Vascular anti-compression peptide recognition method combining prior knowledge embedding and multi-view enhancement

Abstract

The invention relates to the technical field of anti-compression peptides, in particular to a blood vessel anti-compression peptide recognition method combining prior knowledge embedding and multi-view enhancement, which comprises the following steps: constructing an ACEIP sequence data set, a nonACEIP sequence data set and an ACE sequence data set, and preprocessing; constructing feature vector codes of the peptide sequences by utilizing the priori knowledge similarity and the molecular docking adaptation degree; and taking coded vectors of ACEIP and nonACEIP sequences as input, and constructing a Deep4ACE model which comprises a feature enhancement module, a model learning module and a model prediction module and is used for predicting the ACEIP and the nonACEIP. The method solves the problems that an existing method utilizes sequence similarity search, molecular modeling and the like, focuses on a primary structure of the peptide, neglects interaction between a space structure and a receptor, and is high in false positive rate of the identified compression-resistant peptide.

Description

Technical Field

[0001] The present invention relates to the technical field of anti-stress peptides, and in particular to a vascular anti-stress peptide identification method combining prior knowledge embedding and multi-view enhancement. Background Art

[0002] In recent years, active ingredients in food, especially food anti-angiogenic peptides, have attracted extensive attention due to their potential anti-angiogenic effects.

[0003] ACE inhibitory peptides (Angiotensin-Converting Enzyme Inhibitory Peptides, ACEIP) as a class of food vascular antidrug peptides, are angiotensin converting enzyme inhibitory peptides, a class of biologically active peptide substances, which can regulate blood pressure by inhibiting the activity of angiotensin converting enzyme (ACE); ACEIP competitively binds to the active site of ACE, thereby inhibiting the enzymatic activity of ACE, reducing the production of angiotensin II, increasing the level of bradykinin, and ultimately leading to vasodilation and lowering blood pressure.

[0004] At present, there are still many challenges in predicting and screening peptides with vascular anti-stress effects. First, compared with other bioactive peptides, the research on food vascular anti-stress peptides started late, and the basic research is relatively weak, resulting in limited databases and information resources that can be used for prediction. Second, at present, the main methods for predicting food vascular anti-stress peptides rely on sequence similarity searches, molecular modeling, etc. These methods often focus on the primary structure of the peptide, while ignoring its spatial structure and interaction with the receptor. In addition, due to the complex mechanism of action of vascular anti-stress peptides, existing prediction models such as the literature A Web Server for Predicting Antihypertensive Activity of Peptides and A Deep Ensemble Predictor for Identifying Anti-Hypertensive Peptides Using Pretrained Protein Embedding do not consider the mechanism of action information of anti-stress peptides, resulting in a high false positive rate of identified anti-stress peptides. Summary of the invention

[0005] In view of the shortcomings of existing methods, the present invention solves the problems that existing methods such as sequence similarity search and molecular modeling focus on the primary structure of peptides, ignore the spatial structure and interaction with receptors, and have a high false positive rate of identified stress-resistant peptides.

[0006] The technical solution adopted by the present invention is: the vascular anti-compression peptide identification method combining prior knowledge embedding and multi-view enhancement includes the following steps:

[0007] Step 1: Construct ACEIP sequence, nonACEIP sequence and ACE sequence data sets and perform preprocessing;

[0008] As a preferred embodiment of the present invention, the preprocessing includes: removing peptide sequences of amino acids whose similarity is greater than a set threshold and whose length is less than a set threshold.

[0009] Step 2: construct feature vector encoding of peptide sequence using prior knowledge similarity and molecular docking fitness;

[0010] As a preferred embodiment of the present invention, encoding the feature vector using prior knowledge similarity includes:

[0011] Step 211: Encode ACE, ACEIP, and nonACEIP sequences into V using the ESM-2 model ACE 、V ACEIP and V non-ACEIP , calculate the Pearson similarity between ACE and ACEIP, ACE and nonACEIP, and get the similarity matrix SM ACE-ACEIP , SM ACE-nonACEIP ;

[0012] Step 222: Use the ESM-2 model to obtain the characteristic vector of the ACEIP or nonACEIP or ACE sequence as Mat seq , along the sequence length L dimension of Mat seq Find the average and get the eigenvector V seq ;

[0013] Step 213: Based on SM ACE-ACEIP and SM ACE-nonACEIP , construct GS ACE-ACEIP Figure and GS ACE-nonACEIP picture;

[0014] Step 214: Use the ESM-2 model to analyze GS ACE-ACEIP and GS ACE-nonACEIP Node encoding.

[0015] As a preferred embodiment of the present invention, step 214 includes:

[0016] At GS ACE-ACEIP In the process, random walk sampling graph nodes are used to obtain the path, and the nodes in the path are ACEIP sequence and ACE sequence; the ESM-2 model is used to encode the corresponding node sequence in the path into a vector;

[0017] The Skip-Gram model is used to predict the location of the context node, learn the representation of each node, and obtain the vector representation of ACEIP and nonACEIP.

[0018] As a preferred embodiment of the present invention, the molecular docking fitness feature vector encoding includes:

[0019] Step 221: molecularly dock ACEIP and nonACEIP with ACE to obtain docking fitness score matrix DM ACE-ACEIP DM ACE-nonACEIP ;

[0020] Step 222: According to DM ACE-ACEIP and DM ACE-nonACEIP Construction of molecular docking map GD ACE-ACEIP and GD ACE-nonACEIP ; Get the encoding vectors of ACEIP and nonACEIP.

[0021] Step 3: Using the encoded vectors of ACEIP and nonACEIP sequences as input, a Deep4ACE model including a feature enhancement module, a model learning module, and a model prediction module is constructed to predict ACEIP and nonACEIP.

[0022] As a preferred embodiment of the present invention, the feature enhancement module includes:

[0023] Step 311, obtain sequence key point enhanced features: use a certain size window along the input vector V in The first dimension slides; let ACEIP's sliding window Sliding window of nonACEIP by is a positive sample, For negative samples, the prediction probability P is obtained using the RF classifier. i ACEIP , P i nonACEIP ; P i ACEIP and P i nonACEIP As the feature representation of window i; PowerShap is used to filter the first several key windows;

[0024] Step 312: Convert sequence information into position matrix information: traverse each site of the encoding vector, and the first dimension value V of each site vector i 0 ,go through Convert to X-axis coordinate position X i ; The second dimension V of each site vectori 1 ,go through Convert to Y coordinate position Y i ; Set the coordinates (X i , Y i ) is set to 1; traverse each position of the sequence encoding vector and build a position matrix.

[0025] As a preferred embodiment of the present invention, the model learning module includes:

[0026] Sequence keypoint enhancement features Through a fully connected network Get Features Will Each element is mapped to the key window by multiplication to obtain the feature Output local particle feature vectors through LSTM, TDFC, and TCN respectively Context feature information vector Global feature information vector Calculation using self-attention mechanism and The synergy of the output vector After vector expansion, it is input into several fully connected networks and the output vector

[0027] Position Matrix Vector After 3 layers of convolutional network, the output is the feature Will Through the pooling of the average pooling layer, the first two dimensions of the pooled matrix are flattened to obtain the feature The characteristics Convert to Graph

[0028] As a preferred embodiment of the present invention, the feature Convert to Graph include:

[0029] The characteristics Consider it as a vector of several dimensions, calculate the Pearson similarity between the vectors, and set the connection threshold to T 1 , the Pearson similarity between vectors and T 1 Compare and determine whether the channel vectors are connected; then use the graph features The nodes in the graph are fused into dimensional vectors through a 3-layer graph convolutional network and a global maximum pooling operation.

[0030] As a preferred embodiment of the present invention, the convolution kernel size of TCN is set to 5, the step size is 1, the expansion coefficient is 2, the number of layers is set to 3, and the number of channels is set to 128; the output dimension of the LSTM network is 128; and the output dimension of the TDFC network is set to 128.

[0031] As a preferred embodiment of the present invention, the prediction module includes 3 layers of fully connected networks and a sigmoid function, and the output of each layer of the fully connected network is 512, 125, 1.

[0032] Beneficial effects of the present invention:

[0033] 1. The present invention constructs the Deep4ACE model, which predicts ACEIP by building a deep learning network from three aspects: injecting prior knowledge, building graph network embedding features, and multi-view enhancement information. The prediction accuracy of the model is significantly better than that of the existing model;

[0034] 2. Based on the mechanism of action of ACEIP and ACE, the sequence similarity and molecular structure binding degree information of ACE and ACEIP, ACE and nonACEIP are injected. This information is used for the prediction of anti-stress peptides for the first time. Compared with the model that only considers sequence information, it has better model interpretation and accuracy;

[0035] 3. Based on prior knowledge, we construct the ACE and ACEIP graph networks, as well as the nonACE and ACEIP graph network information. We use the random walk and ESM-2 sequence encoding methods to obtain the node context sequence encoding information in the graph. We use the Skip-Gram training strategy to improve the ACE and ACEIP vector encoding information.

[0036] 4. Use the window sliding method to convert window information into probability information, and use the feature selection method to obtain useful information of key windows, avoid information redundancy and improve the robustness of the model;

[0037] 5. Use coordinate transformation strategy to map sequence vectors into two-dimensional space information, obtain the correlation of internal information in sequence representation information, and improve the ability to mine latent correlation information of the model;

[0038] 6. Use long short-term memory network, time-distributed fully connected network, and temporal convolutional network to fully explore the potential fusion information of sequence encoding information from multiple views and improve model accuracy;

[0039] 7. Calculate the correlation of the output vectors of the middle layer of the two-dimensional convolutional neural network, build a graph network, and use the graph convolution method to mine the joint information of the output vectors of the middle layer of the two-dimensional convolutional neural network, enhance the information fusion of the output vectors, and improve the accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is the vascular anti-stress peptide recognition logic frame of the present invention that combines prior knowledge embedding and multi-view enhancement;

[0041] Figure 2 This is the AUROC comparison chart of Deep4ACE and existing models;

[0042] Figure 3 This is a PRROC comparison chart of Deep4ACE and existing models;

[0043] Figure 4 This is a comparison chart of the MCC value distribution of Deep4ACE and existing models. DETAILED DESCRIPTION

[0044] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0045] like Figure 1 As shown in FIG, the vascular anti-stress peptide identification method based on joint prior knowledge embedding and multi-view enhancement includes the following steps:

[0046] Step 1: Construct ACEIP sequence, nonACEIP sequence and ACE sequence data sets and perform preprocessing;

[0047] The present invention collects and identifies angiotensin converting enzyme inhibitory peptides from the DFBP database; the access website of the DFBP database is http: / / www.cqudfbp.net; DFBP contains 6,818 peptide sequences of 31 types from 1,065 references, ensuring the comprehensiveness and diversity of the data; DFBP contains 1,961 ACEIPs and 4857 bioactive peptides from food; ACEIPs are extracted and an ACEIP data set is created; other biological peptide sequences are collected at the same time to create a nonACEIP data set;

[0048] Preprocessing included: using the CD-HIT tool to remove peptide sequences with a similarity greater than 0.7, and deleting extremely similar peptide sequences with a length of less than 2 amino acids; finally, the ACEIP dataset contained 1,470 peptides, and the nonACEIP dataset contained 1,625 peptide sequences; the training dataset and the test dataset were formed by merging the ACEIP dataset and the nonACEIP dataset and randomly dividing the dataset in a ratio of 8:2.

[0049] Step 2: construct feature vector encoding of peptide sequence using prior knowledge similarity and molecular docking fitness;

[0050] Since the peptide sequence is a character sequence, it needs to be converted into a digital vector as the input of the model;

[0051] Step 21, using prior knowledge similarity feature vector encoding;

[0052] Since the protein large language model contains the sequence and structural domain information of proteins and polypeptides, the polypeptide sequence can be better encoded into a digital vector; at the same time, since ACEIP can inhibit ACE activity and reduce the generation of angiotensin II; angiotensin II is a strong vasoconstrictor that can cause vasoconstriction and increase peripheral resistance, thereby raising blood pressure; by inhibiting ACE, it helps to relax blood vessels and lower blood pressure; based on the operating mechanism of ACEIP, the present invention constructs an ACEIP and ACE interaction network, a nonACEIP and ACE interaction network by adding ACE knowledge, and adopts a graph embedding learning method to learn the vector representation of ACEIP and nonACEIP.

[0053] The prior knowledge similarity feature vector encoding includes:

[0054] Step 211, download 275 ACE sequences from the UniProt database, and encode ACE into a feature vector V using the ESM-2 model ACE At the same time, the ACEIP sequence and nonACEIP sequence are encoded into feature vectors V using the ESM-2 model. ACEIP and V non-ACEIP ; Calculate the Pearson similarity between each ACE and ACEIP to obtain the similarity matrix SM ACE-ACEIP ; Calculate the Pearson similarity between each ACE and nonACEIP to obtain the similarity matrix SM ACE-nonACEIP ;

[0055] Step 212, the process of encoding ACEIP, nonACEIP and ACE sequence by ESM-2 model is as follows: Use ESM-2 model to obtain the feature vector of ACEIP or nonACEIP or ACE sequence as Mat seq , whose dimension is L*1024, L is the sequence length, 1024 is the dimension for encoding each amino acid. Due to the different lengths of different sequences, in order to unify the calculation, Mat is calculated along the L dimension. seq Find the average and get the vector representation V of the three sequences seq ;

[0056] Step 213: Based on the matrix SM ACE-ACEIP and the matrix SM ACE-nonACEIP , construct GS ACE-ACEIP Figure and GS ACE-nonACEIP Figure: GS ACE-ACEIP and GSACE-nonACEIP The construction principle is to maintain the connectivity of the graph while making the graph as sparse as possible; first, a small initial threshold T is selected so that SM ACE-ACEIP All nodes in the GS participate in ACE-ACEIP The construction of SM ACE-nonACEIP All nodes in the GS participate in ACE-nonACEIP Then, the connection threshold T is gradually reduced, and the graph segmentation algorithm Girvan-Newman is used to segment the graph so that the number of segmented subgraphs does not increase while maintaining the connectivity of the graph. Finally, the final threshold T is used to construct GS ACE-ACEIP and GS ACE-nonACEIP .

[0057] Step 214: Use the ESM-2 model to analyze GS ACE-ACEIP Figure and GS ACE-nonACEIP The node coding in the figure implements the node coding of ACEIP and nonACEIP;

[0058] First at GS ACE-ACEIP A random walk method is used to sample nodes in the graph and obtain the path, which is composed of the sampled nodes in the graph.

[0059] Secondly, the nodes in the path are ACEIP sequences and ACE sequences, and the ACEIP sequences and ACE sequences are obtained; the ESM-2 model is used to encode the corresponding node sequences in the path into vectors;

[0060] Finally, the Skip-Gram model is used to learn the representation of each node by predicting the position of the context node, thereby obtaining the vector representation of ACEIP. ACE-nonACEIP The same process is used to obtain the vector representation of nonACEIP. In this embodiment, the vector representation of the node is 256-dimensional, so that the ACEIP sequence and the nonACEIP sequence are encoded into digital vectors of 256×1 dimensions respectively.

[0061] Step 22, using molecular docking fitness feature vector encoding;

[0062] Molecular docking is used to predict the optimal binding mode between two or more molecules (such as proteins and small molecules) and to identify potential binding sites on the surface of proteins. By simulating the interaction between ACEIP, ACE sequence and ACE, new binding regions can be revealed and more effective ACE inhibitors can be screened. Molecular docking is used to quantitatively evaluate the binding affinity between ACEIP and ACE, and GD is constructed based on the binding affinity. ACE-ACEIP and GD ACE-nonACEIP,The graph embedding method is used to learn the molecular docking relationship information and encode ACEIP and nonACEIP into vectors.

[0063] The feature vector encoding of molecular docking fitness includes:

[0064] Step 221: molecular docking is performed on ACEIP and nonACEIP in the data set with 275 ACEs downloaded from the UniProt database. The molecular docking is performed using AutoDock Vina software to obtain the docking fitness score matrix DM between ACEIP and ACE. ACE-ACEIP The docking fitness score matrix DM of nonACEIP and ACE ACE-nonACEIP ;

[0065] Step 222: According to DM ACE-ACEIP and DM ACE-nonACEIP , using a graph construction method similar to the encoding method of the prior knowledge similarity feature vector to construct GD ACE-ACEIP and GD ACE-nonACEIP , respectively represent the molecular docking graphs of ACE and ACEIP, and the molecular docking graphs of ACE and nonACEIP; similarly, the graph node embedding learning method in step 214 is used to obtain the encoding vectors of ACEIP and nonACEIP, and the encoding vector dimensions of ACEIP and nonACEIP are set to 256×1, respectively, and the encoding vectors integrate the molecular docking information of ACEIP and ACE and nonACEIP and ACE, respectively;

[0066] Based on the prior knowledge similarity feature vector encoding method and the molecular docking fitness feature vector encoding method, an ACEIP can be encoded into two 256×1 dimensional vectors. By merging these two vectors, an ACEIP can finally be encoded into a 256×2 dimensional vector; similarly, a nonACEIP can also be encoded into a 256×2 dimensional vector.

[0067] Step 3: Using the encoded vectors of ACEIP and nonACEIP sequences as input, a Deep4ACE model is constructed to distinguish ACEIP from nonACEIP.

[0068] The Deep4ACE model consists of three parts: feature enhancement module, model learning module, and model prediction module.

[0069] Step 31: Feature enhancement module construction includes:

[0070] Step 311, using the feature enhancement module to obtain sequence key point enhancement features; by identifying the specific key points in the sequence, the key information of the sequence is enhanced; for an input data vector of Deep4ACE, the dimension is 256×2, defined as V in , using a window size of 32, along V in The first dimension of is slid with a step size of 1, and 225 windows are obtained; the above sliding strategy is used for each sequence in the ACEIP and nonACEIP datasets to obtain the sliding window of ACEIP and the sliding window of nonACEIP respectively; define is the i-th sliding window of ACEIP sequence features, is the i-th sliding window of nonACEIP sequence features, is a positive sample, is a negative sample, and a random forest (RF) is used as the classifier to train the meta-model RF at the i-th position. i , then Input to RF i Get The predicted probability P i ACEIP At the same time, Input to RF i Get The predicted probability P i nonACEIP ; P i ACEIP and P i nonACEIP As the feature representation of window i; repeat the above process to obtain the representation of all windows, and then use the feature selection method PowerShap to screen the top 10 important windows to obtain the key windows; these key windows can well represent the difference between ACEIP and nonACEIP. By extracting the key difference windows and reconstructing the input data, a polypeptide sequence representation vector can be replaced by the key window vector. The size of the replaced vector is 32×10, where 32 corresponds to the window length and 10 is 10 important windows;

[0071] Step 312: converting the sequence information into position matrix information;

[0072] By correlating the information of each position in the sequence, the key information of sequence position association is enhanced; the input is a 256×2 vector, and the dimension corresponding to each sequence position is 2, which can be divided into X and Y axes; the vector V corresponding to each position i 0 As the scale of the X-axis, V i 1As the scale of the Y axis, i is the i-th position in the sequence, and a point distribution is formed by mapping the 256×2 vector to the coordinate axis;

[0073] First, traverse each position of the encoding vector, and the first dimension value V of each position vector i 0 ,go through Convert to X-axis coordinate position X i , v 0 is the first dimension data of the sequence encoding vector; the second dimension V of each site vector i 1 ,go through Convert to Y coordinate position Y i , v 1 is the second dimension data of the sequence encoding vector; and the coordinates (X i ,Y i ) is set to 1; a 32×32 matrix is ​​constructed by traversing each position of the sequence encoding vector;

[0074] Step 32: construct a model learning module;

[0075] Through the feature enhancement module, two new input features are obtained, namely, the sequence key point enhancement feature and the sequence information converted into position matrix information; the sequence key point enhancement feature corresponds to the first branch of the model learning module, and the sequence information converted into position matrix information corresponds to the second branch of the model learning module;

[0076] Step 321: The input of the first branch is The corresponding sequence key point enhanced features have a dimension of 32×10, where 32 corresponds to the window length and 10 is the first 10 key windows; first, Through a fully connected network get The dimension of is 1×10; secondly, Each element is mapped to 10 key windows by multiplication Its dimension is 32×10; the mapping process is an adaptive learning method, which can further determine the importance of each key window; secondly, The long short-term memory network (LSTM), time distributed fully connected network (TDFC), and temporal convolutional network (TCN) are used respectively; the convolution kernel size of TCN is set to 5, the step size is 1, the expansion coefficient is 2, the number of layers is set to 3, and the number of channels is set to 128; the output dimension of the LSTM network is 128; the output dimension of the TDFC network is set to 128; the output vector of the TCN network is Its dimension is 32×128; the LSTM network output vector is Its dimension is 32×128; the TDFC network output vector is Its dimensions are 32×128; Corresponding to the local particle characteristics of the input data; Corresponding to the context feature information of the input data; Corresponding to the global feature information of the input data; secondly, using the self-attention mechanism calculate and The synergy of the three types of input information is increased, and the output vector is Its dimension is 32×128; finally, The vector is expanded and input into three fully connected networks, where the intermediate layer dimensions are 1024, 768 and 512 respectively, and the output vector Its dimension is 512.

[0077] Step 322: Input of the second branch The corresponding sequence information is converted into position matrix information features with a dimension of 32×32; the graph network model of the second branch combines convolutional neural network and graph convolutional network; first, After three layers of convolutional networks, the neighborhood relationship of the position matrix is ​​fully explored; the convolution kernel sizes of the convolutional network are set to 7, 5, and 3 respectively, the number of channels is set to 32, 64, and 128 respectively, and the step size is uniformly set to 1; after three layers of convolutional networks, the output of the network is The dimension is 20×20×128. Through the pooling of the average pooling layer, the pooling size is set to 2, and the size of the pooled matrix is ​​10×10×128. The first two dimensions of the pooled matrix are flattened to obtain a vector of dimension 100×128. Then the innovative Convert to Graph The conversion process is:

[0078] Will Consider 128 vectors of 100 dimensions. By calculating the Pearson similarity between vectors, the connection threshold is set to T. 1 In this embodiment, T 1 = 0.5, when the Pearson similarity between vectors is greater than T 1 Indicates that the two channel vectors are connected, otherwise they are not connected; then Through a three-layer graph convolutional network, the spatial features between channels are learned. The node representation dimension in the graph convolution is set to 512. Finally, the nodes in the graph are fused into a vector with a dimension of 512 through a global maximum pooling operation.

[0079] Step 33: construct a prediction module;

[0080] The attention function of the prediction module is to and Merger Its dimension is 1024, Through a 3-layer fully connected network, the output of each layer is 512, 125, 1, and the output Its size is 1; then After the sigmoid function, its threshold range is converted to 0 to 1. When the model prediction value is greater than the threshold T 3 In this embodiment, T 3 =0.5, indicating that the predicted result is an anti-stress peptide; when the model prediction value is less than 0.5, it means that the predicted result is not an anti-stress peptide;

[0081] The loss function of the model is cross entropy, expressed as Among them, y is the true label, is the predicted label, N is the number of samples, the model optimizer is set to Adam, the learning rate is set to 0.0001, and the momentum is set to 0.9.

[0082] Experimental data:

[0083] In order to prove the effectiveness of the features proposed in the present invention, the prior knowledge similarity feature vector encoding method, the molecular docking fitness feature vector encoding method and the combination performance of the two encoding methods are compared; the sensitivity (SE), specificity (SP), accuracy (ACC), F1-score (F1), Matthews correlation coefficient (MCC), area under the curve (AUC) and average accuracy (AP) are used to evaluate the performance of each encoding method; the test results in the test set are shown in Table 1. It can be seen from Table 1 that the combination of the prior knowledge similarity feature vector encoding method and the molecular docking fitness feature vector encoding method has better prediction results, indicating that the two encoding methods have a synergistic effect.

[0084] Table 1 Performance comparison of two encoding methods

[0085]

[0086]

[0087] Comparison of Deep4ACE with other existing methods:

[0088] The performance of Deep4ACE of the present invention was compared with AHTpin, AHPred-DE, PAAP, pLM4ACE, Ensemble-AHTPpred and AHPNG; AHTpin used four types of descriptors, namely amino acid composition, atomic composition, chemical descriptor and G-scales descriptor, as input features, and used support vector machine classification model as classifier to build a model to predict anti-stress peptides; AHPred-DE ​​used UniRep to pre-train protein embedding, One-Hot feature, K-mer feature, and used a deep ensemble model to predict anti-stress peptides; PAAP used amino acid composition, dipeptide composition, pseudo amino acid composition as encoding features, and used random forest pre-training to predict anti-stress peptides. pLM4ACE uses the ESM-2 large prediction model to encode features, and uses the support vector machine classification model to predict anti-stress peptides; Ensemble-AHTPpred uses amino acid composition, quasi-sequence order descriptors, and secondary structure conformation features, and integrates three machine learning algorithms: random forest, support vector machine, and extreme gradient boosting, and integrates the advantages of each basic model through weighted voting to predict anti-stress peptides; AHPCG uses dipeptide frequency, physicochemical characteristics of amino acids, and uses convolutional neural networks and gated recurrent units to build a deep learning model for identifying anti-stress peptides; Deep4ACE is trained on the training set compared with other existing comparison methods, and its performance is evaluated on the test set.

[0089] The comparison results of Deep4ACE and other methods are shown in Table 2. The AUROC curve comparison is as follows: Figure 2 As shown, the APROC curve is compared with Figure 3 As shown in Table 2, it can be seen that the Deep4ACE model surpasses the comparison model in multiple indicators such as SE, SP, ACC, F1, MCC, AUC, and AP. The MCC index surpasses by 7.61% to 15.06%, the AUC index surpasses by 2.07% to 6.31%, and the AP index surpasses by 3.52% to 12.93%. The surpassing of these comprehensive indicators shows that Deep4ACE does have better performance in predicting anti-stress peptides. In addition, Figure 2 AUROC curve of Deep4ACE and Figure 3 The APROC curves of Deep4ACE in the figure all wrap around the curves of other comparison methods, further illustrating that Deep4ACE has better accuracy in predicting stress-resistant peptides.

[0090] Table 2 Comparison of Deep4ACE evaluation indicators with other comparison models

[0091] SE SP ACC F1 MCC AUC AP AHTpin 0.8125 0.8644 0.8454 0.7938 0.6707 0.9347 0.8902 AHPred-DE 0.7760 0.8554 0.8263 0.7660 0.6281 0.9218 0.8733 PAAP 0.7760 0.8734 0.8377 0.7780 0.6502 0.9290 0.8815 pLM4ACE 0.7552 0.9096 0.8530 0.7901 0.6792 0.9435 0.9064 Ensemble-AHTPpred 0.7656 0.8554 0.8225 0.7596 0.6190 0.9046 0.8335 AHPCG 0.7447 0.8584 0.8167 0.7486 0.6045 0.9011 0.8124 Deep4ACE 0.8177 0.9277 0.8874 0.8418 0.7553 0.9642 0.9416

[0092] In addition, to further demonstrate the performance of Deep4ACE on the entire dataset, the training set and test set are randomly divided by different seeds with a ratio of 8:2. The same training set is used to compare the MCC values ​​of Deep4ACE and the comparison method in the test set, and the MCC value of the comparison method is used as the horizontal axis, and the MCC value of Deep4ACE is used as the vertical axis. The coordinate points are drawn on the coordinate axis. When the coordinate point is on the right diagonal line, it means that the performance of Deep4ACE is better. When the coordinate point is below the right diagonal line, it means that the performance of the comparison method is better. Repeat the above method 100 times to observe the performance of Deep4ACE. Figure 4 It can be found that the MCC value obtained by Deep4ACE under different data divisions is still greater than that of other comparison methods, which further shows that the Deep4ACE method is superior to other comparison methods in predicting anti-stress peptides.

[0093] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A vascular anti-stress peptide identification method based on prior knowledge embedding and multi-view enhancement, characterized in that: The following steps are involved: Step 1: Construct ACEIP, nonACEIP and ACE sequence data sets and perform preprocessing; Step 2: construct feature vector encoding of peptide sequence using prior knowledge similarity and molecular docking fitness; Step 3: Using the encoded vectors of ACEIP and nonACEIP sequences as input, a Deep4ACE model including a feature enhancement module, a model learning module, and a model prediction module is constructed to predict ACEIP and nonACEIP.

2. The vascular anti-compression peptide identification method based on joint prior knowledge embedding and multi-view enhancement according to claim 1 is characterized in that: The similarity feature vector encoding using prior knowledge includes: Use the ESM-2 model to encode ACE, ACEIP, and nonACEIP sequences into V ACE 、V ACEIP and V non-ACEIP , calculate the Pearson similarity between ACE and ACEIP, ACE and nonACEIP, and get the similarity matrix SM ACE-ACEIP , SM ACE-nonACEIP ; Use the ESM-2 model to obtain the feature vector of ACEIP or nonACEIP or ACE sequence as Mat seq , along the sequence length L dimension of Mat seq Find the average and get the eigenvector V seq ; Based on SM ACE-ACEIP and SM ACE-nonACEIP , construct GS ACE-ACEIP Figure and GS ACE-nonACEIP picture; Using ESM-2 model to analyze GS ACE-ACEIP and GS ACE-nonACEIP Node encoding.

3. The vascular anti-compression peptide identification method based on joint prior knowledge embedding and multi-view enhancement according to claim 2 is characterized in that: Step 214 includes: At GS ACE-ACEIP In the process, random walk is used to sample nodes in the graph to obtain the path, and the nodes in the path are ACEIP sequence and ACE sequence; the ESM-2 model is used to encode the corresponding node sequence in the path into a vector; The Skip-Gram model is used to learn the representation of each node by predicting the location of the context node, and to obtain the vector representation of ACEIP and nonACEIP.

4. The vascular anti-compression peptide identification method based on joint prior knowledge embedding and multi-view enhancement according to claim 1 is characterized in that: The molecular docking fitness feature vector encoding includes: Perform molecular docking of ACEIP and nonACEIP with ACE respectively to obtain the docking fitness score matrix DM ACE-ACEIP DM ACE-nonACEIP ; According to DM ACE-ACEIP and DM ACE-nonACEIP Construction of molecular docking map GD ACE-ACEIP and GD ACE-nonACEIP ; Get the encoding vectors of ACEIP and nonACEIP.

5. The vascular anti-compression peptide identification method based on joint prior knowledge embedding and multi-view enhancement according to claim 1 is characterized in that: Feature enhancement modules include: Step 311, obtain sequence key point enhanced features: use a certain size window along the input vector V in The first dimension slides; let ACEIP's sliding window Sliding window of nonACEIP by is a positive sample, For negative samples, the predicted probability is obtained using the RF classifier Will and As the feature representation of window i; PowerShap is used to filter the first several key windows; Step 312: Convert sequence information into position matrix information: traverse each site of the encoding vector, and the first dimension value of each site vector go through Convert to X-axis coordinate position X i ; The second dimension of each site vector go through Convert to Y coordinate position Y i ; Set the coordinates (X i , Y i ) is set to 1; traverse each position of the sequence encoding vector and build a position matrix.

6. The vascular anti-compression peptide identification method based on joint prior knowledge embedding and multi-view enhancement according to claim 5 is characterized in that: The model learning module includes: Sequence keypoint enhancement features Through a fully connected network Get Features Will Each element is mapped to the key window by multiplication to obtain the feature Output local particle feature vectors through LSTM, TDFC, and TCN respectively Context feature information vector Global feature information vector Calculation using self-attention mechanism and The synergy of the output vector After vector expansion, it is input into several fully connected networks and the output vector Position Matrix Vector After 3 layers of convolutional network, the output is the feature Will Through the pooling of the average pooling layer, the first two dimensions of the pooled matrix are flattened to obtain the feature The characteristics Convert to Graph 7. The vascular anti-compression peptide identification method based on joint prior knowledge embedding and multi-view enhancement according to claim 6 is characterized in that: feature Convert to Graph include: The characteristics Consider it as a vector of several dimensions, calculate the Pearson similarity between the vectors, set the connection threshold to T1, compare the Pearson similarity between the vectors with T1 to determine whether the channel vectors are connected; then use the graph feature The nodes in the graph are fused into dimensional vectors through a 3-layer graph convolutional network and a global maximum pooling operation.

8. The vascular anti-compression peptide identification method based on joint prior knowledge embedding and multi-view enhancement according to claim 6 is characterized in that: The convolution kernel size of the TCN network is set to 5, the step size is 1, the expansion coefficient is 2, the number of layers is set to 3, and the number of channels is set to 128; the output dimension of the LSTM network is 128; the output dimension of the TDFC network is set to 128.

9. The vascular anti-compression peptide identification method based on joint prior knowledge embedding and multi-view enhancement according to claim 6, characterized in that: The prediction module includes 3 layers of fully connected networks and sigmoid function, and the output of each layer of fully connected network is 512, 125, 1.

10. The vascular anti-compression peptide identification method based on joint prior knowledge embedding and multi-view enhancement according to claim 1, characterized in that: The preprocessing includes: removing peptide sequences with amino acids whose similarity is greater than a set threshold and whose length is less than a set threshold.

Images