A method, device, medium and equipment for predicting marine phage antibacterial lysin

Abstract

The application discloses a marine bacteriophage antibacterial lysin prediction method and device, medium and equipment, and relates to the fields of biological medicine and computational biology. The method comprises the following steps: constructing a training data set based on global known antibacterial protein sequences; training a protein sequence classification model based on deep learning by using the antibacterial protein data set to obtain a marine antibacterial lysin prediction model. After data standardization processing, the protein sequence of the bacteriophage to be predicted is input into the protein language model, the input sequence is subjected to self-attention modeling by a ProtT5 module, global structure and function related features are extracted, and a first feature vector is obtained; short sequence local dependency features are extracted by a ProtBERT module, and a second feature vector is obtained; the first and second feature vectors are weighted and summed to generate a fusion feature vector; and the fusion feature vector is subjected to probability prediction, and an antibacterial label and a prediction confidence of the marine bacteriophage sequence to be predicted are output.

Description

Technical Field

[0001] This invention relates to the fields of biomedicine and computational biology, and in particular to a method, apparatus, medium, and device for predicting antimicrobial lysins from marine bacteriophages. Background Technology

[0002] In the field of antimicrobial drug development, traditional methods mainly rely on screening antimicrobial components from nature or obtaining antimicrobial drugs through chemical synthesis. Lysozyme, as a natural antimicrobial substance, has been widely used to combat bacterial infections, especially against certain drug-resistant bacteria.

[0003] However, traditional lysozyme development methods cannot overcome the problem of high false positives: In the past, traditional machine learning methods (such as support vector machines, random forests, etc.) were mostly used for antimicrobial protein prediction, but these algorithms have simple logical frameworks, making it difficult to handle massive and complex data, and the prediction results are often accompanied by high false positive rates, resulting in low robustness and reliability of the models. Summary of the Invention

[0004] Based on this, in order to solve the technical problems in the prior art, the present invention provides a method, apparatus, medium and device for predicting marine bacteriophage antimicrobial lysins.

[0005] This invention provides a method for predicting the antilysin of marine bacteriophages, comprising:

[0006] A deep learning network including a ProtT5 module, a ProtBERT module, a weighted module, and a classification module was constructed. A dataset was built based on the collected antimicrobial protein sequence data. The constructed deep learning network was trained using the dataset to obtain a marine bacteriophage antimicrobial lysin prediction model.

[0007] The marine phage antilysin sequence to be predicted is input into the prediction model. The ProtT5 module encodes and decodes the sequence to capture global features and obtain a first feature vector. The ProtBERT module encodes the sequence to capture local sequence dependencies and obtains a second feature vector. The weighting module sums the first and second feature vectors to obtain a fused feature vector. The classification module performs probability prediction on the fused feature vector and outputs the confidence level that the marine phage protein sequence is an antimicrobial protein sequence.

[0008] Furthermore, the encoding and decoding operations of the predicted marine bacteriophage antilysin sequence using the ProtT5 module specifically include:

[0009] In the encoding stage of ProtT5: the position encoding layer in the ProtT5 module encodes the position of the marine bacteriophage antilysin sequence to be predicted, and combines it with protein sequence embedding operations to obtain a protein sequence embedding vector; the multi-head self-attention mechanism layer in the ProtT5 module captures the global dependencies of the protein sequence embedding vector to generate a context-weighted protein sequence vector; the first feedforward neural network layer in the ProtT5 module performs a nonlinear transformation on the weighted protein sequence vector, and combines residual connections and normalization operations to obtain protein-encoded features;

[0010] In the decoding stage of ProtT5, the protein-coding features are reconstructed using the sequence reconstruction layer in the ProtT5 module, generating a predicted amino acid sequence vector. The amino acid sequence vector is then masked using the masked multi-head self-attention mechanism layer in the ProtT5 module to capture local dependencies between amino acids, resulting in an amino acid mask weighted vector. The global dependency between the protein-coding features and the amino acid mask weighted vector is calculated using the interactive attention mechanism layer in the ProtT5 module, yielding a fused weighted amino acid sequence vector. Finally, the fused weighted amino acid sequence vector undergoes a nonlinear transformation using the second feedforward neural network layer in the ProtT5 module, combined with residual connections and normalization operations, to obtain the first feature vector.

[0011] Furthermore, the encoding operation of the predicted marine bacteriophage antilysin sequence using the ProtBERT module specifically includes:

[0012] The position encoding layer in the ProtBERT module aligns the predicted marine bacteriophage antilysin sequence to a preset fixed length and then performs position encoding. Combined with protein sequence embedding operations, a protein sequence embedding vector is obtained.

[0013] The protein sequence embedding vector is weighted by a multi-head self-attention layer in the ProtBERT module to capture the global dependencies of amino acids and obtain a weighted protein sequence vector.

[0014] The weighted protein sequence vector is randomly masked using the mask language modeling layer in the ProtBERT module to generate amino acid mask features.

[0015] The second feature vector is obtained by pooling the amino acid mask features through the pooling layer in the ProtBERT module.

[0016] Furthermore, the step of obtaining a fused feature vector by weighting and summing the first and second feature vectors using a weighting module specifically includes:

[0017] E = α·E ProtT5 +(1-α)·EProtBERT

[0018] Among them, E ProtT5 It is the first eigenvector, E ProtBERT It is the second eigenvector, and α is the adaptive weighting coefficient.

[0019] Furthermore, the step of performing probability prediction on the fused feature vector through the classification module specifically includes:

[0020] The normalized fused feature vector is received by the input layer in the classification module and passed to the hidden layer in the classification module to extract the deep nonlinear relationships in the fused feature vector.

[0021] h i =ReLU(W hid,i ·h i-1 +b hid,i )

[0022] Among them, h i and h i-1 Let W be the output feature vectors of the i-th and (i-1)-th hidden layers, respectively, where i ∈ {1, ..., n}, and n is the total number of hidden layers; hid,i b is the weight matrix of the i-th hidden layer; hid,i Let i be the bias vector of the i-th hidden layer;

[0023] The probability values ​​are output through the output layer in the classification module:

[0024] P=σ(W out ·h n +b out )

[0025] Where σ(·) represents the Sigmoid function; h n W is the output feature vector of the nth hidden layer; out b is the weight matrix of the output layer; out is the bias vector of the output layer; P represents the confidence level that the output marine bacteriophage protein sequence to be predicted is an antibacterial protein sequence.

[0026] Furthermore, the construction of the dataset based on the collected antimicrobial protein sequence data specifically includes:

[0027] Sequence data were obtained from known phage genome databases worldwide, including DDBJ, CHVD, EMBL, Genebank, GOV2, GPD, GVD, IGVD, IMG_VR, MGV, PhagesDB, RefSeq, STV, and TemPhD. Marine metagenomic sequencing data were obtained from marine metagenomic databases, including ENA, SRA, and ProGenomes3. Antimicrobial protein sequence data were extracted from positive antimicrobial protein sequence databases, including NCBI, UniprotKB / Swiss-Prot, PDB, RefSeq, EMBL, GenBank, and PIR. The known phage whole genome data, marine metagenomic sequencing data, and antimicrobial protein sequence data were integrated to obtain the original antimicrobial protein sequence dataset.

[0028] The original antimicrobial protein sequence dataset is preprocessed to obtain a processed antimicrobial protein sequence dataset. The preprocessing includes data integrity verification, format standardization, file segmentation, removal of empty sequences, cleaning of abnormal symbols, sequence pruning, deduplication, redundancy removal, quality control, gene annotation, and species classification.

[0029] Different bioinformatics tools were used to identify phage sequences in the processed antimicrobial protein sequence dataset. The identification results were then combined to obtain the final dataset. The bioinformatics tools included PhiSpy, Iphop, virsorter2, DeepVirFinder, VIBRANT, kraken2, viralVerify, Metaphinder, geNomad, virfinder, PPR-Meta, and Phigaro.

[0030] This invention provides a marine bacteriophage antilysin prediction device, comprising:

[0031] The model building module is used to build a deep learning network including a ProtT5 module, a ProtBERT module, a weighted module, and a classification module. It constructs a dataset based on the collected antimicrobial protein sequence data, and uses the dataset to train the constructed deep learning network to obtain a marine bacteriophage antimicrobial lysin prediction model.

[0032] The prediction module inputs the marine phage antilysin sequence to be predicted into the prediction model. The ProtT5 module encodes and decodes the sequence to capture global features, obtaining a first feature vector. The ProtBERT module encodes the sequence to capture local sequence dependencies, obtaining a second feature vector. The weighting module sums the first and second feature vectors to obtain a fused feature vector. The classification module performs probability prediction on the fused feature vector and outputs the confidence level that the predicted marine phage protein sequence is an antimicrobial protein sequence.

[0033] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting marine bacteriophage antimicrobial lysins.

[0034] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method for predicting marine bacteriophage antimicrobial lysins.

[0035] The above-mentioned at least one technical solution adopted in this invention can achieve the following beneficial effects:

[0036] In the marine bacteriophage antimicrobial lysin prediction method provided by this invention, addressing the problems of high false positive rate and limited model optimization methods in traditional algorithms, this invention employs a deep learning framework based on ProtT5 and ProtBERT to extract contextual features from protein sequences. Combined with multi-dimensional verification methods, this effectively improves the model's prediction accuracy and robustness. Specifically: First, the ProtT5 module encodes and decodes the sequence to generate a first feature vector. Based on ProtT5's global feature extraction capabilities, the first feature vector reflects the protein's overall folding pattern, long-range amino acid interactions, and evolutionary information. The ProtBERT module extracts the sequence's local dependencies, obtaining a second feature vector. This second feature vector reflects short-range amino acid interactions, co-occurrence patterns, and local structural stability. The combination of these two feature vectors, through a weighting module, yields a comprehensive fused feature vector. This fused feature vector contains global protein information and local amino acid analysis, enabling high-confidence prediction of whether a marine bacteriophage protein sequence is an antimicrobial protein through a classification module. Compared with traditional algorithms, this invention enhances the model optimization process through multi-dimensional verification, thereby significantly improving the accuracy and robustness of the prediction model and providing more reliable technical support for the research and application of marine bacteriophage antimicrobial proteins. Attached Figure Description

[0037] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0038] Figure 1 A complete workflow diagram provided for this invention;

[0039] Figure 2 A schematic diagram of an encoder and decoder based on the Transformer architecture provided for this invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments in this specification without creative effort are within the scope of protection of this invention.

[0041] Traditional methods for developing lysozyme have the following main problems:

[0042] Limited biological sources: Existing lysozymes are mainly derived from specific bacteria or bacteriophages. Although these sources can provide a certain amount of lysozyme, their scarcity and limitations in biological acquisition significantly restrict large-scale production and widespread application. Meanwhile, with the exponential growth of omics data, the maintenance and updating of existing models and biological protein databases lag behind, making it difficult to efficiently process and analyze novel and complex antimicrobial lysozyme data. This results in the inability to fully explore the potential sequence features of antimicrobial proteins, further hindering the development and utilization of antimicrobial lysozymes.

[0043] Traditional algorithms cannot overcome the high false positive problem: Past methods based on traditional machine learning (such as support vector machines and random forests) were mostly used for antimicrobial protein prediction. However, these algorithms have simple logical frameworks, making them difficult to handle massive amounts of complex data, and the prediction results are often accompanied by a high false positive rate, leading to low robustness and reliability of the models. Among existing artificial intelligence methods, some natural language processing techniques (such as Word2Vec, FastText, and GloVe) have been applied to feature extraction of target protein sequences. However, these methods do not combine the latest natural language processing techniques and deep learning frameworks specifically for protein sequences, limiting prediction accuracy and the exploration of antimicrobial potential.

[0044] The current methods for model optimization and validation suffer from significant shortcomings. The optimization process lacks a systematic approach, typically focusing only on limited parameter adjustments and training on small datasets, making it difficult to comprehensively improve model performance. Furthermore, validation methods are rudimentary and unreliable, relying heavily on low-throughput, single-dimensional experiments, which fail to adequately assess the accuracy of predictions and the actual activity of antibacterial lysozyme. These limitations not only reduce validation efficiency but also hinder the translation of model predictions into practical medical applications and drug development, significantly impeding the progress of lysozyme development.

[0045] The screening and optimization process is inefficient: Currently, most lysozyme screening and optimization rely on traditional laboratory methods, such as clonal screening and directed evolution. These methods are typically time-consuming, labor-intensive, and costly, and cannot quickly mine potential lysozyme sequences from large-scale biological resources, further limiting the development efficiency and widespread application of lysozymes.

[0046] The potential of antimicrobial lysins in clinical treatment has not been fully explored: As an emerging alternative to antibiotics to combat drug resistance, antimicrobial lysins offer advantages such as broad-spectrum antibacterial activity and low production costs. However, the current state of research on antimicrobial lysins is significantly less advanced compared to other antimicrobial proteins (such as antimicrobial peptides) and antimicrobial prediction models. Since their initial discovery in 1971, less than 2% of lysins have been successfully identified and experimentally validated, with only five lysins entering clinical trials.

[0047] Narrow antimicrobial spectrum: Existing first-line antimicrobial proteins (such as antimicrobial peptides) that primarily combat bacterial infections are mostly effective only against specific types of bacteria, resulting in a limited antimicrobial spectrum that struggles to cover a diverse range of pathogens. Traditional screening methods are inefficient at discovering broad-spectrum antimicrobial proteins, limiting their widespread clinical application. In contrast, the antimicrobial lysin developed in this study not only possesses a broader antimicrobial spectrum but also provides effective solutions against different types of drug-resistant pathogens, expanding the scope of application for antimicrobial drug development.

[0048] Incomplete reference databases: Existing databases are mostly based on terrestrial microbial resources, lacking a systematic integration of marine microorganisms and phage lysins, resulting in limited data sources and insufficient samples for training models. This project mines data from global marine microbiomes and phage corpora to establish the first systematic marine antimicrobial lysin database, providing strong support for the comprehensive development of antimicrobial lysins and promoting the in-depth utilization of this untapped treasure trove of marine resources.

[0049] Problems and Cost Control Deficiencies in Computer-Aided Drug Design in Pharmaceutical Companies: Pharmaceutical companies widely adopt computer-aided design (CADD) technology, but existing tools are mainly geared towards small molecule drug development, lacking dedicated prediction methods for macromolecules (such as antimicrobial lysins). These methods suffer from insufficient prediction accuracy and high computational costs. This project combines deep learning with a natural language processing model specifically for protein sequences, significantly improving the prediction accuracy of antimicrobial lysins and reducing computational costs through efficient model optimization, providing a cost-effective new method for drug development in pharmaceutical companies.

[0050] Lack of standardized functional validation system: Existing functional validation of antimicrobial lysins mostly relies on low-throughput experiments, lacking a unified standardized process, making it difficult to compare results. This project established a standardized antimicrobial lysin activity validation system by introducing high-throughput screening technology and multi-dimensional validation methods, significantly improving validation efficiency and result reliability, and providing strong support for the clinical development of lysins.

[0051] Insufficient data integration and interoperability: Existing data are scattered across multiple omics platforms, lacking a unified format and standards, which restricts the development and mining of antimicrobial lysins. This project integrates metagenomic data, protein sequence data, and phage data, and develops a unified data processing and analysis platform, improving the efficiency of data integration and interoperability, and enabling in-depth mining of antimicrobial lysin resources.

[0052] Phage and lysin research is currently limited to specific species: most studies focus on a few phage and lysin sources, neglecting the diverse global microbial resources. This project focuses on mining high-value lysin resources from the untapped ecosystem of the entire ocean, providing a wealth of novel candidates for research in the field.

[0053] Technological barriers limit the participation of small and medium-sized R&D institutions: the development of antimicrobial lysins is technically complex and has high barriers to entry, limiting the participation of small and medium-sized institutions. The artificial intelligence-assisted method developed in this project is suitable for large-scale data processing, while optimizing the algorithm's versatility and ease of operation, lowering the R&D threshold, and promoting diversified innovation in the field.

[0054] To address these existing problems, this invention significantly improves the efficiency and application potential of antimicrobial lysozyme development, providing a novel technological path and solution for addressing the drug resistance crisis. The technical solution involved in this invention belongs to the fields of biomedicine and computational biology, specifically applied to the development of antimicrobial drugs. This technical solution uses artificial intelligence-assisted methods to screen and optimize lysozymes from marine microorganisms (such as bacteriophages), aiming to develop novel antimicrobial lysozymes for the research and development of antimicrobial drugs. This invention can be widely applied in the field of anti-infection, especially in combating drug-resistant bacteria and developing innovative antimicrobial therapies.

[0055] This invention creates an artificial intelligence-based method for the mining and development of antimicrobial lysins. By integrating global marine microbiome and phage data, it systematically solves the problems of data scarcity, insufficient model optimization, and rudimentary verification system in the development of existing antimicrobial lysins.

[0056] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0057] Figure 1 This is a schematic diagram of a method for predicting antimicrobial lysins from marine bacteriophages in this embodiment, specifically including the following steps.

[0058] S1: Data collection and preprocessing.

[0059] Known phage sequences were collected from databases including DDBJ, CHVD, EMBL, Genebank, GOV2, GPD, GVD, IGVD, IMG_VR, MGV, PhagesDB, RefSeq, STV, and TemPhD. These databases cover the genomic information of various phages worldwide, providing data on phage species, genome structure, and functional genes. Secondly, to expand the application scope of the data and cover more marine ecological contexts, marine-related metagenomic sequencing data were collected from databases such as ENA, DDBJ, SRA, and ProGenomes3 to predict potential phage sequences. Finally, positive antimicrobial protein sequences were collected from previously validated literature, NCBI, UniprotKB / Swiss-Prot, PDB, RefSeq, EMBL, GenBank, and PIR. These databases contain a wealth of validated antimicrobial protein information, providing important references for subsequent model construction, feature extraction, and lysin activity prediction.

[0060] The raw data underwent preprocessing, including steps such as verifying data integrity, format standardization, file splitting, removal of empty sequences, cleaning of outliers, sequence trimming, deduplication, redundancy removal, quality control, gene annotation, and species classification to ensure data accuracy and completeness. Data download utilized tools such as Globus, Entrez Direct, and iSeq. Sequencing data cleaning employed tools such as Fastp, Cutadapt, and Trimmomatic; third-generation sequencing data underwent quality control using Chopper; Shasta and wtdbg2 were used for assembling Nanopore and PacBio data, respectively; Unicycler and Flye were used for long read and mixed read assembly, respectively. Sequence data cleaning used Seqkit, quality control was performed using QUAST and alignment with a self-built library; gene annotation used Bakta and Prokka. Other alignment and binning processes utilized traditional bioinformatics tools.

[0061] Tools such as PhiSpy, Iphop, virsorter2, DeepVirFinder, VIBRANT, kraken2, viralVerify, Metaphinder, geNomad, virfinder, PPR-Meta, and Phigaro were used to predict and identify phage sequences from the assembled data, and the identification results were combined into a union.

[0062] S2: Model building.

[0063] The core of this invention is to embed and extract features from protein sequences using deep learning models and natural language processing techniques, and then construct a classifier based on this to efficiently identify and predict antimicrobial lysins from marine bacteriophage data. The specific technical solution and principles are as follows:

[0064] Protein sequence embedding is a crucial step in model building, with the core objective of converting amino acid sequences into high-dimensional vector representations that can be processed by computers. This study employs two protein language models (pLMs) based on the ProtTrans framework: ProtT5-XXL-BFD and ProtBERT-BFD. Leveraging the powerful contextual modeling capabilities of the Transformer architecture and combined with Natural Language Processing (NLP) techniques, they are trained on specialized large-scale protein databases to efficiently extract global and local features from massive protein datasets, providing high-quality embeddings for downstream tasks.

[0065] S201: Advantages of ProtT5 and ProtBERT

[0066] ProtT5 utilizes a bidirectional Transformer architecture to model protein sequences, capturing global contextual information through a multi-head self-attention mechanism to reveal complex semantic relationships between amino acids. This global modeling approach not only resolves long-range dependencies in sequences but also combines local features to accurately characterize protein structure and function. Furthermore, the highly optimized pre-training and fine-tuning strategies within the T5 architecture enhance the model's generalization ability. Trained on massive protein sequence databases like the Big Fantastic Database (BFD), it learns the evolutionary relationships and structural-functional properties of various proteins, enabling it to extract rich feature representations from large-scale protein data. ProtT5 can also handle protein sequences longer than those of typical machine learning models (e.g., 1024 or more), capturing long-range dependencies in sequences through segmented modeling, making it particularly suitable for biological data with long gene sequences, such as bacteriophages. Simultaneously, ProtT5 employs Masked Language Modeling (MLM), randomly masking 15% of the amino acids in the input sequence during training, allowing the model to predict the masked positions based on context. This mechanism simulates real-world variations in protein sequences while incorporating biological characteristics. For example, the model learns the physicochemical properties of amino acids (such as hydrophobicity, polarity, and spatial conformation) and evolutionary background (such as conservation among homologous sequences) through an attention mechanism, thereby improving the biological relevance of sequence embeddings. This mechanism allows ProtT5 to more accurately capture the biological background of protein sequences, thus improving the model's accuracy in capturing protein sequence features. The embedding vector is then infused with positional encoding information, ultimately outputting a reliable protein embedding vector. This is then fed into the Transformer's multi-head attention mechanism, which captures the contextual relationships of amino acids at each position in the protein sequence, understanding how the function of one amino acid in the sequence is influenced by other amino acids.

[0067]

[0068] Here, Q represents the query vector for the current amino acid, used to retrieve contextual information related to its function. It can be understood as a query vector, such as "What contextual information does the current amino acid need to determine its function?" K represents the informational features provided by all positions in the sequence; that is, the bond value at each amino acid position determines how other positions interact with it. For example, an amino acid at an active site may attract other amino acid positions related to its function. V contains specific feature information about the amino acid; the output features are obtained by weighting the interrelationships between the query matrix and the bond matrix. This represents the attention weight, indicating how each amino acid interacts with other amino acids in the sequence. The higher the weight, the stronger the association between the two amino acids. The correlation between each amino acid and other amino acids in the sequence is quantified. This is used to prevent attention weights from becoming too large when the dimension is large, which could lead to gradient instability.

[0069] Ultimately, the context-aware vector output by the multi-head attention mechanism can generate a feature-enhanced representation by calculating global dependencies within the amino acid sequence, used to describe the structure-function relationship of the protein. This is then stacked with a feed-forward network (FFN). The FNN works by performing independent nonlinear transformations on the features at each location, thereby extracting deeper-level features:

[0070] t i = Layer Normalization (z) i +FFN(z i ))

[0071] Here FFN(z) i ) represents the context-aware feature z processed by a feedforward neural network. i Perform nonlinear transformations. The feedforward network itself typically consists of two fully connected layers and a nonlinear activation function (such as ReLU):

[0072] FFN(z i ) = ReLU(W ff1 z i +b ff1 W ff2 +b ff2

[0073] Here, W1 and W2 map the features to high-dimensional and low-dimensional spaces, respectively, and ff1 and ff2 represent the first and second layers of the feedforward network, respectively. (The last part, "W1z," appears to be an incomplete sentence or fragment.) i +b1 performs a linear transformation on the input features, mapping them to a higher-dimensional feature space. Then, the ReLU activation function is used to enhance the model's non-linear modeling capability, capturing complex relationships within the protein sequence. Subsequently, W2 remaps the high-dimensional features back to the original dimensions. This ensures consistency between the input and output feature dimensions, facilitating subsequent residual connections, i.e., z... i +FFN(z i The original features and the transformed features are added together. This preserves the original input information while adding deeper features. Finally, layer normalization is applied to the residual results. By normalizing the mean and variance of the features in each layer, the vanishing gradient problem is alleviated and the training process is stabilized.

[0074] After completing the above steps, the protein sequence encoder is finished. At this point, the embedding of each amino acid position not only reflects its own information but also incorporates information from other positions in the sequence. Next, a decoder is built. Its function is to extract task-specific features from the contextual representation generated by the encoder and generate task outputs, such as sequence classification or sequence labeling, through additional layers. The decoder consists of an encoder-decoder attention layer and an output generation layer. It generates specific outputs based on the contextual features according to task requirements. In this invention, the task is sequence classification, i.e., whether it is an antimicrobial lysin. The decoder generates specific outputs by analyzing the features at all positions. i Calculate global features using average pooling:

[0075]

[0076] Then, classification results are generated using fully connected layers and activation functions:

[0077] P(y=1=σ(W) cls ·T global +b cls )

[0078] The advantage of the encoder-decoder architecture lies in its ability to flexibly adapt the generated contextual representations to protein classification tasks. The decoder, through a specially designed output module, accurately transforms the contextual features generated by the encoder into task-specific outputs, ensuring that the functional properties of the protein are fully expressed. Simultaneously, this architecture utilizes multi-head self-attention mechanisms and multi-layered stacked structures to efficiently capture long-range dependencies and complex patterns in protein sequences, making it particularly suitable for processing long sequence data such as bacteriophage sequences. Furthermore, ProtT5 is trained on the BFD database, which contains over 200 million protein sequences covering a wide range of biological functions and structural information. This significantly improves the model's ability to model sequence complexity and its prediction accuracy.

[0079] Another language model, ProtBERT, is a pre-trained language model based on BERT (Bidirectional Encoder Representations from Transformers), suitable for extracting local and global features from protein sequences. Similar to ProtT5, ProtBERT also uses a Transformer architecture. However, unlike ProtT5, which includes both an encoder and decoder, ProtBERT does not include a decoder (encoders and decoders are not directly related to the previous point about ProtBERT). Figure 2Instead of directly outputting context-aware features of protein sequences through encoder stacking (as shown), ProtT5 focuses on deeply capturing local dependencies of amino acids in protein sequences. Furthermore, although both ProtT5 and ProtBERT are based on MLM, they differ significantly in implementation and characteristics. For ProtT5, MLM is a subtask; its core design is a text-to-text task, and MLM is only one of the pre-training objectives. The model can be extended for generative tasks such as sequence translation, sequence imputation, and reconstructing complete sequences. This design is more suitable for handling complex context modeling tasks, such as genome-level sequence reconstruction, where the ultimate goal is sequence generation. ProtBERT, on the other hand, is a pure MLM model focused on masked language modeling tasks, disregarding context, thus being more efficient at modeling short sequences and local features. For example, for typical antimicrobial lysin sequences (typically in the 20-200 amino acid range), ProtBERT's embedding efficiency is superior to ProtT5. While ProtT5 can handle long sequences (>1024 amino acids), ProtBERT clearly has an advantage in modeling local features of short sequences.

[0080] ProtBERT adopts the BERT architecture, relying solely on the encoder and lacking a decoder. The input sequence must be aligned to a fixed length to avoid dynamic padding impacting computational efficiency. It employs bidirectional attention to calculate global amino acid dependencies, eliminating the need for unidirectional masking. While computationally intensive, this can be mitigated by reducing the number of attention heads or adjusting the dropout rate to prevent overfitting. ProtBERT uses a masking language for modeling, randomly masking 15% of amino acids. During training, masked sites are predicted based solely on context, without sequence reconstruction. Unlike ProtT5, ProtBERT lacks a decoding process. Its MLM training strategy maintains a fixed attention pattern during inference, requiring layer normalization or batch normalization to stabilize feature distribution. Since the ProtBERT structure lacks Encoder-DecoderAttention, it needs to be replaced with bidirectional self-attention computation. A fixed batch size should be maintained during inference to prevent the normalization layer from becoming sensitive to distribution changes. ProtBERT uses pooling layers for feature aggregation, requiring the addition of global average pooling or max pooling layers in classification tasks to adapt to the output structure, rather than generating predictions through a decoder.

[0081] S202: A strategy combining ProtT5 and ProtBERT.

[0082] ProtBERT's embedding features enhance the model's predictive ability for tasks dominated by short sequences or local features, while ProtT5 compensates for the shortcomings in long sequence modeling. This synergy achieves higher accuracy and more comprehensive functional parsing in the antimicrobial lysin prediction task. To effectively combine the embedding features of ProtT5 and ProtBERT, this invention designs a specific fusion strategy and feature extraction process to generate high-quality feature vectors suitable for classification tasks.

[0083] First, a weighted voting method is employed to maximize the use of the prediction results from ProtT5 and ProtBERT. This invention designs a weighted voting mechanism based on validation set performance. Specifically, two models are used independently to predict the same sequence, and weights w1 and w2 (summing to 1) are assigned based on their performance on the validation set and the length of the input data. This method effectively reduces the potential prediction bias of a single model and enhances the robustness of the classification results. The final fused prediction probability is calculated using the following formula:

[0084] P = w1·P ProtT5 +w2·P ProtBERT

[0085] Then, a weighted average strategy is used to integrate the high-dimensional embedding vectors generated by ProtT5 and ProtBERT. The specific calculation formula is as follows:

[0086] E = α·E ProtT5 +(1-α)·E ProtBERT

[0087] E ProtT5 and E ProtBERT These are the embedding vectors output by the two models, with α being the weighting coefficient. α is optimized through cross-validation to ensure the fused features are more biologically meaningful, balancing global context and fine-grained local features. After embedding fusion, to further improve classifier performance, this invention processes and optimizes the generated feature vectors, mainly including the following steps:

[0088] First, feature normalization is performed. Since the embedded features generated by different models may have different numerical ranges, directly inputting them into the classifier may lead to training instability. Therefore, this invention standardizes the fused embedded features using the following formula:

[0089]

[0090] μ and σ represent the mean and standard deviation of the feature, respectively. Normalization eliminates bias in the feature distribution, ensuring faster convergence of the model during gradient descent. However, during embedding, high-dimensional embedded features may contain redundant information, increasing computational complexity and potentially introducing noise. This invention employs Principal Component Analysis (PCA) to reduce the dimensionality of the features, minimizing redundant information while retaining key features. After normalization and dimensionality reduction, the final extracted feature matrix describes the data dimensions and has biological significance. The output features retain key functional regions and evolutionary context of the protein sequence, making it particularly suitable for the prediction and analysis of antimicrobial lysins.

[0091] ProtT5 performs encoding and decoding operations on the marine phage antilysin sequence to capture the global structural and functional features in the marine phage antilysin sequence to be predicted, and obtains the first feature vector; ProtBERT performs encoding operations on the marine phage antilysin sequence to capture the sequence local dependencies in the marine phage antilysin sequence to be predicted, and obtains the second feature vector.

[0092] The global information in the first feature vector defines the overall function of a protein, such as its folding pattern and sequence function. It may include information about transmembrane regions, specific domains of lysozyme, and enzyme activity, and is biased towards defining the characteristics of long-sequence proteins. The local features in the second feature vector mainly describe the interactions within the protein's primary structure, reflecting the co-occurrence patterns of amino acids over short distances, such as the formation trends of certain conserved secondary structures (α-helices, β-sheets), or catalytic active sites and receptor-binding regions between certain amino acids. However, it cannot reflect the spatial functions of a protein (for example, the active site of a protein may be determined by two amino acids that are far apart in sequence but adjacent in three-dimensional structure), mainly focusing on the interactions between amino acids over short distances. In simple terms, if someone wants to learn a new language, the first feature is like learning grammar and sentence structure, while the second feature is like learning phrases and words. To express "apples are delicious," using only the first feature would result in "I ate an apple today" or "I ate something delicious," without knowing which words are key. Using only the second feature would result in "apple," "eat," and "delicious," without stringing them together into a complete sentence. Therefore, for a large protein language model, it is necessary to consider both local and global aspects, namely the first feature and the second feature.

[0093] S3: Classifier Construction

[0094] Based on extracted protein sequence embedding features, this invention integrates global and local information from these features to construct a deep neural network (DNN) classifier for identifying antilysins from phage protein sequences predicted by marine metagenomics. The choice of DNN as the classifier is based on its significant advantages in handling high-dimensional numerical features and modeling complex relationships. Protein feature vectors generated by ProtT5 and ProtBERT embedding models typically have high dimensionality; the fully connected layers of DNN can naturally adapt to this structured numerical data and flexibly model nonlinear relationships in the feature space.

[0095] Compared to Convolutional Neural Networks (CNNs) or Recurrent Neural Networks (RNNs), Direct Neural Networks (DNNs) are better suited to the data characteristics of this research. CNNs are generally suitable for data with spatial local correlations, and in protein applications, they are best suited for small to medium-sized datasets, requiring convolutional kernels to extract local features. Protein embedding features, however, have already captured global semantic representations through pre-trained models, eliminating the need for additional local feature extraction modules. This allows DNNs to avoid unnecessary computational overhead. Furthermore, RNNs often suffer from vanishing or exploding gradients when processing long sequence data, while the fully connected structure of DNNs, combined with normalization (such as batch normalization) and residual connections, is not only more stable during training but also efficiently handles embedding features generated by pre-trained models. DNNs also offer high flexibility, adapting to different task requirements by increasing network depth or adjusting the number of layers. In this invention, the output of each layer of a DNN can be viewed as a nonlinear transformation in the feature space. This hierarchical representation method can fully capture the complex patterns in protein sequences, including global dependencies (such as the interaction between active sites and remote regulatory regions) and local patterns (such as the functional distribution of amino acids around active sites). This characteristic makes DNNs an efficient and compatible classifier choice.

[0096] In summary, this invention employs a DNN classifier and uses deep nonlinear modeling to precisely associate the embedding features of protein sequences with antibacterial activity tags, thereby achieving high-precision antimicrobial lysin prediction. Its main architecture includes the following parts:

[0097] First, the input layer receives the standardized embedded features and the dimension of each sample vector. These embedded features are then passed to the hidden layers while preserving the integrity of the sequence context information. The hidden layers consist of multiple fully connected layers (FC layers) used to extract deep nonlinear relationships from the embedded features. Each layer efficiently models the input features through a combination of linear transformations and nonlinear activation functions. The output of each layer can be represented as:

[0098] hi =ReLU(W hid,i ·h i-1 +b hid,i )

[0099] h i W is the output feature vector of the i-th hidden layer; i b is the weight matrix of the current layer, representing the weight relationship between features, and is initialized with a Gaussian distribution; i This is the bias vector for the current layer. The ReLU activation function maps the input features to a higher-dimensional feature space and introduces non-linear modeling capabilities, avoiding the gradient vanishing problem and thus enhancing the model's ability to express complex protein patterns. Regarding parameter settings, the number of layers is 3-5, and the number of neurons per layer is 512, 256, and 128, decreasing with each layer. Dropout is used to randomly mask lysin features during iterations (the probability is relatively low at 0.1, because the training data is large enough that overfitting in the predicted classification sequences is acceptable), preventing overfitting and avoiding the loss of too many important features during model training. The output layer uses the Sigmoid activation function to predict whether the protein has antimicrobial lysin activity. The output is a probability value representing the confidence that the sequence belongs to the positive class (antimicrobial lysin), defined as follows:

[0100] P=σ(W out ·h n +b out )

[0101] The probability value P output by the Sigmoid function represents the confidence level that the input sequence is an antimicrobial lysin. To improve the reliability of the prediction, this invention sets a dynamic threshold T; the sequence is labeled as an antimicrobial lysin only when P > T. The classifier training uses the cross-entropy loss function to measure the error between the prediction result and the true label (1 represents antimicrobial lysin, 0 represents non-antimicrobial lysin), defined as:

[0102]

[0103] y i P is the true label of the sample. i The probability is the predicted value. To further improve model performance, this invention employs the following optimization strategy during training: using the Adam optimizer (adaptive momentum estimation) combined with a learning rate decay mechanism, i.e., the initial learning rate is set to 0.001, and it decays exponentially to 0.0001 every 10 epochs. Simultaneously, since antimicrobial lysin is a rare sequence in phage protein sequences, undersampling of non-antimicrobial sequences or data augmentation of antimicrobial lysin sequences are used during training to maintain a balanced sample distribution.

[0104] This classification method can reliably identify potential antimicrobial lysins from large-scale marine metagenomic data, providing precise and efficient technical support for the development of antimicrobial drugs.

[0105] S4: To ensure the reliability and applicability of the model prediction, this invention adopts different performance evaluation strategies to verify the model's performance in the antimicrobial lysin prediction task from two perspectives: evaluation indicators and model comparison.

[0106] First, the Matthews Correlation Coefficient (MCC) method is used. This is a balanced metric suitable for imbalanced datasets, where the proportion of positive to negative antimicrobial protein samples is typically much smaller than that of negative samples. In antimicrobial lysin sequence screening, antimicrobial lysins (positive class) constitute a very small portion of the total samples, while non-antimicrobial lysins (negative class) make up the majority. MCC provides a comprehensive performance summary of the classification model by simultaneously considering true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN), avoiding the bias that may result from relying solely on precision or recall. Compared to other metrics (such as accuracy), a significant advantage of MCC is its insensitivity to changes in the ratio of positive to negative samples in the dataset, as shown in the formula below:

[0107]

[0108] in:

[0109] TP: True cases, i.e., the number of samples that were predicted to be positive and were actually positive;

[0110] TN: True negatives, which are the number of samples that are predicted to be negative and are actually negative.

[0111] FP: False positives, which are the number of samples that are predicted to be positive but are actually negative;

[0112] FN: False negatives, which are the number of samples that were predicted to be negative but were actually positive.

[0113] The value of MCC ranges from -1 to 1:

[0114] 1: This indicates that the model's predictions are completely consistent with the true labels (perfect classification).

[0115] 0: Indicates that the model's predictions are comparable to random predictions and have no classification ability.

[0116] 1: This indicates that the model's prediction is completely opposite to the true label (the prediction result is completely wrong).

[0117] Unlike AUC-ROC or PR-AUC, MCC does not rely on a specific threshold. It directly evaluates the entire confusion matrix, avoiding fluctuations caused by threshold selection. It also exhibits high stability; even with highly imbalanced data distribution, MCC can still provide a reliable evaluation of model performance. Compared to traditional evaluation metrics, such as accuracy, which can produce misleading results in imbalanced data, MCC can reflect the bias caused by this imbalance. For example, when 95% of the samples are negative, predicting only the negative class can achieve 95% accuracy, while MCC can reflect the bias caused by this imbalance. Another example is the F1 score, which only focuses on the performance of positive samples, while MCC evaluates the performance of both positive and negative samples.

[0118] In antimicrobial lysin prediction tasks, due to the scarcity of positive samples and the high cost of detection experiments, MCC (Mean Cross-Sectioning) provides a comprehensive evaluation of both positive and negative samples, helping to screen for high-confidence candidate lysins. Furthermore, MCC has significant optimization potential; in balancing sample distribution, it can be achieved through undersampling of negative samples or oversampling of positive samples (e.g., the SMOTE method). Regularization methods can also improve the model's generalization ability. Therefore, it is a particularly suitable evaluation method for protein sequence prediction tasks.

[0119] S5, Antimicrobial lysin prediction

[0120] After the model is completed, it predicts whether a preprocessed and embedded phage protein sequence is an antilysin. The model calculates the antilysin probability of each input sequence using a classifier, with the output ranging between 0 and 1. A threshold T is used to output a binary classification result (whether it is an antilysin). To ensure the accuracy of the prediction, the output undergoes a series of filtering and parameter selection processes. First, the dynamic threshold T defaults to 0.8. The specific value depends on the characteristics of the user's dataset; for example, decreasing T may increase recall, or increasing T may increase precision. Since this invention is suitable for training and prediction on large-scale datasets, it is recommended that this value not be too low. Then, an alignment tool is used to filter redundant sequences to ensure the uniqueness of candidate lysins. If active site data is available, embedding features can be used to label possible functional regions in the sequence, eliminating low-quality sequences lacking functional regions.

[0121] The final output is presented in a structured form, including the following information: sequence ID, sequence content, antimicrobial lysin probability, whether filtering is used, structural features (optional), data source, and similarity comparison results.

[0122]

[0123] S6, Database Construction

[0124] After completing performance evaluation and screening for high-confidence antimicrobial lysins, this invention constructs a comprehensive antimicrobial lysin database platform for storing, managing, and querying all prediction and validation results. The database design is divided into multiple functional modules, including a lysin sequence information table (recording sequence ID, amino acid sequence, and data source), a functional annotation table (recording active sites, structural features, and mechanisms of action), a prediction result table (recording model confidence and classification results), and a validation experimental data table (containing experimentally validated antimicrobial activity and cytotoxicity information). The database uses MySQL or PostgreSQL to store structured data, while MongoDB is introduced for flexible management of large-scale sequences or unstructured files (such as 3D structures). In terms of technical architecture, the backend is built on FastAPI, Django, or Flask, supporting data storage, user authentication, and API interface functions; the frontend uses React or Vue.js to implement an interactive interface, providing user-friendly data retrieval and display functions. Furthermore, Elasticsearch and GraphQL provide full-text search and advanced query capabilities, and D3.js or Plotly are used to complete visualization functions such as antimicrobial activity distribution and confidence analysis. The database is updated dynamically and regularly, adding screening results and experimental data. Deployed on cloud service platforms (such as AWS or Azure), and utilizing Docker and Kubernetes container technologies, it supports distributed computing and high-concurrency access. Simultaneously, the platform seamlessly integrates with AI models, enabling real-time prediction and automatic storage of results after users upload sequences, providing a complete technological chain for the recommendation and research of antimicrobial lysins. Ultimately, the database will serve as a highly efficient, dynamically updated tool, providing researchers, pharmaceutical companies, and AI developers with convenient support for antimicrobial lysin screening, characterization, and model optimization, accelerating the development and translation of novel antimicrobial drugs.

[0125] This invention addresses the limited biological sources in the development of existing antimicrobial lysozymes through artificial intelligence technology. Traditional lysozymes are mostly derived from specific bacteria or bacteriophages, and their scarcity limits large-scale application. This invention integrates global marine microbiome and bacteriophage data to construct the first systematic marine antimicrobial lysozyme database, significantly expanding the potential sources of antimicrobial lysozymes. Simultaneously, this database integrates multi-omics data, providing abundant resources and laying the foundation for the widespread development of lysozymes in the future.

[0126] To address the issues of high false positive rates and limited optimization methods in traditional algorithms, this invention employs deep learning frameworks (such as ProtT5 and ProtBERT) and natural language processing techniques to extract contextual features from protein sequences. Combined with multi-dimensional validation methods, this effectively improves the model's prediction accuracy and robustness. Furthermore, by optimizing the algorithm and introducing high-throughput screening technology, this invention significantly improves the efficiency of antimicrobial lysin screening and validation, solving the problems of long processing times and high costs associated with existing methods, and providing an efficient solution for large-scale screening and development.

[0127] This invention also significantly lowers the technical barriers to the development of antimicrobial lysins. By optimizing the data integration platform and algorithm design, the interoperability of data processing is improved, promoting in-depth exploration of lysin resources. Addressing the limitation of participation by small research institutions, this invention develops a universal method suitable for datasets of different sizes, platforms, and omics, lowering technical barriers and promoting diverse innovation within the field. Furthermore, by simulating diverse marine environments, lysins adapted to specific environments are screened, providing support for their application under extreme conditions.

[0128] In summary, this invention achieves comprehensive innovation in data integration and model prediction. It effectively solves the problems of insufficient resources, low accuracy, and low efficiency in existing lysozyme development, providing a powerful solution to the antibiotic resistance crisis and promoting the development and widespread application of novel antibacterial drugs.

[0129] The above describes a method for predicting marine phage antilysins based on one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for predicting marine phage antilysins, comprising:

[0130] The model building module is used to construct a deep learning network including a ProtT5 module, a ProtBERT module, a weighted module, and a classification module. It constructs a dataset based on the collected marine phage antilysin protein sequence data, and uses the dataset to train the constructed deep learning network to obtain a marine phage antilysin prediction model.

[0131] The prediction module is used to input the protein sequence of the marine bacteriophage antilysin to be predicted into the prediction model. The ProtT5 module encodes and decodes the protein sequence to be predicted to obtain the first feature vector; the ProtBERT module encodes and decodes the protein sequence to be predicted to obtain the second feature vector; the weighting module performs a weighted summation of the first and second feature vectors to obtain a fused feature vector; and the classification module performs probability prediction on the fused feature vector to output the class label of the protein sequence to be predicted.

[0132] Specific limitations regarding the marine phage antilysin prediction device can be found in the limitations of the marine phage antilysin prediction method described above, and will not be repeated here. Each module in the aforementioned marine phage antilysin prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0133] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described method for predicting marine bacteriophage antilysins.

[0134] This invention also provides the structure of a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to implement the marine bacteriophage antilysin prediction method provided above.

[0135] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0136] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for predicting antilysin from marine bacteriophages, characterized in that, include: A deep learning network including a ProtT5 module, a ProtBERT module, a weighted module, and a classification module was constructed. A dataset was built based on the collected antimicrobial protein sequence data. The constructed deep learning network was trained using the dataset to obtain a marine bacteriophage antimicrobial lysin prediction model. The anti-lysin sequence of the marine phage to be predicted is input into the prediction model. The ProtT5 module is used to encode and decode the anti-lysin sequence of the marine phage to be predicted in order to capture the global features in the anti-lysin sequence of the marine phage to be predicted and obtain the first feature vector. The ProtBERT module encodes the antimicrobial lysin sequence of the marine phage to be predicted, capturing the local sequence dependencies in the sequence and obtaining a second feature vector. The first and second feature vectors are weighted and summed by a weighting module to obtain a fused feature vector. The fused feature vector is then used by a classification module to predict the probability and output the confidence that the marine phage protein sequence is an antimicrobial protein sequence. The encoding and decoding operations of the predicted marine bacteriophage antilysin sequence using the ProtT5 module specifically include: In the encoding stage of ProtT5: the position encoding layer in the ProtT5 module encodes the position of the marine bacteriophage antilysin sequence to be predicted, and combines it with protein sequence embedding operations to obtain a protein sequence embedding vector; the multi-head self-attention mechanism layer in the ProtT5 module captures the global dependencies of the protein sequence embedding vector to generate a context-weighted protein sequence vector; the first feedforward neural network layer in the ProtT5 module performs a nonlinear transformation on the weighted protein sequence vector, and combines residual connections and normalization operations to obtain protein-encoded features; In the decoding stage of ProtT5, the protein-coding features are reconstructed using the sequence reconstruction layer in the ProtT5 module, generating a predicted amino acid sequence vector. The amino acid sequence vector is then masked using the masked multi-head self-attention mechanism layer in the ProtT5 module to capture local dependencies between amino acids, resulting in an amino acid mask weighted vector. The global dependency between the protein-coding features and the amino acid mask weighted vector is calculated using the interactive attention mechanism layer in the ProtT5 module, yielding a fused weighted amino acid sequence vector. Finally, the fused weighted amino acid sequence vector undergoes a nonlinear transformation using the second feedforward neural network layer in the ProtT5 module, combined with residual connections and normalization operations, to obtain the first feature vector. The encoding operation of the predicted marine bacteriophage antilysin sequence using the ProtBERT module specifically includes: The position encoding layer in the ProtBERT module aligns the predicted marine bacteriophage antilysin sequence to a preset fixed length and then performs position encoding. Combined with protein sequence embedding operations, a protein sequence embedding vector is obtained. The protein sequence embedding vector is weighted by a multi-head self-attention layer in the ProtBERT module to capture the global dependencies of amino acids and obtain a weighted protein sequence vector. The weighted protein sequence vector is randomly masked using the mask language modeling layer in the ProtBERT module to generate amino acid mask features. The second feature vector is obtained by pooling the amino acid mask features through the pooling layer in the ProtBERT module.

2. The method for predicting marine bacteriophage antilysins as described in claim 1, characterized in that, The step of weighting and summing the first and second feature vectors using a weighting module to obtain a fused feature vector specifically includes: E=α·E ProtT5 +(1-a)·E ProtBERT Among them, E ProtT5 It is the first eigenvector, E ProtBERT It is the second eigenvector, and α is the adaptive weighting coefficient.

3. The method for predicting marine bacteriophage antilysins as described in claim 1, characterized in that, The step of performing probability prediction on the fused feature vector through the classification module specifically includes: The normalized fused feature vector is received by the input layer in the classification module and passed to the hidden layer in the classification module to extract the deep nonlinear relationships in the fused feature vector. h i =ReLU(W hid,i h i-1 +b hid,i ) Among them, h i and h i-1 Let W be the output feature vectors of the i-th and (i-1)-th hidden layers, respectively, where i ∈ {1, ..., n}, and n is the total number of hidden layers; hid ,i b is the weight matrix of the i-th hidden layer; hid ,i Let i be the bias vector of the i-th hidden layer; The probability values ​​are output through the output layer in the classification module: P=σ(W out ·h n +b out ) Where σ(·) represents the Sigmoid function; h n W is the output feature vector of the nth hidden layer; out b is the weight matrix of the output layer; out is the bias vector of the output layer; P represents the confidence level that the output marine bacteriophage protein sequence to be predicted is an antibacterial protein sequence.

4. The method for predicting marine bacteriophage antilysins as described in claim 1, characterized in that, The dataset constructed based on the collected antimicrobial protein sequence data specifically includes: Sequence data were obtained from known phage genome databases worldwide, including DDBJ, CHVD, EMBL, Genebank, GOV2, GPD, GVD, IGVD, IMG_VR, MGV, PhagesDB, RefSeq, STV, and TemPhD. Marine metagenomic sequencing data were obtained from marine metagenomic databases, including ENA, SRA, and ProGenomes3. Antimicrobial protein sequence data were extracted from positive antimicrobial protein sequence databases, including NCBI, UniprotKB / Swiss-Prot, PDB, RefSeq, EMBL, GenBank, and PIR. The known phage whole genome data, marine metagenomic sequencing data, and antimicrobial protein sequence data were integrated to obtain the original antimicrobial protein sequence dataset. The original antimicrobial protein sequence dataset is preprocessed to obtain a processed antimicrobial protein sequence dataset. The preprocessing includes data integrity verification, format standardization, file segmentation, removal of empty sequences, cleaning of abnormal symbols, sequence pruning, deduplication, redundancy removal, quality control, gene annotation, and species classification. Different bioinformatics tools were used to identify phage sequences in the processed antimicrobial protein sequence dataset. The identification results were then combined to obtain the final dataset. The bioinformatics tools included PhiSpy, Iphop, virsorter2, DeepVirFinder, VIBRANT, kraken2, viralVerify, Metaphinder, geNomad, virfinder, PPR-Meta, and Phigaro.

5. A marine bacteriophage antilysin prediction device, characterized in that, The apparatus is used to implement the method according to any one of claims 1-4, the apparatus comprising: The model building module is used to build a deep learning network including a ProtT5 module, a ProtBERT module, a weighted module, and a classification module. It constructs a dataset based on the collected antimicrobial protein sequence data, and uses the dataset to train the constructed deep learning network to obtain a marine bacteriophage antimicrobial lysin prediction model. The prediction module inputs the marine phage antilysin sequence to be predicted into the prediction model. The ProtT5 module encodes and decodes the sequence to capture global features, obtaining a first feature vector. The ProtBERT module encodes the sequence to capture local sequence dependencies, obtaining a second feature vector. The weighting module sums the first and second feature vectors to obtain a fused feature vector. The classification module performs probability prediction on the fused feature vector and outputs the confidence level that the predicted marine phage protein sequence is an antimicrobial protein sequence.

6. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 4.

7. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 4.

Images