A method for generating and screening targeted antigen peptide sequences based on an active backbone
By using a three-layer MPNN model based on the active backbone and an autoregressive decoding strategy, combined with a pHLA I affinity prediction model to screen antigen peptides, the problems of time-consuming and labor-intensive peptide generation and insufficient accuracy were solved, achieving efficient and accurate targeted antigen peptide generation and screening.
Patent Information
- Application Number
- CN202410667529.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-28
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2044-05-28
AI Technical Summary
Existing methods for peptide generation and screening are time-consuming and labor-intensive. Traditional methods rely on physical energy functions and have limited accuracy and efficiency. Deep learning models have limited generalization ability in peptide generation tasks for specific targets, and experimental data is scarce.
A three-layer MPNN model based on the active backbone was adopted, combined with autoregressive masking technology and disorder decoding strategy. Protein structure data were obtained from the RCSB Protein Data Bank database for training and fine-tuning. Antigenic peptides were screened using the pHLA I affinity prediction model and target-receptor interaction interface analysis tool.
It enables rapid and precise generation and screening of high-affinity targeting antigen peptides, improving the accuracy and efficiency of peptide generation, and accurately restoring the structure and efficiently binding to the target.
Smart Images

Figure CN118522342B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of peptide drug development technology, specifically relating to a method for generating and screening targeted antigen peptide sequences based on an active backbone. Background Technology
[0002] In human cells, peptides are short amino acid sequences of less than 50 amino acids, playing crucial roles in signal transduction and antibacterial activity. Beyond these natural functions, peptides are also used to complement antibody and small molecule therapeutic modalities. Similar to antibodies, peptides can bind to protein surfaces with high affinity and selectivity, targeting disease-related targets that are currently inaccessible to antibodies or small molecules—the so-called "drug-untapped space." Compared to traditional antibody vaccines, peptide vaccines exhibit several advantages: they are easier to synthesize and purify, less expensive, and highly effective in inducing CD8 or CD4 T cell responses in the human body. However, experimental design and peptide screening are time-consuming and labor-intensive; therefore, there is an urgent need for tools and screening methods that can rapidly and reliably generate targeted antigenic peptides to accelerate related research.
[0003] Traditional peptide or protein design relies on frequent calculations of physical energy functions. These physics-based methods depend on the accuracy of these energy functions and require guidance from structural biology expertise, thus limiting their accuracy and efficiency. The rapid development of deep learning technology has revolutionized computational protein design. These deep learning-based models can autonomously generate sequences that fold into a given structural backbone, without using energy functions, by conditionedly simulating the distribution of amino acid sequences based on the backbone structure. This enables rapid and accurate design. The rapid emergence of methods such as SPROF, ProteinMPNN, PiFold, and ProDesignLE further demonstrates the potential and effectiveness of deep learning in structural biology. However, applying protein design models to target antigen peptide generation tasks for specific targets has limitations. Protein sequence generation models are typically trained on a broad range of protein types, thus limiting their generalization ability when used for peptide generation tasks targeting specific targets. While this approach is broad, it can also make it difficult for the model to accurately capture the deeper information specific to the task. Furthermore, peptide design for specific targets faces the challenge of scarce available experimental data. Summary of the Invention
[0004] To address the aforementioned problems, the present invention aims to provide a method for generating and screening targeted antigen peptide sequences based on an active backbone, and specifically demonstrates this method using HLA I target as an example.
[0005] The specific technical solution is as follows:
[0006] A method for generating and screening targeted antigen peptide sequences based on an active scaffold includes the following steps:
[0007] Step 1: From the RCSB Protein Data Bank database, at a resolution smaller than [missing information] The protein structure data was obtained with the condition that the number of amino acids was less than 10,000 and the deadline was August 2, 2021. Then, the mmseqs2 clustering tool was used to cluster the obtained structure data with a 30% sequence identity cutoff value. The obtained protein structure source domain dataset was randomly divided into training set, validation set and test set.
[0008] Step 2: Extract the N and C atoms of the protein backbone from the PDB file obtained in Step 1. α C, O and virtual C β Atoms represent structural features. These skeleton atoms are used as node features, and a set of one-hot codes are added as additional inputs to the encoder. At the same time, the distances between the skeleton atoms and the 48 nearest amino acid atoms in Euclidean space are selected as edge features and input into the encoder. These features are continuously updated by the three-layer MPNN during the backpropagation of the model.
[0009] Step 3: The node features and edge features processed by the three-layer encoder, as well as the encoded protein sequence processed by the autoregressive masking technique, are input into the decoder. The decoder uses a disordered decoding strategy, randomly selecting the next amino acid to be predicted in each decoding step. In any prediction step, the model utilizes all previously predicted amino acid information to ensure that a complete context is provided. For peptide generation tasks with fixed targets, the model can use known partial sequences as background sequences to effectively infer structures with unknown regions. When training the model, the model performance is optimized by minimizing the cross-entropy loss of each residual. In terms of decoding order, the disordered autoregressive decoding method replaces the conventional sequential decoding method from N-end to C-end.
[0010] Step 4: Use the training set obtained in Step 1 to train and optimize the protein sequence generation model;
[0011] Step 5: Taking HLA I as an example, obtain the target domain dataset of the pHLA I complex structure from the RCSB Protein Data Bank database. Further filter and process the obtained PDB files to form a recombinant dataset. Divide each genotype in this dataset into training, validation, and test sets in an 8:1:1 ratio. For genotypes with fewer than 10 data entries, their PDB entries are used only for training and validation and are not included in the test set.
[0012] Step 6: Fine-tune the protein sequence generation model obtained in Step 4 using the training and validation sets obtained in Step 5 to obtain a new model. This model is suitable for generating HLA I antigen peptide sequences. Test the model before and after fine-tuning on the test set.
[0013] Step 7: Use the sequences generated by the pre-fine-tuned model and the fine-tuned model to form a pHLA I complex with the corresponding target sequence. Input the pHLA I complex sequence into the structure prediction model AlphaFold2 for testing. Determine the structure recovery ability of the designed peptide, the structural similarity to the natural peptide, and the binding ability to the target based on the structure confidence score.
[0014] Step 8: Use the pHLA I affinity prediction model MHCfovea and the target-receptor interaction interface analysis tool PDBePISA to create a scoring system for screening antigen peptides;
[0015] Step 9: Taking two HLA I targets as examples, generate 20 antigen peptides for each target antigen peptide using the target antigen peptide sequence generation model obtained in Step 6. Using the scoring system obtained in Step 8, deploy the MHCfovea model on the local computer, input the generated antigen peptide sequences and corresponding target sequences into the affinity prediction model, calculate the predicted binding probability together with the natural peptide sequence, and screen out the design peptides with a binding probability greater than that of the natural peptide.
[0016] Step 10: Use the designed peptides screened in Step 9 to form complexes with the target, input them into the target-receptor interaction interface analysis tool for testing, predict the Gibbs free energy and interaction contact area of the designed peptides and natural peptides, and screen out the designed peptides with Gibbs free energy less than that of the natural peptides.
[0017] As a technical solution of this invention, the protein structure dataset was modified based on the datasets used in relevant literature, resulting in a training set of 23,358 samples, a validation set of 1,464 samples, and a test set of 1,539 samples. The pHLAI complex structure dataset was also adjusted based on the datasets in relevant literature, resulting in a training set of 438 samples, a validation set of 67 samples, and a test set of 51 samples.
[0018] As a technical solution of this invention, the model is first pre-trained on a protein structure dataset, and then fine-tuned on a pHLA I complex structure dataset. Genotype samples with fewer than 10 data points are not included in the training. The model is implemented based on PyTorch, and the optimizer used in the pre-training and fine-tuning processes is Adam, with a learning rate of 5×10⁻⁶. -4The loss function is the cross-entropy loss function. During training, early stopping is performed based on the loss on the validation set. If the loss on the validation set does not improve within 10 epochs, the learning rate is reduced by a factor of 10.
[0019] The beneficial effects of this invention are as follows:
[0020] 1) This invention proposes a method for generating and screening targeted antigen peptide sequences based on an active backbone. Compared with existing methods, this invention has better sequence design capabilities and a more precise screening process, and can generate and screen targeted antigen peptides with higher affinity.
[0021] 2) This invention proposes a scoring system for screening antigen peptides that combines the pHLA I affinity prediction model MHCfovea with the target-receptor interaction interface analysis tool PDBePISA, which can quickly screen design peptides with higher affinity than natural peptides.
[0022] 3) This invention designs a training method for a targeted antigen peptide sequence generation model based on transfer learning. The model is first trained on protein structure data and then fine-tuned using pHLA I complex data, which further improves the generation performance of the model. Attached Figure Description
[0023] Figure 1 Flowchart of the method for generating and screening targeted antigen peptide sequences;
[0024] Figure 2 This is a flowchart of the training process for a transfer learning-based model for generating targeted antigen peptide sequences. Detailed Implementation
[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings and examples. The described embodiments are only a part of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0026] Example
[0027] like Figure 1 and Figure 2 As shown, a method for generating and screening targeted antigen peptide sequences based on an active backbone includes the following steps:
[0028] Source domain datasets of protein structures were obtained from the RCSB Protein Data Bank database, and the resulting PDB files were further processed to obtain a training set of 23,358 samples, a validation set of 1,464 samples, and a test set of 1,539 samples. The N and C atoms of the protein backbone structure were extracted from the obtained PDB files. α C, O and virtual C βAtomic structural features are represented, with these backbone atoms used as node features and a set of one-hot codes added as additional input to the encoder. Simultaneously, the distances between the backbone atoms and their 48 nearest amino acid atoms in Euclidean space are selected as edge features and input to the encoder; these features are continuously updated during model backpropagation. The node and edge features processed by the three-layer encoder, along with the encoded protein sequence processed using autoregressive masking, are input to the decoder. The decoder uses a disordered decoding strategy, randomly selecting the next amino acid to predict at each decoding step. In any prediction step, the model utilizes all previously predicted amino acid information to ensure complete context. The protein sequence generation model ProteinMPNN is trained and optimized using the obtained training set. Taking HLA I as an example, the target domain dataset of the pHLA I complex structure is obtained from the RCSB Protein DataBank database. The obtained PDB files are further filtered and processed to form a recombinant dataset, in which each genotype is divided into training, validation, and test sets in an 8:1:1 ratio. For genotypes with fewer than 10 data entries, their PDB entries were used only for training and validation and were not included in the test set. The final training set consisted of 438 samples, the validation set of 67 samples, and the test set of 51 samples. The obtained protein sequence generation model was fine-tuned using the training and validation sets to obtain a new targeted antigen peptide sequence generation model. The pre-tuned and post-tuned models were then tested on the test set. Sequences generated by the pre-tuned and post-tuned models were combined with corresponding target sequences to form pHLA I complexes. These pHLA I complex sequences were input into the structure prediction model AlphaFold2 for testing. The structure confidence score was used to determine the structural recovery ability of the designed peptide, its structural similarity to the natural peptide, and its binding affinity to the target. A scoring system for screening antigen peptides was created using the pHLA I affinity prediction model MHCfovea and the target-receptor interaction interface analysis tool PDBePISA. Two HLA... Taking target I as an example, 20 antigen peptides are generated using the obtained target antigen peptide sequence generation model. Using the obtained scoring system, the generated antigen peptide sequences and corresponding target sequences are input into the affinity prediction model. Together with the natural peptide sequence, the predicted binding probability is calculated, and design peptides with a binding probability greater than that of the natural peptide are screened out. The selected design peptides and the target form a complex, which is then input into the target-receptor interaction interface analysis tool for testing. The Gibbs free energy and interaction contact area of the design peptide and the natural peptide are predicted, and design peptides with a Gibbs free energy less than that of the natural peptide are screened out.
[0029] The sequence evaluation metrics used in this invention include sequence recovery rate and perplexity. Sequence recovery rate measures the model's ability to recover natural sequences, while perplexity evaluates the certainty of the model's predictions of natural amino acids in a sample. The structural evaluation metrics used in this invention include pLDDT, pTM, and ipTM. pLDDT is a confidence measure based on lDDT, reflecting local confidence in the structure and used to assess confidence within a single structural domain. pTM measures the overall accuracy of the protein complex structure. ipTM reflects the quality of the predicted protein complex interaction interface and the binding potential between the target and receptor.
[0030] In one embodiment of the present invention, taking the HLA I target as an example, the fine-tuned targeted antigen peptide sequence generation model of the present invention is compared with the untuned protein sequence generation model. The results are shown in Tables 1 and 2. Regarding sequence evaluation metrics, the performance of the fine-tuned targeted antigen peptide sequence generation model of the present invention is superior to that of the untuned protein sequence generation model. Specifically, the sequence recovery rate is higher than that of the untuned model, and the perplexity is lower, indicating that the fine-tuned model of the present invention can generate antigen peptides that are more similar to natural peptides. Regarding structural evaluation metrics, the performance of the fine-tuned targeted antigen peptide sequence generation model of the present invention is comparable to that of the protein sequence generation model, indicating that the antigen peptides generated by the pre-trained model of the present invention can accurately recover the required structure.
[0031] Table 1. Performance comparison of the fine-tuned model and the unfine-tuned model on the sequence generation task.
[0032] method Sequence recovery rate Confusion ProteinMPNN 0.254 1.022 This method 0.517 0.486
[0033] Table 2. Performance comparison of the fine-tuned model and the unfine-tuned model on the structure reconstruction task of generated sequences.
[0034] method pLDDT pTM ipTM ProteinMPNN 97.488 0.937 0.904 This method 97.533 0.939 0.902
[0035] In one embodiment of the present invention, taking HLA-A*02:01 and HLA-B*27:05 as examples, the fine-tuned target antigen peptide sequence generation model of the present invention was used for the task of generating HLA I antigen peptides. After generating 20 antigen peptides for each, the binding probability was predicted using the pHLA I affinity prediction model MHCFovea, and the results are shown in Table 3. The present invention has achieved excellent results in the task of generating high-affinity target HLA I antigen peptides.
[0036] Table 3 shows the performance of the HLA I antigen peptides generated by the fine-tuned model on the binding probability prediction task.
[0037]
[0038]
[0039] In one embodiment of the present invention, taking HLA-A*02:01 and HLA-B*27:05 targets as examples, among the 20 antigen peptides generated by the fine-tuned targeted antigen peptide sequence generation model of the present invention, design peptides with a predicted binding probability greater than that of the natural peptide were screened out and input into a target-receptor interaction interface analysis tool for testing. The results are shown in Tables 4 and 5. Finally, design peptides with a Gibbs free energy lower than that of the natural peptide, i.e., design peptides with stronger affinity than the natural peptide, were further screened out. This demonstrates that the present invention can quickly and accurately screen out design peptides with stronger affinity than the natural peptide.
[0040] Table 4. Performance of the HLA-A*02:01 antigenic peptide generated by the fine-tuned model on the interaction prediction task.
[0041]
[0042]
[0043] Table 5. Performance of the HLA-B*27:05 antigenic peptide generated by the fine-tuned model on the interaction prediction task.
[0044]
[0045] The above description is merely a preferred embodiment of the present invention and does not limit the scope of the invention in any way. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention, or any direct or indirect applications to other related technical fields, without departing from the scope of the present invention, shall be deemed to have remained within the scope of protection of the present invention.
Claims
1. A method for generating and screening targeted antigen peptide sequences based on an active backbone, characterized in that, Includes the following steps: Step 1: Obtain the source domain dataset of protein structures from the RCSB Protein Data Bank database and divide it into training, validation, and test sets; Step 2: Extract the backbone structure features of the protein from the PDB file obtained in Step 1. Use these backbone atoms as node features and add a set of one-hot codes as additional input to the encoder. At the same time, select the distance between the backbone atom and the 48 nearest amino acid atoms in Euclidean space as edge features and input them into the encoder. These features are continuously updated during the backpropagation of the model. Step 3: Input the node features and edge features processed by the three-layer encoder, as well as the encoded protein sequence processed by the autoregressive masking technique, into the decoder. The decoder uses a disordered decoding strategy, randomly selecting the next amino acid to be predicted in each decoding step. In any prediction step, the model will utilize all previously predicted amino acid information to ensure that complete context is provided. Step 4: Use the training set obtained in Step 1 to train and optimize the protein sequence generation model; Step 5: Taking the HLAI target as an example, obtain the target domain dataset of the pHLAI complex structure from the RCSB Protein Data Bank database. Further screen and process the obtained PDB files to form a recombinant dataset. Divide each genotype in the dataset into training set, validation set and test set in a ratio of 8:1:
1. For genotypes with less than 10 data entries, their PDB entries are only used for training and validation and are not included in the test set. Step 6: Fine-tune the protein sequence generation model obtained in Step 4 using the training and validation sets obtained in Step 5 to obtain a new target antigen peptide sequence generation model, and test the model before and after fine-tuning on the test set. Step 7: Use the sequences generated by the pre-fine-tuned model and the post-fine-tuned model to form a pHLAI complex with the corresponding target sequence. Input the pHLAI complex sequence into the structure prediction model AlphaFold2 for testing. Evaluate the structure recovery ability of the designed peptide, its structural similarity to the natural peptide, and its binding ability to the target based on the structure confidence score. Step 8: Use the pHLA I affinity prediction model MHCfovea and the target-receptor interaction interface analysis tool PDBePISA to create a scoring system for screening antigen peptides; Step 9: Taking two HLAI targets as examples, use the target antigen peptide sequence generation model obtained in Step 6 to generate 20 antigen peptides for each target. Then, use the scoring system obtained in Step 8 to input the generated antigen peptide sequences and the corresponding target sequences into the affinity prediction model to obtain the predicted binding probability. Select the design peptides with a binding probability greater than that of the natural peptide. Step 10: Use the designed peptides and targets screened in Step 9 to form a complex and input it into AphlaFold2 to predict the structure. Input the complex structure into the target-receptor interaction interface analysis tool for testing and screen out the designed peptides with Gibbs free energy lower than that of the natural peptides.
2. The method for generating and screening targeted antigen peptide sequences based on an active backbone as described in claim 1, characterized in that, The pHLAI complex structure target domain dataset is a small sample dataset, while the protein structure source domain dataset is a large sample dataset for the target domain association task; the source domain dataset and the target domain dataset were obtained from the RCSB Protein Data Bank database and public literature.
3. The method for generating and screening targeted antigen peptide sequences based on an active backbone as described in claim 2, characterized in that, In step three, the unordered autoregressive decoding method is used instead of the conventional sequential decoding method from N to C.
Citation Information
Patent Citations
IgE antigenicity sequence and preparation of branched multiple antigenic peptides
CN101503460A
Polypeptide specifically combined with high expression antigen HLA-E of myeloma cells and application of polypeptide
CN111116716A
