neural network is normalized and residually connected with the input features to generate the final first protein enhancement feature or the second protein semantic feature.

7. The dual-branch protein sequence feature fusion interactive prediction method according to claim 5, characterized in that, The graph isomorphic network uses amino acid residues in protein pairs as nodes, and when the spatial Euclidean distance between any two amino acid residues in a protein pair is less than or equal to 8 angstroms, a connection edge is established between the nodes corresponding to the amino acid residues.

8. The dual-branch protein sequence feature fusion interactive prediction method according to claim 1, characterized in that, Generate protein-protein interaction prediction results, including: The protein pair structure is embedded in a fully connected layer representing the input cascade; Nonlinear feature transformation and feature compression operations are performed through the cascaded fully connected layers to generate structural fusion features; The structural fusion features are input into a multilayer perceptron, and the multilayer perceptron outputs an interaction probability value mapped to the range of 0-1 as the interaction prediction result.

9. A dual-branch protein sequence feature fusion interactive prediction system, characterized in that, The system includes: The data acquisition module acquires at least one pair of target protein sequence data and generates a first protein sequence and a second protein sequence. The feature vector module vectorizes the first protein sequence and the second protein sequence according to the embedding layer of the convolutional neural network, respectively, to generate primary features of the first protein and primary features of the second protein. The enhancement processing module performs dynamic channel attention enhancement processing on the primary features of the first protein to generate enhanced features of the first protein. The semantic encoding module performs semantic embedding encoding on the primary features of the second protein to generate semantic features of the second protein. The feature fusion module performs feature fusion processing on the first protein enhancement feature and the second protein semantic feature to obtain a dual-branch interactive feature set. The embedding representation module constructs nodes of a graph isomorphic network from the elements of the bi-branch interaction feature set, and generates a protein pair structure embedding representation based on the weighted aggregation of features from neighboring nodes. The results prediction module takes the protein pair structure embedding representation as input to the fully connected classifier and outputs the protein interaction prediction results.

10. The dual-branch protein sequence feature fusion interactive prediction system according to claim 9, characterized in that, The result prediction unit includes: The data input unit embeds protein pair structures into a fully connected layer representing the input cascade; The transformation and compression unit performs nonlinear feature transformation and feature compression operations through cascaded fully connected layers to generate structural fusion features; The perceptual mapping unit inputs the structural fusion features into the multilayer perceptron, and the multilayer perceptron outputs interaction probability values ​​mapped to the range of 0-1 as the interaction prediction results.

Images