A protein interaction prediction method based on sequence-structure dual-modal cross-protein attention fusion

Abstract

The present application belongs to the technical field of bioinformatics, and particularly relates to a protein interaction prediction method based on sequence-structure bimodal cross-protein attention fusion. Amino acid sequences and protein skeleton structure information of a protein pair are obtained; a protein language model is used to extract residue-level sequence representation; a structure coding model is used to extract residue-level structure representation; the sequence representation and the structure representation are projected to a shared hidden space, and residue fusion is performed within the same protein; a bidirectional cross-protein attention mechanism is used to model the residue-level interaction dependence between proteins; residue importance weights are generated based on the attention graph, and adaptive attention pooling is performed; protein pair interaction representation is constructed, and a classifier is used to output protein interaction prediction results. The present application simultaneously fuses protein sequence information and structure information, can explicitly capture residue-level interaction patterns between protein pairs, and improves the accuracy and interpretability of protein interaction prediction.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics technology, specifically relating to a protein interaction prediction method based on sequence-structure bimodal cross-protein attention fusion. Background Technology

[0002] Protein-protein interactions (PPIs) play a crucial role in cell signal transduction, transcriptional regulation, metabolic regulation, immune responses, viral infection, and disease development. Accurate identification of these interactions is essential for understanding protein function, constructing biomolecular interaction networks, elucidating disease mechanisms, and aiding in drug target discovery.

[0003] Traditional methods for detecting protein-protein interactions mainly include yeast two-hybrid assays, immunoprecipitation, affinity purification mass spectrometry, and protein microarrays. While these methods provide important evidence for protein-protein interaction studies, they typically suffer from drawbacks such as long experimental cycles, high costs, limited coverage, and high false positive or false negative rates. Therefore, developing efficient, accurate, and interpretable computational prediction methods has become an important direction in protein-protein interaction research.

[0004] In recent years, protein interaction prediction methods based on machine learning and deep learning have received widespread attention. Existing methods typically utilize protein sequence features, physicochemical properties, evolutionary information, structural information, or biological network information for modeling. Among these, protein language models can learn contextual and evolutionary information from large-scale protein sequences, generating effective residue-level sequence representations; simultaneously, protein three-dimensional structures can provide geometric information closely related to molecular binding and spatial conformation. However, many existing methods still primarily rely on single-modal features, or simply fuse sequence and structural information through simple splicing, failing to fully utilize the complementarity between the two.

[0005] Furthermore, protein-protein interactions are typically determined by local residue dependencies on protein surfaces or interfaces, rather than solely by the global representations of the two proteins. Existing methods, when modeling protein-protein pair relationships, simply concatenate the representations of the two proteins and classify them, failing to explicitly capture the residue-level interaction dependencies between the two proteins and providing interpretable information related to potential binding regions. Therefore, it is necessary to design a protein-protein interaction prediction method that can integrate bimodal information on protein sequence and structure and explicitly model the residue-level interactions of protein pairs. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a protein interaction prediction method based on sequence-structure dual-modal cross-protein attention fusion to address the shortcomings of the existing technology, thereby solving the problems of insufficient fusion of sequence and structural information, limited ability of proteins to model residue-level interaction relationships, and insufficient interpretability of prediction results in the existing protein interaction prediction methods.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] The first aspect of this invention discloses a protein interaction prediction method based on sequence-structure bimodal cross-protein attention fusion, comprising the following steps:

[0009] S1. Obtain the amino acid sequence information and protein backbone structure information of the two proteins in the protein pair to be predicted;

[0010] S2. Encode the amino acid sequence information of each of the two proteins using a protein language model, and extract the residue-level sequence characterization of each protein.

[0011] S3. Using a structural coding model, the amino acid sequence information and protein backbone structure information of the two proteins are encoded respectively, and the residue-level structural characterization of each protein is extracted.

[0012] S4. Project the residue-level sequence representation and residue-level structure representation of the two proteins respectively onto a shared latent space of the same dimension to obtain the sequence latent representation and structure latent representation of the two proteins.

[0013] S5. Align the sequence and structural hidden representations of the two proteins after projection, and perform intraprotein fusion by residual addition and layer normalization to obtain the fusion residue-level protein representations of the two proteins respectively.

[0014] S6. Using the fusion residue level protein representation of one protein in a protein pair as the query vector and the fusion residue level protein representation of the other protein as the bond vector and value vector, construct a bidirectional cross-protein attention module to obtain the context-enhanced residue representations of the two proteins and the corresponding cross-protein attention graphs.

[0015] S7. Calculate the importance score of each residue based on the cross-protein attention map, and then use the temperature-scaled softmax function to obtain the residue-level pooling weights of the two proteins.

[0016] S8. Use residue-level pooling weights to perform weighted summation on the context-enhanced residue representations to obtain the protein-level representations of the two proteins respectively.

[0017] S9. Construct a protein-protein pair interaction characterization based on the protein-level representation of the two proteins, and input the protein-protein pair interaction characterization into a multilayer perceptron classifier to output the protein-protein interaction prediction results.

[0018] In S1, the amino acid sequence information has a length of The collection of amino acid residues in a protein; the protein backbone structure information is of length [missing information]. Each residue in the protein retains and The dimension formed by the three-dimensional coordinates of the four main chain atoms is The skeleton coordinate tensor.

[0019] In some embodiments of the present invention, for ease of formula expression, the proteins to be predicted are named protein A and protein B, respectively, for the first and second proteins. It should be noted that protein A and protein B do not represent specific protein names, protein categories, or fixed order.

[0020] In some embodiments of the present invention, in S1, the amino acid sequences of protein A and protein B are respectively represented as follows: and :

[0021]

[0022] in, This indicates the first protein A. One amino acid residue, This indicates the first protein B. One amino acid residue, and These represent the residue lengths of protein A and protein B, respectively.

[0023] In some embodiments of the present invention, in S1, the protein backbone structure information of protein A and protein B are respectively represented as follows: and :

[0024]

[0025]

[0026] in, and Let these represent the skeletal coordinate tensors of protein A and protein B, respectively. This represents the residue corresponding to the i-th residue in protein A. The set of three-dimensional coordinates of the four main chain atoms, This represents the residue corresponding to the j-th residue in protein B. The set of three-dimensional coordinates of the four main chain atoms, and These represent the number of residues in protein A and protein B, respectively. 4 indicates that each residue retains 4 main chain atoms, and 3 indicates that each atom... Three-dimensional coordinate dimension.

[0027] In S2, the protein language model is an ESM-2 protein language model with frozen parameters; the protein language model has a length of The protein sequence is encoded, and the output dimension is... The residue-level sequence characterization.

[0028] In some embodiments of the present invention, in S2, the residue-level sequence characterization is specifically represented as follows:

[0029]

[0030] in, Residue-level sequence characterization of proteins, ESM-2 protein language model representing freezing parameters, This represents the protein sequence.

[0031] In S3, the structure encoding model is a ProteinMPNN encoder with frozen parameters; the structure encoding model is for lengths of... The amino acid sequence information and protein backbone structure information of the protein are encoded, and the output dimension is... The residue-level structure characterization.

[0032] In some embodiments of the present invention, in S3, the residue-level structural characterization is specifically represented as follows:

[0033]

[0034] in, Representation of residue-level structural characterization, This represents the ProteinMPNN encoder with frozen parameters. Represents the protein sequence. This represents the protein backbone coordinate tensor.

[0035] In S4, the residue-level sequence representation and residue-level structure representation are projected to 256 dimensions through linear transformation, layer normalization, and GELU activation function, respectively.

[0036] In some embodiments of the present invention, in S4, the residue-level sequence characterization undergoes linear transformation, layer normalization, and... The activation function maps the 1280-dimensional structure to 256-dimensional structure; the residue-level structure representation undergoes linear transformation, layer normalization, and... After activation, the 128-dimensional mapping is transformed to 256-dimensional, thus placing sequence information and structural information in a shared latent space of the same dimension, specifically as follows:

[0037]

[0038]

[0039] in, This represents the implicit representation of the projected sequence. This represents the implicit representation of the structure after projection. Presentation layer normalization operation, This represents the weight matrix of the sequence modality projection layer. This represents the bias vector of the sequence modality projection layer. This represents the weight matrix of the structural modal projection layer. This represents the bias vector of the structural modal projection layer.

[0040] In some embodiments of the present invention, in S5, the residue-level protein representations of protein A and protein B are specifically represented as follows:

[0041]

[0042]

[0043] in, and They represent proteins respectively. Residue-level protein representation after fusion with protein B, and They represent proteins respectively. and protein Implicit representation of the projected sequence. and They represent proteins respectively. and protein Implicit representation of the structure after projection. Indicates a fill or truncation operation. Presentation layer normalization operation.

[0044] In S6, the bidirectional cross-protein attention module includes an attention direction from protein A to protein B and an attention direction from protein B to protein A. In the first direction, the fused residue-level protein representation of protein A is used as the query vector, and the residue-level protein representation of protein B is used as the bond vector and value vector. In the second direction, the fused residue-level protein representation of protein B is used as the query vector, and the fused residue-level protein representation of protein A is used as the bond vector and value vector.

[0045] In some embodiments of the present invention, in S6, the query vector, key vector, and value vector corresponding to protein A and protein B are specifically represented as follows:

[0046]

[0047]

[0048] in, and Let represent the query vector, key vector, and value vector of protein A under the m-th attention head, respectively. and Let them represent the query vector, key vector, and value vector of protein B under the m-th attention head, respectively; and Let represent the query, key, and value linear mapping weight matrices corresponding to the m-th attention head, respectively.

[0049] In some embodiments of the present invention, in S6, for the direction from protein A to protein B, the calculation is as follows:

[0050]

[0051]

[0052] For the direction from protein B to protein A, the calculation is as follows:

[0053]

[0054]

[0055] in, and They represent the first Protein under attention To protein Direction and protein To protein Cross-protein attention map of directions, and They represent the first Protein under attention and protein Context-enhanced residue representation, This represents the normalization exponential function, used to normalize the inter-residue attention scores into attention weights; The feature dimension of a single attention head. This is used to scale the dot product attention score to avoid the attention score value being too large; Indicates the first One point of attention.

[0056] The bidirectional cross-protein attention module employs a multi-head attention mechanism with a hidden dimension of 256, 4 attention heads, and a dimension of 64 for each attention head.

[0057] In some embodiments of the present invention, in S6, the outputs of multiple attention heads are concatenated and linearly transformed to obtain the final context-enhanced residue representation, specifically as follows:

[0058]

[0059]

[0060] Where M represents the number of attention heads, The weight matrix represents the linear transformation of the multi-head attention output, used to map the concatenated multi-head attention representation back to the preset hidden dimension. This represents the operation of concatenating the outputs of multiple attention heads along the feature dimension.

[0061] In S7, the cross-protein attention maps obtained from multiple attention heads are averaged, and the importance score of each residue in the target protein is calculated based on the attention maps in opposite directions. The residue importance score of protein A is obtained by summing the attention maps from protein B to protein A; the residue importance score of protein B is obtained by summing the attention maps from protein A to protein B.

[0062] In some embodiments of the present invention, in S7, the residue importance scores of protein A and protein B are specifically represented as follows:

[0063]

[0064]

[0065]

[0066]

[0067] in, and These represent the importance scores of each residue in protein A and protein B, respectively.

[0068] In some embodiments of the present invention, in S7, the residue importance score is input into a temperature-scaled softmax function to obtain the pooling weight corresponding to each residue of protein A and protein B. This pooling weight is used to measure the contribution of each residue to the protein-protein interaction prediction, specifically expressed as follows:

[0069]

[0070]

[0071] in, and These represent the residue-level pooling weights of protein A and protein B, respectively. This represents the pooling temperature parameter.

[0072] In some embodiments of the present invention, in S8, the protein levels of protein A and protein B are represented as follows:

[0073]

[0074]

[0075] in, and These represent the protein-level representations of protein A and protein B, respectively.

[0076] In S9, the protein pair interaction representation is composed of the protein-level representations of the two proteins, the element-wise product of the two protein-level representations, the absolute difference between the two protein-level representations, and the cosine similarity of the two protein-level representations.

[0077] In some embodiments of the present invention, in S9, the protein pair interaction of protein A and protein B is characterized as follows:

[0078]

[0079] in, This represents the interaction between proteins.

[0080] In some embodiments of the present invention, in S10, the multilayer perceptron classifier includes a linear layer, a batch normalization layer, a ReLU activation function, and a Dropout layer connected in sequence, and finally outputs the prediction score of protein pairs interacting, specifically expressed as follows:

[0081]

[0082]

[0083] in, This represents the predicted score output by the classifier. This represents the predicted probability of proteins interacting. This represents the interaction characterization of the protein pairs constructed in step S9. This indicates that the multilayer perceptron classifier mapping function is used to map protein-protein pair interaction representations to interaction prediction scores. express An activation function is used to map the predicted score to the [0,1] interval to obtain the probability that protein pairs will interact.

[0084] In the model training process of the protein interaction prediction method, binary focus loss is used as the protein-protein interaction prediction loss, and sequence-structure contrast loss is used to align the global sequence representation and global structure representation of the same protein to enhance the consistency between sequence mode and structure mode.

[0085] In some embodiments of the present invention, the loss function is specifically expressed as:

[0086]

[0087] in, This represents the binary classification focus loss. This represents the sequence-structure contrast loss. This indicates the weight of the comparative loss.

[0088] In some embodiments of the present invention, the protein interaction prediction method provided by Benfang was validated on yeast five-fold cross-validation datasets, multi-species protein interaction datasets, and virus-host protein interaction datasets. The results show that the protein interaction prediction method based on sequence-structure bimodal cross-protein attention fusion provided by the present invention can effectively predict potential interactions between yeast proteins, and exhibits good performance in terms of accuracy, sensitivity, and overall discriminative ability. By comparing with existing prediction methods, it is demonstrated that the prediction method provided by the present invention helps improve the generalization ability of the model in cross-species protein interaction prediction tasks and has strong prediction performance.

[0089] Beneficial effects:

[0090] 1. This invention utilizes both protein amino acid sequence information and protein backbone structure information, and extracts residue-level sequence characterization and structural characterization through pre-trained protein language models and structure coding models, respectively, which can more comprehensively characterize the functional-related features of proteins.

[0091] 2. This invention, through an intraprotein sequence-structure fusion module and a bidirectional cross-protein attention module, achieves residue-level fusion of sequence and structural information and can explicitly model residue-level interaction dependencies between two proteins, thereby improving the discriminative ability of protein-protein interaction prediction.

[0092] 3. This invention calculates residue-level pooling weights based on cross-protein attention maps, enabling the model to highlight residue regions that contribute significantly to interaction prediction, thereby improving the interpretability of the prediction results and exhibiting good generalization ability. Attached Figure Description

[0093] The present invention will be further described in detail below with reference to the accompanying drawings, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0094] Figure 1 This is an overall flowchart of the protein-protein interaction prediction method based on sequence-structure bimodal cross-protein attention fusion in an embodiment of the present invention.

[0095] Figure 2 The ROC and PRC curves of the prediction method provided in this embodiment of the invention are shown on the yeast five-fold cross-validation dataset.

[0096] Figure 3 This is a graph showing the test results of the prediction method provided in this embodiment of the invention on a virus-host dataset. Detailed Implementation

[0097] The present invention will be further described in detail below with reference to specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0098] Unless otherwise specified, the experimental methods described in the following examples are conventional methods; unless otherwise specified, the reagents and materials are commercially available.

[0099] In the following embodiments, for ease of formula expression, protein A and protein B are only used to distinguish two arbitrary input proteins in the protein pair to be predicted, corresponding to the first protein and the second protein in the invention, respectively, and do not represent specific protein names, protein categories or fixed order.

[0100] Example 1:

[0101] This embodiment provides a protein-protein interaction prediction method based on sequence-structure bimodal cross-protein attention fusion. Figure 1 This is an overall flowchart of a protein-protein interaction prediction method based on sequence-structure bimodal cross-protein attention fusion, including the following steps:

[0102] S1. Obtain the amino acid sequence information and protein backbone structure information of the two proteins in the protein pair to be predicted, and represent the protein pair to be predicted as follows: The amino acid sequences of protein A and protein B are represented as follows: and :

[0103]

[0104] in, This indicates the first protein A. One amino acid residue, This indicates the first protein B. One amino acid residue, and These represent the number of residues in protein A and protein B, respectively.

[0105] The protein backbone structure information of protein A and protein B are represented as follows: and For protein A, its residue length is... Each residue retains the three-dimensional coordinates of the four main chain atoms: N, CA, C, and O, forming a dimension of The skeletal coordinate tensor; for protein B, its residue length is... Each residue is retained and The three-dimensional coordinates of the four main chain atoms form a dimension of Skeleton coordinate tensor:

[0106]

[0107]

[0108] in, and Let these represent the skeletal coordinate tensors of protein A and protein B, respectively. This represents the residue corresponding to the i-th residue in protein A. The set of three-dimensional coordinates of the four main chain atoms, This represents the residue corresponding to the j-th residue in protein B. The set of three-dimensional coordinates of the four main chain atoms, and These represent the number of residues in protein A and protein B, respectively. 4 indicates that each residue retains 4 main chain atoms, and 3 indicates that each atom... Three-dimensional coordinate dimension.

[0109] S2. Encode the amino acid sequence of the protein using the ESM-2 protein language model with frozen parameters, and extract the residue-level sequence for characterization of each protein. For a length of... protein sequence The ESM-2 model outputs the context sequence embedding representation corresponding to each residue, resulting in a dimension of The residue-level sequence characterization is specifically represented as follows:

[0110]

[0111] in, Residue-level sequence characterization of proteins, ESM-2 protein language model representing freezing parameters, This represents the protein sequence.

[0112] S3. Input the protein amino acid sequence and the corresponding backbone coordinate tensor into the ProteinMPNN encoder with frozen parameters to obtain the dimension. The residue-level structural characterization is specifically represented as follows:

[0113]

[0114] in, Representation of residue-level structural characterization, This represents the ProteinMPNN encoder with frozen parameters. Represents the protein sequence. This represents the protein backbone coordinate tensor.

[0115] S4. Linear projection, layer normalization, and nonlinear activation are applied to both the residue-level sequence representation and the residue-level structure representation to map them to a shared latent space of the same dimension. Since the dimension of the residue-level sequence representation output by the ESM-2 model differs from the dimension of the structure representation output by the ProteinMPNN encoder, modal projection is performed on both. Specifically, the sequence representation undergoes linear transformation, layer normalization, and... The activation function maps the 1280-dimensional structure to 256-dimensional; the structural representation undergoes linear transformation, layer normalization, and... After activation, the 128-dimensional mapping is transformed to 256-dimensional, thus placing sequence information and structural information in a shared latent space of the same dimension, specifically as follows:

[0116]

[0117]

[0118] in, This represents the implicit representation of the projected sequence. This represents the implicit representation of the structure after projection. Presentation layer normalization operation, This represents the weight matrix of the sequence modality projection layer. This represents the bias vector of the sequence modality projection layer. This represents the weight matrix of the structural modal projection layer. This represents the bias vector of the structural modal projection layer.

[0119] S5. For the sequence and structural representations of the same protein, length alignment is performed, and intraprotein fusion is achieved through residual addition and layer normalization to obtain the fused residue-level protein representation. For proteins A and B, their projected sequence and structural implicit representations are aligned in residue length. When the length of the structural implicit representation differs from that of the sequence implicit representation, the structural implicit representation is padded or truncated to match the sequence implicit representation length. Subsequently, the aligned structural implicit representation and sequence implicit representation are residually added, and layer normalization is performed to obtain the fused residue-level protein representation, specifically as follows:

[0120]

[0121]

[0122] in, and They represent proteins respectively. Residue-level protein representation after fusion with protein B, and They represent proteins respectively. and protein Implicit representation of the projected sequence. and They represent proteins respectively. and protein Implicit representation of the structure after projection. Indicates a fill or truncation operation. Presentation layer normalization operation.

[0123] S6. Represent the fusion residue level protein of protein A obtained in S5. Protein representation at the fusion residue level with protein B Input a bidirectional cross-protein attention module to model residue-level interaction dependencies between two proteins. For the m-th attention head, respectively... and Performing a linear transformation yields the query vector, key vector, and value vector for protein A and protein B, specifically represented as follows:

[0124]

[0125]

[0126] in, and Let represent the query vector, key vector, and value vector of protein A under the m-th attention head, respectively. and Let them represent the query vector, key vector, and value vector of protein B under the m-th attention head, respectively; and Let represent the query, key, and value linear mapping weight matrices corresponding to the m-th attention head, respectively.

[0127] The bidirectional cross-protein attention module includes attention calculations in two directions. In the direction from protein A to protein B, the query vector corresponding to protein A is used as the query, and the bond vector and value vector corresponding to protein B are used as the bond and value. The attention distribution of protein A to protein B residues is calculated, and the context-enhanced residue representation of protein A is obtained. In the direction from protein B to protein A, the query vector corresponding to protein B is used as the query, and the bond vector and value vector corresponding to protein A are used as the bond and value. The attention distribution of protein B to protein A residues is calculated, and the context-enhanced residue representation of protein B is obtained.

[0128] For the direction from protein A to protein B, the calculation is as follows:

[0129]

[0130]

[0131] For the direction from protein B to protein A, the calculation is as follows:

[0132]

[0133]

[0134] in, and They represent the first Protein under attention To protein Direction and protein To protein Cross-protein attention map of directions, and They represent the first Protein under attention and protein Context-enhanced residue representation, This represents the normalization exponential function, used to normalize the inter-residue attention scores into attention weights; The feature dimension of a single attention head. This is used to scale the dot product attention score to avoid the attention score value being too large; Indicates the first One point of attention.

[0135] The bidirectional cross-protein attention module employs a multi-head attention mechanism. In this embodiment, the hidden dimension is set to 256, the number of attention heads is set to 4, and the dimension of each attention head is 64. The outputs of multiple attention heads are concatenated and linearly transformed to obtain the final context-enhanced residue representation, specifically as follows:

[0136]

[0137]

[0138] Where M represents the number of attention heads, The weight matrix represents the linear transformation of the multi-head attention output, used to map the concatenated multi-head attention representation back to the preset hidden dimension. This represents the operation of concatenating the outputs of multiple attention heads along the feature dimension.

[0139] S7. Perform multi-head averaging on the bidirectional cross-protein attention maps, and calculate the importance score of each residue based on the cross-protein attention maps in the opposite direction. Then, use the temperature-scaled softmax function to obtain the residue-level pooling weights. Average the cross-protein attention maps generated by multiple attention heads to obtain the average attention map from protein A to protein B and the average attention map from protein B to protein A. Then, calculate the importance score of each residue in the target protein based on the average attention map in the opposite direction. That is, the residue importance of protein A is obtained by summing the attention maps from protein B to protein A, and the residue importance of protein B is obtained by summing the attention maps from protein A to protein B. Specifically, it is expressed as follows:

[0140]

[0141]

[0142]

[0143]

[0144] in, and These represent the importance scores of each residue in protein A and protein B, respectively.

[0145] The residue importance score is input into a temperature-scaled softmax function to obtain the pooling weight for each residue. This pooling weight is used to measure the contribution of each residue to the prediction of protein-protein interactions.

[0146]

[0147]

[0148] in, and These represent the residue-level pooling weights of protein A and protein B, respectively. This represents the pooling temperature parameter.

[0149] S8. Based on the residue-level pooling weights, the context-enhanced residue representations are weighted and summed to obtain the protein-level representations of the two proteins. Using the residue-level pooling weights obtained in step S7, the context-enhanced residue representations obtained in step S6 are weighted and summed to obtain the protein-level representations of the two proteins. This protein-level representation comprehensively reflects the protein residue features and the interactive contextual information between protein pairs, specifically as follows:

[0150]

[0151]

[0152] in, and These represent the protein-level representations of protein A and protein B, respectively.

[0153] S9. Construct a protein-protein pair interaction representation based on the protein-level representations of the two proteins. The protein-protein pair interaction representation consists of the concatenation of the protein-level representations of the two proteins, the element-wise product of the two protein-level representations, the absolute difference between the two protein-level representations, and the cosine similarity between the two protein-level representations. This method can simultaneously represent the global information, similarity information, and difference information of each protein, specifically as follows:

[0154]

[0155] in, This represents the interaction between proteins.

[0156] S10. Input the protein pair interaction representation into a multilayer perceptron (MLP) classifier and output the protein-protein interaction prediction results. Input the protein pair interaction representation into the MLP classifier to obtain the prediction score of protein pair interaction, and convert the prediction score into an interaction probability using a sigmoid function. When the interaction probability is greater than a preset threshold, it is determined that the protein pair interacts; otherwise, it is determined that the protein pair does not interact. Specifically:

[0157]

[0158]

[0159] in, This represents the predicted score output by the classifier. This represents the predicted probability of proteins interacting. This represents the interaction characterization of the protein pairs constructed in step S9. This indicates that the multilayer perceptron classifier mapping function is used to map protein-protein pair interaction representations to interaction prediction scores. express An activation function is used to map the predicted score to the [0,1] interval to obtain the probability that protein pairs will interact.

[0160] During model training, binary focus loss is used as the protein-protein interaction prediction loss to reduce the impact of easily classified samples on model training; at the same time, sequence-structure contrast loss is used to align the global sequence representation and global structure representation of the same protein, so that the sequence information and structural information of the same protein remain consistent in the representation space.

[0161]

[0162] in, This represents the binary classification focus loss. This represents the sequence-structure contrast loss. This indicates the weight of the comparative loss.

[0163] Example 2:

[0164] To verify the effectiveness of the protein-protein interaction prediction method provided in Example 1, this example uses publicly available protein interaction benchmark datasets for experimental verification. Specifically, the yeast protein interaction dataset is derived from the yeast PPI benchmark dataset constructed from the DIP (Database of Interacting Proteins) database, where positive samples are experimentally verified yeast protein interaction pairs, and negative samples are randomly paired and screened non-interacting protein pairs; the multi-species protein interaction dataset is derived from the publicly available PIPR multi-species PPI benchmark dataset, containing protein interaction data from different species such as Caenorhabditis elegans, Escherichia coli, and Drosophila melanogaster, and is divided according to different sequence similarity thresholds; the virus-host protein interaction dataset is derived from the publicly available human-virus PPI dataset, where positive samples are integrated from publicly available databases such as BioGRID, IntAct, HPDIB, MINT, PHISTO, and VirHostNet, and are divided into C1Virus and C2Virus test settings based on whether the viral protein appears during the training phase. The yeast dataset was used to evaluate the performance of this invention in single-species protein interaction prediction tasks; the multi-species dataset was used to evaluate the generalization ability of this invention in cross-species protein interaction prediction tasks; and the virus-host dataset was used to evaluate the applicability of this invention in human host protein-virus protein interaction prediction tasks. Table 1 shows the performance comparison results of this invention on the yeast five-fold cross-validation dataset. Figure 2 The ROC and PRC curves are shown on the yeast five-fold cross-validation dataset.

[0165] Table 1. Performance comparison results on the yeast five-fold cross-validation dataset.

[0166]

[0167] Through Table 1 and in combination Figure 2 As can be seen, the present invention achieved high and stable predictive performance in the yeast five-fold cross-validation experiment, with an average accuracy of 95.98%, an average sensitivity of 96.27%, an average specificity of 95.65%, an average precision of 95.67%, an average MCC of 91.97%, and an average AUC of 98.94%. These results indicate that the present invention can effectively predict potential interactions between yeast proteins and exhibits good performance in terms of accuracy, sensitivity, and overall discriminative ability.

[0168] Example 3:

[0169] To further verify the generalization ability of the prediction method provided in Example 1 in the cross-species protein interaction prediction task, this example conducted comparative experiments with various existing protein-protein interaction prediction methods on a multi-species protein interaction dataset. Table 2 lists the performance comparison results of the prediction method (CrossPPI) provided in Example 1 and existing methods under the overall setting of a multi-species dataset.

[0170] Table 2. Performance comparison results of different methods in the overall setting of multi-species datasets.

[0171]

[0172] As shown in Table 2, under the overall setting of a multi-species dataset, this invention achieves superior overall performance compared to various existing protein-protein interaction prediction methods. Specifically, this invention achieves high results in accuracy, recall, F1 score, AUC, and MCC, indicating that it can effectively learn the interaction features between protein pairs from different species. These results demonstrate that sequence-structure bimodal feature fusion and a bidirectional cross-protein attention mechanism contribute to improving the model's generalization ability in cross-species protein interaction prediction tasks.

[0173] To further verify the generalization ability of the prediction method provided in Example 1 under low sequence similarity conditions, this example selects protein pairs with homology less than 25% from a multi-species dataset as the test setting and compares them with various existing protein-protein interaction prediction methods. Table 3 lists the performance comparison results of different prediction methods under this setting.

[0174] Table 3. Performance comparison of different prediction methods under the condition of homology less than 25% in multi-species datasets.

[0175]

[0176] As shown in Table 3, even with a low sequence similarity setting of less than 25% homology, this invention still achieved good predictive performance. The invention achieved an accuracy of 99.44%, a precision of 99.83%, a recall of 99.09%, a specificity of 99.82%, an F1 score of 99.46%, an AUC of 99.93%, and a MCC of 98.88%.

[0177] Example 4:

[0178] To further verify the applicability of the prediction method provided in Example 1 to the task of predicting human host protein-viral protein interactions, this example was experimentally verified on a virus-host protein interaction dataset. Figure 3Experimental results of the prediction method provided in Example 1 under C1Virus and C2Virus test settings are presented, wherein, Figure 3 In the figure, A represents the ROC curve under the C1Virus test settings. Figure 3 In the figure, B represents the ROC curve under the C2Virus test setting. The C1Virus test setting is used to evaluate the predictive ability for interactions involving viral proteins present during the training phase, while the C2Virus test setting is used to evaluate the predictive ability for interactions involving viral proteins not present during the training phase. Figure 3 As can be seen, the present invention achieved good prediction results under both C1Virus and C2Virus test settings, with an AUROC of 0.922 on the C1Virus test set and an AUROC of 0.836 on the C2Virus test set.

[0179] This invention provides a method and approach for predicting protein interactions based on sequence-structure bimodal cross-protein attention fusion. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A protein interaction prediction method based on sequence-structure bimodal cross-protein attention fusion, characterized in that, Includes the following steps: S1. Obtain the amino acid sequence information and protein backbone structure information of the two proteins in the protein pair to be predicted; S2. Encode the amino acid sequence information of each of the two proteins using a protein language model, and extract the residue-level sequence characterization of each protein. S3. Using a structural coding model, the amino acid sequence information and protein backbone structure information of the two proteins are encoded respectively, and the residue-level structural characterization of each protein is extracted. S4. Project the residue-level sequence representation and residue-level structure representation of the two proteins respectively onto a shared latent space of the same dimension to obtain the sequence latent representation and structure latent representation of the two proteins. S5. Align the sequence and structural hidden representations of the two proteins after projection, and perform intraprotein fusion by residual addition and layer normalization to obtain the fusion residue-level protein representations of the two proteins respectively. S6. Using the fusion residue level protein representation of one protein in a protein pair as the query vector and the fusion residue level protein representation of the other protein as the bond vector and value vector, construct a bidirectional cross-protein attention module to obtain the context-enhanced residue representations of the two proteins and the corresponding cross-protein attention graphs. S7. Calculate the importance score of each residue based on the cross-protein attention map, and then use the temperature-scaled softmax function to obtain the residue-level pooling weights of the two proteins. S8. Use residue-level pooling weights to perform weighted summation on the context-enhanced residue representations to obtain the protein-level representations of the two proteins respectively. S9. Construct a protein-protein pair interaction characterization based on the protein-level representation of the two proteins, and input the protein-protein pair interaction characterization into a multilayer perceptron classifier to output the protein-protein interaction prediction results.

2. The protein interaction prediction method according to claim 1, characterized in that, In S1, the amino acid sequence information is of length [length missing]. The collection of amino acid residues in a protein; the protein backbone structure information is of length [missing information]. Each residue in the protein retains and The dimension formed by the three-dimensional coordinates of the four main chain atoms is The skeleton coordinate tensor.

3. The protein interaction prediction method according to claim 1, characterized in that, In S2, the protein language model is a frozen parameter ESM-2 protein language model; the protein language model has a length of The protein sequence is encoded, and the output dimension is... The residue-level sequence characterization.

4. The protein interaction prediction method according to claim 1, characterized in that, In S3, the structure encoding model is a ProteinMPNN encoder with frozen parameters; the structure encoding model is for lengths of The amino acid sequence information and protein backbone structure information of the protein are encoded, and the output dimension is... The residue-level structure characterization.

5. The protein interaction prediction method according to claim 1, characterized in that, In S4, the residue-level sequence representation and residue-level structure representation are projected to 256 dimensions through linear transformation, layer normalization, and GELU activation function, respectively.

6. The protein interaction prediction method according to claim 1, characterized in that, In S6, the bidirectional cross-protein attention module is specifically defined as follows: two proteins are referred to as the first protein and the second protein, and the module includes the attention direction from the first protein to the second protein and the attention direction from the second protein to the first protein. In the first direction, the fusion residue-level protein representation of the first protein is used as a query vector, and the residue-level protein representation of the second protein is used as a bond vector and a value vector; in the second direction, the fusion residue-level protein representation of the second protein is used as a query vector, and the fusion residue-level protein representation of the first protein is used as a bond vector and a value vector.

7. The protein interaction prediction method according to claim 6, characterized in that, The bidirectional cross-protein attention module adopts a multi-head attention mechanism with a hidden dimension of 256, 4 attention heads, and a dimension of 64 for each attention head.

8. The protein interaction prediction method according to claim 1, characterized in that, In S7, the cross-protein attention maps obtained from multiple attention heads are averaged, and the importance score of each residue in the target protein is calculated based on the attention maps in opposite directions. The two proteins are denoted as the first protein and the second protein, respectively. The residue importance score of the first protein is obtained by summing the attention maps from the second protein to the first protein. The residue importance score of the second protein is obtained by summing the attention maps from the first protein to the second protein.

9. The protein interaction prediction method according to claim 1, characterized in that, In S9, the protein pair interaction characterization is composed of the protein-level representations of the two proteins, the element-wise product of the two protein-level representations, the absolute difference between the two protein-level representations, and the cosine similarity of the two protein-level representations.

10. The protein interaction prediction method according to claim 1, characterized in that, The protein-protein interaction prediction method employs binary focus loss as the protein-protein interaction prediction loss during model training, and uses sequence-structure comparison loss to align the global sequence representation and global structure representation of the same protein.

Images