Immunogenicity prediction method and device based on multimode fusion and storage medium

Abstract

The invention provides an immunogenicity prediction method and device based on multimode fusion and a storage medium. The method comprises the following steps: constructing a multimode prediction model; obtaining each data set; correspondingly training a multi-modal prediction model based on each data set to obtain a trained multi-modal prediction model; fusing the trained multi-modal prediction models, and constructing an immunogenicity comprehensive evaluation model; forming an intelligent immunogenicity prediction system based on the trained multi-modal prediction model and the immunogenicity comprehensive evaluation model; and performing immunogenicity prediction on the to-be-predicted protein based on the immunogenicity prediction intelligent system to obtain a test result, the test result comprising the immunogenicity risk score of the to-be-predicted protein and the potential immunogenicity T cell epitope on the to-be-predicted protein. According to the method, the effect of directionally optimizing the immunogenicity on the basis of the prediction result can be achieved, and the drug effect and safety of protein biological drugs can be improved.

Description

Technical Field

[0001] The present application relates to the field of bioinformatics technology, and in particular to an immunogenicity prediction method, device and storage medium based on multi-modal fusion. Background Art

[0002] Immunity is a physiological function of the human body. The human body relies on this function to identify self and non-self substances in the body, and destroys and eliminates foreign substances (such as bacteria, viruses, etc.) or damaged cells and tumor cells produced by the human body itself through immune response to maintain human health. Generally speaking, substances that can induce immune response in the human body are called antigens, and proteins that are produced due to antigen stimulation and can specifically bind to the corresponding antigen and have a protective effect are called antibodies. Immunogenicity refers to the ability to cause an immune response. Antigens that enter the human body can stimulate specific immune cells (such as T lymphocytes), causing them to activate, proliferate, and differentiate, and ultimately produce antibodies and primed lymphocytes, and cause immune responses.

[0003] Since antibody drugs, fusion protein drugs, etc. are still exogenous proteins in nature, their immunogenicity has become an issue that cannot be ignored. It refers to the ability of biotechnology drugs to induce the body's immune response to itself or related proteins (such as therapeutic proteins) or cause immune-related events. The induced adverse events include the formation of anti-drug antibodies (ADA) and / or neutralizing antibodies (NAb). The former refers to the production of antibodies that can bind to drugs after medication, which may cause a strong immune response in patients and even endanger their lives; while the latter can neutralize the biological activity of drugs, thereby weakening the efficacy of drugs. For patients, it affects the safety and effectiveness of drugs, and may even cause fatal new diseases to patients due to the cross-reaction between ADA and endogenous proteins. For enterprises, the immunogenicity of many antibody drugs was not discovered until the late stage of clinical development, resulting in the abortion of many products in Phase 3 clinical trials, which greatly increased the R&D costs and risks of pharmaceutical companies. For example, although the PCSK9 antibody bococizumab developed by Pfizer was effective in preclinical studies, it was forced to declare Phase 3 clinical failure because neutralizing antibodies directly affected drug metabolism and efficacy. For drug regulatory authorities, immunogenicity testing is essential, and its evaluation is based on a holistic study of risk assessment, which usually runs through the entire process of biotechnology drug development until after listing. For vaccine drugs, their mechanism of action is to cause the body to respond to the antigen corresponding to the vaccine, so their immunogenicity is the source of efficacy, that is, contrary to the desire of antibody drugs to remove immunogenicity, immunogenicity is required by vaccine drugs. Therefore, whether it is an antibody or a vaccine, immunogenicity is an important factor that cannot be ignored in exerting efficacy.

[0004] There are two main mechanisms for the production of anti-drug antibodies ADA: T cell-dependent and T cell-independent pathways. The former is the main source of immunogenicity of therapeutic proteins such as antibodies or fusion proteins, involving the uptake of biologics by antigen-presenting cells (APCs, such as immature dendritic cells (iDCs) and B cells), which triggers T cell-specific immune responses and produces plasma cells that secrete antigen-specific antibodies, thereby producing complexes of biologic-derived peptides and major histocompatibility complex (MHC) class II molecules, which interact with specific TCR sequences. Peptides that meet these criteria are called immunogenic peptides; the latter involves direct stimulation of B cells by aggregated forms of the organism, mainly plasma cells that secrete IgM antibodies. Specifically, human MHC alleles are translated into human leukocyte antigens (HLA).

[0005] Among the many factors that affect the immunogenicity of biological agents, amino acid sequence variation is one of the most important and basic factors. It is difficult to predict absolute immunogenicity in clinical trials using only amino acid sequences, but the potential immunogenicity caused by sequence differences can be predicted by computer, in vitro or in vivo methods. Since in vitro and in vivo experimental methods are usually simple and difficult to accurately reflect the results of the complex human immune system, the predictions of these methods are not highly correlated with clinical results. Therefore, using computer systems to predict immunogenicity has become an important approach.

[0006] At present, many computer software have been developed to predict the binding (affinity) of peptides to MHC, such as NetMHCIIpan, IEDB website online tools, etc. However, from a biological perspective, the binding of peptides to MHC is only a necessary but insufficient link in the process of the human immune response. The immune response caused by drug molecules is also affected by multiple links and factors such as protease cleavage, peptide presentation, and immune tolerance. Therefore, affinity prediction is not sufficient to accurately predict immunogenicity and has great limitations, such as too strong false positives and poor correlation with real-world clinical data.

[0007] In summary, there is an urgent need to develop more complete methods that can more comprehensively simulate the real immune response process of the immune system and then predict immunogenicity, so that candidate drug molecules can be screened based on immunogenicity, and even immunogenicity can be optimized based on the predicted results, thereby improving the efficacy and safety of protein biological drugs. Summary of the invention

[0008] The present application provides a method, device and storage medium for predicting immunogenicity based on multi-modal fusion to solve the problems existing in the related technologies. The technical solution is as follows:

[0009] In a first aspect, the present application provides an immunogenicity prediction method based on multi-modal fusion, comprising:

[0010] Constructing a multimodal prediction model, the multimodal prediction model includes an affinity prediction model, a presentation prediction model, an MHC weight learning model and an immune tolerance assessment model, the affinity prediction model is used to predict the affinity of a peptide segment to each MHC allele subtype, the presentation prediction model is used to predict the possibility of a peptide segment being naturally processed by antigen presenting cells and presented by each MHC allele subtype, the MHC weight learning model is used to learn the frequency weight of HLA, and the immune tolerance assessment model is used to assess the possibility of a peptide segment being tolerated by the immune system;

[0011] Acquire various data sets, including immunopeptidomics data sets and human proteomics and immunoomics data sets;

[0012] Training the multimodal prediction model based on each of the data sets to obtain a trained multimodal prediction model;

[0013] The trained multimodal prediction models are integrated to construct a comprehensive immunogenicity assessment model, wherein the comprehensive immunogenicity assessment model is used to assess the complete immunogenicity risk based on the full-length protein sequence based on the output data of the affinity prediction model, the presentation prediction model, the MHC weight learning model, and the immune tolerance assessment model, wherein the input data of the MHC weight learning model are the output data of the affinity prediction model and the output data of the presentation prediction model, and the input data of the immune tolerance assessment model are the output data of the affinity prediction model and the output data of the presentation prediction model;

[0014] Based on the trained multimodal prediction model and the immunogenicity comprehensive evaluation model, an immunogenicity prediction intelligent system is formed, wherein the input models of the immunogenicity prediction intelligent system are the affinity prediction model and the presentation prediction model, the intermediate models are the MHC weight learning model and the immune tolerance evaluation model, and the output model is the immunogenicity comprehensive evaluation model;

[0015] Based on the immunogenicity prediction intelligent system, the immunogenicity of the protein to be predicted is predicted to obtain a test result, which includes the immunogenicity risk score of the protein to be predicted and the potential immunogenic T cell epitopes on the protein to be predicted.

[0016] In one implementation, training the multimodal prediction model based on each data set to obtain a trained multimodal prediction model includes:

[0017] Training the affinity prediction model based on the immunopeptidomics data set to obtain a trained affinity prediction model;

[0018] Training the presentation prediction model based on the immunopeptidomics dataset to obtain a trained presentation prediction model;

[0019] Training the MHC weight learning model based on the human proteomics and immunology data sets to obtain a trained MHC weight learning model;

[0020] The immune tolerance assessment model is trained based on the human proteomics and immunology data sets to obtain a trained immune tolerance assessment model.

[0021] In one embodiment, the immunopeptidomics dataset includes MHC-peptide affinity data; training the affinity prediction model based on the immunopeptidomics dataset to obtain the trained affinity prediction model includes:

[0022] The input data of the affinity prediction model are set to be peptide sequences and MHC allele subtype sequences, and the output data are high-affinity peptides, wherein the high-affinity peptides are peptides with high affinity to each MHC allele subtype;

[0023] Tagging the MHC-peptide affinity data in the immunopeptidomics dataset, wherein the tag is the affinity of the peptide-MHC pair;

[0024] The labeled MHC-peptide affinity data is input into the affinity prediction model for training to obtain the trained affinity prediction model.

[0025] In one embodiment, the immunopeptidomics data set includes mass spectrometry elution ligand data; training the presentation prediction model based on the immunopeptidomics data set to obtain a trained presentation prediction model includes:

[0026] The input data of the presentation prediction model are set to be peptide sequences and MHC allele subtype sequences, and the output data are high-likelihood peptides, wherein the high-likelihood peptides are peptides with high probability of being naturally processed by antigen-presenting cells and presented by each MHC allele subtype;

[0027] Tagging the mass spectrometry elution ligand data in the immunopeptidomics data set, wherein the tag is the presentation probability of the peptide-MHC pair;

[0028] The labeled mass spectrometry elution ligand data is input into the presentation prediction model for training to obtain the trained presentation prediction model.

[0029] In one embodiment, the human proteomics and immunology data set includes sequence data related to the human immune system; the MHC weight learning model is trained based on the human proteomics and immunology data set to obtain a trained MHC weight learning model including:

[0030] The input data of the MHC weight learning model is set to be the affinity and presentation probability corresponding to each HLA allele subtype i, and the output data is the weight coefficient corresponding to each HLA allele subtype i;

[0031] Using the affinity prediction model and the presentation prediction model, for each human sequence and non-human sequence in the human immune system related sequence data of the human proteomics and immunogenomics data sets, the affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i are predicted respectively;

[0032] The affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i are input into the MHC weight learning model for training to obtain the weight coefficient corresponding to each HLA allele subtype i.

[0033] In one embodiment, the immune tolerance assessment model comprises: human proteome sequence data; the immune tolerance assessment model is trained based on the human proteomics and immunology data sets, and the trained immune tolerance assessment model comprises:

[0034] The input data of the immune tolerance assessment model is set to be a peptide segment, and the output data is set to be the probability that the peptide segment is tolerated by the immune system;

[0035] Segmenting the human proteome sequence data in the human proteomics and immunology data sets to obtain multiple target peptide segments with non-repeating sequences;

[0036] The multiple target peptide segments are input into the immune tolerance assessment model for training to obtain the trained immune tolerance assessment model.

[0037] In one embodiment, the various data sets further include a human biopharmaceutical ADA data set; the method further includes:

[0038] Use the therapeutic monoclonal antibody data in the human biopharmaceutical ADA dataset as a test set;

[0039] Inputting the test set into the immunogenicity prediction intelligent system, performing immunogenicity prediction on relevant protein molecules, and obtaining an immunogenicity risk score;

[0040] The correlation coefficient of clinical ADA level was calculated based on the immunogenicity risk score to test the predictive ability of the immunogenicity prediction intelligent system for immunogenicity risk.

[0041] In a second aspect, the present application also provides an immunogenicity prediction device based on multi-mode fusion, comprising:

[0042] A first construction unit is used to construct a multimodal prediction model, wherein the multimodal prediction model includes an affinity prediction model, a presentation prediction model, an MHC weight learning model and an immune tolerance assessment model, wherein the affinity prediction model is used to predict the affinity of a peptide segment to each MHC allele subtype, the presentation prediction model is used to predict the possibility of a peptide segment being naturally processed by an antigen presenting cell and presented by each MHC allele subtype, the MHC weight learning model is used to learn the frequency weight of HLA, and the immune tolerance assessment model is used to assess the possibility of a peptide segment being tolerated by the immune system;

[0043] A training unit is used to obtain various data sets, wherein the various data sets include immunopeptidomics data sets and human proteomics and immunoomics data sets; and train the multimodal prediction model based on the various data sets to obtain a trained multimodal prediction model;

[0044] A second construction unit is used to fuse the trained multimodal prediction models to construct an immunogenicity comprehensive evaluation model, wherein the immunogenicity comprehensive evaluation model is used to evaluate the complete immunogenicity risk based on the full-length protein sequence based on the output data of the affinity prediction model, the presentation prediction model, the MHC weight learning model and the immune tolerance evaluation model, wherein the input data of the MHC weight learning model are the output data of the affinity prediction model and the output data of the presentation prediction model, and the input data of the immune tolerance evaluation model are the output data of the affinity prediction model and the output data of the presentation prediction model;

[0045] A third construction unit is used to construct an immunogenicity prediction intelligent system based on the trained multimodal prediction model and the immunogenicity comprehensive evaluation model, wherein the input model of the immunogenicity prediction intelligent system is the affinity prediction model and the presentation prediction model, the intermediate model is the MHC weight learning model and the immune tolerance evaluation model, and the output model is the immunogenicity comprehensive evaluation model;

[0046] The prediction unit is used to perform immunogenicity prediction on the protein to be predicted based on the immunogenicity prediction intelligent system to obtain a test result, wherein the test result includes the immunogenicity risk score of the protein to be predicted and the T cell epitopes of potential immunogenicity on the protein to be predicted.

[0047] In one embodiment, when the training unit is used to train the multimodal prediction model based on the respective data sets to obtain a trained multimodal prediction model, it is specifically used to: train the affinity prediction model based on the immunopeptidomics data set to obtain a trained affinity prediction model; train the presentation prediction model based on the immunopeptidomics data set to obtain a trained presentation prediction model; train the MHC weight learning model based on the human proteomics and immunoomics data sets to obtain a trained MHC weight learning model; train the immune tolerance assessment model based on the human proteomics and immunoomics data sets to obtain a trained immune tolerance assessment model.

[0048] In one embodiment, the immunopeptidomics data set includes MHC-peptide affinity data; when the training unit is used to train the affinity prediction model based on the immunopeptidomics data set to obtain the trained affinity prediction model, it is specifically used to: set the input data of the affinity prediction model to be peptide sequences and MHC allele subtype sequences, and the output data to be high-affinity peptides, wherein the high-affinity peptides are peptides with high affinity to each MHC allele subtype; label the MHC-peptide affinity data in the immunopeptidomics data set, wherein the label is the affinity of the peptide-MHC pair; and input the labeled MHC-peptide affinity data into the affinity prediction model for training to obtain the trained affinity prediction model.

[0049] In one embodiment, the immunopeptidomics dataset includes mass spectrometry elution ligand data; when the training unit is used to train the presentation prediction model based on the immunopeptidomics dataset to obtain a trained presentation prediction model, it is specifically used to: set the input data of the presentation prediction model to be peptide sequences and MHC allele subtype sequences, and the output data to be high-probability peptides, wherein the high-probability peptides are peptides with a high probability of being naturally processed by antigen-presenting cells and presented by each MHC allele subtype; label the mass spectrometry elution ligand data in the immunopeptidomics dataset, wherein the label is the presentation probability of the peptide-MHC pair; and input the labeled mass spectrometry elution ligand data into the presentation prediction model for training to obtain the trained presentation prediction model.

[0050] In one embodiment, the human proteomics and immunology data set includes sequence data related to the human immune system; when the training unit is used to train the MHC weight learning model based on the human proteomics and immunology data set to obtain a trained MHC weight learning model, it is specifically used to: set the input data of the MHC weight learning model to be the affinity and presentation probability corresponding to each HLA allele subtype i, and the output data to be the weight coefficient corresponding to each HLA allele subtype i; use the affinity prediction model and the presentation prediction model to predict the affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i for each human sequence and non-human sequence in the human immune system related sequence data of the human proteomics and immunology data set; input the affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i into the MHC weight learning model for training to obtain the weight coefficient corresponding to each HLA allele subtype i.

[0051] In one embodiment, the immune tolerance assessment model includes: human proteome sequence data; when the training unit is used to train the immune tolerance assessment model based on the human proteomics and immunomics data sets to obtain the trained immune tolerance assessment model, it is specifically used to: set the input data of the immune tolerance assessment model to peptide segments, and the output data to be the probability that the peptide segments are tolerated by the immune system; split the human proteome sequence data in the human proteomics and immunomics data sets to obtain multiple target peptide segments with non-repetitive sequences; input the multiple target peptide segments into the immune tolerance assessment model for training to obtain the trained immune tolerance assessment model.

[0052] In one embodiment, the various data sets also include a human biological drug ADA data set; the immunogenicity prediction device based on multimodal fusion also includes a testing unit, which is used to: use the therapeutic monoclonal antibody data in the human biological drug ADA data set as a test set; input the test set into the immunogenicity prediction intelligent system, perform immunogenicity prediction on related protein molecules, and obtain an immunogenicity risk score; calculate the correlation coefficient of the clinical ADA level based on the immunogenicity risk score to test the immunogenicity prediction intelligent system's ability to predict immunogenicity risk.

[0053] In a third aspect, an embodiment of the present application further provides a computer device, comprising: a memory and a processor, wherein the memory stores instructions, and the instructions are loaded and executed by the processor to implement a method in any one of the above-mentioned aspects, wherein the memory and the processor communicate with each other through an internal connection path.

[0054] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, in which a computer program is stored. When the computer program runs on a computer, the method in any one of the above-mentioned embodiments is implemented.

[0055] The advantages or beneficial effects of the above technical solution include at least:

[0056] The present application can organically integrate multiple prediction models related to immunogenicity to form an efficient immune response simulation system, namely, an immunogenicity prediction intelligent system, which can accurately simulate and predict the strength of the immunogenicity of proteins, especially at the level of clinical immune response caused by the human body, and can identify T cell epitopes with potential immunogenicity, so as to achieve the effect of targeted optimization of immunogenicity based on the prediction results, and can be applied to the prediction of immunogenicity of proteins and their derived molecular forms, including but not limited to vaccines, antigens, antibodies, nanobodies, peptides, antibody conjugates (ADCs), polypeptide conjugates (PDCs), nucleic acids that can be translated into proteins, and other protein biological drugs, thereby improving the efficacy and safety of protein biological drugs.

[0057] The above summary is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments and features described above, further aspects, embodiments and features of the present application will be readily apparent by reference to the accompanying drawings and the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In the accompanying drawings, unless otherwise specified, the same reference numerals throughout the multiple drawings represent the same or similar parts or elements. These drawings are not necessarily drawn to scale. It should be understood that these drawings only depict some embodiments disclosed in the present application and should not be regarded as limiting the scope of the present application.

[0059] Figure 1 A flowchart of an immunogenicity prediction method based on multi-modal fusion provided in an embodiment of the present application;

[0060] Figure 2 An example diagram of the architecture of an immunogenicity prediction intelligent system provided in an embodiment of the present application;

[0061] Figure 3 An example diagram of the distribution of clinical ADA incidence of a therapeutic antibody tested in an embodiment of the present application;

[0062] Figure 4 An example diagram of the distribution of the immunogenicity risk score predicted by an immunogenicity prediction intelligent system and the average value of the clinical ADA incidence rate provided in an embodiment of the present application;

[0063] Figure 5 An example diagram comparing the correlation performance of an immunogenicity prediction intelligent system prediction and the EpiMatrix method provided in an embodiment of the present application;

[0064] Figure 6An example diagram comparing an in vitro experiment provided in an embodiment of the present application with an epitope predicted by an intelligent immunogenicity prediction system;

[0065] Figure 7 A structural block diagram of an immunogenicity prediction device based on multi-mode fusion provided in an embodiment of the present application;

[0066] Figure 8 A structural block diagram of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0067] In the following, only some exemplary embodiments are briefly described. As those skilled in the art will appreciate, the described embodiments may be modified in various ways without departing from the spirit or scope of the present application. Therefore, the drawings and descriptions are considered to be exemplary and non-restrictive in nature.

[0068] One of the main means of targeted optimization of immunogenicity is to adjust the immunogenicity of drugs by modifying the potential T cell epitopes of drug molecules. Therefore, accurate identification of the peptides corresponding to potential T cell epitopes in drug molecules is a prerequisite for targeted optimization of immunogenicity, which can assist in the design of antibody drugs, recombinant proteins, and vaccines.

[0069] In the related art, T cell epitopes can be identified and verified by means of wet experiments such as MHC binding, peripheral blood mononuclear cell (PBMC) activation, and cytokine release, but the efficiency is low.

[0070] Due to the different treatment mechanisms, the optimization direction of antibody drugs is usually to reduce immunogenicity, while the optimization direction of vaccine drugs is usually to increase immunogenicity. In addition, using computational methods for prediction is a more efficient method, which can also be used in combination with experimental methods. Based on this, in the process of antibody and recombinant protein drug development, T cell antigen epitopes on antibodies and recombinant protein drugs can be predicted, and some or all epitopes can be removed by introducing mutations to reduce the probability of ADA production, so as to avoid the metabolism and efficacy of drugs being affected by ADA.

[0071] In summary, in order to improve the efficacy and safety of therapeutic or preventive drugs, the present application proposes relevant technical solutions that can predict protein immunogenicity and identify potential T cell epitopes on proteins, so as to predict or regulate the immune response against the drug itself produced by biological drugs such as antibodies, fusion proteins, peptides, vaccines, etc. after entering the human body, such as predicting the incidence of anti-drug antibodies, the titer of specific antibodies induced by vaccines, or tumor neoantigens, etc., to ultimately guide or accelerate the research and development of therapeutic or preventive drugs.

[0072] The technical solution provided in the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0073] Figure 1 FIG. 4 is a flow chart showing a method for predicting immunogenicity based on multi-modal fusion according to an embodiment of the present application. Figure 1 As shown, the method may include the following steps:

[0074] S110. Construct a multimodal prediction model.

[0075] In one embodiment, the multimodal prediction model may include, but is not limited to: an affinity prediction model, a presentation prediction model, an MHC weight learning model, and an immune tolerance assessment model.

[0076] In specific implementation, the affinity prediction model can be used to predict the affinity of a peptide segment to each MHC allele subtype. The presentation prediction model can be used to predict the possibility of a peptide segment being naturally processed (including enzyme cleavage) by an APC (i.e., antigen presenting cell) and presented by each MHC allele subtype. The MHC weight learning model can be used to learn the frequency weights of HLA. The immune tolerance assessment model is used to assess the possibility of a peptide segment being tolerated by the immune system.

[0077] In an embodiment of the present application, by executing step S110, multiple prediction models related to immunogenicity can be constructed, which facilitates the subsequent organic fusion of these multiple prediction models to form an efficient immune response simulation system for accurately simulating the immunogenicity of biological drugs such as proteins, antibodies, peptides, and vaccines, and can identify T cell epitopes with potential immunogenicity (peptides that induce clinical human immune responses).

[0078] S120: Acquire each data set.

[0079] In one embodiment, each data set may include, but is not limited to: an immunopeptidomics data set, a human proteomics and immunology data set, and a human biopharmaceutical ADA data set.

[0080] In specific implementation, large-scale high-quality immunopeptidome data can be collected, covering major MHC loci and subtypes, to obtain the immunopeptidome data set. The immunopeptidome data may include, but are not limited to: MHC-peptide affinity data (i.e., affinity data (BA) of peptide binding to MHC II) and mass spectrometry (MS) elution ligand data (EL).

[0081] Exemplarily, MHC-peptide affinity data and mass spectrometry elution ligand data can be downloaded from the IEDB immune epitope database to obtain the immunopeptidomics data set, wherein the MHC-peptide affinity data may include about 130,000 data points, which may be used to determine specific binding affinity values ​​through IC50 test experiments; and the mass spectrometry elution ligand data may include about 7.6 million data points.

[0082] In the examples of the present application, by collecting immunopeptidomics data on a large scale and with high quality, it is possible to facilitate the subsequent use of the immunopeptidomics dataset for MHC level prediction.

[0083] In specific implementation, large-scale high-quality human proteomics and immunology data can be collected to obtain the human proteomics and immunology data set, wherein the human proteomics and immunology data may include but are not limited to: B cell receptor (BCR) sequencing data, human immune system related sequence data and human proteome sequence data.

[0084] Exemplarily, B cell receptor (BCR) sequencing data can be downloaded from the OAS database, human immune system related sequence data can be downloaded from the IMGT database, and human proteome sequence data can be downloaded from the NextProt database to obtain the human proteomics and immunology data sets. Among them, the B cell receptor (BCR) sequencing data comes from more than 75 sequencing studies, including more than 1 billion BCR sequences of humans and mice; the human immune system related sequence data can include but are not limited to: human immunoglobulin IgG, T cell receptor TCR, histocompatibility complex MHC.

[0085] In the embodiments of the present application, by collecting large-scale, high-quality human proteomics and immunomics data, it is possible to facilitate the subsequent use of human proteomics and immunomics data sets for immune system level prediction.

[0086] In specific implementation, high-quality human biological drug ADA data can be collected to obtain the human biological drug ADA data set.

[0087] Exemplarily, by searching and collecting clinical ADA incidence data from the label instructions of marketed biological drugs published on the FDA website and the EMA website, and searching the complete sequence of the corresponding drug from the INN website, and searching for immunogenicity-related reports from relevant literature and relevant patent documents using keywords such as Immunogenicity, immunogenicity, and ADA, extracting the drug sequence and ADA incidence data, human biological drug ADA data can be collected. In this way, hundreds of high-quality human biological drug clinical ADA real data (including single / multi-specific antibodies and recombinant proteins, etc.) from the FDA website and the EMA website can be collected.

[0088] In the embodiments of the present application, by collecting high-quality human biological drug ADA data, it is possible to facilitate the subsequent use of the human biological drug ADA data set for immune clinical level prediction.

[0089] It can be understood that each data set is obtained by dividing the MHC-peptide affinity data, mass spectrometry elution ligand data, human immune system related sequence data, B cell receptor sequencing data, human proteome sequence data, and clinical stage ADA data according to three different levels: MHC level, immune system level, and immune clinical level.

[0090] S130 , training a multimodal prediction model based on each data set to obtain a trained multimodal prediction model.

[0091] In one implementation, the implementation process of step S130 may include the following steps:

[0092] S131. Training an affinity prediction model based on the immunopeptidomics dataset to obtain a trained affinity prediction model.

[0093] In specific implementation, the input data of the affinity prediction model can be set as peptide sequence and MHC allele subtype sequence, and the output data can be set as high-affinity peptides. The high-affinity peptides are peptides with high affinity to each MHC allele subtype.

[0094] Exemplarily, the length of the peptide sequence can be set to 9-16 amino acids. The MHC allele subtype sequence can be the amino acid sequence of the MHC molecule. The sequence features of the peptide sequence and the sequence features of the MHC allele subtype sequence can be extracted by one-hot encoding or embedding in an existing protein language model.

[0095] In the examples of the present application, by using both peptide sequences and MHC allele subtype sequences as input data for the affinity prediction model, the affinity prediction model can be made pan-MHC allele subtype specific, and can simultaneously integrate information from multiple alleles, thereby improving the prediction performance, and can not only predict alleles of all known amino acid sequences, but also extrapolate the prediction results to previously uncharacterized MHC molecules. That is, the affinity prediction model in the examples of the present application has pan-specificity.

[0096] In a specific implementation, the MHC-peptide affinity data in the immunopeptidomics data set can be labeled, and the label is the affinity of the peptide-MHC pair.

[0097] In specific implementation, the labeled MHC-peptide affinity data can be input into the affinity prediction model for training to obtain a trained affinity prediction model.

[0098] As an example, in the process of training the affinity prediction model, 80% of the data representative of each HLA allele subtype can be selected from the labeled MHC-peptide affinity data as the training set, and the remaining 20% ​​of the data can be used as the test set. Then, a variety of existing deep learning and machine learning techniques are used to evaluate the prediction performance of the affinity prediction model using five-fold cross-validation of the training set and the test set. For example, in order to achieve the best prediction effect, the same data set can be used to train and test multiple models such as Transformer, LSTM, CNN, XGBOOST, etc., and the prediction performance of each model in each representative HLA allele subtype is compared, and the one with better prediction performance is used as the trained affinity prediction model.

[0099] In the embodiment of the present application, by executing step S131, the affinity of the corresponding peptide segment to each MHC allele subtype can be predicted by the affinity prediction model, and the high-affinity peptide segment can be output as a potential immunogenic T cell epitope.

[0100] S132. Training a presentation prediction model based on the immunopeptidomics dataset to obtain a trained presentation prediction model.

[0101] In specific implementation, the input data of the presentation prediction model can be set as peptide sequence and MHC allele subtype sequence, and the output data can be high-likelihood peptides. The high-likelihood peptides can be peptides with high probability of being naturally processed by APC and presented by each MHC allele subtype.

[0102] In the embodiments of the present application, by taking the peptide sequence and the MHC allele subtype sequence as the input data of the presentation prediction model, the presentation prediction model can predict the possibility of being naturally processed by APC and presented by each MHC allele subtype based on all possible peptides contained in the input amino acid sequence, so that high-probability peptides can be used as potential immunogenic T cell epitopes.

[0103] In a specific implementation, the mass spectrometry elution ligand data in the immunopeptidomics data set can be labeled, and the label is the presentation probability of the peptide-MHC pair.

[0104] In specific implementation, the labeled mass spectrometry elution ligand data can be input into the presentation prediction model for training to obtain a trained presentation prediction model.

[0105] Among them, in the process of training the presentation prediction model, the mass spectrometry elution ligand data can be split in the same or similar splitting method as the above-mentioned MHC-peptide affinity data to train and test the presentation prediction model. Different from the training process of the above-mentioned affinity prediction model, due to the use of mass spectrometry elution ligand data from different cell lines, the multi-allelic spectrum will lead to ambiguity in the exact binding of the peptide to the MHC. Therefore, in the training process of the presentation prediction model, the embodiment of the present application uses deconvolution technology to perform data enhancement, that is, a pre-model is trained using single-allelic data, and then the pre-model is used to predict and mark multi-allelic data. The specific method is: for each data point corresponding to the multi-allelic gene, the pre-model is used to predict the presentation probability of the peptide segment and each possible allele, and then the allele with the highest predicted probability is set as the MHC corresponding to the peptide segment, and used for the final model training.

[0106] In this embodiment of the present application, by executing step S132, the presentation prediction model can be used to predict the likelihood of the corresponding peptide being naturally processed by APC and presented by each MHC allele subtype, and the high-likelihood peptide is output as a potential immunogenic T cell epitope.

[0107] It is understandable that the above-mentioned high affinity prediction model and the above-mentioned presentation prediction model can be used together for MHC level prediction, and the prediction methods of the two are referred to as AlphaMHC hereinafter.

[0108] By using the same data, such as the data in the literature (You, R., (2022). DeepMHCII: a novel binding core-aware deep interaction model for accurate MHC-II peptide binding affinity prediction. Bioinformatics, 30, i220-i228), AlphaMHC, the NetMHCIIpan method widely used for MHC affinity prediction, and the DeepMHCII method based on deep learning are verified, and the comparison results shown in Table 1 can be obtained. As shown in Table 1, AlphaMHC shows the best affinity prediction ability and presentation probability prediction ability compared with the NetMHCIIpan method, DeepMHCII method and other methods.

[0109] Table 1 Performance comparison of AlphaMHC and similar methods

[0110]

[0111] In Table 1, ROC-AUC is the area under the receiver operating curve, and PCC is the Pearson correlation coefficient. The larger the value, the better, and the maximum value is 1; "-" means that this function is not available.

[0112] In the embodiments of the present application, through the above-mentioned high-affinity prediction model and the above-mentioned presentation prediction model, thousands of MHC allele subtypes can be trained. With the support of parallel computing technology, each allele subtype can be calculated simultaneously in a high-throughput manner, so that on the basis of MHC level prediction, the natural immune system physiological mechanisms such as MHC polymorphism and immune tolerance can be further simulated, thereby simulating the final anti-drug immune response formed by the immune system level of a wide range of people to drugs.

[0113] That is, in the present embodiment, the prediction at the MHC level is mainly to predict the interaction between the peptide segment and each MHC allele subtype, which may include affinity, APC processing (such as enzyme cleavage) and presentation probability.

[0114] S133. The MHC weight learning model is trained based on human proteomics and immunogenomics datasets to obtain a trained MHC weight learning model.

[0115] In specific implementation, the input data of the MHC weight learning model can be set as the affinity and presentation probability corresponding to each HLA allele subtype i, and the output data can be set as the weight coefficient corresponding to each HLA allele subtype i; then, the affinity prediction model and the presentation prediction model are used to predict the affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i for each human sequence and non-human sequence in the human immune system related sequence data of the human proteomics and immunogenomics data sets; then, the affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i are input into the MHC weight learning model for training to obtain the weight coefficient corresponding to each HLA allele subtype i. Among them, HLA alleles can include but are not limited to: class I HLA alleles and class II HLA alleles, such as HLA-A, HLA-B, HLA-C, HLA-E, HLA-F, HLA-G, HLA-DQ, HLA-DR, HLA-DP, etc.

[0116] For example, at the level of the human immune system, based on the statistical analysis of HLA alleles in the population in public studies, high-frequency representative type II HLA allele subtypes can be selected as each HLA allele subtype i. For example, the following type II HLA allele subtypes can be selected as each HLA allele subtype i, and can be increased or decreased as appropriate in the subsequent modeling and application process: DRB1*01:01, DRB1*03:01, DRB1*04:01, DRB1*04:04, DRB1*04:05, DRB1*07:01, DRB1*08:02, DRB1*09:01, DRB1*11:01, DRB1*12:01, DRB1*13:01, DRB1*13:02, DRB1*13:02, DRB1*15:01, DRB3*01:01, DRB3*02:02, DRB4*01:01, DRB5*01:01, DPA1*01:03-DPB1*02:01, DPA1*01:03-DPB1*04:01, DPA1*02:01-DPB1*01:01 ,DPA1*02:01-DPB1*04:01,DPA1*02:01-DPB1*05:01,DPA1*02:01-DPB1*14:01 ,DPA1*02:02-DPB1*04:01, DPA1*03:01-DPB1*04:01, DPA1*03:01-DPB1*04:02 , DQA1*01:01-DQB1*05:01, DQA1*01:02-DQB1*06:02, DQA1*03:01-DQB1*03:02, DQA1*04:01-DQB1*04:02, DQA1*05:01-DQB1*02:01, DQA1*05:01-DQB1*03:01. A complete list of these HLA class allele subtypes can be found on existing public websites / public literature.

[0117] Similarly, in non-human immune systems, such as mice or rats, representative subtypes specific to other species (such as cats, dogs, pigs, cows, sheep, alpacas, camels, primates, etc.) can be selected.

[0118] Exemplarily, in the process of training the MHC weight learning model, a series of machine learning tasks related to the overall performance of the immune system can be set according to the sequence data related to the human immune system, and the model outputs of each representative HLA allele subtype or associated with the HLA allele subtype can be used as features. After learning the tasks related to the immune system, combined with the diversity of HLA sequences and the population distribution frequency of HLA allele subtypes worldwide, the MHC weight learning model can assign different weight coefficients to each selected HLA allele subtype, which is conducive to simulating the HLA frequency distribution of the real immune system of a wide range of people worldwide and reducing redundancy.

[0119] For example, the training process of the MHC weight learning model can be: using the aforementioned human immune system related sequence data, respectively calculating the MHC interactions between the human and non-human sequences and each representative HLA allele subtype, constructing a sequence-based species identification model, and then using existing feature attribution or back propagation techniques to obtain the weight coefficient corresponding to each HLA allele subtype. The specific training process can be as follows:

[0120] Step 1: Using the above affinity prediction model and the above presentation prediction model, for each human sequence and non-human sequence in the human immune system related sequence data, respectively predict the affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i as input features of the MHC weight learning model.

[0121] Step 2: Use existing shallow neural networks, gbdt models such as XGBOOST or GBM, multiple linear regression, etc. to train the above affinity prediction model and the above presentation prediction model, and use the human source possibility as the output Y of the above affinity prediction model and the above presentation prediction model, that is, in the form of Y = f(Bi, Ei), i∈[1,32].

[0122] Step 3: Based on different models, the weight coefficients of each MHC subtype can be calculated using methods such as neural network node weight accumulation (shallow neural network), feature importance analysis (XGBOOST) and SHAP (SHapley Additive exPlanations).

[0123] Step 4: Record the weight coefficient combinations of different methods and select the best set of weight coefficient combinations for subsequent immunogenicity prediction.

[0124] In the embodiment of the present application, by executing step S133, the frequency weight of HLA can be learned by using immunogenicity-related tasks and feature selection technology based on the high-frequency representative HLA allele subtypes widely present in the collected population.

[0125] S134. Train the immune tolerance assessment model based on human proteomics and immunogenomics datasets to obtain a trained immune tolerance assessment model.

[0126] In specific implementation, the input data of the immune tolerance assessment model can be set as peptides, and the output data can be set as the probability that the peptides are tolerated by the immune system; then the human proteome sequence data in the human proteomics and immunoomics data sets are segmented to obtain multiple target peptides with non-repeating sequences; finally, the multiple target peptides are input into the immune tolerance assessment model for training to obtain a trained immune tolerance assessment model.

[0127] Exemplarily, during the training process of the immune tolerance assessment model, the collected human proteome sequence data can be segmented to obtain more than 140 million non-repeating target peptides (the length can be set as needed, such as 9 peptides) from more than 120 million sequences, and then the occurrence frequency of these target peptides in the population (range 0-100%) is summarized (repeated peptides) and labeled, and the immune tolerance assessment model is trained in the same way as the affinity prediction model.

[0128] In specific implementation, the immune tolerance assessment model can predict the probability of any target peptide being tolerated by the human immune system and provide the source of the most similar human peptide. After testing, the PCC (Pearson correlation coefficient) of the immune tolerance assessment model for unknown fragments is greater than 0.7.

[0129] In this embodiment of the present application, by executing step S134, the protein sequence data that is commonly present in the human population and is generally tolerated by the immune system can be used as a background reference to evaluate the possibility that the target peptide segment is tolerated by the immune system (does not activate the immune response).

[0130] It can be understood that in terms of prediction of the immune system level, the key prediction mechanism in the embodiments of the present application is to simulate through the MHC weight learning model and the immune tolerance assessment model, both of which are constructed separately using data at the immune system level.

[0131] S140. Construct a comprehensive immunogenicity evaluation model by integrating the trained multimodal prediction models.

[0132] In one embodiment, the comprehensive immunogenicity assessment model is used to assess the complete immunogenicity risk based on the full-length protein sequence based on the output data of the affinity prediction model, the presentation prediction model, the MHC weight learning model and the immune tolerance assessment model, wherein the input data of the MHC weight learning model are the output data of the affinity prediction model and the output data of the presentation prediction model, and the input data of the immune tolerance assessment model are the output data of the affinity prediction model and the output data of the presentation prediction model.

[0133] It can be understood that the comprehensive immunogenicity assessment model summarizes the prediction results of the above-mentioned prediction models based on peptide segments and then conducts a comprehensive assessment to provide a complete immunogenicity risk assessment based on the full-length protein sequence (including monoclonal antibodies, multispecific antibodies, etc.). During this assessment process, if the target protein consists of multiple chains, each chain is calculated in parallel and then summarized.

[0134] In specific implementation, the evaluation process of the comprehensive immunogenicity evaluation model can be as follows:

[0135] Step 1: After the complete sequence of each chain is cut from beginning to end into overlapping peptide segments of a specific length (e.g., 15 amino acids), the following steps are performed in sequence:

[0136] Step 2: Input each peptide segment into the affinity prediction model and the presentation prediction model to simulate the interaction at the MHC level, and output the affinity and presentation probability of the peptide segment for each MHC allele subtype;

[0137] Step 3: Submit the predicted MHC high-affinity fragments or high-presentation probability fragments to the MHC weight learning model and immune tolerance assessment model to simulate the immune system level, and output the interaction weights with each representative HLA allele subtype and the possibility of being tolerated by the immune system;

[0138] Step 4: Record the high-scoring fragments (potentially immunogenic T cell epitopes) and their risk scores after the MHC weight learning model and immune tolerance assessment model are combined and weighted;

[0139] Exemplarily, the risk score of a high-scoring segment may be calculated as follows:

[0140] (Bi+Ei)*Wi (i.e. HLA weight coefficient)*Hi (i.e. the probability of being tolerated by the human immune system), and then summing up the values ​​of each i to get the total risk score of the high-scoring fragment.

[0141] Step 5: Calculate the final immunogenicity risk of the target sequence based on the scores of all potentially immunogenic T cell epitopes. For molecules with multiple chains, further combine the results of each chain.

[0142] For example, when a molecule has three chains, and each chain has 1, 2, and 3 epitopes, the final immunogenicity risk is the sum of the scores of the 6 epitopes. Of course, in practical applications, the scores of each fragment can also be converted into probability values ​​(between 0 and 1), and then 1 minus the probability value of each fragment not occurring is calculated. For example, if there are 2 epitopes in total, with probabilities of 0.3 and 0.4 respectively, the final total probability score is 1-0.7*0.6=0.58, to further improve the scientificity of the calculation.

[0143] It can be understood that in terms of clinical level prediction, the key prediction mechanism in the embodiments of the present application is achieved through a comprehensive immunogenicity evaluation model.

[0144] In an embodiment of the present application, by executing step S140, the outputs of the above-mentioned affinity prediction model, presentation prediction model, MHC weight learning model, and immune tolerance assessment model can be integrated through the comprehensive immunogenicity assessment model to give a final quantitative assessment of the probability and intensity of the final immune response, expressed as a score, and corresponding to the distribution of clinical ADA data, give a high or low ADA risk rating, and output the corresponding T cell epitope.

[0145] S150, based on the trained multimodal prediction model and immunogenicity comprehensive evaluation model, constitute an immunogenicity prediction intelligent system.

[0146] In one embodiment, as shown in Figure 2, the input models of the immunogenicity prediction intelligent system are the affinity prediction model and the presentation prediction model, the intermediate models are the MHC weight learning model and the immune tolerance assessment model, and the output model is the immunogenicity comprehensive assessment model.

[0147] In the implementation of this application, by executing step S150, an immunogenicity prediction intelligent system can be formed by organically combining the five models of affinity prediction model, presentation prediction model, MHC weight learning model, immune tolerance assessment model and immunogenicity comprehensive assessment model. The immunogenicity prediction intelligent system can have end-to-end immunogenicity risk prediction capabilities from sequence to clinic, and the prediction results are generally applicable and representative for most people in the world. In addition, in the immunogenicity prediction intelligent system, each model adopts artificial intelligence technologies such as deep neural network structure, which can effectively extract the characteristics of input information, including many undiscovered hidden feature laws, and can provide high-precision prediction results, and use the self-attention mechanism to provide a certain degree of interpretability, which can not only quantitatively evaluate the overall immunogenicity level of the sequence, but also locate the potential T cell epitopes that produce immunogenicity. At the same time, due to the use of a data set of non-single MHC alleles, the prediction accuracy and scope of application of the immunogenicity prediction intelligent system can be greatly improved.

[0148] S160. Based on the immunogenicity prediction intelligent system, the immunogenicity of the predicted protein is predicted to obtain the test result.

[0149] In one embodiment, before executing step S160, since immunogenicity is clinically manifested as ADA, the correlation between the prediction score and ADA can be used to test the predictive ability of the immunogenicity prediction intelligent system for immunogenicity risk.

[0150] For example, the therapeutic monoclonal antibody data in the human biological drug ADA data set can be used as a test set (as shown in the figure below), and the test set can be input into the immunogenicity prediction intelligent system to predict the immunogenicity of related protein molecules and obtain the immunogenicity risk score; then the correlation coefficient of the clinical ADA level is calculated based on the immunogenicity risk score to test the predictive ability of the immunogenicity prediction intelligent system for immunogenicity risk. Among them, the correlation coefficient can be represented by the Pearson correlation coefficient.

[0151] For example, the correlation between the predicted immunogenicity risk score and the actual clinical ADA incidence can be as follows: Figure 3 and Figure 4 As shown, Figure 3 The distribution of clinical ADA incidence of the tested therapeutic antibodies, with the horizontal axis representing the interval of ADA incidence (in %), the vertical axis representing the number of molecules in this interval, and the risk rating, with red representing high risk (ADA>=20%), orange representing medium risk (5%<=ADA<20%), and green representing low risk (ADA<5%). Figure 4 In the figure, MolScore is the immunogenicity risk score predicted by the immunogenicity prediction intelligent system, and ADA Frequency is the average clinical ADA incidence rate. That is, the immunogenicity prediction intelligent system successfully achieved end-to-end prediction from sequence to clinical immunogenicity risk, and the Pearson correlation coefficient R between the score and the ADA incidence rate reached 0.68, showing good prediction performance. In particular, when 0.2 is used as the threshold of the prediction score, high-risk and low-risk clinical antibody molecules can be effectively distinguished.

[0152] In specific implementation, in order to further test the predictive ability of the immunogenicity prediction intelligent system for immunogenicity risk, the immunogenicity prediction intelligent system can also be compared with the SOTA method EpiMatrix, using another data set consisting of 42 antibody molecules published by EpiVax in 2024. The prediction score of EpiMatrix comes from the literature published by EpiVax (Mattei, AE et al. (2024) 'In silico methods for immunogenicity risk assessment and human homology screening for therapeutic antibodies', mAbs, 16 (1)).

[0153] like Figure 5As shown in the figure, after comparison, the correlation performance of the test immunogenicity prediction intelligent system exceeds that of the well-known EpiMatrix method (R2=0.49 vs R2=0.42, R is the Pearson correlation coefficient between the prediction score and the clinically observed ADA incidence rate, the larger the better). In addition, the score of the test immunogenicity prediction intelligent system has the same scale and range as the ADA incidence rate (both 0-1), which is easier to understand and apply. Among them, Figure 5 Each data point in the figure corresponds to the data of a therapeutic antibody. The horizontal axis represents the clinically observed ADA incidence (what proportion of patients who received the drug produced ADA), and the vertical axis represents the predicted immunogenicity risk score. The left picture shows the results of EpiMatrix, and the right picture shows the results of the immunogenicity prediction intelligent system. Among them, the EpiMatrix data, the software version is V2.0 and the Treg correction parameter is turned on (its official recommended setting for best performance).

[0154] From the above comparison results, it can be seen that the immunogenicity prediction intelligent system in the embodiment of the present application can effectively predict the overall immunogenicity risk.

[0155] In one embodiment, the test results may include, but are not limited to: the immunogenicity risk score of the protein to be predicted and the potentially immunogenic T cell epitopes on the protein to be predicted.

[0156] In the specific implementation, in addition to the overall prediction of immunogenicity risk at the molecular level, in actual applications, in order to further carry out de-immunization or enhance immunogenicity design, it is also necessary to accurately predict the main T cell epitopes in the molecule that cause immunogenicity.

[0157] For example, Biogen's Natalizumab is a humanized monoclonal IgG4 antibody targeting α4 integrin. The incidence of neutralizing antibodies in clinical trials is about 6%, and a single T cell epitope is used to induce the production of multiple sclerosis patients. According to literature reports, the immunogenicity of Natalizumab is mainly caused by a single immunogenic T cell epitope spanning the FR2-CDR2 region of its light chain. The peptide "RLLIHYTSALQPGI" located in this epitope can activate T cells and proliferate them in in vitro experiments, and is an experimentally verified T cell epitope.

[0158] The sequence of Natalizumab was input into the immunogenicity prediction intelligent system for prediction. It scanned the full-length sequences of the heavy chain and light chain of the antibody and its five immune variants, and identified multiple potential T cell epitopes. The epitope with the highest score was the potential immunogenic T cell epitope located in the FR2-CDR2 region of the light chain. The corresponding peptide segment was "PRLLIHYTSALQPGIPS", and most of the amino acids were consistent with the T cell epitopes detected in vitro. The in vitro experimental and predicted epitopes can be shown as follows. Figure 6 At the same time, in order to further verify the immunogenicity of the epitope, the epitope was humanized to generate de-immunogenic mutants (NZM var1-5), which no longer induced immune responses in vitro, and the results predicted by the immunogenicity prediction intelligent system were consistent with this.

[0159] in, Figure 6 In the figure, Unstim is a negative control without immune activation, NZM is Natalizumab monoclonal antibody; var1-5 is a de-immunogenic mutant of NZM.

[0160] That is, in the embodiment of the present application, the immunogenicity of the protein to be predicted is predicted based on the immunogenicity prediction intelligent system, which can accurately simulate the immunogenicity of the protein to be predicted, obtain its immunogenicity risk score, and identify T cell epitopes with potential immunogenicity, so that the immunogenicity can be optimized in a targeted manner based on the prediction results.

[0161] In summary, the immunogenicity prediction method based on multi-mode fusion provided in the embodiment of the present application can organically integrate multiple prediction models related to immunogenicity to form an efficient immune response simulation system, that is, an immunogenicity prediction intelligent system, which can accurately simulate and predict the immunogenicity of proteins, especially at the level of clinical human immune response, and can identify T cell epitopes with potential immunogenicity, so as to achieve the effect of targeted optimization of immunogenicity based on the prediction results, and can be applied to the prediction of immunogenicity of proteins and their derived molecular forms, including but not limited to vaccines, antigens, antibodies, nanobodies, polypeptides, antibody conjugates (ADCs), polypeptide conjugates (PDCs), nucleic acids that can be translated into proteins, etc. The immunogenicity prediction of protein biological drugs can be improved, thereby improving the efficacy and safety of protein biological drugs.

[0162] Figure 7 FIG. 2 shows a structural block diagram of an immunogenicity prediction device based on multi-mode fusion according to an embodiment of the present application. Figure 7 As shown, the device may include:

[0163] A first construction unit 210 is used to construct a multimodal prediction model, which includes an affinity prediction model, a presentation prediction model, an MHC weight learning model and an immune tolerance assessment model. The affinity prediction model is used to predict the affinity of a peptide segment to each MHC allele subtype, the presentation prediction model is used to predict the possibility of a peptide segment being naturally processed by an antigen presenting cell and presented by each MHC allele subtype, the MHC weight learning model is used to learn the frequency weight of HLA, and the immune tolerance assessment model is used to assess the possibility of a peptide segment being tolerated by the immune system;

[0164] The training unit 220 is used to obtain various data sets, including immunopeptidomics data sets and human proteomics and immunoomics data sets; train a multimodal prediction model based on each data set to obtain a trained multimodal prediction model;

[0165] The second construction unit 230 is used to fuse the trained multimodal prediction models to construct an immunogenicity comprehensive evaluation model, and the immunogenicity comprehensive evaluation model is used to evaluate the complete immunogenicity risk based on the full-length protein sequence based on the output data of the affinity prediction model, the presentation prediction model, the MHC weight learning model and the immune tolerance evaluation model, wherein the input data of the MHC weight learning model are the output data of the affinity prediction model and the output data of the presentation prediction model, and the input data of the immune tolerance evaluation model are the output data of the affinity prediction model and the output data of the presentation prediction model;

[0166] The third construction unit 240 is used to construct an immunogenicity prediction intelligent system based on the trained multimodal prediction model and the immunogenicity comprehensive evaluation model, wherein the input models of the immunogenicity prediction intelligent system are the affinity prediction model and the presentation prediction model, the intermediate models are the MHC weight learning model and the immune tolerance evaluation model, and the output model is the immunogenicity comprehensive evaluation model;

[0167] The prediction unit 250 is used to perform immunogenicity prediction on the protein to be predicted based on the immunogenicity prediction intelligent system to obtain a test result, which includes the immunogenicity risk score of the protein to be predicted and the potential immunogenic T cell epitopes on the protein to be predicted.

[0168] In one embodiment, when the training unit 220 is used to train a multimodal prediction model based on each data set to obtain a trained multimodal prediction model, it is specifically used to: train an affinity prediction model based on the immunopeptidomics data set to obtain a trained affinity prediction model; train a presentation prediction model based on the immunopeptidomics data set to obtain a trained presentation prediction model; train an MHC weight learning model based on human proteomics and immunoomics data sets to obtain a trained MHC weight learning model; train an immune tolerance assessment model based on human proteomics and immunoomics data sets to obtain a trained immune tolerance assessment model.

[0169] In one embodiment, the immunopeptidomics data set includes MHC-peptide affinity data; when the training unit 220 is used to train the affinity prediction model based on the immunopeptidomics data set to obtain the trained affinity prediction model, it is specifically used to: set the input data of the affinity prediction model to be peptide sequences and MHC allele subtype sequences, and the output data to be high-affinity peptides, where the high-affinity peptides are peptides with high affinity to each MHC allele subtype; label the MHC-peptide affinity data in the immunopeptidomics data set, where the label is the affinity of the peptide-MHC pair; and input the labeled MHC-peptide affinity data into the affinity prediction model for training to obtain the trained affinity prediction model.

[0170] In one embodiment, the immunopeptidomics data set includes mass spectrometry elution ligand data; when the training unit 220 is used to train the presentation prediction model based on the immunopeptidomics data set to obtain a trained presentation prediction model, it is specifically used to: set the input data of the presentation prediction model to be peptide sequences and MHC allele subtype sequences, and the output data to be high-probability peptides, where high-probability peptides are peptides with a high probability of being naturally processed by antigen-presenting cells and presented by each MHC allele subtype; label the mass spectrometry elution ligand data in the immunopeptidomics data set, and the label is the presentation probability of the peptide-MHC pair; input the labeled mass spectrometry elution ligand data into the presentation prediction model for training to obtain a trained presentation prediction model.

[0171] In one embodiment, the human proteomics and immunology data set includes sequence data related to the human immune system; when the training unit 220 is used to train the MHC weight learning model based on the human proteomics and immunology data set to obtain the trained MHC weight learning model, it is specifically used to: set the input data of the MHC weight learning model to be the affinity and presentation probability corresponding to each HLA allele subtype i, and the output data to be the weight coefficient corresponding to each HLA allele subtype i; use the affinity prediction model and the presentation prediction model to predict the affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i for each human sequence and non-human sequence in the human immune system related sequence data of the human proteomics and immunology data set; input the affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i into the MHC weight learning model for training to obtain the weight coefficient corresponding to each HLA allele subtype i.

[0172] In one embodiment, the immune tolerance assessment model includes: human proteome sequence data; when the training unit 220 is used to train the immune tolerance assessment model based on the human proteomics and immunomics data sets to obtain a trained immune tolerance assessment model, it is specifically used to: set the input data of the immune tolerance assessment model to be peptide segments, and the output data to be the probability that the peptide segments are tolerated by the immune system; split the human proteome sequence data in the human proteomics and immunomics data sets to obtain multiple target peptide segments with non-repetitive sequences; input the multiple target peptide segments into the immune tolerance assessment model for training to obtain a trained immune tolerance assessment model.

[0173] In one embodiment, each data set also includes a human biological drug ADA data set; the immunogenicity prediction device based on multimodal fusion also includes a testing unit 260, which is used to: use the therapeutic monoclonal antibody data in the human biological drug ADA data set as a test set; input the test set into the immunogenicity prediction intelligent system, predict the immunogenicity of related protein molecules, and obtain an immunogenicity risk score; calculate the correlation coefficient of the clinical ADA level based on the immunogenicity risk score to test the immunogenicity prediction intelligent system's ability to predict immunogenicity risk.

[0174] The functions of each unit in the immunogenicity prediction device based on multi-modal fusion in the embodiment of the present application can be found in the corresponding description in the above method and will not be repeated here.

[0175] Figure 8 FIG. 2 shows a block diagram of a computer device according to an embodiment of the present application. Figure 8As shown, the computer device includes: a memory 310 and a processor 320. The memory 310 stores instructions, and the instructions are loaded and executed by the processor 320 to implement the immunogenicity prediction method based on multi-modal fusion in the above embodiment. The number of the memory 310 and the processor 320 can be one or more.

[0176] The computer device also includes:

[0177] The communication interface 330 is used to communicate with external devices and perform data exchange transmission.

[0178] If the memory 310, the processor 320 and the communication interface 330 are implemented independently, the memory 310, the processor 320 and the communication interface 330 can be connected to each other through a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 8 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0179] Optionally, in a specific implementation, if the memory 310, the processor 320 and the communication interface 330 are integrated on a chip, the memory 310, the processor 320 and the communication interface 330 can communicate with each other through an internal interface.

[0180] An embodiment of the present application provides a computer-readable storage medium, in which a computer program is stored. When the computer program runs on a computer, the method provided in the embodiment of the present application is implemented.

[0181] An embodiment of the present application also provides a chip, which includes a processor for calling and executing instructions stored in a memory, so that a communication device equipped with the chip executes the method provided by the embodiment of the present application.

[0182] An embodiment of the present application also provides a chip, including: an input interface, an output interface, a processor and a memory, wherein the input interface, the output interface, the processor and the memory are connected via an internal connection path, and the processor is used to execute the code in the memory. When the code is executed, the processor is used to execute the method provided in the embodiment of the application.

[0183] It should be understood that the processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. It is worth noting that the processor may be a processor supporting the Advanced RISC Machines (ARM) architecture.

[0184] Further, optionally, the above-mentioned memory may include a read-only memory and a random access memory, and may also include a non-volatile random access memory. The memory may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may include a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may include a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available. For example, static RAM (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM) and direct rambus RAM (DR RAM).

[0185] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium.

[0186] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0187] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0188] Any process or method description in the flow chart or otherwise described herein can be understood to represent a module, fragment or portion of a code including one or more executable instructions for implementing the steps of a specific logical function or process. And the scope of the preferred embodiment of the present application includes other implementations, in which the functions may not be performed in the order shown or discussed, including in a substantially simultaneous manner or in a reverse order according to the functions involved.

[0189] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, which can be embodied in any computer-readable medium for use by an instruction execution system, apparatus or device (such as a computer-based system, a system including a processor or other system that can fetch instructions from an instruction execution system, apparatus or device and execute instructions), or used in combination with these instruction execution systems, apparatuses or devices.

[0190] It should be understood that the various parts of the present application can be implemented with hardware, software, firmware or a combination thereof. In the above embodiments, multiple steps or methods can be implemented with software or firmware stored in a memory and executed by a suitable instruction execution system. All or part of the steps of the above embodiment method can be completed by instructing the relevant hardware through a program, which can be stored in a computer-readable storage medium, and when the program is executed, it includes one of the steps of the method embodiment or a combination thereof.

[0191] In addition, each functional unit in each embodiment of the present application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into one module. The above-mentioned integrated module can be implemented in the form of hardware or in the form of a software functional module. If the above-mentioned integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. The storage medium can be a read-only memory, a disk or an optical disk, etc.

[0192] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of various changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A method for predicting immunogenicity based on multimodal fusion, characterized in that: include: Constructing a multimodal prediction model, the multimodal prediction model includes an affinity prediction model, a presentation prediction model, an MHC weight learning model and an immune tolerance assessment model, the affinity prediction model is used to predict the affinity of a peptide segment to each MHC allele subtype, the presentation prediction model is used to predict the possibility of a peptide segment being naturally processed by antigen presenting cells and presented by each MHC allele subtype, the MHC weight learning model is used to learn the frequency weight of HLA, and the immune tolerance assessment model is used to assess the possibility of a peptide segment being tolerated by the immune system; Acquire various data sets, including immunopeptidomics data sets and human proteomics and immunoomics data sets; Training the multimodal prediction model based on each of the data sets to obtain a trained multimodal prediction model; The trained multimodal prediction models are integrated to construct a comprehensive immunogenicity assessment model, wherein the comprehensive immunogenicity assessment model is used to assess the complete immunogenicity risk based on the full-length protein sequence based on the output data of the affinity prediction model, the presentation prediction model, the MHC weight learning model, and the immune tolerance assessment model, wherein the input data of the MHC weight learning model are the output data of the affinity prediction model and the output data of the presentation prediction model, and the input data of the immune tolerance assessment model are the output data of the affinity prediction model and the output data of the presentation prediction model; Based on the trained multimodal prediction model and the immunogenicity comprehensive evaluation model, an immunogenicity prediction intelligent system is formed, wherein the input models of the immunogenicity prediction intelligent system are the affinity prediction model and the presentation prediction model, the intermediate models are the MHC weight learning model and the immune tolerance evaluation model, and the output model is the immunogenicity comprehensive evaluation model; Based on the immunogenicity prediction intelligent system, the immunogenicity of the protein to be predicted is predicted to obtain a test result, which includes the immunogenicity risk score of the protein to be predicted and the potential immunogenic T cell epitopes on the protein to be predicted.

2. The method according to claim 1, characterized in that The multimodal prediction model is trained based on each data set to obtain a trained multimodal prediction model, including: Training the affinity prediction model based on the immunopeptidomics data set to obtain a trained affinity prediction model; Training the presentation prediction model based on the immunopeptidomics dataset to obtain a trained presentation prediction model; Training the MHC weight learning model based on the human proteomics and immunology data sets to obtain a trained MHC weight learning model; The immune tolerance assessment model is trained based on the human proteomics and immunology data sets to obtain a trained immune tolerance assessment model.

3. The method according to claim 2, characterized in that The immunopeptidomics data set includes MHC-peptide affinity data; the affinity prediction model is trained based on the immunopeptidomics data set, and the trained affinity prediction model includes: The input data of the affinity prediction model are set to be peptide sequences and MHC allele subtype sequences, and the output data are high-affinity peptides, wherein the high-affinity peptides are peptides with high affinity to each MHC allele subtype; Tagging the MHC-peptide affinity data in the immunopeptidomics dataset, wherein the tag is the affinity of the peptide-MHC pair; The labeled MHC-peptide affinity data is input into the affinity prediction model for training to obtain the trained affinity prediction model.

4. The method according to claim 2, characterized in that: The immunopeptidomics data set includes mass spectrometry elution ligand data; the presentation prediction model is trained based on the immunopeptidomics data set, and the trained presentation prediction model includes: The input data of the presentation prediction model are set to be peptide sequences and MHC allele subtype sequences, and the output data are high-likelihood peptides, wherein the high-likelihood peptides are peptides with high probability of being naturally processed by antigen-presenting cells and presented by each MHC allele subtype; Tagging the mass spectrometry elution ligand data in the immunopeptidomics data set, wherein the tag is the presentation probability of the peptide-MHC pair; The labeled mass spectrometry elution ligand data is input into the presentation prediction model for training to obtain the trained presentation prediction model.

5. The method according to claim 2, characterized in that: The human proteomics and immunology datasets include sequence data related to the human immune system; The MHC weight learning model is trained based on the human proteomics and immunology data sets to obtain a trained MHC weight learning model, including: The input data of the MHC weight learning model are set to be the affinity and presentation probability corresponding to each HLA allele subtype i, and the output data are set to be the weight coefficient corresponding to each HLA allele subtype i; Using the affinity prediction model and the presentation prediction model, for each human sequence and non-human sequence in the human immune system related sequence data of the human proteomics and immunogenomics data sets, the affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i are predicted respectively; The affinity Bi and presentation probability Ei corresponding to each HLA allele subtype i are input into the MHC weight learning model for training to obtain the weight coefficient corresponding to each HLA allele subtype i.

6. The method according to claim 2, characterized in that The immune tolerance assessment model includes: human proteome sequence data; the immune tolerance assessment model is trained based on the human proteomics and immunology data sets, and the trained immune tolerance assessment model includes: The input data of the immune tolerance assessment model is set to be a peptide segment, and the output data is set to be the probability that the peptide segment is tolerated by the immune system; Segmenting the human proteome sequence data in the human proteomics and immunology data sets to obtain multiple target peptide segments with non-repeating sequences; The multiple target peptide segments are input into the immune tolerance assessment model for training to obtain the trained immune tolerance assessment model.

7. The method according to any one of claims 1 to 6, characterized in that The various data sets also include a human biological drug ADA data set; the method also includes: Use the therapeutic monoclonal antibody data in the human biopharmaceutical ADA dataset as a test set; Inputting the test set into the immunogenicity prediction intelligent system, performing immunogenicity prediction on relevant protein molecules, and obtaining an immunogenicity risk score; The correlation coefficient of clinical ADA level was calculated based on the immunogenicity risk score to test the predictive ability of the immunogenicity prediction intelligent system for immunogenicity risk.

8. An immunogenicity prediction device based on multi-mode fusion, characterized in that: include: A first construction unit is used to construct a multimodal prediction model, wherein the multimodal prediction model includes an affinity prediction model, a presentation prediction model, an MHC weight learning model and an immune tolerance assessment model, wherein the affinity prediction model is used to predict the affinity of a peptide segment to each MHC allele subtype, the presentation prediction model is used to predict the possibility of a peptide segment being naturally processed by an antigen presenting cell and presented by each MHC allele subtype, the MHC weight learning model is used to learn the frequency weight of HLA, and the immune tolerance assessment model is used to assess the possibility of a peptide segment being tolerated by the immune system; A training unit is used to obtain various data sets, wherein the various data sets include immunopeptidomics data sets and human proteomics and immunoomics data sets; and train the multimodal prediction model based on the various data sets to obtain a trained multimodal prediction model; A second construction unit is used to fuse the trained multimodal prediction models to construct an immunogenicity comprehensive evaluation model, wherein the immunogenicity comprehensive evaluation model is used to evaluate the complete immunogenicity risk based on the full-length protein sequence based on the output data of the affinity prediction model, the presentation prediction model, the MHC weight learning model and the immune tolerance evaluation model, wherein the input data of the MHC weight learning model are the output data of the affinity prediction model and the output data of the presentation prediction model, and the input data of the immune tolerance evaluation model are the output data of the affinity prediction model and the output data of the presentation prediction model; A third construction unit is used to construct an immunogenicity prediction intelligent system based on the trained multimodal prediction model and the immunogenicity comprehensive evaluation model, wherein the input model of the immunogenicity prediction intelligent system is the affinity prediction model and the presentation prediction model, the intermediate model is the MHC weight learning model and the immune tolerance evaluation model, and the output model is the immunogenicity comprehensive evaluation model; The prediction unit is used to perform immunogenicity prediction on the protein to be predicted based on the immunogenicity prediction intelligent system to obtain a test result, wherein the test result includes the immunogenicity risk score of the protein to be predicted and the T cell epitopes of potential immunogenicity on the protein to be predicted.

9. A computer device, characterized in that: include: A memory and a processor, wherein the memory stores instructions, and the instructions are loaded and executed by the processor to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed on a computer, the method according to any one of claims 1 to 7 is implemented.

Images