Phage and bacterium interaction prediction model based on deep learning

Abstract

The invention discloses a deep learning-based phage and bacterium interaction prediction model, and relates to a phage and bacterium interaction prediction model. The method aims at solving the problems that an existing model is poor in short sequence prediction effect, and unseen hosts cannot be predicted. The system comprises a data preprocessing module, a data form conversion module, a training module and a prediction module. The data preprocessing module is used for downloading a public data set, obtaining DNA sequences of the bacteriophage and the bacteria and interaction relationship labels of the bacteriophage and the bacteria, preprocessing the DNA sequences and dividing the DNA sequences into three subsets; the feature extraction module is used for extracting a unique embedded vector for each k-mer matrix by using a feature extraction model to serve as vectorization representation of data; the training module is used for updating parameters by using the extracted vector as an initial model so as to realize an optimal effect; and the prediction module is used for predicting candidate bacteria generating interaction for the to-be-predicted bacteriophage by using the trained model pair. The invention belongs to the technical field of metagenome learning.

Description

Technical Field

[0001] The present invention relates to a prediction model for interaction between bacteriophage and bacteria, and belongs to the technical field of metagenomic learning. Background Art

[0002] Bacteriophages are highly specific viruses that target bacteria, infecting only those expressing their receptors while being harmless to other bacteria. This high specificity makes phages potentially useful for bacterial treatment. This therapeutic antibacterial application of phages is known as "phage therapy," particularly in clinical and veterinary settings. In clinical research, phage therapy has the potential to significantly reduce side effects and drug resistance compared to antibiotics. However, precise identification of target bacteria is crucial for selecting the appropriate phage for treatment. Furthermore, the use of phages as biocontrol agents, for example in food sterilization and wastewater treatment to control plant pathogens, also relies on the interaction between phages and bacteria. Therefore, identifying the interactions between phages and bacteria has become an increasingly important task. However, most existing methods rely on complete DNA sequences of both phages and bacteria as initial data, performing poorly on short sequences obtained through metagenomic sequencing and failing to accurately predict the activity of phages whose hosts were not included in the training data.

[0003] The invention patent with publication number CN118866073A and application date of August 29, 2023 discloses a method for predicting phage-host interactions, a method and device for determining the target phage of the host. First, based on the whole genome sequence data of the phage-host pair, the key phage interaction proteins and the key host interaction proteins are determined; according to the protein characteristics and corresponding DNA characteristics of the key phage interaction proteins and the protein characteristics and corresponding DNA characteristics of the key host interaction proteins, the DNA protein feature characterization of the phage and the host is constructed; and the DNA protein feature characterization of the phage and the host is provided to the phage-host interaction prediction model for interaction prediction. Subsequently, based on the interaction prediction results, it is determined whether the phage is the target phage of the host. Using this phage-host interaction prediction method, the interaction between the phage host can be predicted quickly and efficiently, thereby quickly screening the applicable phage of the host based on the interaction prediction results, thereby timely expanding and updating the host's available phage library.

[0004] However, the above methods have poor prediction effects on short sequences and cannot predict unseen hosts. Summary of the Invention

[0005] In order to solve the problem that existing models have poor prediction effect on short sequences and cannot predict unseen hosts, the present invention proposes a phage-bacteria interaction prediction model based on deep learning.

[0006] The technical solution adopted by the present invention to solve the above problems is: the present invention includes a data preprocessing module, a data form conversion module, a training module and a prediction module; The data preprocessing module is used to download public datasets, obtain DNA sequences of phages and bacteria and their interaction relationship labels, perform preprocessing, and divide them into three subsets; The feature extraction module uses the feature extraction model to extract a unique embedding vector for each k-mer matrix as a vectorized representation of the data; The training module uses the extracted vectors to update the parameters of the initial model to achieve the best results; The prediction module uses the trained model to predict candidate bacteria that interact with the phage to be predicted.

[0007] Furthermore, data preprocessing refers to removing redundant sequences that appear repeatedly in the dataset and reshuffling the dataset, dividing it into the training set, validation set, and test set required for subsequent model training and testing in a ratio of 0.8:0.1:0.1.

[0008] Furthermore, obtaining the k-mer matrix means converting each DNA sequence of different lengths into a matrix data of uniform size. First, the k-mer frequency of each sequence is calculated, and the frequency count of each k-mer is filled into a matrix of size 1 according to the arrangement of the specified random seeds. in an all-zero matrix.

[0009] Furthermore, the k-mer frequency calculation method is: , Where, Indicates the The number of times k-mer appears in the entire sequence, Indicates the length of the sequence.

[0010] Furthermore, the feature extraction module is composed of a hybrid model consisting of CNN and BiMamba2. First, CNN convolves, pools, and flattens the matrix to obtain a linear map of the vector to the internal dimension of BiMamba2, performs bidirectional modeling, and outputs fused features that retain bidirectional information.

[0011] Furthermore, the training process of the training module is: Step 1: The initial model outputs the embedding vector and calculates the Euclidean distance between the phage vector and each bacterial vector; Step 2: Calculate the loss by contrastive loss and backpropagate it to the initial model to update the parameters; Step 3: Ultimately, the phages and bacteria that can interact are close to each other in the embedding space, while the non-interactive phages are far away from each other, forming a cluster with the bacteria as the center in the space.

[0012] Furthermore, the prediction process is to receive test data through the trained model and calculate the distance between the phage and the bacteria. If their distance is less than the preset value, it is considered that the phage can infect the bacteria, otherwise it is considered that they cannot interact.

[0013] The beneficial effects of the present invention are: the present invention achieves the highest effect compared with several existing methods on multiple data sets, the model structure is simple and easy to build, and it can make high-accuracy predictions for short sequence fragments and in the presence of unseen hosts. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a design flow chart of the present invention; Figure 2 It is a schematic diagram of the feature extraction model architecture; Figure 3 It is a flow chart of the training and testing process of the present invention. DETAILED DESCRIPTION

[0015] Specific implementation method 1: A prediction model of phage-bacteria interaction based on deep learning, such as Figure 1As shown in the figure, a dataset compiled by Ruohan W was used, redundant data was removed, and the dataset was randomly re-divided into three subsets for training, validation, and testing, respectively, at a ratio of 8:1:1. The dataset contained a total of 5711 bacteriophages and 117 bacteria. The complete bacteriophage sequences in the dataset were randomly cropped into short sequences of 1000bp and 5000bp to construct two short sequence datasets. The k-mer frequencies of all sequences were then calculated. After multiple rounds of testing, a random seed was determined, which was used to fill the frequency counts of each k-mer into an all-zero matrix in a specific order to construct the k-mer matrix. The matrix was then fed into a feature extraction model, initially processed by a convolutional network, and then input into the BiMamba2 network. This network can better capture the dependencies between individual k-mers in the sequence. Its bidirectional processing mechanism fully exploits the directional information in multidimensional data, and a dynamic padding adjustment strategy based on the input data dimension further improves computational efficiency and reduces memory consumption. Finally, the embedding vectors obtained by the feature extraction module learn their spatial characteristics, so that interacting phages and bacteria are close to each other in the embedding space, while non-interacting phages are separated by a distance, forming clusters with the bacteria as the center and a range of a certain threshold. During training, hyperparameters, including the value of K, learning rate, batch size, number of training epochs, and range threshold, are determined through multiple training tests. The model achieved accuracies of 0.8811, 0.5542, and 0.7280 on the whole genome dataset, 1000bp dataset, and 5000bp dataset, respectively, outperforming several existing methods compared with our model. For the case of unseen hosts, we used a dataset compiled by Shang J, where the test set contains 36 bacterial hosts provided only to the model during testing. Our model also achieved the highest accuracy of 0.6073. Compared with existing methods, the present invention constructs a feature extraction module with BiMamba2 as the core. The bidirectional processing characteristics of BiMamba2 improve the prediction accuracy. Then, DNA sequence data of various lengths are converted into a k-mer matrix of uniform size and easy to process and input into the model. By calculating the distance, the prediction accuracy is improved while also giving the model zero-sample capability.

[0016] How it works The present invention designs a phage-bacteria interaction prediction model based on deep learning, which mainly includes a data form conversion module, a feature extraction module, a training module and a prediction module. First, the data form conversion module converts each DNA sequence of different lengths into matrix data of uniform size. The matrix data can be efficiently processed by convolution, and the lower matrix dimension can contain high-dimensional information. For example, when k is 6, the dimension of the one-dimensional data is as high as 4096, while the k-mer matrix size of 6mer is only 64*64. Then, an embedding vector is extracted for each matrix by a hybrid model composed of CNN and BiMamba2. BiMamba2 can better capture the dependency relationship of each k-mer in the long sequence. Finally, the model learns the potential connection between phages and bacteria through these embedding vectors and contrast loss, and finally makes the phages and bacteria that can interact close to each other in the embedding space, and those that cannot interact are far away from each other, forming a cluster with bacteria as the midpoint and a range of a certain threshold in the space. During prediction, the distance between the model phage and bacteria is less than a pre-set threshold, then the phage is considered to be able to infect the bacteria, otherwise it is considered to be non-interactive.

[0017] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the present profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical content disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement of the above embodiments made according to the technical essence of the present invention, within the spirit and principles of the present invention, without departing from the content of the technical solution of the present invention, shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A deep learning-based prediction model for phage-bacteria interactions, characterized by: It includes data preprocessing module, data form conversion module, training module and prediction module; The data preprocessing module is used to download public datasets, obtain DNA sequences of phages and bacteria and their interaction relationship labels, perform preprocessing, and divide them into three subsets; The feature extraction module uses the feature extraction model to extract a unique embedding vector for each k-mer matrix as a vectorized representation of the data; The training module uses the extracted vectors to update the parameters of the initial model to achieve the best results; The prediction module uses the trained model to predict candidate bacteria that interact with the phage to be predicted.

2. A deep learning-based prediction model for phage-bacteria interactions according to claim 1, characterized in that: Data preprocessing refers to removing redundant sequences that appear repeatedly in the dataset and reshuffling the dataset, dividing it into the training set, validation set, and test set required for subsequent model training and testing in a ratio of 0.8:0.1:0.

1.

3. The deep learning-based bacteriophage-bacteria interaction prediction model according to claim 1, characterized in that: Obtaining a k-mer matrix means converting each DNA sequence of different lengths into a matrix data of uniform size. First, the k-mer frequency of each sequence is calculated, and the frequency count of each k-mer is filled into a matrix of size 1 according to the arrangement of the specified random seeds. in an all-zero matrix.

4. A deep learning-based prediction model for phage-bacteria interactions according to claim 3, characterized in that: The k-mer frequency calculation method is: ， Where, Indicates the The number of times k-mer appears in the entire sequence, Indicates the length of the sequence.

5. The deep learning-based prediction model for phage-bacteria interaction according to claim 1, characterized in that: The feature extraction module is composed of a hybrid model consisting of CNN and BiMamba2. First, CNN convolves, pools, and flattens the matrix, and the resulting vector is linearly mapped to the internal dimension of BiMamba2. Bidirectional modeling is performed and the fused features that retain bidirectional information are output.

6. The deep learning-based prediction model for phage-bacteria interaction according to claim 1, characterized in that: The training process of the training module is: Step 1: The initial model outputs the embedding vector and calculates the Euclidean distance between the phage vector and each bacterial vector; Step 2: Calculate the loss by contrastive loss and backpropagate it to the initial model to update the parameters; Step 3: Ultimately, the phages and bacteria that can interact are close to each other in the embedding space, while the non-interactive phages are far away from each other, forming a cluster with the bacteria as the center in the space.

7. The deep learning-based prediction model for phage-bacteria interaction according to claim 1, characterized in that: The prediction process is to receive test data through the trained model and calculate the distance between the phage and the bacteria. If their distance is less than the preset value, it is considered that the phage can infect the bacteria, otherwise it is considered that they cannot interact.

Images