6-hb targeting membrane fusion inhibiting peptide prediction method, apparatus, device, and medium

Abstract

The application relates to the technical field of deep learning, and discloses a 6-HB targeting membrane fusion inhibition peptide prediction method, device, equipment and medium, which comprises the following steps: generating a 6-HB targeting membrane fusion inhibition candidate peptide; inputting the membrane fusion inhibition candidate peptide into a pre-constructed 6-HB targeting membrane fusion inhibition peptide prediction model for classification prediction to obtain the classification result of the membrane fusion inhibition candidate peptide; the membrane fusion inhibition peptide prediction model is a two-stage transfer learning classification model, the first stage is based on sequence general features to perform binary classification prediction of antiviral peptides and non-antiviral peptides; the second stage is combined with spatial conformation features to perform binary classification prediction of membrane fusion inhibition peptides and non-membrane fusion inhibition antiviral peptides; and the classification result is screened to obtain the 6-HB targeting membrane fusion inhibition candidate peptide. The method realizes the leap from coarse-grained antiviral activity recognition to fine-grained 6-HB inhibition activity classification, and improves the prediction accuracy of the 6-HB inhibition peptide specific to the envelope virus type / subtype.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a method, apparatus, device, and medium for predicting 6-HB-targeted membrane fusion inhibitory peptides. Background Technology

[0002] Enveloped viruses are viral types with an external lipid membrane, such as influenza viruses, HIV, and coronaviruses. The infection mechanism of enveloped viruses is highly dependent on the membrane fusion process mediated by viral envelope proteins. The core structural basis of this process is the "six-helix bundle" (6-HB) fusion core formed by the activated viral fusion protein. This structure consists of a stable coiled-helix structure formed by the antiparallel binding of the trimer HR1 (a heptapeptide repeat domain in the envelope protein) core and the HR2 (also a heptapeptide repeat domain in the envelope protein) peptide. This process releases energy, driving virus-cell membrane fusion. Fusion inhibitory peptides targeting 6-HB (such as enfuviride) block the formation of native 6-HB by competitively binding to the HR1 hydrophobic groove, mimicking the HR2 domain, thus becoming an important strategy in antiviral drug design.

[0003] However, traditional antiviral peptide recognition models (such as those based on sequence similarity, physicochemical characteristics, or shallow machine learning) mainly rely on linear sequence information, while the inhibitory activity of 6-HB is highly dependent on the three-dimensional spatial fit between the peptide and the target. Current methods struggle to effectively characterize the conformational matching properties of peptide-target interactions (such as spatial topology, local curvature, dihedral angles, etc.), leading to high false positive rates and an inability to predict subtype-specific inhibitory activity. Summary of the Invention

[0004] In view of this, the present invention provides a method, apparatus, device and medium for predicting 6-HB-targeting membrane fusion inhibitory peptides, so as to efficiently and accurately predict 6-HB-targeting membrane fusion inhibitory candidate peptides.

[0005] In a first aspect, the present invention provides a method for predicting 6-HB-targeted membrane fusion inhibitory peptides, the method comprising:

[0006] Generate membrane fusion inhibitory candidate peptides targeting 6-HB;

[0007] Candidate peptides for membrane fusion inhibition are input into a pre-constructed 6-HB-targeted membrane fusion inhibition peptide prediction model for classification and prediction, and the classification results of the candidate peptides for membrane fusion inhibition are obtained. The 6-HB-targeted membrane fusion inhibition peptide prediction model is a two-stage transfer learning classification model. The first stage is based on the general sequence features to perform binary classification prediction of antiviral peptides and non-antiviral peptides. The second stage incorporates spatial conformation features to perform binary classification prediction of membrane fusion inhibition peptides and non-membrane fusion inhibition antiviral peptides.

[0008] The screening and classification results yielded 6-HB-targeting membrane fusion inhibitory candidate peptides.

[0009] The 6-HB-targeting membrane fusion inhibitory peptide prediction method provided by this invention is an integrated computational framework that combines general sequence features, spatial conformation features, transfer learning enhancement, and a generation-prediction-screening closed loop. It achieves a leap from coarse-grained antiviral activity identification to fine-grained 6-HB inhibitory activity classification, and drives the targeted generation and high-precision screening of functional peptide sequences. This accelerates the development of subtype-specific inhibitory peptides targeting viral fusion cores and improves the prediction accuracy of enveloped virus type / subtype-specific 6-HB inhibitory peptides.

[0010] In one alternative implementation, the 6-HB-targeting membrane fusion inhibitory peptide prediction model is established through the following steps:

[0011] Obtain the first dataset, which includes antiviral peptides and non-antiviral peptides, wherein antiviral peptides are used as the positive dataset and non-antiviral peptides are used as the negative dataset.

[0012] Construct the first initial multi-layer graph convolutional network model;

[0013] Based on the first dataset, a first initial multi-layer graph convolutional network model is trained to obtain a general antiviral peptide activity prediction model after training.

[0014] Obtain a second dataset, which includes 6-HB-targeting membrane fusion inhibitory peptides and non-6-HB-targeting antiviral peptides. The 6-HB-targeting membrane fusion inhibitory peptides are used as the positive dataset, and the non-6-HB-targeting antiviral peptides are used as the negative dataset.

[0015] Construct a second initial multi-layer graph convolutional network model with the same architecture as the first initial multi-layer graph convolutional network model;

[0016] The weights of the general antiviral peptide activity prediction model are frozen and used as a sequence feature extractor. The extracted general sequence features are then fused into the node features of the second initial multi-layer graph convolutional network architecture model.

[0017] Based on the second dataset, the general sequence features extracted by the general antiviral peptide activity prediction model are concatenated with the pre-calculated spatial conformation features to train the second initial multilayer graph convolutional network model and obtain the 6-HB targeting membrane fusion inhibitory peptide prediction model.

[0018] In this embodiment, based on a feature transfer strategy, a first-stage general antiviral peptide activity prediction model is first trained. The weights of the first-stage model are then frozen as a feature extractor. The extracted general sequence features are concatenated with spatial conformation features to form an enhanced feature embedding. A second-stage model is then trained to predict 6-HB-targeting membrane fusion inhibitory peptides, enabling the second-stage model to more accurately identify the key structural information of 6-HB-targeting membrane fusion inhibitory peptides. This effectively utilizes the general features of antiviral peptides while combining the spatial conformation features of the target peptide, significantly improving the accuracy and efficiency of 6-HB-targeting membrane fusion inhibitory peptide prediction. Furthermore, feature transfer and model freezing reduce the number of parameters requiring training, effectively suppressing overfitting, significantly improving model generalization performance, and shortening training time, making it suitable for large-scale antiviral peptide screening and evaluation.

[0019] In one alternative implementation, a first initial multilayer graph convolutional network model is trained based on a first dataset, including:

[0020] Perform data preprocessing on the first dataset;

[0021] Based on the preprocessed first dataset, the general sequence features are calculated, including amino acid type coding features, physicochemical property features, and evolutionary semantic features.

[0022] Based on the preprocessed first dataset, a binary adjacency matrix is ​​constructed using the k-neighborhood method based on sequence distance.

[0023] Using the general characteristics of the sequence as node features and the inverse sequence distance as edge weights, the first graph structure is constructed.

[0024] The first graph structure is input into the first initial multi-layer graph convolutional network model, and a general antiviral peptide activity prediction model is obtained when the preset training indicators are met.

[0025] In this embodiment, a general antiviral peptide activity prediction model is constructed by integrating data preprocessing, sequence feature extraction, and graph convolutional network techniques. The obtained general antiviral peptide activity prediction model can effectively capture global and local features from peptide sequence maps. During subsequent learning transfer, after freezing all learnable parameters, a stable node-level feature embedding representation can be output, thereby effectively enhancing the downstream 6-HB-targeting membrane fusion inhibitory peptide prediction model's ability to express complex relationships in the sequence and improving prediction accuracy.

[0026] In one optional implementation, based on the second dataset, the general sequence features extracted by the general antiviral peptide activity prediction model are concatenated with pre-computed spatial conformation features to train the second initial multilayer graph convolutional network model, including:

[0027] Perform data preprocessing on the second dataset;

[0028] Based on the preprocessed second dataset, spatial configuration features are extracted, including: Rotation and translation invariant coordinates Euclidean distance Normalized inverse Euclidean distance Discrete curvature, Pseudo-dihedral angle, conformation-related AAindx features, and residue solvent-accessible surface area-related features;

[0029] based on Euclidean distance is used to generate a binary adjacency matrix;

[0030] The general sequence features are concatenated along the feature dimension and input into the general antiviral peptide activity prediction model with frozen weights to obtain a high-dimensional sequence feature representation.

[0031] The high-dimensional sequence feature representation and spatial configuration feature are concatenated along the feature dimension to serve as enhanced node features; and the normalized inverse Euclidean distance is used as the edge weight to construct the second graph structure.

[0032] By inputting the second graph structure into the second initial multi-layer graph convolutional network model after feature transfer, a 6-HB targeted membrane fusion inhibitory peptide prediction model is obtained when the preset training indicators are met.

[0033] In this embodiment, by combining the general sequence features of antiviral peptides extracted by the first-stage pre-trained model with the pre-calculated spatial conformation features, the interaction between candidate peptides and target virus 6-HB can be captured more comprehensively at the sequence and conformation levels, thereby effectively improving the second-stage model's ability to identify 6-HB-targeted membrane fusion inhibitory peptides.

[0034] In one alternative implementation, a membrane fusion inhibitory candidate peptide targeting 6-HB is generated, comprising:

[0035] Obtain antiviral peptide data;

[0036] A general antiviral peptide generation model is obtained based on a pre-built model architecture trained on antiviral peptide data.

[0037] Obtain data on 6-HB targeted membrane fusion inhibitory peptide;

[0038] A general antiviral peptide generation model was fine-tuned based on 6-HB-targeted membrane fusion inhibitory peptide data to obtain a 6-HB-targeted membrane fusion inhibitory peptide generation model; the 6-HB-targeted membrane fusion inhibitory peptide generation model was used to generate membrane fusion inhibitory candidate peptides.

[0039] In this embodiment, a two-stage transfer learning framework is employed. The first stage pre-trains the model on an antiviral peptide dataset to obtain a general antiviral peptide sequence generation capability. The second stage fine-tunes all parameters on a membrane fusion inhibitory peptide dataset to train a 6-HB-targeted membrane fusion inhibitory peptide generation model. This effectively improves the generalization ability of the final 6-HB-targeted membrane fusion inhibitory peptide generation model, enabling it to retain general sequence generation capabilities while accurately capturing key features related to membrane fusion when generating 6-HB-targeted membrane fusion inhibitory peptide sequences.

[0040] In one optional implementation, the 6-HB-targeted membrane fusion inhibitory peptide generation model includes: a variable-length sequence preprocessing module for preprocessing the input sequence; the preprocessing of the input sequence includes: sorting the input sequence in descending order according to its actual length; padding the ends of the sorted sequence to a uniform length and generating a corresponding padding mask to mark the valid data position;

[0041] The embedding layer, connected to the output of the variable-length sequence preprocessing module, is used to map the preprocessed integer index sequence into a dense vector representation in a high-dimensional space.

[0042] The LSTM layer, connected to the output of the embedding layer, is used to capture long-range dependencies and contextual information in the preprocessed vector sequence;

[0043] The fully connected layer, the output of the LSTM layer is ultimately mapped to the probability distribution of the output space by the fully connected layer;

[0044] The 6-HB targeted membrane fusion inhibitory peptide generation model adopted a dynamic training control strategy during the training phase, which combined gradient norm constraints and continuous validation loss monitoring.

[0045] The 6-HB targeted membrane fusion inhibitory peptide generation model is equipped with a generation sequence filtering unit during the sequence generation stage. The generation sequence filtering unit performs the following three operations: adjusting the output probability distribution of the Softmax function using a temperature parameter to control generation diversity; dynamically filtering non-amino acid terms to ensure that each generation step only samples from the valid amino acid dictionary; and responding to the end symbol, immediately interrupting the generation process once a special symbol representing the end of the sequence is generated.

[0046] The 6-HB targeted membrane fusion inhibitor peptide generation model constructed in this embodiment can handle variable-length sequence inputs. It reduces invalid computation by using padding masks to mark valid data locations. The dynamic training control module improves the model's generalization performance, training stability, and convergence speed through gradient norm constraints and validation loss monitoring. The generated sequence filtering unit controls the diversity of generated sequences by adjusting the output probability distribution of the Softmax function using temperature parameters. It can also effectively filter non-amino acid terms and respond to end-of-sequence marker interrupts to ensure the reliability of generated sequences. The embedding layer maps the preprocessed integer-indexed sequences to a dense vector representation in a high-dimensional space. The LSTM layer captures long-range dependencies and contextual information in the sequence and maps them to a probability distribution in the output space through a fully connected layer, thereby generating membrane fusion inhibitor peptide sequences with targeting and functional potential.

[0047] In one optional implementation, the classification results are screened to obtain the final 6-HB-targeting membrane fusion inhibitory candidate peptides, including:

[0048] The initial candidate peptides in the classification results were identified as candidate peptides classified as 6-HB-targeting membrane fusion inhibitory peptides.

[0049] Experimental molecular docking was performed on each initial candidate peptide with the 6-HB conformation of the target virus to obtain a docking score corresponding to each initial candidate peptide.

[0050] Determine whether the docking score meets the preset conditions;

[0051] Candidate peptides that meet the preset docking scoring criteria are identified as the final 6-HB-targeted membrane fusion inhibition candidate peptides.

[0052] In this embodiment, by performing experimental docking on each initial candidate peptide and further screening based on the docking score, the success rate of downstream biological validation experiments can be effectively improved.

[0053] In a second aspect, the present invention provides a 6-HB-targeting membrane fusion inhibitory peptide prediction device, the device comprising:

[0054] A sequence generation module is used to generate membrane fusion inhibitory candidate peptides targeting 6-HB;

[0055] The prediction module is used to input membrane fusion inhibition candidate peptides into a pre-constructed 6-HB-targeted membrane fusion inhibition peptide prediction model for classification and prediction, and obtain the classification results of the membrane fusion inhibition candidate peptides. The 6-HB-targeted membrane fusion inhibition peptide prediction model is a two-stage transfer learning classification model. The first stage is based on sequence general features to perform binary classification prediction of antiviral peptides and non-antiviral peptides. The second stage incorporates spatial conformation features to perform binary classification prediction of membrane fusion inhibition peptides and non-membrane fusion inhibition antiviral peptides.

[0056] The screening module is used to screen the classification results and obtain the final 6-HB-targeting membrane fusion inhibitory candidate peptides.

[0057] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the 6-HB-targeted membrane fusion inhibitory peptide prediction method of the first aspect or any corresponding embodiment described above.

[0058] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to perform the 6-HB-targeting membrane fusion inhibitory peptide prediction method of the first aspect or any corresponding embodiment thereof.

[0059] It should be noted that the computer equipment and computer-readable storage medium of the 6-HB targeted membrane fusion inhibitory peptide prediction device provided by this invention correspond to the 6-HB targeted membrane fusion inhibitory peptide prediction method described above. Therefore, for the beneficial effects of the 6-HB targeted membrane fusion inhibitory peptide prediction device, computer equipment, and computer-readable storage medium, please refer to the description of the corresponding beneficial effects of the 6-HB targeted membrane fusion inhibitory peptide prediction method above, and will not be repeated here. Attached Figure Description

[0060] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0061] Figure 1 This is a flowchart illustrating the method for predicting and screening 6-HB-targeted membrane fusion inhibitory peptides according to an embodiment of the present invention.

[0062] Figure 2 This is a schematic diagram of the transfer learning between the 6-HB targeted membrane fusion inhibitory peptide generation model and the prediction model according to an embodiment of the present invention.

[0063] Figure 3 This is a schematic diagram comparing the performance of a general antiviral peptide prediction model according to an embodiment of the present invention with the state-of-the-art model for the same task.

[0064] Figure 4 This is a graph showing the results of ablation experiments using a 6-HB targeted membrane fusion inhibitory peptide prediction model according to an embodiment of the present invention.

[0065] Figure 5This is a graph showing the results of ablation experiments using a 6-HB targeted membrane fusion inhibitory peptide generation model according to an embodiment of the present invention;

[0066] Figure 6 This is a schematic diagram of the 6-HB targeted membrane fusion inhibitory peptide generation model architecture according to an embodiment of the present invention;

[0067] Figure 7 This is a flowchart of sequence preprocessing according to an embodiment of the present invention;

[0068] Figure 8 This is a structural block diagram of a 6-HB-targeted membrane fusion inhibitory peptide prediction device according to an embodiment of the present invention;

[0069] Figure 9 This is a schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Currently, traditional antiviral peptide development technologies face three major bottlenecks: 1. Neglect of spatial conformation dependence: Traditional methods struggle to effectively characterize the conformational matching of peptide-target interactions, leading to high false positive rates and an inability to predict subtype-specific inhibitory activity. 2. Insufficient model generalization due to sample scarcity: The number of experimentally validated membrane fusion inhibitory peptides is extremely limited (especially for specific viral subtypes). Under small sample conditions, conventional machine learning classifiers are prone to overfitting, making it difficult to extract discriminative features from sparse data. Furthermore, the cross-viral genus 6-HB conformational diversity further increases the modeling difficulty, and general antiviral peptide prediction models generally have classification accuracy for membrane fusion inhibition function below acceptable thresholds. 3. Lack of functionality in targeted generation: Existing peptide sequence generation models (such as those based on RNN, VAE, or GPT architectures) primarily focus on sequence probability distribution learning, lacking physicochemical constraints on target biological functions (such as 6-HB binding energy and steric hindrance effects). The generated results often exhibit "ineffective designs" with high sequence novelty but lacking binding activity, failing to efficiently produce functional fusion inhibitory peptides with viral subtype specificity.

[0072] In view of this, according to an embodiment of the present invention, a method for predicting 6-HB-targeted membrane fusion inhibitory peptides is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0073] This embodiment provides a method for predicting 6-HB-targeted membrane fusion inhibitory peptides, specifically a method for predicting membrane fusion inhibitory peptides targeting the six-helix bundle (6-HB) of an enveloped virus fusion core. This method can be executed by devices such as servers, terminals, and mobile terminals. Figure 1 This is a flowchart of a method for predicting 6-HB-targeted membrane fusion inhibitory peptides according to an embodiment of the present invention, as follows: Figure 1 As shown, the process includes the following steps:

[0074] Step S101: Generate 6-HB-targeting membrane fusion inhibitory candidate peptide.

[0075] 6-HB (six-helix bundle) targeting membrane fusion inhibitory peptides are a class of peptide molecules that can specifically interfere with the membrane fusion process mediated by viral envelope proteins (such as HIV gp41 and SARS-CoV-2 S protein). Their mechanism of action lies in targeting the key transitional structure formed by viral fusion proteins during membrane fusion—the six-helix bundle (6-HB). These inhibitory peptides competitively bind to pre-fusion intermediates (such as the HR1 trimer) by mimicking functional regions of viral proteins (such as HR2), thereby preventing endogenous HR1-HR2 interactions, effectively inhibiting the proper formation of 6-HB, and ultimately halting the fusion process between the virus and the host cell membrane, thus inhibiting viral infection.

[0076] In this embodiment, the 6-HB-targeting membrane fusion inhibitory candidate peptide was generated using a pre-constructed 6-HB-targeting membrane fusion inhibitory peptide generation model. The 6-HB-targeting membrane fusion inhibitory peptide generation model was obtained by training an LSTM autoregressive model using a two-stage transfer learning strategy, which will be described in detail below. Of course, the 6-HB-targeting membrane fusion inhibitory candidate peptide in this embodiment can also be extracted from known protein sequences or from sequence libraries; no particular limitation is made here.

[0077] Step S102: Input the membrane fusion inhibition candidate peptides into the pre-constructed 6-HB-targeted membrane fusion inhibition peptide prediction model for classification and prediction, and obtain the classification results of the membrane fusion inhibition candidate peptides. The 6-HB-targeted membrane fusion inhibition peptide prediction model is a two-stage transfer learning classification model, which is trained based on sequence general features and spatial conformation features. The first stage is based on sequence general features to perform binary classification prediction of antiviral peptides and non-antiviral peptides. The second stage incorporates spatial conformation features to perform binary classification prediction of membrane fusion inhibition peptides and non-membrane fusion inhibition antiviral peptides.

[0078] In this embodiment, the pre-constructed 6-HB-targeting membrane fusion inhibitory peptide prediction model is a classification model built based on two-stage transfer learning. Specifically, refer to... Figure 2 As shown, the first stage constructs a graph neural network (GNN) based on a k-contact graph of antiviral peptide sequences. This GNN uses residue type encoding extracted through static feature engineering (which can be one-hot, BLOSUM 62, etc., without specific limitations), ESM2 (Evolutionary Scale Modeling 2) feature embeddings, and physicochemically related amino acid indices as node features, and normalized inverse sequence distance as edge weights. The prediction task of the first stage is binary classification of antiviral peptides and non-antiviral peptides. The second stage constructs a graph neural network based on the k-contact graph of antiviral peptide sequences. carbon atom ( A GNN for a binary contact map using Euclidean distance, where static feature engineering is extracted. Spatial geometric features such as spatial coordinates, pseudo-dihedral angles, and discrete curvature are used as node features. Simultaneously, the general sequence feature representation extracted from the first-stage pre-trained model with frozen parameters is concatenated to form an enhanced node embedding vector. Normalized inverse spatial distance is used as the edge weight. The second-stage prediction task is to classify 6-HB-targeting membrane fusion inhibitory peptides into binary categories of non-6-HB-targeting antiviral peptides.

[0079] In practical use, the identified candidate peptides for 6-HB membrane fusion inhibition can be directly input into the 6-HB-targeting membrane fusion inhibition peptide prediction model to predict whether the candidate peptide is a 6-HB-targeting membrane fusion inhibition peptide.

[0080] Step S103: Screen the classification results to obtain the final 6-HB-targeted membrane fusion inhibitory candidate peptide.

[0081] For example, candidate peptides with a predicted label of 1 (positive) can be selected for protein-peptide flexible docking with the pre-fusion conformation of influenza A H3N2 subtype 6-HB to obtain a docking score. At least several experimentally validated 6-HB-targeting membrane fusion inhibitory peptides for influenza A H3N2 were collected from DRAVP (Antiviral Peptide Database) as positive controls. These peptides were then docked with the target virus 6-HB under identical methods and parameter settings to obtain scores. Then, based on the candidate peptide docking score... docking score with positive control peptide By comparing statistical relationships, candidate peptides with docking affinity superior to the statistical standard of positive control peptides were screened out as highly specific inhibitory peptide candidates for the target virus 6-HB.

[0082] The 6-HB-targeting membrane fusion inhibitory peptide prediction method provided in this embodiment is an integrated computational framework that combines general sequence features, spatial conformation features, transfer learning enhancement, and a generation-prediction-screening closed loop. It achieves a leap from coarse-grained antiviral activity identification to fine-grained 6-HB targeting prediction, and drives the directional generation and high-precision screening of functional peptide sequences, thereby accelerating the development of subtype-specific inhibitory peptides targeting viral fusion cores and improving the prediction accuracy of enveloped virus type / subtype-specific 6-HB inhibitory peptides.

[0083] In some optional implementations, the 6-HB-targeting membrane fusion inhibitory peptide prediction model is established through the following steps:

[0084] Step a1: Obtain the first dataset, which includes antiviral peptides and non-antiviral peptides, wherein antiviral peptides are used as the positive dataset and non-antiviral peptides are used as the negative dataset.

[0085] The following example illustrates this. The positive dataset, i.e., antiviral peptides, labeled 1, consists of 2762 experimentally validated antiviral peptide sequences collected from public databases dbAMP, AVPdb, DRAMP, DBAASP, and HIPdb.

[0086] The negative dataset, i.e., non-antiviral peptides, is labeled 0. A total of 10,089 non-antiviral peptide sequences were screened out. The sources and criteria can be as follows:

[0087] Screening for antimicrobial but not antiviral peptides involves identifying peptide sequences with known antibacterial functions but no confirmed antiviral activity from the dbAMP, DBAASP, and DRAMP databases. Simultaneously, all sequences labeled as targeting viruses or possessing antiviral activity should be explicitly excluded during the screening process.

[0088] To screen for non-antimicrobial peptides, sequences can be selected from the UniProt database. Specifically, this can be done by filtering using attribute keywords. Exclusion keywords can include: "toxin," "membrane," "secretory," "defensive," "antibiotic," "anticancer," "antiviral," and "antifungal." The goal is to retain only peptide sequences that do not contain any of these keywords and whose function is unrelated to antimicrobial activity.

[0089] Step a2: Construct the first initial multi-layer graph convolutional network architecture model. In this embodiment, the first initial multi-layer graph convolutional network model can be a multi-layer graph convolutional network with residual connections.

[0090] Step a3: Based on the first dataset, train the first initial multi-layer graph convolutional network model to obtain the general antiviral peptide activity prediction model after training.

[0091] In some alternative implementations, step a3 above includes:

[0092] Step a31 involves preprocessing the first dataset. This specifically includes:

[0093] Remove non-amino acid markers: Remove unknown amino acid markers (such as X) and non-amino acid markers from all peptide sequences;

[0094] Sequence length filtering: Filter sequences with a length less than 5 or greater than 100 (the filtering length is not limited to this and is not specifically restricted here).

[0095] Remove duplicate samples: Perform sequence duplication checks within both the positive and negative sample datasets and remove identical duplicate sequences to ensure the uniqueness of samples within the datasets;

[0096] Detecting the intersection of positive and negative sample sets: Compare the sequences of positive and negative sample sets, identify and delete cross samples that exist in both sets at the same time, so as to ensure the independence of positive and negative samples in the training data;

[0097] Redundant sequence removal: To prevent training data leakage from distorting model performance evaluation, sequence redundancy removal is performed within both the positive and negative sample sets. The sequence homology analysis tool used is CD-HIT (the specific sequence redundancy removal method is not limited), and the sequence similarity threshold can be set to 80% to 95%. The core parameters are set as follows: similarity threshold "-c" ranges from [0.6, 0.99], preferably 0.95; alignment region constraint "-s" ranges from [0.7, 0.9], preferably 0.8; minimum coverage of representative sequence "-aS" ranges from [0.85, 0.95], preferably 0.9; minimum coverage of current sequence "-aL" ranges from [0.85, 0.95], preferably 0.9; disabling length filtering "-d" is set to 0; and the short sequence-specific alignment mode "-n" is set to 2. Only one representative sequence is retained within each cluster. The specific limitations of the above parameters are not strictly defined here.

[0098] The cleaned and redundant-removed positive and negative sample sets were merged to form the final complete training dataset. The complete dataset was then stratified by label and divided into training, validation, and test sets in a 6:2:2 ratio. The positive and negative sample sets obtained from the first stage of data preparation and preprocessing were used to train the general antiviral peptide prediction model.

[0099] Step a32: Based on the preprocessed first dataset, calculate the general features of the sequence, including amino acid type encoding, physicochemical property features, and evolutionary semantic features.

[0100] This embodiment employs a static pre-computation architecture, meaning it uses an offline pre-computation mode to calculate the aforementioned general features of the sequences. It also utilizes a parallel acceleration mechanism, achieving task-level parallelism through a multi-process pool and dynamically allocating computing resources. When training the first-stage prediction model, only sequence-related features need to be calculated.

[0101] The following describes the general characteristics of sequences.

[0102] Regarding amino acid type coding Establish an injective mapping from 20 standard amino acids to a 20-dimensional discrete vector space; output the feature matrix. Where L is the length of the amino acid sequence, and satisfies , (This embodiment uses onehot encoding as an example, but other encoding methods such as BLOSUM 62 can also be used, and there is no particular limitation.)

[0103] Regarding physicochemical properties Calculation: The predefined AAAindex index system was used to calculate the AAAindex indices related to antiviral peptide activity, including KYTJ820101 (hydrophobicity), FAUJ880111 (positively charged residues), GRAR740102 (polarity), FAUJ880112 (negatively charged residues), ROBB760101 (solvent accessibility), FASG890101 (isoelectric point pI), and other physicochemical indices related to antiviral peptide activity; the feature matrix was output. ， in, Number of selected AAIndex indicators (current embodiment) ).

[0104] ESM evolutionary semantic features calculate:

[0105] Sequence tokenization: Converting amino acid sequences into word segmentation index tensors using a predefined vocabulary;

[0106] Deep representation extraction: In gradient calculation disabled mode, sequence representation is extracted from the intermediate layer (preferably the 33rd layer) or the last layer of the ESM model (esm2_t33_650M_UR50D in this embodiment).

[0107] Residue characterization pruning: Remove feature vectors of task-irrelevant markers such as [CLS] and [SEP] from the beginning and end of the sequence;

[0108] Dimensional constraint: Output feature matrix Dimensions ,Right now Where L represents the amino acid length of the antiviral peptide. This represents the hidden layer dimension of the ESM model (1280 in this example).

[0109] Step a33: Based on the preprocessed first dataset, a binary adjacency matrix is ​​constructed using the k-neighborhood method based on sequence distance. In this embodiment, the k-neighborhood method... Integers; binary adjacency matrices do not contain self-loop edges.

[0110] Step a34: Use the general characteristics of the standardized and spliced ​​sequence as node features and the inverse sequence distance as edge weights to construct the first graph structure.

[0111] Specifically, after constructing the binary adjacency matrix based on the k-neighborhood method using sequence distance, a binarization transformation is applied:

[0112] ;

[0113] in, , The total number of standard amino acids in the peptide chain represents the sequence distance between residue i and residue j. In this embodiment, the nearest neighbor residue number is used. Concatenate the amino acid type encoding matrix, physicochemical property feature matrix, and ESM evolutionary semantic feature matrix along the feature dimension as node features; normalize the inverse sequence distance. As edge weight:

[0114] ;

[0115] in, The sequence distance between residue i and residue j; It is a very small constant 1e-7.

[0116] A k-neighborhood graph was constructed based on sequence distance, and the constructed graph structure data was input into the antiviral peptide activity prediction model for training. The training objective was to classify antiviral peptides into binary categories of antiviral and non-antiviral peptides.

[0117] Step a35: Input the first graph structure into the first initial multi-layer graph convolutional network model, and obtain a general antiviral peptide activity prediction model when the preset training indicators are met.

[0118] The first initial multi-layer graph convolutional network architecture in this embodiment includes two graph convolutional layers, each performing a normalized graph convolution operation:

[0119] ;

[0120] in, Adjacency matrix for adding self-loops; for The degree matrix satisfies ; The trainingable weight matrix for the l-th layer; the optimal hidden dimension for the graph convolutional layer is 64; the inter-layer feature reuse mechanism is as follows:

[0121] ;

[0122] in, For the first Layer input feature matrix, For the first Layer trainable weight matrix, This represents a graph convolution operation containing an activation function.

[0123] During the model training phase, the input features are first standardized and concatenated. Specifically, the mean and standard deviation of the physicochemical property feature moments on the training set are calculated, and z-score standardization is performed on the training set, validation set, and test set respectively. The evolutionary semantic feature matrix is ​​then subjected to LayerNorm standardization and linear transformation dimensionality reduction. The amino acid type encoding matrix, the standardized physicochemical property feature matrix, and the dimensionality-reduced evolutionary semantic feature matrix are concatenated along the feature dimensions to generate a node feature matrix; the output dimension of the linear transformation dimensionality reduction ranges from 16 to 128 dimensions. In this embodiment, the linear projection output dimension is 32. In this embodiment, feature standardization and concatenation are implemented during the model training phase.

[0124] During the model training phase, a two-stage output is adopted. First, the node-level features are aggregated into graph-level feature vectors through global average pooling. Then, the graph-level features are mapped into binary classification logits output using a fully connected layer. In the feature transfer application scenario, all trainable parameters of the model are frozen, and the node-level feature vectors output by the last graph convolution layer are used as the sequence feature embedding representation.

[0125] The model training strategy is as follows:

[0126] Optimizer configuration: Optimizer type AdamW, initial learning rate 1e 3;

[0127] Early stop strategy: Monitor validation loss and trigger early stop when validation loss has not improved for 20 consecutive epochs;

[0128] Performance metrics: Monitoring the comprehensive performance metrics of the validation set during training, including ROC-AUC, accuracy, recall, F1 score, Matthews correlation coefficient (MCC), specificity, and area under the PR curve (AUCPR).

[0129] To mitigate the impact of class imbalance on model performance, this embodiment employs a phased processing strategy: During the model training phase, when the classes of samples in the training set are imbalanced (in this embodiment, the ratio of negative to positive samples in the first phase), In the loss function, class weights are applied to the minority class (positive samples). (This weight is only used for gradient calculation during training and does not change the inference logic) to enhance the model's learning of the minority class (positive samples). After model training, temperature scaling is first applied to the validation set to calibrate the model's output probabilities to improve their confidence accuracy. Secondly, based on the calibrated probabilities, the optimal decision threshold that maximizes the F1 score (or other specified metric) is searched on the validation set. Finally, the trained model parameters, the calibrated temperature parameter T, and the optimal threshold are saved. Make predictions on the test set.

[0130] Specifically, temperature scaling and probability threshold calibration are implemented as follows: the original logits value z of the model is scaled using the trainable temperature parameter T:

[0131] ;

[0132] Scaled logits values Input Sigmoid function Calculate the calibration probability:

[0133] ;

[0134] The trainable temperature parameter T is optimized using gradient descent (L-BFGS optimizer), with the objective of minimizing the validation set cross-entropy loss. Subsequently, a decision threshold search is performed: based on the calibrated probabilities, a decision threshold that maximizes the F1 score is searched on the validation set. The calibration performance metrics are recorded. This process is for performance monitoring only; all weight parameters are frozen, and only the temperature parameter T is optimized. The trained model parameters, along with the calibrated temperature parameter T and the optimal threshold, are saved. During the inference phase, the following calculations are performed on the test samples:

[0135] ;

[0136] If and only if Output positive class tags at the time.

[0137] Among these, the classification performance at different thresholds is evaluated to determine the optimal decision threshold. The recorded performance metrics include at least one of the following: validation set F1 score (which is the harmonic mean of precision and recall), accuracy, recall, Matthews correlation coefficient (MCC), and specificity.

[0138] In this embodiment, a general antiviral peptide activity prediction model is constructed by integrating data preprocessing, sequence feature extraction, and graph convolutional network techniques. The obtained general antiviral peptide activity prediction model can effectively capture global and local features in the graph convolutional network. During subsequent learning transfer, after freezing all learnable parameters, a stable node-level feature embedding representation can be output, thereby effectively enhancing the downstream 6-HB-targeting membrane fusion inhibitory peptide prediction model's ability to express complex relationships in the sequence and improving prediction accuracy.

[0139] In this embodiment, the performance comparison between the first-stage general antiviral peptide prediction model and the state-of-the-art (SOTA) model is as follows: Figure 3As shown. The comparison models include: the model developed by Jiahui Guan et al. (Briefings in Bioinformatics, 2024, DOI: 10.1093 / bib / bbae208), i.e., the first comparison model, and the model developed by Ruifen Cao et al. (Briefings in Bioinformatics, 2023, DOI: 10.1093 / bib / bbad353.), i.e., the second comparison model. To ensure fairness, all comparison models use the publicly available code library of the original authors, strictly follow the officially recommended hyperparameters reported in the literature, and are retrained on the exact same training set as the model in this embodiment, and their performance is evaluated on the same reserved test set. Performance metrics include: AUC (Area Under the Curve, referring to the area under the Receiver Operating Characteristic (ROC) curve), F1 score (the harmonic mean of precision and recall), accuracy, recall, Matthews Correlation Coefficient (MCC), specificity, and AUCPR (Area Under the Precision-Recall Curve, referring to the area under the precision-recall curve).

[0140] Step a4: Obtain the second dataset, which includes 6-HB-targeting membrane fusion inhibitory peptides and non-6-HB-targeting antiviral peptides. The 6-HB-targeting membrane fusion inhibitory peptides are used as the positive dataset, and the non-6-HB-targeting antiviral peptides are used as the negative dataset.

[0141] The following is an example. The positive dataset, namely the 6-HB-targeting membrane fusion inhibitory peptide, is tagged with 1. It can be manually screened based on the antiviral peptide targets, mechanisms, or sources in the dbAMP and AVPdb databases. The membrane fusion inhibitory peptides that have been experimentally confirmed to have antiviral activity and clearly target 6-HB were identified, and 540 positive samples were finally obtained.

[0142] The negative dataset, i.e., antiviral peptides that do not target 6-HB and are labeled 0, can be manually screened based on the same database for peptides that have been experimentally proven to have antiviral activity but do not belong to membrane fusion inhibition mechanisms that target 6-HB (such as blocking receptor binding, inhibiting reverse transcription or replication, etc.). In the end, 490 negative samples were obtained.

[0143] Step a5: Construct a second initial multi-layer graph convolutional network model with the same architecture as the first initial multi-layer graph convolutional network model. In this embodiment, the second initial multi-layer graph convolutional network model still adopts a multi-layer graph convolutional network with residual connections.

[0144] Step a6 involves freezing the weights of the general antiviral peptide activity prediction model, using it as a general sequence feature extractor, and fusing the extracted general sequence features into the node features of the second initial multilayer graph convolutional network model. In other words, all trainable parameters of the general antiviral peptide activity prediction model are frozen as a fixed feature extractor, and the extracted sequence features are input into the second initial multilayer graph convolutional network.

[0145] Specifically, a pre-trained general antiviral peptide activity prediction model can be loaded and all trainable parameters frozen, and node-level sequence embedding feature matrices can be extracted. Then, spatial conformation feature extraction is performed. In this embodiment, a spatial conformation feature extraction system can be used to calculate conformation-related AAindex features. Rotation and translation invariant coordinates Euclidean distance Normalized inverse Euclidean distance, Discrete curvature Pseudo-dihedral angles and solvent-accessible surface area (SASA) characteristics related to residues are considered. Further, a graph structure is constructed based on... The Euclidean distance matrix is ​​used to generate a binary adjacency matrix by setting a fixed distance threshold. The node-level sequence representation extracted from the pre-trained model is then compared with the standardized conformation-related AAindex features. Pseudo-dihedral angle, standardized Discrete curvature Rotation- and translation-invariant coordinates, standardized SASA-related features are concatenated along the feature dimensions to construct a joint node feature matrix; Normalized inverse Euclidean distance is used as the edge weight to construct Spatial proximity map.

[0146] Step a7: Based on the second dataset, the general sequence features extracted by the general antiviral peptide activity prediction model are concatenated with the pre-calculated spatial conformation features to train the second initial multilayer graph convolutional network model and obtain the 6-HB targeting membrane fusion inhibitory peptide prediction model.

[0147] Specifically, the constructed graph structure data is input into a second initial multilayer graph convolutional network architecture to train a prediction model for 6-HB-targeting membrane fusion inhibitory peptides. The training objective is to binary classify 6-HB-targeting membrane fusion inhibitory peptides into non-6-HB-targeting antiviral peptides.

[0148] In some alternative implementations, step a7 above includes:

[0149] Step a71: Perform data preprocessing on the second dataset;

[0150] The data preprocessing for the second dataset is completely consistent with the data preparation in the first stage, including: removal of non-amino acid markers, sequence length filtering, removal of duplicate samples by label subsets, detection of the intersection of positive and negative sample sets, and removal of redundant sequences by label subsets. The preprocessed second dataset is then stratified and randomly sampled according to labels; in this embodiment, it is divided into training, validation, and test sets in a 6:2:2 ratio. The positive and negative sample sets obtained from the second stage of data preparation and preprocessing are used to train the 6-HB-targeting membrane fusion inhibitory peptide prediction model.

[0151] Step a72: Based on the preprocessed second dataset, extract sequence features using the aforementioned method, and simultaneously extract spatial conformation features, which include: Rotation and translation invariant coordinates Euclidean distance Normalized inverse Euclidean distance Discrete curvature, Pseudo-dihedral angle, conformation-related AAindex features, and residue solvent-accessible surface area-related features.

[0152] In this embodiment, before extracting three-dimensional conformational features from the sequence, structural modeling of the input peptide sequence is performed to obtain a PDB file containing structural information. Furthermore, the structure is repaired; that is, PDBFixer is used to complete the main chain and side chains of the input peptide, add physiologically appropriate (pH=7.4) hydrogen atoms, and map non-standard residues to the most similar standard residues. In other words, the three-dimensional structure of the input peptide is improved by completing missing atoms and optimizing the protonation state.

[0153] This embodiment employs a static pre-computation architecture, i.e., an offline pre-computation mode, to synchronously compute general sequence features, namely amino acid type encoding, physicochemical property features, ESM evolutionary semantic features, and spatial conformation features, i.e., conformation-related AAindex features. Rotation and translation invariant coordinates Euclidean distance Normalized inverse Euclidean distance Discrete curvature, Pseudo-dihedral angles and residue solvent-accessible surface area-related features; this embodiment also employs a parallel acceleration mechanism, namely, achieving task-level parallelism through a multi-process pool and dynamically allocating computing resources.

[0154] The extraction of three-dimensional conformational features is as follows.

[0155] Regarding spatial geometric invariant coordinate transformation: this involves eliminating the spatial rotation and translation variance of the input peptide atom coordinates by establishing a local coordinate system along the principal inertial axis. In this embodiment, the spatial geometric invariant coordinate transformation unit and the spatial geometric feature calculation unit use standard residues. Calculate the spatial distribution covariance matrix C and the eigenvector matrix V as follows:

[0156] Based on all input peptides Coordinate calculation of spatial distribution covariance matrix C:

[0157] ;

[0158] in, For the i-th The spatial coordinate vector of an atom; All of the target peptide The mean vector of spatial coordinates, i.e., the centroid coordinates; N is... total.

[0159] Perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvector matrix. and eigenvalue diagonal matrix The decomposition satisfies the following relation:

[0160] ;

[0161] in, It is a diagonal matrix composed of eigenvalues; eigenvector matrix This indicates the direction of the principal axis of inertia of the peptide chain.

[0162] By moving the center of mass to the origin and projecting it onto the principal inertial axis coordinate system, rotation and translation invariance are achieved, eliminating the influence of the spatial pose of the peptide conformation on spatial geometric features.

[0163] ;

[0164] in, For the i-th The spatial coordinate vector of an atom after a rotation- and translation-invariant transformation. .

[0165] When the peptide chain length in the input dataset varies greatly, such as with a coefficient of variation (CV) ≥ 0.3, the following standardized scaling normalization can be used to scale the coordinates to a distribution with a mean of 0 and a standard deviation of 1, so that peptide chains of different lengths have comparable spatial characteristics:

[0166] ;

[0167] Where σ represents all The standard deviation vector on each coordinate axis; ε is a very small constant (e.g., 1e-8) to prevent division by zero.

[0168] Standard deviation vector ( , , Calculated independently by coordinate axis components:

[0169] ;

[0170] ;

[0171] ;

[0172] in, , , For the i-th standard residue of the peptide chain The x, y, z coordinate components in the coordinate system after the principal axis inertial rotation; , , For all standard residues of the peptide chain The mean values ​​of the x, y, z coordinate components in the coordinate system after the principal axis inertial rotation; N is the total number of standard residues in the peptide chain.

[0173] about Euclidean distance and normalized inverse Euclidean distance calculate:

[0174] use Geometrically invariant coordinates ( Calculate the Euclidean distance between residues:

[0175] ;

[0176] Min-Max normalization of the Euclidean distance:

[0177] ;

[0178] Calculate the normalized inverse Euclidean distance:

[0179] ;

[0180] about Discrete curvature calculate:

[0181] ;

[0182] in , , For continuous peptide chains coordinate, .

[0183] The calculated result Discrete curvature vector Aligning with the peptide chain length, we obtain the discrete curvature feature vector of the peptide chain. :

[0184] ;

[0185] about pseudo-dihedron calculate:

[0186] ;

[0187] ;

[0188] in, , , , .

[0189] The calculated result Pseudo-dihedral vector ( ) Aligning with the peptide chain length N, we obtain the pseudo-dihedral feature vector of the peptide chain. ) :

[0190] ;

[0191] in, .

[0192] Regarding the Solvent Accessible Surface Area (SASA) characteristic calculate:

[0193] In this embodiment, a spherical probe algorithm with a probe radius of 1.4 Å is used to calculate the solvent-accessible surface area of ​​residues. Based on the solvent-accessible surface area of ​​residues, residue-level SASA feature vectors are calculated. This includes: absolute total SASA value, relative total SASA value, polar surface area ratio, side chain surface area ratio, and exposure status marker (when the SASA of the residue is >40). .

[0194] Regarding the calculation of conformation-related AAIndex features:

[0195] In this embodiment, a predefined AAAindex index system is used to calculate the combination of AAAindex indices related to the conformation of the 6-HB targeting peptide, including but not limited to: ZHOH04010 (interface tendency), RADA880106 (transmembrane helix tendency), CHAM820101 (α-helix tendency), CHOC760101 (β-turn tendency), FAUJ880103 (residue volume), ROSG850101 (side chain spatial complexity), CHAM830107 (van der Waals volume), and CHAM820102 (β-sheet tendency); the output feature matrix is ​​shown. ,in, The number of selected AAIndex indicators (in this example) ).

[0196] Step a73, based on Euclidean distance is used to generate a binary adjacency matrix.

[0197] In this embodiment, based on Euclidean distance matrix, with a fixed distance threshold. Generate a binary adjacency matrix:

[0198] ;

[0199] in, , The total number of standard amino acids in the peptide chain. Based on the distribution of non-bonded interaction distances in the crystal structure of the 6-HB complex, the preferred method is... In this embodiment =8.0.

[0200] Step a74: Input the general sequence features into the weighted, frozen general antiviral peptide activity prediction model to obtain a high-dimensional sequence feature representation.

[0201] Step a75: Concatenate the high-dimensional sequence feature representation with the pre-computed spatial configuration features along the feature dimension to form the enhanced node features; and use the normalized inverse Euclidean distance... As edge weights, the second graph structure is constructed.

[0202] In this embodiment, the node-level sequence representation extracted by the pre-trained model is compared with the standardized conformation-related AAindex features. Pseudo-dihedral angle, standardized Discrete curvature Rotation- and translation-invariant coordinates, and standardized SASA-related features are concatenated along the feature dimensions to construct a joint node feature matrix; additionally, using... Normalized inverse Euclidean distance As edge weight:

[0203] ;

[0204] in, , where is the Euclidean distance between residue i and residue j in a rotation- and translation-invariant coordinate system; It is a very small constant 1e-7.

[0205] Step a76: Input the second graph structure into the second initial multi-layer graph convolutional network model after feature transfer, and obtain the 6-HB targeted membrane fusion inhibitory peptide prediction model when the preset training index is met.

[0206] The second initial multi-layer graph convolutional network architecture in this embodiment includes two graph convolutional layers, each performing a normalized graph convolutional operation:

[0207] ;

[0208] in, Adjacency matrix for adding self-loops; for The degree matrix satisfies ; For the first Layer trainable weight matrix;

[0209] In this embodiment, the preferred hidden dimension of the graph convolutional layer is 64; the inter-layer feature reuse mechanism is as follows:

[0210] ;

[0211] in, For the first Layer input feature matrix, For the first Layer trainable weight matrix, This represents a graph convolution operation containing an activation function.

[0212] After aggregating node-level features into graph-level feature vectors through global average pooling, a fully connected layer is used to map the graph-level features into binary classification logits output.

[0213] During the model training phase, before splicing the sequence representation extracted from pre-training with the spatial conformation features, the mean and standard deviation of the residue solvent accessible surface area related feature matrix, discrete curvature feature matrix, and conformation related AAindex feature matrix on the training set are calculated respectively, and applied to the training set, validation set, and test set to perform the z-score standard.

[0214] The training strategy in this embodiment is as follows:

[0215] Optimizer configuration: Optimizer type AdamW, initial learning rate 1e 3;

[0216] Early stop strategy: Monitor and validate loss, and trigger early stop when there is no improvement in validation loss for 10 consecutive epochs;

[0217] Performance metrics: Monitoring the comprehensive performance metrics of the validation set during training, including ROC-AUC, accuracy, recall, F1 score, Matthews correlation coefficient, specificity, and area under the PR curve.

[0218] To mitigate the impact of class imbalance on model performance, a phased processing strategy is adopted, with the implementation method being the same as in the first phase.

[0219] In this embodiment, the second-stage classification model, namely the 6-HB targeting membrane fusion inhibitory peptide prediction model, is used. The positive and negative samples are nearly balanced, and no class weighting is applied to the loss function. To optimize decision performance, temperature scaling and decision threshold calibration are implemented, similar to the first stage.

[0220] In this embodiment, by integrating the sequence and conformational features of antiviral peptides, key information related to 6-HB targeting can be captured more comprehensively, and the effective model improves the classification performance of 6-HB targeting membrane fusion inhibitory peptides.

[0221] This invention, based on a feature transfer strategy, combines the sequence and spatial conformation features of antiviral peptides, significantly improving multiple performance metrics of the model on the target task test set. See details below. Figure 4 As shown. Furthermore, by freezing the pre-trained model, this strategy significantly reduces the number of trainable parameters for fine-tuning, effectively reducing the risk of overfitting and shortening model training time.

[0222] In some optional embodiments, step S101 above, namely generating the 6-HB-targeting membrane fusion inhibitory candidate peptide, includes:

[0223] Step S1011: Obtain antiviral peptide data. In this embodiment, the pre-training dataset contains 2051 antiviral peptides, which are randomly divided into training set, validation set, and test set in a 6:2:2 ratio.

[0224] Step S1012: Based on the model architecture pre-trained from antiviral peptide data, a general antiviral peptide generation model is obtained.

[0225] Step S1013: Obtain data on the 6-HB targeted membrane fusion inhibitory peptide.

[0226] In this embodiment, the fine-tuning dataset contains 371 experimentally validated membrane fusion inhibitory peptides, which are randomly divided into training, validation, and test sets in a 6:2:2 ratio.

[0227] Step S1014: Based on the 6-HB-targeted membrane fusion inhibitory peptide data, the general antiviral peptide generation model is fine-tuned to obtain the 6-HB-targeted membrane fusion inhibitory peptide generation model; the 6-HB-targeted membrane fusion inhibitory peptide generation model is used to generate membrane fusion inhibitory candidate peptides.

[0228] In this embodiment, a two-stage transfer learning framework is employed. The first stage pre-trains the model on a broad-spectrum antiviral peptide dataset, enabling it to generate general antiviral peptide sequences. The second stage performs end-to-end fine-tuning on a 6-HB-targeting membrane fusion inhibitory peptide dataset, endowing the model with 6-HB-targeting generation capabilities. This strategy not only enables the model to possess both general and specific target generation capabilities but also effectively suppresses overfitting, significantly improving the model's generalization performance. Figure 5 As shown.

[0229] In some alternative implementations, refer to Figure 6 As shown, the 6-HB-targeted membrane fusion inhibitory peptide generation model includes:

[0230] The variable-length sequence preprocessing module is used to preprocess the input sequence. The preprocessing of the input sequence includes: sorting the input sequence in descending order according to its actual length; padding the end of the sorted sequence to a uniform length and generating a corresponding padding mask to mark the position of valid data. The redundant calculation of padding characters by LSTM is skipped through a dynamic masking mechanism.

[0231] The embedding layer, connected to the output of the variable-length sequence preprocessing module, is used to map the preprocessed integer index sequence into a dense vector representation in a high-dimensional space.

[0232] The LSTM layer, connected to the output of the embedding layer, is used to capture long-range dependencies and contextual information in the preprocessed vector sequence;

[0233] The fully connected layer, the output of the LSTM layer is ultimately mapped to the probability distribution of the output space by the fully connected layer;

[0234] The 6-HB targeted membrane fusion inhibitory peptide generation model adopted a dynamic training control strategy during the training phase. The dynamic training control strategy combined gradient norm constraints and continuous validation loss monitoring to improve sequence stability and model generalization ability.

[0235] The sequence generation stage is equipped with a sequence filtering unit; the sequence filtering unit performs the following three operations: adjusting the output probability distribution of the Softmax function using temperature parameters to control generation diversity; dynamically filtering non-amino acid terms to ensure that each generation step only samples from the valid amino acid dictionary; and responding to the end symbol, immediately interrupting the generation process once a special symbol representing the end of the sequence is generated.

[0236] Specifically, regarding the variable-length sequence preprocessing unit, refer to... Figure 7 As shown:

[0237] Add control characters at the beginning and end: Add start characters at the beginning and end of the sequence respectively. <sos>and end symbol <eos>;

[0238] Lexical Encoding System Construction: Construct a 24-dimensional lexical encoding system, including 20 standard amino acids and 4 control symbols (filler symbols). <pad>, start character <sos>, End symbol <eos>Unknown symbol <unk>);

[0239] Sorting and Filling: Sort sequences in descending order by amino acid count; fill the ends of short sequences. <pad>To a uniform length;

[0240] Variable-length compressed metadata generation: By skipping the calculation of padding positions through the sequence compression algorithm (pack_padded_sequence), only the hidden state output of the valid sequence fragments is retained in the LSTM layer processing.

[0241] Regarding the aforementioned sequence generation filtering unit, this embodiment uses a temperature parameter to adjust the output probability distribution of the Softmax function; when the model outputs a preset sequence terminator ( <eos>If this occurs, the generation of the current sequence is immediately terminated; simultaneously, non-amino acid special terms from the vocabulary are rejected. <pad> ， <sos> ， <eos> ， <unk>It outputs effective amino acids and automatically discards candidate sequences with effective peptide sequences shorter than 5.

[0242] In this embodiment, the embedding layer maps the sequences encoded by 24 types of terms to a 32-dimensional distributed representation; the LSTM layer has 2 stacked layers and 48 hidden layer units (expandable to 32-64), with an inter-layer dropout rate of 0.5; the fully connected classification layer has an output dimension of 24, corresponding to the probability distribution of the vocabulary; and the loss function is cross-entropy with padding masking (with ignore_index=0 set).

[0243] Optimizer configuration: Optimizer type AdamW, initial learning rate 1e 3. Weight decay coefficient: 0.01;

[0244] Gradient clipping mechanism: Gradient norm threshold of 1.0 (L2 constraint) to prevent gradient explosion;

[0245] Early stopping strategy: The early stopping strategy in the pre-training stage is set to terminate training if the validation loss does not improve within 20 consecutive training epochs; the corresponding early stopping strategy in the fine-tuning stage is set to 10 epochs. In this embodiment, early stopping is triggered at epoch 190 in the pre-training stage, with the optimal model having a validation loss of 2.5288 and a validation confusion of 12.5379. Early stopping is triggered at epoch 104 in the fine-tuning stage, with the optimal model having a validation loss of 1.8965 and a validation confusion of 6.6623.

[0246] Gradient masking mechanism: The gradient calculation at the padding position is excluded by the ignore_index parameter of the cross-entropy loss (set to padding index 0).

[0247] The 6-HB-targeted membrane fusion inhibitory peptide generation model proceeded in an unconditional autoregressive mode during the inference phase, meaning that sequence generation was initiated by the start symbol ( <sos>Triggered by the model, it does not rely on externally provided conditional information; the next word predicted at each time step depends only on the model's hidden state before that time step; and the model does not accept any externally input protein structure information or functional labels as generation conditions during the generation process.

[0248] The specific configuration for model inference generation is as follows: the sequence generation technique employs an autoregressive inference mechanism. Among these, the start symbol... <sos>(Index 1) Trigger sequence generation; Temperature sampling: Diversity control coefficient τ = 0.7; Termination condition: Output <eos>(Index 2) or terminated when the maximum length of 50 residues is reached. Then, sequence filtering is performed, i.e., special characters are masked: <pad>(Index 0) <sos>(Index 1) <eos>(Index 2) <unk>(Index 3), only retain standard amino acid symbols (Index 4-23); Length constraint: effective sequence length ≥ 5 residues; automatically discard invalid sequences < 5 residues; Batch output: store effective sequences in FASTA format; default to generating 100 candidate peptides per batch (range adjustable).

[0249] In this embodiment, effective word loss and perplexity are used as performance evaluation metrics for the model:

[0250] Effective lexical loss:

[0251] ;

[0252] in, Let be the cross-entropy loss for the i-th word; For indicator functions, when the lexical Non-filler <pad>where T is 1 if the word is in the vocabulary, otherwise 0, B is the batch size, i.e., the total number of tokens in a batch. The valid token loss on the validation set in this example is 2.4988.

[0253] Perplexity: a general indicator to evaluate the quality of sequence generation, the lower the value, the stronger the prediction ability of the model. The perplexity on the validation set in this example is 12.1680.

[0254] The 6-HB targeting membrane fusion inhibiting peptide generation model constructed in this example can process variable-length sequence input and reduce redundant calculations through a dynamic masking mechanism; the sequence stability and model generalization ability are improved through gradient clipping and continuous validation loss monitoring. In the inference stage, the diversity of the output distribution is regulated through a temperature parameter; and a non-amino acid token dynamic filtering and response terminator response mechanism are integrated to ensure the reliability and effectiveness of the generated sequence.

[0255] In some optional embodiments, the step S103 of screening the classification result to obtain the final 6-HB targeting membrane fusion inhibiting candidate peptide comprises:

[0256] Step S1031, determining an initial candidate peptide in the classification result, the initial candidate peptide being a candidate peptide classified as a membrane fusion inhibiting peptide.

[0257] Step S1032, performing experimental molecular docking between each initial candidate peptide and the 6-HB conformation of the target virus to obtain a docking score corresponding to each initial candidate peptide.

[0258] Step S1033, determining whether the docking score meets a preset condition.

[0259] Step S1034, determining the initial candidate peptide corresponding to the docking score meeting the preset docking score condition as a high specificity candidate peptide targeting 6-HB.

[0260] In the above process, an LSTM autoregressive model under a two-stage transfer learning framework is first used to infer and generate 100 (the specific number is not limited) 6-HB targeting membrane fusion inhibiting candidate peptides. The three-dimensional structure of the predicted candidate peptide is predicted, and the sequence features of the candidate peptide are calculated, including amino acid type encoding, physicochemical property features, evolutionary semantic features, and spatial conformation features, including conformation-related AAindex features, rotation and translation invariant coordinates, Euclidean distance, normalized inverse Euclidean distance, discrete curvature, pseudo-dihedral angle, and residue solvent accessible surface area related features. The 6-HB targeting membrane fusion inhibiting peptide prediction model is used to classify and predict the 6-HB targeting property of the candidate peptide.

[0261] Furthermore, candidate peptides with a predicted tag of 1 (positive) were selected and subjected to protein-peptide flexible docking with the pre-fusion conformation of influenza A H3N2 subtype (or other enveloped viruses or subtypes, as determined according to needs) 6-HB to obtain docking scores. The docking software is HADDOCK (or Rosetta FlexPep Dock or HelixFold3 or other complex prediction software).

[0262] Several experimentally validated membrane fusion inhibitory peptides of influenza A H3N2 were collected from DRAVP (or other sources) as positive controls. These peptides were then docked with the target virus 6-HB under identical methods and parameter settings to obtain scores. ;

[0263] Calculate the candidate peptides in the docking system Spiral content, selection Spiral content greater than or equal to preset threshold Candidate peptides; in this embodiment Set it to 60% (this step is optional).

[0264] Based on candidate peptide docking score Compared with positive control peptide By comparing the docking score statistics, candidate peptides with docking affinity superior to the statistical criteria of the positive control peptide were selected as high-specificity inhibitory peptide candidates for the target virus 6-HB. mean and standard deviation The comparison relationship is the docking score. Candidate peptides that meet the following criteria are identified as highly specific candidate peptides for the target virus 6-HB:

[0265] ;

[0266] in, This is a predefined constant, set to 1 in this embodiment.

[0267] In this embodiment, by performing experimental docking on each initial candidate peptide and further screening based on the docking score, the success rate of downstream biological validation experiments can be effectively improved.

[0268] This embodiment also provides a 6-HB-targeted membrane fusion inhibitory peptide prediction device, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the term "module" can be a combination of software and / or hardware that performs a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0269] The embodiment provides a 6-HB targeting membrane fusion inhibiting peptide prediction device, as shown in the figure, the device comprises: Figure 8 A sequence generation module 201 is configured to generate a 6-HB targeting membrane fusion inhibiting candidate peptide.

[0270] A prediction module 202 is configured to input the membrane fusion inhibiting candidate peptide into a pre-constructed 6-HB targeting membrane fusion inhibiting peptide prediction model to perform classification prediction, and obtain a classification result of the membrane fusion inhibiting candidate peptide; wherein the 6-HB targeting membrane fusion inhibiting peptide prediction model is a two-stage transfer learning classification model, the first stage is based on sequence general features to perform antiviral peptide and non-antiviral peptide binary classification prediction; and the second stage is combined with spatial conformation features to perform membrane fusion inhibiting peptide and non-membrane fusion inhibiting antiviral peptide binary classification prediction.

[0271] A screening module 203 is configured to screen the classification result to obtain a final 6-HB targeting membrane fusion inhibiting candidate peptide.

[0272] In some optional embodiments, the device further comprises:

[0273] A construction module is configured to obtain a first data set, the first data set comprising antiviral peptides and non-antiviral peptides, wherein the antiviral peptides are taken as positive data sets and the non-antiviral peptides are taken as negative data sets; a first initial multi-layer graph convolutional network model is constructed; the first initial multi-layer graph convolutional network model is trained based on the first data set to obtain a trained general antiviral peptide activity prediction model; wherein the method comprises: performing data preprocessing on the first data set; based on the preprocessed first data set, sequence general features are calculated, the sequence general features comprising amino acid type encoding features, physicochemical property features and evolutionary semantic features; based on the preprocessed first data set, a k-neighbor method based on sequence distance is used to construct a binary adjacency matrix; the sequence general features are taken as node features and the reverse sequence distance is taken as edge weight to construct a first graph structure; the first graph structure is input into the first initial multi-layer graph convolutional network model, and when a preset training index is met, the general antiviral peptide activity prediction model is obtained.

[0274]

[0275] ​The construction module is also configured to obtain a second data set, the second data set comprising 6-HB targeting membrane fusion inhibitory peptides and non-6-HB targeting antiviral peptides, wherein the 6-HB targeting membrane fusion inhibitory peptides are taken as positive data sets, and the non-6-HB targeting antiviral peptides are taken as negative data sets; a second initial multi-layer graph convolution network model with the same architecture as the first initial multi-layer graph convolution network model is constructed; the general antiviral peptide activity prediction model weight is frozen as a sequence feature extractor, and the extracted sequence general features are used for fusion to node features of the second initial multi-layer graph convolution network architecture model; the general sequence features extracted by the general antiviral peptide activity prediction model are spliced with the pre-calculated spatial conformation features based on the second data set, the second initial multi-layer graph convolution network model is trained, and a 6-HB targeting membrane fusion inhibitory peptide prediction model is obtained. Wherein, it comprises: data preprocessing of the second data set; based on the preprocessed second data set, spatial conformation features are extracted, the spatial conformation features comprising: rotationally translationally invariant coordinates, Euclidean distance, normalized inverse Euclidean distance, discrete curvature, pseudo-dihedral angle, conformation-related AAindx feature, and residue solvent accessible surface area-related feature; based on Euclidean distance, a binary adjacency matrix is generated; the general sequence features are spliced along the feature dimension and input to the general antiviral peptide activity prediction model with frozen weights to obtain high-dimensional sequence feature representation; the high-dimensional sequence feature representation and the spatial conformation features are spliced along the feature dimension and used as enhanced node features; and the normalized inverse Euclidean distance is taken as an edge weight to construct a second graph structure; the second graph structure is input to the second initial multi-layer graph convolution network model after feature migration, and when a preset training index is met, the 6-HB targeting membrane fusion inhibitory peptide prediction model is obtained.

[0276] The 6-HB targeting membrane fusion inhibition peptide generation model comprises: a variable-length sequence preprocessing module configured to preprocess an input sequence; preprocessing the input sequence comprises: sorting the input sequence in descending order of actual length; padding the sorted sequence at the end to a uniform length, and generating a corresponding padding mask to mark the valid data position; an embedding layer connected to the output end of the variable-length sequence preprocessing module, configured to map the preprocessed integer index sequence to a dense vector representation in a high-dimensional space; an LSTM layer connected to the output end of the embedding layer, configured to capture long-range dependencies and context information in the preprocessed vector sequence; a fully connected layer, the output of the LSTM layer is finally mapped to a probability distribution of the output space by the fully connected layer; the 6-HB targeting membrane fusion inhibition peptide generation model adopts a dynamic training control strategy in the training stage, which combines gradient norm constraint and continuous validation loss monitoring; the 6-HB targeting membrane fusion inhibition peptide generation model is equipped with a generated sequence filtering unit in the sequence generation stage; the generated sequence filtering unit performs the following three operations: using a temperature parameter to adjust the output probability distribution of the Softmax function to control the generation diversity; dynamically filtering non-amino acid tokens to ensure that each step of generation only samples in the valid amino acid dictionary; response terminator, once a special symbol representing the end of the sequence is generated, the generation process is immediately interrupted.

[0277] In some optional embodiments, the sequence generation module 201 comprises:

[0278] The sequence generation unit is configured to: obtain antiviral peptide data; pretrain the pre-constructed model architecture based on the antiviral peptide data to obtain a general antiviral peptide generation model; obtain 6-HB targeting membrane fusion inhibition peptide data; fine-tune the general antiviral peptide generation model based on the 6-HB targeting membrane fusion inhibition peptide data to obtain a 6-HB targeting membrane fusion inhibition peptide generation model; and the 6-HB targeting membrane fusion inhibition peptide generation model is used to generate membrane fusion inhibition candidate peptides.

[0279] In some optional embodiments, the screening module 203 comprises:

[0280] The screening unit is configured to: determine an initial candidate peptide in the classification result, the initial candidate peptide being a membrane fusion inhibition candidate peptide classified as a 6-HB targeting membrane fusion inhibition peptide; perform trial molecular docking of each initial candidate peptide with a target virus 6-HB conformation to obtain a docking score corresponding to each initial candidate peptide; determine whether the docking score meets a preset condition; and determine a candidate peptide meeting the preset docking score condition as a final 6-HB targeting membrane fusion inhibition candidate peptide.

[0281] The 6-HB-targeted membrane fusion inhibitory peptide prediction device in this embodiment is presented in the form of a functional unit. Here, a unit refers to an ASIC circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0282] Further functional descriptions of the above modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0283] This invention also provides a computer device having the above-described features. Figure 8 The device shown is a 6-HB-targeted membrane fusion inhibitory peptide prediction device.

[0284] Please see Figure 9 , Figure 9 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 9 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 9 Take a processor 10 as an example.

[0285] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0286] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0287] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0288] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0289] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0290] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0291] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.< / pad> < / unk> < / eos> < / sos> < / pad> < / eos> < / sos> < / sos> < / unk> < / eos> < / sos> < / pad> < / eos> < / pad> < / unk> < / eos> < / sos> < / pad> < / eos> < / sos>

Claims

1. A method for predicting 6-HB-targeted membrane fusion inhibitory peptides, characterized in that, The method includes: Generate 6-HB-targeting membrane fusion inhibition candidate peptide; The candidate peptides for membrane fusion inhibition are input into a pre-constructed 6-HB-targeted membrane fusion inhibition peptide prediction model for classification and prediction, obtaining the classification results of the candidate peptides. The 6-HB-targeted membrane fusion inhibition peptide prediction model is a two-stage transfer learning classification model. The first stage performs binary classification prediction between antiviral peptides and non-antiviral peptides based on sequence general features. The second stage incorporates spatial conformation features for binary classification prediction between 6-HB-targeted membrane fusion inhibition peptides and non-6-HB-targeted antiviral peptides. Specifically, the first stage constructs a graph neural network based on the k-contact diagram of the antiviral peptide sequence; the second stage constructs a graph neural network based on the k-contact diagram of the antiviral peptide sequence. A graph neural network for a binary contact graph of Euclidean distances between carbon atoms; The classification results were screened to obtain the final 6-HB-targeting membrane fusion inhibitory candidate peptide; The 6-HB-targeted membrane fusion inhibitory peptide prediction model was established through the following steps: Obtain a first dataset, which includes antiviral peptides and non-antiviral peptides, wherein the antiviral peptides are used as a positive dataset and the non-antiviral peptides are used as a negative dataset. Construct the first initial multi-layer graph convolutional network model; Based on the first dataset, train the first initial multilayer graph convolutional network model to obtain a general antiviral peptide activity prediction model after training. Obtain a second dataset, which includes 6-HB-targeting membrane fusion inhibitory peptides and non-6-HB-targeting antiviral peptides, wherein the 6-HB-targeting membrane fusion inhibitory peptides are used as positive datasets and the non-6-HB-targeting antiviral peptides are used as negative datasets. Construct a second initial multi-layer graph convolutional network model with the same architecture as the first initial multi-layer graph convolutional network model; The weights of the general antiviral peptide activity prediction model are frozen and used as a sequence general feature extractor. The extracted sequence general features are used to be fused into the node features of the second initial multilayer graph convolutional network model. The sequence general features include amino acid type encoding features, physicochemical property features, and evolutionary semantic features. Based on the second dataset, the general sequence features extracted by the general antiviral peptide activity prediction model are concatenated with the pre-calculated spatial conformation features to train the second initial multilayer graph convolutional network model, thereby obtaining the 6-HB targeting membrane fusion inhibitory peptide prediction model; the spatial conformation features include: Rotation and translation invariant coordinates Euclidean distance Normalized inverse Euclidean distance Discrete curvature, Pseudo-dihedral angle, conformation-related AAindex features, and residue solvent-accessible surface area-related features.

2. The method according to claim 1, characterized in that, The step of training the first initial multi-layer graph convolutional network model based on the first dataset includes: Perform data preprocessing on the first dataset; Based on the preprocessed first dataset, calculate the general features of the sequence; Based on the preprocessed first dataset, a binary adjacency matrix is ​​constructed using the k-neighborhood method based on sequence distance. The general features of the sequence are used as node features, and the inverse sequence distance is used as edge weights to construct the first graph structure; By inputting the first graph structure into the first initial multi-layer graph convolutional network model, a general antiviral peptide activity prediction model is obtained when the preset training indicators are met.

3. The method according to claim 2, characterized in that, The step of training the second initial multi-layer graph convolutional network model based on the second dataset involves concatenating the general sequence features extracted by the general antiviral peptide activity prediction model with the pre-calculated spatial conformation features. Perform data preprocessing on the second dataset; Based on the preprocessed second dataset, spatial configuration features are extracted; Based on the above Euclidean distance is used to generate a binary adjacency matrix; The general sequence features are concatenated along the feature dimension and input into the general antiviral peptide activity prediction model after weight freezing to obtain a high-dimensional sequence feature representation. The high-dimensional sequence feature representation and the spatial configuration feature are concatenated along the feature dimension to form an enhanced node feature; and the... The normalized inverse Euclidean distance is used as the edge weight to construct the second graph structure; By transferring the second graph structure into the second initial multi-layer graph convolutional network model after feature transfer, the 6-HB targeted membrane fusion inhibitory peptide prediction model is obtained when the preset training index is met.

4. The method according to claim 1, characterized in that, The generated 6-HB-targeting membrane fusion inhibitory candidate peptide includes: Obtain antiviral peptide data; Based on the model architecture pre-trained from the aforementioned antiviral peptide data, a general antiviral peptide generation model is obtained. Obtain data on 6-HB targeted membrane fusion inhibitory peptide; The general antiviral peptide generation model was fine-tuned based on the 6-HB-targeted membrane fusion inhibitory peptide data to obtain a 6-HB-targeted membrane fusion inhibitory peptide generation model; the 6-HB-targeted membrane fusion inhibitory peptide generation model is used to generate the membrane fusion inhibitory candidate peptide.

5. The method according to claim 1, characterized in that, The 6-HB-targeted membrane fusion inhibitory peptide generation model includes: A variable-length sequence preprocessing module is used to preprocess the input sequence; the preprocessing of the input sequence includes: sorting the input sequence in descending order according to its actual length; padding the ends of the sorted sequence to a uniform length and generating a corresponding padding mask to mark the valid data position; An embedding layer, connected to the output of the variable-length sequence preprocessing module, is used to map the preprocessed integer index sequence into a dense vector representation in a high-dimensional space. An LSTM layer, connected to the output of the embedding layer, is used to capture long-range dependencies and contextual information in the preprocessed vector sequence; The fully connected layer is used to map the output of the LSTM layer to a probability distribution in the output space. The 6-HB targeted membrane fusion inhibitory peptide generation model employs a dynamic training control strategy during the training phase, which combines gradient norm constraints and continuous validation loss monitoring. The 6-HB targeted membrane fusion inhibitory peptide generation model is equipped with a generation sequence filtering unit during the sequence generation stage. The generation sequence filtering unit performs the following three operations: adjusting the output probability distribution of the Softmax function using a temperature parameter to control generation diversity; dynamically filtering non-amino acid terms to ensure that each generation step only samples from the valid amino acid dictionary; and responding to the end symbol, immediately interrupting the generation process once a special symbol representing the end of the sequence is generated.

6. The method according to claim 1, characterized in that, The screening of the classification results yields the final 6-HB-targeting membrane fusion inhibitory candidate peptides, including: The initial candidate peptide in the classification results is determined, and the initial candidate peptide is the membrane fusion inhibitory candidate peptide that is classified as a 6-HB-targeting membrane fusion inhibitory peptide. Each of the initial candidate peptides was experimentally docked with the 6-HB conformation of the target virus to obtain a docking score corresponding to each of the initial candidate peptides. Determine whether the docking score meets the preset conditions; The initial candidate peptides that meet the preset docking scoring conditions are determined as the final 6-HB-targeted membrane fusion inhibition candidate peptides.

7. A device for predicting 6-HB-targeted membrane fusion inhibitory peptides, characterized in that, The device includes: A sequence generation module is used to generate membrane fusion inhibitory candidate peptides targeting 6-HB; The prediction module is used to input the membrane fusion inhibition candidate peptides into a pre-constructed 6-HB-targeted membrane fusion inhibition peptide prediction model for classification prediction, and obtain the classification results of the membrane fusion inhibition candidate peptides. The 6-HB-targeted membrane fusion inhibition peptide prediction model is a two-stage transfer learning classification model. The first stage performs binary classification prediction between antiviral peptides and non-antiviral peptides based on sequence general features. The second stage incorporates spatial conformation features to perform binary classification prediction between 6-HB-targeted membrane fusion inhibition peptides and non-6-HB-targeted antiviral peptides. Specifically, the first stage constructs a graph neural network based on the k-contact graph of the antiviral peptide sequence; the second stage constructs a graph neural network based on the k-contact graph of the antiviral peptide sequence. A graph neural network for a binary contact graph of Euclidean distances between carbon atoms; A screening module is used to screen the classification results to obtain the final 6-HB-targeting membrane fusion inhibition candidate peptide; A construction module is used to obtain a first dataset, which includes antiviral peptides and non-antiviral peptides, wherein the antiviral peptides are used as a positive dataset and the non-antiviral peptides are used as a negative dataset. Construct the first initial multi-layer graph convolutional network model; Based on the first dataset, train the first initial multilayer graph convolutional network model to obtain a general antiviral peptide activity prediction model after training. Obtain a second dataset, which includes 6-HB-targeting membrane fusion inhibitory peptides and non-6-HB-targeting antiviral peptides, wherein the 6-HB-targeting membrane fusion inhibitory peptides are used as positive datasets and the non-6-HB-targeting antiviral peptides are used as negative datasets. Construct a second initial multi-layer graph convolutional network model with the same architecture as the first initial multi-layer graph convolutional network model; The weights of the general antiviral peptide activity prediction model are frozen and used as a sequence general feature extractor. The extracted sequence general features are used to be fused into the node features of the second initial multilayer graph convolutional network model. The sequence general features include amino acid type encoding features, physicochemical property features, and evolutionary semantic features. Based on the second dataset, the general sequence features extracted by the general antiviral peptide activity prediction model are concatenated with the pre-calculated spatial conformation features to train the second initial multilayer graph convolutional network model, thereby obtaining the 6-HB targeting membrane fusion inhibitory peptide prediction model; the spatial conformation features include: Rotation and translation invariant coordinates Euclidean distance Normalized inverse Euclidean distance Discrete curvature, Pseudo-dihedral angle, conformation-related AAindex features, and residue solvent-accessible surface area-related features.

8. A computer device, characterized in that, include: A memory and a processor are communicatively connected, the memory stores computer instructions, and the processor executes the computer instructions to perform the 6-HB-targeted membrane fusion inhibitory peptide prediction method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the 6-HB-targeted membrane fusion inhibitory peptide prediction method according to any one of claims 1-6.

Images