Protein-polypeptide binding residue prediction method and system based on pre-training module and fourier network
By using pre-trained modules and Fourier networks, the problems of insufficient utilization of long-range contextual features and sample imbalance in protein-peptide binding residue prediction are solved, achieving higher prediction accuracy and reliability, which is applicable to bioinformatics and drug design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DALIAN UNIV
- Filing Date
- 2026-05-20
- Publication Date
- 2026-07-21
AI Technical Summary
Existing methods for predicting protein-peptide binding residues suffer from problems such as insufficient utilization of long-range contextual features of protein sequences, inadequate extraction of residue order and frequency information, and prediction bias caused by imbalance in binding residue samples.
We employ a method based on pre-trained modules and Fourier networks to extract context-dependent features and evolutionary structural latent features of protein sequences using ProtBert-BFD and ESM models. We combine Fourier layers for feature extraction and construct a supervised contrastive learning module to alleviate the sample imbalance problem.
It improves the accuracy and reliability of protein-peptide binding residue prediction, enabling better identification of binding residues, and is applicable to the fields of bioinformatics and drug design.
Smart Images
Figure CN122435992A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of protein sequence analysis technology, specifically to a method and system for predicting protein-peptide binding residues based on a pre-trained module and a Fourier network. Background Technology
[0002] Protein-peptide interactions are a core component of critical life activities such as cell signal transduction, disease mechanism analysis, and targeted drug discovery. Accurate identification of residues in protein sequences that specifically bind to peptides is of significant theoretical and practical value for elucidating protein functional mechanisms, efficiently assisting in drug molecule design, and targeted screening of potential drug action sites. While traditional wet-lab-based methods for detecting binding residues can provide reliable experimental validation results, they generally suffer from drawbacks such as high detection costs, long experimental cycles, and complex procedures, making it difficult to meet the demands for high-throughput and rapid screening of large-scale protein sequences.
[0003] Current methods for predicting protein-peptide binding residues are mainly divided into two categories: structure-based prediction methods and sequence-based prediction methods. Structure-based methods heavily rely on structural properties such as the accessible surface area, secondary structure, and spatial conformation of proteins. However, peptide molecules are characterized by their small size, weak binding affinity, and high structural flexibility, making the analysis and acquisition of the three-dimensional structure of protein-peptide complexes difficult and with low success rates, severely limiting the practical applicability of these methods. Sequence-based methods can overcome the strong dependence on the three-dimensional structure of protein-peptide complexes and have a wider range of applications. However, most existing methods only extract features from a single dimension such as amino acid composition, physicochemical properties, and evolutionary information, failing to fully explore and characterize the long-range contextual dependencies, implicit structural information, and residue order and frequency distribution patterns within the protein sequence. This inadequate feature expression and incomplete information utilization limit prediction accuracy and generalization ability.
[0004] In addition, protein-peptide binding residue prediction tasks generally suffer from a significant imbalance in the distribution of positive and negative samples, with the number of binding residues in actual data being far less than that of non-binding residues. During model training, this can easily cause the classification model to shift towards the dominant non-binding residue category, resulting in significant prediction bias and drastically reducing the sensitivity of binding residue site identification and the reliability of prediction results. Summary of the Invention
[0005] The purpose of this invention is to propose a protein-peptide binding residue prediction method and system based on a pre-trained module and a Fourier network, in order to solve the problems of insufficient utilization of long-range contextual features of protein sequences, insufficient extraction of residue order and frequency information, and prediction bias caused by imbalance of binding residue samples in existing methods.
[0006] According to a first aspect of the embodiments of this disclosure, a method for predicting protein-peptide binding residues based on a pre-trained module and a Fourier network is provided, comprising the following steps: The amino acid sequence information of the protein to be predicted is obtained, and the amino acid sequence is divided or padded into fixed-length sequence fragments according to a preset maximum length. Word embedding and position embedding are generated for each residue. The fixed-length sequence fragments are input into the pre-trained protein language model ProtBert-BFD and the evolutionary scale modeling pre-trained model ESM in the pre-trained embedding module, respectively, to encode the residues in the sequence fragments and obtain high-dimensional potential protein representations. The residue representations output from the ProtBert-BFD branch and the residue representations output from the ESM branch are fused, and an initial residual connection is introduced to obtain the fused residue representation. A supervised contrastive learning module is constructed, which uses bound residues and non-bound residues as different categories of samples for representation constraints. Contrastive loss is used to make the representations of residues of the same class closer to each other in the feature space and the representations of residues of different classes farther apart, so as to alleviate the problem of positive and negative sample imbalance. The fused residue representation is input into a feature extraction module, which includes at least one Fourier layer; a high-dimensional residue representation is obtained through the Fourier layer. The high-dimensional residue representation input-output module, after feature compression and classification mapping, outputs the probability value of each residue belonging to a protein-peptide binding residue; The training objective is constructed by using cross-entropy loss and contrastive loss together. The pre-trained embedding module, Fourier feature extraction module and output module are jointly optimized, and whether each residue is a protein-peptide binding residue is determined based on the predicted probability of each residue.
[0007] In one embodiment, the pre-trained protein language model ProtBert-BFD is constructed based on the bidirectional encoder representation model BERT from Transformer to capture bidirectional contextual dependencies between residues; the evolutionary scale modeling pre-trained model ESM is used to capture implicit information in protein sequences related to evolution, biochemical properties, secondary structure and tertiary structure.
[0008] In one embodiment, the contrastive learning module represents any two residues in the same batch. and Obtain the distance D between the two: Among them, cos<r1,r2> This represents the cosine similarity between two residues.
[0009] In one embodiment, a contrastive loss function is constructed based on whether residue pairs belong to the same category: in, This indicates the category relationship between residue pairs; when two residues belong to the same category... =0; when the two residues belong to different categories. =1;D max This represents the maximum distance between residue pairs of different categories.
[0010] In one embodiment, each Fourier layer maps the residue representation to the frequency domain via Fourier transform, performs a linear transformation on the frequency domain features to preserve low-frequency patterns and filter high-frequency patterns, and then restores it to the time domain via inverse Fourier transform. Finally, it combines local linear transformation, residual connections, and activation functions to obtain a high-dimensional residue representation.
[0011] In one embodiment, the cross-entropy loss and the contrastive loss together constitute the total loss function. : in, Represents cross-entropy loss, This represents the loss in supervised contrastive learning.
[0012] In one embodiment, the output module is a multi-layer fully connected network with the number of nodes being 64, 32, 16, and 2, respectively.
[0013] According to a second aspect of the present disclosure, a protein-peptide binding residue prediction system based on a pre-trained module and a Fourier network is provided, comprising: The input embedding module obtains the amino acid sequence information of the protein to be predicted, segments or completes the amino acid sequence into fixed-length sequence fragments according to a preset maximum length, and generates word embeddings and position embeddings for each residue. The pre-trained embedding module inputs the fixed-length sequence fragment into the pre-trained protein language model ProtBert-BFD and the evolutionary scale modeling pre-trained model ESM in the pre-trained embedding module, respectively, to encode the residues in the sequence fragment and obtain a high-dimensional potential protein representation. The feature fusion module fuses the residue representations output from the ProtBert-BFD branch with the residue representations output from the ESM branch, and introduces initial residual connections to obtain the fused residue representations. The supervised contrastive learning module is constructed by treating bound residues and non-bound residues as different categories of samples for representation constraints. Through contrastive loss, the representations of residues of the same class are made closer to each other in the feature space, while the representations of residues of different classes are made farther apart, so as to alleviate the problem of positive and negative sample imbalance. The Fourier feature extraction module inputs the fused residue representation into the feature extraction module, which includes at least one Fourier layer; a high-dimensional residue representation is obtained through the Fourier layer; The output module takes the high-dimensional residue representation as input to the output module. The output module includes a multi-layer fully connected network that performs feature compression and classification mapping in sequence, and outputs the probability value of each residue belonging to a protein-peptide binding residue. The joint optimization and prediction module uses cross-entropy loss and contrastive loss to form a training objective, jointly optimizes the pre-trained embedding module, Fourier feature extraction module and output module, and determines whether each residue is a protein-peptide binding residue based on the predicted probability of each residue.
[0014] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network.
[0015] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network.
[0016] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: (1) This invention uses ProtBert-BFD and ESM to extract the context-dependent features and evolutionary structure latent features of protein sequences respectively. By using a dual-branch fusion strategy, the context representation capability of the pre-trained model is fully utilized, which effectively solves the problem of insufficient utilization of long-range context features of protein sequences in existing methods.
[0017] (2) The present invention introduces a Fourier layer to construct a feature extraction module. By mapping the residue representation to the frequency domain and performing linear transformation, it can extract residue order information and frequency pattern at the same time, give full play to the multidimensional feature extraction capability of Fourier network, and make up for the shortcomings of existing methods in extracting residue order and frequency information.
[0018] (3) The present invention constructs a supervised contrastive learning module, which uses a contrastive loss function to make residues of the same class closer to each other and residues of different classes farther apart, and combines a cubic penalty term to enhance the ability to distinguish minority class binding residues, effectively alleviating the problem of positive and negative sample imbalance and improving the accuracy and reliability of binding residue prediction.
[0019] (4) This invention can make full use of the context representation ability of pre-trained protein language models, the multidimensional feature extraction ability of Fourier networks, and the category discrimination ability of contrastive learning, thereby improving the accuracy and reliability of protein-peptide binding residue prediction, and has important application value in the fields of bioinformatics and drug design. Attached Figure Description
[0020] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.
[0021] Figure 1 This is a schematic diagram of a protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network. Detailed Implementation
[0022] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.
[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0025] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.
[0026] Example 1: This embodiment provides a method for predicting protein-peptide binding residues based on a pre-trained module and a Fourier network, including the following steps: S1. Obtain the amino acid sequence information of the protein to be predicted, divide or complete the amino acid sequence into fixed-length sequence fragments according to a preset maximum length, and generate word embeddings and position embeddings for each residue; In this embodiment, the amino acid sequence information of the protein sequence fragment to be predicted is obtained. If the original protein sequence length exceeds a preset maximum length, it is divided into multiple fixed-length sequence fragments according to the preset maximum length; if the original protein sequence length is less than the preset maximum length, it is padded at the end of the sequence to meet the input length requirements of the pre-trained model. Subsequently, each residue is converted into a corresponding word segmentation tag, and a positional encoding is assigned to each position. The word segmentation embedding and positional embedding together form the input representation of the pre-trained embedding module.
[0027] S2. Input the fixed-length sequence fragment into the pre-trained protein language model ProtBert-BFD and the evolutionary scale modeling pre-trained model ESM in the pre-trained embedding module, and encode the residues in the sequence fragment to obtain a high-dimensional potential protein representation. Specifically, fixed-length sequence fragments are input into the ProtBert-BFD model and the ESM model, respectively. The ProtBert-BFD model is built based on the Bidirectional Encoder Representation Model (BERT) from Transformer to capture bidirectional contextual dependencies between residues; the ESM model is used to capture implicit information in the protein sequence related to evolution, biochemical properties, secondary structure, and tertiary structure. By encoding residues using the two pre-trained models, high-dimensional potential protein representations containing both sequence contextual information and structural-functional information can be obtained.
[0028] S3. The residue representations output from the ProtBert-BFD branch and the residue representations output from the ESM branch are fused, and an initial residual connection is introduced to obtain the fused residue representations; Specifically, the residue representations output by the ProtBert-BFD branch and the ESM branch are aligned according to their residue positions and then fused using methods such as splicing, weighted summation, or linear mapping. Simultaneously, initial residual connections are introduced during the fusion process to ensure that the fused representation retains the original embedding information of the input sequence, thereby reducing information loss that may occur during pre-trained representation conversion and deep feature transfer.
[0029] S4. Construct a supervised contrastive learning module, which uses bound residues and non-bound residues as different categories of samples for representation constraints. Through contrastive loss, the representations of residues of the same class are made closer to each other in the feature space, while the representations of residues of different classes are made farther apart, so as to alleviate the problem of positive and negative sample imbalance. In the supervised contrastive learning module, for any two residues in the same batch... and First, calculate the distance D between the two. , Its expression is: Among them, cos<r1,r2> This represents the cosine similarity between two residue pairs. Subsequently, a contrastive loss function is constructed based on whether the residue pairs belong to the same category: in, This indicates the category relationship between residue pairs; when two residues belong to the same category... =0; when the two residues belong to different categories. =1. D maxThis represents the maximum distance between residue pairs of different categories, with a preferred value of 2. Using the above loss function, residues of the same category are constrained to be close to each other, while residues of different categories are constrained to be far apart. Furthermore, a cubic penalty is applied to residue pairs of different categories to enhance the model's ability to distinguish between minority-bound residues.
[0030] S5. Input the fused residue representation into the feature extraction module, the feature extraction module including at least one Fourier layer; obtain a high-dimensional residue representation through the Fourier layer; Specifically, the input feature extraction module is fused with residue representations. For the input residue representation v, it is first mapped from the time domain to the frequency domain using a Fourier transform F to extract the frequency patterns contained in the residue sequence; then, a linear transform R is performed on the frequency domain features to learn low-frequency Fourier patterns and filter out some high-frequency noise; finally, an inverse Fourier transform F is performed... -1 The processed frequency domain representation is restored to the time domain representation; simultaneously, local features at the residue level are further extracted using a local linear transformation W. Finally, the above representation is processed by residual concatenation, summation, and activation function σ to obtain a high-dimensional residue representation. Preferably, the feature extraction module may include four cascaded Fourier layers.
[0031] S6. The high-dimensional residue representation is input to the output module. The output module includes a multi-layer fully connected network, which performs feature compression and classification mapping in sequence, and outputs the probability value of each residue belonging to a protein-peptide binding residue. Specifically, the high-dimensional residues are represented by an input-output module. The output module preferably includes a multi-layer fully connected network, with the number of network nodes set sequentially to 64, 32, 16, and 2, used to progressively compress the feature dimension and complete the classification mapping. The output module generates two predicted scores for each residue, and after normalization, obtains the probability value that the residue belongs to a protein-peptide binding residue.
[0032] S7. The training objective is jointly constructed using the cross-entropy loss and the contrast loss. The pre-trained embedding module, Fourier feature extraction module and output module are jointly optimized. The predicted probability of each residue is used to determine whether it is a protein-peptide binding residue.
[0033] Specifically, during the training phase, the residue category prediction results obtained by the output module are compared with the true labels, and the cross-entropy loss L is used. CE Constrain the residue category prediction results; simultaneously, utilize the contrastive loss L in S4. contrast Constraining the class distribution of residues in the feature space improves the separability of bound and unbound residues in the feature space. The total loss function during model training consists of cross-entropy loss and contrastive loss, and its expression is: Among them, L loss L represents the total training loss of the model. CE L represents the cross-entropy loss. contrast This represents the supervised contrastive learning loss. By simultaneously optimizing the two loss terms mentioned above, the model can improve both the accuracy of residue category prediction and the distinguishability between bound and non-bound residues in the representation space. After training, the protein sequence to be predicted is input into the model to obtain the probability value of each residue belonging to a protein-peptide binding residue, and the model determines whether the residue is a protein-peptide binding residue based on a preset threshold or an optimal threshold.
[0034] To verify the overall performance of the method of this invention, it was tested on test sets TE125 and TE639. The TE125 test set contains 125 protein sequences with a total of 30,870 residues, of which 1,719 are binding residues and 29,151 are non-binding residues. The TE639 test set contains 639 protein sequences with a total of 150,330 residues, of which 8,490 are binding residues and 141,840 are non-binding residues. As shown in Table 1, the method of this invention performs well in terms of sensitivity, specificity, precision, Matthews correlation coefficient (MCC), and area under the receiver operating characteristic (AUC), indicating that the pre-trained protein language model, supervised contrastive learning, and Fourier feature extraction module can effectively improve the prediction performance of protein-peptide binding residues.
[0035] Table 1. Performance results of the method of the present invention on different test sets. Meanwhile, to verify the performance difference between the method of this invention and existing methods, the method of this invention was compared with PepBind and PepBCL on test sets TE125 and TE639, and the evaluation indicators were consistent with those of the above. The comparison results are shown in Tables 2 and 3. The method of this invention performed better in terms of accuracy, specificity, and AUC, indicating that it has certain advantages in reducing false alarms and improving the reliability of protein-peptide binding residue prediction. As shown in Tables 2 and 3, the method of this invention achieved high accuracy, specificity, and AUC on both test sets, indicating that the method can effectively reduce false alarms and improve the reliability of protein-peptide binding residue prediction.
[0036] Table 2 Performance of the present invention and previous methods (partial) on the TE125 test set Table 3 Performance of the present invention and previous methods (partial) on the TE639 test set Example 2: This embodiment provides a protein-peptide binding residue prediction system based on a pre-trained module and a Fourier network, including: The input embedding module obtains the amino acid sequence information of the protein to be predicted, segments or completes the amino acid sequence into fixed-length sequence fragments according to a preset maximum length, and generates word embeddings and position embeddings for each residue. The pre-trained embedding module inputs the fixed-length sequence fragment into the pre-trained protein language model ProtBert-BFD and the evolutionary scale modeling pre-trained model ESM in the pre-trained embedding module, respectively, to encode the residues in the sequence fragment and obtain a high-dimensional potential protein representation. The feature fusion module fuses the residue representations output from the ProtBert-BFD branch with the residue representations output from the ESM branch, and introduces initial residual connections to obtain the fused residue representations. The supervised contrastive learning module is constructed by treating bound residues and non-bound residues as different categories of samples for representation constraints. Through contrastive loss, the representations of residues of the same class are made closer to each other in the feature space, while the representations of residues of different classes are made farther apart, so as to alleviate the problem of positive and negative sample imbalance. The Fourier feature extraction module inputs the fused residue representation into the feature extraction module, which includes at least one Fourier layer; a high-dimensional residue representation is obtained through the Fourier layer; The output module takes the high-dimensional residue representation as input to the output module. The output module includes a multi-layer fully connected network that performs feature compression and classification mapping in sequence, and outputs the probability value of each residue belonging to a protein-peptide binding residue. The joint optimization and prediction module uses cross-entropy loss and contrastive loss to form a training objective, jointly optimizes the pre-trained embedding module, Fourier feature extraction module and output module, and determines whether each residue is a protein-peptide binding residue based on the predicted probability of each residue.
[0037] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.
[0038] Example 3: An electronic device is provided for running the aforementioned "protein-peptide binding residue prediction method based on pre-trained modules and Fourier networks". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S7 of the method described in Embodiment 1, specifically including but not limited to: S1. Obtain the amino acid sequence information of the protein to be predicted, divide or complete the amino acid sequence into fixed-length sequence fragments according to a preset maximum length, and generate word embeddings and position embeddings for each residue; S2. Input the fixed-length sequence fragment into the pre-trained protein language model ProtBert-BFD and the evolutionary scale modeling pre-trained model ESM in the pre-trained embedding module, and encode the residues in the sequence fragment to obtain a high-dimensional potential protein representation. S3. The residue representations output from the ProtBert-BFD branch and the residue representations output from the ESM branch are fused, and an initial residual connection is introduced to obtain the fused residue representations; S4. Construct a supervised contrastive learning module, which uses bound residues and non-bound residues as different categories of samples for representation constraints. Through contrastive loss, the representations of residues of the same class are made closer to each other in the feature space, while the representations of residues of different classes are made farther apart, so as to alleviate the problem of positive and negative sample imbalance. S5. Input the fused residue representation into the feature extraction module, the feature extraction module including at least one Fourier layer; obtain a high-dimensional residue representation through the Fourier layer; S6. The high-dimensional residue representation is input to the output module. The output module includes a multi-layer fully connected network, which performs feature compression and classification mapping in sequence, and outputs the probability value of each residue belonging to a protein-peptide binding residue. S7. The training objective is jointly constructed using the cross-entropy loss and the contrast loss. The pre-trained embedding module, Fourier feature extraction module and output module are jointly optimized. The predicted probability of each residue is used to determine whether it is a protein-peptide binding residue.
[0039] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage (acquiring observation data and outputting inversion results) through a communication interface. The above hardware configuration does not constitute a limitation of the present invention.
[0040] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S7 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.
[0041] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.
[0042] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
[0043] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.
Claims
1. A protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network, characterized in that, Includes the following steps: The amino acid sequence information of the protein to be predicted is obtained, and the amino acid sequence is divided or padded into fixed-length sequence fragments according to a preset maximum length. Word embedding and position embedding are generated for each residue. The fixed-length sequence fragments are input into the pre-trained protein language model ProtBert-BFD and the evolutionary scale modeling pre-trained model ESM in the pre-trained embedding module, respectively, to encode the residues in the sequence fragments and obtain high-dimensional potential protein representations. The residue representations output from the ProtBert-BFD branch and the residue representations output from the ESM branch are fused, and an initial residual connection is introduced to obtain the fused residue representation. A supervised contrastive learning module is constructed, which uses bound residues and non-bound residues as different categories of samples for representation constraints. Contrastive loss is used to make the representations of residues of the same class closer to each other in the feature space and the representations of residues of different classes farther apart, so as to alleviate the problem of positive and negative sample imbalance. The fused residue representation is input into a feature extraction module, which includes at least one Fourier layer; a high-dimensional residue representation is obtained through the Fourier layer. The high-dimensional residue representation input-output module, after feature compression and classification mapping, outputs the probability value of each residue belonging to a protein-peptide binding residue; The training objective is constructed by using cross-entropy loss and contrastive loss together. The pre-trained embedding module, Fourier feature extraction module and output module are jointly optimized, and whether each residue is a protein-peptide binding residue is determined based on the predicted probability of each residue.
2. The protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network according to claim 1, characterized in that, The pre-trained protein language model ProtBert-BFD is built based on the bidirectional encoder representation model BERT from Transformer and is used to capture bidirectional contextual dependencies between residues; the evolutionary scale modeling pre-trained model ESM is used to capture implicit information in protein sequences related to evolution, biochemical properties, secondary structure and tertiary structure.
3. The protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network according to claim 1, characterized in that, The contrastive learning module represents any two residues in the same batch. and Obtain the distance D between the two: Among them, cos<r1,r2> This represents the cosine similarity between two residues.
4. The protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network according to claim 3, characterized in that, Construct a contrastive loss function based on whether residue pairs belong to the same category: in, This indicates the category relationship between residue pairs; when two residues belong to the same category... =0; when the two residues belong to different categories. =1;D max This represents the maximum distance between residue pairs of different categories.
5. The protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network according to claim 1, characterized in that, Each Fourier layer maps the residue representation to the frequency domain via Fourier transform, performs a linear transformation on the frequency domain features to preserve low-frequency modes and filter high-frequency modes, and then restores it to the time domain via inverse Fourier transform. Finally, it combines local linear transformation, residual connections, and activation functions to obtain a high-dimensional residue representation.
6. The protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network according to claim 1, characterized in that, Cross-entropy loss and contrastive loss together constitute the total loss function. : in, Represents cross-entropy loss, This represents the loss in supervised contrastive learning.
7. The protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network according to claim 1, characterized in that, The output module is a multi-layer fully connected network with 64, 32, 16 and 2 nodes respectively.
8. A protein-peptide binding residue prediction system based on a pre-trained module and a Fourier network, characterized in that, include: The input embedding module obtains the amino acid sequence information of the protein to be predicted, segments or completes the amino acid sequence into fixed-length sequence fragments according to a preset maximum length, and generates word embeddings and position embeddings for each residue. The pre-trained embedding module inputs the fixed-length sequence fragment into the pre-trained protein language model ProtBert-BFD and the evolutionary scale modeling pre-trained model ESM in the pre-trained embedding module, respectively, to encode the residues in the sequence fragment and obtain a high-dimensional potential protein representation. The feature fusion module fuses the residue representations output from the ProtBert-BFD branch with the residue representations output from the ESM branch, and introduces initial residual connections to obtain the fused residue representations. The supervised contrastive learning module is constructed by treating bound residues and non-bound residues as different categories of samples for representation constraints. Through contrastive loss, the representations of residues of the same class are made closer to each other in the feature space, while the representations of residues of different classes are made farther apart, so as to alleviate the problem of positive and negative sample imbalance. The Fourier feature extraction module inputs the fused residue representation into the feature extraction module, which includes at least one Fourier layer; a high-dimensional residue representation is obtained through the Fourier layer; The output module takes the high-dimensional residue representation as input to the output module. The output module includes a multi-layer fully connected network that performs feature compression and classification mapping in sequence, and outputs the probability value of each residue belonging to a protein-peptide binding residue. The joint optimization and prediction module uses cross-entropy loss and contrastive loss to form a training objective, jointly optimizes the pre-trained embedding module, Fourier feature extraction module and output module, and determines whether each residue is a protein-peptide binding residue based on the predicted probability of each residue.
9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network as described in any one of claims 1-7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the protein-peptide binding residue prediction method based on a pre-trained module and a Fourier network as described in any one of claims 1-7.