Method and system for predicting range of hosts targeted by phages, and corresponding computer device or medium

Abstract

Provided in the present invention are a method and system for predicting hosts targeted by phages. In the method, host prediction can be performed simply by using tail proteins of phages, thus solving the problem of whole-genome feature redundancy, and moreover, efficient learning is achieved by using both physicochemical properties and contextual information features of tail proteins, such that the prediction accuracy at the family, genus and species levels is higher than that of other machine learning models (with an average improvement of 30%). Thus, the present invention overcomes the technical challenge of achieving both high recall and high precision, greatly expands the prediction range (categories), and solves major technical problems of current prediction tools, thus having broad application prospects in the field of prediction of hosts targeted by phages.

Description

Methods, systems, and corresponding computer equipment or media for predicting the host range of bacteriophage targeting. Technical Field

[0001] This invention belongs to the field of biotechnology, and more specifically, provides a method and system for predicting the target host range of bacteriophages, as well as corresponding computer equipment or media. Background Technology

[0002] Antibiotic therapy has been widely used to treat bacterial infections since its discovery (Langdon et al., 2016). However, due to the broad spectrum of antibiotics, they can have adverse effects on the in vivo microbiome, especially on probiotics (Pérez Cobas et al., 2013). Phage therapy, as another method for treating bacterial infections, has almost no impact on the healthy microbiome because bacteriophages are viruses that specifically infect and kill specific pathogenic bacterial hosts, exhibiting high host specificity. The most common phage therapy involves preparing phage mixtures by mixing different phages to obtain several defined host target ranges (Kelly et al., 2011). The biggest challenge in phage therapy remains pinpointing the phage to its target bacterial host, i.e., defining the host range of a single phage strain. However, current phage resources suffer from problems such as missing or poorly labeled host information, limiting the acquisition of host information and hindering the application of phage therapy.

[0003] Current methods for identifying the host range of bacteriophages in public databases typically rely on traditional experimental methods to confirm the existence of an infectious relationship between the bacteriophage and bacteria that have been successfully isolated experimentally. The most common and widely accepted experimental method for obtaining the host range of bacteriophages is the spot assay (microfluidic-PCR, PhageFish), also known as the double-layer agar plate method. Firstly, this technique involves co-culturing the host bacteria and the suspected bacteriophage sample, amplifying, and then observing lysis. This method is laborious, time-consuming, and expensive, especially in high-throughput analyses that require simultaneous monitoring of multiple bacteriophage-bacterial interactions. Secondly, for bacteria for which physical samples cannot be isolated, co-culturing with the bacteriophage is not possible, making it impossible to determine the true host range using this method.

[0004] Currently, bioinformatics methods used for phage-host prediction are mainly divided into two categories: those that utilize sequence similarity between phage and host genomes and those that use machine learning methods that do not rely on alignment.

[0005] Sequence alignment-based prediction methods primarily utilize the nucleotide similarity between bacteriophage and host genomes. Typically, bacterial and archaea genomes from the NCBI RefSeq database are used as host reference databases. Blastn (nucleotide-nucleotide) and Blastx (nucleotide-protein) alignments are used to identify nucleotide sequence similarities between the input bacteriophage sequence and the candidate host. These similarity regions correspond to integrated proviruses, host-encoded CRISPR spacers, auxiliary metabolic genes (AMGs), or shared tRNAs, reflecting phage-host co-evolution (Edwards et al., 2016; Camacho et al., 2009). Large sequence similarity regions (>5 kbp) in the alignment results are usually due to the integration of prophages very similar to unknown host phages into the host genome (Canchaya et al., 2004). Short, sequence-similar regions resulting from horizontally transferred genes (e.g., AMG) or insertion sites (e.g., tRNA) are often associated with successful phage infection and adaptation of the bacterial host (Breitbart et al., 2018). Conversely, high similarity between phage and bacterial CRISPR spacer regions reflects successful defense by the bacterial host against closely related or similar phages (Makarova et al., 2020).

[0006] Sequence alignment-based methods have shown high accuracy and medium to high recall on the test sets used in their respective benchmarks. However, because these methods rely entirely on sequence similarity as a prerequisite, they struggle to make predictions in the absence of known phage-bacterial pairings, placing high demands on the comprehensiveness and quality of the phage and host inclusion in the reference database. Furthermore, alignment conditions significantly impact the results; strict alignment parameters lead to low recall, while lenient alignment conditions result in unreliable results. This means that while maintaining accuracy, alignment-based methods struggle to achieve high recall. For example, GPD uses the most stringent conditions (0 mismatch), resulting in the highest annotation accuracy but the lowest annotation ratio. Moreover, statistics show that approximately 40% of bacteria possess the CRISPR-Cas system (Horvath and Barrangou, 2010), indicating that a significant portion of bacteria cannot pair with phages using CRISPR spacer alignment methods.

[0007] Phage host prediction methods that do not rely on sequence alignment predict host interactions based on the similarity of sequence composition features between the phage and host genomes. This primarily reflects the adaptation of the phage genome to the host's replication, transcription, and translation mechanisms (Roux et al., 2015). The most common sequence composition features are the composition and frequency of nucleotide k-mers in the genome. Most tools compare k-mer frequencies between phages from unknown hosts and host reference genome databases, based on the assumption that viral and host genomes typically exhibit similar sequence composition biases and k-mer frequency patterns. HostPhinder predicts phage-host interactions by comparing the 16-mer composition between phage genome sequences from unknown hosts and those from known hosts (Villarroel et al., 2016). PHP and VirHostMatcher utilize 4-mer and 6-mer frequency vectors between phages from unknown hosts and reference host genome databases (Lu et al., 2021; Ahlgren et al., 2017). WIsH performs similar k-mer-based comparisons, but to reduce noise in short sequence k-mer frequency signals and better handle short phage sequences, it trains an 8th-order homologous Markov model on each host genome and calculates the maximum likelihood of the phage sequence in each trained Markov model (Galiez et al., 2017). ILMF-VH predicts phage 6-mer frequencies by comparing them with known phage-host species networks (Liu et al., 2019). As an alternative to nucleotide k-mer frequencies, it predicts protein sequences based on characteristics in the phage and host genomes, including calculating the frequency of each amino acid in the protein sequence, the abundance of selected chemical elements in the protein (e.g., carbon, hydrogen, nitrogen, oxygen, and / or sulfur), and the molecular weight of the protein as model features (Boeckaerts et al., 2021; Li et al., 2021; Leite et al., 2018). Boeckaerts' tool uses phage tail proteins as input sequences for host prediction, covering seven host types: Staphylococcus aureus, Acinetobacter baumannii, Escherichia coli, Clostridium difficile, Klebsiella pneumoniae, Pseudomonas aeruginosa, and Salmonella. The other tools use whole genomes and whole proteomes as inputs for host prediction.

[0008] Existing phage host prediction tools, which do not rely on alignment, are less dependent on database performance and sequence similarity compared to methods that rely on sequence similarity between the phage and host genomes. They are also more advantageous for identifying novel phages. However, these methods suffer from problems such as feature redundancy leading to low model accuracy and limited prediction categories. Feature redundancy arises because existing methods select features based on all protein-protein interactions or all nucleic acid frequencies. The entire genome sequence contains a large amount of redundant information, which may not be relevant in the interaction between the phage and its host, thus obscuring the core information of the interaction: phage tail proteins. This results in significant dataset redundancy, overshadowing key sequence information and leading to low prediction accuracy. Boeckaerts' tool uses phage tail proteins as input sequences for host prediction and achieves good accuracy, but its prediction categories are limited to only seven, limiting its application scope.

[0009] Therefore, in this field, the technology for predicting phage target hosts needs to be further improved, especially in balancing accuracy, precision, and recall, and expanding the range of predicted species. Summary of the Invention

[0010] The purpose of this invention is to establish a precise, sensitive, and broad method for predicting the host range of bacteriophages, significantly reducing experimental manpower costs and time, further promoting the application of phage therapy, and advancing research on the correlation between bacteriophages, bacteria, and human health. Here, we explore a multi-level host prediction method based on phage tail protein sequences and establish a phage host prediction framework (PHPF) using a combination of machine learning and deep learning. This framework can simultaneously process the physicochemical and sequence information of protein sequences, thus leveraging the advantages of mixed information. Results show that our framework outperforms other existing and commonly used methods (including the Boeckaerts method). Furthermore, this model can be used for prediction at different host levels, overcoming the limitation of only seven categories in the Boeckaerts et al. study.

[0011] Therefore, in a first aspect, the present invention provides a method for predicting a bacteriophage-targeted host, the method comprising:

[0012] (1) Obtain the amino acid sequence of the tail protein of bacteriophages of known hosts;

[0013] (2) Convert the amino acid sequence described in step (1) into a matrix to obtain the physicochemical information of the amino acid sequence described in step (1), including molecular weight, protein length, isoelectric point, aromaticity, instability, flexibility, aliphatic amino acid index, uncharged polar amino acid index, polar amino acid index, hydrophobic amino acid index, positively charged amino acid index, negatively charged amino acid index, sulfur-containing amino acid index, amide-containing amino acid index, and alcohol-containing amino acid index.

[0014] (3) Use the matrix described in step (2) as input to train the Gradient Boosting Machine (GBM) model, use the physicochemical information described in step (2) as input to train the Convolutional Neural Networks with Attention (CNNA) model, and use the outputs of the Gradient Boosting Machine model and the Convolutional Neural Networks with Attention model as input to train the MLR model to obtain the phage-targeted host prediction model.

[0015] (4) Predict the host of the phage using the phage-targeted host prediction model described in step (3).

[0016] In one implementation, the amino acid sequence described in step (1) is processed:

[0017] a) Delete amino acid sequences that are less than 50 or greater than 1500;

[0018] b) Delete the amino acid sequence containing the unidentified amino acid "X";

[0019] c) Delete the same amino acid sequence that has different hosts;

[0020] d) Delete redundant amino acid sequences.

[0021] In one implementation, step (1) obtains the tail protein nucleotide sequence of a bacteriophage of a known host, and then obtains the amino acid sequence.

[0022] In a further embodiment, the nucleotide sequence is processed as follows:

[0023] Delete nucleotide sequences containing the unidentified amino acid "N".

[0024] In one implementation, the method for converting the amino acid sequence into a matrix in step (2) is as follows: the amino acid sequence is represented by a single-hot matrix, and multiple matrices are filled with the number 0 so that the multiple matrices have the same dimension.

[0025] In one implementation, the gradient booster described in step (3) is LightGBM, XGBoost, CatBoost, HistGradientBoosting, or GPU-Accelerated Gradient Boosting.

[0026] In one implementation, the attention-based convolutional neural network model described in step (3) includes three sub-networks, a convolutional layer, an SE block, and a fully connected layer.

[0027] In a further embodiment, the subnetwork includes convolutional neural layers with attention mechanisms.

[0028] In one implementation, the input to step (4) is the tail protein sequence of the phage to be predicted.

[0029] In one implementation, the input to step (4) is the whole genome sequence of the phage to be predicted, and step (4) also includes first predicting the tail protein sequence of the phage to be predicted.

[0030] In a second aspect, the present invention provides a system for predicting bacteriophage targeting a host, the system comprising:

[0031] The data acquisition module is used to acquire the amino acid sequence of the tail protein of bacteriophages with known hosts;

[0032] The data processing module is used to convert the amino acid sequence described in the data acquisition module into a matrix and obtain the physicochemical information of the amino acid sequence described in the data acquisition module, including molecular weight, protein length, isoelectric point, aromaticity, instability, flexibility, aliphatic amino acid index, uncharged polar amino acid index, polar amino acid index, hydrophobic amino acid index, positively charged amino acid index, negatively charged amino acid index, sulfur-containing amino acid index, amide-containing amino acid index, and alcohol-containing amino acid index.

[0033] The model training module is used to train a gradient booster model by taking the matrix described in the data processing module as input, to train a convolutional neural network model with attention mechanism by taking the physicochemical information described in the data processing module as input, and to train an MLR model by taking the outputs of the gradient booster model and the convolutional neural network model with attention mechanism as input, thereby obtaining a phage-targeted host prediction model.

[0034] The host prediction module is used to predict the host of the bacteriophage using the phage-targeted host prediction model described in the model training module.

[0035] In a third aspect, the present invention provides a computer device including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described in the first aspect.

[0036] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0037] The prediction model of this invention can make predictions using only tail proteins instead of the whole genome sequence for input data, which solves the problem of redundancy of whole genome features. At the same time, it uses the physicochemical and contextual information features of tail proteins for efficient learning. The prediction accuracy is higher than other machine learning models at the family, genus and species level (an average improvement of 30%). It breaks through the technical difficulty of balancing high recall and high accuracy, greatly expands the prediction range (species), and solves the main technical problems in current prediction tools. It has broad application prospects in the field of phage targeting host prediction. Attached Figure Description

[0038] Figure 1 shows a schematic diagram of the overall integrated framework logic of the present invention.

[0039] Figure 2 shows a performance comparison of the model of the present invention on a drug resistance dataset. The ensemble model constructed in this invention performs best when compared with other models.

[0040] Figure 3 shows a performance comparison of the integrated model PHPF, the LightGBM-based module, and the CNNA-based module of the present invention.

[0041] Figure 4 shows the detailed species-level (family, genus, species) results obtained by applying the integrated model of the present invention to a public intestinal phage database for host prediction. The outermost layer indicates that there is a lot of phage host information that cannot be obtained by the original methods, but can be discovered by the integrated model of the present invention.

[0042] Figure 5 illustrates the host prediction using the input tail protein sequence in an embodiment of the present invention. The output model predicts the host species level at different confidence levels, namely family, genus, and species.

[0043] Figure 6 illustrates the host prediction using the input bacteriophage genome sequence in an embodiment of the present invention. The output model predicts the host species level at different confidence levels, namely family, genus, and species. Detailed Implementation

[0044] Based on phage tail proteins, the inventors have developed a model that accurately predicts the host species range of phages using integrated machine learning and deep learning algorithms. This model will directly impact the field of phage research and advance the development of phage therapy. The method uses phage tail proteins for host prediction and integrates machine learning and deep learning to learn and model the physicochemical properties and sequence context information of the tail proteins. It achieves a high accuracy rate of over 90% on drug-resistant phage datasets and constructs host prediction models at different species levels (family, genus, species). Application in publicly available intestinal phage databases can uncover previously undiscovered host information, effectively improving the availability of phage resources, especially in intestinal phage databases. Furthermore, this invention can effectively address the rapid application of phage range in clinical phage cocktail therapy for treating infected bacteria.

[0045] The inventors explored a multi-level host prediction method based on phage tail protein sequences and established a hybrid framework for phage host prediction, PHPF, using a combination of machine learning and deep learning. This framework can simultaneously process both physicochemical and sequence information of protein sequences, thus leveraging the advantages of mixed information. Results show that our framework outperforms other existing and commonly used methods (including the Boeckaerts method). Furthermore, this model can be used for predictions at different host levels, overcoming the limitation of only seven categories in the Boeckaerts et al. study.

[0046] The method of this invention utilizes machine learning and deep learning, requiring only the input of a phage tail sequence (a portion of the phage genome) to complete host prediction. This significantly reduces data requirements, providing convenience for phage host prediction when a complete genome sequence is temporarily unavailable, eliminating the need for re-sequencing and reassembling the complete genome, thus saving time and costs. The method of this invention takes only 5-10 minutes from inputting phage sequence data to outputting predicted hosts and confidence lists at the family, genus, and species levels. Compared to experimental methods, the method of this invention significantly reduces time and labor costs, greatly improving the efficiency of phage host prediction. The family-level dataset of this invention contains 47 categories, the genus-level dataset contains 72 categories, and the species-level dataset contains 120 categories; therefore, the predicted species range can reach 47 families, 72 genera, and 120 species. Compared with previous studies that focused on a few bacterial species (Staphylococcus aureus, Acinetobacter baumannii, Escherichia coli, Clostridium difficile, Klebsiella pneumoniae, Pseudomonas aeruginosa, Salmonella, etc.) (Boeckaerts et al., 2021), the training dataset of this invention has more categories, enabling it to predict a wider range of species with higher accuracy. This breaks through the limitation of predicting only a few species and can be applied to scenarios where more different pathogens are present.

[0047] 1) In this invention, an example of constructing a dataset of drug-resistant bacteriophage tail proteins is as follows.

[0048] Drug-resistant bacteria are a major application area for phage therapy, with a wealth of available data and a rich research background. Similar to Boeckaerts et al. (2021), we first focused on seven common drug-resistant bacteria, including Staphylococcus aureus, Acinetobacter baumannii, Escherichia coli, Clostridium difficile, Klebsiella pneumoniae, Pseudomonas aeruginosa, and Salmonella enterica. To construct the dataset, we first collected whole-genome sequences of bacteriophages from the NCBI database (https: / / www.ncbi.nlm.nih.gov). A total of 3,116 phage genomes were collected from NCBI by searching for the keywords 'Staphylococcus', 'Acinetobacter', 'Escherichia', 'Clostridium', 'Klebsiella', 'Pseudomonas', and 'Salmonella'. Next, we extracted phage tail protein sequences from the whole-genome sequences using a two-step process. The first step was to annotate the proteins from the phage genomes using a bioinformatics workflow, and the second step was to extract the tail protein sequences from the annotation results using the keyword "tail". We then obtained 21,264 tail protein sequences, each consisting of a coding DNA sequence and its corresponding amino acid sequence. We also collected data from a paper published by Boeckaerts et al., which contained a total of 887 phage tail receptor-binding protein sequences. Finally, we combined the phage tail protein datasets from these two sources into the original dataset, containing 22,151 phage tail protein sequences. After collection, three filters were applied to clean the phage tail protein dataset and the phage tail RBP dataset. First, sequences shorter than 50 amino acids were removed to prevent predicted protein regions from being too short. Sequences longer than 1500 amino acids were also removed from the dataset to improve computational efficiency and resource utilization, as we found that amino acid sequences longer than 1500 amino acids accounted for less than 3% of the tail protein dataset. Second, amino acid sequences containing undetermined amino acid "X" and DNA sequences containing undetermined nucleotide "N" were excluded. Finally, identical sequences with different hosts were discarded, and sequence redundancy was removed from each dataset.

[0049] 2) In this invention, an example of constructing a tail protein dataset of all types of bacteriophages at different host species levels is as follows.

[0050] All phage genomes were screened from the NCBI database (https: / / www.ncbi.nlm.nih.gov / genome / browse / #! / viruses / ), removing sequences containing partial or complete CDS. The genome sequences (FASTA format) and annotation files (GBFF format) were downloaded via FTP according to the table. For all phages, the specific host information was determined by extracting the information from the line containing "host=" in the original phage genome annotation file (GBFF format). If the "host=" host information record was empty, the species information before the Phage delimiter in the GeneBank tile (ORGANISM) was used as the host information. For example, for phage AF234172, the "host=" host information record in the annotation file was empty, but the GeneBank tile showed "ORGANISM: Escherichia virus P1", so Escherichia was used as the host information. After obtaining the host information, the NCBI species classification toolkit taxonkit is used to obtain the host's taxonomy id and taxonomy level information (parameter: taxonkit name2taxid--show-rank). Based on the taxonomy id, the host species level information is further obtained and standardized into the format phylum, class, order, family, genus, species, strain (parameter: taxonkit lineage|taxonkit reformat|cut-f 1,3). The phage-host species classification level information is then obtained.

[0051] 3) In this invention, the inventors explored the establishment of a phage host species hierarchy prediction model.

[0052] To find the most suitable method for phage host prediction, we explored different approaches from the perspectives of feature encoding and machine learning algorithms. We first used fixed-length features from Boeckaerts et al. to represent the physicochemical information of protein sequences (i.e., molecular weight, protein length, isoelectric point, aromaticity, instability, flexibility, aliphatic amino acid index, uncharged polar amino acid index, polar amino acid index, hydrophobic amino acid index, positively charged amino acid index, negatively charged amino acid index, sulfur-containing amino acid index, amide-containing amino acid index, and alcohol-containing amino acid index). We then tested five common machine learning algorithms, including K-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), Logistic Regression (LR), Random Forest (RF), and Gradient Boosting (GB). We also applied the machine learning algorithm Light Gradient Boosting Machine (LightGBM) (Ke et al., 2017), a novel gradient boosting algorithm that further optimizes decision tree-based ensemble models from three dimensions: samples, features, and tree growth strategy. Traditional Gradient Boosting Decision Trees (GBDTs), whose base learner, Classification and Regression Trees (CART) (Lewis, 2000), selects the best splitting feature and the best splitting value to the maximum information gain at each node by searching all features and all possible splitting values ​​in the data, which is time-consuming. To address this issue, LightGBM uses Gradient-Based One-Sided Sampling (GOSS), which reduces the number of samples by retaining samples with large gradients and randomly discarding the rest. LightGBM also uses Exclusive Feature Bundling (EFB) to reduce the number of mutually exclusive features, bundling them into a single feature when these features rarely have non-zero values ​​simultaneously. In addition to GOSS and EFB, LightGBM grows leaf-by-leaf rather than checking all previous leaves for each new leaf to grow layer by layer (Al Daoud, 2019), which reduces more errors and achieves better accuracy.

[0053] The methods described above rely on fixed-length features for prediction, neglecting protein sequence information. Therefore, we use one-hot matrices to represent the sequence information of amino acid sequences. Since the amino acid sequences in our dataset have varying lengths, their one-hot matrices also have different shapes. Therefore, we ultimately transform each one-hot matrix into a 1500×20 matrix by padding short sequences with "0". However, data encoded by one-hot matrices cannot be directly used by the aforementioned machine learning methods, while deep learning can handle such high-dimensional data. Therefore, we build a deep learning model based on a Convolutional Neural Network with Attention (CNNA). In CNNA, we first construct three convolutional sub-networks based on multiple convolutional kernels of different sizes. These kernels receive the same one-hot matrix input to capture sequence information of varying lengths. Then, the outputs of the three sub-networks are merged and exported to another convolutional layer, followed by batch normalization. After each convolutional layer, we add an attention mechanism called a Squeeze-and-Excitation Block (SE block). The SE block can reassign weights and recalibrate channel features for each output of the convolutional layer to capture more critical information. The SE block consists of two parts: squeeze and adaptive recalibration. Spatial information is compressed into a channel descriptor that incorporates global information because the filter only focuses on local information. Finally, the output of the SE block is followed by a maxpooling layer, then a fully connected layer, and finally batch normalization to predict the probability of each host.

[0054] The overall model framework is shown in Figure 1. In Figure 1, A shows the LightGBM-based module receiving the physicochemical information of the protein sequence as input; B shows the CNNA-based module processing the sequence information of the amino acid sequence; C shows multivariate logistic regression as a stacked ensemble classifier fusing information from the LightGBM-based module and the CNNA-based module; D shows the sub-network in the CNNA-based module in B containing convolutional neural layers with attention mechanisms, specifically including input, convolution, SE block, convolution, SE block, convolution, SE block, and output.

[0055] Specific model parameters include:

[0056] Model: LightGBM

[0057] Custom layer: SElayer Note: Assume that the input of SElayer is a tensor X and the input channel is C. The output of SElayer is obtained by multiplying X by the output of fc2.

[0058] Model: CNNA

[0059] Model: MLR (Classification Output)

[0060] Example

[0061] Example 1:

[0062] Steps: Input the data described in 1) above into the model of this invention and the Boeckaerts tool (wherein, in addition to the physicochemical information of the amino acid sequence, the input data of Boeckaerts also includes nucleotide frequency, GC content, codon frequency, codon usage preference, amino acid frequency, amino acid index of α-helix, amino acid index of β-sheet, amino acid index of turn, protein composition, protein transition, protein Z-scale), statistically analyze the results and calculate and compare indicators such as accuracy (ACC), F1 score (i.e., the weighted average of precision and recall), MCC (Matthews correlation coefficient).

[0063] The results are shown in Figure 2. Compared with the Boeckaerts tool, the method of this invention significantly improves the accuracy of phage host prediction, reaching 90.48%. The data used are from the data described in section 1) above. The results show that the Boeckaerts tool has an accuracy of 0.8378, an F1 score of 0.8352, and an MCC of 0.7765, while the ensemble model of this invention has an accuracy of 0.9048, an F1 score of 0.9048, and an MCC of 0.8627. All these indicators are superior to the Boeckaerts tool.

[0064] Example 2:

[0065] Steps: Randomly sample 20% subsets from the data described in 2) 10 times and mix them to obtain the test dataset. Input the model of this invention and other common machine learning models (KNN, etc.) into it. Statistically analyze the results and calculate metrics such as accuracy, F1 score, and MCC for comparison.

[0066] Results show that this invention develops a phage host range prediction framework based on tail protein sequences, integrating machine learning and deep learning. This framework can accurately predict phage host families, genera, and species, achieving accuracies of 88.61%, 87.68%, and 89.30%, respectively. Compared to other tools using machine learning models in parallel, these accuracy improvements are 30%, 31%, and 30% (Table 1). At 95% model confidence precision, the accuracy at the family, genus, and species levels reaches 95.97%, 97%, and 99%, with corresponding recalls of 82.58%, 80.72%, and 77.19%. At 69% model confidence precision, the accuracy at the family, genus, and species levels reaches 90.68%, 91.54%, and 96.8%, with corresponding recalls of 87.83%, 86.46%, and 83.33% (Table 2). The above model performance results demonstrate that the present invention significantly outperforms other models in terms of accuracy. It maintains high prediction accuracy and high recall at model confidence levels (high, medium, and low) (95%, 84%, and 69%), with average precision reaching 97.32-93.01% and average recall reaching 80.16-85.87%. The method of this invention possesses the characteristics of high precision and high recall, overcoming the problem in previous tools where recall and precision could not be simultaneously achieved.

[0067] Table 1. Comparison of accuracy performance of this invention with other models at the family, genus, and species levels.

[0068] Table 2. Recall and precision of the present invention at the family, genus, and species levels under different confidence levels.

[0069] Example 3:

[0070] Figure 3 shows the superiority of the integrated machine learning-deep learning model framework design that combines physicochemical information and sequence information. The data used is obtained by mixing the 20% subset that was randomly sampled 10 times as described in 2) above.

[0071] Steps: Randomly select 20% subsets from the data described in 2) above 10 times and mix them to obtain the test dataset. Input the ensemble model of this invention, the CNNA model, and the LightGBM model, and statistically analyze the results and calculate indicators such as accuracy, F1 score, and MCC for comparison.

[0072] Results: To explore the source of the superior performance of the PHPF model of this invention, the inventors designed ablation experiments to detail the contribution of the model ensemble design. Compared with individual modules based on LightGBM and CNNA, PHPF's metrics ACC, F1 score, MCC, ACC (Phage I), and ACC (Phage II) are improved by 1.37%, 1.38%, 1.86%, 1.66%, and 1.39%, respectively. Analysis of the ablation experiment results shows that integrating LightGBM and CNNA significantly outperforms either model alone in prediction performance; that is, the superior prediction quality stems from the stacked-ensemble framework design that applies the prediction inputs.

[0073] Example 4:

[0074] The existing gut phage database (GPD) contains a large number of phage resources with unknown hosts. The integrated model of this invention can predict the host spectrum information of phages with unknown hosts, which greatly improves the availability of phage resources (Figure 4). The data comes from 14,208 phage sequences with unknown hosts in the gut phage database (GPD).

[0075] Steps: Process the above phage sequence using the method described in 3) above, input the processed data into the model of this invention, wait 3-5 minutes, and output the phage name and the family, genus, and species level prediction results of the host it infects.

[0076] Results: The host prediction results were clustered and visualized (Figure 4). The annular diagram has 8 layers from the inside out, which are the model's prediction results at the family, genus, and species levels, the species classification of bacteriophages, the distribution of the number of bacteriophage tail proteins, the distribution of the number of tail proteins such as spike and fiber that have been verified to be closely related to the binding of bacteriophages to their hosts, and whether the host information is previously identified. The category to which the model prediction results belong at the family, genus, and species levels is shown in the figure caption below.

[0077] Figure 4 shows that the model performs well in predicting on the GPD database. Compared with traditional host allocation methods, the model established in this study has discovered more previously unknown host information, with a wide range of prediction results and a maximum prediction resolution at the species level. It also has good prediction results for different bacteriophage species.

[0078] Example 5:

[0079] The integrated model of this invention outputs the different species levels of hosts that the tail protein may infect.

[0080] Steps: Input the sequence of one phage tail protein into the ensemble model, wait 3-5 minutes, and output the host range at different confidence levels. The results are shown in Figure 5.

[0081] Results: Figure 5 illustrates the host prediction results using the input tail protein in this embodiment. The output ensemble model predicts the host species level at different confidence levels, namely family, genus, and species. This is extremely helpful for the application of engineered bacteriophages in synthetic products.

[0082] Example 6:

[0083] The integrated model of this invention outputs the different species levels of hosts that the whole bacteriophage may infect.

[0084] Steps: Input one phage genome sequence into the ensemble model, wait 3-5 minutes, and output the host range at different confidence levels. The results are shown in Figure 6.

[0085] Results: Figure 6 illustrates the host prediction results obtained by inputting bacteriophage genome sequences in this embodiment. The output ensemble model predicts the host species level at different confidence levels, namely family, genus, and species. Compared to inputting tail protein sequences alone, this invention can automatically predict tail proteins from the whole genome first and then predict the host. That is, this invention accepts two input formats, including whole genome sequences and tail protein sequences.

Claims

1. A method for predicting bacteriophage target hosts, characterized in that, The method includes: (1) Obtain the amino acid sequence of the tail protein of bacteriophages of known hosts; (2) Convert the amino acid sequence described in step (1) into a matrix to obtain the physicochemical information of the amino acid sequence described in step (1), including molecular weight, protein length, isoelectric point, aromaticity, instability, flexibility, aliphatic amino acid index, uncharged polar amino acid index, polar amino acid index, hydrophobic amino acid index, positively charged amino acid index, negatively charged amino acid index, sulfur-containing amino acid index, amide-containing amino acid index, and alcohol-containing amino acid index. (3) Use the matrix described in step (2) as input to train the gradient booster model, use the physicochemical information described in step (2) as input to train the convolutional neural network model with attention mechanism, and use the outputs of the gradient booster model and the convolutional neural network model with attention mechanism as input to train the MLR model to obtain the phage target host prediction model. (4) Predict the host of the phage using the phage-targeted host prediction model described in step (3).

2. The method according to claim 1, characterized in that, The amino acid sequence described in step (1) is processed as follows: a) Delete amino acid sequences that are less than 50 or greater than 1500; b) Delete the amino acid sequence containing the unidentified amino acid "X"; c) Delete the same amino acid sequence that has different hosts; d) Delete redundant amino acid sequences.

3. The method according to claim 1 or 2, characterized in that, In step (1), the tail protein nucleotide sequence of the bacteriophage of a known host is obtained, and then the amino acid sequence is obtained.

4. The method according to claim 3, characterized in that, The nucleotide sequence is processed as follows: Delete nucleotide sequences containing the unidentified amino acid "N".

5. The method according to any one of claims 1-4, characterized in that, The method for converting the amino acid sequence into a matrix in step (2) is as follows: the amino acid sequence is represented by a single-hot matrix, and multiple matrices are filled with the number 0 so that the multiple matrices have the same dimension.

6. The method according to any one of claims 1-5, characterized in that, The gradient booster mentioned in step (3) is LightGBM, XGBoost, CatBoost, HistGradientBoosting, or GPU-Accelerated Gradient Boosting.

7. The method according to any one of claims 1-6, characterized in that, The attention-based convolutional neural network model described in step (3) includes three sub-networks, a convolutional layer, an SE block, and a fully connected layer.

8. The method according to claim 7, characterized in that, The subnetwork contains convolutional neural layers with attention mechanisms.

9. The method according to any one of claims 1-8, characterized in that, The input for step (4) is the tail protein sequence of the phage to be predicted.

10. The method according to any one of claims 1-9, characterized in that, The input for step (4) is the whole genome sequence of the phage to be predicted. Step (4) also includes first predicting the tail protein sequence of the phage to be predicted.

11. A system for predicting bacteriophage targeting a host, characterized in that, The system includes: The data acquisition module is used to acquire the amino acid sequence of the tail protein of bacteriophages with known hosts; The data processing module is used to convert the amino acid sequence described in the data acquisition module into a matrix and obtain the physicochemical information of the amino acid sequence described in the data acquisition module, including molecular weight, protein length, isoelectric point, aromaticity, instability, flexibility, aliphatic amino acid index, uncharged polar amino acid index, polar amino acid index, hydrophobic amino acid index, positively charged amino acid index, negatively charged amino acid index, sulfur-containing amino acid index, amide-containing amino acid index, and alcohol-containing amino acid index. The model training module is used to train a gradient booster model by taking the matrix described in the data processing module as input, to train a convolutional neural network model with attention mechanism by taking the physicochemical information described in the data processing module as input, and to train an MLR model by taking the outputs of the gradient booster model and the convolutional neural network model with attention mechanism as input, thereby obtaining a phage-targeted host prediction model. The host prediction module is used to predict the host of the bacteriophage using the phage-targeted host prediction model described in the model training module.

12. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1-9.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-9.

Images