Protein target binding peptide de novo generation method and system based on geometric feature and language model fusion and motif anchor diffusion

Abstract

The application discloses a protein target binding peptide de novo generation method and system based on geometric features and language model fusion and motif anchoring diffusion, and the method comprises the following steps: receiving the 3D structure of a target protein; recognizing a binding pocket on the 3D structure of the target protein based on geometric and language model fusion features by using a pre-trained graph neural network model; generating an anchoring fragment for complementary binding according to the geometric features of the recognized pocket, wherein the anchoring fragment comprises a short peptide motif of a plurality of amino acid residues; expanding the anchoring fragment into a complete binding peptide skeleton spatial structure according to specified topological constraint conditions; deducing the probability distribution of a plurality of preset amino acids at each site on the complete binding peptide skeleton, and selecting an appropriate amino acid for each site according to the probability distribution result, and then obtaining a protein target binding peptide. The application takes binding priority as the organizing principle, takes the geometry of the target site as the main line, and transmits the binding related information throughout the entire design process through stage-by-stage constraint propagation.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of AI and biology, and in particular to a method and system for de novo generation of protein target-binding peptides based on the fusion of geometric features and language models and motif anchoring diffusion, implemented using a computer. Background Technology

[0002] Peptide binders are short peptide sequences consisting of several to dozens of amino acid residues. They can specifically recognize and bind to targets such as protein receptors, nucleic acids, or cell surface markers through non-covalent interactions (i.e., reversible molecular interactions that do not form true chemical bonds), such as hydrogen bonds, hydrophobic interactions, and electrostatic attraction. Due to their high targeting specificity and structural plasticity, peptide binders exhibit diverse applications in the biomedical field: in targeted drug delivery, they can serve as ligands to load therapeutic or diagnostic agents, reducing off-target toxicity; in the functionalization of nucleic acid drugs, they can endow oligonucleotide drugs with tissue / cell targeting and endosome escape capabilities; in co-delivery systems, they are often tandem with cell-penetrating peptides (CPPs, i.e., short peptides that facilitate molecular transmembrane entry into cells), achieving both specific targeted binding and efficient transmembrane internalization. These characteristics make them important molecular components in the development of novel targeted therapies, molecular imaging probes, and diagnostic platforms.

[0003] Peptides occupy a unique intermediate position in molecular design: they are large enough to recognize extended and chemically heterogeneous protein surfaces, yet compact enough to be synthesized, diversified, and optimized in ways that are difficult to achieve with larger biopharmaceuticals. This makes peptides potential scaffolds for therapeutics, biosensors, diagnostic reagents, and programmable biomaterials. Unless otherwise specified, peptides designed for specific binding to specific targets will be referred to as "peptide binders" in the following text.

[0004] However, peptide binding agent design itself is extremely challenging. A useful peptide binding agent must simultaneously meet multiple coupling requirements: it must bind at the correct location, adopt a main chain geometry compatible with the target site, satisfy topological constraints such as linear or cyclic closure (i.e., structural limitations imposed by molecular linkage and closure modes), and its sequence's local dihedral preference must be physically reasonable.

[0005] These requirements create a design space that is not only vast but also structurally complex. Changing the binding site alters the set of feasible backbones; changing the topology changes the allowed torsional trajectory space; and changing the backbone changes the physically or statistically preferred amino acids at each position. In other words, peptide design is not merely a single search for a sequence or structure, but a constrained search for a range of dependent potential objects.

[0006] The disclosure of the above background technical content is only for the purpose of assisting in understanding the concept and technical solution of this application, and does not necessarily provide technical instruction. Summary of the Invention

[0007] The purpose of this invention is to provide a method for de novo generation of protein target-binding peptides based on the fusion of geometric features and language models and motif anchoring diffusion. It prioritizes binding and uses the geometry of the target site as the main thread, and integrates binding-related information throughout the entire design process through staged constraint propagation.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for de novo generation of protein target-binding peptides based on the fusion of geometric features and language models and motif anchoring diffusion includes the following steps: Receives the 3D structure of the target protein; The protein-binding pockets on the 3D structure of the target protein are identified using a pre-trained graph neural network model based on the fusion of geometric and linguistic features. Based on the three-dimensional geometric features of the identified protein binding pocket, a complementary binding anchoring fragment is generated, wherein the anchoring fragment is a short peptide motif comprising one or more amino acid residues. Based on the specified topological constraints, the anchoring fragment is expanded into a complete binding peptide backbone spatial structure; The probability distribution of multiple amino acids at each site on the complete binding peptide backbone spatial structure is inferred, and suitable amino acids are selected for each site on the complete binding peptide backbone spatial structure according to the probability distribution results, thereby obtaining the protein target binding peptide.

[0009] Furthermore, based on any one or a combination of the aforementioned technical solutions, the anchoring segment is generated in the following manner: Using the three-dimensional geometric features of the identified protein binding pocket as input, the trained generative model is invoked to simulate and generate an anchoring fragment with peptide-specific interaction with the protein binding pocket in situ within the physical boundary of the protein binding pocket. The generated anchoring fragment can be detached from its original position, allowing for flexible insertion or removal. The number of residues in the anchoring fragment is determined by the generation model based on the geometry of the protein binding pocket, and the anchoring fragment itself (the number and arrangement of residues) is locked as unmodifiable in the subsequent step of expanding the complete binding peptide backbone spatial structure.

[0010] Furthermore, based on any one or more of the aforementioned technical solutions or combinations thereof, the three-dimensional geometric features of the protein binding pocket include one or more of the following: the Cα coordinates of the protein binding pocket residues, the orientation of the residue side chains, and the local shape description of the protein binding pocket, wherein the local shape description includes the dihedral angle, surface normal vector, curvature, and / or depth of the protein binding pocket. The constraints for the generative model to generate anchored segments include one or more of the following: The Cα coordinates of the anchoring fragment during the generation process are located within the physical boundary of the protein binding pocket; The side chain orientation of the generated anchoring fragment is opposite to the side chain orientation of the hotspot residues in the protein binding pocket; Gradually increase the number of residues in the anchoring fragment until the minimum distance between the newly added residue and the nearest atom in the protein binding pocket is lower than a preset threshold.

[0011] Furthermore, following any one or a combination of the aforementioned technical solutions, the generation model uses the Cα coordinates of the identified protein binding pocket residues and the local dihedral angle as input conditions to simulate and generate the anchoring fragment. The generation model is trained in the following manner: The binding pocket residues and their corresponding contact motif torsion angles were extracted from the PepBDB protein-peptide complex database to construct training sample pairs. Each sample contains the ESM2 embedding vector of the pocket residue and the ground truth value of the contact motif torsion angle. True value of twist angle x 0 Perform forward diffusion noise addition to obtain the noisy twist angle. x t ; The noisy torsion angle x t After being concatenated with temporal embedding, residue position embedding, and pocket embedding, the input is fed into a three-layer fully connected decoder to predict the injected noise ε. θ ( x t ); The generative model was trained using the AdamW optimizer with the mean square error between the predicted noise and the actual noise as the training objective.

[0012] Furthermore, following any one or a combination of the aforementioned technical solutions, the generative model is a diffusion generative model that models in geometric space, the diffusion process uses 1000-2000 time steps, and the optimizer is AdamW. The diffusion generation model projects the input torsion angle tensor (represented by 6-dimensional sin / cos per residue) into a 256-dimensional hidden space, and concatenates it with the temporal embedding, residue position embedding, and pocket embedding after mean pooling to form a joint representation. The prediction noise is then output through a three-layer fully connected decoder. The training data for the diffusion generation model uses known peptide-protein complex binding motifs from PepBDB or PDB as training ground values. The training objective of the diffusion generation model is the mean square error (MSE) loss between the predicted noise and the actual noise.

[0013] Furthermore, following any one or a combination of the aforementioned technical solutions, the specified topological constraint is the peptide backbone topology type input by the user, which includes linear peptides, cyclic peptides, or lasso peptides. The peptide backbone topology type is constrained for subsequent backbone extensions of the anchored fragment.

[0014] Furthermore, based on any one or a combination of the aforementioned technical solutions, the specified topological constraint is the peptide backbone topology type of the linear peptide; Based on the topological constraints, the anchored segment is expanded into a CDR; The CDR is embedded into the frame region backbone of a user-specified antibody, which includes IgG, IgA, IgD, IgE, IgM, and VHH single-domain antibodies.

[0015] Furthermore, following any one or a combination of the aforementioned technical solutions, the probability distribution of amino acids is inferred using a mixture of artificial intelligence pre-trained models and physical priors, including: The probability distribution of multiple amino acids at the i-th site on the peptide backbone is predicted using an artificial intelligence pre-trained model, denoted as P1(a|i). The probability distribution of multiple amino acids at the i-th site on the peptide backbone is output using a physical prior model, denoted as P2(a|i). The probability distributions P1(a|i) and P2(a|i) are weighted and summed to obtain the amino acid probability distribution at the i-th site in the fusion inference. The amino acid with the highest probability in the probability distribution results of the fusion inference is then identified as the amino acid at that site.

[0016] Furthermore, based on any or a combination of the aforementioned technical solutions, the artificial intelligence pre-trained model is an ESM-IF1 inverse folding predictor, and the physical prior model is a Ramachandran prior model. The weighted summation methods include: If the maximum, minimum, mean, or sum of the elements of the amino acid probability distribution matrix at the i-th site output by the ESM-IF1 inverse folding predictor is less than the preset probability threshold p th The probability distribution of the fusion inference is then calculated using the following formula: P final (a|i)=P2(a|i); If the maximum, minimum, mean, or sum of the elements of the amino acid probability distribution matrix at the i-th site output by the ESM-IF1 inverse folding predictor reaches the preset probability threshold p th The probability distribution of the fusion inference is then calculated using the following formula: P final (a|i)=α×P1'(a|i)+β×P2(a|i), where P1'(a|i) is the normalized result of the probability distribution P1(a|i); α and β are weights, which sum to 1, and α>β.

[0017] Furthermore, following any one or a combination of the aforementioned technical solutions, the de novo method for generating protein target-binding peptides provided by this invention further includes the following phased verification steps: After identifying the protein binding pocket based on the fusion features of geometric and language models, residue-level classification indicators are used for evaluation. These residue-level classification indicators include precision, recall, F1 score, area under the receiver operating characteristic curve, and area under the precision-recall curve. After generating the anchored segments, they are evaluated based on geometric validity and structural diversity. After obtaining the complete binding peptide backbone spatial structure, it is evaluated through one or more of the following dimensions: using Laplace plot compliance to check whether the dihedral angles of the backbone are located in the allowed region; using topological satisfaction to verify whether the topological constraints are met; and / or, using motif fidelity to verify whether the reconstruction accuracy of the conservative regions of the backbone meets the standard. After inferring the amino acid probability distribution, the results are evaluated based on sequence recovery rate, rationality of amino acid distribution, and physical validity. After obtaining the protein target-binding peptide, it was evaluated through end-to-end case studies, generation diversity, structural novelty, and computational efficiency relative to the baseline process.

[0018] Furthermore, following any one or a combination of the aforementioned technical solutions, the pre-trained graph neural network model is a GNN model based on graph attention networks. It takes the mixed feature vector of each amino acid residue as input, constructs a multi-scale graph structure on the Cα coordinates of the residues using the k-nearest neighbor algorithm, where k takes values ​​of 5, 10, and 15, and merges the multi-scale graph structure into a single side set. The mixed feature vector includes manually constructed structural features and residue evolutionary context embeddings extracted through a pre-trained protein language model. The input nodes of the GNN model project the hybrid feature vector into a multidimensional hidden space; The main body of the GNN model uses four stacked multi-scale GAT layers for message passing, with each GAT layer followed by a batch normalization layer, an ELU activation layer, and a dropout layer in sequence. The binary classification head of the GNN model outputs a prediction of whether each residue is a peptide-binding interface residue, and the multi-level classification head of the GNN model outputs the spatial hierarchy category to which each residue belongs.

[0019] Furthermore, following any one or a combination of the aforementioned technical solutions, the hybrid feature (node ​​feature) vector of each amino acid residue is a 669-dimensional node feature vector, which includes 29-dimensional manually constructed structural features and 640-dimensional ESM embeddings extracted by the ESM2 model. The manually constructed structural features include pure geometric features, such as dihedral angles and accessible surface area. Each edge in the multi-scale graph structure corresponds to a 21-dimensional edge feature vector, which combines paired geometric quantities and interaction descriptors. The paired geometric quantities include at least one of distance, direction, and relative position.

[0020] Furthermore, following any one or a combination of the aforementioned technical solutions, the training method of the GNN model includes: Raw data were extracted from the PepBDB database and one or more of the following filtering conditions were applied: peptide sequence length was limited to 5-50 amino acid residues, receptor sequence length was not less than 30 residues, crystal structures with a resolution greater than 3.5 Å were excluded, and each complex was required to contain at least 3 interface residues. A combination of distance thresholding and interaction detection methods is used to label interface residues to generate multi-level labels. The distance thresholding method sets corresponding thresholds for the core layer based on heavy atoms, the support layer based on Cβ atoms, and the edge layer based on Cα atoms. The interaction detection method detects hydrogen bonds, salt bridges, and hydrophobic interactions. A multi-task learning strategy is adopted for training, which simultaneously optimizes binary interface prediction and multi-level interface type classification. The AdamW optimizer is used, with focus loss as the loss function, and a dynamic negative sample sampling strategy is introduced. This strategy sets the ratio of negative samples to positive samples in each batch to n:1 starting from the Nth training round, where N is an integer between 5 and 120, and n is between 1.1 and 5.

[0021] Furthermore, following any one or a combination of the aforementioned technical solutions, in the forward propagation computation of the GNN model, the input graph is constructed by three-scale k-nearest neighbor edges with k values ​​of 5, 10, and 15 before entering the GATv2 layer; k=5 is used to capture local close biochemical context or short-range spatial interactions; k=10 is used to capture medium-range secondary structure features; and k=15 is used to capture long-range tertiary structure folding information. Such three-scale fusion can comprehensively characterize the multi-level topological structure of molecules.

[0022] The unnormalized attention coefficient for each layer of GATv2 is calculated using the following formula: e ij = T LeakyReLU[W1· i ||W2· j ||W3· ij ], where LeakyReLU is a function that performs a non-linear mapping on the concatenated vectors. i Let be the node feature vector of node i. j Let be the node feature vector of node j. ij Let W1, W2, and W3 be the feature vectors of the edge between node i and node j, and let W1, W2, and W3 be the learnable weights. Let T be the attention vector, T be the matrix transpose, and e be the unnormalized attention coefficients. ij The final attention coefficient α is obtained after softmax normalization. ij This is used to weight and aggregate the feature information of neighboring nodes. The edge feature vector... ij Features including, but not limited to, the Euclidean distance, spatial angle, and differences in amino acid residue types between node i and node j.

[0023] According to another aspect of the present invention, a de novo system for generating protein target-binding peptides based on the fusion of geometric features and language models and motif anchoring diffusion is provided, comprising the following modules: The pocket recognition module is configured to identify protein-binding pockets on the 3D structure of a target protein by using a pre-trained graph neural network model based on features fused from geometric and linguistic models. An anchoring module is configured to generate complementary anchoring fragments based on the three-dimensional geometry of the identified protein-binding pocket, the anchoring fragments being short peptide motifs comprising one or more amino acid residues. An extension module is configured to extend the anchoring fragment into a complete binding peptide backbone spatial structure according to specified topological constraints. The probability distribution inference module is configured to infer the probability distribution of a preset number of amino acids at each site on the complete binding peptide backbone spatial structure. The generation module is configured to select suitable amino acids for each site on the complete binding peptide backbone spatial structure according to the probability distribution results, thereby obtaining a protein target binding peptide.

[0024] The beneficial effects of the technical solution provided by this invention are as follows: a. The system adopts a process organization method that prioritizes combination, propagating geometric information forward along the entire process. Instead of generating a general peptide and then screening it, the system first identifies the target region on the receptor surface, uses this region to propose the local backbone geometry, expands and optimizes the full-length peptide under topological constraints, and finally outputs the positional amino acid distribution based on the refined backbone. b. Each stage forwards constraint information. The binding site identification stage first locates potential peptide binding interface regions on the receptor surface; the pocket conditionalized backbone stage proposes local torsion angle priors based on this; the topology-aware optimization stage expands the local proposals into the full-length peptide backbone under motif hard constraints; finally, the mixed sequence design stage outputs position-level amino acid probability distributions. This intelligent constraint propagation mechanism, which shrinks the search space layer by layer, is the core feature that distinguishes it from the loosely modular splicing process. c. The goal is not merely to generate physically rational peptides in isolation, but to generate peptides whose backbone geometry is compatible with the target site, and whose sequence is consistent with the local backbone conformation and global design. Figure 1 Candidate peptides that are induced. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 A flowchart illustrating a de novo method for generating protein target-binding peptides based on the fusion of geometric features and language models and motif anchoring diffusion, provided as an exemplary embodiment of the present invention. Figure 2 A schematic block diagram of a de novo protein target binding peptide generation system based on geometric feature and language model fusion and motif anchoring diffusion, provided as an exemplary embodiment of the present invention. Figure 3A schematic diagram of an amino acid heatmap of 18 residues corresponding to the HER2 case provided as an exemplary embodiment of the present invention; Figure 4 A schematic diagram of the AF3 complex model, which is the strongest HER2 candidate, provided as an exemplary embodiment of the present invention; Figure 5 A schematic diagram of the amino acid heatmap corresponding to the 12 residues of the 1XQ8 case provided as an exemplary embodiment of the present invention; Figure 6 A schematic diagram of an AF3 complex model with the highest-ranking 1XQ8 candidate provided for an exemplary embodiment of the present invention; Figure 7 An exemplary embodiment of the present invention provides a method for running a hybrid sequence design phase in a hybrid mode to obtain a position-by-position amino acid heatmap; Figure 8 A schematic diagram illustrating the validation summary of all candidate ranking scores, ipTM, and average pLDDT for case studies of 1XQ8 and HER2 provided as an exemplary embodiment of the present invention; Figure 9 A schematic diagram of the ROC and PRC recognition performance curves of a binding site identifier provided as an exemplary embodiment of the present invention; Figure 10 A comparison diagram of the reconstructed MSE and the MSE after zeroing the pocket embedding of the pocket conditional model provided as an exemplary embodiment of the present invention; Figure 11 A schematic diagram illustrating how the reconstruction gap between conditional and unconditional models widens with increasing noise level, provided as an exemplary embodiment of the present invention. Figure 12 A schematic diagram of the distribution for running inference with a fixed main chain support window and hard protection enabled, provided as an exemplary embodiment of the present invention; Figure 13 The hybrid predictor provided as an exemplary embodiment of the present invention can stably generate a well-formed L×20 probability matrix schematic diagram. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0029] The challenge in peptide binder design lies not in simple sequence or structure generation, but in simultaneously satisfying a series of coupled conditions, including binding site, main chain geometry, topological constraints, and the rationality of local sequence conformation. Existing computational workflows often disperse these issues into independent stages such as binding site identification, main chain generation / optimization, and sequence allocation, making it difficult to continuously transfer geometric information between stages.

[0030] Current computational workflows typically handle these decisions in a fragmented manner: one tool might identify potential pockets (i.e., local regions on the receptor surface where ligands / peptides might be accommodated or contacted), another might generate or optimize the backbone, and a third might propose sequences on the resulting scaffold. While this decomposition is operationally convenient, it mismatches with the actual physics of molecular recognition: binding geometry, backbone feasibility, topology, and sequence compatibility are not independent design subproblems but rather coupled constraints for a single object. This fragmentation also makes it difficult to diagnose the causes of failure; a poor candidate peptide might be due to misaligned pockets, backbone deviation from the expected interface, or inconsistencies between the final sequence and the local geometry, and conventional workflows typically do not preserve the intermediate representations needed to diagnose these discrepancies.

[0031] In one embodiment of the present invention, a method for de novo generation of protein target-binding peptides based on the fusion of geometric features and language models and motif anchoring diffusion is provided. De novo generation means that it does not rely on known peptide sequence templates or pre-existing peptide structures, but directly starts from the 3D structure of the target protein to generate novel binding peptide sequences and structures that do not appear in nature or known databases. Contrary to the design principles of current computational workflows, the peptide binding agent design in this embodiment prioritizes binding, using target site geometry as the main thread, and propagating binding-related information throughout the entire design process through staged constraint propagation, where target information propagates forward along the entire design process. Figure 1As shown, the de novo method for generating protein target-binding peptides provided in this embodiment of the invention includes the following steps: It receives the 3D structure of the target protein, which is the target receptor for generating the binding peptide (i.e., the peptide binder). Identify regions on the receptor that may support peptide binding (referred to as pockets): Identify target peptide binding regions on the 3D structure of the target protein, i.e. protein binding pockets, using a pre-trained graph neural network model based on the fusion of geometric and linguistic features. Using these regions that support peptide binding, short main chain motifs locally compatible with the target pocket are proposed: Based on the three-dimensional geometric features of the identified protein binding pocket, complementary binding anchoring fragments are generated. The anchoring fragment is a short peptide motif comprising one or more amino acid residues, which is the core and indivisible part of the final product binding peptide (Peptide Binder). Under topological constraints (input depending on business needs), these motifs are expanded and optimized into complete peptide backbones: the anchoring fragments are expanded into complete binding peptide backbone spatial structures according to the specified topological constraints. It generates position-level amino acid probability distributions instead of single deterministic sequences, supporting downstream sampling, sorting, and experimental priority determination: it infers the probability distribution of multiple amino acids at each site on the complete binding peptide backbone spatial structure, and selects suitable amino acids for each site on the complete binding peptide backbone spatial structure based on the probability distribution results, thereby obtaining the protein target binding peptide.

[0032] The contribution of this invention lies in organizing the entire de novo peptide generation process around a consistent geometric logic: each stage outputs a structured intermediate, pocket diagram, torsion angle prior, anchor sequence generation, and topologically constrained main chain or sequence heatmap to limit the search space faced by the next stage. This allows it to operate as a coherent peptide design system, rather than a loose collection of independent predictors. The resulting process is interpretable at the system level: each stage can be examined, ablated, or replayed, which is important for scientific understanding and practical applications of iterative design.

[0033] The following sections will describe in detail the process in four main parts: identification of the protein binding pocket, generation of the anchoring fragment, expansion of the spatial structure of the intact binding peptide backbone, and inference of the amino acid probability distribution. The first step is to identify protein-binding pockets, using graph neural networks to address the problem of binding site localization. This stage is not a universal cavity detector: its role is to identify receptor residues that may be involved in peptide recognition so that all subsequent stages can be conditionalized within a specific, localized geometric context. This distinction has significant methodological implications: broad cavity detection tools often identify surface depressions without learning from peptide binding data, while the core requirement of peptide design lies in the representation of binding-related receptor regions, not just any geometrically plausible pocket. The design motivation is to shift from universal pocket geometry to peptide-related interface geometry. For many protein targets, the most useful protein-binding pockets are not deep drug-like ligand pockets; they tend to be extended, irregular, chemically heterogeneous surface patches. Therefore, the question is not merely whether a residue belongs to a cavity, but whether it is located in a receptor region that can reasonably support peptide recognition.

[0034] Methodologically, the core of the site identification stage is not a pre-trained embedding, but a custom residue graph learning framework oriented towards the protein-peptide interface. Its key components include: receptor-only graph construction, consensus interface labeling based on fusion distance rules and interaction heuristics, multi-scale geometric neighborhood design, and a graph attention network (i.e., a network that aggregates information in a graph structure by weighting the importance of neighbors) oriented towards residue-level interface prediction.

[0035] Peptide interface binding site identification is built upon a custom geometric residue mapping method, integrating structural features, interaction-aware features, and ESM2 (Evolutionary Scale Model 2, a protein language model developed by Meta) residue context, and includes ablation analysis of ESM2 and neighborhood selection. ESM2 here provides supplementary residue context information, rather than the method itself.

[0036] In one specific embodiment, the pre-trained graph neural network model is a graph attention network-based GNN model, which is an attention-based graph neural network that uses GATv2 (graph attention v2, GATv2, for aggregating neighbor information on a graph by attention weights) style message passing. In this embodiment, the GNN model takes a mixed feature vector of each amino acid residue as input, constructs a multi-scale graph structure on the Cα coordinates of the residues using the k-nearest neighbor algorithm, where k takes values ​​of 5, 10, and 15, and merges the multi-scale graph structure into a single edge set. The mixed feature vector includes manually constructed structural features and residue evolutionary context embeddings extracted through a pre-trained protein language model. The residue evolutionary context embeddings serve as supplementary contextual information and do not participate in the network architecture design of the GNN model. Specifically, each receptor residue is represented in the current run as a 669-dimensional node feature vector. This representation employs a hybrid design, combining hand-constructed structural features with residue evolutionary context embeddings extracted from a protein language model, rather than relying entirely on a single language model embedding. It includes 29 dimensions of hand-constructed structural features and 640 dimensions of ESM embeddings extracted from the ESM2 model, totaling 669 dimensions. The hand-constructed structural features are purely geometric features, including dihedral angles and accessible surface area. Specifically, the 29 geometric dimensions of the hand-constructed point features are shown in Table 1. Table 1. Geometric features constructed manually in 29 dimensions. In the multi-scale graph structure, each edge corresponds to a 21-dimensional edge feature vector. This edge feature vector combines paired geometric quantities and interaction descriptors. The paired geometric quantities include at least one of distance, direction, and relative position. Specifically, the 21-dimensional edge feature vectors corresponding to each edge are shown in Table 2: Table 2. 21-dimensional edge feature vectors for each edge. Edges are constructed geometrically. The graph builder uses multi-scale k-nearest neighbor connections on the Cα coordinates of residues, k∈{5,10,15}, and merges these scales into a single edge set. The edge feature dimension is 21, combining pairwise geometric quantities and interaction descriptors. Its design aims to allow the model to infer geometric neighborhoods at both the local and mesoscale scales simultaneously, rather than compressing binding site identification into a single fixed contact radius problem.

[0037] The input nodes of the GNN model project the hybrid feature vector into a multidimensional hidden space; The main body of the GNN model uses four stacked multi-scale GAT layers for message passing. Each GAT layer is followed by a batch normalization layer, an ELU activation layer, and a dropout layer in sequence. The batch normalization layer is used as a normalization layer to stabilize training, and the dropout layer randomly masks some neurons during training to reduce overfitting. Residual connections maintain information between the layers.

[0038] The GNN model's binary classification head outputs a prediction of whether each residue is a peptide-binding interface residue, and its multilevel classification head outputs the spatial hierarchy category of each residue. The GNN model has two prediction heads: a binary head for interface classification and a multilevel head for spatial hierarchy prediction. During inference, the binary logits undergo temperature scaling (a method to improve probability calibration by scaling logits, making the predicted probability more consistent with the true accuracy) to achieve calibration.

[0039] In a more detailed embodiment, PocketGNN training employs a multi-task learning strategy, simultaneously optimizing two objectives: binary interface prediction (interface residues vs. non-interface residues) and multi-level interface type classification (edge, support, core). Training uses AdamW (an adaptive optimizer with decoupled weight decay), with an initial learning rate of 0.0003 and a weight decay coefficient of 0.00005. Training lasts for a total of 150 epochs, employing an early stopping strategy (patience=20, min_delta=0.001) to prevent overfitting.

[0040] To address the severe class imbalance problem in peptide-protein interface recognition (interface residues account for only about 14.5% of all residues), this study adopted a three-pronged strategy: First, Focal Loss (for class imbalance) was used instead of standard cross-entropy loss, and by adjusting α=0.25 and γ=2.0, the model was made to focus more on difficult samples that are hard to classify; second, automatic calculation of class weights was implemented, and the weights were dynamically adjusted according to the distribution of each class in the large-scale training data, with the weight range limited to between 0.5 and 10.0; third, a dynamic negative sample sampling strategy was introduced, starting from the 10th epoch, the ratio of negative samples to positive samples in each batch was set to 3:1, forcing the model to learn more discriminative features.

[0041] The learning rate is scheduled using the ReduceLROnPlateau strategy, monitoring the validation set F1 score. If there is no improvement after 10 consecutive epochs, the learning rate is halved, with a minimum learning rate of 1e-6. During end-to-end training, mixed precision acceleration (mixed_precision=true) and gradient clipping (clip=1.0) are used to ensure training stability. Validation is performed every epoch, and the model with the highest validation F1 score is saved as the optimal model.

[0042] The training process of PocketGNN can be broken down into the following key steps: The first step is data batch construction. Since graph neural networks process variable-size graph structures, batch processing employs a concatenation strategy: the node features and edge connections of multiple complexes are concatenated into a large graph, and different samples are distinguished by batch index vectors. This strategy allows for efficient parallel processing but requires all node features to have consistent dimensions. To this end, the geometric features are standardized (mean 0, variance 1).

[0043] The second step is forward propagation computation. The input graph is first constructed using k-NN edges (k=5, 10, 15 at three scales), and then fed into a GATv2 (a second-generation variant of the Graph Attention Network) layer. k=5 is used to capture local, tight biochemical contexts or short-range spatial interactions; k=10 is used to capture medium-range secondary structure features; and k=15 is used to capture long-range tertiary structure folding information. This three-scale fusion can comprehensively characterize the multi-level topological structure of molecules.

[0044] The formula for calculating the attention coefficient for each layer of GATv2 is as follows: e ij = T LeakyReLU[W1· i ||W2· j ||W3· ij ], where LeakyReLU is a function that performs a non-linear mapping on the concatenated vectors. i Let be the node feature vector of node i. j Let be the node feature vector of node j. ij Let W1, W2, and W3 be the feature vectors of the edge between node i and node j, and let W1, W2, and W3 be the learnable weights. Let T be the attention vector, T be the matrix transpose, and e be the unnormalized attention coefficients. ij The final attention coefficient α is obtained after softmax normalization. ij This is used to weight and aggregate the feature information of neighboring nodes. The edge feature vector... ij Features including, but not limited to, the Euclidean distance between node i and node j, spatial angle, differences in amino acid residue types, or biochemical context types.

[0045] The third step is loss calculation and backpropagation. Focal Loss is implemented by multiplying the standard loss by a modulation factor (1-p). γImplementation. Gradients are automatically calculated using PyTorch Autograd. In mixed precision mode, some operations use FP16 to reduce memory usage, but critical accumulations remain at FP32 to ensure numerical accuracy.

[0046] The fourth step is validation and early stopping monitoring. After each epoch, the model performs forward propagation on the validation set (without calculating gradients to save computation), calculating the AUROC, AUPRC (area under the precision-recall curve), and F1 score for each complex, and then calculating the macro-average (i.e., each complex has equal weight). If the macro-average F1 score of the current epoch exceeds the historical high, a model checkpoint is saved. The early stopping counter increments when the F1 improvement is less than 0.001, triggering stopping when it reaches 20.

[0047] The fifth step is dynamic negative sample sampling. Starting from the 10th epoch, each batch of construction first randomly samples an equal number of positive samples (interface residues), then preferentially selects three times the number of difficult cases from the misclassified negative samples of the previous epoch, and finally randomly fills in from the remaining negative samples. This forces the model to learn the most challenging negative samples, significantly improving its ability to identify blurred edge regions.

[0048] Training convergence analysis showed that the validation F1 score improved rapidly in the first 30 epochs (from 0.42 to 0.61), steadily improved from 30 to 80 epochs (to 0.65), improved slowly from 80 to 120 epochs (to 0.6623), and plateaued basically from 120 to 150 epochs. This convergence curve indicates that the model did not suffer from severe overfitting, and the early stopping strategy effectively prevented overtraining.

[0049] The large-scale training data used for GNN model training was constructed as follows: Receptor-peptide complexes derived from the PepBDB (protein-peptide complex database, used for constructing training / validation / test data) were trained using a custom data flow. This flow parsed receptor and peptide structure file formats (PDB format, used to represent 3D structural coordinates), applied quality filters, used multiple complementary standard detection interfaces, calculated residue-level features, and then constructed a PyTorch Geometric graph for each available complex. In the subset-scale run reported in this paper, the flow was executed with max_samples=500; 399 complexes were successfully converted into graphs, and 101 failed the processing filters or interface construction checks. These 399 graphs were divided into 279 training, 59 validation, and 61 test examples. In the fully processed subset, the figure contains 71,961 receptor residues, of which 10,445 (14.5%) are labeled as interface residues.

[0050] The data preprocessing process employs stringent quality control standards: raw data is extracted from the PepBDB database and one or more of the following filtering conditions are applied: peptide sequence length is limited to 5-50 amino acid residues, receptor sequence length is no less than 30 residues, crystal structures with a resolution greater than 3.5 Å are excluded, and each complex must contain at least 3 interface residues; in addition, the presence of non-standard amino acids is allowed to retain more natural peptide variants.

[0051] The labeling of interface residues employs a multi-method consensus strategy, combining distance thresholding and interaction detection. The distance thresholding method sets three threshold layers: a core layer based on heavy atoms (<5 Å), a support layer based on Cβ atoms (<8 Å), and an edge layer based on Cα atoms (<10 Å). The interaction detection method detects hydrogen bonds (distance <3.5 Å), salt bridges (distance <4.0 Å), and hydrophobic interactions (distance <5.0 Å). Final labeling requires confirmation by at least one method (vote_threshold=1), generating multi-level labels: 0 = non-interface, 1 = edge, 2 = support, 3 = core.

[0052] The graph is constructed using the k-nearest neighbor (k-NN) method, where each residue is connected to its five nearest spatial neighbors, while preserving sequential adjacency edges (include_sequential_edges=true). Node features include geometric features (dihedral angles, accessible surface area) and context features from the upper language model (ESM2 embeddings, 640 dimensions). The ESM2 model uses the esm2_t30_150M_UR50D version, inference is performed on the CPU, and feature caching is enabled to avoid redundant computation.

[0053] Interface supervision does not stem from a single hard threshold. Instead, the process combines two types of evidence: distance-based geometric contact rules and interaction-based heuristics. Distance-based labeling uses three nested thresholds: heavy atom contacts define the core interface region (<5 Å), Cβ contacts define the support region (<8 Å), and Cα contacts define the broader edge region (<10 Å). Interaction-based labeling adds presumed hydrogen bonds, salt bridges, and hydrophobic contacts. The final binary interface labels are obtained through consensus voting on the enabled method, while the multi-level labeling preserves the core / support / edge hierarchy for auxiliary supervision. This results in a binary task (interface vs. non-interface) and a coarse spatial organization task.

[0054] The peptide-binding interface is highly imbalanced at the residue level, therefore the training objective is a weighted multi-task loss rather than simple cross-entropy. Binary tasks receive 70% of the loss weight, and spatially hierarchical tasks receive 30%. For binary heads, the model uses focal loss during training with α=0.25 and γ=2.0 to emphasize hard examples and reduce the dominance of the majority non-interface classes. Additional class weights are automatically calculated from the training set and pruned to a bounded range. Optimization uses the adaptive optimizer AdamW (Adam with decoupled weight decay, used to stabilize training and control overfitting) with a learning rate of 3×10⁻⁶. -4 Weight decay 5×10 -5 Gradient clipping (a technique to limit the gradient norm to avoid training instability) is implemented at version 1.0, along with a learning-rate scheduler, ReduceLROnPlateau (which automatically reduces the learning rate when metrics stagnate), monitored on validation F1. Training runs for a maximum of 150 epochs with early stopping (patience=20).

[0055] The second part is the generation of anchor fragments: using the three-dimensional geometric features of the identified protein-binding pocket as input conditions, a trained generation model (i.e., the pocket-conditional backbone generation model) is invoked to simulate and generate anchor fragments with peptide-specific interactions with the protein-binding pocket in situ within the physical boundaries of the protein-binding pocket. The generated anchor fragments can be detached from their original position, allowing for flexible insertion or removal. After generating the anchor fragments, they do not necessarily need to be fixed in situ. Their main purpose is to generate a spatial structure, similar to creating complementary building blocks that can be removed. The removed portion is then used for backbone extension, i.e., the expansion of the complete binding peptide backbone spatial structure, which will be detailed in the next part.

[0056] The number of residues in the anchoring fragment is determined by the generation model based on the geometry of the protein binding pocket, and the anchoring fragment itself (the number and arrangement of residues) is locked as unmodifiable in the subsequent step of expanding the complete binding peptide backbone spatial structure.

[0057] The three-dimensional geometric features of the protein binding pocket include one or more of the following: the Cα coordinates of the protein binding pocket residues, the orientation of the residue side chains, and the local shape description of the protein binding pocket, wherein the local shape description includes the dihedral angle, surface normal vector, curvature, and / or depth of the protein binding pocket. The constraints for the generative model to generate anchored segments include one or more of the following: The Cα coordinates of the anchoring fragment during the generation process are located within the physical boundary of the protein binding pocket (but the generated anchoring fragment can detach from its original position, allowing for flexible insertion or removal). The side chain orientation of the generated anchoring fragment is opposite to the side chain orientation of the hotspot residues in the protein binding pocket; Gradually increase the number of residues in the anchoring fragment until the minimum distance between the newly added residue and the nearest atom in the protein binding pocket is lower than a preset threshold.

[0058] Specifically, the generation model uses the Cα coordinates of the identified protein binding pocket residues and the local dihedral angle as input conditions to simulate the generation of the anchoring fragment. The generation model is trained in the following manner: Binding pocket residues and their corresponding contact motif torsion angles were extracted from the PepBDB protein-peptide complex database to construct training sample pairs. Each sample contained the ESM2 embedding vector of the pocket residue and the ground truth value of the contact motif torsion angle. x 0, where the ESM2 embedding vector is a 640-dimensional pocket representation obtained by mean pooling (including temporal embedding, residue position embedding and pocket embedding), and the torsion angle ground truth is a 6-dimensional representation of each residue encoded by sin / cos; True value of twist anglex 0 Perform forward diffusion noise addition to obtain the noisy twist angle. x t The forward diffusion noise addition method is as follows: x t = Where ε~N(0,1) is standard Gaussian noise; the true value of the torsion angle is given at each time step. x 0 plus noise, This represents the cumulative result of the signal retention rate after adding noise in the first t steps. ,in, Indicates the first i The signal retention rate (its value is less than 1) when noise is added is such that after a certain number of steps, the original signal will decay to near 0.

[0059] The noisy torsion angle x t After being concatenated with temporal embedding, residue position embedding, and pocket embedding, the input is given to a three-layer fully connected decoder (1024→256→128→6) to predict the injected noise ε. θ ( x t ), where the inter-layer activation function of the three-layer fully connected decoder is ReLU; The mean square error between predicted noise and actual noise is ||ε-ε θ ( x t )|| 2 For the training objective, the AdamW optimizer (learning rate 1×10⁻⁶) is used. -4 Weight decay 1×10 -5 The generative model is then trained.

[0060] After one round of training (e.g., 100 epochs), an unconditional control group is constructed by setting the pocket embeddings to zero to verify the contribution of pocket conditionalization to reconstruction accuracy: at the same noise level, the reconstruction error of the conditional model is significantly lower than that of the unconditional model, indicating that the biochemical context information of the pocket residues has a substantial constraining effect on the accurate reconstruction of local torsion angles. Otherwise, the above training steps are repeated.

[0061] The generative model is a diffusion generative model that models in geometric space. The diffusion process uses 1000-2000 time steps, and the optimizer is AdamW. The following explanation uses 1000 time steps as an example. The diffusion generation model projects the input torsion angle tensor (6-dimensional sin / cos representation per residue) into a 256-dimensional hidden space, and concatenates it with the temporal embedding (256-dimensional learnable embedding over 1000 time steps), the residue position embedding (256-dimensional), and the pocket embedding after mean pooling (640-dimensional ESM2 embedding projected to 256-dimensional) into a 1024-dimensional joint representation. The predicted noise is then output through a three-layer fully connected decoder (1024→256→128→6, with ReLU activation used between layers). The training data for the diffusion generation model uses known peptide-protein complex binding motifs from PepBDB or PDB as training ground values. The training objective of the diffusion generation model is to calculate the mean squared error (MSE) loss between the predicted noise and the actual noise. During training, a subset of sample pairs (709 training / 79 validation) are used, each sample containing pocket residue embeddings and the corresponding ground truth contact motif torsion angle. The model learns its ability to reconstruct contact motifs from noisy states by minimizing the MSE on the sin / cos representation of the torsion angle. Once the binding site identification stage identifies candidate receptor regions, the system must propose peptide conformations already biased towards that binding context before full-length topology-aware optimization begins. This is necessary because a main-chain generator that ignores the target pocket would waste most of its search effort on geometrically plausible but biologically irrelevant conformations. Therefore, this stage aims to generate a pocket-conditional torsion angle prior: not the final peptide structure, but a local conformational proposal that narrows the search space for subsequent diffusion stages.

[0062] The generative model is trained using a diffusion model paradigm, but this paper only retains the training facts directly supported by the current repository configuration and implementation. Based on the existing training configuration, the model is trained on uniform peptide structure data, with a diffusion time step T = 1000, an AdamW optimizer, a learning rate of 1 × 10⁻⁴, a weight decay of 0.01, and a total training epoch of 100. The model is the aforementioned three-layer fully connected network with 256 hidden dimensions, using 1000 diffusion time steps. The training objective is the mean squared error (MSE) loss between predicted noise and actual noise.

[0063] The pocket-conditional mainchain generation model is modeled in geometric space rather than sequence space. In the current training configuration, the diffusion process uses 1000 time steps, and the noise schedule is set to linear in the configuration file; the denoising network is the aforementioned three-layer fully connected decoder. Post-training, a geometric refinement step is included, using 3.6–4.0 Å as the validity interval for adjacent Cα bond lengths, an inter-atom conflict threshold of no less than 2.5 Å, and handling loop closure constraints when necessary. Regarding the specific injection method for pocket conditionalization, different script implementations exist in existing repositories; therefore, this paper only confirms that the model uses explicit pocket encoding and conditional generation, without further limiting it to a specific, unverified internal implementation detail. In the current subset-scale experiments, the pocket-conditional mainchain stage is validated using training pairs constructed from protein-peptide complexes. After processing 200 matching PepBDB complexes, 788 usable samples were obtained (709 training, 79 validation). These samples are used to evaluate the reconstruction capability of local mainchain generation under pocket geometry conditions; specific performance results are given in subsequent experimental sections.

[0064] The third part is the extension of the complete peptide backbone spatial structure: the specified topological constraints are the peptide backbone topology type input by the user according to business needs, which includes linear peptides, cyclic peptides or lasso peptides. The peptide backbone topology type is constrained for subsequent backbone extensions of the anchored fragment.

[0065] Before describing the topology-aware optimization phase, a practical but important infrastructure problem must be addressed: peptide backbone data does not exist in a single, model-ready format. The pocket-conditioning backbone phase only provides local, binding-aware torsion proposals for short-contact motifs, which is insufficient for whole peptides, as useful binders must also satisfy sequence length constraints, topology selection, and global backbone consistency outside protected motif regions. Therefore, this phase is responsible for transforming the local motif-conditioning prior into a globally consistent full-length peptide backbone in dihedral space. Linear peptide complexes (from PepBDB) and cyclic peptide sets (from CycPeptMPDB) differ in file structure, topological tags, residue conventions, and the availability of explicit coordinates. Without a normalization layer, any downstream generative model must implicitly learn across mismatched conventions, rather than across the peptide geometry itself. The unified dataset layer is introduced precisely to address this mismatch: it transforms these two structurally heterogeneous data sources—PepBDB linear peptides and CycPeptMPDB cyclic peptides—into a common backbone and torsion angle representation with unified metadata.

[0066] The processing flow follows a phased approach: First, the metadata of the CycPeptMPDB cyclic peptide structure and annotation database is cleaned to extract peptide identifiers, sequences, lengths, and topological annotations. Second, main chain atoms (N, Cα, C) are extracted from the structure file to convert the cyclic peptide structure into a main chain coordinate array (generated via RDKit in 'module3_full_backbone / config.py'). Third, linear peptides are extracted from PepBDB by parsing the peptide.pdb file. Fourth, all main chain coordinate arrays are converted into dihedral angles (φ, ψ, ω) for each residue. Finally, the metadata and arrays are validated and written to a unified output directory. The current version's processing results are stored in 'data / processed / module3_full_backbone_v1', and the linear peptides from PepBDB constitute a unified torsion angle dataset (the current version contains 7,048 linear peptides). The architecture of this data layer supports unifying the two major data sources, PepBDB and CycPeptMPDB, into main chain and torsion angle representations that can be shared by downstream modules.

[0067] The fourth part involves the inference of amino acid probability distributions: In this stage, the topology-sensing optimization phase is implemented as a Transformer-based denoiser (DihedralTransformer) to operate on the peptide dihedral sequence. The state variables are the (φ,ψ) value sequence for each residue (i.e., the two types of key dihedral angles on the main chain; where φ corresponds to C{i-1}-N). i -Cα i -C i ψ corresponds to N i -Cα i -C i -N{i+1}, which together determine the local main chain conformation, and padding to a fixed maximum length. During training, Gaussian noise is added to the target dihedral tensor, and the model is trained to predict the injected noise. The current implementation uses 1000 diffusion steps and a standard denoising MSE target. The core network is a 4-layer Transformer encoder with 8 attention heads and a feedforward dimension 4 times the model width.

[0068] DihedralTransformer is a diffusion model specifically designed for main chain topology optimization. The key difference between it and anchor segment generation lies in its operation space and constraint handling: the anchor segment generation stage generates the initial main chain in the torsion angle space, while the main chain topology optimization stage further denoises and optimizes the topology consistency of the main chain in the same space. Existing implementations explicitly support three types of topology labels: linear, cyclic, and lariat, and preserve the protected motif window during inference. Based on current results, at t=500, the model's noise prediction MSE is approximately 0.04756±0.01156; in the motif fidelity test, the absolute error remains within 10. -7 The magnitude indicates that the protected location can be stably preserved within the range of numerical precision.

[0069] Conditioning is introduced in two ways. First, the model receives peptide length and topological tags as explicit embeddings. Second, motif linear constraints are imposed during inference by identifying protected windows and covering the corresponding φ / ψ positions at each backdiffusion step. Protected motif positions are preserved to floating-point precision. Optimization is performed using AdamW with a learning rate of 10. -4 Weight decay 10 -5 Gradient clipping is set to 1.0.

[0070] The final stage addresses the sequence allocation problem under structural uncertainty. After generating the refined backbone in the topology-sensing optimization stage, the system still faces a many-to-one mapping from geometry to sequence: multiple amino acid choices may be physically reasonable at each position, and this ambiguity is amplified in flexible peptides. Therefore, this stage is designed to output the distribution of residues rather than a single deterministic sequence.

[0071] Formally, the mixed sequence design phase outputs an L×20 matrix P(amino acid|backbone), where each row represents the categorical distribution of 20 standard amino acids at a peptide position. This heatmap is the primary design object. Specific sequences are derived downstream through argmax selection (argmax selects the class with the highest probability; here, it corresponds to selecting the amino acid with the highest probability position-by-position), random sampling, or additional sorting criteria.

[0072] This embodiment utilizes a hybrid sequence of an AI pre-trained model and a physical prior to infer the probability distribution of amino acids. Essentially, it's a probabilistic fusion framework, rather than an end-to-end trainable neural network. Its core task is to infer the probability distribution of 20 amino acids at each position given the main chain geometry. The current implementation comprises two main components: first, the AI ​​pre-trained model, which in this embodiment employs an ESM-IF1 inverse folding predictor to generate amino acid probabilities conditionally from the structure; and second, the physical prior, which in this embodiment is an empirical Laplace plot prior, used to provide a baseline distribution based on dihedral statistics when the ESM-IF1 output is insufficient or ineffective.

[0073] In the current repository implementation, no new neural network is trained; instead, inference-level fusion is performed. Specifically, when ESM-IF1 can output an effective probability distribution of multiple amino acids at the i-th site on the peptide backbone, such as when the sum, mean, maximum, or minimum value of the elements in its probability distribution matrix reaches a preset probability threshold p... th In one specific embodiment of the present invention, the probability threshold p th The set value can be lower than 0.5, or even set to 0.1 or 0.05. For example, when the mean of the probability distribution matrix elements output by ESM-IF1 is greater than or equal to 0.2, the normalized effective probability distribution result output by ESM-IF1 is then fused with the Ramachandran prior according to a preset ratio. The corresponding calculation formula is: P final (a|i)=α×P esm '(a|i)+β×P rama (a|i), where P esm '(a|i) represents the normalized result of the effective probability distribution of amino acids at the i-th site on the ESM-IF1 output peptide backbone; α and β are weights, which sum to 1, and α > β; fixed weight fusion is used in the mixed sequence design stage instead of dynamic entropy weighting.

[0074] In another embodiment, the probability threshold p th The set value is between 0 and 0.9. For example, when the sum of the elements of the probability distribution matrix output by ESM-IF1 is greater than or equal to 0.5, the output of ESM-IF1 is normalized and then fused with the Ramachandran prior according to a preset ratio. In another embodiment, the probability threshold p th The set value is between 0.1 and 0.9. For example, if the minimum value in the probability distribution matrix output by ESM-IF1 is greater than or equal to the probability threshold of 0.15, then the normalized output of ESM-IF1 will be fused with the Ramachandran prior according to a preset ratio. The probability threshold p can also be set... thThe set value is between 0.2 and 1. For example, if the maximum value in the probability distribution matrix output by ESM-IF1 is greater than or equal to the probability threshold of 0.25, then the output of ESM-IF1 is normalized and fused with the Ramachandran prior according to a preset ratio. The above probability threshold p th The specific values ​​are merely examples and should not be used as a basis for limiting the scope of protection of this invention.

[0075] When the output of ESM-IF1 is approximately zero at a certain position, i.e. less than the probability threshold p th Taking the above four embodiments as examples, they are as follows: when the mean of the elements of the probability distribution matrix output by ESM-IF1 is less than 0.2, or when the sum of the elements of the probability distribution matrix output by ESM-IF1 is less than 0.5, or when the minimum value of the elements of the probability distribution matrix output by ESM-IF1 is less than 0.15, or when the maximum value of the elements of the probability distribution matrix output by ESM-IF1 is less than 0.25, then it directly falls back to the Ramachandran prior, i.e., P final (a|i)=P rama (a|i). Fixed residue constraints are achieved through one-hot overlay, which replaces the 20-dimensional probability distribution at a specified position with the one-hot vector of the corresponding amino acid, thereby ensuring that the constraints are precisely satisfied.

[0076] The amino acid with the highest probability in the probability distribution of the fusion inference is determined as the amino acid at that site to generate the binding peptide.

[0077] The de novo method for generating protein target-binding peptides provided by this invention further includes the following phased verification steps: After identifying the protein binding pocket based on the fusion features of geometric and language models, residue-level classification indicators are used for evaluation. These residue-level classification indicators include precision, recall, F1 score, area under the receiver operating characteristic curve, and area under the precision-recall curve. The protein binding pockets predicted by the selected model will be scored accordingly, which will be used by the user to select the target peptide binding pocket. Only after the user selects the target binding pocket can the generation stage of the anchoring fragment begin.

[0078] After the anchoring fragment is generated, it is evaluated based on geometric validity and structural diversity. This evaluation is available for user selection and confirmation. Only after the user selects it will the business input constraint stage begin.

[0079] After obtaining the complete binding peptide backbone spatial structure, it is evaluated through one or more of the following dimensions: using Laplace plot compliance to check whether the dihedral angles of the backbone are located in the allowed region; using topological satisfaction to verify whether the topological constraints are met; and / or, using motif fidelity to verify whether the reconstruction accuracy of the conservative regions of the backbone meets the standard; only after this evaluation is passed can the amino acid probability distribution inference stage be entered. After inferring the amino acid probability distribution, the system evaluates the results based on sequence recovery rate, amino acid distribution rationality, and physical validity; it then provides the user with the amino acid probability for each site. The system recommends binding peptide sequences, and users can also modify sequence combinations based on the probability heatmap. After obtaining the protein target-binding peptide, it was evaluated through end-to-end case studies, generation diversity, structural novelty, and computational efficiency relative to the baseline process.

[0080] In another embodiment of the present invention, a de novo peptide generation system for protein target binding based on the fusion of geometric features and language models and motif anchoring diffusion is provided, such as... Figure 2 As shown, the system includes the following modules: The pocket recognition module is configured to identify protein-binding pockets on the 3D structure of a target protein by using a pre-trained graph neural network model based on features fused from geometric and linguistic models. An anchoring module is configured to generate complementary anchoring fragments based on the three-dimensional geometry of the identified protein-binding pocket, the anchoring fragments being short peptide motifs comprising one or more amino acid residues. An extension module is configured to extend the anchoring fragment into a complete binding peptide backbone spatial structure according to specified topological constraints. The probability distribution inference module is configured to infer the probability distribution of a preset number of amino acids at each site on the complete binding peptide backbone spatial structure. The generation module is configured to select suitable amino acids for each site on the complete binding peptide backbone spatial structure according to the probability distribution results, thereby obtaining a protein target binding peptide.

[0081] It should be noted that the de novo protein target binding peptide generation system based on geometric feature and language model fusion and motif anchoring diffusion provided in this embodiment of the invention belongs to the same inventive concept as the de novo protein target binding peptide generation method based on geometric feature and language model fusion and motif anchoring diffusion provided in the above embodiment. Hereinafter, all the contents of the embodiment of the de novo protein target binding peptide generation method based on geometric feature and language model fusion and motif anchoring diffusion are incorporated into the embodiment of the de novo protein target binding peptide generation system based on geometric feature and language model fusion and motif anchoring diffusion by reference.

[0082] In another embodiment of the present invention, a method for de novo antibody generation targeting a protein is provided, wherein the specified topological constraint is the peptide backbone topology type of a linear peptide. Based on the topological constraints, the anchored segment is expanded into a CDR; Users can embed a CDR into the frame region backbone of the constrained antibody by uploading constraints on the frame region backbone, such as IgG, IgA, IgD, IgE, IgM, or VHH single-domain antibodies.

[0083] This embodiment shares the same inventive concept as the de novo protein target-binding peptide generation method based on geometric feature and language model fusion and motif anchoring diffusion. Therefore, the entire contents of the protein target-binding peptide de novo generation method embodiment are incorporated into this de novo antibody generation method embodiment by reference.

[0084] Using the de novo method for generating protein target-binding peptides based on the fusion of geometric features and language models and motif anchoring diffusion provided in the above embodiments, generating binding peptides requires evaluating whether geometric information can be forward-propagated from target identification to main chain proposal, topology-aware optimization, and final sequence design to generate coherent candidate binders. The corresponding experiments and results are as follows: Case Study 1: HER2 (1N8Z) The HER2 example provides a receptor structure, a saved pocket prediction JSON, the designed peptide backbone product, and a dihedral trajectory. Replaying the downstream process yields an 18-residue amino acid heatmap, as shown below. Figure 3 As shown, this leads to the maximum probability representative sequence and several high-probability alternative candidates. More importantly, this stage outputs a structured and non-uniform sequence probability distribution conditionally determined by the design main chain and the receptor's local binding context, providing a basis for subsequent candidate ranking and experimental screening.

[0085] In validating the AF3 complexes of HER2 candidates, different candidates showed differences in ranking scores, ipTM (interface prediction TM score), average pLDDT (predicted local distance difference test score), and cross-chain PAE (predicted alignment error). Current results indicate that at least some candidates outperform others in overall structural plausibility, but these AF3 outputs should be interpreted as external signals of structural plausibility rather than experimental evidence. A more robust conclusion is that this process can generate geometrically evaluable and rankable HER2 candidate complexes, with the strongest HER2 candidate's AF3 complex model shown below. Figure 4 As shown.

[0086] Case Study 2: α-Synuclein (1XQ8) The 1XQ8 example provides the receptor structure, the saved pocket prediction, and the dihedral trajectory. After replaying the sequence assignment phase, a 12-residue heatmap can be obtained, as shown below. Figure 5 As shown, the maximum probability representative sequence and multiple sampled alternative candidates are derived from this. Similar to the HER2 case, this output is better interpreted as a structured design distribution rather than a single deterministic recommendation, because the main goal at this stage is to provide a candidate space for downstream screening, rather than directly fixing a unique sequence here.

[0087] In the validation of AF3 in 1XQ8 candidates, different candidates did not show complete consistency in overall ranking scores and interface-related signals: some candidates had higher overall complex confidence, while others exhibited relatively more favorable interface-oriented indicators. Therefore, these results are more suitable as a basis for candidate ranking and subsequent experimental prioritization, rather than being directly interpreted as the final biological conclusion for a single candidate. The highest-ranked 1XQ8 candidate's AF3 complex model is shown below. Figure 6 As shown.

[0088] Furthermore, to demonstrate that the system can generate multiple distinct candidates from the same target and motif constraints, this study employed a topology-aware optimization stage to generate five peptide backbones sharing a protected motif window but with different diffusion random seeds. A hybrid sequence design stage was then run in a hybrid mode to obtain position-by-position amino acid heatmaps, such as... Figure 7 As shown. In these five samples, the resulting main chain exhibits statistically significant diversity: the average pairwise Cα RMSD (root mean square deviation) for the off-diagonal lines is 4.62 Å. The corresponding heatmaps yielded distinct greedy representative sequences (sequences obtained by taking the maximum probability digit by digit). These five random samples are merely for illustrative purposes and are separate from the aforementioned real target AF3 case study candidates.

[0089] To obtain external, structure-based plausibility checks, PEPAI2 design peptide candidates from the two case studies were submitted to the AlphaFold3 (AF3) server for complex prediction. The validation summary of all candidate ranking scores, ipTM, and average pLDDT from the 1XQ8 and HER2 case studies is as follows: Figure 8 As shown. These are not experimental binding measurements, nor are they considered proof of binding. Rather, they provide independent plausible signals for the formation of the peptide-receptor complex and help to rank available candidates for further investigation.

[0090] For the two real-world target cases, the AF3 output provides a non-experimental structural plausibility check overall. Different candidates differ in metrics such as ranking score, ipTM, pLDDT, and cross-chain PAE, but overall, no significant overall distortions or patterns severely incompatible with the target were observed. Therefore, these AF3 results are more suitable for plausibility filtering and candidate prioritization than for direct surrogate indicators of binding affinity or experimental validity.

[0091] After establishing system-level rationale through case studies and AF3 validation, the report explains component-level results that illustrate process behavior. Starting from the binding site identification phase, it provides geometrically conditional signals for all downstream generation.

[0092] On the preserved test segmentation (61 images, approximately 1,600 interface residues out of 11,000 total receptor residues), the binding site recognizer achieved AUROC 0.9085 and AUPRC 0.7066 (see [link to test results]). Figure 9 At the selected operational threshold (0.40), the model achieved a precision of 0.6123, a recall of 0.7212, and an F1 score of 0.6623. These metrics indicate that the detector can recover most of the true interfaces without overwhelming downstream stages with false positives.

[0093] From a system architecture perspective, considering the minimum requirements that the site identification stage must meet: the downstream process does not need a perfect pocket oracle, but it does need a sufficiently compact and informative pocket representation to bias subsequent generation. The fact that the main chain stage strongly benefits from pocket conditionation indicates that the learned representation generated in conjunction with the site identification stage is already meaningful for generation.

[0094] To verify the contributions of each component of the GNN in predicting the pocket, the following systematic ablation experiments were conducted. The K-nearest neighbor sensitivity experiment used three configurations: the baseline configuration (k=5, 10, 15) achieved an F1 score of 0.6623 and an AUROC of 0.9085 on the test set; after expanding to (k=10, 20, 30), the F1 score decreased to 0.6307 and the AUROC decreased to 0.8956; after further expanding to (k=15, 30, 45), the F1 score rebounded to 0.6565 and the AUROC was 0.8973. These results indicate that the model has some robustness to the neighborhood range, but (k=5, 10, 15) still performs best overall.

[0095] ESM2 feature ablation experiments show that after removing the ESM embedding, the model's F1 score on the test set drops to 0.5735, AUROC to 0.8820, AUPRC to 0.6460, precision to 0.5271, and recall to 0.6287. These results indicate that the geometric graph method itself already carries a considerable discriminative signal, and the ESM residue context further provides stable performance gains on top of this. Therefore, the site identification stage relies on a complementary combination of geometric representation and language model priors, rather than a complete dependence on any single component.

[0096] From a systems perspective, the significance of these ablation results lies not only in the changes in classification metrics, but also in their direct impact on the quality of downstream conditionalization generation. More stable pocket representations help provide more reliable local geometric priors for the main chain stage of pocket conditionalization and reduce the amplification effect of erroneous pocket signals on the subsequent topology-aware optimization stage.

[0097] The pocket-conditioning backbone stage addresses a key bottleneck in end-to-end peptide design: generating candidate backbones compatible with the local biochemical context of the target-binding pocket. Training pairs constructed from 200 PepBDB complexes yielded 788 samples (709 training / 79 validation). Evaluation on the 79 validation samples at a moderate noise level (t=500) showed... Figure 10 As shown, the MSE of the reconstructed pocket-conditional model is 0.4478±0.1423, with a median of 0.4474; the MSE of the unconditional version after zeroing out the pocket embedding is 1.9302±0.5732, with a median of 1.8852. The corresponding relative error is reduced to 76.8%.

[0098] Across multiple diffusion time steps, the reconstruction gap between the conditional and unconditional models widens with increasing noise levels, such as... Figure 11 As shown, pocket contexts become increasingly valuable when the tortuous state is more weakly specified. This is precisely the regime associated with downstream generation processes where the main chain state is not known in advance and requires local guidance to keep the sampling close to a binding-compatible conformation.

[0099] After using the training index generated by the unified data layer, the DihedralTransformer's evaluation on preserving the test split shows that it can stably learn the noise prediction task in dihedral space. The current unified dataset (`module3_full_backbone_v1`) contains a total of 7,048 linear peptides, divided into 5,638 training peptides, 705 validation peptides, and 705 test peptides according to `split_summary.json`. At t=500, the model's noise prediction MSE is approximately 0.04756±0.01156, with a median of approximately 0.05086. This result demonstrates that the topology-aware dihedral denoiser has stable training performance on the current unified dihedral space dataset and provides a usable foundation for subsequent motif preservation and full-length backbone optimization.

[0100] The training process of DihedralTransformer, as clearly confirmed in the current repository, includes the following: the model uses a separate training script and is trained on a uniform twist angle dataset; the optimizer is AdamW, and the learning rate is 1×10⁻⁶. -4 The weight decays to 1×10 -5 The gradient clipping was set to 1.0. The main body of the model is a 4-layer Transformer encoder with 8 attention heads and an embedding dimension of 128. Training employed 1000 diffusion steps, with the mean squared error of the noise prediction as the direct optimization objective. According to the current results file, at t=500, the model achieved a noise prediction MSE of 0.04756±0.01156 with a median of 0.05086 on the preserved test split; in the order fidelity test, the mean absolute error was 9.27×10⁻⁶. -8 The maximum error is 2.38 × 10⁻⁶. -7 Regarding finer-grained constraint sampling strategies, additional loss designs, or alternating training mechanisms, existing evidence is insufficient, so this paper will not present them as confirmed facts.

[0101] To verify the sequence conditionalization, inference was run with the main chain support window fixed and hard protection enabled. For example... Figure 12 As shown, the absolute sequence error remains at 2.4 × 10⁻⁶. -7 Below radians, it indicates that the sampler essentially preserves the protected motif position precisely under the implemented hard constraint scheme.

[0102] This is an important systemic result: it means that the local binding geometry proposed in the pocket-conditional main chain stage can be incorporated into full-length generation without being diluted or averaged out by the optimization model. From a system perspective, the topology-aware optimization stage is a crucial stage for PEPAI2 to transform from a local binding proposal system into a full-length sequence design system.

[0103] The hybrid sequence design phase transforms the refined backbone into a position-wise amino acid probability distribution by combining empirical Laplace plot priors and ESM-IF1 structure-conditional reverse folding. In the current example run, the hybrid predictor stably generates a well-formed L×20 probability matrix, providing a foundation for subsequent position-wise maximization, sampling generation, and candidate ranking, such as... Figure 13 As shown. The core value of this stage lies in outputting an interpretable probability distribution, rather than prematurely compressing the design results into a uniquely deterministic sequence.

[0104] These metrics indicate that the resulting heatmaps are significantly more informative than a uniform baseline, while still expressing uncertainty in appropriate places. Replays of real-world target examples (HER2, α-synuclein) and end-to-end random runs both produced well-formed heatmaps with reasonable entropy, demonstrating that the mixing formula is stable across different main chain sources. The main failure mode in the mixing sequence design phase is interpretability rather than computation: the heatmap output is a distribution rather than a single sequence, therefore downstream users must decide how to sample or rank candidates.

[0105] The experiments and results above demonstrate that peptide design can be organized as a binding-prioritized, geometry-aware system, where target information is forward-propagated through binding site identification, main chain proposal, topology-aware optimization, and heatmap-based sequence design. The resulting process is not merely a collection of modules; it is a coherent design framework that generates structured intermediates and preserves them across stages. In subset-scale validation, the system achieved strong binding site identification performance (AUROC 0.9085), learned a stable pocket-conditional main chain denoiser (76.8% error reduction), and generated informative amino acid distributions for sequence selection. End-to-end case studies on real targets and AF3 plausibility checks further demonstrate that the system can generate diverse and interpretable peptide candidates for downstream evaluation.

[0106] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0107] The above description is only a specific embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for de novo generation of protein target-binding peptides based on the fusion of geometric features and language models and motif anchoring diffusion, characterized in that, Includes the following steps: Receives the 3D structure of the target protein; The protein-binding pockets on the 3D structure of the target protein are identified using a pre-trained graph neural network model based on the fusion of geometric and linguistic features. Based on the three-dimensional geometric features of the identified protein binding pocket, a complementary binding anchoring fragment is generated, wherein the anchoring fragment is a short peptide motif comprising one or more amino acid residues. Based on the specified topological constraints, the anchoring fragment is expanded into a complete binding peptide backbone spatial structure; The probability distribution of multiple amino acids at each site on the complete binding peptide backbone spatial structure is inferred, and suitable amino acids are selected for each site on the complete binding peptide backbone spatial structure according to the probability distribution results, thereby obtaining the protein target binding peptide.

2. The method for de novo generation of protein target-binding peptides according to claim 1, characterized in that, Anchored fragments are generated using the following method: Using the three-dimensional geometric features of the identified protein binding pocket as input, the trained generative model is invoked to simulate and generate an anchoring fragment with peptide-specific interaction with the protein binding pocket in situ within the physical boundary of the protein binding pocket, and the generated anchoring fragment can detach from the original situ. The number of residues in the anchoring fragment is a variable value, which is determined by the generation model based on the geometry of the protein binding pocket, and the anchoring fragment itself is locked as unmodifiable in the subsequent step of expanding the complete binding peptide backbone spatial structure.

3. The method for de novo generation of protein target-binding peptides according to claim 2, characterized in that, The three-dimensional geometric features of the protein binding pocket include one or more of the following: the Cα coordinates of the protein binding pocket residues, the orientation of the residue side chains, and the local shape description of the protein binding pocket, wherein the local shape description includes the dihedral angle, surface normal vector, curvature, and / or depth of the protein binding pocket. The constraints for the generative model to generate anchored segments include one or more of the following: The Cα coordinates of the anchoring fragment during the generation process are located within the physical boundary of the protein binding pocket; The side chain orientation of the generated anchoring fragment is opposite to the side chain orientation of the hotspot residues in the protein binding pocket; Gradually increase the number of residues in the anchoring fragment until the minimum distance between the newly added residue and the nearest atom in the protein binding pocket is lower than a preset threshold.

4. The method for de novo generation of protein target-binding peptides according to claim 2, characterized in that, The generative model uses the Cα coordinates of the identified protein binding pocket residues and the local dihedral angle as input conditions to simulate the generation of the anchoring fragment. The generative model is trained in the following way: The binding pocket residues and their corresponding contact motif torsion angles were extracted from the PepBDB protein-peptide complex database to construct training sample pairs. Each sample contains the ESM2 embedding vector of the pocket residue and the ground truth value of the contact motif torsion angle. True value of twist angle x 0 Perform forward diffusion noise addition to obtain the noisy twist angle. x t ; The noisy torsion angle x t After being concatenated with temporal embedding, residue position embedding, and pocket embedding, the input is fed into a three-layer fully connected decoder to predict the injected noise ε. θ ( x t ); The generative model was trained using the AdamW optimizer with the mean square error between the predicted noise and the actual noise as the training objective.

5. The method for de novo generation of protein target-binding peptides according to any one of claims 2 to 4, characterized in that, The generative model is a diffusion generative model that models in geometric space. The diffusion process uses 1000-2000 time steps, and the optimizer is AdamW. The diffusion generation model projects the input torsion angle tensor (represented by 6-dimensional sin / cos per residue) into a 256-dimensional hidden space, and concatenates it with the temporal embedding, residue position embedding, and pocket embedding after mean pooling to form a joint representation. The prediction noise is then output through a three-layer fully connected decoder. The training data for the diffusion generation model uses known peptide-protein complex binding motifs from PepBDB or PDB as training ground values. The training objective of the diffusion generation model is the mean square error (MSE) loss between the predicted noise and the actual noise.

6. The method for de novo generation of protein target-binding peptides according to claim 1, characterized in that, The specified topological constraints are the peptide backbone topology type input by the user, which includes linear peptides, cyclic peptides, or lasso peptides. The peptide backbone topology type is constrained for subsequent backbone extensions of the anchored fragment.

7. The method for de novo generation of protein target-binding peptides according to claim 6, characterized in that, The specified topological constraint is the peptide backbone topology type of a linear peptide; Based on the topological constraints, the anchored segment is expanded into a CDR; The CDR is embedded into the frame region backbone of a user-specified antibody, which includes IgG, IgA, IgD, IgE, IgM, and VHH single-domain antibodies.

8. The method for de novo generation of protein target-binding peptides according to claim 1, characterized in that, Inferring the probability distribution of amino acids using a mixture of AI pre-trained models and physical priors, including: The probability distribution of multiple amino acids at the i-th site on the peptide backbone is predicted using an artificial intelligence pre-trained model, denoted as P1(a|i). The probability distribution of multiple amino acids at the i-th site on the peptide backbone is output using a physical prior model, denoted as P2(a|i). The probability distributions P1(a|i) and P2(a|i) are weighted and summed to obtain the amino acid probability distribution at the i-th site in the fusion inference. The amino acid with the highest probability in the probability distribution results of the fusion inference is then identified as the amino acid at that site.

9. The method for de novo generation of protein target-binding peptides according to claim 8, characterized in that, The artificial intelligence pre-trained model is the ESM-IF1 inverse folding predictor, and the physical prior model is the Ramachandran prior model. The weighted summation methods include: If the maximum and minimum values ​​of the elements in the probability distribution matrix of the amino acid at the i-th site output by the ESM-IF1 inverse folding predictor are less than the preset probability threshold p th The probability distribution of the fusion inference is then calculated using the following formula: P final (a|i)=P2(a|i); If the maximum and minimum values ​​of the elements of the amino acid probability distribution matrix at the i-th site output by the ESM-IF1 inverse folding predictor reach the preset probability threshold p th The probability distribution of the fusion inference is then calculated using the following formula: P final (a|i)=α×P1'(a|i)+β×P2(a|i), where P1'(a|i) is the normalized result of the probability distribution P1(a|i); α and β are weights, which sum to 1, and α>β.

10. The method for de novo generation of protein target-binding peptides according to claim 1, characterized in that, It also includes the following phased verification steps: After identifying the protein binding pocket based on the fusion features of geometric and language models, residue-level classification indicators are used for evaluation. These residue-level classification indicators include precision, recall, F1 score, area under the receiver operating characteristic curve, and area under the precision-recall curve. After generating the anchored segments, they are evaluated based on geometric validity and structural diversity. After obtaining the complete binding peptide backbone spatial structure, it is evaluated through one or more of the following dimensions: using Laplace plot compliance to check whether the dihedral angles of the backbone are located in the allowed region; using topological satisfaction to verify whether the topological constraints are met; and / or, using motif fidelity to verify whether the reconstruction accuracy of the conservative regions of the backbone meets the standard. After inferring the amino acid probability distribution, the results are evaluated based on sequence recovery rate, rationality of amino acid distribution, and physical validity. After obtaining the protein target-binding peptide, it was evaluated through end-to-end case studies, generation diversity, structural novelty, and computational efficiency relative to the baseline process.

11. The method for de novo generation of protein target-binding peptides according to claim 1, characterized in that, The pre-trained graph neural network model is a GNN model based on graph attention network. It takes the mixed feature vector of each amino acid residue as input and constructs a multi-scale graph structure on the Cα coordinate of the residue using the k-nearest neighbor algorithm, where k takes values ​​of 5, 10, and 15. The multi-scale graph structure is then merged into a single side set. The mixed feature vector includes manually constructed structural features and residue evolution context embeddings extracted by a pre-trained protein language model. The input nodes of the GNN model project the hybrid feature vector into a multidimensional hidden space; The main body of the GNN model uses four stacked multi-scale GAT layers for message passing, with each GAT layer followed by a batch normalization layer, an ELU activation layer, and a dropout layer in sequence. The binary classification head of the GNN model outputs a prediction of whether each residue is a peptide-binding interface residue, and the multi-level classification head of the GNN model outputs the spatial hierarchy category to which each residue belongs.

12. The method for de novo generation of protein target-binding peptides according to claim 11, characterized in that, The hybrid feature vector for each amino acid residue is a 669-dimensional node feature vector, which includes 29-dimensional hand-constructed structural features and 640-dimensional ESM embeddings extracted by the ESM2 model. The hand-constructed structural features are pure geometric features, including dihedral angles and accessible surface area. In the multi-scale graph structure, each edge corresponds to a 21-dimensional edge feature vector, which combines a pair of geometric quantities and an interaction descriptor. The pair of geometric quantities includes at least one of distance, direction, and relative position.

13. The method for de novo generation of protein target-binding peptides according to claim 11, characterized in that, The training methods for the GNN model include: Raw data were extracted from the PepBDB database and one or more of the following filtering conditions were applied: peptide sequence length was limited to 5-50 amino acid residues, receptor sequence length was not less than 30 residues, crystal structures with a resolution greater than 3.5 Å were excluded, and each complex was required to contain at least 3 interface residues. A combination of distance thresholding and interaction detection methods is used to label interface residues to generate multi-level labels. The distance thresholding method sets corresponding thresholds for the core layer based on heavy atoms, the support layer based on Cβ atoms, and the edge layer based on Cα atoms. The interaction detection method detects hydrogen bonds, salt bridges, and hydrophobic interactions. A multi-task learning strategy is adopted for training, which simultaneously optimizes binary interface prediction and multi-level interface type classification. The AdamW optimizer is used, with focus loss as the loss function, and a dynamic negative sample sampling strategy is introduced. This strategy sets the ratio of negative samples to positive samples in each batch to n:1 starting from the Nth training round, where N is an integer between 5 and 120, and n is between 1.1 and 5.

14. A de novo system for generating protein target-binding peptides based on the fusion of geometric features and language models and motif-anchored diffusion, characterized in that, Includes the following modules: The pocket recognition module is configured to identify protein-binding pockets on the 3D structure of a target protein by using a pre-trained graph neural network model based on features fused from geometric and linguistic models. An anchoring module is configured to generate complementary-binding anchoring fragments based on the three-dimensional geometry of the identified protein-binding pocket, the anchoring fragments being short peptide motifs comprising one or more amino acid residues. An extension module is configured to extend the anchoring fragment into a complete binding peptide backbone spatial structure according to specified topological constraints. The probability distribution inference module is configured to infer the probability distribution of a preset number of amino acids at each site on the complete binding peptide backbone spatial structure. The generation module is configured to select suitable amino acids for each site on the complete binding peptide backbone spatial structure according to the probability distribution results, thereby obtaining a protein target binding peptide.

Images