Vaccine target screening system based on calculation model simulation

Abstract

The invention provides a vaccine target screening system based on calculation model simulation. The vaccine target screening system comprises a multi-source heterogeneous database, wherein the multi-source heterogeneous database integrates and standardizes pathogenic genes, protein structures, literatures and experimental data; the feature calculation module calls a calculation biological model to carry out structural analysis, immunogenicity simulation and stability prediction; the intelligent screening and sorting module applies a multi-objective optimization algorithm to perform parallel evaluation and outputs optimal target spots; a structure iteration optimizer automatically iteratively corrects the optimized target spots to generate a high-potential variant library; and the process suitability simulation module couples the variants with the preparation formula and the process parameters to simulate production storage behaviors and feeds back an optimization target. According to the invention, efficient screening and optimization of vaccine targets can be realized, the accuracy and efficiency of target screening are improved, the research and development cost is reduced, and the research and development process of vaccines is accelerated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical engineering, and more particularly, to a vaccine target screening system based on computational model simulation. BACKGROUND

[0002] In the field of vaccine development, with the deepening of the research on the structure of pathogens and the immune mechanism, screening effective vaccine targets has become a key link in the field. Traditional methods mainly rely on experimental means, such as evaluating the immunogenicity and stability of antigens through cell culture, animal experiments, etc. However, these methods are time-consuming, costly, and difficult to quickly screen a large number of potential targets. In recent years, with the development of computational biology and artificial intelligence technology, vaccine target screening based on computational models has gradually attracted attention. By constructing a computational model, the gene sequence, protein structure, etc. of the pathogen can be analyzed to predict the immunogenicity, stability and production process adaptability of the antigen, thereby accelerating the screening process of vaccine targets.

[0003] In the prior art, although there are some vaccine target screening methods based on computational models, these methods mostly have the following problems or defects: first, the data source is single, lacking integration of multi-source heterogeneous data, which leads to the inability to comprehensively evaluate the characteristics of the target; second, the precision of the computational model is limited, which cannot accurately simulate the interaction between the antigen and the host immune system, as well as its behavior in the production process; in addition, the existing methods lack the ability to iteratively optimize the structure of the target, making it difficult to generate high-potential target variants. These problems limit the application effect of computational models in vaccine target screening, and cannot meet the demand for rapid and efficient screening of vaccine targets. SUMMARY

[0004] The present application provides a vaccine target screening system based on computational model simulation, comprising:

[0005] A multi-source database for storing raw data from pathogen gene databases, protein structure databases, vaccine literature databases and historical experiment databases, and standardizing and indexing the raw data according to a preset entity attribute model to form a unified feature space;

[0006] A feature calculation module for calling computational biology models to analyze molecular structure, simulate immunogenicity and predict stability for candidate vaccine target data stored in the unified feature space, and generating a target simulation feature matrix;

[0007] An ordering module for applying a multi-objective optimization algorithm to parallel evaluation and comprehensive ordering of a candidate vaccine target set based on the target simulation feature matrix, and outputting an optimal target subset;

[0008] an iterative optimizer, configured to generate a high-potential target variant library by iteratively modifying the molecular structure of each target in the preferred target subset according to a preset optimization objective through computational simulation driving, for each target in the preferred target subset;

[0009] a target simulation module, configured to couple the target in the high-potential target variant library with a preset vaccine formulation and production process parameter, simulate its behavior in the production and storage process, and feed back to correct the optimization objective.

[0010] Further, the multi-source database comprises:

[0011] a data acquisition unit, configured to periodically acquire or receive updated pathogen sequence data, antigen three-dimensional structure data, vaccine potency data, and process parameter data from a plurality of the original data sources through an API interface or a web crawler program;

[0012] a feature extraction unit, configured to perform gene annotation and antigen prediction on the acquired pathogen sequence data, surface accessibility calculation and epitope prediction on the antigen three-dimensional structure data, dose-response relationship fitting on the vaccine potency data, and normalization processing on the process parameter data, and extract numerical or vector type primary features;

[0013] a feature fusion unit, configured to associate and align the primary features from different data sources based on a common antigen identifier, and map them to the unified feature space through dimension reduction or embedding technology, so that each candidate vaccine target corresponds to a high-dimensional feature vector.

[0014] Further, in the feature extraction unit, the epitope prediction on the antigen three-dimensional structure data adopts a conformational B cell epitope prediction model based on a deep convolutional neural network, which takes the physicochemical properties of the antigen surface grid points as input, including hydrophobicity, hydrophilicity, flexibility, and solvent accessibility, and outputs the probability of each grid point belonging to an epitope, and the model training objective function is cross-entropy loss:

[0015]

[0016] wherein L represents the loss value; y i represents the true label of grid point i, taking the value of 1 when it belongs to the epitope and 0 when it does not belong to the epitope; p i represents the probability of grid point i belonging to the epitope predicted by the model.

[0017] Further, the feature calculation module comprises:

[0018] a molecular docking simulation unit for simulating a three-dimensional binding process between an antigen in a candidate vaccine target and a preset host immune molecule, the host immune molecule including a major histocompatibility complex (MHC) molecule and a B cell receptor, calculating a binding free energy, a binding interface area, and a number of hydrogen bonds;

[0019] an immune response prediction unit for predicting, based on a molecular docking result and physical and chemical characteristics of an antigen, an initial antibody titer, an antibody affinity maturation potential, and a memory B cell generation level generated by stimulation of the antigen by a trained machine learning regression model;

[0020] a stability prediction unit for predicting, based on an amino acid composition, a glycosylation pattern, and a disulfide bond prediction of an antigen, an aggregation tendency and a degradation half-life under a set temperature and pH stress condition by a simplified model of molecular dynamics simulation.

[0021] Further, in the molecular docking simulation unit, a modified MM-PBSA method for calculating a binding free energy of an antigen and an MHC class II molecule has a calculation formula as follows:

[0022] ΔG bind = ΔE MM + ΔG solv - TΔS

[0023] ΔE MM = ΔE internal + ΔE electrostatic + ΔE vdw

[0024] ΔG solv = ΔG PB + ΔG SA

[0025] wherein ΔG bind is a predicted binding free energy; ΔE MM is a molecular mechanics energy change, including an internal energy ΔE internal , an electrostatic interaction energy ΔE electrostatic , and a van der Waals interaction energy ΔE vdw ; ΔG solv is a solvation free energy change, including a Poisson-Boltzmann term ΔG PB and a solvent accessible surface area term ΔG SA ; T is an absolute temperature; and ΔS is an entropy change caused by binding.

[0026] Further, the sorting module includes:

[0027] a function construction unit for defining a plurality of objective functions including an immunogenicity score, a cross-reactivity score, a sequence conservation score, a developability score, and a safety risk score according to a vaccine design requirement;

[0028] a parallel sorting unit configured to employ a multi-objective evolutionary algorithm framework based on NSGA-II to parallelize fast non-dominated sorting and crowding calculation in the solution space constituted by the plurality of objective functions, taking each target in the target simulation feature matrix as an individual;

[0029] an output unit configured to filter out individuals on the Pareto front, i.e., the preferred target subset, from the sorted population and record the trade-off data of the individuals on different objectives.

[0030] Further, the developability scoring function is as follows:

[0031] Expression Level =w1×Expression Level +w2×Solubility Index +w3×Thermal Stability +w4×Aggregation Propensity ;

[0032] wherein Expression Level is a score of antigen expression level based on codon optimization and host system prediction; Solubility Index is a solubility index based on sequence prediction; Thermal Stability is a thermal stability prediction value from the stability prediction unit; Aggregation Propensity is a predicted aggregation tendency score; and w1, w2, w3, and w4 are weight coefficients preset according to production experience.

[0033] Further, the iterative optimizer comprises:

[0034] a condition parser configured to receive feedback from the target simulation module and preset structure rules, quantify the optimization target into a calculable loss function, and quantify the structure rules into hard constraints or soft penalty terms;

[0035] a structure fine-tuning unit configured to, for protein antigens, employ gradient information based on a protein language model or a three-dimensional convolutional neural network to guide point mutation of the amino acid sequence or fine-tuning of the local conformation, so as to optimize the loss function;

[0036] a modification unit configured to, for polysaccharide antigens, simulate modification of the length, branching, or linkage of the sugar chain according to chemical synthesis or enzyme cutting rules, evaluate the influence of the modification on immunogenicity and stability, and select a positive modification.

[0037] Further, in the structure fine-tuning unit, the process of guiding point mutation by using the protein language model comprises the following steps:

[0038] inputting the original antigen sequence into the pre-trained protein language model to obtain a full-sequence context embedding vector of each amino acid position;

[0039] calculating the gradient of the model output to each position amino acid type under the loss function;

[0040] selecting several positions with the most significant gradient descent direction according to the gradient information, replacing the original amino acid with a candidate amino acid indicated by the gradient and meeting the chemical rule and structure constraint to generate a mutant sequence;

[0041] returning the mutant sequence to the feature calculation module for rapid simulation to evaluate the optimization effect.

[0042] Further, the target simulation module comprises:

[0043] a rapid screening unit for simulating the binding strength and aggregation tendency between high-potential target variants and preset adjuvants, stabilizers based on a simplified model of intermolecular interaction potential, and predicting the physical compatibility in the formulation;

[0044] a behavior simulation unit for establishing a coarse-grained model of ice crystal growth, water migration and freeze-dried cake structure in the freeze-drying process, simulating the glass transition temperature change and reconstitution rate difference of different target variants in the presence of freeze-drying protectants;

[0045] a correlation prediction unit for correlating and analyzing the outputs of the rapid screening unit and the behavior simulation unit with the experience database to predict the long-term storage stability grade and degradation path of the final product, and feeding the predicted weak links back to the iterative optimizer as new constraints.

[0046] The above embodiments of the present application have at least the following beneficial effects:

[0047] 1. By integrating multi-source heterogeneous data and constructing a unified feature space, the system can comprehensively integrate pathogen genes, protein structures, vaccine literature and historical experimental data, overcoming the limitation of single data source in the prior art, providing a richer and more comprehensive information basis for vaccine target screening, and thus improving the accuracy and reliability of target screening.

[0048] 2. By introducing a feature calculation module, the system can accurately simulate the multi-dimensional characteristics of candidate vaccine targets by combining molecular structure analysis, immunogenicity simulation and stability prediction functions, solving the problem that traditional methods cannot accurately evaluate the comprehensive performance of targets. This multi-dimensional simulation analysis enables the system to quickly identify vaccine targets with high potential, greatly improving the screening efficiency.

[0049] 3. The sorting module and the iterative optimizer are designed. The system can not only perform parallel evaluation and comprehensive sorting of candidate target points based on multi-objective optimization algorithms, but also perform iterative correction of preferred target points to generate a high-potential target point variant library. This innovation solves the problem of insufficient target point optimization capability in the prior art, further improves the quality and applicability of target points, and provides higher-quality candidate target points for vaccine research and development. BRIEF DESCRIPTION OF DRAWINGS

[0050] The above and other objects, features and advantages of the exemplary embodiments of the present application will be more apparent from the following detailed description taken in conjunction with the accompanying drawings, in which:

[0051] Figure 1 A structural schematic diagram of a vaccine target point screening system based on computational model simulation is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0052] The principles and spirits of the present application will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are given only to enable those skilled in the art to better understand and implement the present application, and do not limit the scope of the present application in any way.

[0053] Reference will now be made to Figure 1 , Figure 1 A structural schematic diagram of a vaccine target point screening system based on computational model simulation is provided for an embodiment of the present application. As shown in Figure 1 , a vaccine target point screening system based on computational model simulation includes:

[0054] A multi-source database 101 is used to store original data from pathogen gene databases, protein structure databases, vaccine literature databases, and historical experiment databases, and to store and index the original data according to a preset entity attribute model to form a unified feature space;

[0055] A feature calculation module 102 is used to call computational biology models to analyze molecular structures, simulate immunogenicity, and predict stability for candidate vaccine target point data stored in the unified feature space, and to generate a target point simulation feature matrix;

[0056] A sorting module 103 is used to apply multi-objective optimization algorithms to perform parallel evaluation and comprehensive sorting of a candidate vaccine target point set based on the target point simulation feature matrix, and to output a preferred target point subset;

[0057] an iterative optimizer 104, configured to, for each target in the subset of preferred targets, generate a library of high-potential target variants by iteratively modifying its molecular structure according to a preset optimization objective through computational simulation driving;

[0058] a target simulation module 105, configured to couple a target in the library of high-potential target variants with a preset vaccine formulation recipe and production process parameters, simulate its behavior in production and storage, and feedback to modify the optimization objective.

[0059] In some embodiments, the multi-source database comprises:

[0060] a data acquisition unit, configured to periodically acquire or receive updated pathogen sequence data, antigen three-dimensional structure data, vaccine potency data, and process parameter data from a plurality of the original data sources through an API interface or a web crawler program; wherein the data sources can include NCBI, UniProt, PDB, IEDB, PubMed, a patent database, and an internal experimental data management system, etc. The API interface can use RESTful or GraphQL protocols, and the web crawler program can use Python Scrapy or BeautifulSoup libraries to build.

[0061] a feature extraction unit, configured to perform gene annotation and antigen prediction on the acquired pathogen sequence data, surface accessibility calculation and epitope prediction on the antigen three-dimensional structure data, dose-response relationship fitting on the vaccine potency data, and normalization processing on the process parameter data, and extract primary features in numerical or vector types;

[0062] a feature fusion unit, configured to associate and align the primary features from different data sources based on a common antigen identifier, and map them to the unified feature space through dimension reduction or embedding technology, so that each candidate vaccine target corresponds to a high-dimensional feature vector.

[0063] The multi-source database integrates raw data from different data sources, forms a unified feature space through standardization processing, and solves the problems of scattered data sources and inconsistent formats. The data acquisition unit acquires updated data from multiple data sources. These data sources include pathogen gene databases, protein structure databases, vaccine literature databases, and historical experimental databases.

[0064] The pathogen gene database stores the genetic sequence information of pathogens, the protein structure database provides the three-dimensional structure data of proteins, the vaccine literature database covers academic literature related to vaccine research and development, and the historical experimental database records the data and results of past experiments.

[0065] The feature extraction unit processes the acquired data, such as gene annotation and antigen prediction for pathogen sequence data, surface accessibility calculation and epitope prediction for protein structure data, dose-response relationship fitting for vaccine potency data, and normalization for process parameter data. Gene annotation can use existing bioinformatics tools, such as NCBI's BLAST tool for sequence alignment and functional annotation; antigen prediction can be based on machine learning models, input pathogen sequence data, and output potential antigen region probability distribution, which can use online tools such as VaxiJen, ANTIGENpro, or locally trained models. Surface accessibility calculation can use DSSP, NACCESS, etc. Dose-response relationship fitting can use a four-parameter logistic (4PL) model or other nonlinear regression models, fitted by Python's SciPy or R language.

[0066] The feature fusion unit associates and aligns the primary features of different data sources based on common antigen identifiers, and maps them to a unified feature space through dimension reduction or embedding techniques, such as principal component analysis (PCA) or t-SNE for reducing feature dimension, and embedding techniques such as Word2Vec or BERT for converting sequence data to vector representation. In the feature fusion unit, embedding techniques can be used to map feature vectors from different sources to the same dimensional space, such as converting protein sequence amino acid sequences to fixed length vectors through pre-trained language models such as ESM or ProtBERT, and then concatenating or weighted summing with protein structure data feature vectors to obtain high-dimensional feature vectors for each candidate target.

[0067] In some embodiments, in the feature extraction unit, epitope prediction is performed on the antigen three-dimensional structure data, using a conformational B cell epitope prediction model based on deep convolutional neural network, which takes the physicochemical properties of antigen surface grid points as input, including hydrophobicity, hydrophilicity, flexibility and solvent accessibility, and outputs the probability of each grid point belonging to an epitope, and the model training objective function is cross-entropy loss:

[0068]

[0069] where L represents the loss value; y i represents the true label of grid point i, taking value 1 when it belongs to an epitope, and value 0 when it does not belong to an epitope; p i represents the probability of grid point i belonging to an epitope predicted by the model.

[0070] The feature extraction unit performs epitope prediction on the obtained antigen three-dimensional structure data. Epitope refers to a specific region of an antigen molecule that can be recognized and bound by the immune system, usually located on the surface of the antigen. A conformational B cell epitope prediction model based on deep convolutional neural network is used to predict the probability of each grid point belonging to an epitope by analyzing the physicochemical properties of the antigen surface grid points, such as hydrophobicity, hydrophilicity, flexibility, and solvent accessibility. This model uses deep learning techniques to extract features from complex biophysical data and optimizes model performance through a cross-entropy loss function. The training goal of this model is to make the prediction results as close to the true labels as possible, thereby improving the accuracy of epitope prediction.

[0071] In the epitope prediction model, the input data is the physicochemical properties of the antigen surface grid points, reflecting the chemical and physical characteristics of the antigen surface. Hydrophobicity represents the degree of water repulsion of the antigen surface region, hydrophilicity represents the degree of water attraction; flexibility describes the deformability of the antigen surface region, while solvent accessibility indicates whether the region is easily accessible to solvent molecules. By taking physicochemical properties as input, through multiple layers of convolution and pooling operations of convolutional neural network, features related to epitope are extracted. The output of the model is the probability value of each grid point belonging to an epitope, ranging from 0 to 1. The cross-entropy loss function is used to measure the difference between the model prediction value and the true label, where the true label is a binary variable, taking value 1 indicates that the grid point belongs to the epitope, and taking value 0 indicates that it does not belong to the epitope.

[0072] The epitope prediction model is constructed by collecting antigen structure data and its corresponding epitope annotation information as training data, which can be obtained from public databases such as PDB protein data bank, and annotated by experimentally verified epitope information. The input antigen surface grid point data is preprocessed, including normalizing physicochemical property values to unify the range to 0 to 1. The model can use a multi-layer convolutional neural network, for example, use 3 layers of convolutional layers, each layer includes 64 convolutional kernels, the size of the convolutional kernel is 3x3, and the activation function uses ReLU. In the training process, the stochastic gradient descent (SGD) optimizer is used, the learning rate is set to 0.001, and the model performance is evaluated by cross-validation.

[0073] In some embodiments, the feature calculation module generates a target simulation feature matrix by analyzing and simulating the molecular structure of the candidate vaccine target. It includes a molecular docking simulation unit, an immune response prediction unit, and a stability prediction unit, respectively used to simulate the binding process of antigens and host immune molecules, predict immune response, and evaluate the stability of antigens.

[0074] The molecular docking simulation unit is used to simulate the three-dimensional binding process between antigens and host immune molecules, such as MHC molecules and B cell receptors. Binding free energy, binding interface area, and the number of hydrogen bonds are parameters for evaluating binding stability and affinity. The immune response prediction unit predicts immune response indicators, such as initial antibody titer, antibody affinity maturation potential, and memory B cell generation level, based on molecular docking results and physicochemical characteristics of antigens. The stability prediction unit predicts the aggregation tendency and degradation half-life of antigens under specific conditions using a simplified model of molecular dynamics simulation.

[0075] The calculation of binding free energy uses the modified MM-PBSA method, which can accurately evaluate the binding stability between antigens and immune molecules based on factors such as molecular mechanics energy change, solvation free energy change, and entropy change.

[0076] In implementation, select molecular docking software such as AutoDock, GROMACS, or Glide, and input the three-dimensional structure files of antigens and host immune molecules. During simulation, set the search space to cover possible binding sites on the surface of the antigen, and use genetic algorithms or Monte Carlo methods for global optimization to find the optimal binding conformation. The modified MM-PBSA method includes molecular mechanics energy changes, including internal energy, electrostatic interaction energy, and van der Waals interaction energy, and solvation free energy changes, including Poisson-Boltzmann terms and solvent accessible surface area terms.

[0077] In the immune response prediction unit, the machine learning regression model can be trained based on existing experimental data, with input parameters including physicochemical characteristics of antigens and molecular docking results, and output as immune response indicators. In the stability prediction unit, the simplified model of molecular dynamics simulation can be implemented through software such as GROMACS, with input of amino acid sequences and initial conformations of antigens, setting simulation temperature and pH conditions, and output of aggregation tendency and degradation half-life of antigens.

[0078] In some embodiments, the modified MM-PBSA method for calculating the binding free energy of antigens and MHC class II molecules in the molecular docking simulation unit has the following calculation formula:

[0079] ΔG bind =ΔE MM +ΔG solv -TΔS

[0080] ΔE MM =ΔE internal +ΔE electrostatic +ΔE vdw

[0081] ΔG solv =ΔG PB +ΔG SA

[0082] where AG bind is the predicted binding free energy; AE MM is the molecular mechanics energy change, including the internal energy AE internal , electrostatic interaction energy AE electrostatic , and van der Waals interaction energy AE vdw ; AG solv is the solvation free energy change, including the Poisson-Boltzmann term AG PB and the solvent accessible surface area term AG SA ; T is the absolute temperature; and AS is the entropy change caused by binding.

[0083] In practice, the following steps are performed using the GROMACS software: energy minimization of the complex, receptor, and ligand separately; equilibration under constant temperature volume NVT and constant temperature pressure NPT ensembles; production molecular dynamics simulation; extraction of trajectory frames, and calculation of each energy term using the gmx mmxpbxsa module or similar tools. The electrostatic term AG PB is calculated using the APBS software to solve the Poisson-Boltzmann equation. The nonpolar solvation term is generally linearly correlated with the solvent accessible surface area SASA, with the formula:

[0084] AG SA = γ × SASA + b

[0085] where γ and b are empirical parameters, and the entropy change AS can be approximated using quasi-harmonic analysis by gmx covar and gmx anaeig or normal mode analysis.

[0086] In some embodiments, the ranking module ranks the candidate vaccine target set based on the target simulation feature matrix, and outputs a preferred target subset by evaluating and ranking the candidate vaccine target set through a multi-objective optimization algorithm. By constructing a multi-objective evaluation function and using a multi-objective evolutionary algorithm framework based on NSGA-II, multiple conflicting objectives are processed to screen out candidate targets that perform well in immunogenicity, safety, and developability.

[0087] Each of the multiple objective functions corresponds to a requirement in vaccine design, such as an immunogenicity score reflecting the ability of a target to stimulate an immune response, a cross-reactivity score evaluating the risk of a target binding to non-target antigens, a sequence conservation score measuring the degree of conservation of a target among different pathogen strains, a developability score considering the feasibility of a target in production processes, and a safety risk score evaluating the risk of side effects that a target may cause.

[0088] The parallel ranking unit adopts the NSGA-II algorithm framework, considering each target as an individual, and performing fast non-dominated sorting and crowding calculation in the solution space composed of multiple objective functions. Non-dominated sorting is used to identify individuals that perform well on multiple objectives, and crowding calculation is used to maintain population diversity and avoid local optimization. The output unit selects individuals on the Pareto frontier from the sorted population, i.e. the preferred target subset, and records their trade-off data on different objectives, providing a basis for subsequent target optimization.

[0089] In implementation, according to the specific needs of vaccine design, the calculation method and weight distribution of each objective function are determined. For example, the immunogenicity score can be calculated by analyzing the number and quality of epitopes of the target, such as predicted binding affinity; the cross-reactivity score can be evaluated by comparing the target sequence with known allergen sequences, such as using BLAST for comparison, calculating E-value or sequence similarity; the sequence conservation score can be determined by calculating the sequence identity of the target in different pathogen strains, such as using Clustal Omega for multiple sequence alignment, calculating the conservation score; the developability score can be evaluated by combining production process parameters such as expression level, solubility, stability, etc.; the safety risk score can be calculated by analyzing the association of the target with known pathogenic genes or toxic epitopes, such as querying ToxinPred or similar databases.

[0090] The population size, such as 100-500 individuals, and the number of iterations, such as 50-200 generations, are set. The population size can be adjusted according to the number of candidate targets, and the number of iterations is determined according to the convergence of the algorithm. The input parameters of the algorithm include the feature matrix of each target and the weight of the objective function, and the output is the non-dominated sorted population and the trade-off data of each target on different objectives. Specific implementation can use Python DEAP, JMetal or Platypus optimization algorithm library.

[0091] As an example, the steps include: initializing the population, randomly generating or selecting from the candidate target set; calculating all objective function values of each individual; performing non-dominated sorting by fast non-dominated sorting algorithm; calculating crowding; selecting, crossing and mutating to generate offspring population; merging parent and offspring populations, repeating the sorting and selection process until the maximum number of iterations is reached.

[0092] In some embodiments, the developability score function is as follows:

[0093] Expression Level = w1 x Expression Level + w2 x Solubility Index + w3 x Thermal Stability+ w4 x Aggregation Propensity ;

[0094] wherein Expression Level is the antigen expression level score based on codon optimization and host system prediction; Solubility Index is the solubility index based on sequence prediction; Thermal Stability is the thermal stability prediction value from the stability prediction unit; Aggregation Propensity is the predicted aggregation propensity score; w1, w2, w3, w4 are the weight coefficients preset according to production experience.

[0095] The antigen expression level score reflects the expression efficiency of the target in host cells such as CHO, HEK293 or E. coli, which is usually obtained through codon optimization and host system adaptability analysis. Codon optimization can be achieved using tools such as Optimizer, DNAWorks or IDT's Codon Optimization Tool. The solubility index Solubility Index evaluates the solubility of the target protein in solution, which can be predicted by sequence analysis tools such as Protscale, SoluProt or machine learning model-based prediction tools, and the score can be normalized to 0-1. The thermal stability prediction value Thermal Stability indicates the structural stability of the target under high temperature conditions, which is usually obtained through molecular dynamics simulation or experimental determination. The aggregation propensity score Aggregation Propensity evaluates the risk of aggregation of the target during production, which can be obtained through sequence analysis tools such as AGGRESCAN, TANGO or PASTA or experimental data, and the score can be normalized to 0-1 or obtained from experimental data.

[0096] The weight coefficients w1, w2, w3, w4 are parameters preset according to production experience and experimental data, used to balance the influence of different characteristics on developability. For example, w1 is larger because antigen expression level is crucial in production; w4 is negative because aggregation propensity is detrimental to production.

[0097] As an example, the target sequence is analyzed by bioinformatics tools such as codon optimization software to predict the antigen expression level score Expression Level . The solubility index Solubility Index is calculated using protein solubility prediction models such as empirical formulas based on amino acid composition. The thermal stability Thermal Stability and aggregation propensity score AggregationPropensity The weighting coefficients w1, w2, w3, and w4 are adjusted based on the priority of the production process and empirical data. For example, if the production process requires a higher expression level, the weight of w1 can be increased. During the calculation, the above parameter values ​​are substituted into the formula to obtain the exploitability score for each candidate target. For example, assuming w1=0.4, w2=0.3, w3=0.2, w4=-0.1, the antigen expression level score is 80, the solubility index is 70, the thermal stability is 90, and the aggregation tendency score is 20, the exploitability score is:

[0098] Developability Score =0.4×80+0.3×70+0.2×90-0.1×20=75.

[0099] In some embodiments, an iterative optimizer is used to iteratively refine the molecular structure of each target in a preferred subset of targets, generating a library of high-potential target variants. A conditional resolver receives feedback from the target simulation module and preset structural rules, transforming the optimization objective into a specific loss function. For example, the optimization objective might be to improve the immunogenicity or stability of the target, and the loss function could be the difference between the target and the ideal immunogenicity or stability.

[0100] Structural rules can be conservation requirements at certain positions in the amino acid sequence of an antigen. These can be quantified as hard constraints or soft penalties. Hard constraints are conditions that must be met, such as the amino acid type at certain positions being immutable. Soft penalties increase the loss value when a violation is committed; for example, introducing a protease cleavage site increases the penalty value. The structural fine-tuning unit uses gradient information generated by protein language models, such as ESM (Evolutionary Scale Modeling), ProtBERT, or a 3D convolutional neural network (CNN), to guide point mutations in the amino acid sequence. Its output is a context embedding vector for each amino acid position. By calculating the gradient of the loss function with respect to the amino acid type at each position, it determines which positions need to be mutated and which amino acid to mutate to.

[0101] The modification unit simulates modifications to glycan length, branching, or linkages based on the chemical synthesis rules of the polysaccharide antigen, and evaluates the effects of these modifications on immunogenicity and stability. For example, stability can be improved by increasing glycan branching, or immunogenicity can be optimized by changing the type of linkage, such as changing β-1,4 linkages to α-1,6 linkages.

[0102] In some embodiments, the process of using a protein language model to guide point mutation in the structural fine-tuning unit includes:

[0103] The original antigen sequence is input into a pre-trained protein language model to obtain the full-sequence context embedding vector for each amino acid position.

[0104] Computing the gradient of the loss function with respect to the amino acid type at each position;

[0105] According to the gradient information, a number of positions with the most significant gradient descent direction are selected, and the original amino acid is replaced with a candidate amino acid indicated by the gradient and meeting the chemical rule and structural constraint to generate a mutant sequence.

[0106] Returning the mutant sequence to the feature calculation module for rapid simulation to evaluate the optimization effect.

[0107] The protein language model can use ESM-1b, ESM-2 or ProtBERT model to process protein sequence data and generate context embedding vectors for each amino acid position. The embedding vectors include global information of the sequence and reflect the interaction between amino acids and the biological function of the sequence. The protein language model can be obtained from Hugging Face ModelHub or ESM official GitHub repository.

[0108] The gradient of the loss function with respect to the amino acid type at each position is calculated to determine which positions of the amino acid need to be replaced to optimize the target performance. The gradient information indicates the direction and importance of amino acid replacement, helping to select a better amino acid replacement scheme. The gradient of the loss function with respect to the amino acid type at each position can be realized by automatic differentiation function of PyTorch or TensorFlow.

[0109] Local conformation fine-tuning considers the three-dimensional structure of the protein, analyzes the local spatial structure of the protein through three-dimensional convolutional neural network CNN, and guides the fine-tuning of the conformation to optimize its binding properties or stability. The 3D-CNN model can use atomic coordinates or voxelized grid as input, and predict the structure properties or energy by training. When implemented, it needs to meet the chemical rule and structural constraint, such as using Rotamer library to evaluate the rationality of side chain, using Rosetta or FoldX to evaluate the stability of structure, to ensure that the generated mutant sequence is biochemically feasible.

[0110] In some embodiments, the original antigen sequence is input into a pre-trained protein language model such as ESM, and the model outputs a full sequence context embedding vector for each amino acid position. Define the loss function, for example, the immunogenicity loss function can be the difference between the predicted antibody binding affinity and the target value, and the stability loss function can be the difference between the predicted degradation half-life and the target value. The gradient of the loss function with respect to the amino acid type at each position is calculated by backpropagation to determine the positions that need to be mutated.

[0111] According to the gradient information, a number of positions with the most significant gradient descent direction are selected, usually the top 5 positions, and the original amino acid is replaced with the gradient indicated, i.e. the amino acid type corresponding to the most negative gradient value, to generate a mutant sequence. The generated mutant sequence is returned to the feature calculation module for rapid simulation to evaluate its optimization effect. For example, the change in binding affinity of the mutant to the host immune molecule is evaluated by molecular docking simulation, and the change in stability under physiological conditions is evaluated by stability prediction.

[0112] In some embodiments, the target simulation module predicts the performance indicators of the target in the production, storage and use process by simulating the behavior of the target in the vaccine formulation and production process, and feeds back the prediction results to the iterative optimizer for further optimization of the target structure.

[0113] The rapid screening unit simulates the binding strength and aggregation tendency between the target variant and molecules such as adjuvants and stabilizers based on a simplified model of intermolecular interaction potential, and predicts the physical compatibility in the formulation. The behavior simulation unit establishes a coarse-grained model of ice crystal growth, water migration and freeze-dried cake structure in the freeze-drying process, simulates the change in glass transition temperature and the difference in reconstitution rate of different target variants in the freeze-drying protectant. The correlation prediction unit performs correlation analysis on the outputs of the above units with the empirical database to predict the long-term storage stability grade and degradation path of the final product, and feeds back the predicted weak links as new constraints to the iterative optimizer. The parameters involved in these units include intermolecular interaction energy, glass transition temperature, reconstitution rate, etc.

[0114] The rapid screening unit can evaluate the binding strength and aggregation tendency between the target variant and adjuvants, stabilizers, etc. by calculating the intermolecular interaction potential such as van der Waals force and electrostatic interaction energy. For example, use molecular dynamics simulation software such as GROMACS to input the three-dimensional structure of the target variant and adjuvants, set simulation parameters such as temperature, pressure, and solvent environment, and calculate the intermolecular interaction energy by simulation to judge the physical compatibility.

[0115] The behavior simulation unit establishes a coarse-grained model of ice crystal growth, water migration and freeze-dried cake structure in the freeze-drying process, simulates the change in glass transition temperature and the difference in reconstitution rate of different target variants in the freeze-drying protectant. The model can be based on the finite element method, using COMSOL Multi physics or ANSYS Fluent software or lattice Boltzmann method. The ice crystal growth model can be described by the Stefan equation; water migration can be described by Fick's second law; freeze-dried cake structure can be characterized by porosity model. The glass transition temperature T g The Gordon-Taylor equation can be used to estimate: ,

[0116] where w is the mass fraction, Tg1 and T g2 T g , k is a fitting parameter. The reconstitution rate can be described by the Noyes- Whitney equation as shown below:

[0117] ,

[0118] where k s is the dissolution rate constant, A is the surface area, C s is the saturation concentration, and C is the instantaneous concentration.

[0119] The correlation prediction unit correlates the outputs of the above units with empirical databases, such as the intrinsic stability database, degradation kinetics data in literature, to predict the long-term storage stability grade and key degradation pathways of the final product, and feeds the predicted weak links as new constraints back to the iterative optimizer. The parameters involved in these units include intermolecular interaction energy, glass transition temperature, reconstitution rate, etc. The correlation analysis can employ machine learning methods, such as using random forest or gradient boosting tree regression models, inputting the simulated physicochemical parameters, such as T g , aggregation propensity score, and process parameters, such as lyophilization rate, annealing temperature, and outputting the predicted stability grade, such as the stability score at 2-8°C for 24 months, and the main degradation pathways, such as deamidation, oxidation, aggregation. The model can be trained using the scikit-learn or XGBoost library in Python.

[0120] The rapid screening unit can assess the binding strength and aggregation propensity of the target variant with the adjuvant, stabilizer, by calculating the intermolecular interaction potential, such as van der Waals force, electrostatic interaction energy, etc. For example, using molecular dynamics simulation software, such as GROMACS, input the three-dimensional structure of the target variant and the adjuvant, set the simulation parameters, such as temperature 4°C, 25°C, pressure 1 atm, solvent environment: aqueous solution, calculate the intermolecular interaction energy by simulation, analyze the van der Waals and electrostatic terms using gmx energy, and judge their physical compatibility.

[0121] The behavior simulation unit can use finite element analysis methods to establish mathematical models of ice crystal growth and water migration during the lyophilization process, input the physicochemical parameters of the target variant, such as molecular weight, solubility, diffusion coefficient, and lyophilization process parameters, such as cooling rate, heating rate, chamber pressure, simulate the glass transition temperature and reconstitution rate of the target variant during lyophilization. The correlation prediction unit correlates the above simulation results with stability data in the empirical database through machine learning algorithms to predict the long-term storage stability grade and key degradation pathways of the target variant. For example, using random forest or support vector machine models, implemented by scikit-learn, input the simulated physicochemical parameters, such as Tg , aggregation tendency, reconstitution time, and process parameters such as lyophilization cycle, annealing step, output stability grade, classification: stable, moderate, unstable, and degradation path prediction result: probability distribution. Through these detailed operation steps, the target simulation module can provide comprehensive process evaluation for the screening and optimization of vaccine targets, ensuring the feasibility and stability of candidate targets in actual production.

[0122] This embodiment takes the hemagglutinin HA protein of the influenza A virus A / California / 07 / 2009 (H1N1) strain (GenBank accession number ACP41105.1, corresponding to the monomer chain in PDB structure 3LZG) as a specific application object to show the implementation of the present application in actual vaccine target screening.

[0123] The data acquisition unit obtains the amino acid sequence of the above-mentioned HA protein from the NCBI protein database through the API interface, downloads the three-dimensional structure coordinate file thereof from the PDB database, obtains the B cell linear epitope and conformational epitope data of the protein that have been experimentally verified from the IEDB immunogenic epitope database, and calls the relevant experimental records of the immunogenicity of the protein in a mouse model from the internal historical experimental database.

[0124] The feature extraction unit performs gene annotation on the obtained HA sequence, identifies its signal peptide and domain, and calls the antigen prediction tool to confirm its antigenicity score. Surface accessibility calculation is performed on its three-dimensional structure.

[0125] The trained deep convolutional neural network model is used to predict conformational B cell epitopes, including: taking the hydrophobicity, hydrophilicity, flexibility, and solvent accessibility values of the HA protein surface grid points as input, outputting the probability of each grid point belonging to an epitope, and predicting three main potential epitope regions located in the HA1 globular head, which is consistent with the literature reports and IEDB records. The feature fusion unit aligns the primary features from the sequence, structure, literature, and experimental data based on a unified protein identifier, and reduces and fuses them through principal component analysis (PCA) technology, generating a unified feature vector representation including 32 feature dimensions for the HA protein and other homologous alignment sequences.

[0126] The feature calculation module then performs multi-dimensional simulation calculation on the above-featured HA protein. The molecular docking simulation unit uses AutoDock Vina software to simulate the three-dimensional docking between the HA protein receptor binding domain (RBD) and the MHC class II molecule HLA-DRB*101:01, and applies the modified MM-PBSA method to calculate the optimal binding conformation, obtaining its binding free energy The predicted initial IgG antibody titer geometric mean titer (GMT) in mice is 1:2560, and the antibody affinity maturation potential index is 0.78 (range 0-1). The stability prediction unit predicts the degradation half-life of the HA protein under accelerated stability test conditions at 40°C to be about 14 days based on the amino acid composition of the HA protein, predicted glycosylation sites, and disulfide bond networks. The simulation feature matrix of the target is constructed based on the above simulation results.

[0127] The ranking module applies a multi-objective optimization algorithm for comprehensive evaluation based on the simulation feature matrix. In this embodiment, multiple objective functions including immunogenicity, safety, and developability are defined, and the weight coefficients are preset according to production experience. The multi-objective evolutionary algorithm framework based on NSGA-II is used, and the HA protein and several variant sequences obtained by homologous alignment are used as initial population individuals. Fast non-dominated sorting and congestion calculation are performed in a parallel environment. After 100 generations of iteration, three preferred target points on the Pareto frontier are selected from the population to form a preferred target point subset.

[0128] The iterative optimizer iteratively modifies the original HA protein target to improve immunogenicity. The condition resolver quantifies the maximum predicted antibody titer objective as a loss function, and the structure fine-tuning unit inputs the full-length sequence of the HA protein into the pre-trained protein language model ProtBERT to obtain the context embedding vector of each amino acid position and calculate the gradient of the loss function with respect to each position 20 standard amino acid types.

[0129] According to the gradient information, it is identified that the amino acid residues located at positions 145 (lysine, K) and 156 (asparagine, N) of the HA1 head are the positions that have the greatest potential contribution to immunogenicity gain. Under the premise of avoiding damage to the folding or key epitopes, K145 is replaced with arginine R, and N156 is replaced with serine S to generate a double-point mutant K145R / N156S. The mutant sequence is returned to the feature calculation module for rapid simulation, and the results show that the predicted antibody titer is increased to GMT 1:3840, and the stability parameters have not been significantly degraded, thereby verifying the effectiveness of the optimization, and the mutant is stored in the high-potential target variant library.

[0130] The target point simulation module simulates the optimized HA mutant to evaluate and quickly screen the interaction energy between the mutant and the surface of the commonly used aluminum adjuvant (aluminum hydroxide) particles, predicts that its adsorption efficiency is good, and does not show an abnormal self-aggregation tendency. The behavior simulation unit establishes a coarse-grained model describing the freeze-drying process, inputs the predicted value of the glass transition temperature Tg of the mutant, the molecular size and other parameters, simulates the water migration and freeze-dried cake shape change of the mutant in the trehalose / sucrose mixed protective agent system during the freezing, primary drying and secondary drying processes, and predicts that its resolubility time is less than 30 seconds. The correlation prediction unit correlates the above physical compatibility and freeze-drying behavior simulation data with the internally established vaccine stability experience database for correlation analysis, predicts that the antigen activity retention rate of the HA mutant in the aluminum adjuvant adsorption and freeze-dried preparation form is expected to be greater than 95% after 18 months of storage at 2-8°C, and identifies the local water residue in the freeze-drying process as a potential risk point, which is fed back to the iterative optimizer as a new optimization constraint.

[0131] The above description is only some of the preferred embodiments of the present application and an explanation of the principles of the technology used. Those skilled in the art should understand that the scope of the application involved in the embodiments of the present application is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the above features are replaced with each other to form a technical solution with similar functions to the technical features disclosed in the embodiments of the present application (but not limited to).

Claims

1. A vaccine target screening system based on computational model simulation, characterized in that, The system includes: A multi-source database is used to store raw data from pathogen gene databases, protein structure databases, vaccinology literature databases, and historical experimental databases. The raw data is standardized and indexed according to a preset entity attribute model to form a unified feature space. The feature calculation module is used to call computational biology models to perform molecular structure analysis, immunogenicity simulation and stability prediction on candidate vaccine target data stored in the unified feature space, and generate a target simulation feature matrix. The sorting module is used to perform parallel evaluation and comprehensive sorting of the candidate vaccine target set based on the target simulation feature matrix and apply a multi-objective optimization algorithm to output a preferred target subset. An iterative optimizer is used to iteratively modify the molecular structure of each target in the preferred target subset according to a preset optimization objective, through computational simulation, to generate a library of high-potential target variants. The target simulation module is used to couple the targets in the high-potential target variant library with preset vaccine formulations and production process parameters, simulate their behavior during production and storage, and provide feedback to correct the optimization target.

2. The vaccine target screening system based on computational model simulation according to claim 1, characterized in that, The multi-source database includes: The data acquisition unit is used to periodically acquire or receive updated pathogen sequence data, antigen three-dimensional structure data, vaccine potency data, and process parameter data from multiple original data sources via API interfaces or web crawler programs. The feature extraction unit is used to perform gene annotation and antigen prediction on the acquired pathogen sequence data, perform surface accessibility calculation and epitope prediction on the antigen three-dimensional structure data, perform dose-response relationship fitting on the vaccine potency data, normalize the process parameter data, and extract it into numerical or vector-type primary features. The feature fusion unit is used to associate and align the primary features from different data sources based on a common antigen identifier, and map them to the unified feature space through dimensionality reduction or embedding techniques, so that each candidate vaccine target corresponds to a high-dimensional feature vector.

3. The vaccine target screening system based on computational model simulation according to claim 2, characterized in that, In the feature extraction unit, epitope prediction is performed on the three-dimensional structure data of the antigen using a conformational B-cell epitope prediction model based on a deep convolutional neural network. This model takes the physicochemical properties of the grid points on the antigen surface as input, including hydrophobicity, hydrophilicity, flexibility, and solvent accessibility, and outputs the probability that each grid point belongs to an epitope. The model training objective function is cross-entropy loss. Wherein, L represents a loss value; y i represents a real label of the grid point i, and the value of 1 indicates that it belongs to the epitope, and the value of 0 indicates that it does not belong to the epitope; p i represents the probability that the grid point i predicted by the model belongs to the epitope.

4. The vaccine target screening system based on computational model simulation according to claim 1, characterized in that, The feature calculation module includes: The molecular docking simulation unit is used to simulate the three-dimensional binding process between the antigen in the candidate vaccine target and the preset host immune molecules, which include the major histocompatibility complex (MHC) molecule and the B cell receptor, and calculates the binding free energy, binding interface area and number of hydrogen bonds. The immune response prediction unit is used to predict the initial antibody titer, antibody affinity maturation potential, and memory B cell generation level generated by antigen stimulation based on molecular docking results and the physicochemical characteristics of the antigen, using a trained machine learning regression model. The stability prediction unit is used to predict the amino acid composition, glycosylation mode and disulfide bond of the antigen. It uses a simplified model of molecular dynamics simulation to predict the aggregation tendency and degradation half-life under set temperature and pH stress conditions.

5. The vaccine target screening system based on computational model simulation according to claim 4, characterized in that, In the molecular docking simulation unit, the modified MM-PBSA method is used to calculate the binding free energy between the antigen and MHC class II molecules. The calculation formula is as follows: AG bind = DE MM + AG solv - TAS ΔE MM =ΔE internal +ΔE electrostatic +ΔE vdw AG solv = AG PB + AG SA where ΔG bind is the predicted binding free energy; ΔE MM is the change in molecular mechanics energy, including bond stretching ΔE internal , bond information ΔE electrostatic , and van der Waals interactions ΔE vdw ; ΔG solv is the change in solvation free energy, including the Poisson-Boltzmann term ΔG PB and the solvation accessible surface area term ΔG SA ; T is the absolute temperature; and ΔS is the change in entropy upon binding.

6. The vaccine target screening system based on computational model simulation according to claim 5, characterized in that, The sorting module includes: The function building unit is used to define multiple objective functions, including immunogenicity score, cross-reactivity score, sequence conservation score, exploitability score, and safety risk score, according to vaccine design requirements. The parallel sorting unit is used to perform parallel fast non-dominated sorting and crowding calculation in the solution space composed of the multiple objective functions, using the NSGA-II-based multi-objective evolutionary algorithm framework, treating each target point in the target point simulation feature matrix as an individual. The output unit is used to select individuals at the Pareto front from the sorted population, i.e., the preferred target subset, and record their trade-off data on different targets.

7. The vaccine target screening system based on computational model simulation according to claim 6, characterized in that, The exploitability scoring function is shown in the following formula: Expression Level =w1×Expression Level +w2×Solubility Index +w3×Thermal Stability +w4×Aggregation Propensity ; Among them, Expression Level Solubility is a score for antigen expression levels based on codon optimization and host system prediction. Index For solubility index based on sequence prediction; Thermal Stability The predicted thermal stability value comes from the stability prediction unit; Aggregation Propensity The predicted clustering tendency score; w1, w2, w3, and w4 are weighting coefficients preset based on production experience.

8. The vaccine target screening system based on computational model simulation according to claim 1, characterized in that, The iterative optimizer includes: The condition parser is used to receive feedback from the target simulation module and a preset structural pattern, quantize the optimization objective into a computable loss function, and quantize the structural pattern into hard constraints or soft penalty terms. The structural fine-tuning unit is used to guide point mutations in the amino acid sequence or fine-tuning of local conformations for protein antigens using gradient information based on a protein language model or a three-dimensional convolutional neural network, in order to optimize the loss function. The modification unit is used to simulate modifications to the length, branching, or linkage of the sugar chain for polysaccharide antigens, based on chemical synthesis or enzymatic digestion rules, to evaluate their impact on immunogenicity and stability, and to select positive modifications.

9. The vaccine target screening system based on computational model simulation according to claim 8, characterized in that, The structural fine-tuning unit, based on gradient information from a protein language model or a three-dimensional convolutional neural network, guides point mutations in the amino acid sequence or fine-tunes of local conformations, including: The original antigen sequence is input into a pre-trained protein language model to obtain the full-sequence context embedding vector for each amino acid position. Calculate the gradient of the model output for each amino acid type at each position under the loss function; Based on the gradient information, select several positions with the most significant gradient descent direction, replace the original amino acids with candidate amino acids indicated by the gradient and that conform to chemical rules and structural constraints, and generate mutant sequences. The mutant sequence is returned to the feature calculation module for rapid simulation to evaluate the optimization effect.

10. The vaccine target screening system based on computational model simulation according to claim 1, characterized in that, The target simulation module includes: The rapid screening unit is used to simulate the binding strength and aggregation tendency between high-potential target variants and pre-set adjuvant and stabilizer molecules based on a simplified model of intermolecular interaction potential, and to predict the physical compatibility in the formulation. The behavioral simulation unit is used to establish a coarse-grained model of ice crystal growth, water migration and freeze-dried cake structure during freeze-drying, and to simulate the glass transition temperature changes and resolution rate differences of different target variants in the presence of freeze-drying protectant. The correlation prediction unit is used to perform correlation analysis between the outputs of the rapid screening unit and the behavioral simulation unit and the experience database to predict the long-term storage stability level and degradation path of the final product, and to feed back the predicted weak links as new constraints to the iterative optimizer.

Images