A BCE Prediction Method Based on a Hybrid Deep Learning Strategy

Abstract

The present invention relates to the technical field of immunopharmaceutical R & D, and discloses a BCE prediction method based on a hybrid deep learning strategy, which includes the following steps: Step 1, construct a BCE prediction model; Step 2, input a protein sequence into the BCE prediction model; Step 3, use the prediction model to output the result of whether the protein sequence is a BCE. The BCE prediction model in Step 1 includes a database, a feature extraction module, a feature processing module, and a forward neural network module; wherein: the database is used to obtain a training data set; the feature extraction module is used to extract features from the data and obtain four groups of features; the feature processing module is used to process the extracted features. The deep learning prediction method proposed by the present invention represents a protein with sequence features and structural features, and uses a four-channel deep neural network to extract sequence information and structural information respectively, and the same model can be used to predict linear epitopes.

Description

Technical Field

[0001] The present invention relates to the technical field of immunopharmaceutical R & D, and particularly to a BCE prediction method based on a hybrid deep learning strategy. Background Art

[0002] B-cell epitopes (BCEs) are one of the key points in immunopharmaceutical R & D. It refers to the part of an antigen recognized by the variable region of a B-cell antibody. Generally, BCEs can be divided into two categories: linear (continuous) and conformational (discontinuous). Linear BCEs exist on the primary sequence of an antigen molecule and are composed of a continuous sequence fragment of amino acid residues. The identification of epitopes is very important for vaccine design, immunodiagnostic tests, and the production of synthetic antibodies. Although natural linear epitopes are few, there are still many studies on linear epitopes because they are very important for the R & D of synthetic peptide vaccines.

[0003] Since the experimental determination of BCEs is both time-consuming and expensive, computational prediction can play a key role in the development of new vaccines and drugs against common viral pathogens (such as human immunodeficiency virus, hepatitis, or influenza virus).

[0004] At present, generative artificial intelligence technologies represented by large models, through unsupervised learning methods (autoencoders, generative adversarial networks, masked language modeling, contrastive learning, etc.), train large-scale deep learning models through massive data, and can learn the general features and representations of data from unlabeled data. Based on pre-trained or fine-tuned protein language large models, excellent performance has been demonstrated in protein structure, function, and other downstream tasks. Inputting the primary sequence into the protein language large model, the obtained embedded representation can capture some biophysical characteristics of the protein sequence or predict the three-dimensional structure of the protein.

[0005] Currently, for the feature representation of proteins, most use the features extracted from protein sequences, such as the amino acid pair antigenicity scale (AAP), amino acid triplet antigenicity scale (AAT), amino acid composition (AAC), etc. to calculate BCE features, as well as physicochemical properties, structural composition, etc. The computational models are mainly random forests, support vector machines, extreme gradient boosting, etc., and the focus is more on input feature selection or feature dimensionality reduction.

[0006] Starting from the primary sequence of a protein, using generative artificial intelligence technologies such as large models, more sequence features can be obtained. Combining sequence AAP, AAT, AAC, amino acid local structure information, etc., can better represent the features of proteins.

[0007] Epitopes are the basis of protein antigenicity. Determining B-cell epitopes has important guiding significance for vaccine and drug design. Through experimental methods, antigen epitopes are measured, including X-ray diffraction, fluorescence polarization technology, phage display technology, peptide scanning technology, etc. These methods are relatively cumbersome and involve a very large amount of work. With the development of computer technology and the increasing expansion of biological databases, it has become possible to summarize the sequence and structural characteristics of antigen epitopes from existing data and predict possible epitopes through computational means. Computational methods can save a large amount of experimental and labor costs.

[0008] Predicting BCE using computational methods can be described as a binary classification problem of determining whether a protein sequence is an epitope. Summary of the Invention

[0009] To solve the technical problems raised in the background art, the present invention provides a BCE prediction method based on a hybrid deep learning strategy.

[0010] The present invention is implemented by the following technical solutions: A BCE prediction method based on a hybrid deep learning strategy, comprising the following steps:

[0011] Step 1, construct a BCE prediction model;

[0012] Step 2, input a protein sequence into the BCE prediction model;

[0013] Step 3, use the prediction model to output the result of whether the protein sequence is a BCE.

[0014] Specifically, the BCE prediction model in step 1 includes a database, a feature extraction module, a feature processing module, and a forward neural network module;

[0015] Among them:

[0016] The database is used to obtain training data sets;

[0017] The feature extraction module is used to extract features from data and obtain four groups of features;

[0018] The feature processing module is used to process the extracted features;

[0019] The forward neural network module is used to merge multiple processed features and output a prediction value.

[0020] Specifically, the training data in the database comes from the IEDB database and the Bcipep database; the verification data in the database is also selected from the EDB database and the Bcipep database.

[0021] Specifically, the feature extraction module is used for ProT5 representation features, ESM-2 representation features, DSSP features and sequence features, and the sequence features are spliced by orthogonal coding of sequence residues, AAT, AAP, and AAC;

[0022] in:

[0023] AAP is an amino acid antigen pair, AAT is an amino acid trimer, and AAC is an amino acid composition;

[0024] ProtT5 indicates that the features are extracted using the protein language model ProtT5; the specific model is ProtT5-XL-UniRef50, the input of which is the primary sequence of the protein, and the output is the i-th residue R in the corresponding sequence. i The feature representation vector X i , X i Dimension 1024;

[0025] ESM-2 representation features are extracted using the protein language pre-trained large model ESM-2; the specific model is Esm2_t33_650M_UR50D, the input of this model is the primary sequence of the protein, and the output is the i-th residue R in the corresponding sequence. i The feature representation vector X i , X i Dimension 1280;

[0026] DSSP features are extracted using the protein structure prediction model ESMFold; ESMFold is used to predict the three-dimensional structure of the protein, and the protein file is parsed using DSSP software to obtain the secondary structure information, (φ, ψ) dihedral angle information, and solvent accessible surface area information of the eight states of the protein; the secondary structure is represented by 8-dimensional one-hot orthogonal encoding data; a dihedral angle is represented by a pair of sine and cosine functions, as shown in formula (1), and the (φ, ψ) dihedral angle is represented by 4-dimensional data; the solvent accessible surface area is normalized using maximum-minimum, as shown in formula (2), and the minimum value of the residue solvent accessible surface area is 0; the above data has a total of 13 dimensions;

[0027] y1=sin(θ),y2=cos(θ)(1)

[0028]

[0029] Sequence features were extracted using the following method;

[0030] Given that there are too few non-zero data in the above one-hot orthogonal vector encoding:

[0031] First, use formula (3) to map the sparse code to the dense code by the autoencoder; take the parameter W in the encoder as the corresponding new code vector, and the data dimension is 20 dimensions;

[0032] In the dataset of the characteristic representation of amino acid antigen pairs (AAPs), the frequency of adjacent amino acids in the epitope sequence appearing in pairs and the frequency of the amino acid pair in the non-epitope sequence The ratio is calculated as in formula (4) and normalized to [-1, 1] according to the formula The amino acid antigen pair feature is represented by four values: the AAP feature value, the maximum value in AAP, the minimum value in AAP, and the average value of AAP, with a total of 4 dimensions

[0033]

[0034] where and The calculation methods are as in formulas (5) and (6).

[0035]

[0036] where is the number of occurrences of a pair of amino acids in the epitope sequence, is the number of all amino acid pairs in the epitope sequence.

[0037]

[0038] where is the number of occurrences of a pair of amino acids in the non-epitope sequence, is the number of all amino acid pairs in the non-epitope sequence.

[0039] The amino acid triplet (AAT) feature represents the ratio of the occurrence frequency of three consecutive amino acid residues in the epitope sequence in the dataset to the occurrence frequency of the triplet in the non-epitope sequence. The calculation method and normalization method are the same as those of the AAP feature; the AAT feature representation includes 4 sets of values: the AAT feature value, the maximum value, the minimum value, and the average value of the residue, with a total of 4 dimensions;

[0040] The amino acid composition (AAC) feature represents the relative proportion of each amino acid in each protein sequence in the dataset, and the calculation method is as in formula (7).

[0041]

[0042] where R i is the number of amino acids of type i in the sequence, N is the sequence length, and the AAC feature is 1-dimensional;

[0043] Considering the convenience of batch input of data during the training of deep learning models, all sequence lengths are aligned to 71 according to the longest sequence in all test sets, and zeros are padded for sequences shorter than 71.

[0044] The residue features are 71 * 1024 for ProtT5, 71 * 1280 for ESM-2, and 71 * 13 for DSSP. The residue encoding, AAT, AAP, and AAC are concatenated to a total of 71 * 29.

[0045] Specifically, the feature processing module includes: module (a) structure feature processing module, module (b) ESM-2 large model feature processing module, module (c) ProtT5 large model feature processing module, and module (d) sequence feature processing module.

[0046] Specifically, in the structure feature processing module: the three-dimensional structure predicted by ESMFold is parsed into 13-dimensional structure features using DSSP software; the 71 * 13 feature matrix of a single sequence is extracted using a full-size two-dimensional convolutional neural network, and the two-dimensional matrix is converted into a one-dimensional vector; the convolutional kernel size is (71, 13), the activation function is ReLU, and the output is 256 dimensions; the calculation process is as shown in formula (8).

[0047] F1 = ReLU(Conv(W * x 1:71 + b)) (8)

[0048] Specifically, the 1280-dimensional residue representation features output by the ESM-2 large model are input into a two-layer bidirectional LSTM network, and then a one-layer forward attention network is used to map the two dimensions to a one-dimensional vector.

[0049] First, the bidirectional LSTM network focuses on the global information of the protein to capture the long-range dependencies of the protein sequence; the formulaic description of the unidirectional LSTM model is as shown in (9).

[0050]

[0051] Where σ is the activation function, usually the Sigmoid function; ⊙ represents element-wise multiplication of matrices; x t is the network input at time t; i t , f t , o t , c t and h t respectively represent the input gate, forget gate, output gate, internal memory unit, and output at time t; h t-1 is the output of the previous time; c t-1 is the output of the internal memory unit of the previous time; the rest are learnable parameters of the neural network.

[0052] The first-layer LSTM network takes a 1280-dimensional input and outputs 128 dimensions unidirectionally. When the data of the forward LSTM and the backward LSTM converge, a merging operation is performed in the lowest dimension, as shown in formula (10), and the output is 256 dimensions;

[0053] The second-layer LSTM network takes a 256-dimensional input and outputs 128 dimensions unidirectionally. When the forward and backward data converge, a merging operation is also performed in the lowest dimension, and the output is 256 dimensions;

[0054] The forward attention network is mainly used to convert a two-dimensional matrix into a one-dimensional vector; h t ′ is the output of the bidirectional LSTM, representing the t-th amino acid residue in the current sequence; through a forward neural network as shown in formula (11), the attention weight of the residue t in the current feature representation of the sequence is obtained through formula (12); for all residue feature representations in the sequence, weighted summation is performed, as shown in formula (13), to achieve the conversion from two-dimensional feature representation to one-dimensional. The input data dimension of the forward attention network is 71 * 256, and the output is 1 * 256 dimensions.

[0055] e t =h t ′W t (11)

[0056]

[0057]

[0058] Specifically, in the ProtT5 representing the feature extraction module, it includes two layers of bidirectional LSTM networks and one layer of forward attention network. The module input is 1024 dimensions, and the module output F3 is also 256 dimensions, and it also maps two-dimensional features to one-dimensional vectors.

[0059] Specifically, in the sequence feature extraction module, for the encoded features, AAP, AAT, and AAC features of the sequence, a full-size convolutional neural network is used for feature extraction; the input of each sequence is of size 71 * 29, the convolutional kernel size is (71, 29), the activation function is ReLU, and the output feature F4 is 256 dimensions.

[0060] Specifically, the forward neural network module performs a data merging operation on the output one-dimensional vectors of the above four feature processing modules. The output dimension of each module is 256, and the merged dimension is 1024, as shown in formula (14);

[0061] F = concat(F1, F2, F3, F4) (14)

[0062] It includes two-layer forward neural networks, as shown in formulas (15) and (16) respectively. The activation function of the first forward neural network is tanh, and the output dimension is 512; the second uses the Sigmoid function as shown in formula (17), and the output dimension is 1. The output result is 0 or 1, where: 0 represents non-epitope and 1 represents epitope;

[0063] F′ = tanh(W * F + b) (15)

[0064]

[0065] The error loss between the model prediction value and the true value is described by the binary cross-entropy loss function, as shown in formula (18);

[0066]

[0067] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0068] The present invention provides a BCE prediction method based on a hybrid deep learning strategy, using different deep neural networks such as convolutional neural networks, bidirectional long short-term memory networks, and forward attention mechanisms to fit and extract different types of sequence features, realizing the prediction of B-cell linear epitopes for classification tasks. The prediction model trained by the present invention can run on an ordinary user-level computer.

[0069] The deep learning prediction method proposed by the present invention represents a protein using sequence features and structural features, and uses a four-channel deep neural network to extract sequence information and structural information respectively, and can predict linear epitopes using the same model. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a flowchart of the BCE prediction method based on the hybrid deep learning strategy proposed by the present invention;

[0071] Figure 2 It is a structural diagram of the BCE prediction model based on the hybrid deep learning strategy proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0072] Next, in combination with the drawings and specific embodiments, the present invention will be further described. It should be noted that, on the premise of no conflict, the following described embodiments or technical features can be combined arbitrarily to form new embodiments.

[0073] Embodiment:

[0074] Refer to Figure 1 - Figure 2 A BCE prediction method based on a hybrid deep learning strategy proposed in this solution includes the following steps:

[0075] Step 1: Construct a BCE prediction model;

[0076] Step 2: Input a protein sequence into the BCE prediction model;

[0077] Step 3: Use the prediction model to output the result of whether the protein sequence is BCE.

[0078] Specifically, the BCE prediction model in Step 1 includes a database, a feature extraction module, a feature processing module, and a forward neural network module;

[0079] Among them:

[0080] The database is used to obtain training datasets;

[0081] The feature extraction module is used to extract features from the data and obtain four groups of features;

[0082] The feature processing module is used to process the extracted features;

[0083] The forward neural network module is used to merge multiple processed features and output a prediction value.

[0084] Specifically, the training data in the database comes from the IEDB database and the Bcipep database; the validation data in the database is also selected from the EDB database and the Bcipep database.

[0085] Specifically, the feature extraction module is used for ProtT5 representation features, ESM-2 representation features, DSSP features, and sequence features. The sequence features are spliced by sequence residue orthogonal coding, AAT, AAP, and AAC;

[0086] Among them:

[0087] AAP is the amino acid antigen pair, AAT is the amino acid trimer, and AAC is the amino acid composition;

[0088] The ProtT5 feature is extracted using the protein language large model ProtT5; the specific model is ProtT5-XL-UniRef50. The input of this model is the primary sequence of the protein, and the output is the feature representation vector X of the i-th residue R in the corresponding sequence i of, and the dimension of X i is 1024; i

[0089] The ESM-2 feature is extracted using the protein language pre-trained large model ESM-2; the specific model is esm2_t33_650M_UR50D. The input of this model is the primary sequence of the protein, and the output is the feature representation vector X of the i-th residue R in the corresponding sequence i i of, and the X i of, and the Xi with a dimension of 1280;

[0090] The DSSP features are extracted using the large - model ESMFold for protein structure prediction. The three - dimensional structure of the protein is predicted using ESMFold, and the DSSP software is used to parse the protein file to obtain the secondary - structure information of the eight - state of the protein, the (φ,ψ) dihedral - angle information, and the solvent - accessible surface - area information. The secondary structure is represented by 8 - dimensional one - hot orthogonal encoded data; one dihedral angle is represented by a pair of sine and cosine functions, as shown in formula (1), and the (φ,ψ) dihedral angle is represented by 4 - dimensional data; the solvent - accessible surface area is normalized by max - min normalization, as shown in formula (2), and the minimum value of the residue solvent - accessible surface area is 0; the above data is a total of 13 dimensions.

[0091] y1 = sin(θ), y2 = cos(θ)(1)

[0092]

[0093] The sequence features are extracted using the following method;

[0094] Given that there are too few non - zero data in the above one - hot orthogonal vector encoding:

[0095] First, use formula (3) to map the sparse coding to dense coding by the auto - encoder; take the parameter W in the encoder as the corresponding new coding vector, with a data dimension of 20 dimensions;

[0096]

[0097] In the representation of amino - acid antigen - pair (AAP) features, the frequency of adjacent two amino acids appearing in pairs in the epitope sequence and the frequency of the occurrence of this amino - acid pair in the non - epitope sequence The ratio is calculated as shown in formula (4), and is normalized to [-1, 1] according to the formula The amino - acid antigen - pair features are represented by four values: the AAP feature value, the maximum value in AAP, the minimum value in AAP, and the AAP average value, with a total of 4 dimensions

[0098]

[0099] where and The calculation methods are as shown in formulas (5) and (6).

[0100]

[0101] where, is the number of times a pair of amino acids appears in the epitope sequence, is the number of all amino - acid pairs in the epitope sequence.

[0102]

[0103] Among them, is the number of occurrences of a pair of amino acids in the non-epitope sequence, is the number of all amino acid pairs in the non-epitope sequence.

[0104] Amino Acid Triplets (AAT) feature represents the ratio of the occurrence frequency of consecutive three amino acid residues in the epitope sequence in the dataset to the occurrence frequency of this triplet in the non-epitope sequence. The calculation method and normalization method are the same as those of the AAP feature; the AAT feature represents 4 groups of values including the AAT feature value, maximum value, minimum value, and average value of the residue, with a total of 4 dimensions;

[0105] Amino Acid Composition (AAC) feature represents the relative proportion of each amino acid in each protein sequence in the dataset, and the calculation method is as shown in formula (7).

[0106]

[0107] where R i is the number of amino acids of type i in the sequence, N is the sequence length, and the AAC feature is 1-dimensional;

[0108] Considering the convenience of batch input of data during the training of the deep learning model, all sequence lengths are aligned to 71 according to the longest sequence in all test sets, and zeros are padded for sequences shorter than 71.

[0109] The residue feature ProtT5 is 71 * 1024, ESM-2 is 71 * 1280, DSSP is 71 * 13, and the residue encoding, AAT, AAP, and AAC are concatenated together to be 71 * 29;

[0110] Specifically, the feature processing module includes: module (a) structure feature processing module, module (b) ESM-2 large model feature processing module, module (c) ProtT5 large model feature processing module, and module (d) sequence feature processing module.

[0111] Specifically, in the structure feature processing module: the three-dimensional structure predicted by ESMFold is parsed into 13-dimensional structure features using DSSP software; for the 71 * 13 feature matrix of a single sequence, feature extraction is performed using a full-size two-dimensional convolutional neural network, and the two-dimensional matrix is converted into a one-dimensional vector; the convolutional kernel size is (71, 13), the activation function is ReLU, and the output is 256 dimensions; the calculation process is as shown in formula (8).

[0112] F1 = ReLU(Conv(W * x 1:71+b)) (8)

[0113] Specifically, the residue representation features output by the ESM-2 large model are 1280-dimensional and are input into a two-layer bidirectional LSTM network. Then, a forward attention network is used to map the two dimensions to a one-dimensional vector;

[0114] First, the bidirectional LSTM network is used to focus on the global information of the protein and capture the long-range dependencies of the protein sequence; the formula description of the unidirectional LSTM model is shown in (9).

[0115]

[0116] Among them, σ is the activation function, usually the Sigmoid function; ⊙ represents element-wise multiplication of matrices; x t is the network input at time t; i t 、f t 、o t 、c t and h t respectively represent the input gate, forget gate, output gate, internal memory unit and output at time t; h t-1 is the output of the previous time; c t-1 is the output of the internal memory unit of the previous time; the rest are learnable parameters of the neural network.

[0117] The first layer of the LSTM network takes 1280 dimensions as input and outputs 128 dimensions unidirectionally. When the forward LSTM and backward LSTM data converge, a merge operation is performed at the lowest dimension, as shown in formula (10), and the output is 256 dimensions;

[0118] The second layer of the LSTM network takes 256 as input and outputs 128 dimensions unidirectionally. When the forward and backward converge, a merge operation is also performed at the lowest dimension, and the output is 256 dimensions;

[0119] The forward attention network is mainly used to convert a two-dimensional matrix into a one-dimensional vector; h t ′ is the output of the bidirectional LSTM, representing the t-th amino acid residue in the current sequence; through a forward neural network as shown in formula (11), the attention weight of the residue t in the current feature representation in the sequence is obtained through formula (12); for the feature representations of all residues in the sequence, weighted summation is performed, as shown in formula (13), to achieve the conversion from two-dimensional feature representation to one-dimensional. The input data dimension of the forward attention network is 71*256, and the output is 1*256 dimensions.

[0120] e t =h t ′W t (11)

[0121]

[0122] Specifically, in the ProtT5 representing the feature extraction module, it includes two layers of bidirectional LSTM networks and one layer of forward attention network. The module input is 1024 - dimensional, and the module output F3 is also 256 - dimensional, and it also maps two - dimensional features to one - dimensional vectors.

[0123] Specifically, in the sequence feature extraction module for the encoded features, AAP, AAT, and AAC features of the sequence, the full - size convolutional neural network is also used for feature extraction; the input of each sequence is of size 71 * 29, the convolutional kernel size is (71, 29), the activation function is ReLU, and the output feature F4 is 256 - dimensional.

[0124] It should be noted that the forward neural network module performs a data merging operation on the output one - dimensional vectors of the above four feature processing modules. The output dimension of each module is 256, and the merged dimension is 1024, as shown in formula (14);

[0125] F = concat(F1,F2,F3,F4)(14)

[0126] It includes two layers of forward neural networks, as shown in formulas (15) and (16) respectively. The activation function of the first forward neural network is tanh, and the output dimension is 512; the second uses the Sigmoid function as shown in formula (17), and the output dimension is 1, and the output result is 0 or 1, where: 0 represents non - epitope and 1 represents epitope;

[0127] F′ = tanh(W * F + b)(15)

[0128]

[0129] The error loss between the model prediction value and the true value is described by the binary cross - entropy loss function, as shown in formula (18);

[0130]

[0131] In summary, it can be seen that predicting BCE by a computational method can be described as a binary classification problem of judging whether a protein sequence is an epitope.

[0132] The above - mentioned implementation manners are only the preferred implementation manners of the present invention and cannot be used to limit the scope of protection of the present invention. Any non - substantial changes and substitutions made by those skilled in the art based on the present invention belong to the scope of protection required by the present invention.

Claims

1. A BCE prediction method based on a hybrid deep learning strategy, characterized in that: The steps include: Step 1: Build a BCE prediction model; Step 2: inputting the protein sequence into the BCE prediction model; Step 3: Use the prediction model to output whether the protein sequence is a BCE; The BCE prediction model in step 1 includes a database, a feature extraction module, a feature processing module, and a forward neural network module; in: The database is used to obtain training data sets; The feature extraction module is used to extract features from the data and obtain four sets of features; The feature processing module is used to process the extracted features; The forward neural network module is used to merge multiple processed features and output predicted values; The feature extraction module is used to extract ProtT5 expression features, ESM-2 expression features, DSSP features and sequence features, and the sequence features are spliced by sequence residue orthogonal coding, AAT, AAP, and AAC; in: AAP is an amino acid antigen pair, AAT is an amino acid trimer, and AAC is an amino acid composition; ProtT5 indicates that the features are extracted using the protein language model ProtT5; the specific model is ProtT5-XL-UniRef50, the input of which is the primary sequence of the protein, and the output is the mth residue in the corresponding sequence The feature representation vector , Dimension 1024; ESM-2 representation features are extracted using the protein language pre-trained large model ESM-2; the specific model is Esm2_t33_650M_UR50D, the input of this model is the primary sequence of the protein, and the output is the mth residue in the corresponding sequence The feature representation vector , Dimension 1280; DSSP features are extracted using the protein structure prediction model ESMFold; ESMFold is used to predict the three-dimensional structure of the protein, and the protein file is parsed using DSSP software to obtain the secondary structure information of the eight states of the protein. ) dihedral angle information, solvent accessible surface area information; the secondary structure is represented by 8-dimensional one-hot orthogonal encoding data; a dihedral angle is represented by a pair of sine and cosine functions, as shown in formula (1), ( ) The dihedral angle is represented by 4-dimensional data; the solvent accessible surface area is normalized by maximum-minimum, as shown in formula (2), and the minimum value of the residue solvent accessible surface area is 0; the above data has a total of 13 dimensions; （1） （2） Sequence features were extracted using the following method; Given that there are too few non-zero data in the above one-hot orthogonal vector encoding: First, use formula (3) to map the sparse coding of the autoencoder to dense coding; take the parameter W in the encoder as the corresponding new coding vector, and the data dimension is 20 dimensions; （3） AAP features represent the frequency of two adjacent amino acids appearing in pairs in the AAP epitope sequence in the dataset. The frequency of occurrence of the AAP in non-epitope sequences The ratio is calculated as formula (4), according to the formula , normalized to [-1,1]; amino acid antigen pair features are represented by four values: AAP characteristic value, maximum value in AAP, minimum value in AAP and average value of AAP, with a total of 4 dimensions; （4） in and The calculation method is as follows: formula (5) and (6); （5） in, is the number of times a pair of amino acids appears in the epitope sequence, is the number of all amino acid pairs in the epitope sequence; （6） in, is the number of occurrences of an amino acid pair in the non-epitope sequence, is the number of all amino acid pairs in the non-epitope sequence; The AAT feature represents the ratio of the frequency of three consecutive amino acid residues in the epitope sequence to the frequency of the trimer in the non-epitope sequence in the dataset. The calculation and normalization methods are the same as the AAP feature. The AAT feature represents four sets of values, including the AAT characteristic value, maximum value, minimum value and average value of the residue, with a total of four dimensions. AAC features represent the relative proportion of each amino acid in each protein sequence in the dataset and are calculated as shown in formula (7); （7） in In this sequence The number of amino acids of the type, is the sequence length, AAC features are 1-dimensional; Considering the convenience of data batch input during deep learning model training, all sequences are aligned to 71 lengths based on the longest sequence in all test sets, and zeros are added to sequences shorter than 71. The residue feature ProtT5 is 71*1024, ESM-2 is 71*1280, DSSP is 71*13, and the residue code, AAT, AAP, and AAC are spliced together to form a total of 71*29.

2. The BCE prediction method based on a hybrid deep learning strategy according to claim 1, characterized in that: The training data in the database are from the IEDB database and the Bcipep database; the verification data in the database are also selected from the IEDB database and the Bcipep database.

3. The BCE prediction method based on a hybrid deep learning strategy according to claim 1, characterized in that: The feature processing module includes: a structural feature processing module, an ESM-2 large model feature processing module, a ProtT5 large model feature processing module, and a sequence feature processing module.

4. The BCE prediction method based on a hybrid deep learning strategy according to claim 3, characterized in that: In the structural feature processing module: the three-dimensional structure predicted by ESMFold is analyzed using DSSP software to obtain 13-dimensional structural features; A single sequence 71*13 feature matrix is extracted using a full-size two-dimensional convolutional neural network to convert the two-dimensional matrix into a one-dimensional vector; the convolution kernel size is (71, 13), the activation function is ReLU, and the output is 256-dimensional; the calculation process is as shown in formula (8); （8）。 5. The BCE prediction method based on a hybrid deep learning strategy according to claim 3, characterized in that: The residue representation feature output by the ESM-2 large model is 1280-dimensional, which is input into a two-layer bidirectional LSTM network, and then a one-layer forward attention network is used to map the two dimensions to a one-dimensional vector; First, a bidirectional LSTM network is used to focus on the global information of proteins and capture the long-range dependencies of protein sequences; the unidirectional LSTM model is formulated as shown in (9); （9） （10） Among them, σ is the activation function using Sigmoid function; ⊙ represents the matrix bitwise multiplication; x t is the network input at time t; i t 、f t 、o t 、c t and h t Represent the input gate, forget gate, output gate, internal memory unit and output at time t respectively; is the output of the previous moment; is the output of the internal memory unit at the previous moment; the rest are the learnable parameters of the neural network; The first layer of LSTM network has an input of 1280 dimensions and a unidirectional output of 128 dimensions. When the forward LSTM and backward LSTM data converge, a merge operation is performed at the lowest dimension, as shown in formula (10), and the output is 256 dimensions. The second-layer LSTM network has a 256-dimensional input and a 128-dimensional output in one direction. When converging forward and backward, it also performs a merge operation at the lowest dimension and outputs a 256-dimensional output. The forward attention network is mainly used to convert a two-dimensional matrix into a one-dimensional vector; is the output of the bidirectional LSTM, representing the g-th amino acid residue in the current sequence; through a layer of forward neural network as shown in formula (11), the attention weight of the current feature representation of residue g in the sequence is obtained through formula (12); the feature representation of all residues in the sequence is weighted and summed, as shown in formula (13), to achieve the conversion of two-dimensional feature representation to one-dimensional. The input data dimension of the forward attention network is 71*256, and the output dimension is 1*256. （11） （12） （13）。 6. The BCE prediction method based on a hybrid deep learning strategy according to claim 3, characterized in that: The ProtT5 large model feature processing module includes two layers of bidirectional LSTM networks and one layer of forward attention network. The module input is 1024 dimensions, and the module output F3 is also 256 dimensions, and the two-dimensional features are also mapped to one-dimensional vectors.

7. The BCE prediction method based on a hybrid deep learning strategy according to claim 3, characterized in that: In the sequence feature processing module, a full-size convolutional neural network is used to extract the encoding features, AAP, AAT, and AAC features of the sequence. The input of each sequence is 71*29 in size, the convolution kernel size is (71, 29), the activation function is ReLU, and the output feature F4 is 256-dimensional.

8. The BCE prediction method based on a hybrid deep learning strategy according to claim 1, characterized in that: The feedforward neural network module performs a data merging operation on the one-dimensional vectors output by the four feature processing modules. The output dimension of each module is 256, and the merged dimension is 1024, as shown in formula (14); （14） It includes two layers of feedforward neural networks, as shown in formulas (15) and (16), respectively. The first feedforward neural network uses tanh as activation function, and the output dimension is 512; the second one uses Sigmoid function as shown in formula (17), the output dimension is 1, and the output result is 0 or 1, where: 0 represents non-epitope and 1 represents antigen epitope; （15） （16） （17） The error loss between the model prediction value and the true value is described by the binary cross entropy loss function, as shown in formula (18); （18）。

Images