A new antigen immunogenicity prediction screening method based on multi-dimensional feature fusion

Abstract

The application discloses a neoantigen immunogenicity prediction screening method based on multi-dimensional feature fusion, relates to the fields of bioinformatics, computational immunology and neoantigen vaccine design, and comprises the following steps: obtaining MHC binding prediction results of mutant peptides and wild peptides, performing merging and cleaning, performing biological feasibility filtering, extracting affinity scores and mutant-to-wild ratios and calculating natural logarithms of the mutant-to-wild ratios, performing scoring by using a pre-constructed immunogenicity scoring function, outputting immunogenicity probabilities, performing rule correction according to the immunogenicity probabilities and the affinity scores, performing multi-dimensional feature aggregation on positive label candidate peptide segments to generate multi-dimensional aggregation labels, and outputting a multi-level candidate neoantigen list; the application simultaneously considers affinity changes and mutation effects by synergistic modeling of affinity and mutation effects, improves the accuracy, stability and interpretability of a neoantigen screening result, and provides an effective candidate screening scheme for personalized tumor vaccine development.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics, computational immunology, and neoantigen vaccine design technology, and in particular to a neoantigen immunogenicity prediction and screening method based on multi-dimensional feature fusion. Background Technology

[0002] Tumor neoantigens are specific peptides generated by somatic cell mutations. They can be presented by major histocompatibility complex (MHC) molecules and activate T-cell immune responses, making them core targets for cancer immunotherapy and personalized vaccine design. Accurately predicting the immunogenicity of neoantigens is crucial for screening candidate vaccines and designing personalized immunotherapy regimens.

[0003] Currently, the mainstream methods for predicting neoantigens include the following categories: (i) Affinity-based prediction methods: Tools such as NetMHCpan and MHCflurry are used to predict the binding affinity of mutant peptides to MHC molecules, with the half-maximal inhibitory concentration (IC50) value as the screening criterion. These methods only consider the single dimension of peptide-MHC binding, ignoring the differences between mutant and wild-type peptides, as well as other factors affecting immunogenicity such as mutation abundance and gene expression levels.

[0004] (ii) Differential scoring-based methods: These methods assess the presentation advantage of mutant peptides relative to wild-type peptides by calculating the score ratio between mutant and wild-type peptides. These methods use score ratios independently, lack modeling of the interaction between mutant and wild-type peptides, and fail to reflect their synergistic relationship in the antigen presentation process.

[0005] (iii) Multi-step integrated screening methods: Tools such as pVACseq and NeoPredPipe integrate multiple steps for neoantigen screening, linking affinity prediction, gene expression analysis, and other steps, and performing hierarchical filtering based on preset rules. These methods lack a unified scoring framework, with each step using independent thresholds and standards, making it difficult to conduct comprehensive evaluation on the same scale.

[0006] However, the aforementioned existing technologies still have the following problems and shortcomings: First, the screening criteria are too singular: existing screening schemes mainly rely on MHC affinity prediction results to screen candidate peptides. However, due to differences in training data sources, model framework design, and validation datasets, current mainstream affinity prediction tools such as NetMHCpan and MHCflurry, while performing excellently on their respective constructed test sets, often show a certain gap in generalization on external independent test sets compared to publicly reported results. Therefore, relying solely on a single affinity prediction tool for candidate screening can easily introduce model bias, thereby affecting the accuracy and reliability of candidate neoantigen identification.

[0007] Second, there is a lack of interpretability: current screening protocols rely on empirical rules reported in the literature and employ a hierarchical approach to select candidate peptides. However, the biological significance of the relationships between these rules often lacks systematic integration, making it difficult to fully cover the actual needs of complex clinical scenarios. Furthermore, the combination logic and applicable boundaries of different rules are not clearly defined, limiting users' understanding of the protocol and its application in vaccine development. The main reason for this problem is that the formulation and combination of rules must consider not only statistical evidence but also a comprehensive design that integrates biological mechanisms such as neoantigen generation, presentation, and immune recognition.

[0008] Third, the screening mechanism is imperfect: New and neoantigen mutations often induce significant changes in binding characteristics. Traditional methods rely too heavily on affinity, giving it high weight in screening protocols, but often fail to fully consider the biological information implied by these affinity changes in subsequent immunogenicity assessments. Peptides can still elicit effective immune responses; relying solely on affinity limits the number of candidate peptides available for vaccine discovery. Furthermore, the deduplication strategy for neoantigens determines which highly efficient peptides can be retained. Without considering the comprehensive characteristics of peptides, it is difficult to meet the comprehensive judgment requirements of precision medicine.

[0009] Therefore, this invention proposes a novel antigen immunogenicity prediction and screening method based on multi-dimensional feature fusion to solve the problems existing in the prior art. Summary of the Invention

[0010] To address the aforementioned problems, the present invention aims to propose a novel antigen immunogenicity prediction and screening method based on multi-dimensional feature fusion. By fusing affinity scores, mutation effect characteristics, and multi-dimensional biological characteristics, the present invention significantly reduces the amount of candidate peptide data while retaining as many candidate peptides as possible with potential immunogenicity and experimental validation value, thereby improving the accuracy and practicality of the screening results.

[0011] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion, comprising the following steps: Step 1: Obtain the prediction results data of mutant peptides and wild-type peptides after processing by the MHC binding prediction tool. The prediction results data includes at least affinity score, binding tag and common information column for matching mutant peptides and wild-type peptides. According to the common field, the prediction results data of mutant peptides and wild-type peptides are merged by inner join according to the common column information. Data records with missing or invalid values ​​in the merged data are removed. All prediction results are processed with affinity_score(0,1). Step 2: Perform biofeasibility filtering on the data processed in Step 1, and retain candidate peptides that simultaneously meet the following conditions: allele frequency is greater than a preset frequency threshold, transcript expression level is greater than a preset expression level threshold, mutant-wild ratio is greater than zero, and peptide length falls within the MHC-I class molecule presentation length range. Step 3: For the data processed in Step 2, extract the affinity score and mutant-wild ratio for each candidate peptide, then calculate the natural logarithm of the mutant-wild ratio, and then call the pre-constructed immunogenicity scoring function to calculate the immunogenicity probability. The inputs to the immunogenicity scoring function include the affinity score, the natural logarithm of the mutant-wild ratio, and the product of the affinity score and the natural logarithm of the mutant-wild ratio. Step 4: Based on the immunogenicity probability and affinity score, the candidate peptides are corrected according to the rules to obtain the corrected classification label for each candidate peptide, which includes a positive label and a negative label. Step 5: Perform multi-dimensional feature aggregation on the candidate peptides identified as positive tags in Step 4 to generate multi-dimensional aggregated tags; Step 6: Based on the affinity score, immunogenicity probability, and the multidimensional aggregation marker, sort and deduplicate the candidate peptides processed in Step 5, retain the highest-ranked record among multiple rows of data associated with the same peptide, delete the remaining records, and obtain the deduplicated candidate peptide set. Step 7: Generate a multi-level candidate neoantigen list based on the preset screening threshold and the candidate peptide set processed in Step 6.

[0012] A further improvement is that, in step one, removing invalid data rows from the merged data specifically includes deleting data rows where the prediction results for either the mutant peptide or the wild-type peptide are missing.

[0013] The further improvement lies in the following: In step two, the preset frequency threshold is five percent, the preset expression level threshold is zero, and the presentation length of MHC-I molecules ranges from eight to eleven amino acids.

[0014] A further improvement is made in step two, where the mutant-to-wild ratio is the ratio of the affinity score of the mutant peptide to that of the wild peptide. Here, the value after IC50 normalization is used, and the value is between 0 and 1.

[0015] A further improvement is made in step three, where the input features of the immunogenicity scoring function consist of the affinity score, the natural logarithm of the mutant-wild ratio, and the product of the affinity score and the natural logarithm of the mutant-wild ratio. The immunogenicity probability is output after being transformed by the sigmoid function through a linear weighted sum of the input features.

[0016] A further improvement lies in the following: the immunogenicity scoring function is constructed as follows: An immunogenicity prediction model was established using a logistic regression model, and the model parameters were obtained through parameter optimization. The parameter optimization was achieved by combining empirical parameter initialization with grid search algorithm or cross-validation. The logistic regression model took affinity score, natural logarithm of mutant-wild ratio, and the product of affinity score and natural logarithm of mutant-wild ratio as inputs. By learning the contribution relationship of different features to immunogenicity, an immunogenicity prediction function was obtained. The training data for the logistic regression model consists of immunogenic positive samples and immunogenic negative samples. The immunogenic positive samples are mutant peptides that have been experimentally verified to induce specific T cell immune responses. The immunogenic negative samples are corresponding peptides that are from the same patient as the immunogenic positive samples, have the same MHC restriction background, and have not been detected with an immune response under the same experimental conditions.

[0017] By constructing paired samples as described above, the genetic background, MHC presentation environment, and transcriptional expression background of patients in the training samples are kept consistent, thereby reducing the impact of non-mutational factors on the model training results and improving the model's ability to identify mutation-related features. A further improvement is made in step four, where the rule correction includes the following rules: When the affinity score is lower than the preset affinity threshold, the sorting priority of the corrected classification label of the candidate peptide is reduced. When the mutant-wild ratio exceeds the preset upper limit, the sorting priority of the candidate peptide's corrected classification label is reduced. When the absolute value of the difference between the mutant wild ratio and the numerical value is less than the preset near threshold and the affinity score is greater than the preset high affinity threshold, the corrected classification label of the candidate peptide is determined to be negative.

[0018] A further improvement is that, in step five, the multidimensional aggregated markers include: the number of MHC alleles corresponding to each candidate peptide, the number of mutation sources corresponding to each candidate peptide, the number of genes covered by each candidate peptide, the frequency of occurrence of each candidate peptide in all data, and standard transcript markers.

[0019] A further improvement is made in step six, where the specific method for sorting and deduplication is to sort multiple rows of data associated with the same peptide segment according to their priority, with the sorting priority decreasing sequentially. The priority is as follows: The first priority is determined by descending order of affinity scores; The second priority is in descending order of immunogenicity probability; The third priority is ascending order of transcript support level; The fourth priority is the standard transcript marker in the multidimensional aggregated markers, with standard transcripts taking precedence over non-standard transcripts; The fifth priority is the ascending order of the original data rows.

[0020] A further improvement is made in step seven, where the multi-level candidate neoantigen list includes: The first N peptide segments of each mutation site, where N is a preset positive integer; Candidate peptides that simultaneously satisfy the conditions of having an immunogenicity probability greater than a preset probability threshold and an affinity score less than a preset first affinity threshold; Candidate peptides that meet the criteria of immunogenicity probability greater than a preset probability threshold and affinity score less than a preset second affinity threshold, wherein the preset second affinity threshold is less than a preset first affinity threshold; The peptide with the highest immunogenicity probability at each mutation site.

[0021] The beneficial effects of this invention are as follows: (1) This invention introduces an interaction term between affinity score and mutation effect characteristics into the logistic regression model, realizing the synergistic modeling of affinity and mutation effect, breaking through the limitation of relying solely on single affinity prediction, and improving the accuracy of immunogenicity assessment.

[0022] (2) The present invention combines biological rules and probability models for collaborative screening, which makes the screening process have clear judgment basis and interpretability, and improves the interpretability of the screening results.

[0023] (3) The present invention avoids relying solely on fixed affinity thresholds for screening through a rule correction mechanism, thereby retaining some low-affinity candidate peptides with potential immunogenicity and expanding the scope of candidate neoantigen discovery.

[0024] (4) This invention improves the stability and biological rationality of candidate peptide ranking results through multi-dimensional feature aggregation and priority deduplication strategy, providing a more effective and accurate screening scheme for neoantigen vaccine design.

[0025] (5) The present invention uses mutant peptide-wild-type peptide paired samples to construct an immunogenicity training dataset. By controlling factors such as patient background, MHC restriction background and expression background, the model’s ability to learn features related to mutation effects is improved, thereby enhancing the stability and accuracy of immunogenicity prediction results. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the steps of the present invention.

[0027] Figure 2 This is a schematic diagram of the immunogenicity scoring structure of the present invention. Detailed Implementation

[0028] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0029] In this embodiment, the mutant-to-wild-peptide ratio refers to the affinity ratio between the mutant peptide and the wild-type peptide.

[0030] according to Figures 1-2 As shown, this embodiment provides a novel antigen immunogenicity prediction and screening method based on multi-dimensional feature fusion, taking its application in tumor samples from colorectal cancer patients as an example, including the following steps: Step 1: Data Acquisition and Preprocessing Whole-exome sequencing and RNA sequencing data of tumor tissue from a colorectal cancer patient were obtained. After somatic mutation detection and HLA typing, the patient's mutated peptides and their corresponding wild-type peptides were identified. In this embodiment, a total of 425 somatic mutation sites were obtained from the patient, and after full permutation and combination with the patient's HLA, 85,250 candidate raw peptides were obtained.

[0031] All mutant peptides and their corresponding wild-type peptides were input into an MHC binding prediction tool for binding prediction. The prediction tool used in this embodiment is a self-developed MHC-I class molecule binding prediction algorithm, which supports 144 MHC-I class molecule genotypes. Before prediction, the nomenclature format of all MHC alleles was standardized to HLA-A*02:01.

[0032] The input peptide sequences are examined, and only peptide sequences composed of standard amino acids are retained.

[0033] When the prediction tool is executed, the data is grouped according to MHC typing, and the corresponding perceptual model is called for prediction in each group. After the prediction is completed, the prediction results for each peptide are obtained, including common column information such as affinity score, binding tag, peptide sequence, and mutation site information.

[0034] The predicted data for mutant peptides were merged with those for wild-type peptides using inner joins based on common column information. After merging, rows with missing predicted data for either mutant or wild-type peptides were removed. After merging and cleaning, 51,372 valid candidate peptides remained, with peptide lengths ranging from eight to eleven amino acids.

[0035] Step 2: Biological Feasibility Filtration The data processed in the above steps are subjected to biofeasibility filtering, retaining candidate peptides that simultaneously meet the following conditions: If the allele frequency is greater than 5%, the mutation detection data of this patient contains allele frequency information for each mutation site. In this embodiment, 5% is used as the screening threshold to remove low-frequency mutations below this threshold. Transcript expression level is greater than zero. In this embodiment, transcript expression level is greater than zero as the screening criterion, that is, only candidate peptides supported by transcript expression are retained. A mutant-to-wild ratio greater than zero indicates that the predicted affinity score of the mutant peptide is higher than that of the wild peptide. The peptide length is between eight and eleven amino acids.

[0036] Therefore, after biological feasibility filtering, 34,505 candidate peptides remain.

[0037] Step 3: Immunogenicity scoring The immunogenicity scoring function used in this embodiment was obtained by solving a training dataset, which included immunogenicity-positive and immunogenicity-negative samples. Mutant peptides that have been experimentally verified to induce specific T-cell immune responses were considered positive samples, while their corresponding wild-type peptides, derived from the same patient and sharing the same MHC restriction background, were considered negative samples.

[0038] For the data filtered for biofeasibility, the affinity score and mutant-wild ratio of each candidate peptide were extracted, and the natural logarithm of the mutant-wild ratio was calculated.

[0039] Furthermore, a pre-constructed immunogenicity scoring function is used for scoring, the formula of which is: Where A represents the affinity score, R represents the mutant-to-wild ratio, ln(R) represents the natural logarithm of the mutant-to-wild ratio, and A×ln(R) represents the product of the affinity score and the natural logarithm of the mutant-to-wild ratio. , , , These are the calculated model coefficients. In this embodiment, the model coefficients are obtained in advance through training on the IEDB immunogenicity dataset. The value is -2. The value is 2. The value is 6. The value is -3.

[0040] Substituting the affinity score and the natural logarithm of the mutant-wild ratio for each candidate peptide into the above formula, the immunogenicity probability p is calculated, which is between zero and one.

[0041] Step 4: Rule Revision and Priority Assessment Candidate peptides are corrected according to rules based on immunogenicity probability and affinity score, resulting in corrected classification labels for each candidate peptide, including positive and negative labels.

[0042] The specific rules are as follows: When the affinity score is lower than a preset affinity threshold, the revised classification label of the candidate peptide is lowered in sorting priority. In this embodiment, the IC50 value corresponding to the preset affinity threshold is 500 nM, meaning that peptides with affinity scores lower than this threshold are shifted back during sorting.

[0043] When the mutant-to-wild-type ratio exceeds a preset upper limit, the ranking priority of the candidate peptide's corrected classification label is reduced. In this embodiment, the preset upper limit is 20; for peptides with a mutant-to-wild-type ratio exceeding 20, their ranking priority is downgraded.

[0044] When the absolute value of the difference between the mutant-wild-type ratio and the value of 1 is less than a preset proximity threshold and the affinity score is greater than a preset high affinity threshold, the corrected classification label of the candidate peptide is determined to be negative. In this embodiment, the preset proximity threshold is 0.1 and the preset high affinity threshold is 50 nM. This rule is used to exclude wild-type background peptides with a mutant-wild-type ratio close to 1 (i.e., the mutant peptide is not significantly different from the wild-type peptide) and high affinity.

[0045] Step 5: Multi-dimensional feature aggregation Candidate peptides identified as positive tags undergo multi-dimensional feature aggregation to generate multi-dimensional aggregated tags. Specifically, this includes the following statistical dimensions: The number of MHC alleles corresponding to each candidate peptide is counted to generate allele affinity binding capacity markers. If a peptide is predicted to bind by multiple MHC alleles, its allele binding capacity is marked as a multi-allelic genotype.

[0046] The number of mutation sources corresponding to each candidate peptide is counted to generate a multi-mutation source marker. If a peptide originates from multiple different somatic mutations, its multi-mutation source marker is designated as multi-mutation source type.

[0047] The number of genes covered by each candidate peptide is counted to generate cross-gene coverage markers. If the transcript corresponding to a peptide is marked as a standard transcript, then its standard transcript is marked as yes; otherwise, it is a non-standard transcript.

[0048] Count the frequency of each candidate peptide in all data to generate peptide frequency markers.

[0049] After multi-dimensional feature aggregation, each positive candidate peptide was appended with the aforementioned multi-dimensional aggregation marker.

[0050] Step Six: Peptide Sorting and Deduplication Based on affinity scores, immunogenicity probabilities, and multidimensional aggregation markers, candidate peptides aggregated with multidimensional features were sorted and deduplicated.

[0051] The specific sorting method is as follows: Multiple rows of data associated with the same peptide segment are sorted according to the following priority, with the priority decreasing in descending order: The first priority is determined by descending order of affinity scores; The second priority is in descending order of immunogenicity probability; The third priority is ascending order of transcript support level; The fourth priority is the standard transcript marker in the multidimensional aggregated marker, with standard transcripts taking precedence over non-standard transcripts; The fifth priority is the ascending order of the original data rows.

[0052] After sorting, the highest-ranked record among multiple rows of data associated with the same peptide is retained, and the remaining records are deleted, resulting in a deduplicated set of candidate peptides. After further deduplication, 198 candidate peptides are obtained.

[0053] Step 7: Multi-level candidate output Based on a preset screening threshold and a deduplicated set of candidate peptides, a multi-level candidate neoantigen list is generated. In this embodiment, the output multi-level candidate neoantigen list includes: The first three peptide segments of each mutation site; Candidate peptides that simultaneously meet the requirements of immunogenicity probability greater than 0.5 and affinity score less than 500 nM; Candidate peptides that meet the requirements of immunogenicity probability greater than 0.5 and affinity score less than 50 nM; The peptide with the highest immunogenicity probability at each mutation site.

[0054] The list of candidate neoantigens obtained in this embodiment was compared with the reported data on the patient's immune response validation. The optimal ranking of each mutation site at each screening stage is shown in Table 1 below: Table 1. Optimal ranking of validation mutations at each stage In Table 1, the best ranking refers to the highest-ranking peptide containing the current mutation site. Despite undergoing rigorous immunogenicity scoring and affinity screening, the API5, RNF10, and PHLPP1 mutations were retained and placed at a high position in the final candidate list. This demonstrates that this embodiment does not mechanically retain all known mutations in NeoReward, but rather redefines candidate priorities under the constraints of biological feasibility and immune function, effectively improving the overall ranking of validated candidate peptides to within the top thirty (API5 R243Q at number twenty-four and PHLPP1 G566E at number twenty-one), providing valuable biological insights for vaccine design and immune mechanism research.

[0055] The above results demonstrate that the method provided in this embodiment can significantly reduce the number of candidates while retaining and prioritizing mutation candidates with validated evidence. Finding even one genuine neoantigen in a patient is clinically significant, and this embodiment achieves effective screening and ranking improvement at all three validated mutation sites.

[0056] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the present invention without departing from its framework and scope of application, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion, characterized in that: Includes the following steps: Step 1: Obtain the prediction results data of mutant peptides and wild-type peptides after processing by the MHC binding prediction tool. The prediction results data includes affinity score, binding tag and common column information. Then, the prediction results data of mutant peptides and wild-type peptides are merged by inner linking according to the common column information, and invalid data rows in the merged data are removed. Step 2: Perform biofeasibility filtering on the data processed in Step 1, and retain candidate peptides that simultaneously meet the following conditions: allele frequency is greater than a preset frequency threshold, transcript expression level is greater than a preset expression level threshold, mutant-wild ratio is greater than zero, and peptide length falls within the MHC-I class molecule presentation length range. Step 3: For the data processed in Step 2, extract the affinity score and mutant-wild ratio for each candidate peptide, then calculate the natural logarithm of the mutant-wild ratio, and then call the pre-constructed immunogenicity scoring function to calculate the immunogenicity probability. The inputs to the immunogenicity scoring function include the affinity score, the natural logarithm of the mutant-wild ratio, and the product of the affinity score and the natural logarithm of the mutant-wild ratio. Step 4: Based on the immunogenicity probability and affinity score, the candidate peptides are corrected according to the rules to obtain the corrected classification label for each candidate peptide, which includes a positive label and a negative label. Step 5: Perform multi-dimensional feature aggregation on the candidate peptides identified as positive tags in Step 4 to generate multi-dimensional aggregated tags; Step 6: Based on the affinity score, immunogenicity probability, and the multidimensional aggregation marker, sort and deduplicate the candidate peptides processed in Step 5, retain the highest-ranked record among multiple rows of data associated with the same peptide, delete the remaining records, and obtain the deduplicated candidate peptide set. Step 7: Generate a multi-level candidate neoantigen list based on the preset screening threshold and the candidate peptide set processed in Step 6.

2. The method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion according to claim 1, characterized in that: In step one, removing invalid data rows from the merged data specifically includes deleting data rows where the prediction results for either the mutant peptide or the wild-type peptide are missing.

3. The method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion according to claim 1, characterized in that: In step two, the preset frequency threshold is 5%, the preset expression level threshold is zero, and the MHC-I class molecule presentation length ranges from eight to eleven amino acids.

4. The method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion according to claim 1, characterized in that: In step two, the mutant-to-wild ratio is the ratio of the affinity score of the mutant peptide to the affinity score of the wild peptide.

5. The method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion according to claim 1, characterized in that: In step three, the input features of the immunogenicity scoring function consist of the affinity score, the natural logarithm of the mutant-wild ratio, and the product of the affinity score and the natural logarithm of the mutant-wild ratio. The immunogenicity probability is output after linear weighting of the input features and transformation by the sigmoid function.

6. The method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion according to claim 1, characterized in that: In step three, the immunogenicity scoring function is constructed as follows: An immunogenicity prediction model was established using a logistic regression model, and the model parameters were obtained through parameter optimization. The parameter optimization was achieved by combining empirical parameter initialization with grid search algorithm or cross-validation. The logistic regression model took affinity score, natural logarithm of mutant-wild ratio, and the product of affinity score and natural logarithm of mutant-wild ratio as inputs. By learning the contribution relationship of different features to immunogenicity, an immunogenicity prediction function was obtained. The training data for the logistic regression model consists of immunogenic positive samples and immunogenic negative samples. The immunogenic positive samples are mutant peptides that have been experimentally verified to induce specific T cell immune responses. The immunogenic negative samples are corresponding peptides that are from the same patient as the immunogenic positive samples, have the same MHC restriction background, and have not been detected with an immune response under the same experimental conditions.

7. The method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion according to claim 1, characterized in that: In step four, the rule correction includes the following rules: When the affinity score is lower than the preset affinity threshold, the sorting priority of the corrected classification label of the candidate peptide is reduced. When the mutant-wild ratio exceeds the preset upper limit, the sorting priority of the candidate peptide's corrected classification label is reduced. When the absolute value of the difference between the mutant wild ratio and the numerical value is less than the preset near threshold and the affinity score is greater than the preset high affinity threshold, the corrected classification label of the candidate peptide is determined to be negative.

8. The method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion according to claim 1, characterized in that: In step five, the multidimensional aggregated markers include: the number of MHC alleles corresponding to each candidate peptide, the number of mutation sources corresponding to each candidate peptide, the number of genes covered by each candidate peptide, the frequency of occurrence of each candidate peptide in all data, and standard transcript markers.

9. The method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion according to claim 1, characterized in that: In step six, the specific method for sorting and deduplication is to sort multiple rows of data associated with the same peptide segment according to their priority, with the sorting priority decreasing sequentially. The priority is as follows: The first priority is determined by descending order of affinity scores; The second priority is in descending order of immunogenicity probability; The third priority is ascending order of transcript support level; The fourth priority is the standard transcript marker in the multidimensional aggregated markers, with standard transcripts taking precedence over non-standard transcripts; The fifth priority is the ascending order of the original data rows.

10. The method for predicting and screening the immunogenicity of new antigens based on multi-dimensional feature fusion according to claim 1, characterized in that: In step seven, the multi-level candidate neoantigen list includes: The first N peptide segments of each mutation site, where N is a preset positive integer; Candidate peptides that simultaneously satisfy the conditions of having an immunogenicity probability greater than a preset probability threshold and an affinity score less than a preset first affinity threshold; Candidate peptides that meet the criteria of immunogenicity probability greater than a preset probability threshold and affinity score less than a preset second affinity threshold, wherein the preset second affinity threshold is less than a preset first affinity threshold; The peptide with the highest immunogenicity probability at each mutation site.

Images