CN118522342A - Active framework-based targeted antigen peptide sequence generation and screening method - Google Patents

Abstract

The invention discloses a method for generating and screening a targeting antigen peptide sequence based on an active framework, which comprises the following specific steps: acquiring a source domain dataset of a protein structure and a target domain dataset of a pHLA I complex structure; generating a model based on the source domain data set pre-training protein sequence to obtain new parameters and the model; obtaining pHLA I composite structure independent test datasets; testing on the obtained target domain data set by using the pre-fine tuning model and the obtained post-fine tuning model; testing the pre-fine tuning model and the obtained post-fine tuning model by utilizing a structure prediction model; creating a scoring system for screening the antigenic peptides based on pHLA I affinity prediction models and interaction interface analysis tools; taking two HLA I targets as an example, generating antigen peptide by using a fine-tuned model; screening was performed on the resulting antigenic peptides using a scoring system. The invention has more excellent generation capacity of the target antigen peptide sequence, and provides a scoring system for screening antigen peptides with higher affinity.

Description

A method for generating and screening targeted antigen peptide sequences based on active skeletons

Technical Field

The present invention belongs to the technical field of polypeptide drug research and development, and specifically relates to a method for generating and screening targeted antigen peptide sequences based on an active skeleton.

Background Art

In human cells, peptides are short-chain amino acid sequences of less than 50 amino acids in length, which play key roles in signal transduction, antibacterial and other fields. In addition to these natural functions, peptides are also used to supplement the therapeutic model of antibodies and small molecules. Similar to antibodies, peptides can bind to the surface of proteins with high affinity and high selectivity, and can target disease-related targets that are currently incapable of antibodies or small molecules, the so-called "undruggable space". Compared with traditional antibody vaccines, peptide vaccines show many advantages. They are easy to synthesize and purify, low cost, and very effective in inducing CD8 or CD4 T cell responses in the human body. However, the process of experimental design and screening of peptides is very time-consuming and labor-intensive. Therefore, there is an urgent need for a tool and screening method that can quickly and reliably generate targeted antigen peptides to accelerate the development of related research.

Traditional peptide or protein design relies on frequent calculations of physical energy functions. These traditional methods based on physical principles rely on the accuracy of physical energy functions and require the guidance of structural biology expertise. Therefore, traditional methods have limitations in terms of design accuracy and efficiency. The rapid development of deep learning technology has brought revolutionary changes to the field of computational protein design. These deep learning-based models can simulate the conditional distribution of amino acid sequences as conditions without using energy functions, and autonomously generate sequences that can fold into a given structural skeleton, thereby achieving fast and accurate design. The rapid emergence of methods such as SPROF, ProteinMPNN, PiFold, and ProDesignLE further marks the application potential and effectiveness of deep learning in the field of structural biology. However, there are certain limitations in applying protein design models to the task of generating targeted antigenic peptides for specific targets. Protein sequence generation models are usually trained based on a wide range of protein types, so when used for peptide generation tasks for specific targets, the generalization ability of the model is limited. Although this method covers a wide range, it also makes it difficult for the model to accurately capture the deep information of a specific task. In addition, peptide design for specific targets also faces the challenge of scarce available experimental data.

Summary of the invention

In view of the above problems, the purpose of the present invention is to provide a method for generating and screening targeted antigen peptide sequences based on an active skeleton, and to specifically demonstrate it using the HLA I target as an example.

The specific technical solutions are as follows:

A method for generating and screening targeted antigen peptide sequences based on an active skeleton comprises the following steps:

Step 1: From the RCSB Protein Data Bank database, select The number of amino acids was less than 10,000, and the protein structure data were obtained with a cutoff date of August 2, 2021. Then, the mmseqs2 clustering tool was used to cluster the obtained structural data with a sequence identity of 30% as the cutoff value, and the obtained protein structure source domain dataset was randomly divided into training set, validation set, and test set.

Step 2: Extract the N, C _α , C, O and virtual C _β atoms of the protein skeleton from the PDB file obtained in step 1 to represent the structural features. Use these skeleton atoms as node features and add a set of one-hot encoding as additional input to the encoder. At the same time, select the distance between the skeleton atoms and the 48 amino acid atoms closest to them in Euclidean space as edge features and input them into the encoder. These features are continuously updated by the three-layer MPNN during the back propagation process of the model.

Step 3: Input the node features and edge features processed by the three-layer encoder and the encoded protein sequence processed by the autoregressive masking technology into the decoder. The decoder uses a disordered decoding strategy to randomly select the next amino acid to be predicted in each decoding step. In any prediction step, the model will use all previously predicted amino acid information to ensure that a complete context is provided. For the peptide generation task of a fixed target, the model can use the known partial sequence as the background sequence to effectively infer the structure 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 conventional sequential decoding method from N-terminus to C-terminus is replaced by the disordered autoregressive decoding method.

Step 4: Use the training set obtained in step 1 to train and optimize the protein sequence generation model;

Step 5: Taking the HLA I target as an example, the target domain dataset of the pHLA I complex structure was obtained from the RCSB Protein Data Bank database, and the obtained PDB files were further screened and processed to form a recombinant dataset. Each genotype in the dataset was divided into a training set, a validation set, and a test set in a ratio of 8:1:1. For genotypes with less than 10 data records, their PDB entries were only used for training and validation and were not included in the test set;

Step 6: Use the training set and validation set obtained in step 5 to fine-tune the protein sequence generation model obtained in step 4 to obtain a new model, which is suitable for the generation of targeted HLA I antigen peptide sequences, and use the pre-fine-tuning model and the fine-tuning model to test on the test set;

Step 7: Use the sequences generated by the pre-fine-tuning model and the fine-tuning 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, and judge the structural recovery ability of the designed peptide, the structural similarity with the natural peptide, and the binding ability with the target according to the structural 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 HLA I targets as an example, 20 antigen peptides were generated using the targeted antigen peptide sequence generation model obtained in step 6, and the scoring system obtained in step 8 was used to deploy the MHCfovea model on a local computer. The generated antigen peptide sequences and the corresponding target sequences were input into the affinity prediction model, and the predicted binding probability was calculated together with the natural peptide sequence to screen out the designed peptides with a binding probability greater than that of the natural peptides.

Step 10: Use the designed peptide screened in step 9 to form a complex with the target, input it into the target-receptor interaction interface analysis tool for testing, predict the Gibbs free energy and interaction contact area of the designed peptide and the natural peptide, and screen out the designed peptide with a Gibbs free energy smaller than that of the natural peptide.

As a technical solution of the present invention, the protein structure dataset was reduced to a certain extent according to the corresponding dataset used in the relevant literature, and finally 23,358 samples were obtained for the training set, 1,464 samples for the validation set, and 1,539 samples for the test set. The pHLAI complex structure dataset was also adjusted to a certain extent according to the dataset of the relevant literature, and finally 438 samples were obtained for the training set, 67 samples for the validation set, and 51 samples for the test set.

As a technical solution of the present 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 less than 10 data are not trained. The model is implemented based on PyTorch. The optimizer used in the pre-training and fine-tuning process is Adam, the learning rate is 5×10 ^-4 , and the loss function is the cross entropy loss function. During the training process, early stopping is performed according to the loss of the validation set. If the loss of the validation set does not improve within 10 epochs, the learning rate is reduced by 10 times.

The beneficial effects of the present invention are:

1) The present invention proposes a method for generating and screening targeted antigen peptide sequences based on an active skeleton. Compared with existing methods, the present invention has better sequence design capabilities and precise screening processes, and can generate and screen targeted antigen peptides with higher affinity.

2) The present invention proposes a scoring system for screening antigen peptides that combines the pHLA I affinity prediction model MHCfovea and the target-receptor interaction interface analysis tool PDBePISA, which can quickly screen designed peptides with higher affinity than natural peptides.

3) The present 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 with pHLA I complex data, further improving the generation performance of the model.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for generating and screening targeting antigen peptide sequences;

Figure 2 is a training flowchart of the targeted antigen peptide sequence generation model based on transfer learning.

DETAILED DESCRIPTION

The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings and embodiments of the specification. The described embodiments are only part of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Example

As shown in FIG1 and FIG2 , a method for generating and screening a targeting antigen peptide sequence based on an active skeleton comprises the following steps:

The source domain dataset of protein structure was obtained from the RCSB Protein Data Bank database and the obtained PDB files were further processed, finally obtaining 23358 samples in the training set, 1464 samples in the validation set, and 1539 samples in the test set; the N, C _α , C, O and virtual C _β atoms of the protein skeleton structure were extracted from the obtained PDB files to represent the structural features, and these skeleton atoms were used as node features, and a set of one-hot encoding was added as an additional input to the encoder. At the same time, the distance between the skeleton atoms and the 48 amino acid atoms closest to them in the Euclidean space was selected as the edge feature input to the encoder, and these features were continuously updated during the back propagation of the model; the node features and edge features processed by the three-layer encoder, as well as the encoded protein sequence processed by the autoregressive masking technology, were input into the decoder. The decoder uses a disordered decoding strategy to randomly select 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 a complete context is provided; the obtained training set is used to train and optimize the protein sequence generation model ProteinMPNN; taking the HLA I target as an example, the target domain data set of the pHLA I complex structure is obtained from the RCSB Protein DataBank database, and the obtained PDB files are further screened and processed to form a recombinant data set, and each genotype in the data set is divided into training set, validation set and test set in a ratio of 8:1:1. For genotypes with less than 10 data records, their PDB entries were only used for training and validation and were not included in the test set. The final training set consisted of 438 samples, the validation set consisted of 67 samples, and the test set consisted of 51 samples. The training set and validation set were used to fine-tune the protein sequence generation model to obtain a new targeted antigen peptide sequence generation model, and the pre-fine-tuning model and the fine-tuned model were used to test the model on the test set. The sequences generated by the pre-fine-tuning model and the fine-tuned model were combined with the corresponding target sequences to form a pHLA I complex, and the pHLA I complex sequence was input into the structure prediction model AlphaFold2 for testing. The structural confidence score was used to determine the structural recovery ability of the designed peptide, the structural similarity with the natural peptide, and the binding ability with the target. The pHLA I affinity prediction model MHCfovea and the target-receptor interaction interface analysis tool PDBePISA were used to create a scoring system for screening antigen peptides. Two HLA Taking target I as an example, 20 antigen peptides were generated using the obtained targeted antigen peptide sequence generation model, and the generated antigen peptide sequences and the corresponding target sequences were input into the affinity prediction model using the obtained scoring system, and the predicted binding probability was calculated together with the natural peptide sequence to screen out the designed peptides with a binding probability greater than that of the natural peptides; the screened designed peptides were used to form a complex with the target, and the complex was input into the target-receptor interaction interface analysis tool for testing, and the Gibbs free energy and interaction contact area of the designed peptides and the natural peptides were predicted, and the designed peptides with a Gibbs free energy less than that of the natural peptides were screened out.

The sequence evaluation indicators used in the present invention include sequence recovery rate and perplexity. The sequence recovery rate is an indicator for measuring the ability of the model to recover the natural sequence, and the perplexity is an indicator for evaluating the certainty of the model in predicting the natural amino acids of the sample. The structural evaluation indicators used in the present invention include pLDDT, pTM and ipTM. pLDDT is a confidence metric based on lDDT, which reflects the local confidence in the structure and is used to evaluate the confidence within a single domain. pTM is an indicator for measuring the accuracy of the overall protein complex structure. ipTM is an indicator reflecting the quality of the predicted protein complex interaction interface and the binding potential between the target and the receptor.

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 protein sequence generation model before fine-tuning, and the results are shown in Tables 1 and 2. In terms of sequence evaluation indicators, the performance of the fine-tuned targeted antigen peptide sequence generation model of the present invention is better than the protein sequence generation model before fine-tuning, wherein the sequence recovery rate is higher than the model before fine-tuning, and the perplexity is lower than the model before fine-tuning, which indicates that the fine-tuned model of the present invention can generate antigen peptides that are more similar to natural peptides. In terms of structural evaluation indicators, 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, which indicates that the antigen peptides generated by the pre-training model of the present invention can accurately restore the required structure.

Table 1 Performance comparison of the fine-tuned model and the pre-fine-tuned model on the sequence generation task

method Sequence recovery rate Perplexity ProteinMPNN 0.254 1.022 This method 0.517 0.486

Table 2 Performance comparison of the fine-tuned model and the pre-fine-tuned model on the structure recovery task of the generated sequence

method pDDT pTM ipTM ProteinMPNN 97.488 0.937 0.904 This method 97.533 0.939 0.902

In one embodiment of the present invention, taking HLA-A*02:01 target and HLA-B*27:05 target as examples, the fine-tuned targeted antigen peptide sequence generation model of the present invention is used for the generation task of targeted HLA I antigen peptides, and after generating 20 antigen peptides respectively, the pHLA I affinity prediction model MHC Fovea is used to predict the binding probability, and the results are shown in Table 3. The present invention has achieved excellent results in the task of generating high-affinity targeted HLA I antigen peptides.

Table 3 Performance of the targeted HLA I antigen peptides generated by the fine-tuned model in the binding probability prediction task

In one embodiment of the present invention, taking HLA-A*02:01 target and HLA-B*27:05 target as examples, among the 20 antigen peptides generated by the fine-tuned targeted antigen peptide sequence generation model in the present invention, the designed peptides with a predicted binding probability greater than that of the natural peptides were screened out, and were input into the target-receptor interaction interface analysis tool for testing, and the results are shown in Tables 4 and 5. Finally, the designed peptides with a Gibbs free energy less than that of the natural peptides, that is, the designed peptides with a stronger affinity than that of the natural peptides, were further screened out. This shows that the present invention can quickly and accurately screen out designed peptides with a stronger affinity than that of the natural peptides.

Table 4 Performance of the targeted HLA-A*02:01 antigen peptides generated by the fine-tuned model in the interaction prediction task

Table 5 Performance of the targeted HLA-B*27:05 antigen peptides generated by the fine-tuned model in the interaction prediction task

The above description is only a preferred specific embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, or directly or indirectly applies it to other related technical fields, which belongs to the content that does not depart from the technical solution of the present invention and is still included in the protection scope of the present invention.

Claims (3)

1. A method for generating and screening a targeting antigen peptide sequence based on an active skeleton, characterized in that it comprises the following steps: Step 1: Obtain the source domain dataset of protein structure from the RCSB Protein Data Bank database and divide it into training set, validation set, and test set; 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, add a set of one-hot encodings as additional input, and input them into the encoder. At the same time, select the distance between the backbone atoms and the 48 amino acid atoms closest to them in Euclidean space and input them into the encoder as edge features. These features are continuously updated during the back propagation process of the model. Step 3: Input the node features and edge features processed by the three-layer encoder and the encoded protein sequence processed by the autoregressive masking technology into the decoder. The decoder uses a disordered decoding strategy to randomly select the next amino acid to be predicted in each decoding step. In any prediction step, the model will use all previously predicted amino acid information to ensure that a 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, the target domain dataset of the pHLAI complex structure was obtained from the RCSB Protein Data Bank database. The obtained PDB files were further screened and processed to form a recombinant dataset. Each genotype in the dataset was divided into a training set, a validation set, and a test set at a ratio of 8:1:1. For genotypes with less than 10 data records, their PDB entries were only used for training and validation and were not included in the test set. Step 6: Use the training set and validation set obtained in step 5 to fine-tune the protein sequence generation model obtained in step 4 to obtain a new target antigen peptide sequence generation model, and use the pre-fine-tuning model and the post-fine-tuning model to test on the test set; Step 7: Use the sequences generated by the pre-fine-tuning model and the post-fine-tuning model and the corresponding target sequence to form a pHLAI complex, input the pHLAI complex sequence into the structure prediction model AlphaFold2 for testing, and evaluate the structural recovery ability of the designed peptide, the structural similarity with the natural peptide, and the binding ability with the target according to the structural 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 an example, 20 antigen peptides were generated using the targeted antigen peptide sequence generation model obtained in step 6, and the generated antigen peptide sequences and the corresponding target sequences were input into the affinity prediction model using the scoring system obtained in step 8 to obtain the predicted binding probability, and the designed peptides with a binding probability greater than that of the natural peptide were screened out; Step 10: Use the designed peptides screened in step 9 to form a complex with the target and input it into AphlaFold2 to predict the structure. Input the complex structure into the target-receptor interaction interface analysis tool for testing to screen out designed peptides with Gibbs free energy less than that of natural peptides. 2. A method for generating and screening targeted antigen peptide sequences based on an active skeleton as described in claim 1, characterized in that the pHLAI complex structure target domain dataset is a small sample dataset, and 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 are obtained from the RCSB Protein Data Bank database and public literature. 3. A method for generating and screening a targeting antigen peptide sequence based on an active skeleton as described in claim 2, characterized in that in step 3, in terms of decoding order, a disordered autoregressive decoding method is used to replace the conventional sequential decoding method from N-terminus to C-terminus.