A method and system for identifying antimicrobial peptides and anti-inflammatory peptides

By designing the PepNet network model, using the pre-trained feature and residual expansion convolution module and the Transformer module, the problem of feature extraction and insufficient network architecture in the recognition of antimicrobial peptides and anti-inflammatory peptides is solved, and higher recognition accuracy and consistency are achieved .

CN118486376BActive Publication Date: 2024-11-15SHANDONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410938993.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-15
Publication Date
2024-11-15
Estimated Expiration
2044-07-15

AI Technical Summary

Technical Problem

The existing antimicrobial peptide and anti-inflammatory peptide recognition methods have shortcomings in feature extraction and network architecture, and cannot effectively capture complex patterns and relationships in the peptide sequence, resulting in inaccurate recognition results.

Method used

The peptide sequence features are extracted using a pre-trained protein language model, and combined with the residual expansion convolution module and the residual Transformer module, the PepNet network model is designed, and the classification is carried out through a multi-layer perceptron to achieve efficient identification of antimicrobial peptides and anti-inflammatory peptides. .

Benefits of technology

It improves the recognition accuracy of antimicrobial peptides and anti-inflammatory peptides, can better capture complex patterns and relationships in the peptide sequence, significantly better than other advanced methods, and achieves high accuracy and consistency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118486376B_ABST
    Figure CN118486376B_ABST
Patent Text Reader

Abstract

The present invention proposes a method and system for identifying antimicrobial peptides and anti-inflammatory peptides, and relates to the technical field of biological peptide identification. The method includes obtaining peptide sequences of antimicrobial peptides and anti-inflammatory peptides, extracting pre-trained features related to functions in peptide sequences; extracting the amino acid type and physicochemical properties of the peptide sequences as original amino acid features, inputting the original amino acid features into a residual dilated convolution module, extracting features from the interval sequence neighbors of the original amino acid features, and obtaining sequence features; fusing the sequence features and the pre-trained features to obtain combined features; inputting the combined features into a residual Transformer module, and using the residual Transformer to extract the global sequence information of the combined features to obtain the final amino acid representation features; based on the final amino acid representation features, the peptide sequences are classified. The present invention does not require manual feature design, and can well capture the complex patterns and relationships hidden in the peptide sequences, as well as the global relationships of the peptide sequences, thereby improving the accuracy of the recognition results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of biological peptide recognition, and in particular relates to an antimicrobial peptide and anti-inflammatory peptide recognition method and system. Background Art

[0002] Bacterial infection and inflammation play a vital role in human health. Bacterial infections include a range of diseases such as respiratory tract infections, skin infections, urinary tract infections, and gastrointestinal infections, which can cause serious pathological consequences. Inflammation is a basic physiological response of the body to injury or infection, aimed at protecting tissues and promoting repair. When inflammation persists for a long time and is not controlled, it may develop into chronic inflammation, which is closely related to a variety of diseases, including neurodegenerative diseases, cardiovascular diseases, cancer, and autoimmune diseases. For bacterial infections, antibiotics are usually used as a treatment method, which can effectively kill or inhibit the growth and reproduction of bacteria. However, the overuse of antibiotics has led to the emergence of bacterial resistance, and the spread of infections caused by resistant pathogens has caused major global health problems.

[0003] For chronic inflammation, nonsteroidal anti-inflammatory drugs (NSAIDs), corticosteroids, and immunosuppressants are currently the main treatments, which can suppress the immune system and reduce the level of inflammation. However, similar to antibiotics, their overuse often leads to many adverse reactions and may induce drug resistance. Therefore, there is an urgent need to discover and rationally design new and effective antimicrobial and anti-inflammatory drugs.

[0004] Antimicrobial peptides (AMPs) and anti-inflammatory peptides (AIPs) have been shown to have less drug resistance and have wider applications and advantages in treating bacterial infections and inflammation. Antimicrobial peptides can penetrate the bacterial cell membrane, destroy its membrane structure or directly interact with biological molecules inside the bacteria, leading to bacterial death. On the other hand, anti-inflammatory peptides can inhibit the production and release of inflammatory mediators and reduce tissue inflammatory responses. Therefore, the effective identification of bioactive antimicrobial peptides and anti-inflammatory peptides is crucial for the development of candidate drugs. Traditional experimental methods for the identification of antimicrobial peptides and anti-inflammatory peptides are time-consuming, costly, and labor-intensive. Therefore, computational methods, especially traditional machine learning-based methods and the currently popular deep learning-based methods, have received widespread attention due to their ability to quickly, efficiently, and high-throughput identify antimicrobial peptides and anti-inflammatory peptides.

[0005] Existing computational methods for antimicrobial and anti-inflammatory peptide identification can be roughly divided into machine learning-based and deep learning-based methods. However, the performance of traditional machine learning-based methods, such as support vector machine (SVM) and random forest (RF), is severely affected by handcrafted features. Deep learning-based methods have shown excellent performance in identifying antimicrobial and anti-inflammatory peptides due to their unique network architecture and powerful learning ability. Well-known deep learning models, including convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory (LSTM), bidirectional long short-term memory (Bi-LSTM), gated recurrent unit (GRU), and Transformer have achieved success in this field. CNN is able to extract local features from peptide sequences, while RNN, LSTM, Bi-LSTM, and GRU are specifically used to cyclically encode the position information of peptide sequences. Transformer can capture the relationship between different amino acid residues in various high-dimensional feature subspaces.

[0006] Although existing deep learning methods have achieved great success in predicting antimicrobial and anti-inflammatory peptides, they still have many shortcomings, including feature extraction and network architecture:

[0007] In terms of feature extraction, current deep learning-based methods usually use manually designed features such as amino acid composition (AAC), dipeptide composition adaptive jump graph (ASDC), and physicochemical properties. However, these handcrafted features often cannot fully capture the complex patterns and relationships hidden in peptide sequences and may overlook key information related to antibacterial or anti-inflammatory activity.

[0008] In terms of network structure, CNN cannot extract global or long-range dependencies of peptide sequences, and is prone to gradient vanishing or exploding. LSTM and GRU models may consume too much memory or computing time, especially when processing large-scale data. Therefore, combining a reasonable and interpretable deep learning model to fully extract function-related peptide sequence features is crucial for predicting antimicrobial and anti-inflammatory peptides. Summary of the invention

[0009] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a method and system for identifying antimicrobial peptides and anti-inflammatory peptides, which selects two features as input features, does not require manual feature design, and can well capture the complex patterns and relationships hidden in peptide sequences, as well as the global or long-range dependencies of peptide sequences, thereby improving the accuracy of the recognition results.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0011] The first aspect of the present invention provides a method for identifying antimicrobial peptides and anti-inflammatory peptides.

[0012] A method for identifying antimicrobial peptides and anti-inflammatory peptides comprises the following steps:

[0013] Obtain peptide sequences of antimicrobial peptides and anti-inflammatory peptides, and use pre-trained protein language models to extract pre-trained features related to functions in peptide sequences;

[0014] The peptide sequence and pre-trained features are input into the peptide function prediction model, the amino acid type and physicochemical properties of the peptide sequence are extracted as the original amino acid features, the original amino acid features are input into the residual dilated convolution module, and features are extracted from the interval sequence neighbors of the original amino acid features to obtain sequence features;

[0015] Combine sequence features and pre-trained features to obtain combined features;

[0016] The combined features are input into the residual Transformer module, and the residual Transformer containing the encoder-decoder structure is used to extract the global sequence information of the combined features to obtain the final amino acid representation features;

[0017] Based on the final amino acid representation features, the peptide sequences are classified.

[0018] The second aspect of the present invention provides an antimicrobial peptide and anti-inflammatory peptide recognition system.

[0019] An antimicrobial peptide and anti-inflammatory peptide recognition system, comprising:

[0020] The preprocessing module is configured to: obtain peptide sequences of antimicrobial peptides and anti-inflammatory peptides, and extract pre-trained features related to functions in the peptide sequences using a pre-trained protein language model;

[0021] The original feature and sequence feature extraction module is configured to: input the peptide sequence and the pre-trained features into the peptide function prediction model, extract the amino acid type and physicochemical properties of the peptide sequence as the original amino acid features, input the original amino acid features into the residual dilation convolution module, extract features from the interval sequence neighbors of the original amino acid features, and obtain sequence features;

[0022] The fusion module is configured to: fuse the sequence features and the pre-trained features to obtain the combined features;

[0023] The global sequence information extraction module is configured to: input the combined features into the residual Transformer module, extract the global sequence information of the combined features using the residual Transformer including an encoder-decoder structure, and obtain the final amino acid representation features;

[0024] The classification module is configured to classify the peptide sequence based on the final amino acid representation features.

[0025] One or more of the above technical solutions have the following beneficial effects:

[0026] The present invention provides an antimicrobial peptide and anti-inflammatory peptide identification method and system, designs a polypeptide function prediction model including a residual dilation convolution module and a residual Transformer module, namely a PepNet network model, the input features of the PepNet network model come from two parts, respectively, a pre-trained protein language model extracts pre-trained features related to functions in peptide sequences, and original amino acid features including amino acid types and physicochemical properties of peptide sequences, the original amino acid features are processed by the residual dilation convolution module, after obtaining sequence features, the sequence features are fused with the pre-trained features to obtain combined features, and finally the combined features are processed by the residual Transformer module to obtain the final amino acid representation features. The present invention does not require manual design and production of features, selects the above two parts of features as input features, can better capture the complex patterns and relationships hidden in the peptide sequence, will not ignore the key information related to antimicrobial or anti-inflammatory activity, so that the final recognition result is more accurate.

[0027] The present invention uses a residual dilated convolution module to encode the original amino acid features, including the one-hot encoding and physicochemical properties of the amino acid types, to obtain multi-order neighborhood information of each amino acid in the sequence. The three designed dilated convolution layers gradually expand the receptive field and can capture information from the separated amino acid neighbors.

[0028] The present invention inputs the original amino acid features encoded by the residual dilated convolution module together with the high-dimensional pre-trained features into the residual Transformer module. The Transformer encoder and decoder can extract information from all positions in the peptide sequence and calculate the dependencies between amino acids at different positions in the peptide sequence, thereby capturing global sequence information.

[0029] The present invention designs a peptide sequence classifier, which generates a representation of the peptide sequence through average pooling of the learned amino acid features, and then inputs it into a multi-layer perceptron for peptide activity classification, thereby determining whether the peptide sequence belongs to the positive class or the negative class, thereby achieving binary classification of the peptide sequence.

[0030] The experimental results show that the pre-trained features of amino acids, the residual dilated convolution module and the residual Transformer module have a significant impact on the effect of PepNet. It also proves that the performance of the PepNet network in the recognition of antimicrobial peptides AMP and anti-inflammatory peptides AIP is significantly better than other advanced methods, and it has achieved high accuracy and consistency in the prediction of peptide functions.

[0031] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0033] Figure 1 Schematic diagram of the PepNet network structure.

[0034] Figure 2 This is the overall flow chart of PepNet.

[0035] Figure 3 Schematic diagram of the residual dilated convolution module.

[0036] Figure 4 Schematic diagram of the residual Transformer module.

[0037] Figure 5 Schematic diagram of the comparison between PepNet and other methods on the AMP test set.

[0038] Figure 6 A schematic diagram showing the comparison of PepNet with other methods on the AIP test set.

[0039] Figure 7 The figure is a schematic diagram of the effect of PepNet ablation experiments on the AMP and AIP test sets, where the letter AD represents the amino acid feature ablation experiment, EJ represents the residual dilated convolution module ablation experiment, KP represents the residual Transfomer module ablation experiment, and Q and R represent the pooling method ablation experiment. DETAILED DESCRIPTION

[0040] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0041] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0042] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0043] Embodiment 1

[0044] The existing deep learning methods proposed in the above background technology have deficiencies in feature extraction and network architecture for predicting antimicrobial peptides and anti-inflammatory peptides. In terms of feature extraction, manually designed features often cannot fully capture the complex patterns and relationships hidden in peptide sequences, and may ignore key information related to antimicrobial or anti-inflammatory activity; in terms of network architecture, CNN cannot extract global or long-range dependencies of peptide sequences, and is prone to gradient vanishing or exploding, while LSTM and GRU models may consume too much memory or computing time.

[0045] In view of this, in a specific embodiment of the present invention, a method for identifying antimicrobial peptides and anti-inflammatory peptides is disclosed, which mainly has the following technical innovations:

[0046] (1) By using a pre-trained protein language model to fully extract peptide sequence features related to function;

[0047] (2) A residual dilated convolution module is designed to capture the amino acid types and physicochemical properties of the neighbors;

[0048] (3) A residual Transformer module is designed to encode the amino acid pre-training features and use its multi-head attention mechanism to obtain the global information of the peptide sequence;

[0049] (4) Design a peptide sequence classifier, average-pool the learned amino acid features into peptide sequence features, and classify them.

[0050] Based on the above technical innovation, this embodiment proposes a peptide function prediction model (PepNet). The PepNet network model is an antimicrobial peptide and anti-inflammatory peptide recognition algorithm based on deep learning and pre-training models and has strong interpretability. Based on the proposed PepNet network model, the following processing is mainly performed on the peptide sequence to be processed:

[0051] 1. Extract multiple peptide sequence features.

[0052] For a given peptide sequence, PepNet extracts its amino acid type and physicochemical properties, while extracting high-dimensional pre-trained features related to function from the protein large language model ProtT5-XL-U50.

[0053] 2. Residual dilated convolution module.

[0054] This embodiment encodes the original amino acid features through the residual dilated convolution module, including the one-hot encoding and physicochemical properties of the amino acid types, and obtains the multi-order neighborhood information of each amino acid in the sequence; inspired by the time domain convolutional network (TCN), this embodiment designs three dilated convolution layers to gradually expand the receptive field and capture information from the interval amino acid neighbors.

[0055] 3. Residual Transformer module.

[0056] The original amino acid features encoded by the residual dilated convolution module are input into the residual Transformer module together with the high-dimensional pre-trained features. The Transformer encoder and decoder can extract information from all positions in the peptide sequence and calculate the dependencies between amino acids at different positions in the peptide sequence, thereby capturing global sequence information.

[0057] 4. Peptide sequence classifier. The learned amino acid features are average-pooled to generate representations of peptide sequences, which are then fed into a multi-layer perceptron (MLP) for peptide activity classification.

[0058] Furthermore, Figure 1 , Figure 2 , Figure 3 and Figure 4 As shown, the antimicrobial peptide and anti-inflammatory peptide identification method proposed in this embodiment may specifically include the following steps:

[0059] Obtain peptide sequences of antimicrobial peptides and anti-inflammatory peptides, and use pre-trained protein language models to extract pre-trained features related to functions in peptide sequences;

[0060] The peptide sequence and pre-trained features are input into the PepNet network, the amino acid type and physicochemical properties of the peptide sequence are extracted as the original amino acid features, the original amino acid features are input into the residual dilated convolution module, and features are extracted from the interval sequence neighbors of the original amino acid features to obtain sequence features;

[0061] Combine sequence features and pre-trained features to obtain combined features;

[0062] The combined features are input into the residual Transformer module, and the residual Transformer containing the encoder-decoder structure is used to extract the global sequence information of the combined features to obtain the final amino acid representation features;

[0063] Based on the final amino acid representation features, the peptide sequences are classified.

[0064] like Figure 1 As shown, the overall structure diagram of the PepNet network of this embodiment is shown. After the peptide sequence is obtained, the amino acid types and physicochemical properties are extracted respectively, and the amino acid types and physicochemical properties are processed by the residual dilated convolution module and then input into the residual Transformer module;

[0065] At the same time, the peptide sequence is extracted based on the pre-trained model, and the pre-trained features are transferred and learned and also input into the residual Transformer module;

[0066] In the residual Transformer module, the features processed by the residual dilated convolution module and the pre-trained features are fused and processed, and finally classified to obtain the classification results.

[0067] Next, the technical solution of this embodiment will be described in detail:

[0068] 1. Extraction of multiple peptide sequence features

[0069] For a given peptide sequence, this embodiment extracts amino acid types and physicochemical properties as original features. In addition, high-dimensional pre-trained features derived from a large protein language model are also obtained.

[0070] The original features include one-hot coding of amino acid types (20 standard amino acids) and 14 physicochemical properties of amino acids, including 8 amino acid indicators (T4 lysozyme 44 Helical tendency BLAM930101, accessibility information value BIOV880101, Helix normalized frequency MAXF760101, Volume of crystal water etc. using ProtOr TSAJ990101, Amino acid composition of multi-spanning proteins NAKH920108, Amino acid composition of intracellular proteins CEDJ970104, All The folding conformational preference LIFS790101 and the optimal relative partition energy MIYS990104), as well as six specific physicochemical properties (atomicity, polarity, polarizability, net charge, hydrophobicity, and β-sheet conformational tendency).

[0071] The high-dimensional pre-trained features are 1024-dimensional embedding features generated for each amino acid by ProtT5-XL-U50 (ProtT5-XL-U5033 is a widely used pre-trained protein language model based on the Transformer architecture).

[0072] In order to batch encode peptide sequences in a neural network, this example constructs a shape of L ×(34+1024) characteristic matrix X ,in L represents a hyperparameter of fixed peptide sequence length (in the present invention, L =40). When the peptide sequence length is less than L If the length exceeds L , then truncate.

[0073] (II) Residual dilation convolution module obtains interval neighbor information

[0074] The obtained raw features of amino acids are encoded by the residual dilated convolutional block, as Figure 3 As shown in Figure 2, the residual dilated convolution block contains 3 dilated convolution layers, which gradually expand the receptive field and obtain information from the interval sequence neighbors.

[0075] Assume that the input feature matrix is , then for the first i Amino acid characteristics The dilated convolution operation F Defined as:

[0076]

[0077] in, i =1,2,…, L , L represents the hyperparameter of the fixed peptide sequence length; 2m+1 is the convolution kernel size (i.e., the number of neighbors received), k is the expansion coefficient (in this embodiment, the three expansion convolution layers k 1, 2, 4 respectively); and is a learnable parameter; and are the input dimension and output dimension of the dilated convolutional layer respectively.

[0078] Therefore, the residual dilated convolution block containing three dilated convolution layers can be summarized as:

[0079]

[0080] in is the original feature matrix of amino acids.

[0081] Then, the outputs of the three convolutional layers are It is connected with the input original feature residual and then passed through a multi-layer perceptron (MLP) operation as shown below:

[0082]

[0083] in cat Represents a splicing operation, d is the size of the transform dimension.

[0084] 3. Transfer learning pre-training features

[0085] For the high-dimensional pre-trained features generated by the pre-trained model ProtT5-XL-U50, this embodiment reuses them through MLP. Assume that the pre-trained feature matrix is ​​of shapeL ×1024 ,but:

[0086]

[0087] in d is the size of the transform dimension.

[0088] (IV) Residual Transformer module obtains global information of peptide sequence

[0089] Sequence features encoded by residual dilated convolution module and reuse pre-trained feature embeddings In this embodiment, the combined feature matrix is ​​first obtained by merging them and performing an MLP operation ,as follows:

[0090]

[0091] in cat Represents a splicing operation, d is the size of the transform dimension.

[0092] Then, if Figure 4 As shown, the combined feature matrix Input to the residual Transformer module to capture global sequence information. The residual Transformer module contains a residual Transformer of the encoder-decoder structure, where the encoder inputs the sequence matrix Map to Z , decoder integrated and Z To generate the output sequence representation .

[0093] The encoder is N The structure of the network is composed of 2 identical layers, each of which contains two sub-layers, a multi-head self-attention layer and a feed-forward network. N It is a hyperparameter, and the present invention sets it to 1. Residual connections are performed before and after the two sub-layers, and layer normalization is performed.

[0094] Similar to the encoder, the decoder consists of N The encoder consists of 2 identical layers, each of which contains three sublayers, namely two multi-head self-attention layers and a feedforward network. The difference from the encoder is that the additional multi-head attention layer acts on the output of the encoder and the output of the first multi-head self-attention layer.

[0095] The core component of Transformer is the multi-head self-attention mechanism, which can effectively capture the relationship between long-range amino acids and facilitate the extraction of global representation information of peptide sequences. The multi-head self-attention mechanism can be summarized as follows:

[0096]

[0097]

[0098] Among them, Q, K, and V are the query, key, and value of the multi-head self-attention layer respectively. cat For splicing operation, h is the number of parallel attention heads (in this embodiment h =4), W Q , W K , W V , W o are the learnable weights applied to Q, K, V, and multi-head outputs; Attention ( q , k , v ) is the attention mechanism, which is performed as follows:

[0099]

[0100] in q , k , v are the query, key, and value in the self-attention mechanism.

[0101] The process of the feedforward network is as follows:

[0102]

[0103] in X is the input feature matrix. and is the learnable parameter of the first layer of the neural network in the feedforward network; and is the learnable parameter of the second layer of the neural network in the feedforward network, and ReLU is the activation function.

[0104] Therefore, the residual Transformer module can be expressed as:

[0105]

[0106]

[0107] in and It is the encoder and decoder in Transformer.

[0108] In addition, in the combined feature matrix And the output of the Transformer module A residual connection operation is applied between them, which is expressed as follows:

[0109]

[0110] in cat Represents a concatenation operation.

[0111] In addition, the Transformer model is different from the inherent sequential nature of the recurrent neural network (RNN), which processes all residue information in the peptide sequence at the same time. Therefore, this embodiment combines position encoding to describe sequence features. This embodiment uses sine and cosine functions in different feature dimensions, as shown below:

[0112]

[0113] in POS is the position of the amino acid in the peptide sequence, i is the feature dimension, d is the size of the transformation dimension.

[0114] (V) Generate peptide binary predictions.

[0115] For the learned amino acid representation matrix ,The present invention first average-pools them into a sequence representation and then classifies it using MLP.

[0116]

[0117] in, score It is a probability between 0 and 1. The closer the score is to 1, the greater the possibility that the input peptide sequence belongs to the positive class. The closer the score is to 0, the greater the possibility that the input peptide sequence belongs to the negative class.

[0118] 6. Evaluation plan

[0119] This example evaluates our model mainly from the following two aspects: (1) Comparing this example with other algorithms on various evaluation indicators. (2) Conducting ablation experiments on the core technology proposed in this example.

[0120] The identification of antimicrobial peptides and anti-inflammatory peptides is actually a binary classification problem. Therefore, the present invention uses ACC (accuracy), Recall, Precision, F1-score, and MCC (Matthews correlation coefficient) as evaluation indicators. The calculation of each indicator is as follows:

[0121]

[0122]

[0123]

[0124]

[0125]

[0126] Among them, TP (True Positives) represents true positive examples, that is, the number of samples that the model predicts as positive; TN (True Negatives) represents true negative examples, that is, the number of samples that the model predicts as negative; FP (False Positives) represents false positive examples, that is, the number of samples that the model predicts as positive; FN (False Negatives) represents false negative examples, that is, the number of samples that the model predicts as positive.

[0127] In order to balance the precision and recall, the present invention adopts F1-score and MCC as the main evaluation criteria.

[0128] First, the effect of the PepNet model proposed in this embodiment is compared with other models for identifying antimicrobial peptides and anti-inflammatory peptides in terms of the above indicators. Secondly, in order to further explore the high-dimensional pre-training features, as well as the residual dilated convolution module and residual Transformer module designed in this embodiment, PepNet was subjected to ablation experiments on these three parts.

[0129] (VII) Experimental results

[0130] (1) PepNet has great advantages over other advanced methods.

[0131] PepNet was trained and tested on two datasets: antimicrobial peptides and anti-inflammatory peptides. AMPlify, AMP Scanner Vr.1 (RF Precision), AMP Scanner Vr.1 (EarthPrecision), AMP Scanner Vr.2, AMP Scanner Vr.2 (retrain), and TriNet (retrain) were compared on the AMP test set, and AIPStack, PTPP (Class feature), PPTPP (Probabilityfeature), PPTPP (Fusion feature), AIP_MDL, and TriNet (retrain) were compared on the AIP test set.

[0132] The performance comparison results on the AMP test set are shown in Table 1 and Figure 5 As shown in the figure. Based on the comparison results, we found that PepNet showed the best performance on the AMP test set, outperforming all other compared methods. PepNet achieved an accuracy of 0.950, a recall of 0.954, a precision of 0.947, an F1-score of 0.951, and an MCC of 0.901. In terms of accuracy, recall, precision, F1-score, and MCC, PepNet achieved improvements of 3.8% to 22.3%, 6.5% to 21.2%, 1.6% to 29.7%, 4.2% to 21.3%, and 8.4% to 60.3% over other methods, respectively. Specifically, PepNet achieved a significant improvement in recall, F1-score, and MCC values, which were 6.5%, 4.2%, and 8.4% higher than the suboptimal model, respectively. In addition, PepNet was the only model with all evaluation indicators above 0.9, indicating its strong ability to accurately identify AMP.

[0133] Table 1 shows the performance comparison of PepNet and other methods on the AMP test set.

[0134] Accuracy Recall Accuracy F1 MCC AMPlify 0.890 0.867 0.913 0.887 0.780 AMP Scanner Vr.1 (RF Precision) 0.787 0.787 0.788 0.787 0.575 AMP Scanner Vr.1 (Earth Precision) 0.781 0.795 0.773 0.784 0.562 AMP Scanner Vr.2 (Feb2020) 0.777 0.880 0.730 0.798 0.567 AMP Scanner re-trained 0.903 0.896 0.909 0.902 0.806 TriNet 0.915 0.896 0.932 0.913 0.831 PepNet 0.950 0.954 0.947 0.951 0.901

[0135] The performance comparison results on the AIP test set are shown in Table 2 and Figure 6 As shown in the figure, it can be seen that PepNet's performance on the AIP test set is the best among all evaluation methods. PepNet has an accuracy of 0.819, a recall of 0.940, a precision of 0.705, an F1-score of 0.806, and an MCC of 0.666.

[0136] In addition, compared with other methods, PepNet achieved improvements of 8.2% to 30.6% in precision, 28.6% to 889.5% in recall, 14.0% to 374.1% in F1-score, and 33.2% to 276.3% in MCC. Compared with the second-best model, PepNet achieved improvements of 28.6%, 14%, and 33.2% in recall, F1-score, and MCC, respectively.

[0137] Although PepNet's precision is lower than PPTPP (Fusion feature) and AIP_MDL, its other indicators are quite high. In addition, considering that recall and precision are two trade-off indicators, PepNet's recall is significantly higher than other algorithms when the precision decreases relatively little, which shows that PepNet is more sensitive than other algorithms in identifying positive samples. Therefore, considering all factors, PepNet achieves the best performance.

[0138] Table 2 shows the performance comparison of PepNet and other methods on the AIP test set.

[0139] Accuracy Recall Accuracy F1 MCC AIPStack 0.740 0.605 0.704 0.650 0.448 PPTPP (Class feature) 0.711 0.524 0.681 0.592 0.382 PPTPP (Probability feature) 0.727 0.562 0.698 0.623 0.419 PPTPP (Fusion feature) 0.627 0.095 0.784 0.170 0.177 AIP_MDL 0.757 0.731 0.774 0.707 0.500 TriNet 0.669 0.450 0.620 0.521 0.287 PepNet 0.819 0.940 0.705 0.806 0.666

[0140] In summary, PepNet significantly outperformed other state-of-the-art methods in AMP and AIP identification. The high values ​​of F1-score and MCC indicate that PepNet achieved high accuracy and consistency in peptide function prediction.

[0141] (2) PepNet conducts ablation experiments on core technologies

[0142] In this example, we will delve into each component of the PepNet model to clarify its role and verify its contribution. Through a series of ablation experiments, we systematically study the impact of changing individual model components or hyperparameters, adhering to the approach of modifying only one component or hyperparameter at a time. Based on the structure of the PepNet framework, the ablation experiments examine the impact of feature selection, the contribution of the residual dilated convolution module, the contribution of the residual Transformer module, and the impact of pooling amino acid features into peptide sequence features.

[0143] In order to study the contribution of amino acid features to PepNet, this embodiment trains and tests PepNet by removing amino acid type features, physicochemical property features and pre-training features. In order to study the contribution of the residual dilated convolution module to PepNet, the present invention trains and tests PepNet by removing the residual convolution module and the residual connection operation, and replacing the dilated convolution (CNN) with a bidirectional long short-term memory model (Bi-LSTM), a long short-term memory model (LSTM) and a gated recurrent unit (GRU). In order to study the contribution of the residual Transformer module to PepNet, the present invention trains and tests PepNet by removing the residual Transformer module, the residual connection operation, and the selection of the number of layers of the Transformer hyperparameter. The smile experiment results on the AMP and AIP test sets are shown in Tables 3 and 4, respectively.

[0144] It can be found that the pre-trained features of amino acids, the residual dilated convolution module, and the residual Transformer module have a significant impact on the performance of PepNet. Specifically, the pre-trained features contribute the most to PepNet on the AMP and AIP test sets. Removing this feature causes a significant drop in F1 scores of 4.7% and 11.3% on the AMP and AIP test sets, and a significant drop in MCC scores of 9.5% and 27.3% compared to using all features. Removing the residual dilated convolution module causes a drop in F1 score and MCC performance of 5.2% and 10.1% on the AMP test set, and a drop in performance of 11.8% and 31.8% on the AIP test set. Removing the residual Transformer module causes a significant drop in the performance of PepNet, especially in the AIP test set, where the F1 score and MCC drop significantly by 27.7% and 121.3%, respectively. Max pooling causes a drop in F1 score of 2.7% and MCC of 5.9% on the AMP test set, and a drop in F1 score of 12.6% and MCC of 32.4% on the AIP test set.

[0145] Table 3 shows the ablation experiment on the AMP test set.

[0146]

[0147] Table 4 shows the ablation experiment on the AIP test set.

[0148]

[0149] like Figure 7As shown, it is a schematic diagram of the accuracy, recall, precision, F1 and MCC effects of PepNet's ablation experiments on antimicrobial peptides and anti-inflammatory peptide test sets. The dotted line represents the result of PepNet on the antimicrobial peptide data, the asterisk line represents the result of PepNet on the anti-inflammatory peptide data, the horizontal axis represents the various ablation experiments performed in the present invention, wherein the letters AD represent the amino acid feature ablation experiment, EJ represents the residual dilated convolution module ablation experiment, KP represents the residual Transfomer module ablation experiment, Q and R represent the pooling ablation experiment, and the vertical axis represents the value of each indicator.

[0150] Embodiment 2

[0151] This embodiment discloses an antimicrobial peptide and anti-inflammatory peptide recognition system.

[0152] An antimicrobial peptide and anti-inflammatory peptide recognition system, comprising:

[0153] The preprocessing module is configured to: obtain peptide sequences of antimicrobial peptides and anti-inflammatory peptides, and extract pre-trained features related to functions in the peptide sequences using a pre-trained protein language model;

[0154] The original feature and sequence feature extraction module is configured to: input the peptide sequence and the pre-trained features into the PepNet network, extract the amino acid type and physicochemical properties of the peptide sequence as the original amino acid features, input the original amino acid features into the residual dilated convolution module, extract features from the interval sequence neighbors of the original amino acid features, and obtain sequence features;

[0155] The fusion module is configured to: fuse the sequence features and the pre-trained features to obtain the combined features;

[0156] The global sequence information extraction module is configured to: input the combined features into the residual Transformer module, extract the global sequence information of the combined features using the residual Transformer including an encoder-decoder structure, and obtain the final amino acid representation features;

[0157] The classification module is configured to classify the peptide sequence based on the final amino acid representation features.

[0158] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0159] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for identifying antimicrobial peptides and anti-inflammatory peptides, characterized in that: The following steps are involved: Obtain peptide sequences of antimicrobial peptides and anti-inflammatory peptides, and use pre-trained protein language models to extract pre-trained features related to functions in peptide sequences; The peptide sequence and pre-trained features are input into the peptide function prediction model, the amino acid type and physicochemical properties of the peptide sequence are extracted as the original amino acid features, the original amino acid features are input into the residual dilated convolution module, and features are extracted from the interval sequence neighbors of the original amino acid features to obtain sequence features; Combine sequence features and pre-trained features to obtain combined features; The combined features are input into the residual Transformer module, and the residual Transformer containing the encoder-decoder structure is used to extract the global sequence information of the combined features to obtain the final amino acid representation features; Based on the final amino acid representation features, the peptide sequences are classified; The peptide sequences to be processed were processed as follows: Extract multiple peptide sequence features, among which the original amino acid features include one-hot codes of 20 standard amino acids and 14 physicochemical properties of amino acids; The residual dilated convolution module encodes the original amino acid features, including the one-hot encoding and physicochemical properties of the amino acid types, and obtains the multi-order neighborhood information of each amino acid in the sequence. Inspired by the time domain convolutional network, three dilated convolution layers are designed to gradually expand the receptive field and capture information from the neighbors of the interval amino acids. In the residual Transformer module, the original features of amino acids encoded by the residual dilated convolution module are input into the residual Transformer module together with the high-dimensional pre-trained features. The Transformer encoder and decoder can extract information from all positions in the peptide sequence and calculate the dependencies between amino acids at different positions in the peptide sequence, thereby capturing global sequence information. Peptide sequence classifier,The learned amino acid features are average-pooled to generate a representation of the peptide sequence, which is then input into a multi-layer perceptron for peptide activity classification.

2. The method for identifying antimicrobial peptides and anti-inflammatory peptides according to claim 1, characterized in that: The 14 physical and chemical properties include 8 amino acid indicators: T4 lysozyme 44 Helical tendency BLAM930101, accessibility information value BIOV880101, Helix normalized frequency MAXF760101, Volume of crystallographic water using ProtOr TSAJ990101, Amino acid composition of multi-spanning proteins NAKH920108, Amino acid composition of intracellular proteins CEDJ970104, All The folding conformational preferences LIFS790101 and the optimal relative partition energy MIYS990104, as well as six specific physicochemical properties: atomicity, polarity, polarizability, net charge, hydrophobicity and β-sheet conformational tendency.

3. The method for identifying antimicrobial peptides and anti-inflammatory peptides according to claim 1, characterized in that: In the residual dilated convolution module, features are extracted from the interval sequence neighbors using a plurality of dilated convolution layers with different dilation coefficients connected in series, and outputs of the plurality of dilated convolution layers with different dilation coefficients connected in series are obtained, specifically: Assume that the input feature matrix is , then for the first i Amino acid characteristics The dilated convolution operation F Defined as: ; in, i =1,2,…, L, L represents the hyperparameter of the fixed peptide sequence length; 2m+1 is the convolution kernel size, i.e., the number of neighbors received; k is the expansion coefficient; and is a learnable parameter; and are the input dimension and output dimension of the dilated convolutional layer respectively.

4. The method for identifying antimicrobial peptides and anti-inflammatory peptides according to claim 3, characterized in that: Also includes: The outputs of multiple dilated convolutional layers with different dilation coefficients connected in series are residually connected with the original features of amino acids, and then passed through a multi-layer perceptron to obtain the sequence features.

5. The method for identifying antimicrobial peptides and anti-inflammatory peptides according to claim 1, characterized in that: The pre-trained protein language model is the ProtT5-XL-U50 model.

6. The method for identifying antimicrobial peptides and anti-inflammatory peptides according to claim 1, characterized in that: The combined features are input into the encoder and decoder of the residual Transformer module respectively, and the encoder is used to map the combined features to Z , using the decoder to integrate the combined features and Z To generate the output sequence representation, specifically expressed as: ; ; in and It is the encoder and decoder in Transformer; is a combination of features; Z The encoder maps the combined feature to obtain the feature representation; It is a feed-forward network; It is a multi-head self-attention mechanism; is the output sequence representation of the Transformer module.

7. The method for identifying antimicrobial peptides and anti-inflammatory peptides according to claim 6, characterized in that: A residual connection operation is applied between the combined features and the output sequence representation of the Transformer module to obtain the final amino acid representation features.

8. The method for identifying antimicrobial peptides and anti-inflammatory peptides according to claim 6, characterized in that: Also includes: Before the combined features are fed into the encoder of the residual Transformer module, sine and cosine functions are used on different feature dimensions in combination with positional encoding to characterize the sequence features.

9. The method for identifying antimicrobial peptides and anti-inflammatory peptides according to claim 1, characterized in that: The final amino acid representation features are averaged and pooled into a sequence representation, and then MLP is used to classify the average pooled sequence representation. The closer the score is to 1, the greater the possibility that the input peptide sequence belongs to the positive class; conversely, the closer the score is to 0, the greater the possibility that the input peptide sequence belongs to the negative class.

10. An antimicrobial peptide and anti-inflammatory peptide recognition system, characterized in that: include: The preprocessing module is configured to: obtain peptide sequences of antimicrobial peptides and anti-inflammatory peptides, and extract pre-trained features related to functions in the peptide sequences using a pre-trained protein language model; The original feature and sequence feature extraction module is configured to: input the peptide sequence and the pre-trained features into the peptide function prediction model, extract the amino acid type and physicochemical properties of the peptide sequence as the original amino acid features, input the original amino acid features into the residual dilation convolution module, extract features from the interval sequence neighbors of the original amino acid features, and obtain sequence features; The fusion module is configured to: fuse the sequence features and the pre-trained features to obtain the combined features; The global sequence information extraction module is configured to: input the combined features into the residual Transformer module, extract the global sequence information of the combined features using the residual Transformer including an encoder-decoder structure, and obtain the final amino acid representation features; The classification module is configured to: classify the peptide sequence based on the final amino acid representation features; The peptide sequences to be processed were processed as follows: Extract multiple peptide sequence features, among which the original amino acid features include one-hot codes of 20 standard amino acids and 14 physicochemical properties of amino acids; The residual dilated convolution module encodes the original amino acid features, including the one-hot encoding and physicochemical properties of the amino acid types, and obtains the multi-order neighborhood information of each amino acid in the sequence. Inspired by the time domain convolutional network, three dilated convolution layers are designed to gradually expand the receptive field and capture information from the neighbors of the interval amino acids. In the residual Transformer module, the original features of amino acids encoded by the residual dilated convolution module are input into the residual Transformer module together with the high-dimensional pre-trained features. The Transformer encoder and decoder can extract information from all positions in the peptide sequence and calculate the dependencies between amino acids at different positions in the peptide sequence, thereby capturing global sequence information. Peptide sequence classifier,The learned amino acid features are average-pooled to generate a representation of the peptide sequence, which is then input into a multi-layer perceptron for peptide activity classification.