Bombyx mori protein function prediction multi-modal fusion method based on deep residual network

Abstract

The invention discloses a bombyx mori protein function prediction multi-modal fusion method based on a deep residual network, and the method comprises the steps: collecting the sequences and high-resolution three-dimensional structure data of bombyx mori proteins of different varieties and at different development stages from databases of bombyx mori genomes, protein crystal structures and the like by employing a data crawling and interface calling technology; and constructing a multi-modal data source. The sequence data are cleaned and coded in a one-hot mode, and the structural data are optimized and converted into a graph structure. Sequence and structure features are respectively extracted through a convolutional neural network and a GraphSAGE graph neural network, and the features are dynamically fused through an adaptive weight fusion strategy. And building a deep residual network training model, and realizing accurate prediction of the protein function category and confidence of the bombyx mori with unknown functions. According to the method, the precision and the high robustness of silkworm protein function prediction are improved through full-process technical optimization, and an efficient tool is provided for silkworm molecular breeding and functional genomics research.

Description

Technical Field

[0001] The present invention relates to the field of protein function prediction, and in particular to a multimodal fusion method for silkworm protein function prediction based on a deep residual network. Background Art

[0002] The prediction of silkworm protein function is of great significance to sericulture science and related fields. With the deepening of molecular biology research, understanding the function of silkworm proteins can help in the selection of superior varieties, disease prevention and control, etc. However, traditional prediction methods are difficult to meet the needs. On the one hand, silkworm protein data are scattered in various databases, and data from different varieties and developmental stages lack systematic integration, making it difficult to obtain comprehensive and high-quality data. On the other hand, past studies have mostly focused on a single data type. Whether relying solely on protein sequences or analyzing protein structures alone, they cannot fully capture the complexity of silkworm protein functions, resulting in limited prediction accuracy.

[0003] Existing technologies for predicting protein functions in silkworms have numerous deficiencies in data processing. Efficient data collection from multiple authoritative databases is difficult, and accurate classification and labeling for different silkworm species and developmental stages is impossible. During the preprocessing phase, conventional methods struggle to effectively remove noise and erroneous sequences from protein sequence data, and simple encoding methods fail to effectively preserve sequence features. For protein structural data, traditional methods struggle to optimize the structure, eliminate irrational conformations, or transform the structure into a form conducive to analysis, resulting in a weak foundation for subsequent analysis.

[0004] From the perspective of model construction and application, the feature extraction network of existing technologies is not sophisticated enough. When extracting sequence features, it is difficult to mine key patterns through parameter optimization; structural feature extraction cannot fully utilize the relationship between nodes and edges. Fusion strategies are mostly fixed weights and cannot adapt to the dynamic changes of different data features. During model training, traditional deep networks are prone to gradient problems and have difficulty learning complex features. When faced with data with imbalanced categories, ordinary loss functions result in poor prediction of rare functional categories. In addition, existing methods lack effective data enhancement methods, resulting in insufficient model generalization capabilities. When faced with new silkworm protein data, the prediction accuracy and reliability are greatly reduced. Summary of the Invention

[0005] In response to the defects of the existing technology, the present invention provides a multimodal fusion method for silkworm protein function prediction based on deep residual network.

[0006] The present invention provides a multimodal fusion method for predicting silkworm protein function based on a deep residual network, comprising the following steps:

[0007] S1. Using data crawling and interface call technology, we collected sequence data and high-resolution 3D structure data of silkworm proteins from the silkworm genome database and protein crystal structure database. We also classified and annotated protein data from different silkworm species and developmental stages to construct a multimodal protein data source.

[0008] S2. We use a sequence cleaning algorithm on protein sequence data, filtering out noise and erroneous sequences through regular expression matching and statistical models, and converting them into digital format using one-hot encoding. For protein structural data, we use molecular dynamics simulation software to optimize the structure according to specific force field parameters and represent the protein as a graph.

[0009] S3. Use convolutional neural networks to extract features from sequence data. Build a graph neural network based on GraphSAGE for structural feature extraction. Utilize node and edge relationships to obtain features. Through an adaptive weight fusion strategy, calculate the weights of each modal feature based on a small fully connected neural network and dynamically fuse them into a unified feature vector.

[0010] S4. Build a deep residual network, use fusion features to train the network, construct a silkworm protein function prediction model, predict silkworm proteins with unknown functions, and output functional categories and confidence levels.

[0011] Furthermore, in step S2, the silkworm protein sequence data is pre-processed by one-hot encoding, and the silkworm protein sequence is s=s1s2…s n , where n is the sequence length, representing that the protein is composed of n amino acid residues connected in sequence, s i represents the i-th amino acid residue in the sequence, i = 1, 2, ... n, the number of amino acid species is m, m = 20, after one-hot encoding, each amino acid residue s i is encoded as a vector v of length m si , if s i Corresponding to the jth amino acid, the vector v si The jth element of is 1, and the rest of the elements are 0. The mathematical expression is

[0012] Furthermore, in step S2, the silkworm protein structure data is preprocessed to represent the protein three-dimensional structure as a graph structure G = (V, E), where V is a node set and each node v i ∈V corresponds to an amino acid residue, E is an edge set, and edge e ij ∈E represents amino acid residue v i and v j There is interaction between them. By mining the structural data features and calculating the degree features of the nodes, let the node v i The degree is d(v i), the calculation formula is |V| represents the total number of nodes in the node set V. The formula means that for node v i The count of all connected edges, d(v i ) represents the node v i The number of other directly connected amino acid residues.

[0013] Furthermore, in step S3, a sequence feature extraction network based on a convolutional neural network for extracting features from silkworm protein sequence data is constructed, comprising multiple convolutional layers and pooling layers. The convolution kernel size of the convolutional layer is set to k×1, where k is an odd number. The input of the lth convolutional layer is set to X l , is a two-dimensional matrix with dimension h l ×w l ,h l Indicates the input feature map height, w l Represents width, the convolution kernel is W l , a two-dimensional matrix with dimensions k×w l , used for sliding convolution on the input feature map, with a bias of b l , is a scalar, the convolution layer outputs Y l By formula Y l =f(X l *W l +b l ) is calculated, f is the ReLU function, the expression is The pooling layer uses maximum pooling, and the pooling window size is p×1.

[0014] Furthermore, in step S3, a graph neural network structure feature extraction network based on GraphSAGE is used for the feature extraction of silkworm protein structure data. In the GraphSAGE network, for each node v i , update its own features by aggregating neighbor node features, let node v i The feature at layer t is The set of neighbor nodes is N(v i ), then node v i The feature update formula at the t+1th layer is: Among them, σ is the activation function, W t is the weight matrix, the dimension of which matches the dimension of the feature vector, b t is the bias vector, CONCAT represents the concatenation operation, which is to connect the node v i Its own features at layer t and the average value of neighbor node features Concatenate them into a new feature vector in sequence.

[0015] Furthermore, in step S3, when fusing the extracted silkworm protein sequence features and structural features, an adaptive weight fusion strategy is adopted, and the features extracted from the sequence data are set as F s , dimension d s , the feature extracted from the structural data is F t , dimension d t , calculate the sequence feature F through a small fully connected neural network s The importance weight w s and structural features F t The importance weight w t , the calculation process is: w s =softmax(W s1 F s +b s1 ), w t =softmax(W t1 F t +b t1 ), where W s1 , W t1 is the weight matrix, b s1 , b t1 is the bias vector, and the softmax function converts the calculation result into a probability distribution form with a value range between 0 and 1, and w s +w t =1, the fused feature F is calculated by the formula F=w s F s +w t F t get.

[0016] Furthermore, in step S4, each residual block consists of two convolutional layers and a skip connection. Suppose the feature input to the residual block is x, and the output after the first convolutional layer is y1=f(x*W1+b1), where W1 is the convolution kernel of the first convolutional layer, the dimension is adapted to the input feature x, b1 is its bias, f is the ReLU function, and the output after the second convolutional layer is y2=f(y1*W2+b2), W2 is the convolution kernel of the second convolutional layer, b2 is its bias, and the skip connection directly adds the input x to y2 to obtain the output y=y2+x of the residual block. At the same time, a batch normalization operation is introduced in the residual block to normalize the output of each convolutional layer.

[0017] Furthermore, in step S4, when training the silkworm protein function prediction model, a weighted cross entropy loss function is used to measure the difference between the model prediction result and the true label. The calculation formula of the weighted cross entropy loss function L is: Among them, N is the number of training samples, C is the number of silkworm protein functional categories, y ijIndicates that the true label of the i-th sample belongs to the j-th category, with a value of 0 or 1, p ij It indicates the probability that the model predicts that the i-th sample belongs to the j-th class, and its value range is between 0 and 1. ij is the weight of sample i belonging to category j, which is determined by analyzing the distribution of the number of samples in each category in the training data.

[0018] Furthermore, after the prediction is completed in step S4, the prediction results are evaluated, using the accuracy Acc, recall Rec, F1 value and Matthews correlation coefficient MCC as evaluation indicators. The calculation formula of the accuracy Acc is: The calculation formula of recall rate Rec is The calculation formula of F1 value F1 is The calculation formula of Matthews correlation coefficient (MCC) is: Among them, TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, and FN represents the number of false negative samples.

[0019] Furthermore, after completing the multimodal data acquisition in step S1, before performing data preprocessing in step S2, a data enhancement operation is added. For the silkworm protein sequence data, random insertion, deletion, and replacement of amino acid residues are adopted, combined with a simulated evolutionary algorithm, and the sequence is mutated according to a specific mutation probability. For the silkworm protein structure data, random rotation and translation of the protein structure are adopted, and molecular docking technology is used to simulate the interaction between proteins and other molecules according to a specific docking algorithm to generate new structural conformation data.

[0020] The beneficial effects of the present invention are:

[0021] This paper proposes a multimodal fusion method for predicting silkworm protein function based on a deep residual network. This method significantly improves prediction performance through multi-step technological innovation. This method first constructs a multimodal dataset of silkworm proteins covering multiple species and developmental stages, integrating sequence data with high-resolution three-dimensional structural data, thus overcoming the limitations of single-modal information. In the preprocessing stage, a proprietary sequence cleaning algorithm, combined with molecular dynamics simulation optimization, effectively eliminates data noise and standardizes structural representations, laying the foundation for subsequent analysis. The feature extraction stage utilizes customized CNN and GraphSAGE networks to mine sequence patterns and structural associations, respectively. CNN captures local sequence features using odd-numbered convolution kernels and ReLU activation functions, while GraphSAGE utilizes a neighborhood aggregation mechanism to dynamically fuse node and neighborhood information. An innovative adaptive weight fusion strategy dynamically balances modal contributions through learnable weights, offering greater flexibility than traditional fixed-weight methods. Deep residual networks address the gradient degradation problem through skip connections and batch normalization techniques. Combined with a weighted cross-entropy loss function, they effectively mitigate the impact of sample imbalance and improve the prediction accuracy of rare functional categories. The data augmentation phase employs sequence variation and structural perturbation techniques to expand the diversity of training data and enhance model generalization. Multi-dimensional evaluation metrics, including precision, recall, F1 value, and Matthews correlation coefficient, are employed to ensure comprehensive quantification of model performance. This method, through comprehensive process optimization, achieves high accuracy and robustness in predicting protein function in silkworms, providing a highly effective tool for silkworm molecular breeding and functional genomics research. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flowchart of the multimodal fusion method for silkworm protein function prediction based on deep residual network provided by the present invention. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0024] like Figure 1 As shown in FIG, a multimodal fusion method for silkworm protein function prediction based on a deep residual network includes the following steps:

[0025] Step S1. Multimodal data acquisition: Using data crawling and interface calling technology from the silkworm genome database and protein crystal structure database, sequence data and high-resolution three-dimensional structure data of silkworm proteins are collected. Protein data of different silkworm species and developmental stages are classified and annotated to construct a multimodal data source of proteins.

[0026] Specifically, this step integrates authoritative database resources (such as the silkworm genome library and protein crystal structure library) and adopts automated data crawling and API interface calling technology to achieve efficient collection of silkworm protein sequence data and high-resolution three-dimensional structure data. In particular, protein data of different silkworm varieties (such as Lepidoptera model species) and different developmental stages (eggs, larvae, pupae, and adults) are classified and annotated to ensure comprehensive coverage of data sources. By constructing a multimodal dataset containing sequence primary structure, spatial conformation, and functional annotations, multi-dimensional feature input is provided for subsequent analysis, solving the problem of insufficient information in single-modality data and laying a data foundation for model training.

[0027] Step S2. Data preprocessing: Protein sequence data is cleaned using a sequence cleaning algorithm, filtering out noise and erroneous sequences through regular expression matching and statistical models, and converted into digital format using one-hot encoding. For protein structure data, molecular dynamics simulation software is used to optimize the structure according to specific force field parameters, and the protein is represented as a graph structure.

[0028] Specifically, the sequence data processing uses a self-developed cleaning algorithm, which matches abnormal sequence patterns (such as too short length, non-standard amino acids) through regular expressions, combines statistical models to identify low-quality data, and effectively removes noise. The cleaned sequence is then converted into a one-hot encoding form, and the 20 natural amino acids are mapped into a 20-dimensional binary vector to retain the precise information of the amino acid composition. Structural data processing uses molecular dynamics simulation software (such as GROMACS) to optimize the crystal structure based on the CHARMM force field parameters, eliminate unreasonable conformations such as atomic clashes, and convert the optimized structure into a graph structure representation, with each amino acid residue as a node and the interactions between residues (such as hydrogen bonds, van der Waals forces) as edges, providing structured input for graph neural network processing.

[0029] Step S3. Feature extraction and fusion: Use convolutional neural networks to extract features from sequence data. Build a graph neural network based on GraphSAGE for structural feature extraction. Utilize node and edge relationships to obtain features. Through an adaptive weight fusion strategy, calculate the weights of each modality feature based on a small fully connected neural network and dynamically fuse them into a unified feature vector.

[0030] Specifically, sequence feature extraction uses a customized CNN network, which captures local patterns in amino acid sequences (such as functional domains and motifs) by alternating stacking of multiple layers of 3×1 / 5×1 convolution kernels (step size 1) and ReLU activation functions. The output of the convolution layer is reduced in dimension by maximum pooling (window 2×1) to generate a sequence feature vector. Structural feature extraction uses the GraphSAGE graph neural network, which dynamically fuses the node's own features with the average of the neighboring features through a neighborhood aggregation mechanism to learn the spatial interaction pattern between residues. An adaptive weight mechanism is introduced in the fusion stage, and the importance weights of sequence and structural features are calculated separately through a small fully connected network (dynamically adjusted between 0-1), and the weighted sum is used to generate a fused feature vector, solving the problem that traditional fixed weight fusion cannot adapt to data modality differences.

[0031] Step S4. Model construction and prediction: Build a deep residual network, use fusion features to train the network, build a silkworm protein function prediction model, predict silkworm proteins with unknown functions, and output functional categories and confidence levels.

[0032] Specifically, the deep residual network contains 10-20 residual blocks, each of which consists of two 3×3 convolutional layers, a batch normalization layer, and a skip connection to solve the gradient vanishing problem in deep network training. The network input is a fused feature vector, and the output layer predicts 5-10 categories of protein functions (such as enzyme activity, structural proteins, etc.) through the softmax function. The training process uses a weighted cross-entropy loss function, adjusts the impact of sample imbalance through category weights, and combines the Adam optimizer (learning rate 1e-4) for end-to-end training. In the prediction stage, multimodal data of the protein to be analyzed is input, and the model outputs the probability distribution of functional categories and confidence scores, supporting threshold screening (such as confidence ≥ 0.8) to improve prediction reliability. This architecture significantly improves the learning ability of complex features while maintaining computational efficiency.

[0033] Preferably, in step S2, the silkworm protein sequence data is pre-processed by one-hot encoding. Suppose the silkworm protein sequence is S=s1s2…s n , where n is the sequence length, indicating that the protein consists of n amino acid residues connected in sequence. i represents the i-th amino acid residue in the sequence, i = 1, 2, ... n. The number of common amino acid species is m, usually m = 20. After one-hot encoding, each amino acid residue s i is encoded as a vector v of length m si If s i Corresponding to the jth amino acid, the vector v si The jth element of is 1, and the rest of the elements are 0. The mathematical expression is For example, in the sequence "ATG", "A" corresponds to the first amino acid, and after one-hot encoding, vs1 =(1,0,0,…,0), "T" corresponds to the second amino acid, v s2 =(0,1,0,…,0), "G" corresponds to the third amino acid, v s3 =(0,0,1,…,0).

[0034] Preferably, in step S2, the silkworm protein structure data is preprocessed to represent the protein three-dimensional structure as a graph structure G = (V, E). Where V is a node set, and each node v i ∈V corresponds to an amino acid residue. E is an edge set, edge e ij ∈E represents amino acid residue v i and v j There is interaction between them. To mine the structural data features, calculate the degree features of the nodes. Let the node v i The degree is d(v i ), and its calculation formula is Here |V| represents the total number of nodes in the node set V. The formula means that for node v i Count of all connected edges, i.e. d(v i ) represents the node v i The number of other directly connected amino acid residues. For example, if a node v i The degree d(v i )=4, indicating that the amino acid residue interacts with the other four amino acid residues.

[0035] Preferably, in step S3, the sequence feature extraction network based on convolutional neural network (CNN) for extracting features from silkworm protein sequence data comprises multiple convolution layers and pooling layers. The convolution kernel size of the convolution layer is set to k×1, where k is an odd number, such as k=3, 5, 7, etc. Assume that the input of the lth convolution layer is X l , is a two-dimensional matrix with dimension h l ×w l ,h l Indicates the height of the input feature map (feature dimension related to the sequence), w l Represents width (corresponding to protein sequence length). The convolution kernel is W l , a two-dimensional matrix with dimensions k×w l , used for sliding convolution on the input feature map. The bias is b l , is a scalar. The convolution layer outputs Y l By formula Y l =f(X l *W l +b l ) is calculated, f is the ReLU function, the expression is The pooling layer uses maximum pooling, and the pooling window size is p×1, such as p=2, 3, etc.

[0036] Preferably, in step S3, a graph neural network structure feature extraction network based on GraphSAGE is used for extracting the feature of silkworm protein structure data. In the GraphSAGE network, for each node v i , updates its own features by aggregating neighbor node features. Let node v i The feature at layer t is The set of neighbor nodes is N(v i ). Then node v i The feature update formula at the t+1th layer is: Where σ is the activation function, such as ReLU, Sigmoid, etc. W t is the weight matrix, and its dimension matches the dimension of the feature vector. t is the bias vector. CONCAT represents the concatenation operation, which is to connect the node v i Its own features at layer t and the average value of neighbor node features Concatenate them into a new feature vector in sequence.

[0037] Preferably, in step S3, when fusing the extracted Bombyx mori protein sequence features and structural features, an adaptive weight fusion strategy is adopted. s , dimension d s , the feature extracted from the structural data is F t , dimension d t . Compute the sequence feature F through a small fully connected neural network s The importance weight w s and structural features F t The importance weight w t , the calculation process is: w s =softmax(W s1 F s +b s1 ), w t =softmax(W t1 F t +b t1 ), where W s1 , W t1 is the weight matrix, b s1 , b t1 is the bias vector, and the softmax function converts the calculation result into a probability distribution form with a value range between 0 and 1, and w s +w t = 1. The fused feature F is obtained by the formula F = w s Fs +w t F t get.

[0038] Preferably, in step S4, the deep residual network used for training contains multiple carefully designed residual blocks. Each residual block consists of two convolutional layers and a skip connection. Suppose the feature input to the residual block is x, and the output after the first convolutional layer is y1=f(x*W1+b1), where W1 is the convolution kernel of the first convolutional layer, the dimension is adapted to the input feature x, b1 is its bias, and f is the ReLU function. The output after the second convolutional layer is y2=f(y1*W2+b2), W2 is the convolution kernel of the second convolutional layer, and b2 is its bias. The skip connection directly adds the input x to y2 to obtain the output y=y2+x of the residual block. At the same time, a batch normalization operation is introduced in the residual block to normalize the output of each convolutional layer.

[0039] Preferably, in step S4, when training the silkworm protein function prediction model, a weighted cross entropy loss function is used to measure the difference between the model prediction result and the true label. The calculation formula of the weighted cross entropy loss function L is: Where N is the number of training samples and C is the number of silkworm protein functional categories. ij Indicates that the i-th sample belongs to the true label of the j-th class, and the value is 0 or 1. ij It indicates the probability that the model predicts that the i-th sample belongs to the j-th class, and its value range is between 0 and 1. ij is the weight of sample i belonging to category j, which is determined by analyzing the distribution of the number of samples in each category in the training data.

[0040] Preferably, after the prediction is completed in step S4, the prediction result is evaluated. Accuracy Acc, recall Rec, F1 value F1 and Matthews correlation coefficient (MCC) are used as evaluation indicators. The calculation formula of accuracy Acc is The calculation formula of recall rate Rec is The calculation formula of F1 value F1 is The calculation formula of Matthews correlation coefficient (MCC) is: Among them, TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, and FN represents the number of false negative samples.

[0041] Preferably, after completing multimodal data acquisition in step S1 and before performing data preprocessing in step S2, data enhancement is performed. For Bombyx mori protein sequence data, random insertion, deletion, and substitution of amino acid residues are employed in conjunction with a simulated evolutionary algorithm to mutate the sequence according to specific mutation probabilities. For Bombyx mori protein structure data, random rotation and translation of the protein structure are employed, utilizing molecular docking technology and specific docking algorithms to simulate protein interactions with other molecules and generate new structural conformational data.

[0042] This data enhancement phase uses sequence variation and structural perturbation techniques to expand the diversity of training data before data preprocessing. Specifically, sequence data enhancement employs a three-pronged strategy: first, amino acid insertions (randomly selected from 20 natural amino acids), deletions, or substitutions are randomly selected at sequence positions with a 5%-15% probability of variation. Second, a simulated evolutionary algorithm is used to generate evolutionary variant sequences based on the evolutionary tree of silkworm protein sequences and substitution models such as the JTT matrix. Finally, the biological plausibility of the variant sequences is ensured by controlling the number of inserted / deleted amino acids (e.g., 1-3 residues per operation) and the conservatism of the substitutions (e.g., preferentially replacing amino acids of the same class). Structural data enhancement includes two parts: spatial transformation and molecular docking. First, the three-dimensional structure is randomly rotated (±15° around the x / y / z axes) and translated (1-5 angstroms) to simulate the conformational fluctuations of the protein in solution. Second, molecular docking software (such as AutoDock) is used to dock the target protein with randomly selected ligand molecules (such as metabolites and nucleic acid fragments) to generate new complex conformations. Conformational diversity is ensured by adjusting docking parameters (such as a search space radius of 20 angstroms and 1 million energy evaluations). By introducing biologically relevant perturbations, this data enhancement technology effectively improves the model's adaptability to sequence variations and conformational changes, prevents overfitting, and enables the model to learn more universal protein functional characteristics. In particular, it enhances the prediction ability for rare functional categories and atypical conformations.

[0043] During the data acquisition stage, this method uses efficient technology to accurately collect silkworm protein sequences and high-resolution three-dimensional structural data from authoritative multi-source databases, and carefully classifies and annotates them according to different varieties and developmental stages. The comprehensive and high-quality multimodal data source constructed greatly enriches the data dimensions, provides a solid foundation for subsequent analysis, and overcomes the disadvantage of insufficient information from a single data source.

[0044] During data preprocessing, the self-developed sequence cleaning algorithm combines regular expressions and statistical models to effectively filter out noise and erroneous sequences, and one-hot encoding retains the precise information of the amino acid sequence. With the help of molecular dynamics simulation software, the protein structure is optimized according to specific force field parameters, unreasonable conformations are eliminated and converted into graph structures, significantly improving data quality and availability.

[0045] Feature extraction and fusion are highly innovative. A customized convolutional neural network, through carefully configured parameters such as kernel size and step length, can deeply mine key sequence patterns. A GraphSAGE-based graph neural network leverages node-edge relationships to effectively capture protein structural features. An innovative adaptive weight fusion strategy utilizes a small, fully connected neural network to dynamically calculate and fuse the weights of each modal feature. This allows the model to flexibly adjust to different data, resulting in greater accuracy and efficiency than fixed-weight fusion.

[0046] The model is constructed using a deep residual network. Residual blocks combine skip connections with batch normalization to address the vanishing gradient problem in deep networks and enhance the ability to learn complex features. A weighted cross-entropy loss function is used during training, with weights determined based on the number of sample categories. This effectively addresses category imbalance and improves the accuracy of predictions for rare functional categories.

[0047] After the prediction is complete, a comprehensive evaluation is conducted using multiple metrics, including accuracy, recall, F1 value, and Matthews correlation coefficient, to ensure accurate quantification of model performance. Furthermore, the data enhancement phase utilizes various innovative methods for sequence and structural data, such as random sequence mutation combined with simulated evolutionary algorithms, and random structural transformation combined with molecular docking techniques. This expands data diversity, effectively prevents overfitting, and significantly improves the model's generalization and robustness. Overall, this method comprehensively improves the accuracy, reliability, and adaptability of silkworm protein function prediction, providing a powerful and effective tool for related research and applications.

[0048] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. A multimodal fusion method for silkworm protein function prediction based on deep residual network, characterized by: include: S1. Using data crawling and interface call technology, we collected sequence data and high-resolution 3D structure data of silkworm proteins from the silkworm genome database and protein crystal structure database. We also classified and annotated protein data from different silkworm species and developmental stages to construct a multimodal protein data source. S2. We use a sequence cleaning algorithm on protein sequence data, filtering out noise and erroneous sequences through regular expression matching and statistical models, and converting them into digital format using one-hot encoding. For protein structural data, we use molecular dynamics simulation software to optimize the structure according to specific force field parameters and represent the protein as a graph. S3. Use convolutional neural networks to extract features from sequence data. Build a graph neural network based on GraphSAGE for structural feature extraction. Utilize node and edge relationships to obtain features. Through an adaptive weight fusion strategy, calculate the weights of each modal feature based on a small fully connected neural network and dynamically fuse them into a unified feature vector. S4. Build a deep residual network, use fusion features to train the network, construct a silkworm protein function prediction model, predict silkworm proteins with unknown functions, and output functional categories and confidence levels.

2. The multimodal fusion method for silkworm protein function prediction based on deep residual network according to claim 1, characterized in that: In step S2, the silkworm protein sequence data is pre-processed by one-hot encoding, and the silkworm protein sequence is S=s1s2…s n , where n is the sequence length, representing that the protein is composed of n amino acid residues connected in sequence, s i represents the i-th amino acid residue in the sequence, i = 1, 2, ... n, the number of amino acid species is m, m = 20, after one-hot encoding, each amino acid residue s i is encoded as a vector v of length m si , if s i Corresponding to the jth amino acid, the vector v si The jth element of is 1, and the rest of the elements are 0. The mathematical expression is 3. The multimodal fusion method for silkworm protein function prediction based on deep residual network according to claim 1, characterized in that: In step S2, the protein structure data of silkworm is preprocessed and the three-dimensional structure of the protein is represented as a graph structure G = (V, E), where V is a node set and each node v i ∈V corresponds to an amino acid residue, E is an edge set, and edge e ij ∈E represents amino acid residue v i and v j There is interaction between them. By mining the structural data features and calculating the degree features of the nodes, let the node v i The degree is d(v i ), the calculation formula is |V| represents the total number of nodes in the node set V. The formula means that for node v i The count of all connected edges, d(v i ) represents the node v i The number of other directly connected amino acid residues.

4. The multimodal fusion method for silkworm protein function prediction based on deep residual network according to claim 1, characterized in that: In step S3, a sequence feature extraction network based on a convolutional neural network for extracting features from silkworm protein sequence data is constructed, which includes multiple convolutional layers and pooling layers. The convolution kernel size of the convolutional layer is set to k×1, where k is an odd number. The input of the lth convolutional layer is X l , is a two-dimensional matrix with dimension h l ×w l ,h l Indicates the input feature map height, w l Represents width, the convolution kernel is W l , a two-dimensional matrix with dimensions k×w l , used for sliding convolution on the input feature map, with a bias of b l , is a scalar, the convolution layer outputs Y l By formula Y l =f(X l *W l +b l ) is calculated, f is the ReLU function, the expression is The pooling layer uses maximum pooling, and the pooling window size is p×1.

5. The multimodal fusion method for silkworm protein function prediction based on deep residual network according to claim 1, characterized in that: In step S3, a graph neural network structure feature extraction network based on GraphSAGE is used to extract the feature of silkworm protein structure data. In the GraphSAGE network, for each node v i , update its own features by aggregating neighbor node features, let node v i The feature at layer t is The set of neighbor nodes is N(v i ), then node v i The feature update formula at the t+1th layer is: Among them, σ is the activation function, W t is the weight matrix, the dimension of which matches the dimension of the feature vector, b t is the bias vector, CONCAT represents the concatenation operation, which is to connect the node v i Its own features at layer t and the average value of neighbor node features Concatenate them into a new feature vector in sequence.

6. The multimodal fusion method for silkworm protein function prediction based on deep residual network according to claim 1, characterized in that: In step S3, when fusing the extracted silkworm protein sequence features and structural features, an adaptive weight fusion strategy is adopted, and the features extracted from the sequence data are set as F s , dimension d s , the feature extracted from the structural data is F t , dimension d t , calculate the sequence feature F through a small fully connected neural network s The importance weight w s and structural features F t The importance weight w t , the calculation process is: w s =softmax(W s1 F s +b s1 ), w t =softmax(W t1 F t +b t1 ), where W s1 , W t1 is the weight matrix, b s1 , b t1 is the bias vector, and the softmax function converts the calculation result into a probability distribution form with a value range between 0 and 1, and w s +w t =1, the fused feature F is calculated by the formula F=w s F s +w t F t get.

7. The multimodal fusion method for silkworm protein function prediction based on deep residual network according to claim 1, characterized in that: In step S4, each residual block consists of two convolutional layers and a skip connection. Suppose the feature input to the residual block is x, and the output of the first convolutional layer is y1=f(x*W1+b1), where W1 is the convolution kernel of the first convolutional layer, the dimension is adapted to the input feature x, b1 is its bias, f is the ReLU function, and the output of the second convolutional layer is y2=f(y1*W2+b2), W2 is the convolution kernel of the second convolutional layer, b2 is its bias, and the skip connection directly adds the input x to y2 to obtain the output y=y2+x of the residual block. At the same time, a batch normalization operation is introduced in the residual block to normalize the output of each convolutional layer.

8. The multimodal fusion method for silkworm protein function prediction based on deep residual network according to claim 1, characterized in that: In step S4, when training the silkworm protein function prediction model, the weighted cross entropy loss function is used to measure the difference between the model prediction result and the true label. The calculation formula of the weighted cross entropy loss function L is: Among them, N is the number of training samples, C is the number of silkworm protein functional categories, y ij Indicates that the true label of the i-th sample belongs to the j-th category, with a value of 0 or 1, p ij It indicates the probability that the model predicts that the i-th sample belongs to the j-th class, and its value range is between 0 and 1. ij is the weight of sample i belonging to category j, which is determined by analyzing the distribution of the number of samples in each category in the training data.

9. The multimodal fusion method for silkworm protein function prediction based on deep residual network according to claim 1, characterized in that: After the prediction is completed in step S4, the prediction results are evaluated, and the accuracy Acc, recall Rec, F1 value and Matthews correlation coefficient MCC are used as evaluation indicators. The calculation formula of the accuracy Acc is: The calculation formula of recall rate Rec is The calculation formula of F1 value F1 is The calculation formula of Matthews correlation coefficient (MCC) is: Among them, TP represents the number of true positive samples, TN represents the number of true negative samples, FP represents the number of false positive samples, and FN represents the number of false negative samples.

10. The multimodal fusion method for silkworm protein function prediction based on deep residual network according to claim 1, characterized in that: After completing the multimodal data acquisition in step S1, before performing data preprocessing in step S2, a data enhancement operation is added. For the silkworm protein sequence data, random insertion, deletion, and replacement of amino acid residues are adopted, combined with a simulated evolutionary algorithm, and the sequence is mutated according to a specific mutation probability. For the silkworm protein structure data, random rotation and translation of the protein structure are adopted, and molecular docking technology is used to simulate the interaction between proteins and other molecules according to the docking algorithm to generate new structural conformation data.

Images