A method for designing targeted polypeptides with dynamic feedback of protein folding code and three-dimensional structure
By combining protein language models and structural encoders, and utilizing self-attention and REINFORCE algorithms to optimize peptide generation, the problems of unreasonable and unstable structures in peptide design are solved, resulting in highly specific and structurally stable peptide sequences.
Patent Information
- Application Number
- CN202611282055.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-06
AI Technical Summary
Existing peptide design methods fail to effectively incorporate the three-dimensional structural information of target proteins, resulting in peptide sequences with unreasonable or unstable structures after folding. Furthermore, insufficient sampling efficiency makes it difficult to generate peptides with high specificity and structural stability.
The protein language model ESM-C and the ProteinMPNN encoder are used to encode the sequence and structural information of the target protein, respectively. Interactive modeling is performed through a self-attention mechanism, and the peptide sequence decoder is optimized by combining the REINFORCE algorithm to generate peptide sequences that meet the constraints of the target protein.
The generated polypeptide sequences have stronger binding ability, closely resemble natural distribution characteristics, and have improved structural rationality and stability, providing a reliable foundation for subsequent structural design and functional verification.
Smart Images
Figure CN123053221A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of peptide design technology, and more particularly to a targeted peptide design method based on protein folding code and dynamic feedback of three-dimensional structure. Background Technology
[0002] Peptide drugs exert important regulatory functions in various physiological processes by binding to specific target proteins, such as ion channel regulation, antibacterial defense, neurotransmitter transmission, and immune responses. In recent years, peptide drugs have demonstrated significant application value in the treatment of various diseases as a class of promising therapeutic molecules. Compared to traditional small molecule drugs and large molecule antibodies, peptides offer comprehensive advantages in terms of specificity, safety, and syntheticability. Therefore, de novo design of highly specific peptides targeting specific targets is of great significance for advancing drug discovery and disease treatment.
[0003] In recent years, protein language models have further driven a technological revolution in peptide design. Protein language models (such as ESM2 and ProGen) effectively capture the co-evolutionary patterns and statistical characteristics among amino acid residues through large-scale self-supervised pre-training. However, most protein language models do not consider the specificity between peptides and target proteins. Although some studies have attempted to incorporate target information into protein language models to enhance the specificity of peptide design through conditional masking modeling and contrastive learning, these methods still mainly remain at the sequence level of conditional modeling. Because they fail to explicitly introduce the three-dimensional structural information of the target protein, the models struggle to fully characterize the spatial geometric complementarity upon which the interaction between the target and the peptide depends. Furthermore, the lack of explicit modeling of stereochemical constraints (such as bond lengths and bond angles) in protein language models often results in structurally inappropriate peptide sequences after folding, such as atomic conflicts or unstable conformations, further limiting their practical application potential.
[0004] Traditionally, targeted peptides are primarily obtained through experimental or high-throughput screening techniques, supplemented by computational methods such as molecular docking and molecular dynamics simulations for optimization. However, these methods are often limited by insufficient sampling efficiency when dealing with the vast peptide sequences and conformational spaces, making it difficult to comprehensively cover potential high-quality candidate molecules.
[0005] Existing deep learning methods still face several key limitations in targeted peptide design. First, structure-based generative methods heavily rely on large-scale, high-quality protein-peptide complex data, but the limited availability of such data restricts model performance improvement. Second, protein language model-based methods are mainly limited to statistical modeling of one-dimensional amino acid sequence space, simplifying proteins into discrete symbolic sequences and capturing co-evolutionary signals between residues solely through contextual dependencies, failing to fully capture the spatial complementarity upon which protein-peptide interactions depend. Finally, existing generative models often fail to explicitly integrate the three-dimensional structural information of proteins and lack effective geometric and physical constraints on structural rationality during the generation process, easily leading to problems such as atomic conflicts or conformational instability.
[0006] Therefore, there is an urgent need to develop a generative framework that can simultaneously model sequence semantics and three-dimensional structural constraints in order to achieve the design of peptide molecules with high specificity and structural stability for target proteins, thereby improving one or more of the problems existing in the above-mentioned related technical solutions.
[0007] It should be noted that this section is intended to provide background or context for the technical solutions of the invention as set forth in the claims. The description herein does not imply acceptance of prior art simply because it is included in this section. Summary of the Invention
[0008] The purpose of this invention is to provide a targeted peptide design method based on protein folding code and dynamic feedback of three-dimensional structure, thereby solving at least to some extent one or more problems caused by the limitations and defects of related technologies.
[0009] This invention provides a method for designing targeted peptides based on protein folding codes and dynamic feedback of three-dimensional structure, comprising: S1 encodes the sequence and structural information of the target protein, respectively, to obtain the sequence characterization and structural characterization. S2, after aligning the sequence representation and the structural representation, they are spliced together to obtain a unified multimodal sequence representation. The multimodal sequence representation is then input into a multimodal encoder based on the protein folding code to perform interactive modeling of sequence information and structural information, thereby generating a multimodal fusion representation of the target. S3, construct a peptide sequence decoder. The peptide sequence decoder is based on a self-attention mechanism and includes a ProGen2 block, a multimodal target sensing adapter, and a normalization layer. The current peptide sequence and the multimodal fusion representation of the target are input into the sequence decoder. The multimodal target sensing adapter is used to perform interactive modeling of the peptide sequence representation and the multimodal fusion representation of the target. The ProGen2 block is used to generate a peptide sequence that meets the constraints of the target protein. S4, The peptide sequence decoder is trained and optimized using an autoregressive loss function; the peptide sequence decoder is then optimized using the REINFORCE algorithm; S5, the optimized polypeptide sequence decoder is used to predict the target protein to be processed, and a polypeptide sequence that meets the target protein constraints is obtained.
[0010] In this invention, in S1, the protein language model ESM-C is used as a polypeptide sequence encoder to encode the sequence information of the target protein; the ProteinMPNN encoder is used to encode the structural information.
[0011] In this invention, in S2, the multimodal encoder adopts a protein folding coding-guided attention network and includes a normalization and feedforward neural network module. It realizes the interaction modeling between sequence information and structural information through self-attention calculation, thereby capturing the correlation between the two and generating a multimodal fusion representation.
[0012] In this invention, in S3, the peptide sequence decoder uses a cross-modal attention mechanism to interactively model the characterization of the current peptide sequence and the multimodal fusion characterization of the target, thereby enhancing the conditional dependency between the peptide sequence and the target and guiding the generation of a peptide sequence that satisfies the constraints of the target protein.
[0013] In this invention, in S4, during the sequence decoder training and optimization process, the network parameters of the ProGen2 block are frozen, and the sequence decoder and the multimodal encoder based on the protein folding code are randomly initialized and participate in the optimization. The training objective is to minimize the autoregressive loss function.
[0014] In this invention, step S4, the process of optimizing the peptide sequence decoder using the REINFORCE algorithm, includes: A peptide sequence decoder is used as a policy network to generate sequence trajectories from the initial state to the final state. For each state in the sequence trajectory, the likelihood of the generated polypeptide sequence is decoded into a polypeptide sequence. ESMFold is used to predict its three-dimensional structure and pLDDT score. The pLDDT score is used as the reward signal for the current state, and the pLDDT scores are accumulated and summed as the reward. The optimization objective is constructed based on the REINFORCE algorithm, and the polypeptide sequence decoder is optimized. The optimization objective is the cumulative sum of the logarithmic probability and the reward.
[0015] In this invention, S4 further includes: The peptide sequence decoder is optimized using a weighted combination of sequence reconstruction loss and reinforcement learning loss as the training objective.
[0016] This invention also provides a targeted peptide design system based on protein folding code and dynamic feedback of three-dimensional structure, comprising: The basic coding module is used to encode the sequence information and structural information of the target protein, respectively, to obtain the sequence characterization and structural characterization. The multimodal fusion module is used to align and splice sequence representations and structural representations to obtain a unified multimodal sequence representation. The multimodal sequence representation is then input into a multimodal encoder based on the protein folding code to perform interactive modeling of sequence information and structural information, thereby generating a multimodal fusion representation. The peptide sequence decoding module constructs a peptide sequence decoder based on a self-attention mechanism and includes a ProGen2 block, a multimodal target sensing adapter, and a normalization layer. The current peptide sequence and the multimodal fusion representation of the target are input into the sequence decoder. The multimodal target sensing adapter is used to perform interactive modeling of the peptide sequence representation and the multimodal fusion representation of the target. The ProGen2 block is used to generate a peptide sequence that meets the constraints of the target protein. The optimization module uses an autoregressive loss function to train and optimize the peptide sequence decoder; and uses the REINFORCE algorithm to optimize the peptide sequence decoder.
[0017] The technical solution provided by this invention may include the following beneficial effects: This invention constructs a multimodal encoder based on the protein folding code, models cross-modal dependencies between sequence representation and structural representation through a self-attention mechanism, and achieves deep interaction between target protein sequence and structural information, thereby obtaining a unified and semantically rich multimodal fusion representation. Target protein-related constraints are injected into the generation process to guide the fine-tuning of peptide sequence decoder parameters, thereby enabling the design of targeted peptide sequences. The REINFORCE algorithm is used to optimize the peptide sequence decoder, making the generated peptide sequences more reasonable. It can generate peptide candidate sequences with stronger binding ability and closer to the natural distribution characteristics, providing a more reliable foundation for subsequent structural design and functional verification. Attached Figure Description
[0018] 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.
[0019] Figure 1 This diagram illustrates a flowchart of a targeted peptide design method based on dynamic feedback of protein folding codes and three-dimensional structures in an exemplary embodiment of the present invention. Figure 2 The diagram illustrates the architecture of a targeted peptide design system based on protein folding code and dynamic feedback of three-dimensional structure in an exemplary embodiment of the present invention. The meanings of each part in the diagram are as follows: A is a schematic diagram of the overall architecture; B is a schematic diagram of the core process of peptide sequence generation and structure prediction; C is a schematic diagram of the sequence decoder framework; and D is a schematic diagram of the construction of the reinforcement learning loss function. Figure 3 This illustrates the binding energy distribution of peptides generated by different methods under different Top-k (k=1,3,5,10) test paradigms in an exemplary embodiment of the present invention. Detailed Implementation
[0020] 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 the invention 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.
[0021] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of the present invention 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.
[0022] This example implementation first provides a targeted peptide design method based on protein folding code and dynamic feedback of three-dimensional structure, referencing... Figure 1 As shown, the method includes the following steps: S1 encodes the sequence and structural information of the target protein, respectively, to obtain the sequence characterization and structural characterization. S2, after aligning the sequence representation and the structural representation, they are spliced together to obtain a unified multimodal sequence representation. The multimodal sequence representation is then input into a multimodal encoder based on the protein folding code to perform interactive modeling of sequence information and structural information, thereby generating a multimodal fusion representation of the target. S3, construct a peptide sequence decoder. The peptide sequence decoder is based on a self-attention mechanism and includes a ProGen2 block, a multimodal target sensing adapter, and a normalization layer. The current peptide sequence and the multimodal fusion representation of the target are input into the sequence decoder. The multimodal target sensing adapter is used to perform interactive modeling of the peptide sequence representation and the multimodal fusion representation of the target. The ProGen2 block is used to generate a peptide sequence that meets the constraints of the target protein. S4, The peptide sequence decoder is trained and optimized using an autoregressive loss function; the peptide sequence decoder is then optimized using the REINFORCE algorithm; S5, the optimized polypeptide sequence decoder is used to predict the target protein to be processed, and a polypeptide sequence that meets the target protein constraints is obtained.
[0023] This embodiment constructs a multimodal encoder based on the protein folding code, models cross-modal dependencies between sequence representation and structural representation through a self-attention mechanism, and achieves deep interaction between target protein sequence and structural information, thereby obtaining a unified and semantically rich multimodal fusion representation. Target protein-related constraints are injected into the generation process to guide the fine-tuning of sequence decoder parameters, thereby realizing the design of targeted peptide sequences. The REINFORCE algorithm is used to optimize the peptide sequence decoder, making the generated peptide sequences more reasonable. It can generate peptide candidate sequences with stronger binding ability and closer to the natural distribution characteristics, providing a more reliable foundation for subsequent structural design and functional verification.
[0024] The specific steps of the targeted peptide design method based on protein folding code and three-dimensional structure dynamic feedback in this application are described below.
[0025] In S1, such as Figure 2 As shown in Figure A, the target information is processed. Specifically, for the sequence information of the target protein, this application uses the Protein Language Model (ESM-C) as a sequence encoder to learn the potential biological representation of the protein sequence. The sequence encoder is based on the Transformer architecture and captures long-range dependencies between any two amino acids through a self-attention mechanism, thereby learning a high-quality representation that reflects protein evolutionary information. Mask language modeling is used as a pre-training target to improve its ability to represent protein sequences. The encoded sequence representation serves as the sequence semantic feature input for the downstream multimodal encoder, providing target sequence-level semantic information for feature interaction and peptide sequence generation.
[0026] To extract structural representation information of the target protein, this application employs the ProteinMPNN encoder. Based on a message-passing neural network, the ProteinMPNN encoder performs representation learning on the graph structure, aggregating and updating node and edge information through a multi-layer message-passing mechanism to obtain the structural representation information of the protein backbone. This structural representation serves as the three-dimensional structural feature input for the downstream multimodal encoder, providing target spatial structural information for feature interaction and peptide sequence generation.
[0027] In S2, a multimodal encoder based on the protein folding code is used to jointly model protein sequence information (one-dimensional lexicon) and three-dimensional structural information (three-dimensional lexicon) to obtain the fused protein characterization.
[0028] The sequence representation obtained in S1 is used as a one-dimensional sequence term, and the structural representation obtained in S1 is a residue-level structural representation, which is used as a three-dimensional structural term. Both the one-dimensional sequence term and the three-dimensional structural term are used as initial representation vectors, maintaining alignment in sequence position, but the representation vectors of the sequence term and the structural term encode different biological information respectively. Based on this, the structural term is concatenated with the sequence term according to predetermined rules to form a unified multimodal sequence. Then, the representation of the multimodal sequence is input into a protein folding coding-guided attention network (i.e., a multimodal encoder based on the protein folding code), and the interaction model between sequence information and structural information is realized through self-attention computation, thereby capturing the correlation between the two and generating a multimodal fusion representation of the target.
[0029] In S3, the multimodal fusion characterization of the target output from S2 is used to constrain the peptide sequence generation process, thereby generating a peptide sequence that meets the requirements of the target protein. The specific process is as follows: Please refer to Figure 2 In Part B of this embodiment, the constructed sequence decoder is based on a pre-trained protein language model. This sequence decoder is built based on a Transformer decoding structure and includes a ProGen2 block, a multimodal target sensing adapter, and a normalization layer. This sequence decoder generates sequences by modeling the generated sequences and predicting the amino acid at the next position. The multimodal target sensing adapter in this application employs a cross-attention network structure, which has the function of generating polypeptide sequences through multimodal conditional sensing. The multimodal target sensing adapter structure is used to receive the target multimodal fusion representation, interactively model the current polypeptide sequence representation with the target multimodal fusion representation, thereby establishing a conditional dependency between the polypeptide sequence and the target, and guiding the generation of polypeptide sequences that satisfy the target protein constraints.
[0030] The parameters of the peptide sequence decoder are frozen during training and used only as the base generative model.
[0031] The data processing procedure of the multimodal target sensing adapter is represented by the following formula:
[0032]
[0033] in, This represents the characterization of a polypeptide sequence, used to describe the high-dimensional representation of the polypeptide sequence in terms of amino acid composition, contextual semantics, and potential structural information, and is denoted as . ,in Expressed as the length of the polypeptide sequence. is the dimension of the hidden layer. The multimodal fusion representation of the target is used to characterize the joint semantic representation of the target across multiple modal information such as sequence and structure, and is denoted as . ,in Represented as target length, For the hidden layer dimension. It represents the correlation strength between peptide and target multimodal information, and is used to characterize the conditional dependence between target multimodal fusion characterization and peptide sequence characterization. , and These represent the learnable projection matrices for the query, key, and value, respectively. This ensures that the generated peptide characterization space can dynamically extract key structural and chemical environment information from the target multimodal fusion characterization, thereby generating peptide sequences that meet the constraints of the target protein.
[0034] S4, during the training process of target-sensing peptide generation, such as Figure 2 As shown in Figures B and C, all parameters in the ProGen2 network are frozen, and all multimodal target sensing adapters and protein folding code-based multimodal encoders are randomly initialized and used in the optimization. The training objective is to minimize the autoregressive loss function, which is defined as follows:
[0035] 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 is calculated. This loss function effectively constrains the distribution of peptide sequence generation by maximizing the log-likelihood of the true amino acid sequence, thereby improving the accuracy and stability of sequence generation under target conditions.
[0036] S5, such as Figure 2As shown in Figure D, the multimodal target sensing adapter can generate peptide sequences for a given target. However, the generated peptide sequences may have structural instability or unreasonable conformation in three-dimensional space. This embodiment proposes a reinforcement learning optimization method based on 3D dynamic feedback to further optimize the peptide generation process of the peptide sequence decoder.
[0037] Specifically, for the likelihood of the peptide sequence generated in each state of the sequence trajectory, it is decoded into a peptide sequence. ESMFold is used to predict its three-dimensional structure and pLDDT score. The pLDDT score is used as the reward signal for the current state, and the sum of the pLDDT scores is used as the reward. The REINFORCE algorithm is then used to further optimize the peptide sequence decoder. This method mainly includes three key components: a policy network, a reward mechanism, and an optimization strategy. Typically, the peptide sequence decoder network acts as the policy network, generating sequence trajectories from the initial state to the final state. Subsequently, for the peptide sequence generated in each state of the trajectory, the reward function is used to evaluate the structural rationality and conformational stability of the generated peptide. The optimization strategy updates the policy network parameters by maximizing the reward.
[0038] Peptide sequence decoder network as a policy network ,exist The sequence trajectory generated during the step is as follows:
[0039] in, Indicates the first The state-action pairs are generated in each step. For the likelihood of the peptide sequence generated in each state, the sequence is decoded into a peptide sequence. This application uses ESMFold to predict its three-dimensional structure and pLDDT score. The pLDDT score reflects the structural quality of the peptide sequence and serves as a reward signal for the current state. From the first step... The sum of pLDDT scores reaching the final state is considered the reward:
[0040] Based on the REINFORCE algorithm, the following optimization objective is constructed:
[0041] The goal is to accumulate the log probability at each step. With returns The product of these factors maximizes long-term returns, allowing the peptide sequence decoder to gradually learn the optimal strategy.
[0042] Furthermore, this application aims to alleviate the inconsistency between local rewards and global structural quality. This application utilizes subsequent pLDDT signals to correct the rewards allocated for early amino acid selection. In this way, later structural quality feedback can guide early generation decisions and improve confidence at the overall sequence level.
[0043] Considering that the peptide sequence decoder may generate a large number of repetitive sequences during the free generation process, a reconstruction loss is introduced to constrain the reinforcement learning process. The specific loss function is defined as follows:
[0044] The overall training objective is constructed as a weighted combination of sequence reconstruction loss and reinforcement learning loss:
[0045] in, and This is a balancing coefficient used to control the relative contribution of sequence reconstruction loss to the reinforcement learning objective. For example, and The values are 0.4 and 0.6 respectively, but they are not limited to these.
[0046] This embodiment describes a multimodal encoder based on protein folding codes, multimodal condition-aware peptide sequence generation, and 3D dynamic feedback reinforcement learning structure optimization. The method in this application constructs a closed-loop framework integrating data preprocessing, model generation, and reinforcement learning. By constructing a dynamic evaluation mechanism, the generated peptide sequences are structurally predicted and evaluated, achieving synergistic optimization of sequence evolution and structural constraints in conformational space, thereby improving the rationality and controllability of targeted peptide design. This method uses sequence and structure as target protein constraints to generate targeted peptide sequences and dynamically controls the peptide generation process inversely based on the quality of the structure.
[0047] The implementation effect of this application will be illustrated below through specific experiments.
[0048] I. Experimental conditions: Hardware platform: NVIDIA A40 GPU (48GB 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. II. Experiment Content: In this experiment, the proposed method was systematically evaluated from the perspectives of affinity and properties, and compared with existing mainstream methods such as RFdiffusion, PPFlow, and PepPrClip. Experimental results are as follows: Figure 3 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.
[0049] from Figure 3 Based on the results in Part A, under the HDOCK evaluation paradigm, the peptides generated by the method in this application generally exhibit lower binding energies than the baseline method for targets of different lengths. Since lower binding energy indicates higher binding affinity, this phenomenon demonstrates that the method in this application has a significant advantage in generating high-affinity candidate peptides.
[0050] from Figure 3 Looking at the results in Part B, for different affinity thresholds, most of the samples generated by the method in this application are distributed in the lower binding energy region, indicating that the generated results have high binding affinity overall, and this advantage is not due to individual samples, but has a certain degree of universality. This shows that the proposed method can stably generate high-quality candidate peptides, rather than relying on a few optimal results.
[0051] from Figure 3 The results in section C show that the distribution of key physicochemical properties such as hydrophobic moments in this application is highly consistent with the distribution of natural peptides. This indicates that the generated peptides are closer to the natural sequences in terms of overall physicochemical characteristics, and can better maintain biological rationality and structural feasibility, thus verifying the effectiveness of the method in improving the naturalness of generated peptides.
[0052] In summary, the above analysis demonstrates that the method presented in this application exhibits a consistent and stable advantage in optimizing binding affinity. It not only effectively enhances the potential binding capacity between the peptide and the target but also displays characteristics highly consistent with the distribution of natural peptides in key physicochemical properties such as hydrophobic moments, indicating that the generated sequences possess good biological rationality and structural feasibility at the physicochemical level. Based on this, the method of this application can generate peptide candidate sequences with stronger binding affinity and closer approximation to natural distribution characteristics, providing a more reliable foundation for subsequent structural design and functional verification.
[0053] This embodiment also provides a targeted peptide design system based on protein folding code and dynamic feedback of three-dimensional structure, including: The basic coding module is used to encode the sequence information and structural information of the target protein, respectively, to obtain the sequence characterization and structural characterization. The multimodal fusion module is used to align and splice sequence representations and structural representations to obtain a unified multimodal sequence representation. The multimodal sequence representation is then input into a multimodal encoder based on the protein folding code to perform interactive modeling of sequence information and structural information, thereby generating a multimodal fusion representation. The peptide sequence decoding module constructs a peptide sequence decoder based on a self-attention mechanism and includes a ProGen2 block, a multimodal target sensing adapter, and a normalization layer. The current peptide sequence and the multimodal fusion representation of the target are input into the sequence decoder. The multimodal target sensing adapter is used to perform interactive modeling of the peptide sequence representation and the multimodal fusion representation of the target. The ProGen2 block is used to generate a peptide sequence that meets the constraints of the target protein. The optimization module uses an autoregressive loss function to train and optimize the peptide sequence decoder; and uses the REINFORCE algorithm to optimize the peptide sequence decoder.
[0054] Regarding the system in the above embodiments, the specific manner in which each module performs its operations has been described in detail in the embodiments related to the method, and will not be elaborated upon here.
[0055] 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.
[0056] 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 method for designing targeted peptides based on protein folding codes and dynamic feedback of three-dimensional structure, characterized in that, include: S1 encodes the sequence and structural information of the target protein, respectively, to obtain the sequence characterization and structural characterization. S2, after aligning the sequence representation and the structural representation, they are spliced together to obtain a unified multimodal sequence representation. The multimodal sequence representation is then input into a multimodal encoder based on the protein folding code to perform interactive modeling of sequence information and structural information, thereby generating a multimodal fusion representation of the target. S3, construct a peptide sequence decoder. The peptide sequence decoder is based on a self-attention mechanism and includes a ProGen2 block, a multimodal target sensing adapter, and a normalization layer. The current peptide sequence and the multimodal fusion representation of the target are input into the sequence decoder. The multimodal target sensing adapter is used to perform interactive modeling of the peptide sequence representation and the target multimodal fusion representation. The ProGen2 block is used to generate a peptide sequence that meets the target protein constraints. S4, The peptide sequence decoder is trained and optimized using an autoregressive loss function; The peptide sequence decoder was optimized using the REINFORCE algorithm; S5, the optimized polypeptide sequence decoder is used to predict the target protein to be processed, and a polypeptide sequence that meets the target protein constraints is obtained.
2. The method according to claim 1, characterized in that, In S1, the protein language model ESM-C is used as a peptide sequence encoder to encode the sequence information of the target protein; the ProteinMPNN encoder is used to encode the structural information.
3. The method according to claim 1, characterized in that, In S2, the multimodal encoder uses a protein folding coding-guided attention network and includes normalization and feedforward neural network modules. It achieves interactive modeling between sequence information and structural information through self-attention calculation, thereby capturing the correlation between the two and generating a multimodal fusion representation.
4. The method according to claim 1, characterized in that, In S3, the peptide sequence decoder uses a cross-modal attention mechanism to interactively model the characterization of the current peptide sequence and the multimodal fusion characterization of the target, thereby enhancing the conditional dependency between the peptide sequence and the target and guiding the generation of a peptide sequence that satisfies the constraints of the target protein.
5. The method according to claim 1, characterized in that, In S4, during the training and optimization of the peptide sequence decoder, the network parameters of the ProGen2 block are frozen, and the peptide sequence decoder and the multimodal encoder based on the protein folding code are randomly initialized and participate in the optimization. The training objective is to minimize the autoregressive loss function.
6. The method according to claim 5, characterized in that, In S4, the process of optimizing the peptide sequence decoder using the REINFORCE algorithm includes: A peptide sequence decoder is used as a policy network to generate sequence trajectories from the initial state to the final state. For each state in the sequence trajectory, the likelihood of the generated polypeptide sequence is decoded into a polypeptide sequence. ESMFold is used to predict its three-dimensional structure and pLDDT score. The pLDDT score is used as the reward signal for the current state, and the pLDDT scores are accumulated and summed as the reward. The optimization objective is constructed based on the REINFORCE algorithm, and the polypeptide sequence decoder is optimized. The optimization objective is the cumulative sum of the logarithmic probability and the reward.
7. The method according to claim 6, characterized in that, S4 also includes: The peptide sequence decoder is optimized using a weighted combination of sequence reconstruction loss and reinforcement learning loss as the training objective.
8. A targeted peptide design system based on protein folding code and dynamic feedback of three-dimensional structure, characterized in that, include: The basic coding module is used to encode the sequence information and structural information of the target protein, respectively, to obtain the sequence characterization and structural characterization. The multimodal fusion module is used to align and splice sequence representations and structural representations to obtain a unified multimodal sequence representation. The multimodal sequence representation is then input into a multimodal encoder based on the protein folding code to perform interactive modeling of sequence information and structural information, thereby generating a multimodal fusion representation. The peptide sequence decoding module constructs a peptide sequence decoder based on a self-attention mechanism and includes a ProGen2 block, a multimodal target sensing adapter, and a normalization layer. The current peptide sequence and the multimodal fusion representation of the target are input into the sequence decoder. The multimodal target sensing adapter is used to perform interactive modeling of the peptide sequence representation and the multimodal fusion representation of the target. The ProGen2 block is used to generate a peptide sequence that meets the constraints of the target protein. The optimization module uses an autoregressive loss function to train and optimize the peptide sequence decoder; and uses the REINFORCE algorithm to optimize the peptide sequence decoder.