A method and system for predicting halophilic proteins based on a hybrid deep learning architecture

Abstract

The application provides a kind of based on hybrid deep learning architecture Halophilic protein prediction method and system, it is related to bioinformatics analysis technical field, the method comprises: obtaining the protein sequence to be predicted;Use protein language model to extract its global semantic embedding feature;At the same time, based on the specific reduction amino acid alphabet, the sequence is re-encoded, and the reduction dipeptide frequency in three interval modes is counted to generate a local composition feature vector;After splicing two kinds of features, input the hybrid deep learning model composed of one-dimensional convolutional neural network and light weight Transformer encoder in series for processing, and finally output the prediction result by full connection layer.The application significantly improves the accuracy, generalization ability and computing efficiency of halophilic protein identification by fusing multi-scale sequence features and hierarchical hybrid network architecture, and provides an effective tool for intelligent mining of large-scale extreme environment enzyme resources.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics analysis, and particularly relates to a halophilic protein prediction method and system based on a hybrid deep learning architecture. BACKGROUND

[0002] In high-salt environments such as salt lakes and oceans, halophilic microorganisms have evolved the special ability to express halophilic proteins, allowing them to maintain structural and functional stability under extreme salinity. These proteins have important application value in industrial biological catalysis and environmental remediation. With the popularization of high-throughput sequencing technology, a large amount of unlabeled protein sequences have accumulated in public databases, which contain a large amount of potential halophilic protein resources. However, traditional wet experimental verification methods are costly and time-consuming, and cannot meet the needs of large-scale screening.

[0003] In recent years, deep learning and protein language models have shown great potential in protein function prediction. In practice, researchers either use language models that can learn contextual representations from large-scale sequence data to obtain global information, or rely on traditional features based on amino acid composition statistics to describe local properties. However, these methods use a single type of feature or model architecture, making it difficult to capture both global semantic information and local physicochemical properties of halophilic proteins, resulting in limitations in prediction accuracy and generalization ability when facing new sequences, which hinders the efficient and accurate identification of halophilic protein resources from massive data. SUMMARY

[0004] To solve the technical problem of the existing prediction methods being unable to balance global semantics and local physicochemical information due to single features, resulting in insufficient prediction accuracy and generalization ability in the background art, the present application provides a halophilic protein prediction method and system based on a hybrid deep learning architecture. The method fuses global semantic features extracted by a protein language model and local composition features generated based on a specific reduction scheme, and constructs a hybrid model combining one-dimensional convolutional neural networks and lightweight Transformer encoders for joint learning, achieving intelligent identification of halophilic proteins with high precision and robustness.

[0005] To achieve the above purpose, a first aspect of the present application provides a halophilic protein prediction method based on a hybrid deep learning architecture, comprising:

[0006] obtaining a protein amino acid sequence to be predicted;

[0007] inputting the protein amino acid sequence into a protein language model to extract global semantic embedding features of the sequence;

[0008] recode the protein amino acid sequence based on a preset reduced amino acid alphabet, and count reduced dipeptide frequencies in three modes of continuity, single interval and double interval to generate a local composition feature vector; wherein the reduced amino acid alphabet divides 20 natural amino acids into multiple functionally equivalent classes;

[0009] concatenate the global semantic embedding feature and the local composition feature vector to obtain a fusion feature vector;

[0010] input the fusion feature vector into a hybrid deep learning model for processing, the hybrid deep learning model comprising a one-dimensional convolutional neural network, a lightweight Transformer encoder and a fully connected layer connected in sequence;

[0011] determine, according to an output of the fully connected layer, whether the protein amino acid sequence is a prediction result of a halophilic protein.

[0012] Further, the reduced amino acid alphabet is "AGNPQST#DE#FWY#ILMV#C#K#R#H", wherein "#" is used to separate different amino acid classes; counting the reduced dipeptide frequency includes: respectively counting the frequencies of reduced dipeptides in three modes of continuity, interval of one residue and interval of two residues in the recoded sequence, and concatenating the frequency vectors in the three modes to generate the local composition feature vector.

[0013] Further, the one-dimensional convolutional neural network comprises at least two convolutional layers and corresponding pooling layers, for extracting local spatial pattern features from the fusion feature vector; the lightweight Transformer encoder comprises a multi-head self-attention mechanism, for modeling the output of the one-dimensional convolutional neural network to capture long-range dependencies in the sequence.

[0014] Further, the convolution operation of the lth layer of the one-dimensional convolutional neural network is defined as follows: let the input feature vector be wherein N= , is the length of the global semantic embedding feature, is the number of classes of the reduced amino acid alphabet; the convolution kernel weight matrix is wherein is the convolution kernel size of the lth layer; the bias term is ; then the output feature value of the lth position is calculated as follows:

[0015] ;

[0016] wherein, ​​The value range of x is , is a ReLU activation function;

[0017] Each convolutional layer is followed by a max-pooling layer, the pooling kernel size of which is 7, the step is 7, and the output of the first pooling unit is calculated as follows:

[0018] ;

[0019] Where [7] represents a set of integers from 1 to 7.

[0020] Further, before the protein amino acid sequence to be predicted is obtained, a model training step is further included, and the training step comprises:

[0021] An initial data set containing halophilic protein and non-halophilic protein sequences is obtained;

[0022] The sequences in the initial data set are screened to reduce sequence similarity and balance the number of positive and negative samples;

[0023] The screened data set is divided into a training set and a test set;

[0024] The training set is used to train the mixed deep learning model, wherein the training process comprises: inputting the protein sequence in the training set into the protein language model to extract global semantic embedding features, generating a local composition feature vector based on the pre-set reduced amino acid alphabet, inputting the two types of features into the mixed deep learning model after splicing into a fusion feature vector, and adjusting the model parameters according to the difference between the model output and the true label until the model converges.

[0025] Further, the determination of the prediction result according to the output of the full connection layer comprises:

[0026] The full connection layer outputs a vector containing multiple class probabilities;

[0027] The class with the maximum probability value in the vector is selected as the final prediction result of the protein amino acid sequence.

[0028] The second aspect of the application provides a halophilic protein prediction system based on a mixed deep learning architecture, comprising:

[0029] A sequence acquisition module for obtaining a protein amino acid sequence to be predicted;

[0030] A global feature extraction module for inputting the protein amino acid sequence into a protein language model to extract global semantic embedding features of the sequence; ​

[0031] a local feature generation module configured to re-encode the protein amino acid sequence based on a preset reduced amino acid alphabet, and count reduced dipeptide frequencies in three modes of continuous, single-interval and double-interval, to generate a local composition feature vector; wherein the reduced amino acid alphabet divides 20 natural amino acids into multiple functionally equivalent classes;

[0032] a feature fusion module configured to concatenate the global semantic embedding feature and the local composition feature vector to obtain a fusion feature vector;

[0033] a deep learning calculation module configured to input the fusion feature vector into a hybrid deep learning model for processing, the hybrid deep learning model comprising a one-dimensional convolutional neural network, a lightweight Transformer encoder and a fully connected layer connected in sequence;

[0034] a classification prediction module configured to determine, according to an output of the fully connected layer, whether the protein amino acid sequence is a prediction result of a halophilic protein.

[0035] A third aspect of the present application provides a halophilic protein prediction device based on a hybrid deep learning architecture, comprising a memory, a processor and a program stored on the memory and running on the processor, wherein the processor implements the steps of the halophilic protein prediction method based on the hybrid deep learning architecture according to the first aspect of the present application.

[0036] A fourth aspect of the present application provides a computer readable storage medium having a program stored thereon, wherein the program is executed by a processor to implement the steps of the halophilic protein prediction method based on the hybrid deep learning architecture according to the first aspect of the present application.

[0037] A fifth aspect of the present application provides a computer program product comprising software code, wherein the program in the software code implements the steps of the halophilic protein prediction method based on the hybrid deep learning architecture according to the first aspect of the present application.

[0038] Compared with the prior art, the halophilic protein prediction method and system based on the hybrid deep learning architecture provided by the present application have the following beneficial effects:

[0039] (1) The application fuses global semantic embedding features and local component feature vectors, and by splicing the global evolutionary semantic information captured by the protein language model and the dipeptide frequency features containing local physicochemical rules (such as charge and hydrophobicity) based on a specific reduction scheme (such as “AGNPQST#DE#FWY#ILMV#C#K#R#H”), a multi-scale and heterogeneous fusion feature representation is constructed. The feature describes the sequence from different dimensions, so that the model can simultaneously perceive the global functional context and the local key mode, thereby significantly improving the representation ability and discrimination accuracy of the model for the complex characteristics of halophilic proteins.

[0040] (2) The application is based on a hybrid deep learning model architecture of a one-dimensional convolutional neural network, a lightweight Transformer encoder and a fully connected layer connected in turn. The one-dimensional convolutional neural network (CNN) is used to efficiently capture the local spatial patterns and short-range correlations (such as the co-occurrence of specific amino acid pairs) between adjacent feature points in the fusion features, then the long-range dependency between the features abstracted by the CNN is modeled through the multi-head self-attention mechanism in the lightweight Transformer encoder, and finally a comprehensive decision is made by the fully connected layer. This hierarchical and well-defined hybrid architecture can fully exploit the local and global information in the fusion features, enhancing the expressiveness and prediction robustness of the model.

[0041] (3) The application reduces 20 amino acids to 8 functionally equivalent classes according to their key physicochemical properties such as hydrophobicity, charge and polarity, and counts the dipeptide frequencies in three modes of continuous, single interval and double interval on the re-encoded sequence. On the one hand, the feature dimension is greatly compressed from 400 dimensions (20x20) to 192 dimensions (8x8x3), reducing the computational complexity and the risk of overfitting; on the other hand, this reduction based on physicochemical properties makes the generated local features more focused on residue interaction patterns closely related to halophilic adaptation (such as the distribution of surface negative charge clusters), filtering out irrelevant noise, thereby improving the discriminability of the features and the interpretability of the model.

[0042] (4) The application defines the convolution and pooling specific operations and parameters, and designs a lightweight feature extraction process. Among them, the setting of the convolution kernel size (such as 5 and 3) aims to balance the receptive field and the amount of calculation to effectively capture local patterns; while the maximum pooling operation with a kernel size of 7 and a step size of 7, while retaining the most significant feature responses, down-samples the feature map with a larger step size, greatly reducing the sequence length and computational load that the subsequent Transformer module needs to process, so that the entire hybrid model has faster training and inference speed while maintaining high performance, and is suitable for large-scale sequence analysis.

[0043] (5) The model training step provided by the application carries out strict screening (including CD-HIT redundancy removal, length filtering, positive and negative sample balancing, etc.) on the initial data set before training, and constructs a high-quality, low-similarity and class-balanced training set and test set. This data preprocessing process provides a data basis for the model to learn the universal salt-tolerant protein discrimination rule, and provides a guarantee for training a model with high accuracy and high generalization ability. BRIEF DESCRIPTION OF DRAWINGS

[0044] The accompanying drawings, which form a part of this disclosure, are intended to provide further understanding of the disclosure and are incorporated herein for illustrative purposes. The schematic embodiments of the disclosure and their descriptions are used to explain the disclosure, and do not constitute improper limitations on the disclosure.

[0045] Figure 1 is a flowchart of the salt-tolerant protein prediction method based on the mixed deep learning architecture provided by the first embodiment of the application;

[0046] Figure 2 is a schematic diagram of the key evaluation index performance of the model provided by the first embodiment of the application;

[0047] Figure 3 is a schematic diagram of the transverse performance comparison provided by the first embodiment of the application;

[0048] Figure 4 is a schematic diagram of the explainability analysis result provided by the first embodiment of the application;

[0049] Figure 5 is a schematic diagram of the salt-tolerant protein prediction system based on the mixed deep learning architecture provided by the second embodiment of the application;

[0050] Figure 6 is a structural schematic diagram of the salt-tolerant protein prediction device based on the mixed deep learning architecture provided by the third embodiment of the application. DETAILED DESCRIPTION

[0051] It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the application. Unless otherwise specified, all technical and scientific terms used in the application have the same meaning as generally understood by those skilled in the art to which the application belongs.

[0052] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0053] The embodiments in the present application and the features in the embodiments can be combined with each other in the case of no conflict.

[0054] All data acquisition of the embodiments is based on the compliance with laws and regulations and user consent, and the legal application of data.

[0055] Embodiment one

[0056] As Figure 1 The embodiment provides a halophilic protein prediction method based on a hybrid deep learning architecture, which comprises the following steps:

[0057] Acquiring a protein amino acid sequence to be predicted;

[0058] Inputting the protein amino acid sequence into a protein language model to extract global semantic embedding features of the sequence;

[0059] Re-encoding the protein amino acid sequence based on a preset reduced amino acid alphabet, and counting reduced dipeptide frequencies in three modes of continuity, single interval and double interval to generate a local composition feature vector; wherein the reduced amino acid alphabet divides 20 natural amino acids into multiple functionally equivalent classes;

[0060] Splicing the global semantic embedding features and the local composition feature vector to obtain a fusion feature vector;

[0061] Inputting the fusion feature vector into a hybrid deep learning model for processing, wherein the hybrid deep learning model comprises a one-dimensional convolutional neural network, a lightweight Transformer encoder and a fully connected layer connected in sequence;

[0062] According to the output of the fully connected layer, determining whether the protein amino acid sequence is a halophilic protein prediction result.

[0063] Specifically, the reduced amino acid alphabet is "AGNPQST#DE#FWY#ILMV#C#K#R#H", where "#" is used to separate different amino acid categories; the reduced dipeptide frequency is counted by counting the frequencies of consecutive, one-residue-interval and two-residue-interval reduced dipeptides in the recoded sequence respectively, and the frequency vectors in the three modes are spliced to generate the local composition feature vector.

[0064] The application proposes a reduced amino acid alphabet based on multi-dimensional physicochemical properties, which divides 20 natural amino acids into 8 functionally equivalent categories: AGNPQST, DE, FWY, ILMV, C, K, R, and H. This grouping is not a random aggregation, but a comprehensive consideration of key physicochemical parameters such as hydrophobicity, polarity, charge, aromaticity, thiol group, and pKa value, and is directed to the optimization of the structure and function characteristics of halophilic proteins. Specifically as follows:

[0065] AGNPQST group: covering 7 amino acids such as glycine (G), alanine (A), and serine (S), with moderate hydrophobicity and moderate polarity, representing common flexible hydrophilic residues on the surface of proteins, suitable for modeling non-specific solvent exposed regions;

[0066] DE group: containing aspartic acid (D) and glutamic acid (E), both carrying negative charge and having the highest polarity, directly corresponding to the acid residue cluster enriched on the surface of halophilic proteins, which is a key feature to maintain protein stability in high salt environment;

[0067] FWY group: phenylalanine (F), tryptophan (W), and tyrosine (Y) share aromatic ring structure and moderate hydrophobicity, which are used to capture π-π stacking and hydrophobic core formation tendency;

[0068] ILMV group: isoleucine (I), leucine (L), methionine (M), and valine (V) constitute typical aliphatic side chains with high hydrophobicity and low polarity, reflecting the conservation of the hydrophobic core inside the halophilic protein;

[0069] C group: cysteine (C) can form disulfide bonds due to the presence of thiol group, which is crucial for structural rigidity in extreme environments;

[0070] K / R / H group: representing high basic small volume (K, Charge=+1), high basic large volume (R, Volume=124), and near neutral pKa (H, pKa≈6.0) basic residues, respectively, which are independently grouped to distinguish their differentiated roles in salt bridge formation and protonation state.

[0071] Through the above grouping strategy, the original 20-dimensional amino acid encoding is compressed into an 8-dimensional functionally equivalent representation, which significantly reduces the dimension of the feature space while retaining the key physicochemical signals of halophilic proteins.

[0072] In particular, in one embodiment, the protein sequences are first re-encoded using the reduction scheme "AGNPQST#DE#FWY#ILMV#C#K#R#H". In this scheme, the amino acids separated by "#" are re-encoded into a uniform amino acid symbol: for example, all aspartic acid (D) and glutamic acid (E) residues in the sequence are re-encoded as D after reduction.

[0073] Let be the original sequence of a halophilic protein, then the re-encoded sequence can be computed as follows:

[0074] ;

[0075] For each sequence, all possible dipeptide patterns are enumerated. Assuming the size of the amino acid alphabet is , three types of dipeptide patterns are considered:

[0076] 1. contiguous (two amino acids separated by 0, denoted as A1A2);

[0077] 2. single-interval pair (two amino acids separated by 1, denoted as A1XA2, where X represents any amino acid);

[0078] 3. double-interval pair (two amino acids separated by 2, A1XXA2, where X represents any amino acid).

[0079] For each pattern, the sequence is encoded into a frequency vector of dimension n 2 , representing all possible ordered pairs of re-encoded residues. The final feature representation is obtained by concatenating the three vectors, resulting in a feature vector of length 3 n 2 . Finally, for each sequence, the frequencies of each re-encoded dipeptide pair are normalized to obtain a fixed-length feature vector representing the dipeptide composition bias in the context of halophilic adaptation.

[0080] The statistics of the reduced dipeptide frequencies include the frequencies of the reduced dipeptides separated by 0 (contiguous), 1 (one residue apart), and 2 (two residues apart) in the re-encoded sequence, respectively. These three separation patterns can capture residue associations at different spatial distances and more comprehensively describe the local sequence environment. Finally, the frequency vectors under the three patterns are concatenated to generate the final local composition feature vector.

[0081] Specifically, the one-dimensional convolutional neural network comprises at least two convolutional layers and corresponding pooling layers, which are used to extract local spatial pattern features from the fusion feature vector.

[0082] A lightweight convolutional neural network is constructed by using two layers of stacked one-dimensional convolutional layers. The convolutional kernel size used in the first layer is 5, and the convolutional kernel used in the second layer is 3. The ReLU function is used as the activation function in each convolution operation.

[0083] The convolutional layer uses a small-size convolutional kernel (such as a size of 5 in the first layer and a size of 3 in the second layer) to perform sliding calculation on the fusion feature vector. Through weight sharing and local connection, the convolutional layer automatically learns and extracts the local spatial patterns (for example, the combination rule of the residues of specific rationalized attributes in a local window of the sequence) that are discriminative in the features. The ReLU activation function connected after each convolutional layer introduces nonlinearity. The subsequent max-pooling layer (pooling kernel 7, step 7) down-samples the convolutional output, retains the most significant feature response in each local region, and gradually reduces the data size, thereby reducing the model complexity and enhancing the translation invariance of the features. This design enables the CNN to efficiently capture the key local signals related to the halophilic characteristics from the high-dimensional fusion features.

[0084] The lightweight Transformer encoder comprises a multi-head self-attention mechanism, which is used to model the output of the one-dimensional convolutional neural network to capture long-range dependencies in the sequence.

[0085] In one specific embodiment, a lightweight Transformer model is constructed by using 2 layers of stacked encoding layers. The model dimension of each encoding layer is set to 8, and a 4-head self-attention mechanism is used. This model reduces the computational complexity by reducing the model dimension, and enhances the information mining ability for different feature subspaces by using the multi-head attention mechanism. Therefore, the model can be used in low-computing scenarios while ensuring the lightweight model.

[0086] Each attention head is calculated as follows:

[0087] ;

[0088] wherein, , , represent the query, key and value matrices generated by the input value. In addition, matrices are the learnable weight matrices corresponding to , , in each attention head, which map the input matrix from dimensions to and The dimension supports parallel computation of attention scores across multiple feature subspaces. Additionally, .

[0089] The output of each attention head is:

[0090] ;

[0091] The outputs of each attention head are concatenated, and the calculation expression is as follows:

[0092] ;

[0093] Finally, the concatenated output is weighted by the weight matrix. The final output is obtained after linear transformation. .

[0094] The feature sequence extracted and pooled by the CNN is used as input. The self-attention mechanism allows any two positions in the sequence (even if they are far apart) to interact directly and calculate the association weights between them. Through a multi-head design, the model can learn dependencies in parallel in different representation subspaces (such as focusing on charge patterns, hydrophobic patterns, etc.). This enables the Transformer to effectively model long-range dependencies between features abstracted by the CNN, integrate global contextual information, and compensate for the limited receptive field of the CNN. The lightweight design (such as fewer layers and smaller model dimensions) ensures overall computational efficiency.

[0095] Specifically, the convolution operation of the l-th layer of the one-dimensional convolutional neural network is defined as follows: Let the input feature vector be... ,in N= , The length of the global semantic embedding feature. The number of categories in the reduced amino acid alphabet; the convolution kernel weight matrix is... ,in For the first Layer convolution kernel size; bias term is Then the first Output feature value at each position The calculation is as follows:

[0096] ;

[0097] in, The range of values ​​is , It is the ReLU activation function;

[0098] Each convolutional layer is followed by a max-pooling layer, the max-pooling layer having a kernel size of 7 and a stride of 7. Output of a pooling unit The calculation is as follows:

[0099] ;

[0100] wherein [7] represents a set of integers from 1 to 7.

[0101] The convolution network uses small size convolution kernels and large step pooling operations based on lightweight design, guarantees the extraction of feature ability while compressing the number of parameters to one fifth of the traditional convolution network, and is suitable for scenes with limited computing resources.

[0102] In the present application, the mixed expression feature of the protein sequence integrates the key biological characteristics of halophilic proteins. It is known that halophilic proteins are rich in negatively charged amino acid residues on their surface, and among the 20 common amino acids, only aspartic acid (Asp, D) and glutamic acid (Glu, E) have negative electric properties. The simplified amino acid alphabet used in the present application classifies D and E into the same category (denoted as "#DE"), abstractly models the typical physicochemical characteristics of the protein surface from the perspective of charge properties, and this information is encoded in the reduced dipeptide frequency feature. In addition, the reduced dipeptide frequency feature further integrates the commonalities of amino acids in multiple key physicochemical dimensions such as polarity, hydrophobicity, aromaticity, volume, and thiol group, thereby systematically retaining the structure and function signals closely related to halophilic adaptation. Further, through multiple sequence alignment analysis of the halophilic protein family, it is found that there is a high frequency of homotypic substitution between D and E at the conserved sites, indicating that they have high interchangeability in function and structure. The ESM-2 protein language model learns this evolutionarily conserved pattern during pre-training, and the global semantic embedding vector output by the model implicitly contains the context-dependent relationship of D / E homotypic substitution. Through the feature fusion module, the above two types of information, i.e., the global semantic feature provided by ESM-2 and the local reduced dipeptide frequency feature based on the simplified alphabet statistics, are spliced to form a mixed expression feature, which is used as the input of the convolutional neural network. The network contains two one-dimensional convolution layers with convolution kernels of lengths 5 and 3, respectively: in the embodiment, the first five reduced dipeptide frequency , , , , For example, when using a one-dimensional convolution kernel of size 5 [ , , , , ] for convolution operation, the relevant mathematical operation is:

[0103] ;

[0104] , , , , The information of polarity, charge and the like is carried in the amino acid grouping, so the result of the convolutional neural network operation can also capture the spatial patterns of the protein surface negative charge distribution, polarity distribution, hydrophobicity and the like. The convolution operation on the global semantic part of the fusion feature can effectively capture the homologous substitution clues of D / E residues in the local sequence window; the max-pooling operation (pooling kernel size is 7, and step is 7) can inhibit low-significance features by retaining the maximum value of each channel response, so as to highlight the most relevant evolutionary conservation signals and charge enrichment regions. Therefore, the convolutional neural network of the application can efficiently mine the discriminative structural features of halophilic proteins in the two key dimensions of evolutionary conservation and surface charge distribution, and significantly improve the prediction accuracy and biological interpretability.

[0105] Specifically, before the protein amino acid sequence to be predicted is acquired, a model training step is further included, and the training step comprises:

[0106] An initial data set containing halophilic protein and non-halophilic protein sequences is acquired;

[0107] The sequences in the initial data set are screened to reduce sequence similarity and balance the number of positive and negative samples;

[0108] Specifically, the screening of the sequences in the initial data set comprises three steps, wherein the first screening operation is to count the number of positive sample data and negative sample data, if the number of one type of data exceeds 1.5 times the number of the other type of data, it is considered that the two types of data are imbalanced, in this case, stratified sampling is used to select the class with more number until the number of sequences of the two classes is consistent; the second screening operation is to use CD-HIT software to remove sequences with too high sequence similarity in the two classes by using a certain percentage value as a threshold, generally, the percentage threshold can be a value in the range of 40% to 60%; the third screening operation is to count the length of the protein sequence, only select the protein sequence with a length meeting a certain range, and obtain a data set with smaller sequence length dispersion.

[0109] The screened data set is divided into a training set and a test set;

[0110] The training set is used to train the hybrid deep learning model, wherein the training process comprises: inputting the protein sequence in the training set into the protein language model to extract global semantic embedding features, generating a local composition feature vector based on the pre-set reduced amino acid alphabet, inputting the two types of features into the hybrid deep learning model after splicing into a fusion feature vector for processing, and adjusting the model parameters according to the difference between the model output and the real label until the model converges.

[0111] The protein language model is specifically an ESM-2 model, which is a large language model developed by Meta AI focusing on protein analysis. The pre-trained ESM-2 model can capture the evolutionary model, structural features and functional information of proteins from protein sequences. Specifically, N sequences are input into the ESM-2 model as a batch, and the ESM-2 model uses a mask language modeling objective function to output a feature embedding representation corresponding to each protein sequence after mask processing of the amino acid residues. These feature embedding representations include global semantic features and local sequence information of the protein. The embedding vectors are subjected to average pooling operation, and a global semantic feature vector of a fixed dimension (1280 dimensions) can be extracted.

[0112] The protein global semantic embedding from the ESM-2 language model and the three-mode dipeptide frequency features based on the reduced alphabet "AGNPQST#DE#FWY#ILMV#C#K#R#H" are combined to construct a multi-scale hybrid feature representation system.

[0113] Specifically, let the embedding feature of ESM-2 be , where is the sequence length. Let the reduced dipeptide feature be . The fused feature vector is defined as:

[0114] ;

[0115] where || represents element-level concatenation along the feature dimension.

[0116] The embedding feature of the ESM-2 model is 1280-dimensional, the reduced dipeptide feature is 192-dimensional, and after the fusion operation, a 1472-dimensional dataset is obtained.

[0117] The above two types of features are integrated into a high-dimensional hybrid vector through feature dimension concatenation, and L2 normalization is used to eliminate differences in different feature scales, and finally a composite feature space with global biological semantics and local structure characteristics is formed.

[0118] The hybrid numerical feature can:

[0119] Enhance the model's ability to recognize rare salt-loving protein patterns: the evolutionary signal of ESM-2 can compensate for the feature sparsity under small sample data, and the dipeptide feature can strengthen the explicit modeling of known salt-loving mechanisms (such as surface charge enrichment and specific residue co-occurrence);

[0120] Biological interpretability of prediction results: visually link prediction results with key evolutionary conserved sites or structural preference patterns in the sequence;

[0121] Reduction of homology bias: the generalization ability of global semantic embedding and the robustness of local dipeptide features work together to effectively alleviate the overfitting problem caused by the clustering of homologous sequences.

[0122] Specifically, the determination of the prediction result according to the output of the fully connected layer comprises:

[0123] The fully connected layer outputs a vector containing multiple class probabilities;

[0124] Select the class with the maximum probability value in the vector as the final prediction result of the protein amino acid sequence.

[0125] Specifically, a weight matrix of size 1280x4 is used to perform linear transformation on the output of the Transformer network. The features output by the Transformer network are flattened into a 1280-dimensional vector, which is used as the input of the current network layer. After linear transformation, a 4-dimensional log probability vector is generated. The components in this vector correspond to the four target classes. Determine the maximum component in the log probability vector. The sequence class corresponding to this component is the prediction result of the current sequence by the model. Among them, if the index of the maximum component value in the vector is 0, the current protein sequence is predicted to be a non-halophilic protein; if the maximum value is 1, the current protein sequence is predicted by the model to be a halophilic protein; if the maximum value component index is 2 and 3, it means that the current protein sequence is an unknown protein, and the current model has insufficient prediction ability. A class is randomly generated from "halophilic protein" and "non-halophilic protein" as the prediction result of the model for the protein sequence.

[0126] The embodiment further comprises the steps of experimentally verifying and performance evaluating the halophilic protein prediction model, specifically as follows:

[0127] Constructing a benchmark dataset: Collect halophilic protein sequences and non-halophilic protein sequences using database retrieval or publicly available databases, divide them into training set, test set according to certain proportion, and pre-screen the sequences;

[0128] Setting up a comparison model: Select the current mainstream method as the baseline, including: a classifier using only ESM-2 embedding, a SVM model using only traditional dipeptide frequency features, a publicly available halophilic protein prediction tool HPClas, and a Blast method.

[0129] Evaluation metrics tools: The following metrics are calculated on the independent test set: Accuracy (ACC), Sensitivity (Sn), Specificity (Sp), Matthews Correlation Coefficient (MCC), and the harmonic mean F1 of precision.

[0130] Ablation experiment design: Remove reduced dipeptide features (retain only ESM-2 embeddings), remove convolutional neural network modules, and use unreduced dipeptide frequency features to quantify the contribution of each module to the final performance.

[0131] Results analysis:

[0132] Experiments demonstrate that the classification accuracy (ACC) of this invention reaches 93.0% on the independent test set, significantly outperforming the contrastive model using only ESM-2 global semantic features (ACC = 88.5%) and the baseline model relying solely on reduced dipeptide frequency features (ACC = 80.5%). These results indicate that this invention, by fusing multi-source heterogeneous features, exhibits a significant advantage in predictive performance.

[0133] Further ablation experiments showed:

[0134] After removing the reduced dipeptide frequency feature, the model accuracy decreased by 4.85 percentage points (from 93.0% to 88.15%).

[0135] After removing the convolutional neural network module, the accuracy dropped by approximately 4.3 percentage points (to 88.7%).

[0136] The results validate the key role of multi-scale feature fusion mechanism and hierarchical spatial pattern modeling strategy in improving the model's discriminative ability.

[0137] Furthermore, this invention employs an 8-class simplified amino acid alphabet (AGNPQST#DE#FWY#ILMV#C#K#R#H) based on physicochemical properties, compressing the traditional dipeptide frequency feature dimension from 400 dimensions (20×20) to 64 dimensions (8×8), a reduction of 84% in feature dimension, significantly improving inference efficiency. Notably, under the same model architecture, the accuracy obtained using the simplified alphabet is 93.0%, while the dipeptide frequency feature calculated using the original 20-alphabet alphabet only achieves 40.5% accuracy—an absolute improvement of 52.5 percentage points and a relative improvement of approximately 1.3 times. This fully demonstrates that the simplified scheme proposed in this invention not only does not lose key biological information but also enhances the discriminative power of the features through functionally equivalent clustering.

[0138] Please seeFigure 2 In addition to the accuracy, the model also performs well in other key evaluation indicators:

[0139] The Matthews correlation coefficient (MCC) is 0.86,

[0140] The sensitivity (Sn) is 0.92,

[0141] The specificity (Sp) is 0.94,

[0142] The F1 score (F1) is 0.929.

[0143] See Figure 3 In the comparison of horizontal performance, the present application and the existing public salt-tolerant protein prediction method perform as follows in the classification evaluation indicators on the independent test set:

[0144] The sensitivity (Sn), specificity (Sp), accuracy (ACC), Matthews correlation coefficient (MCC), and F1 score (F1) of the disclosed tool HPClas are 0.880, 0.810, 0.845, 0.692, and 0.850, respectively;

[0145] The BLAST method based on sequence homology alignment is 0.590, 0.700, 0.645, 0.292, and 0.624, respectively, in the same indicators.

[0146] In comparison, the present application achieves better results in the above five indicators. The MCC improvement rate is 24.3% (vs. HPClas) and 194.5% (vs. BLAST), respectively, indicating that the present application has significantly stronger discriminant stability and generalization ability in the class imbalance scenario. The experimental results fully prove that the present application has substantial progress and leading advantage in comprehensive classification performance compared with the current mainstream salt-tolerant protein prediction method.

[0147] The above comprehensive indicators show that the present application has strong negative sample exclusion ability and overall classification stability while maintaining high recall ability, and is suitable for high-precision identification of salt-tolerant proteins in actual scenarios.

[0148] The present embodiment also includes the step of performing structural analysis on the salt-tolerant protein prediction model to reveal the biological basis of model decision-making and verify its correlation with the physicochemical properties of salt-tolerant proteins, as follows:

[0149] (1) Key feature identification:

[0150] The SHAP (SHapley Additive exPlanations) model and other model explanation methods are used to calculate the contribution of each input feature to the prediction result; the top-N features with the highest contribution are selected as the Top-N feature set based on the absolute value of the contribution in descending order.

[0151] (2) Local site positioning of global semantic features:

[0152] If the Top-N features include the global semantic feature output by the ESM-2 model (such as the kth embedding vector), the corresponding key local site of the protein sequence is determined by the attention weight analysis method. Specifically:

[0153] First, the ESM-2 model is used to encode the target protein sequence, and the self-attention weight matrix output by the last layer of the Transformer module is extracted;

[0154] For the kth semantic feature, the average attention weight of each amino acid residue position on all attention heads is calculated;

[0155] The top M residue sites with the highest average attention weight are selected as the key local sites associated with the ESM-2 semantic feature.

[0156] (3) Local site mapping of reduced dipeptide features:

[0157] If the Top-N features include the reduced dipeptide frequency feature (such as "D*D", "DA", or "D*A"), according to the composition of the amino acid types, the 8-class simplified alphabet (AGNPQST#DE#FWY#ILMV#C#K#R#H) used, and the amino acid spacing information, all dipeptide instances that meet the pattern are backtracked and located in the original protein sequence, and the corresponding position interval is marked as a key local site.

[0158] (4) Structure and physicochemical property verification:

[0159] Based on the above positioning results, a three-dimensional structure model of the target protein is generated, and the visualization analysis of hydrophobicity and charge distribution is superimposed on the key local sites.

[0160] Please refer to Figure 4 In the explainability analysis of the training data set in the present application, it is found that:

[0161] ESM-Feat836 (i.e., the 836th dimension of the ESM-2 embedding vector) is one of the top ten key features in terms of contribution;

[0162] The reduced dipeptide features D*D, DA, and D*A are also among the top-10 high-contribution features.

[0163] Furthermore, representative samples were selected from both the halophilic and non-halophilic protein sets for comparative analysis:

[0164] Halophilic proteins: UniProt ID AUV80949 and EPX85750;

[0165] Non-salamiophilic proteins: UniProt ID ALG69294 and OQY49972.

[0166] exist Figure 4 In the 3D structure shading scheme shown:

[0167] Green indicates areas with strong hydrophobicity.

[0168] White indicates the hydrophobic / hydrophilic neutral region.

[0169] Yellow indicates areas with weak hydrophobicity.

[0170] The key local sites identified by the above method are highlighted in red and purple.

[0171] The analysis results show that:

[0172] For ESM-Feat836, key local sites in halophilic proteins mainly exhibit strong hydrophobicity (green), while in non-halophilic proteins it exhibits a uniform mixture of strong and weak hydrophobicity.

[0173] For D*D and DA features, the number of key local sites identified in halophilic proteins is significantly greater than that in non-halophilic proteins, reflecting the typical feature of halophilic proteins being enriched with negatively charged amino acids (D / E) on their surface.

[0174] Regarding the D*A characteristic, the proportion of weakly hydrophobic amino acids at key sites of halophilic proteins is lower than that of non-halophilic proteins, further confirming the functional adaptation of halophilic proteins in terms of surface charge and hydrophobic distribution.

[0175] In summary, this embodiment not only verifies that the multi-source features (ESM-2 semantic features and reduced dipeptide frequencies) fused in this invention have clear biological significance, but also proves that this invention can effectively capture the discriminative structural features of halophilic proteins in terms of surface charge enrichment, hydrophobic core conservation, and residue co-occurrence patterns under the deep learning framework, thereby providing interpretable and verifiable technical support for high-precision halophilic protein prediction.

[0176] Example 2

[0177] like Figure 5 This embodiment provides a halophilic protein prediction system based on a hybrid deep learning architecture, including:

[0178] a sequence acquisition module configured to acquire a protein amino acid sequence to be predicted;

[0179] a global feature extraction module configured to input the protein amino acid sequence into a protein language model and extract global semantic embedding features of the sequence;

[0180] a local feature generation module configured to re-encode the protein amino acid sequence based on a preset reduced amino acid alphabet, and count reduced dipeptide frequencies in three modes of continuity, single interval and double interval to generate a local composition feature vector; wherein the reduced amino acid alphabet divides 20 natural amino acids into multiple functionally equivalent classes;

[0181] a feature fusion module configured to concatenate the global semantic embedding features and the local composition feature vector to obtain a fusion feature vector;

[0182] a deep learning calculation module configured to input the fusion feature vector into a hybrid deep learning model for processing, the hybrid deep learning model comprising a one-dimensional convolutional neural network, a lightweight Transformer encoder and a fully connected layer connected in sequence;

[0183] a classification prediction module configured to determine, according to an output of the fully connected layer, whether the protein amino acid sequence is a prediction result of a salt-tolerant protein.

[0184] In one specific embodiment, the prediction system further comprises a sequence preprocessing module configured to perform resampling, sequence similarity calculation and sequence length statistics, etc. on the input sequence to obtain a data set with smaller similarity and more balanced sequence quantity. The global feature extraction module and the local feature generation module are further configured to: only load an ESM-2 model instance through a central processing unit or through a central processing unit + graphics processing unit in parallel to realize feature extraction of batch sequences, and simultaneously utilize CUDA threads to optimize vectorization calculation of dipeptide frequency statistics. The feature fusion module is further configured to: concatenate the global semantic features and the reduced dipeptide frequency features in front and back to obtain an expanded one-dimensional fusion feature data set. The dimension of the one-dimensional fusion feature data set is the sum of the dimensions of the global semantic features and the reduced dipeptide frequency feature vector. The deep learning calculation module is further configured to: input the one-dimensional fusion feature data set into a convolutional neural network to capture hierarchical spatial patterns with biological significance in the protein sequence; input the output of the convolutional neural network into a Transformer model to model long-range interactions in amino acid residues and reduced dipeptides through a multi-head attention mechanism, thereby effectively mining potential information in local features and global semantics; and realize high-speed data transmission between the central processing unit and the graphics processing unit through a PCIe bus to ensure real-time synchronization of intermediate results of the convolutional layer and the Transformer layer in the video memory.

[0185] Embodiment three

[0186] Embodiment three of the present application provides a halophilic protein prediction device based on a hybrid deep learning architecture, comprising a memory, a processor and a program stored in the memory and running on the processor, wherein the processor implements the steps of the halophilic protein prediction method based on the hybrid deep learning architecture as described in embodiment one of the present application.

[0187] The detailed steps are the same as the halophilic protein prediction method based on the hybrid deep learning architecture provided in embodiment one, which will not be repeated here.

[0188] In one specific embodiment, as Figure 6 The prediction device includes one or more central processing units 10, one or more graphics processing units 11, a main memory 12 and a secondary memory 13. The secondary memory 13 is used to persistently store the ESM-2 model weight file, the reduced dipeptide feature calculation program, the trained hybrid deep learning model parameters and the protein sequence data to be predicted; the main memory is used to temporarily cache the output data of the sequence preprocessing module, the global semantic features and local dipeptide features of the global feature extraction module, the splicing results of the feature fusion module and the intermediate calculation results of the deep learning calculation module; the central processing unit is used to perform the logic control of the sequence preprocessing module (including resampling algorithm, sequence similarity calculation and length statistics), coordinate the data flow of each module, and schedule the programs and data in the secondary memory through the main memory; the graphics processing unit is used to accelerate the convolutional neural network operation (including two layers of one-dimensional convolution and maximum pooling operation) of the deep learning calculation module and the multi-head self-attention mechanism calculation of the Transformer model, and efficiently process the hierarchical spatial patterns and long-range interaction modeling of protein sequences through parallel computing capability.

[0189] Embodiment four

[0190] Embodiment four of the present application provides a computer readable storage medium having a program stored thereon, wherein the program is executed by a processor to implement the steps of the halophilic protein prediction method based on the hybrid deep learning architecture as described in embodiment one of the present application.

[0191] The detailed steps are the same as the halophilic protein prediction method based on the hybrid deep learning architecture provided in embodiment one, which will not be repeated here.

[0192] The medium storing program codes (auxiliary storage) includes: a mechanical hard disk, a solid state disk, an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), an optical storage device, a magnetic storage device, a flash memory, a magnetic disk or an optical disk, and / or a combination of the above devices. The main storage medium is a random access memory (SROM).

[0193] Embodiment five

[0194] Embodiment five of the present application provides a computer program product comprising software codes, wherein the programs in the software codes perform the steps in the method for predicting halophilic proteins based on a hybrid deep learning architecture as described in embodiment one of the present application.

[0195] The detailed steps are the same as the method for predicting halophilic proteins based on a hybrid deep learning architecture provided in embodiment one, which will not be repeated here.

[0196] The computer product is stored on a data carrier and is designed to perform the method for predicting halophilic proteins based on a hybrid deep learning architecture and a protein language model as described above. Therefore, the computer product according to the present application has the advantages described in detail with reference to the device of the present application. The computer program product can be coded as computer executable instruction codes and can be executed by a suitable programming language such as Python, etc. In addition, the computer program product can be provided on a network (such as the Internet), and Internet users can download the computer program product through the network (such as the Internet) when needed. The computer program product can be implemented by means of software or one or more special electronic circuits (i.e. hardware), or in the form of a combination of software and hardware.

[0197] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes. The solutions in the embodiments of the present application can be implemented in various computer languages, such as object-oriented programming languages Java and interpreted scripting language JavaScript, etc.

[0198] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart Figure 1 one or more functions specified in the flowchart Figure 1 one or more functions specified in the flowchart

[0199] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart Figure 1 one or more functions specified in the flowchart Figure 1 one or more functions specified in the flowchart

[0200] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart Figure 1 one or more functions specified in the flowchart Figure 1 one or more functions specified in the flowchart

[0201] The above only describes the preferred embodiments of the present application and is not used to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting halophilic proteins based on a hybrid deep learning architecture, characterized in that, include: Obtain the amino acid sequence of the protein to be predicted; The protein amino acid sequence is input into a protein language model to extract global semantic embedding features of the sequence; The protein amino acid sequence is recoded based on a preset reduced amino acid alphabet, and the frequencies of reduced dipeptides in three modes—continuous, single-spaced, and double-spaced—are counted to generate a local composition feature vector. The reduced amino acid alphabet divides 20 natural amino acids into multiple functionally equivalent classes. The reduced amino acid alphabet is "AGNPQST#DE#FWY#ILMV#C#K#R#H", where "#" is used to separate different amino acid categories. The count of reduced dipeptide frequencies includes: counting the frequencies of reduced dipeptides occurring consecutively, spaced by one residue, and spaced by two residues in the recoded sequence, and concatenating the frequency vectors from the three modes to generate the local composition feature vector. The global semantic embedding features are concatenated with the local component feature vectors to obtain a fused feature vector; The fused feature vector is input into a hybrid deep learning model for processing. This model comprises a one-dimensional convolutional neural network, a lightweight Transformer encoder, and a fully connected layer connected sequentially. The one-dimensional convolutional neural network includes at least two convolutional layers and corresponding pooling layers, used to extract local spatial pattern features from the fused feature vector. The lightweight Transformer encoder includes a multi-head self-attention mechanism to model the output of the one-dimensional convolutional neural network to capture long-range dependencies in the sequence. The l-th layer convolution operation of the one-dimensional convolutional neural network is defined as follows: Let the input feature vector be... ,in N= , The length of the global semantic embedding feature. The number of categories in the reduced amino acid alphabet; the convolution kernel weight matrix is... ,in For the first Layer convolution kernel size; bias term is Then the first Output feature value at each position The calculation is as follows: ； in, The range of values ​​is , It is the ReLU activation function; Each convolutional layer is followed by a max-pooling layer, the max-pooling layer having a kernel size of 7 and a stride of 7. The output of each pooling unit The calculation is as follows: ； Where, [7] represents the set of integers from 1 to 7; Based on the output of the fully connected layer, determine whether the predicted amino acid sequence of the protein is a halophilic protein.

2. The method as described in claim 1, characterized in that, Before obtaining the amino acid sequence of the protein to be predicted, a model training step is also included, which includes: Obtain an initial dataset containing sequences of halophilic and non-halophilic proteins; The sequences in the initial dataset are filtered to reduce sequence similarity and balance the number of positive and negative samples; The filtered dataset is divided into training and test sets; The training set is used to train the hybrid deep learning model, wherein the training process includes: inputting the protein sequences in the training set into the protein language model to extract global semantic embedding features, generating local composition feature vectors based on the preset reduced amino acid alphabet, concatenating the two types of features into a fusion feature vector and inputting it into the hybrid deep learning model for processing, and adjusting the model parameters according to the difference between the model output and the real label until the model converges.

3. The method as described in claim 1, characterized in that, The step of determining the prediction result based on the output of the fully connected layer includes: The fully connected layer outputs a vector containing probabilities for multiple classes; The category with the highest probability value in the vector is selected as the final prediction result of the protein amino acid sequence.

4. A halophilic protein prediction system based on a hybrid deep learning architecture, characterized in that, include: The sequence acquisition module is used to obtain the amino acid sequence of the protein to be predicted. A global feature extraction module is used to input the protein amino acid sequence into a protein language model and extract the global semantic embedding features of the sequence. The local feature generation module is used to recode the protein amino acid sequence based on a preset reduced amino acid alphabet, and to count the frequencies of reduced dipeptides in three modes: continuous, single-spaced, and double-spaced, to generate a local composition feature vector. The reduced amino acid alphabet divides 20 natural amino acids into multiple functionally equivalent classes. The reduced amino acid alphabet is "AGNPQST#DE#FWY#ILMV#C#K#R#H", where "#" is used to separate different amino acid categories. Counting the frequencies of the reduced dipeptides includes: counting the frequencies of reduced dipeptides occurring in the recoded sequence in continuous, single-residue-spaced, and double-residue-spaced modes, respectively, and concatenating the frequency vectors from the three modes to generate the local composition feature vector. The feature fusion module is used to concatenate the global semantic embedding features with the local component feature vectors to obtain a fused feature vector; A deep learning computation module is used to input the fused feature vector into a hybrid deep learning model for processing. The hybrid deep learning model includes a one-dimensional convolutional neural network, a lightweight Transformer encoder, and a fully connected layer connected sequentially. The one-dimensional convolutional neural network includes at least two convolutional layers and corresponding pooling layers, used to extract local spatial pattern features from the fused feature vector. The lightweight Transformer encoder includes a multi-head self-attention mechanism to model the output of the one-dimensional convolutional neural network to capture long-range dependencies in the sequence. The l-th layer convolution operation of the one-dimensional convolutional neural network is defined as follows: Let the input feature vector be... ,in N= , The length of the global semantic embedding feature. The number of categories in the reduced amino acid alphabet; the convolution kernel weight matrix is... ,in For the first Layer convolution kernel size; bias term is Then the first Output feature value at each position The calculation is as follows: ； in, The range of values ​​is , It is the ReLU activation function; Each convolutional layer is followed by a max-pooling layer, the max-pooling layer having a kernel size of 7 and a stride of 7. The output of each pooling unit The calculation is as follows: ； Where, [7] represents the set of integers from 1 to 7; The classification prediction module is used to determine whether the protein amino acid sequence is a halophilic protein based on the output of the fully connected layer.

5. A device for predicting halophilic proteins based on a hybrid deep learning architecture, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of the halophilic protein prediction method based on a hybrid deep learning architecture as described in any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the halophilic protein prediction method based on a hybrid deep learning architecture as described in any one of claims 1 to 4.

7. A computer program product, comprising software code, characterized in that, The program in the software code performs the steps of the halophilic protein prediction method based on a hybrid deep learning architecture as described in any one of claims 1 to 4.

Images