A method for targeted polypeptide design based on conditional latent diffusion model and specificity feedback
By combining a conditional latent diffusion model with a specific feedback mechanism, and integrating target structure characterization and peptide sequence encoding, a conditional latent diffusion network was constructed. This solved the problems of global consistency and target specificity in targeted peptide design, and enabled the precise generation of high-affinity peptides.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHWESTERN POLYTECHNICAL UNIV
- Filing Date
- 2026-04-27
- Publication Date
- 2026-07-31
AI Technical Summary
Existing peptide design methods struggle to balance global sequence consistency and target structure specificity in targeted peptide design, and the affinity of peptide-target interactions cannot be assessed in real time during the generation process, resulting in insufficient design efficiency and accuracy.
A conditional latent diffusion model and a specific feedback mechanism were adopted. The target pocket structure characterization was extracted by the ProteinMPNN model and peptide sequence encoding was combined with the ESMC protein language model to construct a conditional latent diffusion network. Self-attention and residual-level cross-attention networks were introduced for noise reduction, and a specificity monitor was used for multi-task training to optimize the peptide generation process.
This technology improves the structural matching and affinity of peptides to target pockets in targeted peptide design, generating highly accurate and controllable peptide sequences, thereby enhancing design efficiency and effectiveness.
Smart Images
Figure CN122493952A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, and in particular to a targeted peptide design method based on a conditional latent diffusion model and specific feedback. Background Technology
[0002] Peptides are biomolecules composed of 50 or fewer amino acid residues and have gradually developed into an important class of therapeutic drugs in recent years, especially targeted peptide drugs, which play a vital role in disease treatment due to their high specificity and safety. However, traditional peptide design methods mainly rely on high-throughput experimental screening, a strategy with limited efficiency when faced with the vast peptide space. Recently, generative artificial intelligence, represented by protein language models and diffusion models, has driven a technological revolution in targeted peptide design.
[0003] With the significant success of pre-trained models in the field of natural language processing, protein language models have gradually become a key technology in peptide design and protein engineering. Protein language models such as ESM2 and ProGen, through large-scale self-supervised pre-training, effectively capture the implicit co-evolutionary patterns between amino acid residues. However, existing protein language models have significant limitations: on the one hand, while masked language models possess excellent sequence representation and long-range dependency modeling capabilities, their generation capabilities are limited; autoregressive models, while achieving high-throughput sequence generation, struggle to maintain global sequence consistency due to unidirectional causal constraints. On the other hand, diffusion models (such as RFdiffusion) perform well in protein structure design, but current structure-based diffusion strategies (such as HYDRA and PepMimic) often overemphasize geometric topological relationships, neglecting evolutionary constraints at the sequence level. In recent years, latent diffusion models have emerged, combining the representational capabilities of protein language models with the generation capabilities of diffusion models. However, existing latent diffusion models are mainly applied to functional peptide design, failing to explicitly introduce target structure specificity, especially neglecting the three-dimensional geometric constraints of the target pocket, thus limiting their application in targeted peptide design. Furthermore, most generative frameworks struggle to assess the affinity of peptide-target interactions in real time during the denoising process, thus making it impossible to dynamically adjust and optimize the trajectory.
[0004] Therefore, it is necessary to improve one or more of the problems existing in the above-mentioned related technical solutions.
[0005] It should be noted that this section is intended to provide background or context for the technical solutions of this disclosure as set forth in the claims. The description herein does not constitute an admission that it is prior art simply because it is included in this section. Summary of the Invention
[0006] The purpose of this invention is to provide a targeted peptide design method based on a conditional latent diffusion model and specific feedback, thereby overcoming, to at least some extent, one or more problems caused by the limitations and defects of related technologies.
[0007] This invention provides a targeted peptide design method based on a conditional latent diffusion model and specificity feedback, comprising: S1, Target pocket structure characterization extraction: The three-dimensional structure of the protein pocket is encoded by the ProteinMPNN model to obtain the characterization vector of the pocket structure. S2, Natural polypeptide sequence characterization and extraction: Natural polypeptide sequences are encoded using the ESMC protein language model to obtain the latent space characterization of the polypeptide sequence. S3, Construction of Conditional Latent Diffusion Network Model: Based on the diffusion transformer framework, a conditional latent diffusion network model is constructed using the representation of the three-dimensional structure of the protein pocket as conditional information. This conditional latent diffusion network model includes multiple denoising modules. Each denoising module includes a self-attention network, a residual-level cross-attention network, and a pointwise feedforward network. The residual-level cross-attention network aims to map the representation of the protein pocket as a condition to the diffusion network. S4, Forward Diffusion and Reverse Denoising: Based on the conditional latent diffusion model framework, Gaussian noise is gradually added to the latent space representation of the peptide sequence; then, the conditional latent diffusion network is used to perform reverse iterative denoising along the time step, thereby gradually estimating and eliminating the Gaussian noise in the latent space representation of the peptide sequence to generate the latent space representation of the target peptide. S5, Sequence Decoding: Construct a sequence decoder, which includes a one-dimensional convolutional layer and a multilayer perceptron, with each layer using the ReLU function as the activation function; then, use the pre-trained sequence decoder to convert the latent space representation of the target peptide output in S4 into the corresponding target peptide sequence. S6, Specificity Assessment and Optimization: Based on the Transformer framework, a specificity monitor is constructed, which includes a self-attention mechanism, a cross-attention mechanism, and a multilayer perceptron. Then, the specificity of the peptide space generated by the model is evaluated using the pre-trained specificity monitor to obtain the affinity loss value. The affinity loss and diffusion loss are fused together for multi-task joint training to achieve iterative optimization of the model. Finally, the trained conditional latent diffusion network model and sequence decoder are used to obtain a target peptide sequence with high affinity.
[0008] In this invention, in step S4, the latent space representation of the peptide is progressively denoised at the diffusion time step, and a reverse denoising process is performed through a conditional latent diffusion network model. Specifically, the latent space representation of the peptide is modeled by a self-attention network and input into the conditional latent diffusion network model, and the pocket structure representation is introduced into the conditional latent diffusion network model through a residue-level cross-attention network to apply conditional constraints during the denoising process.
[0009] In this invention, in S3, the self-attention network, the residual-level cross-attention network, and the point-by-point feedforward network are all provided with residual connection layers and normalization layers.
[0010] In this invention, in S4, during the diffusion denoising process, the mean square error between the predicted noise and the actual noise is used as the loss function.
[0011] In this invention, in step S5, before training the diffusion model, the sequence decoder is trained from scratch using the cross-entropy loss function based on the PepMerge dataset.
[0012] In this invention, the training process of the specificity monitor in step S6 includes: Based on the PepMerge dataset, the natural peptide sequence is randomly perturbed, and the peptide structure is predicted using ESMfold. At the same time, combined with the database in ppbench2024, affinity scores are calculated using HDOCK software to construct a protein pocket-peptide and protein target-peptide affinity database. Pocket structure characterization and peptide sequence input into a specific monitor; Attention in a specificity monitor is used to capture the dependence and interaction patterns between peptides and protein pockets, and multilayer perceptron is used to predict the affinity score of peptide-target pockets. Using the root mean square error as the loss function, the difference between the predicted score and the HDOCK score is minimized to train the specificity monitor until convergence.
[0013] In this invention, the sequence decoder and specificity monitor are frozen during the training process of the denoising model, and the mean square error function of the denoising process and the affinity loss function of the specificity monitor are used as the total loss function of the diffusion model.
[0014] The technical solution provided by this invention may include the following beneficial effects: This invention presents a targeted peptide design method based on a conditional latent diffusion model and specific feedback. It constructs a conditional diffusion process within the latent representation space obtained from protein language model learning, using pocket target structural information as a conditional constraint to guide peptide sequence generation. Simultaneously, a specific feedback mechanism is introduced during the diffusion denoising training phase to dynamically evaluate and regulate the generated results, thereby improving the peptide's structural matching and affinity to the pocket target, achieving more precise and controllable targeted peptide design. Attached Figure Description
[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.
[0016] Figure 1 A flowchart illustrating a targeted peptide design method based on a conditional latent diffusion model and specificity feedback in an exemplary embodiment of this disclosure is shown. Figure 2 This diagram illustrates the overall architecture of the targeted peptide design method based on the conditional latent diffusion model and specificity feedback in an exemplary embodiment of this disclosure. Figure 3 (a) is a flowchart illustrating the forward diffusion and denoising processes for peptide characterization using the method of this application; (b) is a schematic diagram illustrating the core framework of the conditional diffusion model of this application, namely, the multi-level attention network structure. Figure 4 The following comparative experimental results are shown in the exemplary embodiments of this disclosure: (a) For 158 target proteins, 10 high-affinity peptides were designed for each target using this method and three benchmark models, namely RFdiffusion, ESM3, and PepPrCLIP, and the affinity scores of different methods were compared based on HDOCK software; the lower the HDOCK score, the higher the affinity. The sample below the diagonal line represents the affinity of this method being higher than that of the benchmark model. (b) The affinity score distribution of the peptides generated by this method and the benchmark model is shown. Detailed Implementation
[0017] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.
[0018] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.
[0019] In this example implementation, a target S1 based on a conditional latent diffusion model and specific feedback is first provided. The target pocket structure characterization extraction is performed by encoding the three-dimensional structure of the protein pocket using the ProteinMPNN model to obtain the pocket structure characterization vector. S2, Natural polypeptide sequence characterization and extraction: Natural polypeptide sequences are encoded using the ESMC protein language model to obtain the latent space characterization of the polypeptide sequence. S3, Construction of Conditional Latent Diffusion Network Model: Based on the diffusion transformer framework, a conditional latent diffusion network model is constructed using the representation of the three-dimensional structure of the protein pocket as conditional information. This conditional latent diffusion network model includes multiple denoising modules. Each denoising module includes a self-attention network, a residual-level cross-attention network, and a pointwise feedforward network. The residual-level cross-attention network aims to map the representation of the protein pocket as a condition to the diffusion network. S4, Forward Diffusion and Reverse Denoising: Based on the conditional latent diffusion model framework, Gaussian noise is gradually added to the latent space representation of the peptide sequence; then, the conditional latent diffusion network is used to perform reverse iterative denoising along the time step, thereby gradually estimating and eliminating the Gaussian noise in the latent space representation of the peptide sequence to generate the latent space representation of the target peptide. S5, Sequence Decoding: Construct a sequence decoder, which includes a one-dimensional convolutional layer and a multilayer perceptron, with each layer using the ReLU function as the activation function; then, use the pre-trained sequence decoder to convert the latent space representation of the target peptide output in S4 into the corresponding target peptide sequence. S6, Specificity Assessment and Optimization: Based on the Transformer framework, a specificity monitor is constructed, which includes a self-attention mechanism, a cross-attention mechanism, and a multilayer perceptron. Then, the specificity of the peptide space generated by the model is evaluated using the pre-trained specificity monitor to obtain the affinity loss value. The affinity loss and diffusion loss are fused together for multi-task joint training to achieve iterative optimization of the model. Finally, the trained conditional latent diffusion network model and sequence decoder are used to obtain a target peptide sequence with high affinity.
[0020] In this embodiment, a conditional diffusion process is constructed in the latent representation space obtained by the protein language model, using target structural information as a conditional constraint to guide the generation of peptide sequences. Simultaneously, a specific feedback mechanism is introduced during the diffusion denoising training phase to dynamically evaluate and regulate the generation results, thereby improving the structural matching degree and affinity of the peptide to the pocket target, achieving more precise and controllable targeted peptide design.
[0021] The specific process of each step in the above embodiments will be described below.
[0022] S1 and S2 employ ProteinMPNN to extract structural representations of protein pockets, while the ESM C protein language model is used for deep embedding modeling of natural peptide sequences. ESMC uses masked language modeling as a pre-training target to enhance its semantic representation ability of protein sequences. In this method, for each natural peptide sample, the continuous vector representation output by ESMC (i.e., the continuous embedding representation of the peptide in the latent space) is used as the initial state input for the diffusion process, ensuring that the diffusion evolution process always exists within a biologically meaningful representation space, thereby guaranteeing the biological rationality and feasibility of the generated results at the structural and functional levels.
[0023] S3, Construct a conditional latent diffusion network model.
[0024] This application designs a multi-layered attention network as the core framework for conditional diffusion. For example... Figure 3 As shown in (b), a conditional latent diffusion network model is constructed based on the diffusion transformer framework, using the representation of the three-dimensional structure of the protein pocket as conditional information. This model includes multiple denoising modules, each comprising a self-attention network, a residual-level cross-attention network, and a pointwise feedforward network. Specifically, the self-attention network focuses on capturing long-range interactions between residues under noisy conditions, thus providing biological justification for the generated peptides. Importantly, the residue-level cross-attention network allows peptide characterization to directly focus on the specific geometric constraints of each amino acid within the pocket, thereby generating complementary characterizations based on a given target. Similar to the Transformer, this application incorporates residual connections and layer normalization in each sub-network and employs a multi-head attention mechanism to improve the model's stability and performance.
[0025] In S4, forward diffusion and reverse denoising are used.
[0026] Please refer to Figure 3 In (a), the latent space representation of the peptide output by S2 is perturbed with Gaussian noise and then fed into the conditional latent diffusion network model. The protein pocket structure representation is used as a condition to guide the denoising diffusion process. Figure 3In the first block of (b), the latent space representation of the peptide after Gaussian noise perturbation is used as the input to the self-attention layer, and is denoted as... Where N represents the polypeptide sequence length and d is the hidden layer dimension. Similarly, the pocket structure characterization is represented as... , and serve as a guiding condition for the denoising process, in which This indicates the number of amino acids in the pocket.
[0027] The representation of the 3D pocket is mapped to a diffusion network model through a residue-level cross-attention network, as follows:
[0028] in, This indicates the strength of the association between the peptide and the binding pocket. , and Let represent the learnable projection matrices for the query, key, and value, respectively, and the attention score. Used to characterize the conditional dependence between peptides and pockets. This ensures that the peptide characterization space being generated can dynamically extract key structural and chemical environment information from the pocket information, thereby generating peptide sequences that meet the structural constraints of the pocket space.
[0029] During the diffusion process, the mean square error (MSE) between the predicted noise and the actual noise is used as the loss function, and its formula is as follows:
[0030] in, This represents the real Gaussian noise added through the forward diffusion process, while This indicates that the denoising network operates at a specific time step. Predicted noise.
[0031] S5, build the sequence decoder module. Please refer to... Figure 2 This application designs a sequence decoder that maps the continuous representation (peptide latent space representation) output by S2 to discrete amino acids. The sequence decoder consists of a one-dimensional convolutional layer and a multilayer perceptron, with ReLU as the activation function for each layer. Based on the PepMerge dataset, the sequence decoder is trained from scratch before training the diffusion model. The loss is calculated using the cross-entropy function, with the specific formula as follows:
[0032] in, Indicates the length of the polypeptide sequence. This indicates the size of the amino acid vocabulary. Indicates position The real amino acids at the location One-hot encoding. Indicates position amino acids The predicted probability. The sequence decoder can convert the generated peptide representation into an amino acid sequence. During the training of the conditional latent diffusion model, the parameters of the sequence decoder are frozen.
[0033] S6, based on specific feedback for multi-task learning.
[0034] To dynamically optimize the peptide generation process, this application develops a specific feedback mechanism within a multi-task learning framework. Specifically, based on the PepMerge dataset, the native peptide sequence is randomly perturbed, and ESMfold is used to predict the peptide structure. Simultaneously, using the database in ppbench2024, affinity scores are calculated through HDOCK software to construct protein pocket-peptide and protein target-peptide affinity databases for pre-training a specificity monitor.
[0035] like Figure 2 As shown on the right, the monitor also employs a Transformer architecture, which includes self-attention mechanisms, cross-attention mechanisms, and multilayer perceptrons. The attention mechanisms in the monitor are able to capture global dependencies and complex interaction patterns between peptides and protein pockets.
[0036] To minimize the predicted affinity score Affinity score obtained from HDOCK software We use the root mean square error (RMSE) as the loss function to measure the difference between them.
[0037] During the training of the denoising model, the sequence decoder and specificity monitor are frozen. The affinity loss predicted by the specificity monitor is used as an auxiliary constraint for denoising training and is added to the diffusion denoising process through the following multi-task joint strategy:
[0038] in, This is used to balance the impact of the specific feedback module on the overall target. Notably, the monitor parameters are frozen during multi-task training. The joint loss guides the diffusion model to optimize for pocket target specificity and high-affinity regions while maintaining consistency with the native peptide.
[0039] The training process for the specificity monitor includes: Constructing a protein pocket-peptide and protein target-peptide affinity database; Input pocket structure characterization and peptide sequence latent space characterization into a specific monitor; Attention in a specificity monitor is used to capture the dependence and interaction patterns between peptides and protein pockets, and multilayer perceptron is used to predict the affinity score of peptide-target pockets. Using the root mean square error as the loss function, the difference between the predicted score and the HDOCK score is minimized to train the specificity monitor until convergence.
[0040] The following simulation experiments further illustrate the beneficial effects of the targeted peptide design method based on the conditional latent diffusion model and specific feedback proposed in this application.
[0041] 1. Experimental conditions: Hardware platform: NVIDIA RTX 4090 GPU (24GB VRAM) is used for model training and inference, while CPU is used for auxiliary computing; The software platform is: Ubuntu 20.04.3 LTS operating system, PyCharm development platform; Software language: Python; The deep learning framework used is PyTorch.
[0042] 2. Experiment Content: This application focuses on a systematic evaluation of the proposed PocketPep method (a targeted peptide design method based on conditional latent diffusion model and specificity feedback) from the perspective of binding affinity, and compares it with existing mainstream methods such as RFdiffusion, ESM3, and PepPrClip. Experimental results are as follows: Figure 4 As shown, the binding energy distribution of peptides generated by each method is illustrated under different Top-k (k=1,3,5,10) test paradigms.
[0043] from Figure 4 (a) The results show that, under the HDOCK evaluation paradigm, most data points are located below the diagonal, indicating that the peptides generated by PocketPep generally have lower binding energies than the baseline method. Since lower binding energy indicates higher affinity, this phenomenon demonstrates that the method of this application has a significant advantage in generating high-affinity candidate peptides.
[0044] from Figure 4 (b) The results show that most of the samples generated by PocketPep are distributed in the lower binding energy region, indicating that the generated peptides have high binding affinity overall, and this advantage is not due to individual samples, but rather has a certain degree of universality. This demonstrates that the method proposed in this application can stably generate high-quality candidate peptides.
[0045] The above analysis shows that PocketPep demonstrates a continuous and stable advantage in binding affinity optimization, generating peptide candidate sequences with stronger binding ability, providing a more reliable foundation for subsequent structural design and functional verification.
[0046] In summary, the beneficial effects of this application include: (1) Using pocket structure features as a condition, the complex relationship and interaction between peptides and target pockets are adaptively captured by combining attention mechanism, thereby generating targeted peptides with higher affinity and stronger specificity. The latent diffusion model transfers the denoising process to a latent space that is semantically highly compressed and approximately follows a Gaussian distribution. This significantly reduces the computational cost while ensuring the mathematical stability of the generation process and achieving high-quality semantically controllable generation.
[0047] Specifically, this application, based on a diffusion framework, progressively injects Gaussian noise into the characterization of natural peptides through a forward diffusion process, achieving a gradual mapping of their latent characterization to a random distribution. Subsequently, in the reverse denoising process, the protein pocket structure characterization is used as conditional information input to guide a multi-level attention network to progressively denoise the noise characterization, thereby generating a latent space characterization of the peptide with target awareness capabilities. Based on the DiT architecture, this application constructs a multi-level attention network as the core framework of the conditional diffusion model. This model consists of multiple stacked denoising modules with identical structures, each containing a self-attention network, a residue-level cross-attention network, and a feedforward network. Among them, the self-attention network aims to capture long-range dependencies between residues under noise perturbation conditions, thereby enhancing the biological rationality of the generated peptide sequence. The residue-level cross-attention network can focus on the geometric constraint information of each amino acid in the protein pocket during the generation process, thereby constructing a complementary characterization space under given target conditions. Similar to the Transformer structure, residual connections and layer normalization mechanisms are introduced in each sub-network, and a multi-head attention mechanism is adopted to improve the training stability and overall performance of the model.
[0048] (2) This application designs a sequence decoder to decode continuous representations onto discrete amino acids. The sequence decoder consists of a one-dimensional convolutional layer and a multilayer perceptron, with ReLU activation function for each layer. Based on the PepMerge dataset, the sequence decoder is trained from scratch before training the diffusion model. During the diffusion denoising training process, the parameters of the sequence decoder are frozen.
[0049] (3) By using a specific feedback mechanism based on multi-task learning, the current state of the diffusion model is monitored in real time, and the generation process is dynamically evaluated and optimized, thereby improving the ability of the denoising model to generate high-affinity peptide sequences under conditional constraints.
[0050] Specifically, to dynamically optimize the peptide generation process, this application develops a specific feedback mechanism based on a multi-task learning framework to assist the diffusion model in generating a peptide representation space with high affinity during the denoising model training process. First, a protein pocket-peptide affinity database is constructed for pre-training the specificity monitor. This specificity monitor employs a Transformer architecture, which includes a self-attention mechanism, a cross-attention mechanism, and a multilayer perceptron. The cross-attention mechanism can capture the dependencies and complex interaction patterns between peptides and protein pockets. During diffusion denoising training, the parameters of the sequence decoder and the specificity monitor are frozen.
[0051] It should be noted that although several modules of the system for executing actions are mentioned in the detailed description above, this division is not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules described above can be embodied in one module. Conversely, the features and functions of one module described above can be further divided into multiple modules for embodiment. Components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any inventive effort.
[0052] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A targeted peptide design method based on a conditional latent diffusion model and specificity feedback, characterized in that, include: S1, Target pocket structure characterization extraction: The three-dimensional structure of the protein pocket is encoded by the ProteinMPNN model to obtain the characterization vector of the pocket structure. S2, Natural polypeptide sequence characterization and extraction: Natural polypeptide sequences are encoded using the ESMC protein language model to obtain the latent space characterization of the polypeptide sequence. S3, Construction of Conditional Latent Diffusion Network Model: Based on the diffusion transformer framework, a conditional latent diffusion network model is constructed using the representation of the three-dimensional structure of the protein pocket as conditional information. This conditional latent diffusion network model includes multiple denoising modules. Each denoising module includes a self-attention network, a residual-level cross-attention network, and a pointwise feedforward network. The residual-level cross-attention network aims to map the representation of the protein pocket as a condition to the diffusion network. S4, Forward Diffusion and Reverse Denoising: Based on the conditional latent diffusion model framework, Gaussian noise is gradually added to the latent space characterization of peptide sequences. Then, the conditional latent diffusion network is used to perform reverse iterative denoising along the time step, thereby gradually estimating and eliminating Gaussian noise in the latent space representation of the peptide sequence to generate the latent space representation of the target peptide. S5, Sequence Decoding: Construct a sequence decoder, which includes a one-dimensional convolutional layer and a multilayer perceptron, with each layer using the ReLU function as the activation function; then, use the pre-trained sequence decoder to convert the latent space representation of the target peptide output in S4 into the corresponding target peptide sequence. S6, Specificity Assessment and Optimization: Based on the Transformer framework, a specificity monitor is constructed, which includes a self-attention mechanism, a cross-attention mechanism, and a multilayer perceptron. Then, the specificity of the peptide space generated by the model is evaluated using the pre-trained specificity monitor to obtain the affinity loss value. The affinity loss and diffusion loss are fused together for multi-task joint training to achieve iterative optimization of the model. Finally, the trained conditional latent diffusion network model and sequence decoder are used to obtain a target peptide sequence with high affinity.
2. The targeted peptide design method based on conditional latent diffusion model and specificity feedback according to claim 1, characterized in that, In S4, the latent space representation of the peptide is progressively denoised at the diffusion time step, and a reverse denoising process is performed through a conditional latent diffusion network model. Specifically, the latent space representation of the peptide is modeled by a self-attention network and input into the conditional latent diffusion network model, and the pocket structure representation is introduced into the conditional latent diffusion network model through a residue-level cross-attention network to apply conditional constraints during the denoising process.
3. The targeted peptide design method based on conditional latent diffusion model and specificity feedback according to claim 1, characterized in that, In S3, the self-attention network, residual-level cross-attention network, and point-by-point feedforward network all have residual connection layers and normalization layers.
4. The targeted peptide design method based on conditional latent diffusion model and specificity feedback according to claim 1, characterized in that, In S4, during the diffusion denoising process, the mean square error between the predicted noise and the actual noise is used as the loss function.
5. The targeted peptide design method based on conditional latent diffusion model and specificity feedback according to claim 1, characterized in that, In S5, before training the diffusion model, the sequence decoder is trained from scratch using the cross-entropy loss function based on the PepMerge dataset.
6. The targeted peptide design method based on conditional latent diffusion model and specificity feedback according to claim 1, characterized in that, In S6, the training process for the specificity monitor includes: Based on the PepMerge dataset, the natural peptide sequence is randomly perturbed, and the peptide structure is predicted using ESMfold. At the same time, combined with the database in ppbench2024, affinity scores are calculated using HDOCK software to construct a protein pocket-peptide and protein target-peptide affinity database. Pocket structure characterization and peptide sequence input into a specific monitor; Attention in a specificity monitor is used to capture the dependence and interaction patterns between peptides and protein pockets, and multilayer perceptron is used to predict the affinity score of peptide-target pockets. Using the root mean square error as the loss function, the difference between the predicted score and the HDOCK score is minimized to train the specificity monitor until convergence.
7. The targeted peptide design method based on conditional latent diffusion model and specificity feedback according to any one of claims 1 to 6, characterized in that, During the training of the denoising model, the sequence decoder and the specificity monitor are frozen, and the mean square error function of the denoising process and the affinity loss function of the specificity monitor are used as the total loss function of the diffusion model.