Method for predicting polypeptide hemolytic activity through graph attention fusion method based on dynamic characteristics

Abstract

The invention relates to the cross technical field of bioinformatics and computational biology, in particular to a method for predicting polypeptide hemolytic activity by a graph attention fusion method based on dynamic characteristics, which comprises the following steps: acquiring a data set, and constructing polypeptide sequence dynamic characteristics based on the data set; the dynamic features of the polypeptide sequence comprise N * 1500 * 5-dimensional feature tensors, and N is a residue base number in the polypeptide sequence; based on the dynamic features of the polypeptide sequence, constructing a time sequence diagram attention convolutional network; based on a time sequence diagram attention convolutional network, a diagram attention network and time sequence multi-head self-attention are fused, multi-scale feature extraction and fusion are achieved, and classification prediction is completed through global pooling and a full-connection neural network. According to the method, the dynamic feature tensor of the polypeptide sequence is constructed through a molecular dynamics simulation technology, and efficient prediction of the hemolytic polypeptide is realized in combination with graph data structure conversion and a time sequence graph attention convolutional network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics and computational biology, in particular to a method for predicting polypeptide hemolytic activity based on dynamic characteristics and graph attention fusion. BACKGROUND

[0002] Hemolytic polypeptides play an important role in the development of drug delivery systems, antibacterial agents and biological pesticides. The evaluation of hemolytic activity is a key indicator for measuring biological activity and safety. Currently, traditional experimental methods such as in vitro hemolysis test and animal model detection can provide direct data, but they have problems such as long experimental period, high cost, limited sample size, etc., which cannot meet the demand of high-throughput screening, and the experimental results are easily affected by sample differences and environmental factors, and the repeatability is poor.

[0003] In the field of computational prediction, early static feature analysis methods based on amino acid composition and physicochemical properties cannot capture the dynamic changes of polypeptides in biological environments and the interaction process with cell membranes. With the development of deep learning technology, some models attempt to extract features from polypeptide sequences for prediction, but still face many challenges: first, the data samples of hemolytic polypeptides are relatively scarce, which leads to insufficient training data for the model and limited generalization ability; second, existing pre-trained models have limitations in protein field application, and it is difficult to analyze the complex nonlinear relationship between polypeptide sequence, structure and hemolytic activity; third, traditional models lack effective mining of dynamic information such as conformational evolution and thermodynamic action of polypeptides in membrane environments, making it difficult to meet the actual application demand of prediction accuracy. Therefore, it is urgent to innovate technical means and break through the bottleneck of existing prediction methods to realize accurate prediction of hemolytic polypeptides. SUMMARY

[0004] The embodiment of the present application provides a method for predicting the hemolytic activity of polypeptides based on the graph attention fusion method of dynamic characteristics, which constructs the dynamic feature tensor of polypeptide sequence by molecular dynamics simulation technology, combines graph data structure conversion and time series graph attention convolution network, and realizes efficient prediction of hemolytic polypeptides.

[0005] To solve the above technical problems, in a first aspect, the embodiment of the present application provides a method for predicting the hemolytic activity of polypeptides based on the graph attention fusion method of dynamic characteristics, which comprises the following steps: first, obtaining a data set and constructing dynamic features of polypeptide sequences based on the data set; the dynamic features of polypeptide sequences include an N x 1500 x 5 dimensional feature tensor, N being the number of residues in the polypeptide sequence; then, constructing a time series graph attention convolution network based on the dynamic features of polypeptide sequences; finally, based on the time series graph attention convolution network, fusing the graph attention network and the time series multi-head self-attention to realize multi-scale feature extraction and fusion, and completing classification prediction through global pooling and fully connected neural network.

[0006] In some exemplary embodiments, acquiring a dataset and constructing dynamic features of the peptide sequence based on the dataset includes: performing a 150 ns high-resolution molecular dynamics simulation of the peptide in a human physiological simulation environment using the GROMACS simulation platform; extracting 5-dimensional features for each residue with a 2 fs step size and a 100 ps sampling interval, and constructing an N×1500×5-dimensional feature tensor to record the dynamic evolution of the peptide in the membrane environment; the feature tensor is represented as: X∈R 1500×N×5 Where X is the feature tensor and R is XX.

[0007] In some exemplary embodiments, the simulation environment employs a mixture of aqueous solution and phospholipid bilayer membrane. The aqueous solution uses the SPC / E water model and contains approximately 10,000 water molecules to simulate a physiological solution environment. The phospholipid bilayer membrane is constructed using 1-palmitoyl-2-oleoyl-sn-glycerol-3-phosphocholine lipids, forming a bilayer membrane containing 128 lipid molecules to simulate the erythrocyte membrane environment. The system temperature is controlled at 310 K using the Nosé-Hoover temperature control algorithm, and the pressure is controlled at 1 bar using the Parrinello-Rahman pressure control algorithm. A CHARMM36 force field is used to ensure accurate modeling of protein-lipid interactions.

[0008] In some exemplary embodiments, the extracted 5-dimensional features are dynamic features associated with the hemolytic peptide, including root mean square fluctuation, solvent-accessible surface area, number of hydrogen bonds, van der Waals interaction energy, and solvation free energy.

[0009] In some exemplary embodiments, a temporal graph attention convolutional network is constructed based on the dynamic features of the peptide sequence, including: converting the dynamic features of the peptide sequence into a dynamic graph sequence, and constructing the temporal graph attention convolutional network through Z-score normalization and adjacency matrix.

[0010] In some exemplary embodiments, the dynamic features of the peptide sequence are transformed into a dynamic graph sequence, and a temporal graph attention convolutional network is constructed through Z-score normalization and adjacency matrix, including: based on the feature tensor X∈R in the dynamic features of the peptide sequence. 1500×N×5 Construct 1500 dynamic graph sequences Among them, each graph G t =(V t E t ,X t Characterizing the topological structure and dynamic properties of the peptide at time step t, V t For node set, E t For edge set, X t The node feature matrix is ​​used; Z-score normalization is performed on each feature dimension d; the normalization formula is:

[0011]

[0012] wherein X t,i,d represents the d-th dimension of the i-th residue at time t, μ d is the mean of the d-th dimension, σ d is the variance of the d-th dimension; G t The corresponding adjacency matrix is:

[0013]

[0014] wherein A t (i,j) represents whether there is an edge between the i-th residue and the j-th residue at time t.

[0015] In some example embodiments, based on the time-series graph attention convolutional network, the graph attention network and the time-series multi-head self-attention are fused, including: processing the dynamic graph sequence based on the time-series graph attention convolutional module The graph attention network and the time-series multi-head self-attention are fused; the calculation process of the graph attention network includes: for each time step t, the computer attention coefficient is calculated, which is used to measure the importance of node j to node i at time step t, and the calculation formula is:

[0016]

[0017] wherein, is a learnable weight matrix used to project the input features to a new feature space F', F' is 32, is a learnable attention vector, || represents a splicing operation, two vectors are spliced, and LcakyRcLU is an activation function.

[0018] In some example embodiments, the calculation process of the time-series multi-head self-attention includes: using the softmax function for normalization; then, the feature of node i at time step t is updated by aggregating the features of the neighbor nodes through the attention coefficient; subsequently, the time-series multi-head self-attention is calculated; the graph attention network and the time-series multi-head self-attention are combined to obtain the feature matrix obtained at different scales; the feature matrix is fused through weighted summation; the fused feature matrix is globally pooled to obtain the feature representation of the graph; and the feature representation of the graph is input to a fully connected neural network for classification.

[0019] In some example embodiments, in the process of fusing the graph attention network and the time-series multi-head self-attention, a loss function is constructed for learning the network parameters of the time-series graph attention convolutional network; the loss function includes a classification loss function and a graph structure preservation loss function.

[0020] In some example embodiments, the classification loss function employs a cross-entropy loss function as follows:

[0021]

[0022] where N is the number of samples, y n is the true label of the nth sample, is the probability that the model predicts the nth sample belongs to hemolytic polypeptide; the graph structure preservation loss function is as follows:

[0023]

[0024] where T is the number of time steps, |V t | is the number of nodes at time step t, and m is a positive boundary value for controlling the minimum distance between non-adjacent node features;

[0025] The loss function is: L = λ cls L cls + λ struct L struct , where λ cls = 0.5, and λ struct = 0.5.

[0026] The technical scheme provided by the embodiments of the present application has at least the following advantages:

[0027] The method for predicting the hemolytic property of a polypeptide based on a graph attention fusion method of kinetic characteristics provided by the embodiments of the present application comprises the following steps: first, a data set is obtained, and based on the data set, a dynamic feature of a polypeptide sequence is constructed; the dynamic feature of the polypeptide sequence comprises an N x 1500 x 5-dimensional feature tensor, and N is the number of residues in the polypeptide sequence; then, based on the dynamic feature of the polypeptide sequence, a time-series graph attention convolutional network is constructed; finally, based on the time-series graph attention convolutional network, a graph attention network and a time-series multi-head self-attention are fused to realize multi-scale feature extraction and fusion, and classification prediction is completed through global pooling and a fully connected neural network. The present application realizes efficient prediction of hemolytic polypeptides by constructing a dynamic feature tensor of a polypeptide sequence through molecular dynamics simulation technology, combining graph data structure conversion and a time-series graph attention convolutional network, and can be widely applied in the fields of drug development, biological pesticide design and biomedical toxicity evaluation.

[0028] The hemolytic polypeptide prediction method DeepHemo provided in the present application is based on molecular dynamics simulation, combined with deep learning and graph data processing technology. First, the GROMACS simulation platform is used to perform 150ns high-resolution molecular dynamics simulation of the polypeptide in a human physiological simulation environment, with a 2fs step and a 100ps sampling interval. The 5-dimensional dynamic features of each residue are extracted to construct an N x 1500 x 5-dimensional feature tensor, which records the dynamic evolution of the polypeptide in the membrane environment. Then the feature tensor is converted into 1500 dynamic graph sequences, which are standardized by Z-score and constructed into an adjacency matrix to provide structured data for graph convolutional networks. Then, the time series graph attention convolutional network is used to fuse the graph attention network (GAT) and the time series multi-head self-attention (MHSA) to realize multi-scale feature extraction and fusion. Finally, the global pooling and fully connected neural network are used for classification and prediction. At the same time, a composite loss function including classification loss and graph structure preservation loss is designed to optimize the model training process. The method fills the gap in the prior art and has important practical application value. BRIEF DESCRIPTION OF DRAWINGS

[0029] One or more embodiments are illustrated by way of example in the figures that form a part of this patent document, and in which is shown a number of specific embodiments in which embodiments can be practiced. In addition, the drawings are not necessarily drawn to scale, except as otherwise noted in the following

[0030] Figure 1 is a flowchart of a method for predicting the hemolytic property of a polypeptide based on a graph attention fusion method of kinetic characteristics provided by an embodiment of the present application.

[0031] Figure 2 is an AUROC curve comparison of the method for predicting the hemolytic property of a polypeptide provided by the present application on the DS1, DS2, DS3 and DS4 data sets.

[0032] Figure 3 is a PRROC curve comparison chart of the method for predicting the hemolytic property of a polypeptide provided by the present application and other methods on the DS1, DS2, DS3 and DS4 data sets. DETAILED DESCRIPTION

[0033] As known from the background, existing hemolytic polypeptide prediction models attempt to extract features from polypeptide sequences for prediction, but still face many challenges. Therefore, there is an urgent need for innovative technical means to break through the bottleneck of existing prediction methods and achieve accurate prediction of hemolytic polypeptides.

[0034] Hemolytic polypeptides are a class of polypeptide molecules with potential biomedical applications, which have the ability to dissolve cell membranes in living organisms, and therefore have wide application prospects in the fields of drug development, biological pesticides, etc. However, accurately predicting the hemolytic activity of polypeptides is still a challenging task. Traditional polypeptide sequence feature extraction methods cannot fully capture the semantic and contextual information in polypeptide sequences, and the application of existing pre-trained models in the protein field still has some limitations. In order to overcome these problems and improve the accuracy and reliability of hemolytic polypeptide prediction, the present application proposes an innovative method that combines the advantages of traditional feature extraction and pre-trained models, and uses deep network and multi-path data fusion strategy to achieve efficient prediction of hemolytic polypeptides. The goal of the present application is to provide a more reliable and effective method for polypeptide sequence analysis and related fields of research and application.

[0035] To solve the above technical problems, the present application provides a method for predicting the hemolytic activity of polypeptides based on the fusion of dynamic characteristics and graph attention, which comprises the following steps: first, obtaining a data set and constructing polypeptide sequence dynamic characteristics based on the data set; the polypeptide sequence dynamic characteristics include an N x 1500 x 5 dimensional feature tensor, and N is the number of residues in the polypeptide sequence; then, based on the polypeptide sequence dynamic characteristics, a time series graph attention convolutional network is constructed; finally, based on the time series graph attention convolutional network, the graph attention network and the time series multi-head self-attention are fused to realize multi-scale feature extraction and fusion, and the classification prediction is completed through global pooling and fully connected neural network. The present application provides a method for predicting the hemolytic activity of polypeptides based on the fusion of dynamic characteristics and graph attention, which uses machine learning algorithm to couple hydrological model to construct a water quality evaluation model that is convenient to use and suitable for local conditions, and realizes efficient and accurate prediction of future water quality.

[0036] The embodiments of the present application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art can understand that in the embodiments of the present application, many technical details are proposed in order to enable the reader to better understand the present application. However, the technical solutions claimed by the present application can be realized even without these technical details and various changes and modifications based on the following embodiments.

[0037] Referring to Figure 1 , the present application provides a method for predicting the hemolytic activity of polypeptides based on the fusion of dynamic characteristics and graph attention, which comprises the following steps:

[0038] Step S101, obtaining a data set and constructing polypeptide sequence dynamic characteristics based on the data set; the polypeptide sequence dynamic characteristics include an N x 1500 x 5 dimensional feature tensor, and N is the number of residues in the polypeptide sequence.

[0039] Step S102, constructing a time series graph attention convolutional network based on the polypeptide sequence dynamic characteristics.

[0040] In step S103, based on the time sequence diagram attention convolution network, the graph attention network and the time sequence multi-head self-attention are fused to realize multi-scale feature extraction and fusion, and classification prediction is completed through global pooling and full connection neural network.

[0041] The method for predicting the hemolytic activity of polypeptides based on the dynamic feature fusion method of the graph attention provided by the application is a method for comprehensively utilizing traditional features and pre-training models. The prediction accuracy and reliability of hemolytic polypeptides are improved through the strategy of deep network and multi-path data fusion. This innovative method has wide potential in polypeptide sequence analysis and related fields. The goal of the application is to provide a more reliable and effective method for polypeptide sequence analysis and related fields. The steps of the method include: first, constructing the dynamic features of the polypeptide sequence. In order to capture the dynamic behavior of the polypeptide sequence in the biological environment, the application uses molecular dynamics (MD) simulation to generate a comprehensive time series feature tensor. The tensor is represented as T ∈ R N×T×M , where N is the number of residues in the polypeptide sequence, T is the time step, and M is the feature dimension of each residue. The feature set covers physicochemical properties, structural properties, dynamic behavior, and interaction characteristics, aiming to reveal the complex relationship between polypeptide dynamics and hemolytic activity. The generated tensor provides high-resolution time evolution information, providing key support for the hemolytic polypeptide prediction model.

[0042] In some embodiments, in step S101, a data set is obtained, and based on the data set, the dynamic features of the polypeptide sequence are constructed, including the following steps:

[0043] In step S1011, the GROMACS simulation platform is used to perform 150ns high-resolution molecular dynamics simulation of the polypeptide in a human physiological simulation environment.

[0044] In step S1012, 5-dimensional features of each residue are extracted at a 2fs step and a 100ps sampling interval, and an N x 1500 x 5-dimensional feature tensor is constructed to record the dynamic evolution of the polypeptide in the membrane environment. The feature tensor is represented as: X ∈ R 1500×N×5 ; where X is the feature tensor and R is a real matrix.

[0045] The application uses GROMACS to perform 150ns molecular dynamics simulation, with a simulation step of 2fs and a sampling interval of 100ps, generating T = 1500 time steps.

[0046] In some embodiments, the simulation environment in step S1011 adopts a mixed system of an aqueous solution and a phospholipid bilayer membrane; wherein the aqueous solution adopts an SPC / E water model, contains about 10,000 water molecules, and simulates a physiological solution environment; the phospholipid bilayer membrane adopts 1-palmitoyl-2-oleoyl-sn-glycero-3-phosphocholine lipid, constructs a bilayer membrane containing 128 lipid molecules, simulates a red blood cell membrane environment, the system temperature is controlled at 310K, the Nosé-Hoover temperature control algorithm is adopted, the pressure is controlled at 1 bar, the Parrinello-Rahman pressure control algorithm is adopted, and the CHARMM36 force field is used to ensure accurate modeling of protein-lipid interactions.

[0047] Specifically, the initial polypeptide structure is obtained from the three-dimensional structure (PDB format) predicted by AlphaFold3, and the confidence threshold pLDDT is greater than 85, which ensures the biological rationality of the conformation. Before simulation, the system is preprocessed as follows: (1) energy minimization: 5000 steps of energy minimization are performed using the steepest descent method to eliminate bad contacts; (2) equilibration: 100 ps of NVT equilibration (constant particle number, volume, temperature) is performed, followed by 200 ps of NPT equilibration (constant particle number, pressure, temperature); boundary conditions: periodic boundary conditions are applied in x, y, and z directions, long-range electrostatic interactions are handled using the particle mesh Ewald (PME) method, and the cutoff distance is set to This setting captures the dynamic behavior of the polypeptide in the interaction between the aqueous solution and the cell membrane through long-term simulation combined with the membrane environment, providing high-resolution dynamic data for hemolytic prediction.

[0048] In some embodiments, the 5-dimensional features extracted in step S1012 are dynamic features related to hemolytic polypeptides, including root mean square fluctuation, solvent accessible surface area, number of hydrogen bonds, van der Waals interaction energy, and solvation free energy.

[0049] Specifically, 5-dimensional features are extracted for each residue to generate an N x 1500 x 5-dimensional basic feature tensor. The following is the calculation method for each feature.

[0050] 1) Root mean square fluctuation (RMSF)

[0051] RMSF quantifies the positional fluctuation of the Cα atom of each residue within a 100 ps time window, reflecting local flexibility and dynamic stability. Hemolytic polypeptides usually exhibit specific flexibility patterns in membrane-binding regions, and higher RMSF values indicate that surface conformation flexibility promotes membrane destruction.

[0052] The calculation formula of RMSF is: r i,j The Cα atom coordinates (x, y, z) of residue (i) at frame (j) are r i,tAverage C a coordinates of residue (i) at time step (t) (100 ps window), N f Number of frames within the time window (50 frames within 100 ps, based on 2 fs step size). Finally, normalize where μ RMSF , σ RMSF is the mean and standard deviation of the RMSF of all residues.

[0053] 2) Solvent accessible surface area (SASA)

[0054] SASA reflects the solvent exposure property of residues, the hydrophobic SASA of hemolytic polypeptides decreases when embedded in the membrane, and the hydrophilic SASA remains stable, which is related to the membrane destruction ability. The solvent exposure degree of each residue within a 100 ps time window is quantified by SASA, and the membrane destruction ability of the polypeptide is quantified. The calculation method of SASA of residue (i) is where A a is the theoretical maximum surface area of atom (a), Acc a,t is the solvent accessibility of atom a at time step t, and is normalized to obtain SASA.

[0055] 3) Number of hydrogen bonds (HBond)

[0056] The number of hydrogen bonds counts the average number of hydrogen bonds of each residue with surrounding molecules (water, lipid, other residues) within a 100 ps time window, reflecting the dynamic changes of hydrogen bond network. Hydrogen bond network affects the stability and membrane interaction of polypeptides, and hemolytic polypeptides usually form dynamic hydrogen bonds on the membrane surface to promote membrane insertion. The number of hydrogen bonds of residue i is where d i,k,j is the hydrogen bond distance between residue i and molecule k at frame j, θ i,k,j is the hydrogen bond angle, and δ is the indicator function which is 1 when the hydrogen bond condition is met, otherwise 0. N f is the number of frames within the time window. Finally, normalize where max(HBond) is the maximum number of hydrogen bonds.

[0057] 4) Van der Waals interaction energy (VDW)

[0058] Van der Waals interaction energy quantifies the non-covalent interaction strength of each residue with surrounding molecules (water, lipid, other residues), reflecting the binding characteristics of polypeptides in membrane environment. Van der Waals interaction drives the membrane embedding of polypeptides, and hemolytic polypeptides usually exhibit strong Van der Waals interaction in the hydrophobic core of the membrane. The calculation formula of Van der Waals interaction energy is ò ik , σ ik are the Lennard-Jones parameters of residue i and molecule k, r ik，jThe distance between residue i and molecule k in frame j is determined by normalization. The van der Waals interaction energy of residue i at time t is obtained.

[0059] 5) Solvation free energy (ΔG_solv)

[0060] The solvation free energy (SFE) quantifies the thermodynamic interaction of each residue with the aqueous environment, reflecting the influence of solvation on the dynamic behavior of peptides. A low SFE typically indicates that the residue is more likely to detach from the aqueous phase and embed into the hydrophobic core of the membrane, thereby enhancing the membrane-disrupting ability of hemolytic peptides. The SFE reflects the affinity of the residue for the aqueous environment. Residues with low SFEs are more likely to detach from the aqueous phase and embed into the hydrophobic core of the phospholipid bilayer, thus enhancing the peptide's membrane-disrupting ability. This characteristic is particularly crucial in the membrane insertion and pore formation mechanisms of hemolytic peptides.

[0061] The method for calculating solvation free energy is as follows: ΔG solv,i,t The solvation free energy of residue i at time step t, where R is the gas constant, has a value of R = 8.314 J·mol⁻¹. -1 ·K -1 T is the absolute temperature, T = 340 K. For thermodynamic inverse temperature, ΔU is the potential energy difference between the solvated state (solv) and the vacuum state (vac), <·> solv 、<·> vac These represent the ensemble averages over a 100 ps time window under solvation and vacuum conditions, respectively. The values ​​were normalized using the min-max method. Finally, the solvation free energy of residue i at time t is obtained.

[0062] This application utilizes molecular dynamics processes to obtain dynamic features associated with hemolytic peptides, including root-mean-square fluctuations, solvent-accessible surface area, number of hydrogen bonds, van der Waals interaction energies, and solvation free energy. For a peptide sequence of length N, to construct a dynamic feature tensor associated with hemolytic peptides, molecular dynamics (MD) simulations are used to generate an N×1500×5 dimensional matrix, characterizing the dynamic behavior of an N-length peptide sequence in a 150 ns simulation (step size 2 fs, sampling interval 100 ps, ​​T = 1500 time steps). Each residue can be represented as a 1500×5 dimensional matrix, and these features can dynamically represent the process of peptide hemolysis.

[0063] The molecular dynamics (MD) simulation is used to extract the root mean square fluctuation (RMSF), the solvent accessible surface area (SASA), the number of hydrogen bonds, the van der Waals interaction energy, and the solvation free energy, and the like dynamic characteristics, which provide a unique advantage for the prediction of the hemolytic polypeptide. These characteristics dynamically capture the atomic-level behavior of the polypeptide in the water solution and the POPC membrane environment through 150 ns high-resolution simulation (2 fs step, 100 ps sampling), generate an N x 1500 x 5 tensor, and accurately depict the dynamic evolution of the conformational flexibility, membrane surface adsorption, hydrophobic core embedding, and thermodynamic driving in the hemolysis process. Compared with static characteristics or in vitro experiments, the obtained dynamic characteristics can reflect the nonlinear interaction between the polypeptide and the membrane in real time.

[0064] For a sequence of length N, the vector representation of each residue at 1500 time points can be obtained through the above process. The present application will use a time series graph convolution network to predict whether a polypeptide is a hemolytic polypeptide, and therefore needs to construct a graph data structure to represent the polypeptide sequence corresponding to each time point at each time point according to the vector representation of the residue. Therefore, in order to make full use of the dynamic characteristics generated by the molecular dynamics (MD) simulation and capture the conformational evolution and membrane interaction behavior of the polypeptide in the hemolytic related environment, the present application constructs a time series graph based on the feature tensor X ∈ R 1500×N×5 The 1500 dynamic graph sequences are constructed.

[0065] In some embodiments, the step S102 of constructing a time series graph attention convolution network based on the dynamic characteristics of the polypeptide sequence comprises: converting the dynamic characteristics of the polypeptide sequence into a dynamic graph sequence, and constructing a time series graph attention convolution network through Z-score standardization and an adjacency matrix.

[0066] In some embodiments, the step of converting the dynamic characteristics of the polypeptide sequence into a dynamic graph sequence and constructing a time series graph attention convolution network through Z-score standardization and an adjacency matrix comprises: based on the feature tensor X ∈ R 1500 ×N×5 The 1500 dynamic graph sequences are constructed Each graph G t = (V t , E t , X t ) represents the topology and dynamic characteristics of the polypeptide at time step t, V t is a node set, E t is an edge set, and X t is a node feature matrix; Z-score standardization is performed on each feature dimension d; the standardization formula is:

[0067]

[0068] X t,i,ddenotes the d-th dimension of the i-th residue at time t, μ d denotes the mean of the d-th dimension, σ d denotes the variance of the d-th dimension; G t The corresponding adjacency matrix is:

[0069]

[0070] where A t denotes whether there is an edge between the i-th residue and the j-th residue at time t.

[0071] In some embodiments, the step S103 of fusing the graph attention network and the time sequence multi-head self-attention based on the time sequence graph attention convolutional network comprises: processing the dynamic graph sequence based on a time sequence graph attention convolutional module The graph attention network (GAT) and the time sequence multi-head self-attention (MHSA) are fused; the calculation process of the graph attention network comprises: for each time step t, a computer attention coefficient is used to measure the importance of the j-th node to the i-th node at the time step t, and the calculation formula is:

[0072]

[0073] wherein, is a learnable weight matrix used to project the input features to a new feature space F', F' is 32, is a learnable attention vector, and || represents a splicing operation of splicing two vectors, and LcakyRcLU is an activation function.

[0074] In some embodiments, the calculation process of the time sequence multi-head self-attention comprises: using a softmax function to normalize; then, through node feature updating, the features of the neighbor nodes are aggregated through the attention coefficient to update the features of the i-th node at the time step t; subsequently, time sequence multi-head self-attention calculation is performed; the graph attention network and the time sequence multi-head self-attention are combined to obtain a feature matrix obtained at different scales; the feature matrix is fused through weighted summation; the fused feature matrix is globally pooled to obtain a feature representation of the graph; and the feature representation of the graph is input to a fully connected neural network for classification.

[0075] Specifically, in order to make the attention coefficient comparable between neighbor nodes, a softmax function is used for normalization wherein is a set of neighbor nodes of the i-th node at the time step t.

[0076] Then, through node feature updating, the features of the neighbor nodes are aggregated through the attention coefficient to update the features of the i-th node at the time step t, Where σ is the ReLU activation function.

[0077] Subsequently, temporal multi-head self-attention (MHSA) computation is performed. Let the graph feature matrix after GAT processing be... Where V t This represents the number of nodes at time step t. Let X... t Projected onto the query, key, and value spaces respectively, i.e. in d k This is the dimension for both keys and queries, set to 128. The attention score is calculated as follows: For the multi-head case, there are h = 8 heads, each calculated in the same way. Finally, the outputs of all heads are concatenated and processed by a linear transformation. in, F″ is a learnable weight matrix, and F″ is the output feature dimension. Combining GAT and MHSA yields the output Z of the temporal graph attention convolutional module. t =MHSA(X t ), where X t This is the graph feature matrix after GAT processing. By combining GAT and MSHA, feature matrices at different scales can be obtained as follows: Fusion is achieved through a weighted summation method. Where β i It is a learnable weight that satisfies The fused feature matrix Perform global pooling to obtain the feature representation of the graph. Finally, the feature representation of the graph is input into a fully connected neural network for classification: y = softmax(W cls g+b cls ), where W cls ∈R C×F″ It is the weight matrix of the classifier, b cls ∈R C C=2 is the number of categories.

[0078] In some embodiments, during the process of fusing the graph attention network and the temporal multi-head self-attention in step S103, a loss function is constructed to learn the network parameters of the temporal graph attention convolutional network; the loss function includes a classification loss function and a graph structure preservation loss function.

[0079] In order to combine the classification loss and the graph structure preservation loss to better adapt to the graph classification task, a new loss function is constructed for learning the parameters of the DeepHemo network; (1) classification loss, the classification loss measures the difference between the predicted class probability distribution and the true class label, and the cross-entropy loss function is used, as follows:

[0080]

[0081] Where N is the number of samples, y n is the true label of the nth sample (0 or 1), is the probability that the model predicts the nth sample to belong to the hemolytic polypeptide.

[0082] (2) Graph structure preservation loss, by depicting the time series graph structure characteristics that adjacent nodes have similar feature representations and non-adjacent nodes have different feature representations, to strengthen the learning of network parameters.

[0083] The graph structure preservation loss function is:

[0084]

[0085] Where T is the number of time steps, |V t | is the number of nodes at time step t, and m is a positive boundary value for controlling the minimum distance between non-adjacent node features.

[0086] The final comprehensive loss function is L = λ cls L cls + λ struct L struct , where λ cls = 0.5 and λ struct = 0.5.

[0087] To verify the feasibility and technical effects of the present application, the model of the present application is compared with a plurality of existing hemolytic polypeptide prediction models, and significant performance improvement is achieved in a plurality of evaluation indicators. The evaluation indicators include accuracy (ACC), Matthews correlation coefficient (MCC), area under the curve (AUC), average accuracy (AP) and F1 value. The accuracy measures the consistency between the predicted results and the true results, the Matthews correlation coefficient considers the balance of true positives and true negatives in the classification results, the area under the curve measures the performance of the classification model, the average accuracy reflects the average value of the accuracy of the model at different thresholds, and the F1 value combines the accuracy and recall rate, and comprehensively considers the accuracy and integrity of the classification model. By comparing the existing models in these evaluation indicators, the model of the present application shows better performance, exceeding the prediction ability of the existing models. The model can more accurately predict hemolytic polypeptides and has better performance in the classification task.

[0088] In summary, the application proposes a method based on molecular dynamics (MD) simulation and time series graph attention convolution network combined with multi-scale feature fusion for predicting the hemolysis of polypeptides. Compared with traditional methods and existing models, the application DeepHemo has the following significant beneficial effects:

[0089] (1) Breakthrough dynamic feature acquisition technology.

[0090] The application uses 150ns high-resolution molecular dynamics simulation (2fs step, 100ps sampling interval), combines GROMACS simulation platform and CHARMM36 force field, constructs an SPC / E water model containing about 10,000 water molecules and a bilayer membrane system containing 128 POPC lipid molecules, and accurately simulates the human physiological environment. By extracting 5-dimensional features such as root mean square fluctuation and solvent accessible surface area for each residue, an N x 1500 x 5-dimensional dynamic feature tensor is generated. Compared with traditional static feature analysis methods, it can capture the conformational changes, thermodynamic driving and other nonlinear interaction processes of polypeptides in membrane environment in real time, providing dynamic data support with biological significance for hemolysis prediction.

[0091] (2) Innovative sequence-graph data conversion mechanism.

[0092] The application first converts the dynamic features of polypeptide sequences into 1500 time step dynamic graph sequences, each graph structure contains a node set, an edge set and a 5-dimensional node feature matrix, and fully presents the evolution process of polypeptide topological structure and dynamic characteristics. Through Z-score standardization processing and adjacency matrix construction, the problem that traditional sequence analysis cannot capture context association is effectively solved, providing structured data input conforming to biological interaction logic for subsequent graph convolution network, significantly improving the learning efficiency of the model on polypeptide hemolysis related features.

[0093] (3) Efficiently cooperative hybrid network architecture.

[0094] The application constructs a time series graph attention convolution network, which combines graph attention network (GAT) and time series multi-head self-attention (MHSA) mechanism. GAT calculates the attention coefficient between nodes, dynamically aggregates neighbor node features, and strengthens local structure information extraction; MHSA excavates sequence dependency relationship from time dimension, realizes multi-scale feature fusion. This architecture breaks through the limitations of single model, extracts and integrates features from graph data at different scales, and significantly enhances the representation ability of the model for complex biological data.

[0095] (4) Targeted optimization of loss function design.

[0096] The application proposes a composite optimization objective function containing a classification loss and a graph structure preservation loss. The classification loss uses the cross-entropy algorithm to ensure high consistency between the predicted results and the true labels; the graph structure preservation loss forces the model to learn the spatial topological relationship of the polypeptide sequence by constraining the feature distance of adjacent and non-adjacent nodes. The two losses are optimized together, so that the model can balance classification accuracy and structure information preservation during training, and significantly outperforms existing prediction models in terms of accuracy, Matthews correlation coefficient and other indicators, providing reliable technical support for polypeptide drug research and development, biological pesticide design and other fields.

[0097] The method for predicting the hemolytic property of polypeptides based on the kinetic feature graph attention fusion method provided by the application is described in detail below through specific examples.

[0098] First, the data set is obtained.

[0099] (1) Collection and preprocessing of hemolytic polypeptide sequence and non-hemolytic polypeptide sequence data

[0100] According to existing databases and literature, the application collects four main hemolytic polypeptide sequence data sets, named HemoPI-2, HemoPI-3, HAPPENN and EnDL-HemoLyt. For ease of reference, we denote them as DS1, DS2, DS3 and DS4, respectively. In order to reduce redundant data, we use CD-HIT de-duplication algorithm to remove sequences with sequence similarity of more than 50%. The specific de-duplication results are shown in Table 1.

[0101] Table 1 Hemolytic polypeptide sequence and non-hemolytic polypeptide sequence data set

[0102]

[0103] (2) Polypeptide sequence dynamic feature extraction

[0104] Specifically, 5-dimensional features are extracted for each residue to generate an N x 1500 x 5-dimensional basic feature tensor. The features include root mean square fluctuation, solvent accessible surface area, number of hydrogen bonds, van der Waals interaction energy, and solvation free energy. For a polypeptide sequence with a length of N, to construct a dynamic feature tensor related to hemolytic polypeptides, a N x 1500 x 5-dimensional matrix is generated using molecular dynamics (MD) simulation to represent the dynamic behavior of a polypeptide sequence with a length of N in 150 ns simulation (step size 2 fs, sampling interval 100 ps, T = 1500 time steps). Each residue can be represented as a 1500 x 5-dimensional matrix, and these features can dynamically represent the process of polypeptide hemolysis.

[0105] The dynamic characteristics such as root mean square fluctuation (RMSF), solvent accessible surface area (SASA), number of hydrogen bonds, van der Waals interaction energy and solvation free energy are extracted by molecular dynamics simulation, which provides a unique advantage for predicting the hemolytic of polypeptides. These characteristics dynamically capture the atomic-level behavior of polypeptides in water solution and POPC membrane environment through 150ns high-resolution simulation (2fs step, 100ps sampling), generate Nx1500x5 tensor, and accurately describe the conformational flexibility, membrane surface adsorption, hydrophobic core embedding and thermodynamic driving in the process of hemolysis. Compared with static characteristics or in vitro experiments, the dynamic characteristics obtained can reflect the nonlinear interaction between polypeptides and membranes in real time.

[0106] (3) Constructing a time series graph attention convolutional network to predict hemolytic polypeptides

[0107] For a sequence of length N, the vector representation of each residue at 1500 time points can be obtained through the above process. The time series graph convolutional network will be used to predict whether the polypeptide is a hemolytic polypeptide, so at each time point, a graph data structure is constructed according to the vector representation of the residues to represent the polypeptide sequence at each time point. Therefore, in order to make full use of the dynamic characteristics generated by molecular dynamics (MD) simulation, the conformational evolution and membrane interaction behavior of polypeptides in the hemolytic related environment are captured.

[0108] (3) Time series graph attention convolutional network combined with multi-scale feature fusion.

[0109] The time series graph attention convolutional module is used to process the graph sequence The graph attention network (GAT) and the time series multi-head self-attention (MHSA) are combined. Through the combination of GAT and MSHA, the feature matrices obtained at different scales are The fusion is performed by weighted summation where β i is a learnable weight, The fused feature matrix is globally pooled to obtain the feature representation of the graph Finally, the feature representation of the graph is input into a fully connected neural network for classification.

[0110] In order to combine the classification loss and the graph structure preservation loss to better adapt to the graph classification task, the application constructs a new loss function for learning the DeepHemo network parameters, and the loss function includes the classification loss and the graph structure preservation loss; the final comprehensive loss function is L = λ cls L cls + λ struct L struct , where λ cls = 0.5, λ struct= 0.5, through Adma optimizer, set learning rate to 0.001, batch data size to 32, through back propagation method, iteratively learn network model parameters until the loss function is less than 0.001 or the iteration number reaches 256 times, stop training, obtain the final trained model, and then evaluate the performance of the model on the training set.

[0111] (4) Model evaluation and verification of the present application

[0112] The model of the present application is compared with a plurality of existing hemolytic polypeptide prediction models, and significant performance improvement is achieved in a plurality of evaluation indicators. The evaluation indicators include specificity (Specificity, SEP), sensitivity (Sensitivity, SEN), accuracy (Accuracy, ACC), Matthews correlation coefficient (Matthews correlation coefficient, MCC), F1-score (F1), area under the curve (AUC) and average accuracy (Average precise, AP), which are defined as follows, wherein TP is the label of the model predicted as a hemolytic polypeptide, TN is the label of the model predicted as a non-hemolytic polypeptide, FP is the model predicted as a hemolytic polypeptide but actually a non-hemolytic polypeptide, FN is the model predicted as a non-hemolytic polypeptide but the actual label is a hemolytic polypeptide, p(r) is the Precision value for at the prediction threshold r,

[0113]

[0114] Through comparison with existing models in these evaluation indicators, the model of the present application exhibits better performance, exceeding the prediction ability of existing models. The model can more accurately predict hemolytic polypeptides and has better performance in classification tasks. Table 2, Table 3, Table 4, and Table 5 are comparisons of the present application with HemoPI, HemoPred, HLPpred-Fuse, HAPPENN, HemoPImod, HemoNet, AMPDeep, and EnDL-HemoLyt methods on D1, D2, D3, and D4 data sets. As can be seen from Table 2, Table 3, Table 4, and Table 5, the present application has better performance in SPE, SEN, ACC, MCC, F1, AUC, and AP indicators. In addition, Figure 2 , Figure 3AUROC curve and PRROC curve of the present application and other methods are respectively shown, and the curve of the present application is higher than and wrapped in other curves, indicating that the present application has stronger ability to predict hemolytic polypeptides and non-hemolytic polypeptides. Moreover, in order to better show the prediction probability value distribution of the comparison method, the present application compares the prediction probability of DeepHemo with that of each comparison method, and the results show that DeepHemo of the present application exhibits smaller prediction errors, again showing that the present application can well distinguish hemolytic polypeptides and non-hemolytic polypeptides.

[0115] Table 2 Comparison of the present application and other methods in the DS1 data set

[0116]

[0117]

[0118] Table 3 Comparison of the present application and other methods in the DS2 data set

[0119]

[0120] Table 4 Comparison of the present application and other methods in the DS3 data set

[0121]

[0122] Table 5 Comparison of the present application and other methods in the DS4 data set

[0123]

[0124]

[0125] According to the above technical solution, the embodiment of the present application provides a method for predicting polypeptide hemolysis based on a graph attention fusion method of kinetic characteristics, which includes the following steps: first, obtaining a data set, and based on the data set, constructing a polypeptide sequence dynamic feature; the polypeptide sequence dynamic feature includes an N×1500×5-dimensional feature tensor, and N is the number of residues in the polypeptide sequence; then, based on the polypeptide sequence dynamic feature, a time series graph attention convolutional network is constructed; finally, based on the time series graph attention convolutional network, a graph attention network and a time series multi-head self-attention are fused to realize multi-scale feature extraction and fusion, and classification prediction is completed through global pooling and full-connection neural network. The present application realizes efficient prediction of hemolytic polypeptides by constructing a polypeptide sequence dynamic feature tensor through molecular dynamics simulation technology, combining graph data structure conversion and a time series graph attention convolutional network, and can be widely applied in the fields of drug development, biological pesticide design and biomedical toxicity evaluation.

[0126] The hemolytic polypeptide prediction method DeepHemo proposed in the present application is based on molecular dynamics simulation, combined with deep learning and graph data processing technology. First, the GROMACS simulation platform is used to simulate the polypeptide in the human physiological simulation environment for 150 ns high-resolution molecular dynamics simulation, with a 2 fs step and a 100 ps sampling interval. The 5-dimensional dynamic features of each residue are extracted to construct an N x 1500 x 5-dimensional feature tensor, which records the dynamic evolution of the polypeptide in the membrane environment. Then the feature tensor is converted into 1500 dynamic graph sequences, which are standardized by Z-score and constructed into an adjacency matrix to provide structured data for graph convolution network. Then the time series graph attention convolution network is used to fuse the graph attention network (GAT) and the time series multi-head self-attention (MHSA) to realize multi-scale feature extraction and fusion. Finally, the global pooling and fully connected neural network are used to complete the classification prediction. At the same time, a composite loss function containing classification loss and graph structure preservation loss is designed to optimize the model training process. The method fills the gap in the prior art and has important practical application value.

[0127] Those skilled in the art can understand that the above-mentioned embodiments are specific examples for implementing the present application, and in actual application, various changes can be made in form and detail without departing from the spirit and scope of the present application. Any person skilled in the art can make changes and modifications without departing from the spirit and scope of the present application, therefore the protection scope of the present application should be limited by the scope defined in the claims.

Claims

1. A method for predicting peptide hemolytic activity using graph attention fusion based on kinetic characteristics, characterized in that, Includes the following steps: Obtain a dataset and construct dynamic features of peptide sequences based on the dataset; the dynamic features of peptide sequences include an N×1500×5-dimensional feature tensor, where N is the number of residues in the peptide sequence; Based on the dynamic features of peptide sequences, a temporal graph attention convolutional network is constructed. Based on a temporal graph attention convolutional network, a graph attention network and a temporal multi-head self-attention network are fused to achieve multi-scale feature extraction and fusion. Classification prediction is then completed through global pooling and a fully connected neural network.

2. The method for predicting peptide hemolytic activity using graph attention fusion based on kinetic characteristics according to claim 1, characterized in that, Acquire a dataset and, based on the dataset, construct dynamic features of peptide sequences, including: Using the GROMACS simulation platform, high-resolution molecular dynamics simulations of peptides were performed at 150 ns in a human physiological simulation environment. With a step size of 2fs and a sampling interval of 100ps, 5-dimensional features of each residue were extracted to construct an N×1500×5-dimensional feature tensor to record the dynamic evolution of the peptide in the membrane environment. The characteristic tensor is represented as: X e R 1500×N×5 Where X is the feature tensor and R is a real matrix.

3. The method for predicting peptide hemolytic activity using graph attention fusion based on kinetic characteristics according to claim 2, characterized in that, The simulated environment employs a mixture of aqueous solution and phospholipid bilayer membrane; wherein, The aqueous solution uses the SPC / E water model and contains approximately 10,000 water molecules to simulate the physiological solution environment. The phospholipid bilayer membrane was constructed using 1-palmitoyl-2-oleoyl-sn-glycerol-3-phosphocholine lipids, forming a bilayer membrane containing 128 lipid molecules to simulate the environment of the erythrocyte membrane. The system temperature was controlled at 310K using the Nosé-Hoover temperature control algorithm. The pressure was controlled at 1 bar, using the Parrinello-Rahman pressure control algorithm and the CHARMM36 force field to ensure accurate modeling of protein-lipid interactions.

4. The method for predicting peptide hemolytic activity using graph attention fusion based on kinetic characteristics according to claim 2, characterized in that, The extracted five-dimensional features are dynamic features related to hemolytic peptides, including root mean square fluctuation, solvent-accessible surface area, number of hydrogen bonds, van der Waals interaction energy, and solvation free energy.

5. The method for predicting peptide hemolytic activity using graph attention fusion based on kinetic characteristics according to claim 1, characterized in that, Based on the dynamic features of peptide sequences, a temporal graph attention convolutional network is constructed, including: The dynamic features of peptide sequences are transformed into dynamic graph sequences, and a temporal graph attention convolutional network is constructed by Z-score normalization and adjacency matrix.

6. The method for predicting peptide hemolytic activity using graph attention fusion based on kinetic characteristics according to claim 5, characterized in that, The dynamic features of peptide sequences are transformed into dynamic graph sequences. A temporal graph attention convolutional network is constructed using Z-score normalization and adjacency matrix, including: Based on the feature tensor X∈R in the dynamic features of polypeptide sequences 1500×N×5 Construct 1500 dynamic graph sequences Among them, each graph G t =(V t E t ,X t Characterizing the topological structure and dynamic properties of the peptide at time step t, V t For node set, E t For edge set, X t The node feature matrix; Z-score normalization is performed on each feature dimension d; The standardized formula is: Among them, X t,i,d μ represents the d-th dimension of residue i at time t. d The mean of the d-th dimension, σ d Let d be the variance in each of the d dimensions; G t The corresponding adjacency matrix is: Among them, A t (i,j) indicates whether there is a boundary between residue i and residue j at time t.

7. The method for predicting peptide hemolytic activity using graph attention fusion based on kinetic characteristics according to claim 1, characterized in that, Based on a temporal graph attention convolutional network, this method fuses graph attention networks with temporal multi-head self-attention, including: Processing dynamic graph sequences based on a temporal graph attention convolution module Fusion of graph attention networks and temporal multi-head self-attention; The computation process of graph attention networks includes: For each time step t, the computer attention coefficient The formula used to measure the importance of node j to node i at time step t is: in, F' is a learnable weight matrix used to project the input features onto a new feature space F', where F' is 32. It is a learnable attention vector, || represents the concatenation operation, which concatenates the two vectors, and LcakyRcLU is the activation function.

8. The method for predicting peptide hemolytic activity using graph attention fusion based on kinetic characteristics according to claim 7, characterized in that, The computational process of temporal multi-head self-attention includes: Use the softmax function for normalization; Then, by updating the node features, the features of neighboring nodes are aggregated through the attention coefficient, and the features of node i at time step t are updated. Subsequently, temporal multi-head self-attention computation is performed; By combining graph attention networks and temporal multi-head self-attention, feature matrices are obtained at different scales. The feature matrices are fused using a weighted summation method; The fused feature matrix is ​​then subjected to global pooling to obtain the feature representation of the graph. The feature representation of the graph is input into a fully connected neural network for classification.

9. The method for predicting peptide hemolytic activity using graph attention fusion based on kinetic characteristics according to claim 1, characterized in that, In the process of fusing graph attention network and temporal multi-head self-attention, a loss function is constructed to learn the network parameters of the temporal graph attention convolutional network; The loss functions include a classification loss function and a graph structure preservation loss function.

10. The method for predicting peptide hemolytic activity using graph attention fusion based on kinetic characteristics according to claim 9, characterized in that, The classification loss function uses the cross-entropy loss function, as shown below: Where N is the number of samples, y n It is the true label of the nth sample. It is the probability that the nth sample is a hemolytic peptide, as predicted by the model. The graph structure retains the loss function, as shown below: Where T is the number of time steps, |V t | is the number of nodes at time step t, and m is a positive boundary value used to control the minimum distance between features of non-adjacent nodes; The loss function is: L = λ cls L cls +λ struct L struct , where λ cls =0.5, λ struct =0.5.

Images