Evaluation Methods and Systems for Predicting the Immunogenicity of Vaccine Adjuvants Using Machine Learning

Abstract

This invention discloses an assessment method and system for predicting the immunogenicity of vaccine adjuvants using machine learning, relating to the field of vaccine adjuvant R&D technology. The method includes dividing antigen homology clusters and adjuvant backbone clusters based on candidate sample units, and adding cluster labels to the candidate sample units to obtain standard sample units. This invention jointly models adjuvant structural features with antigen protein sequence features and introduces the consistency of immune response stages into the historical experimental sample screening process. This expands the selection criteria for historical reference samples from static similarity to matching immune evolution stages. Based on local adaptive measurement and local linear reconstruction, the main predicted immunogenicity value is structurally corrected, enabling the influence of historical samples on current candidate samples to adaptively change with the degree of local structural adaptation. This makes the immunogenicity classification and adjuvant selection more targeted and stable.

Description

Technical Field

[0001] This invention relates to the field of vaccine adjuvant development technology, and in particular to an evaluation method and system for predicting the immunogenicity of vaccine adjuvants using machine learning. Background Technology

[0002] As a key functional component regulating the intensity and type of immune response, vaccine adjuvants are being transformed from a traditional trial-and-error model relying on animal experiments and empirical formulation optimization to a predictive model based on multi-source data modeling, driven by both bioinformatics and machine learning technologies.

[0003] Existing machine learning-based vaccine adjuvant immunogenicity prediction methods treat samples as static objects for unified modeling, lacking the combined use of adjuvant structural features and antigen protein sequence features, as well as a historical sample screening and correction mechanism based on the consistency of the immune response process stages. This makes it difficult for historical experimental samples to form an effective correspondence with current candidate samples, affecting the correspondence between predicted immunogenicity values ​​and the strength of the actual immune response, and weakening the stability of candidate adjuvant screening results. Summary of the Invention

[0004] In a first aspect, the present invention provides an assessment method for predicting the immunogenicity of vaccine adjuvants using machine learning, comprising, Obtain raw data from vaccine experiments, construct candidate sample units and historical candidate sample units, perform antigen homology cluster division and adjuvant backbone cluster division based on candidate sample units, and add cluster labels to candidate sample units to obtain standard sample units; Descriptor vectors, structural fingerprint vectors, and protein sequence feature vectors are extracted from standard sample units to construct modeling feature vectors. These vectors are then used to make predictions through a machine learning model to obtain predicted immunogenicity values. Finally, the main predicted immunogenicity value is obtained by summing and averaging these values. Based on candidate sample units and historical experimental subsets, the local evolution rate, cumulative evolution amount, and immune dynamic phase are calculated respectively. A standard phase anchor point set is constructed, a standard phase anchor point alignment state sequence is generated, the phase synchronization difference throughout the process is calculated, and phase matching is performed to obtain the historical experimental subset. Based on candidate sample units and historical experimental subsets, local dispersion values, local metric coefficients, and local adaptive distances are calculated to screen local reconstruction neighborhoods, construct local reconstruction matrices, and solve local linear reconstruction weight vectors by least squares optimization. The modeling feature vectors of the local reconstruction neighborhoods are then reconstructed using weighted reconstruction. The residual intensity is calculated using the weighted quadratic norm and mapped to the neighborhood confidence coefficient. The predicted immunogenicity value is then weighted and corrected to obtain the corrected predicted immunogenicity value. Based on the predicted immunogenicity value, the consistency deviation is calculated, and the credibility of the candidate sample unit is calculated by combining the neighborhood credibility coefficient. The credibility score and the corrected predicted immunogenicity value are divided by rules to obtain the immunogenicity level.

[0005] As a preferred embodiment of the evaluation method for predicting the immunogenicity of vaccine adjuvants using machine learning as described in this invention, the step of extracting descriptor vectors, structural fingerprint vectors, and protein sequence feature vectors from standard sample units, constructing modeling feature vectors, performing prediction through a machine learning model to obtain predicted immunogenicity values, and obtaining the master predicted immunogenicity value by summing and averaging, includes: Descriptor vectors, structural fingerprint vectors, protein sequence feature vectors, and encoding vectors are extracted from standard sample units and concatenated to obtain the total feature vector. The total feature vector is then divided into training and validation sets. The features in the training set are filtered using GainRatio to obtain the modeling feature vector; The machine learning model is trained using a training set and a validation set. The modeling feature vector is then input into the machine learning model to obtain the predicted immunogenicity value for each machine learning model. The predicted immunogenicity values ​​are summed and averaged to obtain the master predicted immunogenicity value.

[0006] As a preferred embodiment of the evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in this invention, the step of calculating the phase synchronization difference throughout the entire process and performing phase matching to obtain a subset of historical experiments includes: Define a neighborhood size and calculate the validation error using a validation subset. Sum and average the validation errors of all validation subsets to obtain the comprehensive color difference function corresponding to the neighborhood size. Select the neighborhood size with the smallest comprehensive color difference function and set it as the optimal neighborhood size. ; For the For each candidate sample unit, the local evolution rate is calculated and accumulated to obtain the cumulative evolution amount, which is then used as a normalization benchmark to generate the immune dynamics phase. Based on immune dynamics phase, the number of standard phase anchors is calculated, a set of standard phase anchors is constructed, and standard phase anchors that meet the phase screening criteria are searched. The first standard phase anchor is then generated through linear interpolation. The alignment state vector of each candidate sample unit at the standard phase anchor point is calculated, and the phase synchronization difference throughout the process is calculated. The historical experimental subset was selected by using the phase synchronization difference throughout the process to screen historical candidate sample units.

[0007] As a preferred embodiment of the evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in this invention, the step of constructing a local reconstruction matrix and solving for the local linear reconstruction weight vector through least squares optimization includes: Based on historical experimental subsets and candidate sample units, local dispersion values ​​are calculated, the dimensionality of the statistical modeling feature vectors is determined, and the local dispersion values ​​across all dimensions are summed and averaged to obtain a locally stable baseline. Local metric coefficients are then calculated, and a locally adaptive metric matrix is ​​generated through matrix construction. The local adaptive distance is calculated and sorted in ascending order before screening. Each is set as a locally reconstructed neighborhood; Based on the local adaptive metric matrix and the local reconstruction neighborhood, a local reconstruction matrix is ​​constructed and the diagonal elements are summed and averaged to obtain the stable adjustment coefficient. The stable reconstruction matrix is ​​calculated, an optimization objective function is constructed, and the local linear reconstruction weight vector is solved.

[0008] As a preferred embodiment of the evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in this invention, the step of weighted reconstruction of the modeling feature vectors of the locally reconstructed neighborhood, and calculating the residual intensity using the weighted quadratic norm, includes: The modeling feature vectors of neighbor candidate sample units in the local reconstruction neighborhood are weighted and summed using the local linear reconstruction weight vector to obtain the neighborhood reconstruction vector. The modeling feature vector of the candidate sample unit is subtracted from the neighborhood reconstruction vector to obtain the reconstruction residual vector, and the residual strength is calculated.

[0009] As a preferred embodiment of the evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in this invention, wherein: the weighted correction of the predicted immunogenicity value to obtain the corrected predicted immunogenicity value includes: The residual intensity is used as a measure of sample structure consistency. It is mapped to the neighborhood confidence coefficient through a monotonically decreasing function. The main predicted immunogenicity value is then corrected to obtain the corrected predicted immunogenicity value.

[0010] As a preferred embodiment of the evaluation method for predicting the immunogenicity of vaccine adjuvants using machine learning as described in this invention, the step of performing rule-based classification on the confidence score and the corrected predicted immunogenicity value to obtain the immunogenicity level includes: The consistency deviation of the predicted immunogenicity value is calculated, and the credibility of the candidate sample unit is generated by combining the neighborhood credibility coefficient. Immunogenicity levels are classified using fixed rules based on confidence scores and corrected predicted immunogenicity values. Candidate sample units are extracted based on the immunogenicity levels to obtain the preferred adjuvant set.

[0011] As a preferred embodiment of the evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in this invention, the step of dividing antigen homology clusters and adjuvant backbone clusters based on candidate sample units, and adding cluster labels to the candidate sample units to obtain standard sample units, includes: Obtain raw data from vaccine trials, construct candidate sample units, construct historical candidate sample units according to the same field structure as the candidate sample units, and set observation points; The amino acid sequences of vaccine antigen proteins in the raw data of vaccine experiments were locally aligned using BLAST to identify antigen homology clusters. The Bemis–Murcko skeleton extraction method was used to define adjuvant skeleton clusters in the standardized structure of the raw vaccine experimental data; Cluster labels are added to candidate sample units to obtain standard sample units.

[0012] Secondly, the present invention provides an assessment system for predicting the immunogenicity of vaccine adjuvants using machine learning, comprising, The vaccine trial data construction module is used to obtain raw vaccine trial data through the API interface, construct candidate sample units, and construct historical candidate sample units according to the same field structure as the candidate sample units. The standard sample unit generation module is used to divide antigen homology clusters and adjuvant backbone clusters based on candidate sample units, and add cluster labels to candidate sample units to obtain standard sample units. The modeling feature construction module is used to extract descriptor vectors, structural fingerprint vectors, and protein sequence feature vectors from standard sample units, and perform joint modeling to obtain modeling feature vectors and construct a sample set. The immunogenicity prediction module is used to input a sample set into a machine learning model and output predicted immunogenicity values. The immunodynamics phase matching module is used to construct a standard set of phase anchor points based on candidate sample units and historical candidate sample units, calculate the phase synchronization difference throughout the process, and perform phase matching to obtain a subset of historical experiments. The local confidence correction module is used to screen local reconstruction neighborhoods based on candidate sample units and historical candidate sample units, construct a local reconstruction matrix, solve the local linear reconstruction weight vector by least squares optimization, perform weighted reconstruction on the modeling feature vector of the local reconstruction neighborhood, calculate the residual intensity by weighted quadratic norm and map it to neighborhood confidence coefficient, and perform weighted correction on the predicted immunogenicity value to obtain the corrected prediction value. The credibility assessment and grading module is used to calculate the consistency deviation based on the predicted immunogenicity value, calculate the credibility of candidate sample units by combining the neighborhood credibility coefficient, and perform rule-based grading on the credibility score and corrected predicted value to obtain the immunogenicity level.

[0013] The beneficial effects of this invention are as follows: This invention jointly models the adjuvant structural features and antigen protein sequence features, and introduces the consistency of the immune response process stages into the historical experimental sample screening process. This expands the selection criteria for historical reference samples from static similarity to matching the immune evolution stages. On this basis, based on local adaptive measurement and local linear reconstruction, the main predicted immunogenicity value is structurally corrected, so that the effect strength of historical samples on current candidate samples can adaptively change with the degree of local structural adaptation. Furthermore, by combining the consistency of multi-model outputs and the neighborhood confidence coefficient to form a confidence score, the immunogenicity level classification and the selection of optimal adjuvants have stronger targeting and stability. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of the evaluation method for predicting the immunogenicity of vaccine adjuvants using machine learning in Example 1.

[0016] Figure 2 This is a schematic diagram of the evaluation system for predicting the immunogenicity of vaccine adjuvants using machine learning in Example 1. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0018] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides an assessment method for predicting the immunogenicity of vaccine adjuvants using machine learning, comprising the following steps: S1. Obtain the original data of the vaccine experiment, construct candidate sample units and historical candidate sample units, divide antigen homology clusters and adjuvant backbone clusters based on the candidate sample units, and add cluster labels to the candidate sample units to obtain standard sample units; Specifically, antigen homology clusters and adjuvant backbone clusters are segmented based on candidate sample units, and cluster labels are added to the candidate sample units to obtain standard sample units, including: Obtain the raw data from the vaccine experiment, construct candidate sample units, and construct historical candidate sample units according to the same field structure as the candidate sample units; If the original sampling points of the historical candidate sample unit are inconsistent with the preset observation time series, linear interpolation is used to generate corresponding observation values ​​for the target observation points between two adjacent original sampling points. If a target preset observation point has only one side of valid sampling points, or if the target preset observation point does not exist before or after it, and interpolation cannot be performed through adjacent sampling, then the historical candidate sample unit is deleted. The aforementioned preset observation time series refers to a set of multiple fixed observation points pre-set relative to the time when the last immunization was completed; The sources of the aforementioned raw data from vaccine experiments include, for example, experimental databases, literature repositories, and internal research records; The raw data from the aforementioned vaccine trials include adjuvant data, antigen data, immunization configuration data, and immunization outcome data. Each type of data includes its corresponding trial number. The above adjuvant data includes standardized structural formulas and carrier types; The aforementioned antigen data includes the amino acid sequence of the vaccine antigen protein; The above immunization configuration data includes the route of administration and animal species; The above-mentioned immune results data include the geometric mean titer of IgG, the titer of neutralizing antibodies, the number of IFN-γ ELISpot spots, and the proportion of memory T cells under the preset observation time series; The aforementioned candidate sample units consist of an adjuvant object (a specific data point in the adjuvant data), an antigen object (a specific data point in the antigen data), a set of immunization configurations, and immunization results under a preset observation time series, all under the same experiment number. Each candidate sample unit is assigned a unique number. , For sample index; The amino acid sequences of vaccine antigen proteins are compared using BLAST (Basic Local Alignment Search Tool) to calculate the similarity measure between any two sequences. When the similarity measure is greater than or equal to the comparison threshold, the two sequences are marked as the same antigen homology cluster. The aforementioned comparison threshold refers to sorting all similarity measures in ascending order and setting the lower quartile as the comparison threshold. The Bemis–Murcko skeleton extraction method was used to extract the skeleton from the standardized structural formulas, and adjuvant molecules with the same skeleton were labeled as the same adjuvant skeleton cluster. Cluster labels are added to candidate sample units to obtain standard sample units.

[0019] S2. Extract descriptor vectors, structural fingerprint vectors, and protein sequence feature vectors from standard sample units, construct modeling feature vectors, and use machine learning models to make predictions to obtain predicted immunogenicity values. By summing and averaging, the main predicted immunogenicity value is obtained. Specifically, descriptor vectors, structural fingerprint vectors, and protein sequence feature vectors are extracted from standard sample units to construct modeling feature vectors. These are then used in a machine learning model to predict immunogenicity values. Finally, the master predicted immunogenicity value is obtained by summing and averaging these values, including: Based on the standardized structure in the standard sample unit, extract the descriptor vector and the structural fingerprint vector; The aforementioned descriptor vector includes molecular weight, number of hydrogen bond donors, number of hydrogen bond acceptors, topological polar surface area, number of rotatable bonds, number of ring structures, and hydrophobicity index. The molecular weights mentioned above are obtained by summing the standard relative atomic masses of all atoms. The number of hydrogen bond donors is the number of hydroxyl and amino groups. The number of hydrogen bond acceptors is the number of oxygen and nitrogen atoms that can accept hydrogen bonds. The topological polar surface area is the sum of the TPSA contributions of polar atoms. The number of rotatable bonds is the number of single bonds and non-cyclic structures. The number of ring structures is obtained by molecular ring detection algorithms, such as the DFS backtracking method and the Hopcroft-Tarjan algorithm. The hydrophobicity index is calculated using existing LogP methods, such as Wildman–Crippen LogP and Wildman–Crippen lipid-water partition coefficient. The above TPSA contribution values ​​were obtained by looking up the atomic contribution table using Topological Polar Surface Area. The aforementioned polar atoms include oxygen and nitrogen; The aforementioned structural fingerprint vector refers to the result obtained by using the Morgan fingerprint algorithm with parameters fixed at a radius of 2 and a bit depth of 1024 on the standardized structure. Based on the amino acid sequence of the vaccine antigen protein in the standard sample unit, the sequence matrix is ​​extracted, and the sequence matrix is ​​subjected to ACC transformation, Auto-Cross Covariance, and Auto-Cross Covariance to obtain the autocorrelation value. The autocorrelation values ​​are arranged in ascending order of lag order to obtain the protein sequence feature vector. The sequence matrix mentioned above refers to the 5-dimensional E-descriptor vector obtained by encoding each amino acid in the amino acid sequence of the vaccine antigen protein using E-descriptor and sorting them according to their sequence positions. The above five physicochemical dimensions are the standard amino acid physicochemical properties represented by the existing E-descriptor encoding system, including hydrophobicity, molecular size value, α-helix tendency value, residue index, and β-sheet tendency value. The lag order mentioned above is a positive integer and needs to meet the basic constraints of the ACC method, namely, the lag order must be less than the length of the amino acid sequence of the vaccine antigen protein; The immunogenicity of vaccine adjuvants is not only determined by the adjuvant itself, but also affected by the background of antigen proteins. Using only adjuvant molecular features will lose the constraints of antigen sequence on immune recognition and adjuvant enhancement boundaries. Using only protein features cannot express the chemical driving force of adjuvants. Therefore, dual-channel modeling is necessary to retain both the chemical differences of small molecule, lipid and polysaccharide adjuvants and the modulating effect of antigen sequence differences on adjuvant performance, so that the model is more in line with the actual development logic of vaccine formulation. This effectively solves the problem of feature alignment difficulties between antigen sequences and adjuvant structures caused by different data modalities in vaccine development; The immunogenicity data in the standard unit samples were normalized using Min–Max methods to obtain normalized immunogenicity data. The normalized immunogenicity data were then summed and averaged to obtain the continuous immunogenicity score. ; The descriptor vector, structural fingerprint vector, protein sequence feature vector, and immunization configuration data are concatenated end to end to obtain the total feature vector. For the total eigenvector Consistency screening was performed, combined with continuous immunogenicity scores. Construct a sample set , The total number of samples; The above consistency screening refers to consistent feature dimensions, consistent feature order, and no missing values. If the dimensions are inconsistent, the inconsistent features are deleted. The feature order refers to the order of descriptor vector, structural fingerprint vector, protein sequence feature vector, and coding vector. If there are missing values, they are filled using the mean. Dividing the sample set into training and validation sets refers to using a random partitioning method, such as dividing it into a 70% training set and a 30% validation set. Each sample in the sample set is assigned a cluster identifier, and all samples are divided into clusters based on the cluster identifiers. The training and validation sets are then divided into clusters within each cluster. The cluster identifiers mentioned above refer to the antigen homology cluster number and the adjuvant backbone cluster number; Using features from the training set as input and continuous immunogenicity scores from the training set as supervision targets, the gain ratio of each feature is obtained through feature filtering using the GainRatio method. The features are then sorted in descending order of gain ratio, retaining the top-ranked features. The feature index corresponding to each feature is used to extract features from the total feature vector of the training set, validation set, and candidate sample units to be predicted, resulting in the modeling feature vector. ; After feature selection is completed, the training and validation sets will not be re-divided. The modeling feature vector obtained in this step The static feature vector of the candidate sample unit is used to reflect the change process of the immune response of the candidate sample unit under different observation points. For each observation point of the candidate sample unit in the preset observation time series, the normalized immune result data of that point is read and concatenated with the modeling feature vector to obtain the modeling state vector. The modeling state vector is then arranged in the order of the preset observation time series to obtain the modeling state vector sequence. This sequence is used for the local evolution speed, cumulative evolution amount, immune dynamic phase, standard phase anchor point alignment state sequence, and phase synchronization difference throughout the process in subsequent steps. The same processing is performed on historical candidate sample units. The above To preserve the feature ratio, the feature gain ratio is adaptively set to meet the following requirements. The formula is: ; ; in, For the front The cumulative contribution rate of each feature For the first The gain ratio of each feature The total number of features, The cumulative contribution rate threshold; The aforementioned cumulative contribution threshold refers to sorting the cumulative contribution rates according to the gain ratio to obtain a cumulative contribution rate sequence, performing second-order difference on the cumulative contribution rate sequence, and selecting the cumulative contribution rate corresponding to the largest absolute value of the second-order difference as the cumulative contribution threshold. The position of the largest second-order difference absolute value in the cumulative contribution rate sequence represents the inflection point where the marginal contribution of the feature changes from rapid increase to significant slowdown. Retaining features before this position can balance the preservation of effective information and the removal of low-contribution features. Each machine learning model is trained independently using the training set, and the prediction error of each set of model parameters is calculated using the mean squared error formula using the validation set. The model parameters with the smallest prediction error are selected for use to obtain the trained machine learning model. The modeling feature vector is input into each trained machine learning model for independent prediction to obtain the predicted immunogenicity value of each machine learning model. The predicted immunogenicity values ​​of all machine learning models are summed and averaged to obtain the main predicted immunogenicity value. The aforementioned machine learning models include random forest regression model, XGBoost regression model, and multilayer perceptron regression model; The above model parameters refer to the number of decision trees and maximum depth for random forests, the learning rate, tree depth and subsampling ratio for XGBoost regression models, and the hidden layer structure and learning rate for multilayer perceptron regression models. Using a configuration-independent model system, rather than modeling each inoculation condition separately, can avoid the sample being fragmented. Using a multi-model ensemble with equal weights instead of a single model can reduce the bias of a single algorithm, enabling the model to handle diverse experimental configurations while reducing the sensitivity of a single model to sample distribution, thereby improving the predictive stability for novel adjuvants, antigens, and inoculation combinations.

[0020] S3. Based on candidate sample units and historical experimental subsets, calculate local evolution rate, cumulative evolution amount and immune dynamic phase respectively, construct standard phase anchor point set, generate standard phase anchor point alignment state sequence, calculate phase synchronization difference throughout the process, and perform phase matching to obtain historical experimental subset. Specifically, the phase synchronization difference throughout the entire process is calculated, and phase matching is performed to obtain a subset of historical experiments, including: Define a neighborhood size and calculate the validation error using a validation subset. Sum and average the validation errors of all validation subsets to obtain the comprehensive color difference function corresponding to the neighborhood size. Select the neighborhood size with the smallest comprehensive color difference function and set it as the optimal neighborhood size. The number of neighbors used for local linear reconstruction is given by the formula: ; in, To verify the error, To verify the sample size of the subset, For neighborhood scale, and For the first Continuous immunogenicity scores and corrected predicted immunogenicity values ​​for each validation sample; Subsequent training and validation sets are used to complete the training and validation of the machine learning model and the generation of the master predictor immunogenicity value. The above is used to determine the optimal neighborhood size. The validation subset is independent of the training set and validation set settings; The above-mentioned validation subset was obtained separately from historical candidate sample units that meet the criteria of consistent animal species, consistent antigen homology clusters, consistent route of administration, or consistent vector category through stratified random partitioning. The stratification criteria include at least two of the following: continuous immunogenicity score interval, antigen homology clusters, and adjuvant skeleton category. The upper limit of the aforementioned neighborhood size The neighborhood size is set to the square root of the number of historical candidate samples, rounded down. ; For the Each candidate sample unit is located at the observation point. With observation point Between these points, the local evolution rate is calculated using the following formula: ; in, For the first The local evolution rate of each candidate sample unit For the first Each candidate sample unit at the observation point The modeling state vector, For the first Each candidate sample unit at the observation point The modeling state vector; By calculating the intensity of state changes between adjacent observation points, the local evolution rate is obtained, so that the speed of immune response is characterized by actual state changes rather than absolute time length. This overcomes the problems of existing methods that use observation dates to replace the immune process and make it difficult to identify differences in the evolution rhythm of different samples. It is beneficial to establish a dynamic correspondence between samples based on the actual degree of immune response rather than the surface time synchronization relationship. For the Each candidate sample unit at the observation point The local evolution rates are accumulated to obtain the cumulative evolution amount, which is then used as a normalization benchmark to generate the immune dynamics phase, as shown in the formula: ; in, For the first Immunokinetic phase of each candidate sample unit For the first The cumulative evolution of each candidate sample unit. To prevent extremely small positive numbers with a denominator of 0; The local evolution rates at each time point are summed to form a cumulative evolution amount, and then normalized to generate the original phase trajectory. This allows each sample to obtain a continuous phase sequence describing the progression of its immune response, no longer relying on a single endpoint value or absolute number of days to characterize the stage. This overcomes the problem of existing methods lacking full-process stage coordinates and provides a continuous and comparable dynamic scale for subsequent phase alignment and synchronization difference calculation between different samples. Based on the observation points, the number of standard phase anchor points is calculated, and a set of standard phase anchor points is constructed, representing the sequence from the first standard phase anchor point to the [nth / nth] standard phase anchor point. The standard phase anchor points are sorted and obtained using the following formula: ; ; in, For the first One standard phase anchor point, The number of standard phase anchor points, This represents the number of historical candidate sample units. For the first The total number of observation points for each historical candidate sample unit; A unified set of standard phase anchor points is constructed to map samples with different numbers of observation points and inconsistent time distributions to the same set of standard phase positions. This solves the problem that it is difficult to establish a step-by-step comparison relationship between different experiments due to inconsistent observation rhythms in existing methods. This enables all subsequent samples to be aligned, compared and screened under a consistent immune evolution progress coordinate, thereby improving the dynamic comparability of samples across experiments. Search for standard phase anchor points that meet the phase screening criteria, and generate the first phase anchor point through linear interpolation. The alignment state vector of each candidate sample unit at the standard phase anchor point is used to generate a standard phase anchor point alignment state sequence. The same operation is performed on historical candidate sample units to synchronously generate historical standard phase anchor point alignment state sequences. The formula is as follows: ; ; in, For the first The number of candidate sample units is less than or equal to The most recent previous observation point number, For the first The number of candidate sample units is greater than or equal to The number of the most recent observation point, For the first Alignment state vector of candidate sample units; By interpolating the state vector at the standard phase anchor point, a phase-aligned state vector sequence is formed, so that different samples are no longer mechanically compared according to the day or the sampling point, but compared according to the position of the same immune evolution progress. This improves the stage misalignment problem caused by direct comparison based on absolute time points in existing methods, and improves the matching accuracy of historical samples at the dynamic level and the reliability of subsequent immunogenicity correction. Define the phase synchronization difference throughout the process and construct the phase synchronization weight, using the following formula: ; ; ; in, For the first The candidate sample unit and the first Phase synchronization differences throughout the entire process of each historical candidate sample unit For the first The candidate sample unit and the first Phase synchronization weights of historical candidate sample units, For phase difference scale parameters, For the first The number of samples in the historical experimental subset of each candidate sample unit; The mean state difference between candidate samples and historical samples is calculated at all standard phase anchor points to obtain the phase synchronization difference throughout the process. This makes the similarity of historical samples no longer depend solely on static features or the final immune result, but rather on whether the entire immune response process is synchronized. This compensates for the shortcomings of existing methods that neglect the consistency of evolutionary paths, improves the ability of historical sample screening to characterize the similarity of the real immune process, and reduces the probability of process mismatched samples entering the subsequent correction stage. The aforementioned historical experimental subset refers to historical candidate sample units whose phase synchronization difference throughout the entire process is less than or equal to the phase synchronization screening threshold; If the historical experiment subset obtained according to the phase synchronization screening threshold is empty, then select the first P historical candidate sample units in ascending order of phase synchronization difference throughout the process to form the historical experiment subset. The above P is the square root of the number of historical candidate sample units and rounded down, so that the number of samples in the historical experimental subset adapts to the overall size of the historical candidate sample units, and reduces the use of samples with large phase synchronization differences throughout the process while ensuring the number of local reference samples. The aforementioned phase synchronization screening threshold is determined based on the quantile of the phase synchronization difference throughout the entire process, for example, taking the 75th quantile of all phase synchronization differences throughout the entire process; By mapping the phase synchronization difference throughout the process to phase synchronization weights, different historical samples entering the phase synchronization historical experiment subset can be continuously weighted according to their dynamic synchronization with the current candidate samples, rather than simply being retained or removed. This addresses the problem of existing methods having a coarse characterization of the contribution of historical samples, allowing historical samples that are more consistent with the immune evolution process of the current samples to play a greater role in subsequent local distance correction and local reconstruction, thereby improving the precision of historical information utilization.

[0021] S4. Based on candidate sample units and historical experimental subsets, calculate local dispersion values, local metric coefficients, and local adaptive distances to screen local reconstruction neighborhoods, construct local reconstruction matrices, and solve for local linear reconstruction weight vectors through least squares optimization. Perform weighted reconstruction on the modeling feature vectors of local reconstruction neighborhoods, calculate residual strength through weighted quadratic norm, and map it to neighborhood confidence coefficients. Perform weighted correction on predicted immunogenicity values ​​to obtain corrected predicted immunogenicity values. Specifically, a local reconstruction matrix is ​​constructed, and the local linear reconstruction weight vector is solved by least squares optimization, including: Based on historical experimental subsets and candidate sample units, the local dispersion value is calculated using the following formula: ; in, For the first The candidate sample unit in the th ... Local dispersion values ​​across each feature dimension The number of sample units in the historical experiment subset. For the first The candidate sample unit in the th ... The values ​​taken on each feature dimension For the first The historical candidate sample unit in the first Values ​​can be taken in each feature dimension; The values ​​mentioned above in terms of dimension refer to the values ​​of the first, second, and third dimensions of the modeling feature vector. One component; The number of dimensions of the statistical modeling feature vector is calculated by summing and averaging the local dispersion values ​​of all dimensions to obtain a locally stable baseline quantity. A local metric coefficient is then calculated to enhance the ability to distinguish dimensions with more significant differences in subsequent distance calculations. The formula is as follows: ; in, For the first The candidate sample unit in the th ... Local metric coefficients on each feature dimension To prevent tiny positive numbers with a denominator of 0, The number of dimensions, For the first The candidate sample unit in the th ... Local dispersion values ​​across each feature dimension; The local metric coefficients are sequentially placed on the main diagonal of the matrix, with all off-diagonal elements set to 0, resulting in a local adaptive metric matrix. Based on this matrix, the ... The candidate sample unit and the first The local adaptive distance between historical candidate sample units is given by the formula: ; in, and For the first The candidate sample unit and the first Modeling feature vectors of historical candidate sample units, For the first The candidate sample unit and the first Local adaptive distance of historical candidate sample units, For transpose, For the first Local adaptive metric matrix of candidate sample units; The key features affected by the current adjuvant sample are not the same under different experimental backgrounds. Therefore, it is necessary to re-evaluate the similarity between the current sample and the historical sample according to the current local experimental environment. This will enable us to more accurately identify the historical experiments that are truly similar to the current adjuvant sample and improve the relevance of subsequent neighborhood priors. Candidate sample units are sorted in ascending order according to the magnitude of their local adaptive distance before screening. Each is set as a locally reconstructed neighborhood; The number of historical neighbors for the current adjuvant sample should not be fixed, but should be automatically determined based on the actual correction effect achievable in historical experiments. This will allow the most suitable historical reference range to be selected for the current adjuvant sample, avoiding immunogenicity correction bias caused by a neighborhood that is too narrow or too wide.

[0022] Based on the local adaptive metric matrix and the local reconstruction neighborhood, the local reconstruction matrix is ​​constructed as follows: ; ; in, For the first The candidate sample unit in the initial local reconstruction matrix is ​​the th . Line 1 Column elements, For the first The modeling feature vector of each candidate sample unit, and For the local reconstruction of the neighborhood of the first The and the first The modeling feature vector of each neighbor candidate sample unit, For the first The nth candidate sample unit in the local reconstruction matrix Line 1 Column elements, and For the first The first historical candidate sample unit and the first The historical candidate sample unit relative to the first Phase synchronization weights for each candidate sample unit; By using phase synchronization weights to weight the local reconstruction matrix, the solution of local reconstruction weights is constrained not only by the static feature proximity relationship but also by the degree of synchronization of the immune dynamics stage. This directly embeds dynamic synchronization into the linear reconstruction process of historical samples to current samples, overcoming the problem that existing methods only reconstruct based on static neighborhoods and have difficulty suppressing cross-stage sample interference. This allows historical samples with more consistent phases to make a greater contribution to the formation of the corrected predicted immunogenicity value, and improves the fit of the correction results to the real immune response stage. The stability adjustment coefficient is obtained by summing and averaging the diagonal elements of the local reconstruction matrix within the local reconstruction neighborhood. The stable reconstruction matrix is ​​then calculated, an optimization objective function is constructed, and the local linear reconstruction weight vector is solved using the following formula: ; ; ; in, For the first Local linear reconstruction weight vector of candidate sample units, To reconstruct a stable matrix, For the first The length of each candidate sample unit is A column vector of all 1s represents a column vector where each element is 1. For the first The local reconstruction matrix of each candidate sample unit. To stabilize the adjustment coefficient, It is the identity matrix. The number of samples in the local reconstruction neighborhood. It is the set of real numbers; The constraints for solving the local linear reconstruction weight vector are as follows: The 1 on the right side of the equation is the number 1, which is used to ensure that the reconstructed weights satisfy affine invariance. The aforementioned identity matrix refers to a square matrix whose main diagonal elements are 1s and all other elements are 0s.

[0023] Furthermore, the modeling feature vectors of the locally reconstructed neighborhood are reconstructed using weighted quadratic norm, and the residual strength is calculated, including: The modeling feature vectors of neighbor candidate sample units in the local reconstruction neighborhood are weighted and summed using the local linear reconstruction weight vector to obtain the neighborhood reconstruction vector. Before correcting the master predicted immunogenicity value, it is necessary to evaluate the structural consistency of candidate sample units in the feature space. This includes subtracting the neighborhood reconstruction vector from the modeling feature vector of the candidate sample unit to obtain the reconstruction residual vector, and calculating the residual strength using the following formula: ; in, For the first The residual strength of each candidate sample unit represents the degree of deviation of the candidate sample unit from the historical sample manifold structure in the feature space. For the first The reconstructed residual vector of each candidate sample unit; If the residual strength is small, it means that the candidate sample unit can be linearly represented by the neighboring samples and belongs to the local structure range of the historical sample distribution. If the residual strength is large, it means that the candidate sample unit cannot be effectively linearly reconstructed by the neighboring samples and deviates from the historical sample distribution structure. Before using historical experiments to correct the current adjuvant sample, it is necessary to obtain the historical reference conclusions and their fit with the current sample. This provides both a historical prior immunogenicity reference value for the current adjuvant sample and a basis for subsequent judgment on the reliability of the reference value.

[0024] Furthermore, a weighted correction is applied to the predicted immunogenicity value to obtain the corrected predicted immunogenicity value, including: Using the residual intensity as a measure of sample structural consistency, it is mapped to a neighborhood confidence coefficient through a monotonically decreasing function. This coefficient is then used to correct the main predicted immunogenicity value, yielding the corrected predicted immunogenicity value, as shown in the formula: ; ; in, For the first The corrected predicted immunogenicity value of each candidate sample unit. For the first The neighborhood confidence coefficient of each candidate sample unit. For the first The principal predicted immunogenicity value for each candidate sample unit. For the first The neighborhood prior values ​​of each candidate sample unit; The aforementioned monotonically decreasing function is characterized by its simple calculation and stable value of continuous differentiable trace. The neighborhood confidence coefficient is calculated by the monotonically decreasing function, so that the smaller the residual strength, the larger the neighborhood confidence coefficient, indicating that the neighborhood prior score reference is high, and vice versa. The structural consistency information in the feature space is mapped to the fusion weight in the score space, realizing the adaptive coupling between the feature structure and the score response. The aforementioned neighborhood prior value is obtained by weighted summation of the continuous immunity scores of neighbor candidate sample units in the locally reconstructed neighborhood and the local linear reconstruction weight vector; Some adjuvant samples are better suited to be corrected by historical experiments, while others should retain the judgment of the master model. Therefore, the correction strength should change automatically with the degree of local matching, so as to retain the generalization ability of the master model to new adjuvant samples while utilizing historical experimental experience, thereby improving the practicality of immunogenicity prediction results. A closed-loop feedback loop from features to results was established, enabling the model to automatically learn and strengthen feature dimensions that are strongly correlated with high immunogenicity, while weakening irrelevant dimensions, thereby improving the prediction confidence of the immunogenicity of candidate adjuvants.

[0025] S5. Based on the predicted immunogenicity value, calculate the consistency deviation, combine the neighborhood confidence coefficient to calculate the confidence of the candidate sample unit, and perform rule division on the confidence score and the corrected predicted immunogenicity value to obtain the immunogenicity level. Specifically, the confidence score and corrected predicted immunogenicity value are divided according to rules to obtain the immunogenicity level, including: The consistency deviation for predicting immunogenicity values ​​is calculated using the following formula: ; in, For the first Consistency deviation of each candidate sample unit , as well as For the first The predicted immunogenicity value of each candidate sample unit, where 3 represents the number of machine learning models used; The confidence score of candidate sample units, representing the true immunogenicity of the candidate sample, is calculated using the following formula: ; in, For the first The credibility of each candidate sample unit It is a natural exponential function.

[0026] Immunogenicity levels are classified using fixed rules based on confidence scores and corrected predicted immunogenicity values ​​to generate a preferred set of adjuvants; The above fixed rules include setting the immunogenicity level to 2 if the corrected predicted immunogenicity value is greater than or equal to the high correction threshold and the confidence score is greater than or equal to the confidence threshold; setting it to 1 if the corrected predicted immunogenicity value is greater than or equal to the low correction threshold and the confidence score is greater than or equal to the confidence threshold; otherwise, setting it to 0. The aforementioned preferred adjuvant set refers to all candidate sample units with a level of 2; The aforementioned high correction threshold is set based on the 75th percentile of the predicted immunogenicity value in the training set, used to screen out candidate adjuvants with the highest immunogenicity scores as priority recommendations. The low correction threshold is set based on the median of the predicted immunogenicity value in the training set, used to distinguish between moderate and low immunogenicity, retaining potential candidate adjuvants for further evaluation. The confidence threshold is selected based on the validation set optimization, choosing the value that minimizes the validation error, balancing screening efficiency and misjudgment risk, and ensuring that the output candidate adjuvants have high predictive reliability.

[0027] This embodiment also provides an assessment system for predicting the immunogenicity of vaccine adjuvants using machine learning, including: The vaccine trial data construction module is used to obtain raw vaccine trial data through the API interface, construct candidate sample units, and construct historical candidate sample units according to the same field structure as the candidate sample units. The standard sample unit generation module is used to divide antigen homology clusters and adjuvant backbone clusters based on candidate sample units, and add cluster labels to candidate sample units to obtain standard sample units. The modeling feature construction module is used to extract descriptor vectors, structural fingerprint vectors, and protein sequence feature vectors from standard sample units, and perform joint modeling to obtain modeling feature vectors and construct a sample set. The immunogenicity prediction module is used to input a sample set into a machine learning model and output predicted immunogenicity values. The immunodynamics phase matching module is used to construct a standard set of phase anchor points based on candidate sample units and historical candidate sample units, calculate the phase synchronization difference throughout the process, and perform phase matching to obtain a subset of historical experiments. The local confidence correction module is used to screen local reconstruction neighborhoods based on candidate sample units and historical candidate sample units, construct a local reconstruction matrix, solve the local linear reconstruction weight vector by least squares optimization, perform weighted reconstruction on the modeling feature vector of the local reconstruction neighborhood, calculate the residual intensity by weighted quadratic norm and map it to neighborhood confidence coefficient, and perform weighted correction on the predicted immunogenicity value to obtain the corrected prediction value. The credibility assessment and grading module is used to calculate the consistency deviation based on the predicted immunogenicity value, calculate the credibility of candidate sample units by combining the neighborhood credibility coefficient, and perform rule-based grading on the credibility score and corrected predicted value to obtain the immunogenicity level.

[0028] In summary, this invention jointly models adjuvant structural features and antigen protein sequence features, and introduces the consistency of immune response stages into the historical experimental sample screening process. This expands the selection criteria for historical reference samples from static similarity to matching immune evolution stages. Based on this, structural correction is performed on the master predicted immunogenicity value using local adaptive measurement and local linear reconstruction. This allows the influence of historical samples on current candidate samples to adaptively change with the degree of local structural adaptation. Furthermore, by combining the consistency of multi-model outputs and the neighborhood confidence coefficient to form a confidence score, the immunogenicity level classification and adjuvant selection have stronger targeting and stability.

[0029] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for assessing the immunogenicity of vaccine adjuvants using machine learning, characterized in that, include: Obtain raw data from vaccine experiments, construct candidate sample units and historical candidate sample units, perform antigen homology cluster division and adjuvant backbone cluster division based on candidate sample units, and add cluster labels to candidate sample units to obtain standard sample units; Descriptor vectors, structural fingerprint vectors, and protein sequence feature vectors are extracted from standard sample units to construct modeling feature vectors. These vectors are then used to make predictions through a machine learning model to obtain predicted immunogenicity values. Finally, the main predicted immunogenicity value is obtained by summing and averaging these values. The aforementioned machine learning models include random forest regression models, XGBoost regression models, and multilayer perceptron regression models; Based on candidate sample units and historical experimental subsets, the local evolution rate, cumulative evolution amount, and immune dynamic phase are calculated respectively. A standard phase anchor point set is constructed, a standard phase anchor point alignment state sequence is generated, the phase synchronization difference throughout the process is calculated, and phase matching is performed to obtain the historical experimental subset. Based on candidate sample units and historical experimental subsets, local dispersion values, local metric coefficients, and local adaptive distances are calculated to screen local reconstruction neighborhoods, construct local reconstruction matrices, and solve local linear reconstruction weight vectors by least squares optimization. The modeling feature vectors of the local reconstruction neighborhoods are then reconstructed using weighted reconstruction. The residual intensity is calculated using the weighted quadratic norm and mapped to the neighborhood confidence coefficient. The predicted immunogenicity value is then weighted and corrected to obtain the corrected predicted immunogenicity value. Based on the predicted immunogenicity value, the consistency deviation is calculated, and the credibility of the candidate sample unit is calculated by combining the neighborhood credibility coefficient. The credibility score and the corrected predicted immunogenicity value are divided by rules to obtain the immunogenicity level.

2. The evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in claim 1, characterized in that: The process involves extracting descriptor vectors, structural fingerprint vectors, and protein sequence feature vectors from standard sample units to construct modeling feature vectors. These are then used in a machine learning model to predict immunogenicity values. Finally, the main predicted immunogenicity value is obtained by summing and averaging these values. Descriptor vectors, structural fingerprint vectors, protein sequence feature vectors, and encoding vectors are extracted from standard sample units and concatenated to obtain the total feature vector. The total feature vector is then divided into training and validation sets. The features in the training set are filtered using GainRatio to obtain the modeling feature vector; The machine learning model is trained using a training set and a validation set. The modeling feature vector is then input into the machine learning model to obtain the predicted immunogenicity value for each machine learning model. The predicted immunogenicity values ​​are summed and averaged to obtain the master predicted immunogenicity value.

3. The evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in claim 2, characterized in that: The calculation of phase synchronization differences throughout the process, followed by phase matching, yields a subset of historical experiments, including: Define a neighborhood size and calculate the validation error using a validation subset. Sum and average the validation errors of all validation subsets to obtain the comprehensive color difference function corresponding to the neighborhood size. Select the neighborhood size with the smallest comprehensive color difference function and set it as the optimal neighborhood size. ; For the For each candidate sample unit, the local evolution rate is calculated and accumulated to obtain the cumulative evolution amount, which is then used as a normalization benchmark to generate the immune dynamics phase. Based on immune dynamics phase, the number of standard phase anchors is calculated, a set of standard phase anchors is constructed, and standard phase anchors that meet the phase screening criteria are searched. The first standard phase anchor is then generated through linear interpolation. The alignment state vector of each candidate sample unit at the standard phase anchor point is calculated, and the phase synchronization difference throughout the process is calculated. The historical experimental subset was selected by using the phase synchronization difference throughout the process to screen historical candidate sample units.

4. The evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in claim 3, characterized in that: The construction of the local reconstruction matrix and the solution of the local linear reconstruction weight vector by least squares optimization include: Based on historical experimental subsets and candidate sample units, local dispersion values ​​are calculated, the dimensionality of the statistical modeling feature vectors is determined, and the local dispersion values ​​across all dimensions are summed and averaged to obtain a locally stable baseline. Local metric coefficients are then calculated, and a locally adaptive metric matrix is ​​generated through matrix construction. The local adaptive distance is calculated and sorted in ascending order before screening. Each is set as a locally reconstructed neighborhood; Based on the local adaptive metric matrix and the local reconstruction neighborhood, a local reconstruction matrix is ​​constructed and the diagonal elements are summed and averaged to obtain the stable adjustment coefficient. The stable reconstruction matrix is ​​calculated, an optimization objective function is constructed, and the local linear reconstruction weight vector is solved.

5. The evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in claim 4, characterized in that: The weighted reconstruction of the modeling feature vectors of the local reconstruction neighborhood, and the calculation of the residual strength using the weighted quadratic norm, includes: The modeling feature vectors of neighbor candidate sample units in the local reconstruction neighborhood are weighted and summed using the local linear reconstruction weight vector to obtain the neighborhood reconstruction vector. The modeling feature vector of the candidate sample unit is subtracted from the neighborhood reconstruction vector to obtain the reconstruction residual vector, and the residual strength is calculated.

6. The evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in claim 5, characterized in that: The weighted correction of the predicted immunogenicity value to obtain the corrected predicted immunogenicity value includes: The residual intensity is used as a measure of sample structure consistency. It is mapped to the neighborhood confidence coefficient through a monotonically decreasing function. The main predicted immunogenicity value is then corrected to obtain the corrected predicted immunogenicity value.

7. The evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in claim 6, characterized in that: The process of dividing the confidence score and corrected predicted immunogenicity value into rules to obtain the immunogenicity level includes: The consistency deviation of the predicted immunogenicity value is calculated, and the credibility of the candidate sample unit is generated by combining the neighborhood credibility coefficient. Immunogenicity levels are classified using fixed rules based on confidence scores and corrected predicted immunogenicity values. Candidate sample units are extracted based on the immunogenicity levels to obtain the preferred adjuvant set.

8. The evaluation method for predicting vaccine adjuvant immunogenicity using machine learning as described in claim 7, characterized in that: The process of dividing antigen homology clusters and adjuvant backbone clusters based on candidate sample units, and adding cluster labels to candidate sample units to obtain standard sample units, includes: Obtain raw data from vaccine trials, construct candidate sample units, construct historical candidate sample units according to the same field structure as the candidate sample units, and set observation points; The amino acid sequences of vaccine antigen proteins in the raw data of vaccine experiments were locally aligned using BLAST to identify antigen homology clusters. The Bemis–Murcko skeleton extraction method was used to define adjuvant skeleton clusters in the standardized structure of the raw vaccine experimental data; Cluster labels are added to candidate sample units to obtain standard sample units.

9. An evaluation system for predicting the immunogenicity of vaccine adjuvants using machine learning, based on the evaluation method for predicting the immunogenicity of vaccine adjuvants using machine learning as described in any one of claims 1 to 8, characterized in that: include, Vaccine trial data construction module; Standard sample unit generation module; Modeling feature construction module; Immunogenicity prediction module; Immunokinetic phase-matching module; Locally reliable correction module; Credibility assessment and rating module.

Images