A method for predicting affinity between major histocompatibility complex and peptide

Abstract

The present invention relates to a method for predicting the affinity between a major histocompatibility complex (MHCI) and a polypeptide, belonging to the field of data prediction technology. The method comprises obtaining data on the affinity between the major histocompatibility complex (MHCI) and a polypeptide; preprocessing the data on the affinity between the major histocompatibility complex (MHCI) and a polypeptide; embedding MHCI sequences based on amino acid embedding rules, generating an MHCI sequence matrix for each MHCI sequence; embedding polypeptide sequences based on amino acid embedding rules, generating a polypeptide sequence matrix for each polypeptide sequence, and normalizing the size of the polypeptide sequence matrix; constructing a pretrained prediction model for the affinity between MHCI and polypeptides; training the prediction model based on the MHCI sequence matrix and the preprocessed polypeptide sequence matrix; and predicting affinity based on the trained prediction model. The method of the present invention exhibits better performance in predicting the affinity between MHCI class molecules and 10-mer peptides and 11-mer peptides, resulting in more accurate predictions.

Description

Technical Field

[0001] The present invention relates to the technical field of data prediction, and in particular to a method for predicting the affinity between a major histocompatibility complex and a polypeptide, which can be used for vaccine design, immunotherapy and disease prevention. Background Art

[0002] Recent advances in deep learning techniques have led to the emergence of numerous algorithms for predicting the binding affinity of MHC class I peptides. Most existing computational methods can be categorized into two categories: pan-specific and allele-specific. Allele-specific methods have limited generalization to unseen MHC class I molecules because they use a separate model for each molecule. In contrast, pan-specific methods integrate information about both MHC molecules and peptides into a single model, allowing for simultaneous learning of binding specificities for all MHC molecules.

[0003] Previous prediction methods have mostly used data from the International Immune Epitope Database (IEDB). This data primarily targets binding assays for peptides up to nine amino acids in length, leading algorithms to perform better when processing peptides of this specific length. This bias is problematic because, in real-world biological processes, peptides are not always strictly nine amino acids in length. Many biological phenomena and therapeutic processes involve peptides that may be longer or shorter, and existing algorithms often lack the ability to predict these peptides, failing to meet the needs of practical applications.

[0004] Some algorithms have been developed to more accurately predict the affinity of MHCI and peptides, especially non-9-mer peptides.

[0005] In 2021, Jing Jin et al. proposed an algorithm called DeepAttentionPan in PROTEINS: Structure, Function, and Bioinformatics. It uses convolutional neural networks and attention mechanisms to predict the affinity of MHCI and peptides. The input MHCI and peptide sequences are passed through a submodule called DCNN and an attention module, and the predicted affinity score is finally obtained. With the help of the attention mechanism, higher and more stable prediction performance is achieved.

[0006] In 2023, Mahmood Kalemati et al. proposed a deep learning algorithm called CapsNet-MHC in Communications Biology. CapsNet-MHC first used Blosum 62 to encode and embed MHCI and peptide sequences, and then extracted features through the CNN-Attention module. Then, CapsuleNet multiplied the weight matrix to encode spatial dependencies and extract high-level features from the input vector. Ultimately, it achieved more accurate and interpretable MHCI peptide binding prediction.

[0007] In 2023, Wei Qu et al. proposed a new deep learning-based method, DeepMHCI, in Bioinformatics, which predicts MHC-I peptide binding affinity by incorporating biological knowledge into model design. Through the gating layer and the residual combined with the interactive convolutional layer, the model shows more accurate motifs at different lengths to understand the MHCI peptide binding pattern.

[0008] The above algorithms failed to show good performance in predicting MHCI and non-9 specific peptides, especially peptides with lengths of 10 and 11. Summary of the Invention

[0009] The technical problems to be solved by the present invention are:

[0010] To overcome the shortcomings of the prior art, the present invention provides a method for predicting the affinity of major histocompatibility complex (MHC) peptides to accurately predict the affinity of MHC I with 9-mer, 10-mer, and 11-mer peptides, thereby improving the interpretability and generalizability of the prediction method.

[0011] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0012] A method for predicting the affinity between a major histocompatibility complex and a polypeptide, comprising:

[0013] Obtain data on major histocompatibility complex MHCI and peptide affinity;

[0014] Preprocessing of major histocompatibility complex MHCI and peptide affinity data;

[0015] The MHCI sequence is embedded based on the embedding rules of amino acids, and an MHCI sequence matrix is ​​generated for each MHCI sequence;

[0016] The peptide sequence is embedded based on the embedding rules of amino acids, and a peptide sequence matrix is ​​generated for each peptide sequence, and the size of the peptide sequence matrix is ​​unified;

[0017] Constructing a pre-trained prediction model for the affinity between MHCI and peptides; the prediction model sequentially comprises a convolutional layer, a bidirectional long short-term memory network layer, a pooling layer, a fully connected layer, and an output layer;

[0018] The prediction model was trained based on the MHCI sequence matrix and the preprocessed peptide sequence matrix;

[0019] Predict the affinity of major histocompatibility complex to peptides based on the trained prediction model.

[0020] A further technical solution of the present invention: the pre-processing of the data of major histocompatibility complex MHC I and polypeptide affinity comprises:

[0021] Major histocompatibility complex MHCI and peptide affinity IC 50 The value is converted to an affinity score in [0,1]: Wherein, nM is nanomolar;

[0022] For each MHCI sequence, 25 random natural peptides with lengths of 8, 9, 10, and 11 were introduced, and the affinity score after introduction was set to 0.1. They were used as artificial negative samples to participate in the training of the prediction model, but not in the verification of the prediction model.

[0023] A further technical solution of the present invention: The embedding rules for amino acids are specifically as follows:

[0024] Different combinations of 3 0s and 3 1s are used to represent 20 amino acids;

[0025] Different combinations of two 0s and two 1s are used to distinguish the four charge properties of amino acids: the four charge properties include: if it is non-polar, it is represented by [0, 1, 1, 1]; if it is polar and uncharged, it is represented by [0, 1, 0, 1]; if it is polar and positively charged, it is represented by [0, 1, 1, 0]; if it is polar and negatively charged, it is represented by [1, 0, 0, 1];

[0026] For unknown residues X, specify the embedding representation method.

[0027] A further technical solution of the present invention: the size unification processing of the polypeptide sequence matrix includes:

[0028] If the length of the peptide sequence is less than 11, the matrix is ​​padded with 0 to fill the matrix size to 11×10

[0029] If the peptide sequence length is equal to 11, no matrix operation is performed;

[0030] If the length of the polypeptide sequence is greater than 11, the first 11 rows of the matrix are truncated to obtain an 11×10 matrix.

[0031] A further technical solution of the present invention is: the prediction model of the affinity between MHCI and polypeptide is specifically:

[0032] Convolutional layer: includes four parallel convolution kernels of different sizes. All layers use the ReLU function as the activation function. Before each convolution kernel is convolved, the MHCI and peptide sequence matrices are padded with zeros in two different ways. The MHCI and peptide sequence information obtained by convolution with the same zero-padding method are concatenated. The data obtained by the four convolution kernels are concatenated and output.

[0033] Bidirectional LSTM network layer: It includes two Bidirectional LSTM network blocks, one of which consists of a single-layer Bidirectional LSTM network and the other consists of a two-layer Bidirectional LSTM network. The output of the convolutional layer is input into the two Bidirectional LSTM network blocks respectively, and the output of the two Bidirectional LSTM network blocks is concatenated.

[0034] Pooling layer: performs maximum pooling on the output of the feature extraction layer;

[0035] Fully connected layer: includes four hidden layers of different dimensions. All four hidden layers use the ReLU function as the activation function, and a dropout ratio of 0.5 is added between the first and second layers.

[0036] Output layer: The results of the four hidden layers of the fully connected layer are cascaded to output the affinity scores of MHCI and peptides. This layer uses the sigmoid function as the activation function.

[0037] A further technical solution of the present invention is: the training of the prediction model based on the MHCI sequence matrix and the pre-processed polypeptide sequence matrix comprises:

[0038] The MHCI sequence matrix, the preprocessed peptide sequence matrix and the corresponding affinity scores were divided into five groups, each group was used as a training set, and the remaining four groups were used as validation sets for cross-training.

[0039] A further technical solution of the present invention is to introduce Mean Squared Error as a loss function to update network parameters during the training process until the loss function MSE is minimized. The MSE formula is as follows:

[0040]

[0041] Among them, y i Indicates the true affinity score between MHCI and the peptide, It represents the predicted affinity score of the prediction model for the i-th MHCI and peptide, and n is the size of the training set.

[0042] A further technical solution of the present invention is: during the training process, an Adam optimizer is introduced to optimize the network parameters, the initial learning rate is 0.001, and the learning rate is adjusted to half of the original every 10 rounds during the training process.

[0043] A computer system, characterized in that it includes: one or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the above-mentioned method.

[0044] A computer-readable storage medium is characterized by storing computer-executable instructions, which are used to implement the above method when executed.

[0045] The beneficial effects of the present invention are:

[0046] The method provided by the present invention for predicting the affinity between a major histocompatibility complex and a polypeptide has the following advantages over the prior art:

[0047] 1. The present invention uses an equal number of 0s and 1s to embed amino acids, which can make the characteristic information of different sequences more obvious.

[0048] 2. Since the present invention uses four convolution kernels of different sizes, it can capture the information and mutual dependence between different numbers of adjacent amino acids.

[0049] 3. The present invention uses two different zero-filling methods for four different sizes of convolution kernels before the convolution operation and connects them after the convolution, so as to better capture the first and last amino acids and the overall continuous sequence information.

[0050] The method of the present invention shows better performance in predicting the affinity of MHC class I molecules with 10-mer peptides and 11-mer peptides, and the prediction is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The accompanying drawings are only for the purpose of illustrating particular embodiments and are not to be considered limiting of the present invention. Like reference symbols denote like parts throughout the drawings.

[0052] Figure 1 Flowchart of the implementation of the method of the present invention.

[0053] Figure 2 A framework diagram for implementing the method of the present invention.

[0054] Figure 3Detailed diagram of the convolutional layer of our invention. DETAILED DESCRIPTION

[0055] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only intended to illustrate the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

[0056] like Figure 1 As shown, the embodiment of the present invention provides a method for predicting the affinity between the major histocompatibility complex and a polypeptide, which is implemented by the following steps:

[0057] Step 1: Obtain the data of MHCI binding to peptides from the International Immune Epitope Database (IEDB);

[0058] This example uses a dataset of MHCI-peptide binding affinities obtained from the International Immune Epitope Database (IEDB) as an example. The dataset contains 185,985 records of MHCI-peptide binding. The dataset includes 153 MHCI molecules, including 7 BoLA, 1 Gogo, 7 H-2, 104 HLA, 19 Mamu, 11 Patr, and 4 SLA molecules.

[0059] Step 2: Preprocessing of the acquired MHCI-peptide binding data

[0060] (2a) The affinity IC of the obtained MHCI binding to the peptide 50 The value is converted into an affinity score in [0,1] using the following formula:

[0061] 1-log(IC 50 nM) / log(50000)

[0062] Wherein, nM is nanomolar;

[0063] (2b) After converting all affinities to [0,1], 25 random natural peptides with lengths of 8, 9, 10, and 11 were introduced for each MHCI molecule, and the affinity score after conversion was set to 0.1. These samples were used as artificial negative samples to participate in the training of the model, but not in the validation of the model;

[0064] Step 3: Data Partitioning

[0065] The 2101285 data items are divided into five groups in order, of which the first group contains 40534 items, the second group contains 40956 items, the third group contains 39186 items, the fourth group contains 40048 items, and the fifth group contains 40561 items.

[0066] Step 4: Embedding MHCI and peptide sequences

[0067] The amino acids are embedded using a combination of tens 0s and 1s, as follows:

[0068] Different combinations of 3 0s and 3 1s are used to represent 20 amino acids;

[0069] Different combinations of two 0s and two 1s are used to distinguish the four charge properties of amino acids: if it is non-polar, it is represented by [0,1,1,1]; if it is polar and uncharged, it is represented by [0,1,0,1]; if it is polar and positively charged, it is represented by [0,1,1,0]; if it is polar and negatively charged, it is represented by [1,0,0,1].

[0070] In particular, for unknown residue X, specify the embedding representation method;

[0071] The final amino acid embedding rules are as follows:

[0072] A:[1,1,1,0,0,0,0,0,1,1],R:[1,1,0,1,0,0,0,1,1,0],N:[1,1,0,0,1,0, 0,1,0,1],D:[1,1,0,0,0,1,1,0,0,1],C:[1,0,1,1,0,0,0,1,0,1],Q:[1,0 ,1,0,1,0,0,1,0,1],E:[1,0,1,0,0,1,1,0,0,1],G:[1,0,0,1,1,0,0,0,1, 1],H:[1,0,0,1,0,1,0,1,1,0],I:[1,0,0,0,1,1,0,0,1,1],L:[0,1,1,1,0 ,0,0,0,1,1],K:[0,1,1,0,1,0,0,1,1,0],M:[0,1,1,0,0,1,0,0,1,1],F:[ 0,1,0,1,1,0,0,0,1,1],P:[0,1,0,1,0,1,0,0,1,1],S:[0,1,0,0,1,1,0,1 ,0,1],T:[0,0,1,1,1,0,0,1,0,1],W:[0,0,1,1,0,1,0,0,1,1],Y:[0,0,1, 0,1,1,0,0,1,1],V:[0,0,0,1,1,1,0,0,1,1],X:[1,1,1,1,0,0,0,0,0,0].

[0073] The MHCI sequence is embedded according to the embedding rules of amino acids, and a 34×10 matrix is ​​generated for each MHCI sequence;

[0074] The peptide sequence is embedded according to the embedding rules for amino acids. Each peptide sequence generates an L×10 matrix, where L is the length of the peptide. The shape of the generated peptide matrix is ​​unified. For peptides with a length of less than 11, the corresponding matrix is ​​padded to 11×10 by zero padding. For peptides with a length greater than 11, the matrix consisting of the first 11 amino acids is used as the corresponding peptide embedding matrix. For peptides with a length equal to 11, no peptide embedding matrix operation is performed.

[0075] Step 5: Construct a prediction model for the affinity between MHCI and peptides, such as Figure 2 As shown in the figure, it includes convolutional layer, bidirectional long short-term memory network layer, pooling layer, fully connected layer, and output layer in sequence; the details are as follows:

[0076] (5a) For four different convolution kernels, the two zero-filling methods corresponding to each convolution kernel are as follows:

[0077]

[0078] (5b) Convolutional layers with 128 kernels and kernel sizes of 1×10, 3×10, 5×10, and 7×10 are established. All layers use the ReLU function as the activation function.

[0079] (5c) The MHCI and peptide matrices embedded in step 3 are padded with zeros according to the method of step 5a. After padding, they are passed through the convolution layer and connected. Finally, four data with a dimension of 128 and a shape of 45×1 are obtained. The data output from these four convolutions are connected in the order of the convolution kernel size from small to large, and finally the data with a dimension of 128 and a shape of 360×1 is obtained; Figure 3 As shown;

[0080] (5d) Create a feature extraction component with an input dimension of 128, consisting of two bidirectional LSTM blocks and a pooling layer. The first block consists of a single-layer bidirectional LSTM network, and the second block consists of two bidirectional LSTM layers. Each layer contains 128 hidden units. The results obtained by passing through the two bidirectional LSTM blocks are concatenated to obtain an output data with a dimension of 512 and a shape of 360×1. Finally, the input data is passed through a pooling layer with a kernel size of 360×1.

[0081] (5e) Establish a combined score output layer consisting of a cascade of 512-dimensional input, 1024-dimensional first hidden layer, 768-dimensional second hidden layer, 640-dimensional third hidden layer, 340-dimensional fourth hidden layer, and 1-dimensional output. Except for the last layer, the remaining layers use the ReLU function as the activation function, and the last layer uses the sigmoid activation function. Dropout with a ratio of 0.5 is added between the first and second layers.

[0082] Step 6: Train a model to predict MHCI peptide affinity

[0083] (6a) According to the grouping in step (3), each group is used as a validation set, and the remaining four groups are used as training sets. The MHCI and peptide embedding matrix obtained in step (4) and the true score converted in step (2a) are input into the model, and finally five models and the validation results of the complete data set are obtained;

[0084] (6b) Introducing Mean Squared Error as a loss function to update network parameters during model training until the loss function MSE is minimized, thus completing model training; the MSE formula is as follows:

[0085]

[0086] Among them, y i Indicates the true affinity score between MHCI and the peptide, represents the predicted affinity score of the model for the i-th MHCI and peptide, and n is the size of the training set.

[0087] (6c) The Adam optimizer (Adaptive Moment Estimation) is introduced to optimize network parameters during model training. The initial learning rate is 0.001, and the learning rate is adjusted to half of the original value every 10 rounds during training.

[0088] The following is a detailed description of the effects of this application in conjunction with simulation:

[0089] 1. Simulation conditions:

[0090] The computer hardware CPU of the simulation experiment is Intel Core(TM) i7 and the computer hardware memory is 32G;

[0091] Computer software: Python 3.8 and Pycharm integrated development software on WINDOWS11 system.

[0092] 2. Simulation content:

[0093] Simulation 1: Classify peptides by length, using AUROC (AUC) and Pearson correlation coefficient (PCC) as indicators to evaluate the predictive ability of the model. The evaluation indicators of the present invention and three existing methods, DeepMHCI, Deep AttentionPan, and CapsNet-MHC, were compared on the MHCI-peptide binding dataset obtained from IEDB. The results are shown in Table 1:

[0094] Table 1 Evaluation of the present invention and three existing methods under different polypeptide lengths

[0095]

[0096] As can be seen from Table 1, the performance of the present invention in predicting the affinity of MHCI for 10-mer and 11-mer peptides is significantly higher than that of the other three methods. When predicting 9-mer peptides, both the present invention and DepMHCI significantly outperform DeepAttentionPan and CapsNet-MHC. Good performance is maintained when predicting 8-mer peptides and peptides with ≥12-mer affinity. These simulation results demonstrate that the present invention maintains high accuracy when predicting the affinity of peptides of varying lengths for MHCI.

[0097] Simulation 2: To demonstrate the generalization and universal applicability of the present invention, the present invention was applied to predict the affinity of MHCII and the peptide, and a similar five-fold split was performed for validation. Similarly, AUROC and Pearson correlation coefficient (PCC) were used as indicators to evaluate the predictive ability of the model. The present invention was compared with DeepAttentionPan and CapsNet-MHC in terms of evaluation indicators. The results are shown in Table 2:

[0098] Table 2 Evaluation of the present invention and two existing methods in predicting the affinity between MHCII and peptides

[0099]

[0100] As can be seen from Table 2, the performance of the present invention in predicting the affinity between MHCII and peptides is significantly higher than that of the other two methods. The simulation results show that the model structure of the present invention has strong generalization ability and universal applicability.

[0101] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present invention, and these modifications or replacements should all be included in the scope of protection of the present invention.

Claims

1. A method for predicting the affinity between a major histocompatibility complex and a polypeptide, characterized in that: include: Obtain data on major histocompatibility complex MHCI and peptide affinity; Preprocessing of major histocompatibility complex MHCI and peptide affinity data; The MHCI sequence is embedded based on the embedding rules of amino acids, and an MHCI sequence matrix is ​​generated for each MHCI sequence; The peptide sequence is embedded based on the embedding rules of amino acids, and a peptide sequence matrix is ​​generated for each peptide sequence, and the size of the peptide sequence matrix is ​​unified; Constructing a pre-trained prediction model for the affinity between MHCI and peptides; the prediction model sequentially comprises a convolutional layer, a bidirectional long short-term memory network layer, a pooling layer, a fully connected layer, and an output layer; The prediction model of the affinity between MHCI and polypeptide is specifically: Convolutional layer: includes four parallel convolution kernels of different sizes. All layers use the ReLU function as the activation function. Before each convolution kernel is convolved, the MHCI and peptide sequence matrices are padded with zeros in two different ways. The MHCI and peptide sequence information obtained by convolution with the same zero-padding method are concatenated. The data obtained by the four convolution kernels are concatenated and output. Bidirectional LSTM network layer: It includes two Bidirectional LSTM network blocks, one of which consists of a single-layer Bidirectional LSTM network and the other consists of a two-layer Bidirectional LSTM network. The output of the convolutional layer is input into the two Bidirectional LSTM network blocks respectively, and the output of the two Bidirectional LSTM network blocks is concatenated. Pooling layer: performs maximum pooling on the output of the feature extraction layer; Fully connected layer: includes four hidden layers of different dimensions. All four hidden layers use the ReLU function as the activation function, and a dropout ratio of 0.5 is added between the first and second layers. Output layer: cascades the results of the four hidden layers of the fully connected layer to output the affinity scores of MHCI and peptides. This layer uses the sigmoid function as the activation function; The prediction model was trained based on the MHCI sequence matrix and the preprocessed peptide sequence matrix; Predict the affinity of major histocompatibility complex to peptides based on the trained prediction model.

2. The method for predicting the affinity between a major histocompatibility complex and a polypeptide according to claim 1, wherein: The pre-processing of the data of major histocompatibility complex MHCI and polypeptide affinity includes: Major histocompatibility complex MHCI and peptide affinity The value is converted to an affinity score in [0, 1]: = ;in, for Namor; For each MHCI sequence, 25 random natural peptides with lengths of 8, 9, 10, and 11 were introduced, and the affinity score after introduction was set to 0.

1. They were used as artificial negative samples to participate in the training of the prediction model, but not in the verification of the prediction model.

3. The method for predicting the affinity between a major histocompatibility complex and a polypeptide according to claim 1, wherein: The embedding rules for amino acids are specifically as follows: Different combinations of 3 0s and 3 1s are used to represent 20 amino acids; Different combinations of two 0s and two 1s are used to distinguish the four charge properties of amino acids: the four charge properties include: if it is non-polar, it is represented by [0, 1, 1, 1]; if it is polar and uncharged, it is represented by [0, 1, 0, 1]; if it is polar and positively charged, it is represented by [0, 1, 1, 0]; if it is polar and negatively charged, it is represented by [1, 0, 0, 1]; For unknown residues X, specify the embedding representation method.

4. The method for predicting the affinity between a major histocompatibility complex and a polypeptide according to claim 1, wherein: The size unification processing of the polypeptide sequence matrix includes: If the length of the peptide sequence is less than 11, the matrix is ​​padded with 0 to fill the matrix size to 11 10 If the peptide sequence length is equal to 11, no matrix operation is performed; If the length of the peptide sequence is greater than 11, the first 11 rows of the matrix are cut off to obtain 11 10 matrix.

5. The method for predicting the affinity between a major histocompatibility complex and a polypeptide according to claim 1, wherein: The method of training the prediction model based on the MHCI sequence matrix and the pre-processed polypeptide sequence matrix includes: The MHCI sequence matrix, the preprocessed peptide sequence matrix and the corresponding affinity scores were divided into five groups, each group was used as a training set, and the remaining four groups were used as validation sets for cross-training.

6. The method for predicting the affinity between a major histocompatibility complex and a polypeptide according to claim 5, characterized in that: During the training process, the Mean Squared Error is introduced as the loss function to update the network parameters until the loss function MSE is minimized. The MSE formula is as follows: in, Indicates the true affinity score between MHCI and the peptide, Indicates the prediction model for i The predicted affinity scores of the MHCIs and peptides are shown in Figure 2, where n is the size of the training set.

7. The method for predicting the affinity between a major histocompatibility complex and a polypeptide according to claim 5, characterized in that: During the training process, the Adam optimizer was introduced to optimize the network parameters. The initial learning rate was 0.001, and the learning rate was adjusted to half of the original value every 10 rounds during the training process.

8. A computer system, characterized in that include: One or more processors, and a computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are enabled to implement the method of claim 1.

9. A computer-readable storage medium, characterized in that Computer-executable instructions are stored, and when the instructions are executed, they are used to implement the method of claim 1.

Images