A bacteriophage promoter prediction method based on multi-source transfer learning

Abstract

The application discloses a bacteriophage promoter prediction method based on multi-source transfer learning, which comprises the following steps: data set construction: collecting positive samples of target domains and multi-source domains, using an intergroup shuffle method to reconstruct different domain positive samples, obtaining corresponding target domain and multi-source domain negative samples, and finally obtaining a target domain and multi-source domain data set with balanced positive and negative sample ratios; multi-source transfer learning model construction: including four functional modules of data preprocessing, feature extraction, feature alignment and target classification; model training: dividing the data set into a training set and a validation set, using the training set to build and train the model, using the validation set to adjust the model parameters, and obtaining an optimal model; model prediction. The application adopts a multi-source transfer learning mode for small sample data to construct an optimal model for predicting and classifying bacteriophage promoters, improves the prediction effect, and makes contributions to subsequent bacteriophage promoter prediction and prediction development of other regulatory elements.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics computing technology, and in particular to a phage promoter prediction method based on multi-source transfer learning. Background Art

[0002] Bacteriophages are specialized viruses that infect and kill bacteria. Therefore, applying bacteriophages to address clinical diseases caused by bacteria holds high research value and practical significance. Over the past few decades, the number of sequenced phage genomes has increased exponentially. This wealth of genomic data requires accurate and efficient tools to assist biologists in genome annotation. Phage genome annotation encompasses both regulatory and functional elements. Promoters, as crucial regulatory elements, have a significant impact on phage gene expression.

[0003] Currently, there is very little known data on bacteriophage promoters. For example, the phiSITE database provides only 488 experimentally verified phage promoter sequences. Collection of phage promoter data at NCBI is challenging, and some data has not been experimentally verified. Due to this limited data, tools built using common prediction models have poor generalization capabilities. Furthermore, phage genomes are short and compact, and their promoter recognition differs from that of other prokaryotes. While many prokaryotic promoter prediction tools exist, these interspecies differences make it difficult to achieve satisfactory results using prokaryotic promoter prediction models for phage promoter prediction.

[0004] Therefore, in order to solve the above technical problems, it is urgent for those skilled in the art to propose a new technical means. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention proposes a phage promoter prediction method based on multi-source transfer learning, the purpose of which is to solve the technical problem of how to use multi-source transfer learning on small sample data to construct an optimal model to predict and classify phage promoters.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A phage promoter prediction method based on multi-source transfer learning comprises the following steps:

[0008] (1) Dataset construction: Collect positive samples from the target domain and multi-source domains, reconstruct the positive samples from different domains using the inter-group shuffling method, and obtain the corresponding negative samples from the target domain and multi-source domains. Finally, a target domain and multi-source domain dataset with a balanced ratio of positive and negative samples is obtained.

[0009] (2) Construction of multi-source transfer learning model: The multi-source transfer learning model includes four functional modules: data preprocessing, feature extraction, feature alignment, and target classification. The specific functions of each module are as follows:

[0010] Data preprocessing: converting nucleotide sequence data in the target domain and multi-source domain datasets into digital data that can be input into the feature extractor;

[0011] Feature extraction: Use feature extractors to obtain features of target domain and multi-source domain data;

[0012] Feature alignment: Calculate the distance between the target domain and multi-source domain features, and transfer the knowledge of multi-source domain data to the target domain data by minimizing the distance between the same category data in the target domain and multi-source domains and maximizing the distance between the different category data in the target domain and multi-source domains;

[0013] Target classification: Use a classifier to classify the target domain and multi-source domain feature data;

[0014] (3) Multi-source transfer learning model training: The target domain and multi-source domain datasets are divided into training sets and validation sets according to the five-fold cross-validation method. The training set is used to build and train the multi-source domain transfer learning model, and the validation set is used to adjust the model parameters to finally obtain the optimal model.

[0015] (4) Prediction: The external test set is input into the trained multi-source transfer learning model for prediction to obtain the probability value of it being a phage promoter.

[0016] Preferably, the specific steps of the inter-group scrambling method in step (1) are as follows:

[0017] S1. Split the positive sample into multiple segments, each with a length of k. If the sequence length is not divisible by k, the length of the last segment is the remainder of the promoter sequence length divided by k.

[0018] S2. Permutate and combine the fragments generated by each positive sample to obtain a new sequence. Use the globalxx function of the bioinformatics tool Biopython to calculate the similarity between the new sequence and the original sequence. If the similarity between the new sequence and the original sequence is greater than or equal to 0.7, permutate and combine the fragments again until the similarity between the generated sequence and the original sequence is less than 0.7.

[0019] Preferably, the specific steps of data preprocessing in step (2) are as follows:

[0020] Dictionary encoding: Since each sequence in the dataset consists of four nucleotides: A, C, G, and T, this step converts the nucleotides in the sequence into 1, 2, 3, and 4, respectively;

[0021] Sequence padding: Use the zero-padding method to fill each sequence with the number 0 to the length of the longest sequence in the data set;

[0022] Word embedding: Convert words into vector representations that can express the connection between each element in the sequence.

[0023] Preferably, the specific steps of feature extraction in step (2) are as follows:

[0024] First, a convolutional neural network (CNN) is used to perform a convolution operation on the data, and then the convolved data is input into a bidirectional gated recurrent unit (BiGRU) and output, and finally the target domain and multi-source domain features are obtained respectively.

[0025] Preferably, when aligning the features in step (2), the formula for calculating the distance between the target domain features and the multi-source domain features is:

[0026]

[0027] Where, Represents the feature vector of the i-th data in the k-th source domain dataset, represents the eigenvector of the jth data in the target domain dataset, and ||·|| represents the Frobenius normal form, which is to calculate the square root of the sum of the squares of the elements in the same position.

[0028] Preferably, when classifying the targets in step (2), the classifier used is a fully connected neural network (FCNN).

[0029] Compared with the existing technology, the present invention provides a phage promoter prediction method based on multi-source transfer learning, which has the following beneficial effects:

[0030] 1. The negative samples generated by the inter-group shuffling method in the present invention can effectively reduce the false positive rate of the model compared with the negative samples obtained by the previous random interception method.

[0031] 2. The present invention introduces knowledge of multi-source domain data, which is beneficial for the classifier to classify target domain data and improves the generalization ability of the model.

[0032] The present invention uses multi-source transfer learning on small sample data to construct an optimal model to predict and classify phage promoters. It has strong generalization ability, improves the prediction effect, and contributes to the subsequent development of phage promoter prediction and prediction of other regulatory elements. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a schematic diagram of the overall framework flow of the present invention;

[0034] Figure 2 Schematic diagram of the multi-source transfer learning model in the present invention. DETAILED DESCRIPTION

[0035] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0036] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside", "outside", etc., indicating directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore should not be understood as limiting the present invention.

[0037] like Figure 1 and Figure 2 As shown, a phage promoter prediction method based on multi-source transfer learning includes the following steps:

[0038] (1) Dataset construction:

[0039] The dataset in this embodiment consists of two parts: divided into target domain and multi-source domain datasets, 781 target domain positive samples, i.e., phage promoter sequences, were collected from the datasets of phiSITE, NCBI, and the existing phage promoter prediction tool PhagePromoter, and 12,575 multi-source domain positive samples, i.e., Bacillus subtilis promoter and Escherichia coli promoter, were collected from the Escherichia coli promoter dataset provided in the fifth edition of DBTBS for the generation and expression prediction of Bacillus subtilis promoter and Escherichia coli promoter; 781 target domain negative samples and 12,575 multi-source domain negative samples were obtained by using the inter-group shuffling method, and finally a complete dataset was obtained. The composition of the dataset used in this embodiment is shown in the following table:

[0040]

[0041] (2) Multi-source transfer learning model construction:

[0042] The multi-source transfer learning model includes four functional modules: data preprocessing, feature extraction, feature alignment, and target classification. The specific functions of each module are as follows:

[0043] Data preprocessing: Since each sequence in the dataset consists of four nucleotides: A, C, G, and T, dictionary encoding is used to convert the nucleotides in the sequence into 1, 2, 3, and 4, respectively. Sequence padding is then used to pad each sequence with the number 0 to the length of the longest sequence in the dataset. Since digital representation cannot reflect the positional relationship between each element in the sequence, word embedding is used to convert words into vector representation, which can effectively represent the relationship between each element in the sequence for input into the feature extractor.

[0044] Feature extraction: First, a convolutional neural network (CNN) is used to convolve the data, and then the convolved data is input into a bidirectional gated recurrent unit (BiGRU) and output, finally obtaining the target domain and multi-source domain features respectively.

[0045] Feature alignment: Calculate the distance between the target domain and multi-source domain features. By minimizing the distance between the same-category data in the target domain and multi-source domains and maximizing the distance between the different-category data in the target domain and multi-source domains, the knowledge of the multi-source domain data can be transferred to the target domain data. This helps the classifier classify the target domain data and improves the generalization ability of the model. The formula for calculating the distance between the target domain and multi-source domain features is:

[0046]

[0047] Where, Represents the feature vector of the i-th data in the k-th source domain dataset, represents the eigenvector of the jth data in the target domain dataset, and ||·|| represents the Frobenius normal form, which is to calculate the square root of the sum of the squares of the elements in the same position.

[0048] Target classification: Use a fully connected neural network (FCNN) as a classifier to classify the target domain and multi-source domain feature data;

[0049] (3) Multi-source transfer learning model training: The target domain and multi-source domain datasets are divided into training sets and validation sets according to the five-fold cross-validation method. The training set is used to build and train the multi-source domain transfer learning model, and the validation set is used to adjust the model parameters to finally obtain the optimal model.

[0050] (4) Prediction: The external test set is input into the trained multi-source transfer learning model for prediction to obtain the probability value of it being a phage promoter.

[0051] The above content describes in detail how the phage promoter prediction method based on multi-source transfer learning provided by the present invention uses multi-source transfer learning on small sample data to construct an optimal model to predict and classify phage promoters. The experimental results of the present invention are described below.

[0052] 1. Verification Standards

[0053] The validation criteria for measuring the overall performance of the model include sensitivity (SEN), specificity (SPE), precision (PRE), accuracy (ACC), Matthews correlation coefficient (MCC), and F1 score (F1), which are calculated as follows:

[0054]

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] Among them, TP (True positive) represents the number of true positives, that is, the number of real phage promoter sequences that are correctly predicted as phage promoter sequences; TN (True negative) represents the number of true negatives, that is, the number of real non-phage promoter sequences that are correctly predicted as non-phage promoter sequences; FP (False positive) is the number of false positives, that is, the number of phage non-promoter sequences that are predicted as phage promoter sequences; FN (False negative) is the number of false negatives, that is, the number of phage promoter sequences that are originally phage promoter sequences that are predicted as non-phage promoter sequences.

[0061] In addition, this example also uses AUC and AUPR to measure the overall performance of the model. Generally, the six indicators given in the above formula are affected by the threshold. That is, samples greater than or equal to the threshold are predicted as positive samples, while samples less than the threshold are considered negative samples. The default threshold is 0.5, but it can be adjusted manually. AUC and AUPR, on the other hand, are not affected by the threshold and range between 0 and 1. The closer to 1, the better the overall model performance. Therefore, they are often considered more important evaluation indicators.

[0062] 2. Verify the results

[0063] In this embodiment, in order to verify the superiority of the present invention, two groups of experiments were conducted: the first group of experiments did not use the multi-source domain dataset, but only used the target domain dataset to train the model. This model is a traditional deep learning model and is named Base; the second group of experiments used the multi-source domain dataset and the target domain dataset to train the model. This model is the multi-source transfer learning model proposed in the present invention and is named PromoterTL. The performance comparison of the above two models based on eight evaluation indicators (SEN, SPE, PRE, F1, MCC, ACC, AUC and AUPR) is shown in the following table:

[0064]

[0065] As shown in the table above, the PromoterTL model achieves better results in all aspects except SEN, which is lower than the Base model. In particular, the AUC improves by 0.04. This shows that multi-source transfer learning, by introducing multi-source domain datasets, can address the problem of small target domain samples and improve the model's predictive performance.

[0066] 3. Prediction effect

[0067] In order to test the prediction effect of the present invention, a test set was collected from NCBI. Ten and 61 phage promoter sequences were found in two phage genome sequences (Pseudomonas Phage PhiPsa17 and Staphylococcus Phage), respectively. Then, a corresponding number of non-promoter sequences were intercepted in the phage genome to obtain two test sets with balanced positive and negative samples, respectively referred to as test set 1 and test set 2.

[0068] The confusion matrix of the prediction results of the present invention on test set 1 is shown in the following table:

[0069]

[0070] The confusion matrix of the prediction results of the present invention on test set 2 is shown in the following table:

[0071]

[0072] On test set 1, the multi-source transfer learning model proposed in the present invention correctly predicted all phage promoter sequences and non-phage promoter sequences. On test set 2, the multi-source transfer learning model proposed in the present invention was able to predict most phage promoters, thus proving the effectiveness of the multi-source transfer learning method.

[0073] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A phage promoter prediction method based on multi-source transfer learning, characterized in that: The following steps are involved: (1) Dataset construction: Collect positive samples from the target domain and multi-source domains, reconstruct the positive samples from different domains using the inter-group shuffling method, and obtain the corresponding negative samples from the target domain and multi-source domains. Finally, a target domain and multi-source domain dataset with a balanced ratio of positive and negative samples is obtained. The specific steps of the inter-group scrambling method described in step (1) are as follows: S1. Split the positive sample into multiple segments, each with a length of k. If the sequence length is not divisible by k, the length of the last segment is the remainder of the promoter sequence length divided by k. S2. Permutate and combine the fragments generated from each positive sample to generate a new sequence. Calculate the similarity between the new sequence and the original sequence using the globalxx function of the bioinformatics tool Biopython. If the similarity between the new sequence and the original sequence is greater than or equal to 0.7, permutate and combine the fragments again until the similarity between the new sequence and the original sequence is less than 0.

7. (2) Construction of multi-source transfer learning model: The multi-source transfer learning model includes four functional modules: data preprocessing, feature extraction, feature alignment, and target classification. The specific functions of each functional module are as follows: Data preprocessing: converting nucleotide sequence data in the target domain and multi-source domain datasets into digital data that is input into the feature extractor; Feature extraction: Use feature extractors to obtain features of target domain and multi-source domain data; When aligning the features in step (2), the formula for calculating the distance between the target domain features and the multi-source domain features is: ； Where, Represents the feature vector of the i-th data in the k-th source domain dataset, Represents the feature vector of the jth data in the target domain dataset, It represents the Frobenius normal form, which is to calculate the square root of the sum of the squares of the elements in the same position; Feature alignment: Calculate the distance between the target domain and multi-source domain features, and transfer the knowledge of multi-source domain data to the target domain data by minimizing the distance between the same category data in the target domain and multi-source domains and maximizing the distance between the different category data in the target domain and multi-source domains; Target classification: Use a classifier to classify the target domain and multi-source domain feature data; (3) Multi-source transfer learning model training: The target domain and multi-source domain datasets are divided into training sets and validation sets according to the five-fold cross-validation method. The training set is used to build and train the multi-source domain transfer learning model, and the validation set is used to adjust the model parameters to finally obtain the optimal model. (4) Prediction: The external test set is input into the trained multi-source transfer learning model for prediction to obtain the probability value of it being a phage promoter.

2. The phage promoter prediction method based on multi-source transfer learning according to claim 1, characterized in that: The specific steps of data preprocessing in step (2) are as follows: Dictionary encoding: Since each sequence in the dataset consists of four nucleotides: A, C, G, and T, this step converts the nucleotides in the sequence into 1, 2, 3, and 4, respectively; Sequence padding: Use the zero-padding method to fill each sequence with the number 0 to the length of the longest sequence in the data set; Word embedding: Convert words into vector representations that can express the connection between each element in the sequence.

3. The method for predicting bacteriophage promoters based on multi-source transfer learning according to claim 1, characterized in that: The specific steps of feature extraction in step (2) are as follows: First, a convolutional neural network (CNN) is used to perform a convolution operation on the data, and then the convolved data is input into a bidirectional gated recurrent unit (BiGRU) and output, and finally the target domain and multi-source domain features are obtained respectively.

4. The method for predicting bacteriophage promoters based on multi-source transfer learning according to claim 1, characterized in that: When classifying the target in step (2), the classifier used is a fully connected neural network FCNN.

Images