A structure-driven affinity regression and binding classification joint prediction method for protein-peptide complexes
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-15
- Publication Date
- 2026-08-11
Smart Images
Figure CN122551868A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence-driven drug design and development technology, specifically relating to a structure-driven affinity regression and binding classification joint prediction method for protein-peptide complexes. Background Technology
[0002] Protein-peptide interactions are one of the complex molecular regulatory mechanisms in living organisms. As a unique type of drug, peptide molecules have significant advantages in bridging the technological gap between traditional small-molecule and large-molecule drugs. Specifically, peptide molecules can mimic the secondary conformation of natural proteins, achieving multivalent interactions with targets by increasing the interfacial contact area. When dealing with protein interfaces with flat surfaces and lacking deep binding pockets, peptide molecules exhibit superior binding affinity and target specificity compared to small-molecule drugs; compared to large-molecule antibodies, peptide molecules have higher tissue penetration and lower immune risk, and possess good flexibility in chemical modification. With the development of artificial intelligence technology, the accuracy of predicting the three-dimensional structure of protein-peptide complexes has been greatly improved. However, existing static structure-based modeling methods still cannot meet the actual needs of drug screening because static structures only describe the spatial distribution of atoms and cannot directly characterize the binding strength and kinetic stability of molecules. Since binding affinity is a core technical indicator for evaluating the drug potential of peptide drugs, how to achieve accurate quantitative perception of protein-peptide binding strength based on structural geometric information has become a key technical problem that urgently needs to be solved in this field.
[0003] In practical applications of protein-peptide affinity prediction, existing technologies face multiple technical bottlenecks. First, the extremely high flexibility of peptide molecules causes them to exhibit complex dynamic characteristics during binding, making it impossible for static modeling methods to quantify the contribution of conformational entropy change to binding strength, resulting in prediction bias. Second, the heterogeneity of interfacial interactions makes it extremely difficult to construct energy mapping relationships. Furthermore, limited by data sparsity and sample noise, existing algorithms generally lack generalization ability on unseen targets.
[0004] Currently, existing affinity prediction methods are mainly divided into three categories, but all have significant limitations. Sequence-based methods extract features using pre-trained protein language models, which are computationally efficient but lack explicit spatial geometric information, making it difficult to accurately characterize complex interfacial atomic contacts. Structure-based graphical models construct residue contact maps using graph neural networks, which can capture local interactions but are often limited to local neighborhoods, making it difficult to effectively model long-range dependencies and global topological constraints. While end-to-end structure prediction scoring modules have frameworks like Boltz2 that perform well in protein-small molecule systems, their affinity prediction modules are mainly designed for rigid small molecule ligands, and their energy mapping functions do not fully consider the high flexibility of peptides and the special interfacial physical environment, resulting in poor prediction performance when directly transferred to peptide systems.
[0005] In summary, existing technologies have significant limitations in handling flexible conformation modeling of peptides, cross-molecular system transfer applications, and small-sample generalization. Therefore, there is an urgent need for a novel deep learning framework that can effectively utilize the learning features of large-scale biomolecular pre-training and adapt to the physical properties of peptide interfaces to achieve end-to-end, high-precision protein-peptide affinity prediction. Summary of the Invention
[0006] To address the aforementioned problems, the present invention aims to provide a structure-driven affinity regression and binding classification joint prediction method for protein-peptide complexes. This invention constructs a structure-driven multi-task dataset, introduces a three-dimensional structural representation based on a pre-trained large model, and employs a transfer learning strategy of "freezing the main branch + fine-tuning the branches" and a trapezoidal mask mechanism for accurate annotations. This enables the model to achieve joint prediction of protein-peptide binding states and binding strength while maintaining stable structural representation capabilities, thereby improving the efficiency of protein-peptide interaction analysis and peptide drug screening.
[0007] The specific technical solution is as follows:
[0008] A structure-driven affinity regression and binding classification joint prediction method for protein-peptide complexes includes the following steps:
[0009] Step 1, Dataset Construction and Preprocessing: Integrate multi-source protein-peptide complex structure data and experimental affinity data to construct a unified dataset containing classification labels (binding / non-binding) and regression labels (affinity values). Obtain complex structure data from the PepPrCLIP database and the protein-peptide complex database constructed in the DiffPepBuilder project to construct a protein-peptide complex affinity classification dataset. Simultaneously, collect protein-peptide experimental affinity data from the PPIKB database, including dissociation constant (Kd), inhibition constant (Ki), or half-maximal inhibitory concentration (IC50), and construct a protein-peptide complex affinity regression dataset accordingly.
[0010] Step two involves filtering and classifying the dataset from Step one. Specifically, for the protein-peptide complex affinity classification dataset, the interface buried surface area is calculated based on the solvent-accessible surface area, and this metric is used as the structural criterion for determining whether a protein-peptide complex has formed a stable bond. For the protein-peptide complex affinity regression dataset, only linear peptide samples with a length less than 50 amino acids are retained, while cyclic peptides and sequences containing non-natural amino acids are removed. Furthermore, for threshold data where experimental affinity is reported in the form of "≥" or "≤", records equal to or less than the measured value are uniformly retained to ensure data consistency.
[0011] Step 3: Construct negative sample data. While keeping the peptide sequence unchanged, replace the receptor protein in the positive sample with a protein sequence from a different protein cluster to generate a new protein-peptide combination. The protein clusters are clustered based on sequence similarity to ensure that the replaced protein has a significant difference from the original protein in sequence space, thereby reducing the possibility of forming a true binding conformation. During the construction of negative samples, the peptide length distribution is kept consistent with that of the positive samples, and length perturbation is allowed within ±1 residue range to maintain the consistency of data distribution. One negative sample is generated for each positive sample to construct a 1:1 positive-to-negative sample ratio, while avoiding the generation of protein-peptide combinations that are duplicates of existing positive samples.
[0012] Step four involves structural quality filtering and redundancy removal of the dataset. The final dataset contains 20,950 protein-peptide complex structures, including 10,435 positive samples and 10,515 negative samples. The dataset is then divided into training and testing sets. The training set contains 18,429 samples, of which 4,500 have affinity numerical labels; 8,992 are positive samples and 9,437 are negative samples. The testing set contains 2,521 samples, of which 516 have affinity numerical labels; 1,443 are positive samples and 1,078 are negative samples.
[0013] Step 5: Inherit the backbone network design from Boltz2, and model the geometry and spatial constraints of the complex using both residue-level and residue-pair-level representations. Let the total number of residues in the complex be... , residues The monomer is denoted as residue pairs The paired representation of is denoted as In the structural backbone network, paired representations are iteratively updated through multiple layers of PairFormer modules:
[0014]
[0015] This module utilizes an attention mechanism and a triangular update operation to model higher-order dependencies between residues, enabling information to propagate globally within the residue pair network. Simultaneously, the monomer representation is updated based on the pairwise representation, thus forming a bidirectional coupling optimization between the monomer and pairwise representations. After the structural representation update is complete, the 3D coordinates of the complex are predicted using a denoising generation module.
[0016]
[0017] This module iteratively optimizes the residue coordinates to ensure that the predicted structure satisfies both geometric rationality and energy consistency.
[0018] Step six involves inputting the protein-peptide complex structure predicted in step five into the affinity prediction module, enabling the module to learn structural features. The affinity prediction module is based on residue pairs output from the structural backbone network. And extract interface residue pairs using a cross-chain mask matrix:
[0019]
[0020] Interface-level representation is defined as:
[0021]
[0022] in This represents a global pooling or attention convergence operation, used to compress residue-level information into a complex-level vector representation.
[0023] This interface indicates that the inputs are then fed into two parallel prediction branches: a classification branch to predict whether a valid binding occurs, and a regression branch to predict the affinity value on a logarithmic scale. The two branches share a common interface-level structure. However, independent parameter mapping is used in the output layer to characterize the probability of binding and the binding strength separately.
[0024] Step 7: Fine-tune the affinity prediction module using the protein-peptide dataset obtained in Step 4; use Huber loss for the regression task; let the model output affinity prediction be... The actual pAffinity value is For samples with missing numerical labels, their true values are recorded as NaN, and the model's predicted values are used to replace these positions during loss calculation, so that they do not contribute to the gradient of the regression loss; Huber loss is defined as:
[0025]
[0026] in , For threshold hyperparameters; The threshold hyperparameter is preferably set to 0.5 in this invention.
[0027] For the combined state classification task, focus loss is used. Let the combination probability of the classification branch output be... The real label is First, define:
[0028]
[0029] To predict the binding probability, This is a real label.
[0030] The focal loss is written as:
[0031]
[0032] in For difficult cases, focus coefficient, To balance the category weights. Take... and will Let it be a constant ( The final joint loss function is:
[0033]
[0034] Step 8: Evaluate the trained model using a test set to verify its effectiveness in the protein-peptide complex binding prediction task.
[0035] The advantages of this invention are twofold: First, by introducing a three-dimensional structural representation of the complex and combining residue-level and residue-pair-level feature modeling, the model's ability to characterize the spatial structure of the protein-peptide interface is improved. Simultaneously, by constructing a multi-task learning framework, the model can simultaneously predict whether protein-peptide binding occurs stably and the binding strength, thereby improving the reliability of the prediction results. Second, by constructing a large-scale structural dataset, introducing a negative sample generation strategy, and employing transfer learning fine-tuning methods, the model can maintain stable prediction performance in cross-protein system scenarios. This invention successfully implements a structure-driven affinity regression and binding classification joint prediction method for protein-peptide complexes, capable of rapidly predicting binding probabilities and affinity values at the protein and peptide sequence levels, providing an effective computational tool for peptide drug screening and protein-peptide interaction research. Attached Figure Description
[0036] Figure 1 This is a flowchart illustrating the combined prediction of structure-driven affinity regression and binding classification for specific protein-peptide complexes according to the present invention. Detailed Implementation
[0037] The present invention will now be described in detail with reference to the accompanying drawings and embodiments, providing a clear and complete description of the technical solutions in the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0038] Example
[0039] like Figure 1 As shown, a structure-driven affinity regression and binding classification joint prediction method for protein-peptide complexes includes the following steps:
[0040] Step 1, Dataset Construction and Preprocessing: Integrate multi-source protein-peptide complex structure data and experimental affinity data to construct a unified dataset containing classification labels (binding / non-binding) and regression labels (affinity values). Obtain complex structure data from the PepPrCLIP database and the protein-peptide complex database constructed in the DiffPepBuilder project to construct a protein-peptide complex affinity classification dataset. Simultaneously, collect protein-peptide experimental affinity data from the PPIKB database, including dissociation constant (Kd), inhibition constant (Ki), or half-maximal inhibitory concentration (IC50), and construct a protein-peptide complex affinity regression dataset accordingly.
[0041] Step two involves filtering and classifying the dataset from Step one. Specifically, for the protein-peptide complex affinity classification dataset, the interface buried surface area is calculated based on the solvent-accessible surface area, and this metric is used as the structural criterion for determining whether a protein-peptide complex has formed a stable bond. For the protein-peptide complex affinity regression dataset, only linear peptide samples with a length less than 50 amino acids are retained, while cyclic peptides and sequences containing non-natural amino acids are removed. Furthermore, for threshold data where experimental affinity is reported in the form of "≥" or "≤", records equal to or less than the measured value are uniformly retained to ensure data consistency.
[0042] Step 3: Construct negative sample data. While keeping the peptide sequence unchanged, the receptor protein in the positive sample is replaced with a protein sequence from a different protein cluster, thereby generating a new protein-peptide combination. The protein clusters are clustered based on sequence similarity to ensure that the replaced protein has a significant difference from the original protein in sequence space, thereby reducing the possibility of forming a true binding conformation. During the construction of negative samples, the peptide length distribution is kept consistent with that of the positive samples, and length perturbation is allowed within ±1 residue range to maintain the consistency of data distribution. One negative sample is generated for each positive sample to construct a positive-to-negative sample ratio of approximately 1:1, while avoiding the generation of protein-peptide combinations that are duplicates of existing positive samples.
[0043] Step four involves structural quality filtering and redundancy removal of the dataset. The final dataset contains 20,950 protein-peptide complex structures, including 10,435 positive samples and 10,515 negative samples. The dataset is then divided into training and testing sets. The training set contains 18,429 samples, of which 4,500 have affinity numerical labels; 8,992 are positive samples and 9,437 are negative samples. The testing set contains 2,521 samples, of which 516 have affinity numerical labels; 1,443 are positive samples and 1,078 are negative samples.
[0044] Step 5: Inherit the backbone network design from Boltz2, and model the geometry and spatial constraints of the complex using both residue-level and residue-pair-level representations. Let the total number of residues in the complex be... , residues The monomer is denoted as residue pairs The paired representation of is denoted as In the structural backbone network, paired representations are iteratively updated through multiple layers of PairFormer modules:
[0045]
[0046] This module utilizes an attention mechanism and a triangular update operation to model higher-order dependencies between residues, enabling information to propagate globally within the residue pair network. Simultaneously, the monomer representation is updated based on the pairwise representation, thus forming a bidirectional coupling optimization between the monomer and pairwise representations. After the structural representation update is complete, the 3D coordinates of the complex are predicted using a denoising generation module.
[0047]
[0048] This module iteratively optimizes the residue coordinates to ensure that the predicted structure satisfies both geometric rationality and energy consistency.
[0049] Step six involves inputting the protein-peptide complex structure predicted in step five into the affinity prediction module, enabling the module to learn structural features. The affinity prediction module is based on residue pairs output from the structural backbone network. And extract interface residue pairs using a cross-chain mask matrix:
[0050]
[0051] Interface-level representation is defined as:
[0052]
[0053] in This represents a global pooling or attention convergence operation, used to compress residue-level information into a complex-level vector representation.
[0054] This interface indicates that the inputs are then fed into two parallel prediction branches: a classification branch to predict whether a valid binding occurs, and a regression branch to predict the affinity value on a logarithmic scale. The two branches share a common interface-level structure. However, independent parameter mapping is used in the output layer to characterize the probability of binding and the binding strength separately.
[0055] Step 7: Fine-tune the affinity prediction module using the protein-peptide dataset obtained in Step 4; use Huber loss for the regression task; let the model output affinity prediction be... The actual pAffinity value is For samples with missing numerical labels, their true values are recorded as NaN, and the model's predicted values are used to replace these positions during loss calculation, so that they do not contribute to the gradient of the regression loss; Huber loss is defined as:
[0056]
[0057] in , For threshold hyperparameters; The threshold hyperparameter is preferably set to 0.5 in this invention.
[0058] For the combined state classification task, focus loss is used. Let the combination probability of the classification branch output be... The real label is First, define:
[0059]
[0060] To predict the binding probability, This is a real label.
[0061] The focal loss is written as:
[0062]
[0063] in For difficult cases, focus coefficient, To balance the category weights. Take... and will Let it be a constant ( The final joint loss function is:
[0064]
[0065] Step 8: Evaluate the trained model using a test set to verify its effectiveness in the protein-peptide complex binding prediction task.
[0066] In embodiments of the present invention, the proposed deep learning algorithm was tested on a test set. Experimental results show that the model achieves optimal performance for all values on the test set, demonstrating that the protein-peptide affinity prediction method proposed in this invention can effectively improve the accuracy of protein-peptide interaction analysis.
[0067] Table 1. Model performance on the classification test set
[0068] Boltz2 0.6287 0.5695 0.8001 0.7511 Boltz2_Scratch 0.7989 0.7779 0.8690 0.8401 PPAP 0.5724 0.5000 0.7862 0.7281 Proaffinity-GNN 0.5728 0.5005 0.7863 0.7282 ESM+Cross Attention 0.7447 0.7584 0.7565 0.7844 This method 0.8128 0.7925 0.8767 0.8508
[0069] Table 2. Model performance on the regression test set
[0070] Boltz2 2.2096 2.6373 0.3724 0.3138 Boltz2_Scratch 1.0291 1.3619 0.7143 0.7612 PPAP 2.2039 2.6299 -0.5436 -0.5582 Proaffinity-GNN 1.8851 2.2452 0.0220 0.0888 This method 0.9335 1.2237 0.7203 0.7437
[0071] The above description is only a preferred embodiment of the present invention and does not limit the scope of patent protection of the present invention. Any equivalent substitutions or modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.
Claims
1. A structure-driven affinity regression combined with binding classification prediction method for protein-polypeptide complex, characterized in that, Includes the following steps: Step 1, Data Construction and Preprocessing: Integrate multi-source protein-peptide complex structure data and experimental affinity data to construct a dataset containing classification labels and regression labels; Step 2, structural representation learning: The complex is processed using a pre-trained 3D structural prediction model to extract pairwise representation features containing higher-order dependencies of residue pairs; Step 3, fine-tuning of transfer learning: Freeze the backbone network parameters of the pre-trained 3D structure prediction model and only update the parameters of the subsequently constructed affinity prediction module; Step 4, Interface Feature Extraction and Multi-Task Inference: Based on the cross-chain mask matrix, protein-peptide interface residue pair features are extracted from the pairwise representation. An interface-level vector representation is constructed through pooling operations. The interface-level vector representation is then input in parallel to the binding classification branch and the affinity regression branch, and the binding probability and affinity prediction value are output. Step 5, Joint Loss Optimization: Optimize using a joint loss function that includes regression loss and classification loss. For samples lacking affinity labels in the regression task, gradient masking is performed to block their backpropagation contribution.
2. The structure-driven affinity regression and binding classification combined prediction method for protein-polypeptide complex according to claim 1, characterized in that, The data preprocessing described in step one specifically includes: 1) For classification datasets: Calculate the surface area buried in the interface, and use whether the surface area exceeds a preset threshold as the criterion for determining whether protein-peptide binding has occurred stably; 2) For regression datasets: screen linear natural polypeptide sequences with a length of less than 50 amino acids, and remove cyclic peptides and non-natural amino acid sequences; for threshold data with "≥" or "≤" signs in experimental affinity, uniformly retain records that are equal to or less than the measured value.
3. The structure-driven affinity regression and binding classification combined prediction method for protein-polypeptide complex according to claim 1, characterized in that, Step one also includes negative sample construction and redundancy removal: while keeping the peptide sequence unchanged, the receptor protein in the positive sample is replaced with a protein sequence from a different protein cluster to generate a negative sample; the protein clusters are divided based on sequence similarity clustering; the peptide length in the negative sample can be randomly perturbed within ±1 residues based on the positive sample to maintain data distribution consistency.
4. The structure-driven affinity regression and binding classification combined prediction method for protein-polypeptide complex according to claim 1, characterized in that, The update process of pairwise representation features in step two includes: Residue pair level representations are updated iteratively by a multi-layer PairFormer module with update formula: ， wherein, is the residue monomer representation; the updated pairwise representation is input to the denoising generation module to predict the three-dimensional spatial coordinates of the complex residues.
5. The structure-driven affinity regression and binding classification combined prediction method for protein-polypeptide complex according to claim 1, characterized in that, The interface level vector representation in step four The extraction process is: Using cross-chain mask matrix Identify interface residue pairs, where: ， Through pooling operations The interface displayed is as follows: ; The classification branch shares interface-level vector representations with the regression branch and predicts the binding probability p and the affinity value in log scale separately through independent output layer parameters .
6. The structure-driven affinity regression and binding classification combined prediction method for protein-polypeptide complex according to claim 1, characterized in that, Regression loss function in step five Classification loss with Huber loss Classification loss with focal loss Joint loss function is defined as: ， For samples with missing true values, their labels are marked as NaN, and their contribution to the regression loss is set to zero when calculating the regression loss.
7. The method for joint prediction of structure-driven affinity regression and binding classification for protein-peptide complexes according to claim 6, characterized in that, The calculation of the focal loss introduces a hard example focusing coefficient and a class weight balancing factor defined as: ， wherein, , is the predicted probability of binding, is the true label.