CN118629514B - Sequence immunogenicity prediction method, device, electronic device and storage medium - Google Patents

Abstract

The invention provides a sequence immunogenicity prediction method, a sequence immunogenicity prediction device, electronic equipment and a storage medium, and relates to the technical field of immunogenicity prediction. According to the sequence immunogenicity prediction method, the sequence immunogenicity prediction device, the electronic equipment and the storage medium, related information of an existing immunogenicity database is utilized to realize the excavation of related characteristics of the immunogenicity, so that a linear relation between sequence characteristics of a known peptide fragment and the immunogenicity of the sequence is established, a sequence immunogenicity prediction model is established based on the linear relation, and then the sequence characteristics of a sequence to be predicted and the sequence immunogenicity prediction model are utilized to realize the prediction of the immunogenicity of the sequence to be predicted. Compared with the prior art, the method can improve the interpretability and the accuracy of the immunogenicity prediction result, has great significance in meeting the supervision requirement and improving the research and development efficiency, and is beneficial to improving the safety evaluation accuracy and accelerating the drug marketing process.

Description

Sequence immunogenicity prediction method, device, electronic device and storage medium

Technical Field

The present invention relates to the technical field of immunogenicity prediction, and in particular to a sequence immunogenicity prediction method, device, electronic device and storage medium.

Background Art

Immunogenicity prediction refers to the prediction of the immunogenicity of proteins or peptides through bioinformatics methods, that is, the prediction of their possibility of inducing an immune response. It can be used in peptides, antibodies, PDC (peptide-drug conjugate), antibody-drug conjugate (ADC), etc. Immunogenicity prediction is of great significance in the fields of vaccine development, allergen identification, autoimmune disease research, etc.

At present, there are two main methods for immunogenicity prediction: sequence-based and spatial structure-based. Among them, the sequence-based immunogenicity prediction method mainly realizes immunogenicity prediction by mining and extracting the rules and characteristics of the sequence itself, while the spatial structure-based immunogenicity prediction method relies on the peptide- MHC crystal structure and scoring function to predict the peptide- MHC interaction, thereby realizing immunogenicity prediction.

However, although the sequence-based method has certain advantages in accuracy, its accuracy depends on a large amount of high-quality training data, and this method is a black box model with poor interpretability; while the structure-based method does not rely on training data and has strong interpretability, its prediction accuracy is poor. It can be seen that it is difficult for existing immunogenicity prediction methods to take into account both prediction accuracy and interpretability.

Summary of the invention

1. Technical issues to be resolved

In view of the deficiencies in the prior art, the present invention provides a sequence immunogenicity prediction method, device, electronic device and storage medium, which at least solves the problem that the existing immunogenicity prediction methods cannot take into account both prediction accuracy and interpretability.

(II) Technical solution

To achieve the above objectives, the present invention is implemented through the following technical solutions:

In a first aspect, the present invention provides a method for predicting sequence immunogenicity, the method comprising:

Extracting sequence features of peptide sequences with known immunogenicity, and constructing a feature matrix based on the sequence features; the sequence features include k-mer features; wherein k is a positive integer;

Establishing a regression relationship between the feature matrix and immunogenicity, and obtaining weights corresponding to different sequence features based on the regression relationship;

Constructing a sequence immunogenicity prediction model based on the weights;

The immunogenicity of the sequence to be predicted is predicted based on the sequence features of the sequence to be predicted and the sequence immunogenicity prediction model.

In some embodiments, the sequence to be predicted is selected from a polypeptide or an antibody.

Furthermore, antibodies include ADCs.

In some embodiments, the polypeptide comprises a peptide segment of 8-13 amino acids in length.

In some embodiments, the sequence features further include: CPP features of the peptide sequence and thermodynamic information features of the peptide sequence.

In some embodiments, the thermodynamic information features of the peptide sequence include: the aliphatic index of the protein sequence, the BLOSUM62 index of the protein sequence, the Boman index, the theoretical net charge of the protein sequence, the Cruciani property of the protein sequence, the FASGAI vector of the protein sequence, the hydrophobic moment of the protein sequence, the hydrophobic index of the protein sequence, the instability index of the protein sequence, the Kidela factor of the protein sequence, the isoelectric point of the protein sequence, the protFP descriptor of the protein sequence, the VHSE scale of the protein sequence, and the Z-scale parameter of the protein sequence.

In some embodiments, the regression relationship is a linear regression relationship.

Furthermore, in one embodiment, obtaining weights corresponding to different sequence features based on the regression relationship includes: obtaining weights corresponding to different sequence features based on elastic network regression.

In some embodiments, extracting the k-mer features comprises:

Customize the value of k ; where k represents the length of the peptide sequence fragment to be considered, and k is a positive integer;

Starting from one end of a given protein sequence, all k amino acid combinations in the protein sequence are traversed, and the types and quantities of all amino acid combinations are counted.

Furthermore, in some embodiments, k is any positive integer between 2 and 10.

Furthermore, in one embodiment, k =2 or k =3.

In some embodiments, extracting the CPP features of the peptide sequence comprises:

Each peptide sequence was globally aligned with the TCR sequence, and the quantile of the alignment score was used as the CPP feature.

Furthermore, in some embodiments, each peptide sequence is globally aligned with the TCR sequence based on an amino acid scoring matrix.

Furthermore, in one embodiment, the 10% and 90% quantiles of the alignment score are used as CPP features.

In a second aspect, the present invention provides a sequence immunogenicity prediction device, the device comprising:

The immunogenicity prediction model building unit is used to extract sequence features of peptide sequences with known immunogenicity, and construct a feature matrix based on the sequence features; establish a regression relationship between the feature matrix and immunogenicity, and obtain weights corresponding to different sequence features based on the regression relationship; and construct a sequence immunogenicity prediction model based on the weights; wherein the sequence features include k-mer features; wherein k is a positive integer;

The immunogenicity prediction unit is used to predict the immunogenicity of the sequence to be predicted based on the sequence characteristics of the sequence to be predicted and the sequence immunogenicity prediction model.

In some embodiments, the sequence to be predicted is selected from a polypeptide or an antibody.

Furthermore, antibodies include ADCs.

In some embodiments, the polypeptide comprises a peptide segment of 8-13 amino acids in length.

In some embodiments, the immunogenicity prediction model building unit also includes a CPP feature extraction subunit and a thermodynamic related feature extraction subunit; the CPP feature extraction subunit is used to extract the CPP features of the peptide sequence, and the thermodynamic related feature extraction subunit is used to extract the thermodynamic information features of the peptide sequence.

In some embodiments, the thermodynamic information features of the peptide sequence include: the aliphatic index of the protein sequence, the BLOSUM62 index of the protein sequence, the Boman index, the theoretical net charge of the protein sequence, the Cruciani property of the protein sequence, the FASGAI (Factor Analysis Scales of Generalized AminoAcid Information) vector of the protein sequence, the hydrophobic moment of the protein sequence, the hydrophobic index of the protein sequence, the instability index of the protein sequence, the Kidela factor of the protein sequence, the isoelectric point (pI) of the protein sequence, the protFP descriptor of the protein sequence, the VHSE scale of the protein sequence, and the Z-scale parameter of the protein sequence.

Furthermore, in one embodiment, the thermodynamic information characteristics of the peptide sequence are calculated using the corresponding function in the peptide package of R software.

In some embodiments, the regression relationship is a linear regression relationship.

Furthermore, in some embodiments, obtaining weights corresponding to different sequence features based on the regression relationship includes: obtaining weights corresponding to different sequence features based on elastic network regression.

Furthermore, in some embodiments, when obtaining weights corresponding to different sequence features based on elastic network regression, a 5-fold cross validation method is used to perform parameter estimation.

In some embodiments, the immunogenicity prediction model building unit includes a k-mer feature extraction subunit, and the k-mer feature extraction subunit extracts the k-mer feature including:

Customize the value of k ; where k represents the length of the peptide sequence fragment to be considered, and k is a positive integer;

Starting from one end of a given protein sequence, all k amino acid combinations in the protein sequence are traversed, and the types and quantities of all the amino acid combinations are counted.

Furthermore, in some embodiments, k is any positive integer between 2 and 10.

Furthermore, in one embodiment, k =2 or k =3.

In some embodiments, the CPP feature extraction subunit extracts the CPP feature of the peptide sequence including:

Each peptide sequence was globally aligned with the TCR sequence, and the quantile of the alignment score was used as the CPP feature.

Furthermore, in some embodiments, each peptide sequence is globally aligned with the TCR sequence based on an amino acid scoring matrix.

Furthermore, in one embodiment, the 10% and 90% quantiles of the alignment score are used as CPP features.

In a third aspect, the present invention provides an electronic device comprising a processor, a communication interface, a memory and a bus, wherein the processor, the communication interface and the memory communicate with each other via the bus, and the processor can call logic instructions in the memory to execute the steps of the method provided in the first aspect.

In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

(III) Beneficial effects

The present invention provides a sequence immunogenicity prediction method, device, electronic device and storage medium. Compared with the prior art, it has the following beneficial effects:

The present invention provides a method, device, electronic device and storage medium for predicting sequence immunogenicity, which mines immunogenicity-related features through relevant information of an existing immunogenicity database, thereby establishing a linear relationship between the sequence features and immunogenicity of a known peptide sequence, and using this relationship as a sequence immunogenicity prediction model, and then using the sequence features of the sequence to be predicted and the above sequence immunogenicity prediction model to predict the immunogenicity of the sequence to be predicted. Compared with the prior art, the immunogenicity prediction results of the present invention are highly interpretable and accurate.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

FIG1 is a flow chart of a method for predicting sequence immunogenicity of the present invention;

FIG2 is a flow chart of constructing sequence features in Example 1 of the present invention;

FIG3 is a flow chart of feature extraction in Example 1 of the present invention;

FIG4 is a curve showing auc changes during the cross validation process in Example 1 of the present invention;

FIG5 is a schematic diagram of the structure of a sequence immunogenicity prediction device in Example 2 of the present invention;

FIG6 is a schematic diagram of the structure of an electronic device for predicting sequence immunogenicity in Example 3 of the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Antibody-Drug Conjugate (ADC) can combine the specificity and targeting of antibodies with the efficacy of cytotoxic drugs to achieve precise attacks on tumor cells. ADC combines the specificity of antibodies with the high efficiency of drugs, which is of great significance in improving the accuracy of cancer treatment and reducing toxicity to normal tissues. However, due to its complex structure and mechanism of action, ADC may trigger specific immune responses (such as the production of anti-drug antibodies ADA, etc.), resulting in reduced efficacy or aggravated adverse reactions. Therefore, it is urgent to develop an accurate and highly interpretable method for evaluating and predicting the immunogenicity of ADC drugs.

At present, pharmaceutical regulatory standards are gradually improving, and the stringent requirements for the evaluation of the immunogenicity of ADC drugs are also increasing. There is an urgent need to develop prediction methods that are both accurate and highly interpretable.

To date, the immunogenicity prediction methods are mainly sequence-based immunogenicity prediction methods and spatial structure-based immunogenicity prediction methods. Among them, sequence-based immunogenicity prediction methods, such as SMM (B. Peters, A. Sette, BMC bioinformatics 6, 132 (2005)), use a matrix-based algorithm and a site-specific scoring matrix to determine whether a peptide sequence matches the binding motif of a specific MHC allele; NetMHC uses an artificial neural network to train potential patterns or features with binding prediction capabilities through the data of each allele (e.g., "HLA-A*02:01", "H2-Db", etc.). Although sequence-based immunogenicity prediction methods are simple, intuitive, easy to implement, and have high accuracy, they rely on a large amount of training data and have certain requirements for data volume and data quality. The processing of a large amount of data will lead to slow model operation, high cost, and difficulty in large-scale implementation; at the same time, sequence-based methods are black box models, and their results are poorly interpretable and difficult to trace. Structure-based methods rely on peptide- MHC crystal structures and scoring functions to predict peptide-MHC by energy minimization. MHC interactions, thereby achieving immunogenicity prediction. Although structure-based methods do not rely on training data and have certain interpretability, their accuracy is poor.

In summary, the existing immunogenicity prediction methods cannot balance explainability and accuracy, and are slow to run, costly, and difficult to implement on a large scale, and cannot effectively assist in the development of ADC drugs.

In response to the above problems, the present application uses the relevant information and data in the existing immunogenicity database to mine immunogenicity-related features, and then establishes the relationship between the above features and immunogenicity, and uses this relationship to predict the immunogenicity of the sequence to be predicted.

It should be noted that the immunogenicity prediction technology proposed in the present invention can be used for the immunogenicity prediction of peptides, antibodies, PDC (peptide-drug conjugate), ADC and other aspects, because the immunogenicity prediction is only a prediction of peptides or antibodies, so in theory, as long as the drug contains a peptide, the immunogenicity prediction technology proposed in the present invention can actually be applied. However, in order to facilitate the elaboration of the technical solution, ADC is used as an example in the embodiments of the present invention, but this is not regarded as limiting the scope of protection of the technical solution of the present invention.

Based on this, the present invention proposes a sequence immunogenicity prediction method, device, electronic device and storage medium. In order to facilitate the understanding of the technical solution of this application, the nouns and related concepts appearing in this application are explained:

Epitope metadata refers to data that contains information about the interactions between biological molecules (such as proteins, antibodies, etc.).

Peptides are short-chain molecules composed of amino acid residues, usually consisting of 2 to 50 amino acids. Peptides are formed when amino acid residues are linked together by peptide bonds. Peptides are the building blocks of proteins, which are long-chain molecules composed of one or more peptides.

MHC (Major Histocompatibility Complex) is a group of genes that encode specific proteins. Its sequence is very complex because MHC genes are polymorphic, that is, there are various different alleles. The proteins encoded by MHC genes play an important role in the immune system, especially in the process of antigen presentation.

MHC Ligand Assays (Class II) is an experimental method used to study the interaction between MHC (Major Histocompatibility Complex) Class II molecules and their ligands. MHC Class II molecules mainly play a role in the antigen presentation process. They can bind and present exogenous antigens to CD4+T cells, thereby triggering an immune response.

TCR data refers to "Tissue-Resident Memory" cell data, that is, tissue-resident memory cell data. These cells are a special type of immune cells that are mainly found in tissues rather than circulating in the blood. They play an important role in immune response, can quickly respond to reinfection, and provide long-term immune protection. TCR data are often used to study the response and memory mechanisms of the immune system.

The truncated mean is to remove the maximum and minimum values when calculating the average, and then average the remaining values.

Quantiles are values that divide a data set into several equal parts after arranging it in order of size. Common quantiles include quartiles (dividing data into four equal parts), medians (dividing data into two equal parts), deciles (dividing data into ten equal parts), etc. Quantiles are a commonly used descriptive statistic in statistics and data analysis. They can help better understand the distribution and central tendency of data, and can also be used to identify outliers and assess the degree of dispersion of data.

R software is an open source statistical analysis software, which is widely used in data analysis, statistical modeling and visualization.

K-mer refers to all possible sequence fragments (or "substrings") of length k , which can be part of DNA, RNA or protein sequences. In protein sequence analysis, the k-mer method involves dividing the protein sequence into all possible fragments of length k and using these fragments for further analysis, such as sequence alignment, pattern recognition, function prediction, etc.

ElasticNet Regression is a linear regression method that combines L1 regularization (Lasso) and L2 regularization (Ridge).

The implementation process of each embodiment of the present invention is described in detail below with reference to the accompanying drawings, as well as explanations of specific steps and structures.

Embodiment 1:

In the first aspect, the present invention first proposes a method for predicting sequence immunogenicity, as shown in FIG1 , the method comprising:

S1. Extracting sequence features of peptide sequences with known immunogenicity, and constructing a feature matrix based on the sequence features; the sequence features include k-mer features; wherein k is a positive integer;

S2. establishing a regression relationship between the feature matrix and immunogenicity, and obtaining weights corresponding to different sequence features based on the regression relationship;

S3, constructing a sequence immunogenicity prediction model based on the weights;

S4. Predicting the immunogenicity of the sequence to be predicted based on the sequence features of the sequence to be predicted and the sequence immunogenicity prediction model.

This embodiment provides a sequence immunogenicity prediction method, which mines immunogenicity-related features through relevant information in an existing immunogenicity database, thereby establishing a linear relationship between known peptide sequence features and their immunogenicity, and then predicts the immunogenicity of the sequence to be predicted using the sequence features of the sequence to be predicted and the above linear relationship. Compared with the prior art, the immunogenicity prediction results of this embodiment are highly accurate and interpretable.

The implementation process of one embodiment of the present invention is described in detail below with reference to Figures 1-3 and specific explanations of steps S1-S4.

S1. Extract sequence features of peptide sequences with known immunogenicity, and construct a feature matrix based on the sequence features; the sequence features include k-mer features; wherein k is a positive integer.

S11, obtain epitope metadata. In this embodiment, the epitope metadata is to obtain the MHC Ligand Assays (Class II) database from IEDB, and then randomly extract 6338 data from Class II, which contains the immunogenicity information of peptide segments and peptide segments. The following Table 1 is an example of some MHC Ligand Assas data.

Table 1 Some examples of MHC Ligand Assas data

The above 6338 data are marked as samples 1-6338. Among samples 1-6338, there are 3315 samples with positive immunogenicity and 3023 samples with negative immunogenicity. Further, the immunogenicity content of the above samples is converted into a one-hot encoding, that is, the positive immunogenicity samples are marked as 1 and the negative immunogenicity samples are marked as 0. The immunogenicity vector of the above sample peptides is It is expressed as:

Among them, the immunogenicity of each sample peptide ,satisfy ; i represents the sample number, and n represents the total number of samples.

S12. Sequence features of peptide sequences with known immunogenicity are extracted based on epitope metadata, and a feature matrix is constructed based on the sequence features.

Among the existing immunogenicity prediction technologies, sequence-based immunogenicity prediction methods generally mine the rules of the sequence itself, or extract hydrophobicity, charge and other characteristics from the sequence itself, and then combine affinity or other parameters that can represent immunogenicity to build a model, and then predict the immunogenicity of the sequence based on the constructed model. In another immunogenicity prediction technology, the structure-based immunogenicity prediction method, it mainly analyzes the spatial binding of two molecules through the spatial structure of the MHC crystal to predict immunogenicity. However, the above two methods have their own advantages, but both have the problem that the accuracy and interpretability of sequence immunogenicity prediction cannot be taken into account at the same time.

In this embodiment, referring to Figures 2-3, in order to solve this problem, when extracting sequence features of peptide sequences, the selected sequence features include two parts: features based on the sequence itself and features based on the sequence structure. Among them, the features based on the sequence itself include: thermodynamic information features of the sequence and k-mer features of the sequence, and the features based on the sequence structure include CPP features of the sequence. Specifically, for the 6338 sample data obtained above, the CPP feature data set, thermodynamic information feature data set, and k-mer feature data set of the training sample features are constructed respectively.

S121. Extract the thermodynamic information characteristics of peptide sequences based on epitope metadata.

Thermodynamic information-related features of peptide sequences are commonly used physical and chemical indicators in immunogenicity prediction technology. When extracting thermodynamic information-related features of sequences, the corresponding functions are generally used to calculate the thermodynamic information of the sequences.

In one embodiment, the following thermodynamic-related characteristics of the peptide sequence are calculated using the corresponding functions in the peptide package of the R software. The thermodynamic information of the peptide sequence and the corresponding calculation function are specifically shown in Table 2 below.

Table 2 Thermodynamic characteristics and their corresponding functions

The thermodynamic information features of the peptide sequence are calculated according to the functions in Table 2, and a thermodynamic information feature matrix is constructed. In this embodiment, the peptide sequence contains a total of 14 thermodynamic information features, which can be expressed as follows:

in, A dataset representing thermodynamic information features of sample peptide sequences.

S122. Extract k-mer features of peptide sequences based on epitope metadata.

When constructing peptide sequence feature engineering, in addition to using the thermodynamic information characteristics of the peptide sequence, in order to improve the generalization ability of the sequence immunogenicity prediction model, the regular characteristics of the sequence itself must also be learned.

In one embodiment, the k-mer model is used to extract the regular features of the sequence itself. Specifically, the steps for calculating the k-mer of a peptide sequence are as follows:

Define the value of k : k is a predetermined positive integer representing the length of the sequence segment to be considered. The larger the k value, the more complex and sparse the model is. When the model is too complex, the amount of calculation is too large to be implemented. In this embodiment, in order to simplify the model, k = 2 is set, that is, by traversing two amino acid combinations and then counting the number of combinations, the purpose of mining sequence features is achieved.

However, in other embodiments, the value of k is not specifically limited, as long as k is a positive integer, it is within the scope of the present application. For example, in one embodiment, k is any positive integer between 2 and 10. More preferably, the value of k is set to 2 or 3.

Generate k-mer : For a given protein sequence, by starting from one end of the sequence and moving one residue at a time, all possible continuous sequence fragments ( k-mers ) of length k are intercepted, and the number of each fragment type is counted. This step will generate a large number of k-mers , the number of which is related to the length of the sequence and the value of k . For example, there is a peptide numbered SEQID NO:5, and its sequence is: AAAKLD. When k = 2, its k-mer characteristics are shown in Table 3 below.

Table 3 k-mer features of peptide “AAAKLD” when k = 2

In this embodiment, taking 20 amino acids as an example, a total of 400 k-mer features can be obtained, and the constructed k-mer features can be expressed by the formula:

in, The k-mer feature dataset represents the sample peptide sequence. Some examples of k-mer features of the sample peptide sequence are shown in Table 4.

Table 4 Examples of k-mer features of sample peptide sequences

S123. Extract CPP features of peptide sequences based on epitope metadata.

In order to further improve the accuracy and interpretability of peptide sequence immunogenicity prediction, this embodiment not only extracts the characteristics of the peptide sequence itself (i.e., the thermodynamic information characteristics of the peptide sequence and the k-mer characteristics of the peptide sequence), but also adds the structural information of the peptide sequence, and extracts the CPP characteristics of the peptide sequence by comparing the peptide sequence with the TCR data.

Specifically, first, TCR data are directly obtained from existing literature, and some example TCR data are shown in Table 5 below.

Table 5 Some examples of TCR data

Then, through TCR -peptide contact potential profiling, the CPP feature data set was obtained. Specifically, a specific amino acid scoring matrix was used for the 6338 sample peptides obtained above, and each sample peptide sequence was globally compared with the TCR sequence in the database, and then the comparison score was recorded. Examples of some comparison scores are shown in Table 6 below.

Table 6 Example of score results for comparison between sample peptide sequences and TCR sequences

In one embodiment, in order to ensure the effectiveness of the CPP feature and thus improve the accuracy of subsequent immunogenicity prediction, the arithmetic mean, 10% tailed mean, standard deviation, standard error, median, MAD, skewness, abundance, and IQR of the scoring results are statistically analyzed for each peptide sequence, and finally the 10% and 90% quantiles are used as CPP features.

In this embodiment, there are 35 specific amino acid scoring matrices. Each of the above amino acid scoring matrices can construct 11 features, and all CPP features are combined into a CPP feature matrix, which contains a total of 385 CPP features, which can be expressed as follows:

in, CPP feature dataset representing sample peptide sequences.

S124. Generate feature matrix based on thermodynamic information feature data set of peptide sequence, k-mer feature, and CPP feature.

The thermodynamic characteristics, k-mer characteristics, and CPP characteristics of the peptide sequences obtained above are integrated to form a feature matrix , ,in, is the total number of features after integrating the three features, the vector is the feature variable corresponding to each sample, and T represents the transpose of the vector matrix.

In this step, after extracting sequence-based and structure-based features for the peptide sequence, a large number of features are obtained. Generally speaking, 500-8000 features will be obtained.

In this embodiment, for the sequence, its thermodynamic information characteristics, k-mer characteristics, and CPP characteristics are mined at the same time. Not only the physical and chemical index characteristics commonly used in sequence immunogenicity prediction are extracted, but also the regular characteristics of the sequence itself are learned. Furthermore, the structural information of the sequence is also mined. The integration of multiple features not only improves the generalization ability of the immunogenicity prediction model, but also can find the characteristics that affect the immunogenicity results, making the sequence immunogenicity prediction results explainable and traceable.

S2. Establishing a regression relationship between the feature matrix and immunogenicity, and obtaining weights corresponding to different sequence features based on the regression relationship.

S21. Establishing a regression relationship between the characteristic matrix and immunogenicity.

In the sequence immunogenicity prediction technology, the current mainstream method for modeling and analyzing sequence features is still deep learning. However, when using deep models to model and analyze sequence features, although it is relatively easy to debug deep model parameters, the overall amount of computation is huge, and the model has poor interpretability, making it difficult to interpret and trace the results. And because the number of sequence features obtained in the S1 step is huge, this will undoubtedly increase the amount of computation and difficulty of subsequent feature modeling and analysis.

To address this problem, in some embodiments, a linear model is used to model the characteristic matrix of peptide sequences and their immunogenicity. Specifically, a regression relationship is established between the characteristic matrix of the above sample and the immunogenicity result of the sample, which can be expressed as:

in, represents the probability of a positive immunogenicity result; , , ..., is a sequence feature, specifically represented by a vector; is the regression coefficient, which indicates the degree of relationship between immunogenicity and the characteristic.

S22. Obtain weights corresponding to different sequence features based on the regression relationship.

In this example, the immunogenicity of the sample peptides is known, i.e. The value of (for example, the immunogenicity of the above sample peptide is ), then the scoring weight vector can be obtained by elastic network regression

When the value in the brackets around argmin is the minimum value, The value of is the estimated value we need. represents the regression coefficient estimated by the model; represents the sequence feature matrix, that is ; Represents the L1 regularization coefficient, which is used for feature selection; Represents the L2 regularization coefficient, which is used to prevent the model from overfitting.

and Mathematically, it is the same parameter. Generally, this parameter is estimated and does not actually exist. When describing a theory, it is usually used , when describing the estimated results of the parameters, we use .

Re-order , and substitute it into the formula of the above elastic net regression model, then:

The parameter estimation was performed using a 5-fold cross validation method. As shown in Figure 4, the maximum AUC was obtained. Value. Among them, in Figure 4 and are the values corresponding to the immunogenicity and characteristic matrix of the sample peptides respectively. Finally, the elastic net regression formula is used to calculate .

This embodiment uses a sparse linear regression model to solve the MHC prediction problem. Compared with the deep learning model of the existing mainstream technology, the model has a smaller amount of calculation and a faster analysis speed when targeting extremely large feature data. In the process of synthesizing libraries that require large-scale screening, a lot of time can be saved. At the same time, the adjustment of parameters such as thresholds in the model is more flexible.

S3. Constructing a sequence immunogenicity prediction model based on the weights.

Based on the linear relationship between sequence immunogenicity and sequence features obtained in step S2 and the weights corresponding to different sequence features, a sequence immunogenicity prediction model is constructed, which can be expressed as follows:

in, The probability that the immunogenicity of the peptide sequence to be predicted is Positive; is the scoring weight vector corresponding to the features of the sequence to be predicted; The feature matrix representing the peptide sequence to be predicted.

S4. Predicting the immunogenicity of the sequence to be predicted based on the sequence features of the sequence to be predicted and the sequence immunogenicity prediction model.

According to the above step S12, the feature matrix of the sequence to be predicted is constructed , wherein the feature matrix of the sequence to be predicted includes the CPP feature of the sequence to be predicted, the thermodynamic feature of the sequence to be predicted, and the k-mer feature of the sequence to be predicted.

Then, the sequence immunogenicity prediction model obtained in step S3 can be used to predict the immunogenicity of the sequence to be predicted.

Among them, the immunogenicity prediction method provided in this embodiment is applicable to peptides with a length of 8-13 amino acids and can be used for the prediction of immunogenicity. In one embodiment, the sequence to be predicted is selected from a polypeptide or an antibody peptide sequence. More preferably, in another embodiment, the antibody includes ADC.

At this point, the entire process of the sequence immunogenicity prediction method of this embodiment is completed.

In order to verify the superiority of the sequence immunogenicity prediction method of the above embodiment (hereinafter referred to as "this method") in terms of prediction accuracy and prediction efficiency, the following experiments are performed to prove that:

1) Verification of the accuracy of this method.

Substitute the characteristic matrix of the sample peptide obtained in step S1 into the formula for predicting the immunogenicity of the peptide sequence to be predicted: Then the immunogenicity prediction results of each sample peptide are obtained. ,when > 0, the sequence is judged to be an immunogenic sequence. When the value is 0, the sequence is judged to be a non-immunogenic sequence. Then the predicted result is compared with the real immunogenicity result of the original known sample peptide segment, and the results shown in Table 7 can be obtained:

Table 7 Evaluation of the accuracy of immunogenicity prediction results

In the above Table 7,

,

,

,

.

in:

TP (True Positive) means: the prediction is 1 and the true value is also 1;

FP (False Positive) means: the prediction is 1 and the true value is 0;

TN (True Negative) means: the prediction is 0 and the true value is also 0;

FN (False Negative) means: the prediction is 0 and the true value is 1.

From the results in Table 7 above, it can be seen that the accuracy of this method can meet the current requirements of ADC drug development, and the entire process has good traceability, can better interpret the results, and has a large detection throughput and high efficiency.

2) Comparison of the accuracy of this method with existing immunogenicity prediction methods.

Re-obtain 6540 test sequences, named test sequence 1 to test sequence 6540. Use the process described in step S12 to construct the features of the sequence, and obtain the feature matrix Then use the value calculated in step S3 Substitute into the formula .when > 0, the sequence is judged to be an immunogenic sequence. When the value is 0, the sequence is judged to be a non-immunogenic sequence. The prediction results are then compared with the standard results, and the accuracy is shown in Table 8:

Table 8 Evaluation of the accuracy of immunogenicity test results

At the same time, the mainstream tool netMHCPan4.0 was used to predict the immunogenicity of the above 6540 test sequences, and its prediction results were compared with the standard results. The accuracy results are shown in Table 9.

Table 9 Accuracy of netMHCPan immunogenicity prediction results

Comparing Table 8 with Table 9, it can be seen that the accuracy and other indicators of this method are higher than those of netMHCPan4.0, indicating that this method can be used to predict immunogenicity, and the accuracy of immunogenicity prediction is better than the existing mainstream immunogenicity prediction methods.

3) Evaluation of the prediction efficiency of this method.

The time taken by this method and netMHCPan to predict immunogenicity in Experiment 1) and Experiment 2) is compared, and the following Table 10 is obtained. As shown in Table 10, the operating efficiency of this method is much higher than that of netMHCPan, and it can significantly improve the research and development efficiency in the relevant research and development process of ADC drugs. In the experiment to evaluate the prediction efficiency of this method, all tests were run using a single thread, and the CPU used was Intel(R) Xeon(R) Silver 4214R.

Table 10 Comparison of running time between this method and netMHCPan

Embodiment 2:

In a second aspect, the present invention provides a sequence immunogenicity prediction device, as shown in FIG5 , the device comprising:

The immunogenicity prediction model building unit is used to extract sequence features of peptide sequences with known immunogenicity, and construct a feature matrix based on the sequence features; establish a regression relationship between the feature matrix and immunogenicity, and obtain weights corresponding to different sequence features based on the regression relationship; and construct a sequence immunogenicity prediction model based on the weights; wherein the sequence features include k-mer features; wherein k is a positive integer;

The immunogenicity prediction unit is used to predict the immunogenicity of the sequence to be predicted based on the sequence characteristics of the sequence to be predicted and the sequence immunogenicity prediction model.

The present embodiment provides a sequence immunogenicity prediction device, which mines immunogenicity-related features for relevant information of an existing immunogenicity database through an immunogenicity prediction model modeling unit, thereby establishing a linear relationship between known peptide sequence features and their immunogenicity, and using it as a sequence immunogenicity prediction model; then the immunogenicity prediction unit uses the sequence features of the sequence to be predicted and the above-constructed sequence immunogenicity prediction model to predict the immunogenicity of the sequence to be predicted. Compared with the prior art, the immunogenicity prediction results of this embodiment are highly accurate and interpretable.

In some embodiments, the sequence to be predicted is selected from a polypeptide or an antibody.

Furthermore, antibodies include ADCs.

In some embodiments, the polypeptide comprises a peptide segment of 8-13 amino acids in length.

In some embodiments, the immunogenicity prediction model building unit also includes a CPP feature extraction subunit and a thermodynamic related feature extraction subunit; the CPP feature extraction subunit is used to extract the CPP features of the peptide sequence, and the thermodynamic related feature extraction subunit is used to extract the thermodynamic information features of the peptide sequence.

In some embodiments, the thermodynamic information features of the peptide sequence include: the aliphatic index of the protein sequence, the BLOSUM62 index of the protein sequence, the Boman index, the theoretical net charge of the protein sequence, the Cruciani property of the protein sequence, the FASGAI (Factor Analysis Scales of Generalized AminoAcid Information) vector of the protein sequence, the hydrophobic moment of the protein sequence, the hydrophobic index of the protein sequence, the instability index of the protein sequence, the Kidela factor of the protein sequence, the isoelectric point (pI) of the protein sequence, the protFP descriptor of the protein sequence, the VHSE scale of the protein sequence, and the Z-scale parameter of the protein sequence.

Furthermore, in one embodiment, the thermodynamic information characteristics of the peptide sequence are calculated using the corresponding function in the peptide package of R software.

In some embodiments, the regression relationship is a linear regression relationship.

Furthermore, in some embodiments, obtaining weights corresponding to different sequence features based on the regression relationship includes: obtaining weights corresponding to different sequence features based on elastic network regression.

Furthermore, in some embodiments, when obtaining weights corresponding to different sequence features based on elastic network regression, a 5-fold cross validation method is used to perform parameter estimation.

In some embodiments, the immunogenicity prediction model building unit includes a k-mer feature extraction subunit, and the k-mer feature extraction subunit extracts the k-mer feature including:

Customize the value of k ; where k represents the length of the peptide sequence fragment to be considered, and k is a positive integer;

Starting from one end of a given protein sequence, all k amino acid combinations in the protein sequence are traversed, and the types and quantities of all the amino acid combinations are counted.

Furthermore, in some embodiments, k is any positive integer between 2 and 10.

Furthermore, in one embodiment, k =2 or 3.

In some embodiments, the CPP feature extraction subunit extracts the CPP feature of the peptide sequence including:

Each peptide sequence was globally aligned with the TCR sequence, and the quantile of the alignment score was used as the CPP feature.

Furthermore, in some embodiments, each peptide sequence is globally aligned with the TCR sequence based on an amino acid scoring matrix.

Furthermore, in one embodiment, the 10% and 90% quantiles of the alignment score are used as CPP features.

It is understandable that the sequence immunogenicity prediction device provided in this embodiment corresponds to the above-mentioned sequence immunogenicity prediction method, and the explanations, examples, beneficial effects, etc. of the relevant contents can refer to the corresponding contents in the sequence immunogenicity prediction method, which will not be repeated here.

Embodiment 3:

In a third aspect, an embodiment of the present invention provides an electronic device, see Figure 6, the electronic device includes a processor, a communication interface, a memory and a bus, wherein the processor, the communication interface and the memory communicate with each other through the bus, and the processor can call the logic instructions in the memory to execute the steps of the method provided in the first aspect.

It is understandable that the sequence immunogenicity prediction electronic device provided in this embodiment corresponds to the above-mentioned sequence immunogenicity prediction method, and the explanations, examples, beneficial effects, etc. of its relevant contents can refer to the corresponding contents in the sequence immunogenicity prediction method, which will not be repeated here.

Embodiment 4:

In a fourth aspect, an embodiment of the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

It is understandable that the computer-readable storage medium for sequence immunogenicity prediction provided in this embodiment corresponds to the above-mentioned sequence immunogenicity prediction method, and the explanations, examples, beneficial effects, etc. of the relevant contents can refer to the corresponding contents in the sequence immunogenicity prediction method, which will not be repeated here.

In summary, compared with the prior art, the present invention has the following beneficial effects:

1. The present invention provides a method, device, electronic device and storage medium for predicting sequence immunogenicity. Through the relevant information of the existing immunogenicity database, the invention realizes the mining of immunogenicity-related features, thereby establishing a linear relationship between the known peptide sequence features and their immunogenicity, and using it as a sequence immunogenicity prediction model. Then, the sequence features of the sequence to be predicted and the above sequence immunogenicity prediction model are used to predict the immunogenicity of the sequence to be predicted. Compared with the prior art, the immunogenicity prediction results of the present invention are highly interpretable and accurate.

2. The present invention provides a method, device, electronic device and storage medium for predicting sequence immunogenicity. In addition to using commonly used physical and chemical indicators, the k-mer model is used to extract the regularity and characteristics of the sequence itself when predicting the immunogenicity of the sequence. At the same time, the CPP characteristics of the sequence are obtained by comparing the peptide segment with the TCR data, replacing the MHC crystal structure to reflect the structural information of the sequence. Compared with the prior art, the present invention can interpret and trace the immunogenicity prediction results on the basis of ensuring the accuracy of the prediction results by mining and analyzing the characteristics of the peptide sequence itself and the structural characteristics of the sequence.

3. The present invention provides a sequence immunogenicity prediction method, device, electronic device and storage medium. When predicting the immunogenicity of a sequence, a sparse linear regression model is used to assist in immunogenicity prediction. Compared with the deep model in the prior art, it has a small amount of calculation, a fast analysis speed, and strong interpretability of the results. In addition, the model parameters can be flexibly adjusted according to actual needs during the research and development process.

It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (8)

1. A method of predicting sequence immunogenicity, the method comprising:

extracting sequence characteristics of peptide segment sequences with known immunogenicity, and constructing a feature matrix based on the sequence characteristics; the sequence features include k-mer features; wherein k is a positive integer;

Establishing a regression relation between the feature matrix and immunogenicity, and acquiring weights corresponding to different sequence features based on the regression relation;

Constructing a sequence immunogenicity prediction model based on the weights;

predicting the immunogenicity of a sequence to be predicted based on sequence features of the sequence to be predicted and the sequence immunogenicity prediction model;

The sequence feature further comprises: CPP characteristics of the peptide fragment sequence and thermodynamic information characteristics of the peptide fragment sequence;

extracting the k-mer feature comprises:

customizing the value of k; wherein k represents the length of the peptide fragment sequence fragment to be considered, and k is a positive integer;

Starting from one end of a given protein sequence, traversing all k amino acid combinations in the protein sequence, and counting the types and numbers of all amino acid combinations;

extracting the CPP feature of the peptide fragment sequence includes:

Each peptide fragment sequence was globally aligned with the TCR sequence and the quantiles of the alignment score were used as CPP features.

2. The method of claim 1, wherein the regression relationship is a linear regression relationship.

3. The method of claim 2, wherein obtaining weights corresponding to different sequence features based on the regression relationship comprises: and acquiring weights corresponding to the different sequence features based on elastic network regression.

4. A sequence immunogenicity prediction device, the device comprising:

the immunogenicity prediction model modeling unit is used for extracting sequence characteristics of peptide segment sequences with known immunogenicity and constructing a feature matrix based on the sequence characteristics; establishing a regression relation between the feature matrix and immunogenicity, and acquiring weights corresponding to different sequence features based on the regression relation; constructing a sequence immunogenicity prediction model based on the weights; wherein the sequence features comprise k-mer features; wherein k is a positive integer;

an immunogenicity prediction unit for predicting the immunogenicity of a sequence to be predicted based on sequence characteristics of the sequence to be predicted and the sequence immunogenicity prediction model;

The immunogenicity prediction model modeling unit comprises a k-mer feature extraction subunit, a CPP feature extraction subunit and a thermodynamically related feature extraction subunit; the k-mer feature extraction subunit is used for extracting k-mer features of the peptide fragment sequence, the CPP feature extraction subunit is used for extracting CPP features of the peptide fragment sequence, and the thermodynamically related feature extraction subunit is used for extracting thermodynamic information features of the peptide fragment sequence;

The k-mer feature extraction subunit, when extracting the k-mer feature, comprises:

customizing the value of k; wherein k represents the length of the peptide fragment sequence fragment to be considered, and k is a positive integer;

traversing all k amino acid combinations in a given protein sequence starting from one end of the protein sequence, and counting the type and number of all the amino acid combinations;

The CPP feature extraction subunit extracts the CPP feature of the peptide fragment sequence, including:

Each peptide fragment sequence was globally aligned with the TCR sequence and the quantiles of the alignment score were used as CPP features.

5. The apparatus of claim 4, wherein the regression relationship is a linear regression relationship.

6. The apparatus of claim 5, wherein obtaining weights corresponding to different sequence features based on the regression relationship comprises: and acquiring weights corresponding to the different sequence features based on elastic network regression.

7. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor performs the steps of the sequence immunogenicity prediction method according to any of claims 1 to 3.

8. A non-transitory computer readable storage medium having stored thereon a computer program, characterized in that the computer program when executed by a processor implements the steps of the sequence immunogenicity prediction method according to any of claims 1 to 3.