An anti-virus peptide prediction method and system based on federated learning and neural network

Abstract

The application discloses an antiviral peptide prediction method and system based on federated learning and a neural network, utilizes multiple independent data sets to respectively perform training on a prediction model, constructs a double-channel deep neural network through the model, extracts features of different dimensions from original variable-length sequence data, analyzes peptide segment sequences from two aspects of sequence data and evolution levels, and finally transmits model gradients and parameters obtained to an intermediate server through encryption technology; the server aggregates the transmitted model parameters to form an intermediate model without contacting the data itself; the server transmits parameters of the intermediate model to participants, completes updating of participant models, and the above steps are circularly performed until model training is completed. The combination of federated learning and the neural network can accurately and quickly predict the antiviral activity of peptide segments under the premise of guaranteeing data security.

Description

Technical Field

[0001] This invention belongs to the field of biopeptide recognition technology, specifically relating to an antiviral peptide prediction method and system based on federated learning and neural networks. Background Technology

[0002] In recent years, a large number of antiviral peptide prediction models based on machine learning and deep learning have been developed. Machine learning-based models can achieve good prediction results on relatively small datasets.

[0003] With the rapid growth of sequencing data, machine learning is somewhat lacking in prediction accuracy and generalization ability. Deep learning-based models can achieve accurate prediction results when faced with massive amounts of data, but these models generally need to be trained on multiple independent datasets. When using the data, data leakage is inevitable and it may be maliciously stolen and used. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide an antiviral peptide prediction method and system based on federated learning and neural networks, which can minimize data leakage and make accurate and fast predictions, in order to solve the problem of data silos caused by the inability to share data due to data privacy, and the technical problem of repetitive workload caused by separating feature extraction and model learning in traditional methods.

[0005] The present invention adopts the following technical solution:

[0006] An antiviral peptide prediction method based on federated learning and neural networks includes the following steps:

[0007] S1. The client uses peptide sequence data to train the prediction model, and then encrypts the gradient and parameters of the prediction model.

[0008] S2. The intermediate server aggregates the gradients and parameters of the prediction model encrypted in step S1 to form an intermediate model. The intermediate model then transmits the gradients and parameters to the client for updating.

[0009] S3. Repeat steps S1 and S2, evaluate the intermediate model using the binary cross-entropy loss function, obtain the shared model and pass it to the client; input the peptide sequence to be predicted into the client, perform one-hot encoding after preprocessing, input the one-hot encoding into the prediction model, and output the prediction result to complete the antiviral peptide prediction.

[0010] Specifically, peptide sequences of different lengths are encoded according to the one-hot encoding rules to obtain one-hot codes; the one-hot codes are fed into a dual-channel deep neural network containing a first neural network and a second neural network respectively, the corresponding features obtained are fused, and the fused features are used as input to construct two fully connected layers and a Softmax layer for identifying antiviral peptides.

[0011] Furthermore, before encoding, redundant sequences and sequences with a similarity greater than 90% in the peptide sequence are deleted.

[0012] Furthermore, the peptide sequence has a length of L, which is encoded into a binary matrix of dimension L*20, where 20 represents the types of amino acids.

[0013] Furthermore, the first neural network is a variant of a recurrent neural network, combining the outputs of two recurrent neural networks, one processing the sequence from right to left and the other from left to right. Both recurrent neural networks contain multiple LSTM units, which are used to remember the values ​​of sequences of arbitrary length. The length of the input sequence is used as the time step to obtain the output of the final time step in both directions. The number of LSTM units is dynamically adjusted by the length of the input sequence. Each LSTM unit includes an input gate, a forget gate, and an output gate, and selects the hidden state vector h. t As output, 128-dimensional feature data is obtained, and a dropout layer is added after the output with keep-prob set to 0.8.

[0014] Furthermore, the second neural network includes a feature extraction layer, an average block layer, and a convolutional layer;

[0015] The feature extraction layer initializes 20 20*1 convolutional kernels using a 20*20 Blosum matrix, and constructs a position-specific scoring matrix by performing convolution operations on the input variable-length one-hot code.

[0016] The average block layer is used to divide the position-specific scoring matrix into average blocks, converting variable evolutionary data into fixed-length evolutionary features. By setting width=4, the 20*L position-specific scoring matrix is ​​divided into four parts, and the average value of each part is taken according to the algorithm, finally resulting in a 20*4 average block matrix.

[0017] The convolutional layer is used to perform convolution operations on the obtained blocks to obtain a 100-dimensional output, and a dropout layer is added after the output with keep-prob set to 0.8.

[0018] Furthermore, the convolution operation is as follows:

[0019]

[0020] Where X is the input data, i is the position index, k is the kernel index, and W... k is the convolution kernel, M is the window size, N is the number of input channels, and m and n are the lower bounds of the accumulator.

[0021] Furthermore, the fully connected layer specifically refers to:

[0022] fulCN(X)=ReLU(WX+b)

[0023] Where X is the input data, W is the M×N weight matrix, b is the n-dimensional bias vector, fulCN is the fully connected function, and ReLU is the activation function;

[0024] The Softmax layer is specifically as follows:

[0025]

[0026] Where V is a given logical vector, i and k are position indices, and K = 2.

[0027] Furthermore, Paillier encryption specifically involves:

[0028] Each participating research institution encrypts the parameters of the prediction model using the public key published by the server and transmits the model parameters to the intermediate server. The intermediate server decrypts the received encrypted parameters using its private key, and the model updates the corresponding input parameters. Two large prime numbers p and q are selected, where p ≈ q. N = p × q is calculated, and λ = lcm(p-1, q-1), where lcm() is the least common multiple function or λ = φ(N), where φ() is Euler's totient function. A random number g is selected. The public key PK is (N, g), and the private key SK is (φ(N), μ), where μ = φ(N). -1 modN; B randomly selects R, C = g M ×R N modN 2 Encryption is performed using A's public key; then A obtains B's C to complete decryption, M = L(C) λ modN 2 )×μmodN, C is the encrypted data, R is a random number that satisfies gcd(r,n)=1, and M represents the original text.

[0029] Secondly, embodiments of the present invention provide an antiviral peptide prediction system based on federated learning and neural networks, comprising:

[0030] The encryption module is used by the client to train the prediction model using peptide sequence data and then encrypt the gradient and parameters of the prediction model.

[0031] The aggregation module is used by the intermediate server to aggregate the gradients and parameters of the prediction model encrypted by the encryption module to form an intermediate model. The intermediate model then passes the gradients and parameters to the client for updates.

[0032] The prediction module is used to repeatedly execute the encryption and aggregation modules. It evaluates the intermediate model using the binary cross-entropy loss function, obtains a shared model, and passes it to the client. The peptide sequence to be predicted is input into the client, preprocessed, and then one-hot encoded. The one-hot encoded sequence is then input into the prediction model, and the prediction result is output to complete the antiviral peptide prediction.

[0033] Compared with the prior art, the present invention has at least the following beneficial effects:

[0034] This invention discloses an antiviral peptide prediction method based on federated learning and neural networks. It employs a dual-channel deep neural network to predict variable-length peptide sequences. When processing sequence data, it eliminates the need for separate feature extraction; instead, it uses two channels to analyze peptide sequences at both the sequence and evolutionary levels, embedding feature extraction into the neural network. This allows for dynamic optimization of the feature extraction method during model training. This method is simpler and more efficient than traditional methods that require separating feature extraction and model learning. Furthermore, it utilizes federated learning to connect data silos, improving prediction accuracy by fully leveraging fragmented data while ensuring data privacy and security. This solves the technical problem of data silos arising from the inability to share data due to privacy concerns in deep learning applications of biopeptide identification technology.

[0035] Furthermore, peptide sequences of different lengths are encoded according to the one-hot encoding rule, which solves the problem of classifiers' inability to handle data attributes and also plays a role in expanding features to a certain extent. The one-hot codes are then introduced into the first neural network; the bidirectional recurrent neural network can capture long-term dependencies and make better use of sequence data. The second neural network analyzes local evolutionary information by constructing a dynamic neural network. The outputs of the two neural networks are fused, and then applied to the prediction of antiviral peptides through two fully connected layers and a softmax layer. This method is far superior to existing methods for predicting antiviral peptides.

[0036] Furthermore, CD-HIT is a tool for clustering protein or nucleic acid sequences. It clusters sequences based on their similarity to remove redundant sequences, and is generally used to construct non-redundant datasets for subsequent experimental analysis. Here, CD-HIT is used to remove redundant sequences and sequences with a similarity greater than 90% from the dataset, which is beneficial for data preprocessing and subsequent model training.

[0037] Furthermore, the peptide sequence is converted into a binary matrix. By converting the data format, it becomes easier to input as training data into the neural network for training.

[0038] Furthermore, the bidirectional recurrent neural network used in the first neural network can capture long-term dependencies, making better use of sequence data and facilitating the identification of antiviral peptides.

[0039] Furthermore, the second neural network initializes the convolution kernels using the Blosum matrix, extracts features using the convolution kernels, and analyzes the evolutionary information of peptide sequences by constructing a dynamic neural network, which is beneficial for the identification of antiviral peptides.

[0040] Furthermore, the features output by the two neural networks are fused, and then applied to the prediction of antiviral peptides through two fully connected layers and a Softmax layer. This method is far superior to existing methods for predicting antiviral peptides.

[0041] Furthermore, using the Paillier encryption method, the data itself is not transmitted in plaintext, thus making it virtually impossible to leak the original data, guaranteeing data privacy. Federated learning using the Paillier encryption algorithm can ensure that data participants can train neural network models without exchanging data, achieving the goal of data privacy protection.

[0042] It is understandable that the beneficial effects of the second aspect mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0043] In summary, this invention can analyze data from both sequence and evolutionary levels, embedding feature extraction into neural networks without requiring separate feature extraction, thus greatly reducing training costs and enabling rapid and accurate identification of antiviral peptides. On the other hand, it utilizes federated learning to connect data silos, making full use of scattered data to improve prediction accuracy while ensuring data privacy and security.

[0044] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0046] Figure 2 This is a flowchart of the algorithm model structure for training in this invention;

[0047] Figure 3 This is a model diagram of the first neural network of the present invention;

[0048] Figure 4 This is a model diagram of the second neural network of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] In the description of this invention, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0051] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0052] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" relationship.

[0053] It should be understood that although terms such as first, second, third, etc., may be used in the embodiments of the present invention to describe the preset range, these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from one another. For example, without departing from the scope of the embodiments of the present invention, the first preset range may also be referred to as the second preset range, and similarly, the second preset range may also be referred to as the first preset range.

[0054] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0055] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0056] Federated learning is a privacy-preserving distributed machine learning approach. Its core idea is to build a machine learning model based on datasets distributed across multiple devices, effectively preventing data leakage. Therefore, this invention combines federated learning with deep learning, distributing each independent dataset across different servers. Each server independently trains its neural network model without direct data exchange; only encrypted aggregation of model parameters is performed. This effectively protects private data from leakage without affecting prediction accuracy.

[0057] This invention provides an antiviral peptide prediction method based on federated learning and neural networks. It trains prediction models using multiple independent datasets. The models construct dual-channel deep neural networks to extract features of different dimensions from the original variable-length sequence data and analyze peptide sequences from both sequence data and evolutionary level perspectives. Finally, the obtained model gradients and parameters are transmitted to an intermediate server using encryption technology. The server aggregates the transmitted model parameters to form an intermediate model without accessing the data itself. The server then transmits the parameters of the intermediate model to participating parties to update their models. This process is repeated until a high-performance shared model is obtained. This shared model is then used to predict the peptide sequences to be predicted. By combining federated learning and neural networks, this method can accurately and quickly predict the antiviral activity of peptides while ensuring data security, minimizing data leakage, and providing accurate and rapid predictions.

[0058] Please see Figure 1 This invention discloses an antiviral peptide prediction method based on federated learning and neural networks, comprising the following steps:

[0059] S1. The training of the model is distributed across multiple clients, each client having a portion of peptide sequence data. The client uses its own dataset to train the prediction model and transmits the gradient and parameters of the prediction model to the intermediate server through encryption technology.

[0060] S101. Preprocess the dataset by using CD-HIT software to remove redundant sequences and sequences with a similarity greater than 90%.

[0061] S102. Encode the peptide sequence obtained in step S101 using a one-hot encoding.

[0062] Peptide sequences of varying lengths are directly encoded using the one-hot encoding rules to obtain one-hot codes. The length of the peptide sequence to be predicted is defined as L. After encoding, a binary matrix of dimension L*20 is obtained, where 20 represents the types of amino acids.

[0063] S103. Import the data obtained in step S102 into a dual-channel deep neural network for prediction, output the prediction results and save the model parameters.

[0064] By combining federated learning with neural networks, and constructing a dual-channel deep neural network, the system analyzes variable-length sequences from both sequential and evolutionary perspectives to achieve identification and prediction. Federated learning addresses the data privacy security issues faced when using artificial intelligence, and because it employs a multi-segment parameter transmission method, it can handle larger amounts of data, allowing for more parameter calculations without compromising algorithm performance.

[0065] Please see Figure 2 The prediction process is as follows:

[0066] S1031. The one-hot code obtained in step S102 is imported into the first neural network and the second neural network to obtain the corresponding feature outputs respectively. Overfitting of the model can be prevented by setting a dropout layer.

[0067] Please see Figure 3 The first neural network is a variant of the recurrent neural network (RNN), combining the outputs of two RNNs: one processes the sequence from right to left, and the other from left to right. Both RNNs contain many LSTM units, which can remember the values ​​of sequences of arbitrary length. Using the length of the input sequence as the time step, they produce outputs for the final time steps in both directions. The number of LSTM units is dynamically adjusted by the length of the input sequence. Each LSTM unit includes an input gate, a forget gate, and an output gate, as shown in the following formula:

[0068] f t =σ(W f x t +U f b t-1 +b f )

[0069] i t =σ(W i x t +U i b t-1 +b i )

[0070] o t=σ(W o x t +U o b t-1 +b o )

[0071]

[0072]

[0073] Where, x t It is the input vector, f t It is the activation vector of the forget gate, i t It is the activation vector of the input gate, o t It is the activation vector of the output gate, C t It is the unit state vector, h t is the hidden state vector, W and U are parameter matrices, and b is the bias vector.

[0074] In the first neural network, h is selected. t The output yields 128-dimensional feature data. Furthermore, a dropout layer is added after the output, with keep-prob set to 0.8 to prevent overfitting.

[0075] Please see Figure 4 The second neural network is a dynamic convolutional neural network used to process score matrices at specific locations; it is a variant of convolutional neural networks (CNNs). It includes feature extraction layers, average block layers, and convolutional layers.

[0076] In the feature extraction layer, 20 20*1 convolutional kernels are initialized using a 20*20 Blosum matrix. The position-specific scoring matrix PSSM is obtained by performing a convolution operation on the input variable-length one-hot code. The convolution operation formula is as follows:

[0077]

[0078] Where M is the input data, i is the position index, and k is the kernel index. Each convolutional kernel W k It is an M×N weighted matrix, where M is the window size and N is the number of input channels, where M=1 and N=20.

[0079] In the average block layer, the position-specific scoring matrix (PSSM) obtained from the feature extraction layer is divided into average blocks, transforming variable evolutionary data into fixed-length evolutionary features. The PSSM matrix is ​​a commonly used method to reflect the evolutionary information of biological sequences. For a sequence of length L, the PSSM can be represented by a 20*L matrix.

[0080] The blocks are dynamically generated based on the input length, and the average value of each block is calculated. The specific pseudocode implementation is as follows:

[0081] function average_block(list,width)

[0082] while t i ∈list do

[0083] block = length(t) i ) / width

[0084] for s = 1:width do

[0085] nt i (s)=avg{t i [block×(s-1)+1,block×s]}

[0086] end for

[0087] new_list.add(nt i )

[0088] end while

[0089] reture new_list

[0090] end function.

[0091] By setting width=4, the 20*L PSSM is divided into four parts, and the average value of each part is taken according to the algorithm, finally resulting in a 20*4 average block matrix.

[0092] In the convolutional layer, the blocks obtained above are subjected to convolution operations, as shown in the following formula:

[0093]

[0094] Where X is the input data, i is the position index, and k is the kernel index. Each convolution kernel W k It is an M×N weighted matrix, where M is the window size, N is the number of input channels, m and n are the lower bounds of the accumulator, and step is the convolution stride. Where M = 4, N = 4, and step = 4;

[0095] ReLU is an activation function that helps overcome the vanishing gradient problem and speeds up training. Its formula is as follows:

[0096]

[0097] A 100-dimensional output is obtained in the second neural network.

[0098] Similarly, a dropout layer was added after the output, with keep-prob set to 0.8, to finally obtain the corresponding feature data.

[0099] S1032. The features obtained in step S1031 are fused to facilitate subsequent calculations;

[0100] S1033. Using the fused features obtained in step S1032 as input, construct two fully connected layers and one Softmax layer to identify antiviral peptides.

[0101] The fully connected layer is constructed based on a nonlinear combination of the sequence and evaluation features, and its formula is expressed as follows:

[0102] fulCN(X)=ReLU(WX+b)

[0103] Where X is the input data, W is the M×N weight matrix, and b is the n-dimensional bias vector.

[0104] The Softmax layer normalizes a numerical vector into a probability distribution vector, where the sum of all probabilities is 1. The Softmax function is typically used as the last layer in a neural network, taking input values ​​from the previous layer and converting them into probabilities, as expressed in the following formula:

[0105]

[0106] Where V represents a given logical vector, i and k represent position indices, and K = 2.

[0107] This invention addresses security concerns regarding parameters like weights in federated learning during forward and backward propagation. Considering both data security and the specific application scenario, it employs the Paillier homomorphic encryption algorithm to encrypt these parameters. Furthermore, it allows for efficient manipulation of the ciphertext, including addition and multiplication, without requiring a decryption key. During this process, data from participating parties is not transmitted to the server, preventing data leakage.

[0108] Paillier homomorphic encryption is used to encrypt parameters such as weights applied in federated learning. Paillier encryption is a public-key encryption algorithm, and its specific usage is as follows:

[0109] Each participating research institution uses the public key published by the server to encrypt the parameters of the prediction model and then transmits the model parameters to the intermediate server.

[0110] The server uses its private key to decrypt the received encrypted parameters, and the model updates the corresponding input parameters.

[0111] Choose two large prime numbers p and q, where p ≈ q;

[0112] Calculate N = p × q, and λ = lcm(p-1, q-1), where lcm() is the least common multiple function or λ = φ(N), where φ() is the Euler totient function;

[0113] Select a random number g;

[0114] The public key PK is (N, g), and the private key SK is (φ(N), μ), where μ = φ(N). -1 modN;

[0115] Encryption: B randomly selects R, C = g M ×R N modN 2 Using A's public key;

[0116] Deciphering: A obtains B's C, M = L(C) λ modN 2 )×μmodN, where

[0117] Where C is the encrypted data, R is a random number that satisfies gcd(r,n)=1, and M represents the original text.

[0118] S2. The encrypted model parameters obtained in step S1 are aggregated to form an intermediate model. Then, the parameters of the intermediate model are passed to each client to complete the client model update.

[0119] The method for model parameter aggregation is as follows:

[0120] First, the central server initializes the model parameters. Then, it performs several rounds of parameter aggregation, selecting at least one and at most k clients to participate in training in each round. In round t, the server will process the model W for that round. t The data is then sent to the client, and each selected client simultaneously processes the data locally based on model W. t Train your own model using private data After training is complete, the model is uploaded to the server. The server aggregates the models collected from each client using a weighted average based on the number of samples from each party, resulting in the next model W. t+1 The formula is as follows:

[0121]

[0122] Where, n k Let be the number of samples on client k, and n be the total number of samples from all selected clients.

[0123] The computational load on the client side is controlled by three key parameters:

[0124] C: refers to the proportion of clients participating in the calculation in each round;

[0125] S: refers to the number of training steps performed by each client on the local dataset in each round;

[0126] B: Refers to the batch size used when updating the client. When B = ∞, it means that the entire local dataset is treated as a single batch.

[0127] With C=1, S=1 and B=∞, each client puts all its local data into training at once in each round, and uses full-batch gradient descent on all the data owned by all clients.

[0128] S3. Repeat steps S1-S2. The intermediate server evaluates the model's performance using a binary cross-entropy loss function, ultimately obtaining a high-performance shared model. Each client receives the same shared model, and clients do not communicate or depend on each other; each client can predict independently during prediction. The client inputs the peptide sequence to be predicted, which is preprocessed and then one-hot encoded. The one-hot encoding is then input into the prediction model for prediction, and the prediction result is output.

[0129] In another embodiment of the present invention, an antiviral peptide prediction system based on federated learning and neural networks is provided. This system can be used to implement the above-mentioned antiviral peptide prediction method based on federated learning and neural networks. Specifically, the antiviral peptide prediction system based on federated learning and neural networks includes an encryption module, an aggregation module, and a prediction module.

[0130] The encryption module is used by the client to train the prediction model using peptide sequence data and encrypt the gradient and parameters of the prediction model.

[0131] The aggregation module is used by the intermediate server to aggregate the gradients and parameters of the prediction model encrypted by the encryption module to form an intermediate model. The intermediate model then passes the gradients and parameters to the client for updates.

[0132] The prediction module executes the encryption and aggregation modules. It evaluates the intermediate model using the binary cross-entropy loss function, obtains the shared model, and passes it to the client. The peptide sequence to be predicted is input into the client, preprocessed, and then one-hot encoded. The one-hot encoded sequence is then input into the prediction model, and the prediction result is output to complete the antiviral peptide prediction.

[0133] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0134] To further illustrate, this implementation uses the dataset studied by Nishant Thakur et al. [1]([1]Thakur N,Qureshi A,Kumar M.AVPpred:collection and prediction of highly effective antiviral peptides.Nucleic Acids Res.2012;40(Web Server issue):W199-W204.doi:10.1093 / nar / gks450). The dataset we use contains 604 antiviral peptides and 1056 non-antiviral peptides, of which 544 antiviral peptides and 951 non-antiviral peptides are divided into the training set, and the remaining 60 antiviral peptides and 105 non-antiviral peptides are divided into the test set.

[0135] Step 1: Randomly shuffle the 1495 peptides in the training set and divide them into 5 equal parts, each containing 299 samples, which will serve as training data for five clients. Each client first filters the input peptide sequences and encodes them using one-hot encoding before training. During the training phase, cross-validation is used to find the optimal hyperparameter values. The batch size is set to 128, the training epoch is 20, and the Adam optimizer and binary cross-entropy are used. After training, each client encrypts and uploads the model parameters to the intermediate server.

[0136] Step 2: The intermediate server aggregates the encrypted model parameters obtained in Step 1 to form an intermediate model, and then transmits the parameters of the intermediate model to each client to complete the client model update.

[0137] Step 3: Repeat steps 1-2 until a high-performance shared model is obtained, and each client will receive the same shared model. Select one client and input the test set. The performance parameters of the shared model on the test set are shown in Table 1.

[0138] ACC (Accuracy) SN (Sensitivity) SP (Specificity) MCC value AUC value 0.947 0.945 0.950 0.894 0.916

[0139] As shown in Table 1, by combining federated learning and neural networks in this embodiment, the accuracy of the final result is greatly improved. It can effectively utilize fragmented data while protecting data privacy, and can dynamically adjust the feature extraction method when training the learning model. This provides a fast and accurate method for the identification of antiviral peptides, which is convenient for modern medical research and treatment.

[0140] In summary, this invention presents an antiviral peptide prediction method and system based on federated learning and neural networks. By combining federated learning and neural networks, it can accurately and quickly identify antiviral peptides while ensuring data security. Federated learning addresses the data privacy security issues faced when using artificial intelligence, and the use of multi-segment parameter transmission allows for larger data volumes and more parameter calculations without compromising algorithm performance. A dual-channel deep neural network is constructed, comprising a first neural network and a second neural network, to extract features of different dimensions from the original variable-length sequence data. In the first neural network, a bidirectional recurrent neural network (BLSTM) is used to extract sequence features from one-hot encoding. BLSTM can capture the dependencies between sequence data, thus effectively extracting data. In the second neural network, a dynamic convolutional neural network is used to extract evolutionary features. Finally, two fully connected layers are constructed to connect the first and second neural networks, ultimately completing the prediction and identification of antiviral peptides. Compared to other models, this model can process sequence data without requiring feature extraction methods and can analyze peptide sequences from both sequence data and evolutionary level perspectives, significantly improving the speed and accuracy of prediction.

[0141] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0142] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0143] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0144] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0145] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. An anti-virus peptide prediction method based on federated learning and neural network, characterized in that, The method comprises the following steps: The client trains the prediction model using peptide sequence data, encrypts the gradient and parameters of the prediction model, encodes different lengths of peptide sequences according to the one-hot coding rule to obtain one-hot codes, and performs the following steps: The one-hot code is sent into a double-channel deep neural network comprising a first neural network and a second neural network respectively, the obtained corresponding features are fused, the fused features are taken as inputs, two fully connected layers and a Softmax layer are constructed for identifying the antiviral peptides, the first neural network is a variant of a recurrent neural network, the outputs of two recurrent neural networks are combined, one of which processes the sequence from right to left, and the other processes the sequence from left to right; both of the two recurrent neural networks comprise a plurality of LSTM units, the LSTM units are used to remember the values of the sequence data of any length, the length of the input sequence is taken as a time step, and the outputs of the final time steps in two directions are obtained; the number of LSTM units is dynamically adjusted according to the length of the input sequence, each LSTM unit comprises an input gate, a forgetting gate and an output gate, and a hidden state vector As the output, 128-dimensional feature data is obtained, and a dropout layer is added after the output, with keep-prob set to 0.8; the second neural network comprises a feature extraction layer, an average block layer and a convolutional layer; The feature extraction layer initializes 20 20*1 convolution kernels using a 20*20 Blosum matrix, and constructs a position-specific scoring matrix by performing convolution operation on the input variable-length one-hot code; The average block layer is used to average block the position-specific scoring matrix, convert the variable evolutionary data into fixed-length evolutionary features, set width=4, divide the 20*L position-specific scoring matrix into four parts, take the average value of each part according to the algorithm, and finally obtain a 20*4 average block matrix; The convolution layer is used to perform convolution operation on the obtained blocks to obtain a 100-dimensional output, and a dropout layer is added after the output, and the keep-prob is set to 0.8; The intermediate server aggregates the gradient and parameters of the encrypted prediction model to form an intermediate model, and the intermediate model transmits the gradient and parameters to the client for updating; Repeat the above steps, evaluate the intermediate model using the binary cross-entropy loss function, obtain a shared model, and transmit the shared model to the client; input the peptide sequence to be predicted into the client, perform one-hot encoding after preprocessing, input the one-hot code into the prediction model, and output the prediction result to complete the prediction of the antiviral peptide. 2.The anti-virus peptide prediction method based on federated learning and neural network according to claim 1, wherein, Before encoding, the redundant sequences in the peptide sequence and the sequences with a similarity greater than 90% are deleted. 3.The anti-virus peptide prediction method based on federated learning and neural network according to claim 1, wherein, The length of the peptide sequence is L, and the binary matrix with a dimension of L*20 is obtained after encoding, and 20 represents the type of amino acid. 4.The anti-virus peptide prediction method based on federated learning and neural network according to claim 1, wherein, The convolution operation is: wherein, is the input data, is the position index, is the index of the kernel, is the convolution kernel, is the window size, is the number of input channels, and is the lower bound of the cumulative symbol.

5. The federated learning and neural network-based antiviral peptide prediction method according to claim 1, wherein, The full connection layer is specifically: wherein, is an input data, is a weight matrix of, is an n-dimensional bias vector, is a fully connected function, is an activation function; The Softmax layer is specifically: wherein for a given logic vector, and is a position index, .

6. The federated learning and neural network-based antiviral peptide prediction method according to claim 1, wherein, The Paillier encryption is specifically: Each research institution participant respectively encrypts the parameters of the prediction model with the public key of the server, and respectively transmits the model parameters to the intermediate server; the intermediate server decrypts the received encrypted parameters with the private key, and updates the corresponding input parameters of the model; two large prime numbers , and are selected , and , is the least common multiple function or , is the Euler function; a random number is selected , the public key PK is , the private key SK is , , , and the public key of A is used for encryption; then A obtains the public key of B , completes decryption, , ; is the encrypted data, is a random number and satisfies , , and the original text is represented. 7.A system for predicting antiviral peptides based on federated learning and neural networks, characterized in that, It comprises: The encryption module is used for the client to train the prediction model using peptide sequence data, encrypt the gradient and parameters of the prediction model, encode different lengths of peptide sequences according to the one-hot coding rule to obtain one-hot codes; The one-hot code is respectively sent into a double-channel deep neural network comprising a first neural network and a second neural network, the obtained corresponding features are fused, the fused features are taken as inputs, two fully connected layers and a Softmax layer are constructed for identifying the antiviral peptide, the first neural network is a variant of a recurrent neural network, the outputs of two recurrent neural networks are combined, one of which processes a sequence from right to left, and the other processes a sequence from left to right; both of the two recurrent neural networks comprise a plurality of LSTM units, the LSTM units are used to remember values of sequence data of an arbitrary length, the length of the input sequence is taken as a time step, and the outputs of the final time steps in two directions are obtained; the number of the LSTM units is dynamically adjusted according to the length of the input sequence, each LSTM unit comprises an input gate, a forgetting gate and an output gate, and a hidden state vector is selected As an output, 128-dimensional feature data is obtained, and a dropout layer is added after the output, with keep-prob set to 0.8; the second neural network comprises a feature extraction layer, an average block layer and a convolutional layer; The feature extraction layer initializes 20 20*1 convolution kernels using a 20*20 Blosum matrix, and constructs a position-specific scoring matrix by performing convolution operation on the input variable-length one-hot code; The average block layer is used to average block the position-specific scoring matrix, convert the variable evolutionary data into fixed-length evolutionary features, set width=4, divide the 20*L position-specific scoring matrix into four parts, take the average value of each part according to the algorithm, and finally obtain a 20*4 average block matrix; The convolution layer is used to perform convolution operation on the obtained blocks to obtain a 100-dimensional output, and a dropout layer is added after the output, and the keep-prob is set to 0.8; The aggregation module is used for the intermediate server to aggregate the gradient and parameters of the encrypted prediction model of the encryption module to form an intermediate model, and the intermediate model transmits the gradient and parameters to the client for updating; The prediction module is used for repeatedly executing the encryption module and the aggregation module, evaluating the intermediate model by using a binary cross entropy loss function, obtaining a shared model and delivering the shared model to the client, inputting a to-be-predicted peptide sequence into the client, performing one-hot encoding after pre-processing, inputting the one-hot encoding into the prediction model, and outputting a prediction result to complete the prediction of the antiviral peptide.

Images