A method for tumor neoantigen identification based on molecular fingerprint comparison

Abstract

The application discloses a tumor neoantigen identification method based on molecular fingerprint comparison. Traditional tumor neoantigen identification does not consider the influence of polypeptide molecular fingerprint on the immunogenicity of mutant peptides, compared with the method, the method fully utilizes the molecular fingerprint information of polypeptides, can significantly filter out non-immunogenic mutant peptides, and improves the accuracy of tumor neoantigen prediction; the method can provide precise tumor neoantigen targets for tumor vaccines and tumor cell therapy.

Description

Technical Field

[0001] This invention relates to the field of biotechnology, and in particular to a method for identifying neotumor antigens based on molecular fingerprint comparison. Background Technology

[0002] Cancer is a malignant disease that can impair the function of many bodily systems and can be life-threatening. Traditional cancer treatments mainly rely on methods such as chemotherapy, radiotherapy, and surgery to directly kill cancer cells or remove tumor tissue. These methods are often non-specific, meaning they cannot accurately distinguish between cancer cells and normal cells. Immunotherapy, an emerging treatment approach, fights cancer by activating or enhancing the patient's own immune system, and is often more specific than traditional methods.

[0003] When genes mutate within tumor cells, their expression products can be processed by the cells to form a mutant peptide. This mutant peptide can be presented to the surface of tumor cells or dendritic cells by human leukocyte antigen (HLA) molecules, and then recognized by T cell receptor (TCR) molecules on the surface of T cells, initiating the body's anti-tumor immune response. Because neoantigens exist only on tumor cells, are tumor-specific, and do not induce immune tolerance or autoimmunity, they are considered promising targets and are widely used in immunotherapy fields such as tumor vaccines and cell therapy.

[0004] In recent years, with the development of high-throughput sequencing technology, mutant peptides from various sources of variation have been discovered, including genomic single nucleotide variants, insertion / deletion variants, and gene fusion variants. Screening for mutant peptides with high affinity for HLA molecules using machine learning models is currently the main method used to identify tumor neoantigens. While tumor neoantigens identified by this method may have strong affinity for HLA molecules and be presented on the cell surface, the resulting mutant peptides often show little difference in molecular characteristics from wild-type peptides within the cell, failing to effectively activate T-cell immune responses and leading to low efficacy. Therefore, how to analyze the molecular characteristics of neoantigens and accurately and efficiently identify immunogenic mutant peptides from numerous candidate peptides is a crucial problem facing the application of neoantigens.

[0005] This invention employs a method for identifying tumor neoantigens based on molecular fingerprinting. The method first utilizes a deep neural network to construct the three-dimensional conformation of a mutant peptide that binds to HLA molecules. Then, it analyzes the molecular fingerprint of the mutant peptide, predicts its immunogenicity by comparing the differences between the molecular fingerprints of the mutant peptide and the wild-type peptide, and uses the mutant peptide with high immunogenicity as the final neoantigen. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a method for identifying neotumor antigens based on molecular fingerprint comparison.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A method for identifying tumor neoantigens based on molecular fingerprint comparison includes the following steps:

[0009] (1) Perform whole-genome sequencing on tumor samples and compare them with human reference genomes to detect mutations in the samples. Generate mutant peptides and wild peptides as candidate neoantigens centered on mutation sites.

[0010] (2) Based on the amino acid sequence of the HLA molecule corresponding to the tumor sample and the amino acid sequence of the mutant peptide and wild peptide obtained in step (1), the three-dimensional structure data of the peptide molecule bound to the HLA molecule are calculated by the peptide three-dimensional structure prediction model.

[0011] (3) Based on the three-dimensional structural data of the mutant peptide / wild peptide obtained in step (2), calculate its molecular fingerprint and calculate the difference between the two sets of molecular fingerprints to obtain a set of numerical vectors.

[0012] (4) Based on the numerical vector obtained in step (3), calculate the immunogenicity score of the mutant peptide using the immunogenicity prediction model, and determine whether the mutant peptide has immunogenicity based on the score.

[0013] The mutant peptides and wild peptides mentioned in step (1) are 7 to 13 amino acids in length.

[0014] The polypeptide three-dimensional structure prediction model described in step (2) includes a sequence data processing module, an embedding module, a vector conversion module, an encoding module, and a decoding module.

[0015] The sequence data processing module is used to read the amino acid sequences of peptides and HLA molecules and form vectors; preferably, it reads the amino acid sequence data of peptides and HLA molecules, with lengths p and h respectively, and merges the amino acid sequences of peptides and HLA molecules to form a vector of length p+h.

[0016] The aforementioned embedding module is used to map the vectors formed by the data processing module to a d-dimensional vector space.

[0017] The vector transformation module is used to transform vectors mapped to a d-dimensional vector space into interaction matrices; preferably, PyTorch broadcast addition operations are used to transform vectors into interaction matrices; more preferably, the interaction matrix contains interaction information within peptide molecules, interaction information within HLA molecules, and interaction information between peptide molecules and HLA molecules.

[0018] The encoding module is used to extract features from the matrix obtained by the embedding module; preferably, the encoding module includes 8 blocks, each block is composed of 4 computation layers connected in sequence, namely attention computation layer, layer normalization layer, feature optimization layer, and layer normalization layer.

[0019] The decoding module is used to predict the coordinates of atomic groups from the feature matrix obtained by the encoding module. Preferably, the decoding module includes three blocks, each block consisting of seven computational layers connected in sequence: a linear layer, a ReLU activation layer, a Batch normalization layer, a linear layer, a ReLU activation layer, a linear layer, and a coordinate transformation layer.

[0020] The training process of the polypeptide three-dimensional structure prediction model described in step (2) includes the following steps:

[0021] (A1) Obtain the training dataset for the three-dimensional structure of peptides;

[0022] (A2) The amino acid sequence data of peptides and HLA molecules are formatted and used as input features. The corresponding three-dimensional structure data of peptides are used as training labels to construct a dataset for training deep neural network models.

[0023] (A3) Select several dataset samples and feed them into a deep neural network model to output the prediction results of the three-dimensional structure of the polypeptide;

[0024] (A4) Calculate the error between the predicted three-dimensional structure of the polypeptide and the actual structure, and optimize the model parameters of the deep neural network based on the Adam optimization method.

[0025] (A5) Repeat steps (A3) to (A5) until the model parameters of the deep neural network converge, thus completing the training process.

[0026] The formatting described in step (A2) includes the following steps:

[0027] The amino acid sequences are formatted as numerical sequence data from 0 to 19 in the order of [A,R,N,D,C,Q,E,G,H,I,L,K,M,F,P,S,T,W,Y,V].

[0028] The three-dimensional structure described in step (2) or (3) is three-dimensional structural data in the form of rigid body transformation.

[0029] The rigid body transformation form of the three-dimensional structural data is obtained by using the Gram-Schmidt orthogonalization algorithm to transform the spatial coordinates. The calculation includes the following steps:

[0030]

[0031]

[0032]

[0033]

[0034]

[0035]

[0036] Rotation matrix Translation vector Rigid body transformation This yields three-dimensional structural data in the form of a rigid body transformation.

[0037] The molecular fingerprint described in step (3) includes data on four dimensions of the polypeptide molecule surface: hydrophobicity, electrostatic potential, free electrons, and hydrogen bond donors; preferably, it is calculated using MaSIF software.

[0038] The method for calculating the difference in molecular fingerprints in step (3) is to standardize the maximum value of the data and then subtract them by dimension to finally form a set of vectors; preferably, the maximum value of the molecular fingerprint is standardized in the four dimensions of hydrophobicity, electrostatic potential, free electrons and hydrogen bond donors respectively, and then subtracted by dimension to finally form a set of four-dimensional vectors.

[0039] The immunogenicity prediction model described in step (4) includes a vector data processing module and a prediction module.

[0040] The vector data processing module is used to read the vector formed by the differences in molecular fingerprints.

[0041] The prediction module is used to predict the immunogenicity score based on the vector formed by the differences in molecular fingerprints. Preferably, the prediction module sequentially passes through a linear layer, a ReLU activation layer, a Batch normalization layer, and the linear layer processes the differential data. The last linear layer maps multiple neurons into one neuron, which is the predicted immunogenicity score.

[0042] The training process of the immunogenicity prediction model described in step (4) includes the following steps:

[0043] (B1) Immunogenicity data of mutant peptides were extracted from the Immunoeptope Database (IEDB), and wild-type peptides corresponding to these mutant peptides were obtained by sequence alignment.

[0044] (B2) The three-dimensional structure data of mutant peptides and wild peptides were calculated using a peptide three-dimensional structure prediction model;

[0045] (B3) Based on the obtained three-dimensional structural data, calculate the data of hydrophobicity, electrostatic potential, free electrons and hydrogen bond donors on the surface of mutant peptides and wild peptides. Standardize the maximum value of each, and then subtract them by dimension to finally form a set of four-dimensional vectors.

[0046] (B4) Using the four-dimensional vector obtained in step (B3) as the input feature and the immunogenicity data of the mutant peptide as the training label, a dataset for training the neural network model is constructed.

[0047] (B5) Select several dataset samples and feed them into the neural network model to output the predicted immunogenicity score of the mutant peptide.

[0048] (B6) Calculate the error between the predicted immunogenicity score of the mutant peptide and the actual result, and optimize the model parameters of the neural network based on the Adam optimization method.

[0049] (B7) Repeat steps (B5) to (B6) until the parameters of the neural network model converge, thus completing the training process.

[0050] The mutant peptides and wild-type peptides described in step (B1) are 7 to 13 amino acids in length.

[0051] The method for obtaining immunogenicity data described in step (B1) includes the following steps:

[0052] Experimental data for different epitopes were downloaded from the Immune Epitope Database (IEDB). Data with unclear HLA phenotypes were removed, and linear epitopes, epitopes with human host, epitopes with T cell experimental validation results, and epitopes with disease types of cancer or infectious diseases were selected. If a peptide had multiple experimental data sets with consistent results, only one data set was retained, and the remaining duplicate data were removed. If a peptide had multiple experimental data sets with both positive and negative results, only the positive data was retained, and the remaining data were removed. Data with positive T cell responses were marked as IS=1, and data with negative T cell responses were marked as IS=0.

[0053] The method for obtaining the wild-type peptide described in step (B1) includes the following steps:

[0054] Using the blastp software, the sequence of the mutant peptide was compared with the human reference genome sequence. The sequence with a mismatch number of 1 in the comparison results was the wild peptide.

[0055] The method for calculating the error described in step (B6) is mean square error loss (MSE loss).

[0056] The Adam optimization method described in step (B6) uses an adaptive learning rate as the input to the optimizer, employs a mini-batch gradient descent algorithm, sets the batch size and the maximum number of iterations, and obtains the final prediction model when the loss function of the validation set no longer increases.

[0057] A novel tumor antigen identification system based on molecular fingerprint comparison includes a polypeptide three-dimensional structure prediction model and an immunogenicity prediction model. The three-dimensional model data is obtained by inputting the data of the antigen to be tested into the polypeptide three-dimensional structure prediction model, and then the three-dimensional model data is input into the immunogenicity prediction model to obtain the immunogenicity score.

[0058] The present invention has the following advantages and effects compared with the prior art:

[0059] Traditional tumor neoantigen identification methods do not consider the influence of peptide molecular fingerprints on the immunogenicity of mutant peptides. In contrast, the method described in this invention fully utilizes the molecular fingerprint information of peptides, which can significantly filter out non-immunogenic mutant peptides and improve the accuracy of tumor neoantigen prediction. The method described in this invention can provide precise tumor neoantigen targets for tumor vaccines and tumor cell therapies. Attached Figure Description

[0060] Figure 1 This is a flowchart of the method of the present invention.

[0061] Figure 2 This is a statistical chart showing the Immunogenicity Score (IS) of non-immunogenic mutant peptides and immunogenic mutant peptides; there are 417 non-immunogenic mutant peptide samples and 259 immunogenic mutant peptide samples. Based on the statistical results, it can be seen that the IS value of immunogenic mutant peptides is significantly higher than that of non-immunogenic mutant peptides (P<0.001).

[0062] Figure 3 This is a statistical chart showing the proportion of samples with low IS values ​​(IS<0.2) and high IS values ​​(IS>=0.2) among non-immunogenic mutant peptides and immunogenic mutant peptides.

[0063] Figure 4 This is a schematic diagram of a peptide three-dimensional structure prediction model; where a is the model framework for calculating the peptide three-dimensional structure, and b is a schematic diagram of the interaction matrix.

[0064] Figure 5 This is a schematic diagram of the model framework for calculating IS values ​​in an immunogenicity prediction model.

[0065] Figure 6These are the predicted peptide structures from Example 3; where a is the structure of peptide ASLNLPAVSW bound to HLA-B*57:03, the blue molecule is the actual peptide structure, the red molecule is the peptide structure predicted by the model, and the gray molecule is HLA-B*57:03; b is the structure of peptide GTSGSPIINR bound to HLA-A*11:01, the blue molecule is the actual peptide structure, the red molecule is the peptide structure predicted by the model, and the gray molecule is HLA-A*11:01. Detailed Implementation

[0066] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0067] Unless otherwise specified in the following implementation plan, the test conditions are generally as per standard test conditions or the test conditions recommended by the reagent company. Unless otherwise specified, all materials and reagents used are commercially available.

[0068] Example 1: Establishment of a three-dimensional structure prediction model for polypeptides

[0069] 1.1 Establishing a three-dimensional structure prediction model for peptides

[0070] To accurately predict the three-dimensional structure formed after an unknown antigen sequence binds to an HLA molecule, this embodiment obtains peptide information of known antigens and HLA, as well as three-dimensional structure data of antigen formation, from publicly available databases. This information is then input into a deep learning model to obtain a deep learning model capable of predicting the three-dimensional structure of unknown antigens. The specific steps are as follows:

[0071] (1) Obtain the training dataset of polypeptide three-dimensional structure

[0072] Enter the keyword "MHC" into the Protein Structure Database (PDB); screen for complex data containing complete HLA and peptide information; limit the HLA length range to 260–290 and the peptide length range to 7–13; use the biopython package to extract peptide 3D structure data, peptide and HLA molecular sequence data; finally, 1018 samples were obtained, of which 90% were used for model training and validation, and 10% were used for model testing.

[0073] (2) The amino acid sequences of peptides and HLA molecules in the training set are formatted into numerical sequence data from 0 to 19 in the order [A,R,N,D,C,Q,E,G,H,I,L,K,M,F,P,S,T,W,Y,V]. Since each atomic group of peptide amino acid residues can be treated as a rigid body, the three-dimensional structure data of the peptide is converted into rigid body transformation (T...i The transformation is achieved using the Gram-Schmidt orthogonalization algorithm, specifically: for any three-dimensional structural coordinates of a group of atoms... Calculate separately:

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080] Rotation matrix Translation vector Rigid body transformation The three-dimensional structural data mentioned below all refer to rigid body transformation (T) i Three-dimensional structural data in the form of )

[0081] (3) Using the numerical sequence data of peptides and HLA molecules as input features and the three-dimensional structure data of peptides as training labels, a dataset for training deep neural network models is constructed.

[0082] (4) Select several dataset samples and feed them into a deep neural network model to output the prediction results of the polypeptide three-dimensional structure data. The method for constructing the deep learning model specifically includes the following steps:

[0083] 1. A model was constructed using PyTorch to read the numerical sequence data of peptides and HLA molecules, with lengths p and h, respectively. The numerical sequences of the peptides and HLA molecules were merged to form a vector of length p+h.

[0084] 2. Define an embedder to map the above vectors to a d-dimensional vector space;

[0085] 3. Use PyTorch's broadcast addition operation to transform the above vector into an interaction matrix. This interaction matrix contains three parts: interaction information within peptide molecules, interaction information within HLA molecules, and interaction information between peptide molecules and HLA molecules. Figure 4(As shown in b) The interaction information between peptide molecules and HLA molecules is one of the key factors in optimizing model performance, ensuring that the model can construct the three-dimensional structure of the peptide under the constraint of HLA molecular information. Existing peptide homology modeling methods (such as PANDORA, which currently has the best performance) require manually specifying the anchor points of the peptide and HLA molecules to constrain the peptide structure during homology modeling, resulting in lower accuracy compared to this model.

[0086] 4. Define an encoder to extract features from the matrix above. The encoder consists of 8 blocks, each block is composed of 4 computational layers connected in sequence, including: attention computation layer, layer normalization layer, feature optimization layer, and layer normalization layer.

[0087] 5. Define a decoder to predict the coordinates of atomic groups from the feature matrix. The decoder consists of 3 blocks, each of which is composed of 7 computational layers connected in sequence, including: linear layer, ReLU activation layer, Batch normalization layer, linear layer, ReLU activation layer, linear layer, and coordinate transformation layer.

[0088] The final output of the model is the three-dimensional structural data of the peptide;

[0089] (5) Calculate the error between the predicted three-dimensional structure of the polypeptide and the actual structure, and optimize the model parameters of the deep neural network based on the Adam optimization method. The calculation method for the error between the predicted structure and the actual structure is as follows:

[0090] By comparing the 3D structural data output by the model with the real 3D structural data in the database, the binary cross-entropy loss of the distance distribution of internal atomic groups is calculated. The model is optimized by the Adam optimizer, using an adaptive learning rate as the input of the optimizer and a mini-batch gradient descent algorithm. The batch size and the maximum number of iterations are set. When the loss function of the validation set no longer increases, the final prediction model is obtained.

[0091] (6) Repeat steps (4) to (5) until the model parameters of the deep neural network converge, completing the training process and obtaining the three-dimensional structure prediction model of the polypeptide.

[0092] 1.2 Using a three-dimensional peptide structure prediction model to predict the structure formed after peptides bind to HLA molecules

[0093] (1) Download the amino acid sequence information of the target HLA type from the Immuno Polymorphism Database (IPD) https: / / www.ebi.ac.uk / ipd / imgt / hla / alleles / ;

[0094] (2) Format the amino acid sequence information of the peptide and the corresponding HLA into numerical sequence data;

[0095] (3) Input the above numerical sequence data into the polypeptide three-dimensional structure prediction model obtained in 1.1;

[0096] (4) Generate the three-dimensional structure of the polypeptide.

[0097] 1.3 Validation of the predicted structure formed by peptide ASLNLPAVSW binding to HLA-B*57:03

[0098] Following the method described in section 1.2, using the peptide ASLNLPAVSW as an example, its structure after binding with HLA-B*57:03 was predicted. Sample 6v2p from the PDB database was used to assess the accuracy of the generated structure. The root mean square error (RMSD) between the predicted and actual structures was [value missing]. (like Figure 6 As shown in a), the predicted structure has good consistency with the actual structure.

[0099] 1.4 Validation of the predicted structure formed after peptide GTSGSPIINR binds to HLA-A*11:01

[0100] Following the method described in section 1.2, using the peptide GTSGSPIINR as an example, its structure after binding with HLA-A*11:01 was predicted. Sample 5wkh from the PDB database was used to assess the accuracy of the generated structure. The root mean square error (RMSD) between the predicted and actual structures was [value missing]. (like Figure 6 As shown in b), the predicted structure has a slight misalignment of isoleucines at positions 7 and 8.

[0101] Example 2: Establishment of an immunogenicity prediction model

[0102] 2.1 Establishing an immunogenicity prediction model

[0103] (1) Immunogenicity data of mutant peptides were extracted from the Immunoeptope Database (IEDB), and wild peptides corresponding to these mutant peptides were obtained by sequence alignment.

[0104] The method for obtaining the immunogenicity dataset is as follows:

[0105] Download experimental data for different epitopes from the Immune Epitope Database (IEDB); remove data with unclear HLA phenotypes; screen for linear epitopes; screen for epitopes with human host; screen for epitopes with T cell experimental validation results; screen for epitopes with disease types of cancer or infectious diseases; if a peptide has multiple experimental data and the results are consistent, retain one data point and remove the remaining duplicate data; if a peptide has multiple experimental data and the results are inconsistent, retain only the data that is validated as positive and remove the remaining data; the inconsistency rate refers to the experimental conclusions of the same antigen peptide and the same HLA typing, including both positive and negative results; data with positive T cell reactions are marked as IS=1; data with negative T cell reactions are marked as IS=0.

[0106] The method for obtaining the wild-type peptide is as follows:

[0107] Using blastp software, mutant peptides were compared with human proteome sequences, and sequences with a mismatch count of 1 were retained as wild peptides. A total of 6,761 samples were obtained, of which 90% were used for model training and validation, and 10% were used for model testing.

[0108] (2) The three-dimensional structure prediction model of the polypeptide obtained in Example 1 was used to calculate the three-dimensional structure formed by the above mutant peptide and wild peptide after binding to the HLA molecule;

[0109] (3) Calculate the molecular fingerprints of the mutant peptides and wild peptides mentioned above. The molecular fingerprints include data on four dimensions of the surface of the polypeptide molecule: hydrophobicity, electrostatic potential, free electrons, and hydrogen bond donors. The molecular fingerprints are analyzed and calculated using MaSIF software.

[0110] (4) Calculate the molecular fingerprint differences between the mutant peptide and the wild-type peptide. The method for calculating the molecular fingerprint differences is as follows: standardize the molecular fingerprints of the mutant peptide and the wild-type peptide according to the maximum values ​​in four dimensions: hydrophobicity, electrostatic potential, free electrons, and hydrogen bond donors. Then, subtract them according to the dimensions to finally form a set of four-dimensional vectors. The comparison of the molecular fingerprints of the mutant peptide and the wild-type peptide is one of the keys to the optimization of model performance. It ensures that the model can predict the immunogenicity of the mutant peptide from the perspective of molecular physicochemical properties. There are two existing methods for predicting immunogenicity: one is to calculate the sequence similarity between the mutant peptide and the pathogenic microorganism (…). Secondly, the sequence differences between mutant peptides and human autopeptides are calculated (Richman et al., 2017); both methods are based on sequence information comparison, and their accuracy is lower than that of this model.

[0111] (5) Using molecular fingerprint differences as input features and immunogenicity data of mutant peptides as training labels, a dataset for training a neural network model is constructed, wherein the neural network model is specifically:

[0112] A model was built using PyTorch. Molecular fingerprint differences between mutant and wild-type peptides were then analyzed.

[0113] The differential data are processed sequentially through a linear layer, a ReLU activation layer, a batch normalization layer, and another linear layer.

[0114] The last linear layer maps multiple neurons to a single neuron, which is the predicted IS value;

[0115] (6) Select several dataset samples and feed them into the neural network model to output the predicted immunogenicity score of the mutant peptide.

[0116] (7) Calculate the error between the predicted immunogenicity score of the mutant peptide and the actual result, and optimize the model parameters of the above neural network based on the Adam optimization method.

[0117] (8) Repeat steps (6) to (7) until the parameters of the above neural network model converge, complete the training process, and obtain the immunogenicity prediction model.

[0118] Example 3: Screening and identification of antigens by combining polypeptide three-dimensional structure prediction models and immunogenicity prediction models.

[0119] 3.1 Application of deep learning models to screen and identify tumor neoantigens

[0120] (1) Whole-genome sequencing was performed on tumor tissue samples and normal tissue samples from the patient. The obtained sequencing data were aligned to the human reference genome, and mutations in the tumor tissue samples were detected by sample comparison. Mutant peptide fragments and wild-type peptide fragments were generated with the mutation site as the center. The length of the peptide fragments was 7-13 amino acids.

[0121] (2) Based on the patient's HLA typing information and the polypeptide sequence information obtained in step (1), the data is formatted into numerical sequence data of a certain length and imported into the polypeptide three-dimensional structure prediction model obtained in Example 1. The three-dimensional structures formed by the mutant peptide and wild peptide after binding to the HLA molecule are calculated respectively.

[0122] (3) Calculate their molecular fingerprints based on the three-dimensional structures formed after mutant peptides and wild peptides bind to HLA molecules;

[0123] (4) Calculate the difference in molecular fingerprints between mutant peptides and wild peptides, and represent the difference as a set of numerical vectors;

[0124] (5) Based on the difference vector of the molecular fingerprint above, import the immunogenicity prediction model obtained in Example 2, calculate the immunogenicity score (IS) of the mutant peptide, and the mutant peptide with IS>=0.2 is the final tumor neoantigen.

[0125] 3.2 Screening antigens and predicting their immunogenicity using human ovarian cancer mutant protein KDM5C as the target.

[0126] Taking the human ovarian cancer mutant protein KDM5C as an example, the detected mutation in this protein is E656K (at position 656 of the protein, glutamic acid E is mutated to lysine K). The mutant peptide sequence is KMAACP[K]KL, the wild peptide sequence is KMAACP[E]KL, and the patient has an HLA-A*A0211 genotype. The calculated IS value of this mutant peptide is 0.51, satisfying the condition IS>=0.2; therefore, this predicted peptide can be considered a candidate peptide for tumor neoantigens.

[0127] Further literature data (PMID:29643231) demonstrates that this antigen (KMAACP[K]KL) can bind to HLA-A*A0211 and activate CD8+ T cells, making it a novel antigen with immunogenicity.

[0128] 3.3 Screening antigens targeting human melanoma mutant protein RFX1 and predicting their immunogenicity

[0129] Taking the human melanoma mutant protein RFX1 as an example, the detected mutation in this protein is S94L (at position 94 of the protein, a mutation from threonine S to leucine L). The mutant peptide sequence is QPTGAPTP[L], and the wild-type peptide sequence is QPTGAPTP[S]. The patient has an HLA-A*B0702 genotype. The calculated IS value of this mutant peptide is 0.08, which does not meet the condition IS>=0.2. Therefore, this predicted polypeptide is a non-immunogenic polypeptide and should not be considered as a candidate neoantigen.

[0130] Further literature data (PMID:27198675) proves that this antigen (QPTGAPTP[L]) cannot activate CD8+ T cells and is a non-immunogenic peptide.

[0131] 3.4 Screening antigens and predicting their immunogenicity using the human non-small cell lung cancer mutant protein HERC1 as a target.

[0132] Taking the human non-small cell lung cancer mutant protein HERC1 as an example, the mutation detected in this protein is P3278S (at position 3278 of the protein, proline P is mutated to threonine S). The mutant peptide sequence is ASNA[S]SAAK, the wild-type peptide sequence is ASNA[P]SAAK, and the patient has an HLA-A*A1101 genotype. The calculated IS value of this mutant peptide is 0.44, satisfying the condition IS>=0.2; therefore, this predicted peptide can be considered a candidate peptide for tumor neoantigens.

[0133] Further literature data (PMID:25765070) demonstrates that this antigen (ASNA[S]SAAK) can bind to HLA-A*A1101 and activate CD8+ T cells, making it a novel immunogenic antigen.

[0134] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for identifying tumor neoantigens based on molecular fingerprint comparison, characterized in that... Includes the following steps: (1) Perform whole-genome sequencing on tumor samples and compare them with human reference genomes to detect mutations in the samples. Generate mutant peptides and wild peptides as candidate neoantigens centered on mutation sites. (2) Based on the amino acid sequence of the HLA molecule corresponding to the tumor sample and the amino acid sequence of the mutant peptide and wild peptide obtained in step (1), the three-dimensional structure data of the peptide molecule bound to the HLA molecule are calculated by the peptide three-dimensional structure prediction model. (3) Based on the three-dimensional structural data of the mutant peptide / wild peptide obtained in step (2), calculate its molecular fingerprint and calculate the difference between the two sets of molecular fingerprints to obtain a set of numerical vectors. (4) Based on the numerical vector obtained in step (3), calculate the immunogenicity score of the mutant peptide using the immunogenicity prediction model, and determine whether the mutant peptide has immunogenicity based on the score. The method for calculating the difference in molecular fingerprints in step (3) is to standardize the maximum value of the molecular fingerprint in four dimensions: hydrophobicity, electrostatic potential, free electrons and hydrogen bond donors, and then subtract them according to the dimensions to finally form a set of four-dimensional vectors. The training process of the immunogenicity prediction model described in step (4) includes the following steps: (B1) Immunogenicity data of mutant peptides were extracted from an immune epitope database, and wild-type peptides corresponding to these mutant peptides were obtained by sequence alignment. (B2) Calculate the three-dimensional structure data of mutant and wild peptides using a peptide three-dimensional structure prediction model; (B3) Based on the obtained three-dimensional structural data, calculate the data of hydrophobicity, electrostatic potential, free electrons and hydrogen bond donors on the surface of mutant peptides and wild peptides. Standardize the maximum value of each, and then subtract them by dimension to finally form a set of four-dimensional vectors. (B4) Using the four-dimensional vector obtained in step (B3) as the input feature and the immunogenicity data of the mutant peptide as the training label, a dataset for training the neural network model is constructed. (B5) Select several dataset samples and feed them into the neural network model to output the predicted immunogenicity score of the mutant peptide; (B6) Calculate the error between the predicted immunogenicity score of the mutant peptide and the actual result, and optimize the model parameters of the neural network based on the Adam optimization method; (B7) Repeat steps (B5) to (B6) until the parameters of the neural network model converge, thus completing the training process.

2. The method according to claim 1, characterized in that: The mutant peptide and wild peptide mentioned in step (1) are 7 to 13 amino acids in length; The polypeptide three-dimensional structure prediction model described in step (2) includes a sequence data processing module, an embedding module, a vector conversion module, an encoding module, and a decoding module; The three-dimensional structure described in step (2) or (3) is three-dimensional structural data in the form of rigid body transformation; The molecular fingerprint described in step (3) includes data on four dimensions of the polypeptide molecule surface: hydrophobicity, electrostatic potential, free electrons, and hydrogen bond donors. The molecular fingerprint was calculated using MaSIF software; The immunogenicity prediction model described in step (4) includes a vector data processing module and a prediction module.

3. The method according to claim 2, characterized in that: The sequence data processing module is used to read the amino acid sequence data of peptides and HLA molecules, with lengths p and h respectively, and to merge the amino acid sequences of peptides and HLA molecules to form a vector of length p + h. The aforementioned embedding module is used to map the vectors formed by the data processing module to a d-dimensional vector space; The vector transformation module is used to transform vectors mapped to d-dimensional vector space into interaction matrices. The vector conversion module uses PyTorch broadcast addition operations to convert vectors into interactive matrices; The interaction matrix includes information on interactions within peptide molecules, interactions within HLA molecules, and interactions between peptide molecules and HLA molecules. The encoding module is used to extract features from the matrix obtained by the embedding module; The encoding module consists of 8 blocks, each block being composed of 4 computational layers connected in sequence: attention computation layer, layer normalization layer, feature optimization layer, and layer normalization layer. The decoding module is used to predict the coordinates of atomic groups from the feature matrix obtained by the encoding module; The decoding module comprises three blocks, each consisting of seven computational layers connected in sequence: a linear layer, a ReLU activation layer, a Batch normalization layer, a linear layer, a ReLU activation layer, a linear layer, and a coordinate transformation layer.

4. The method according to claim 2, characterized in that: The vector data processing module is used to read the vector formed by the differences in molecular fingerprints; The prediction module is used to predict the immunogenicity score based on the vector formed by the differences in molecular fingerprints. The prediction module sequentially passes through a linear layer, a ReLU activation layer, a batch normalization layer, and a linear layer to process differential data. The last linear layer maps multiple neurons into one neuron, which is the predicted immunogenicity score.

5. The method according to claim 2, characterized in that: The rigid body transformation form of the three-dimensional structural data is obtained by using the Gram-Schmidt orthogonalization algorithm to transform the spatial coordinates. , , The calculation includes the following steps: ； Rotation matrix R i = concat( , , Translation vector = Rigid body transformation T i = ( R i , i This yields the three-dimensional structural data in the form of a rigid body transformation.

6. The method according to claim 1, characterized in that: The training process of the polypeptide three-dimensional structure prediction model described in step (2) includes the following steps: (A1) Obtain the training dataset for the three-dimensional structure of peptides; (A2) The amino acid sequences of peptides and HLA molecules are formatted into numerical sequence data from 0 to 19 in the order of [A, R, N, D, C, Q, E, G, H, I, L, K, M, F, P, S, T, W, Y, V] and used as input features. The corresponding three-dimensional structure data of the peptides are used as training labels to construct a dataset for training deep neural network models. (A3) Select several dataset samples and feed them into a deep neural network model to output the prediction results of the three-dimensional structure of the polypeptide; (A4) Calculate the error between the predicted three-dimensional structure of the polypeptide and the actual structure, and optimize the model parameters of the deep neural network based on the Adam optimization method; (A5) Repeat steps (A3) to (A5) until the model parameters of the deep neural network converge, thus completing the training process.

7. The method according to claim 1, characterized in that: The mutant peptides and wild-type peptides described in step (B1) are 7 to 13 amino acids in length; The method for obtaining immunogenicity data described in step (B1) includes the following steps: Experimental data for different epitopes were downloaded from the immune epitope database. Data with unclear HLA phenotypes were removed, and linear epitopes, epitopes with human host, epitopes with T cell experimental validation results, and epitopes with disease types of cancer or infectious diseases were selected. If a peptide had multiple experimental data sets with consistent results, only one data set was retained, and the remaining duplicate data were removed. If a peptide had multiple experimental data sets with both positive and negative results, only the positive data was retained, and the remaining data were removed. Data with a positive T-cell response were marked as IS=1, and data with a negative T-cell response were marked as IS=0; The method for obtaining the wild-type peptide described in step (B1) includes the following steps: Using blastp software, the sequence of the mutant peptide was compared with the human reference genome sequence. The sequence with a mismatch number of 1 in the comparison results was the wild peptide. The error calculation method described in step (B6) is the mean square error loss; The Adam optimization method described in step (B6) uses an adaptive learning rate as the input to the optimizer, employs a mini-batch gradient descent algorithm, sets the batch size and the maximum number of iterations, and obtains the final prediction model when the loss function of the validation set no longer increases.

8. A tumor neoantigen identification system based on molecular fingerprint comparison, characterized in that: The method includes constructing a three-dimensional structure prediction model and an immunogenicity prediction model for peptides using the method of claim 1. The three-dimensional model data is obtained by inputting the data of the antigen to be tested into the three-dimensional structure prediction model of peptides, and then the three-dimensional model data is input into the immunogenicity prediction model to obtain the immunogenicity score.

Images