Antibacterial cyclic peptide screening system and method based on multi-modal cross attention mechanism

Abstract

The invention discloses an antibacterial cyclic peptide screening system and method based on a multi-modal cross attention mechanism, and belongs to the field of antibacterial cyclic peptide screening, the antibacterial cyclic peptide screening system comprises the antibacterial cyclic peptide screening system of the multi-modal cross attention mechanism, the antibacterial cyclic peptide screening system of the multi-modal cross attention mechanism comprises a multi-modal cross attention screening model and a multi-modal cross attention screening model, the method is used for screening antibacterial cyclic peptides; data of the data input module comprises a one-dimensional sequence, a two-dimensional structure and a three-dimensional structure, data sources in the data input module comprise but are not limited to APD3, DRAMP and a Un i Prot public database, and the data input module is used for conveying data of the antibacterial cyclic peptide; through cooperative use of the devices, the cross attention fusion module is arranged, and weights of different modals are dynamically allocated, so that a key active region of the antibacterial cyclic peptide is more obvious, and a multi-modal cross attention screening model can more accurately identify related characteristics of the antibacterial activity of the antibacterial cyclic peptide; the screening of the antibacterial cyclic peptide is more accurate.

Description

Technical Field

[0001] The present invention relates to the technical field of antimicrobial cyclic peptide screening, and in particular to an antimicrobial cyclic peptide screening system and method based on a multimodal cross-attention mechanism. Background Art

[0002] Antibacterial peptides have a unique bactericidal mechanism and can exert a strong antibacterial effect by directly destroying bacterial cell membranes or interfering with their internal physiological functions. Unlike traditional antibiotics, antibacterial peptides not only have broad-spectrum antibacterial activity against a variety of bacteria, but also are less likely to produce drug resistance. Antimicrobial peptides are an important part of antibacterial drug research. Antimicrobial cyclic peptides have become the main development direction in antimicrobial peptide research due to their excellent biological and chemical properties.

[0003] However, the R&D of antimicrobial peptide drugs is costly, time-consuming, and risky. Due to the special cyclic structure of antimicrobial cyclic peptides, the cost, cycle, and risk conflicts in the synthesis, verification, and analysis stages of the R&D cycle are particularly prominent, thus reducing the efficiency of the later screening of antimicrobial cyclic peptides.

[0004] At present, most of the research work on antimicrobial cyclic peptide drugs can be divided into traditional methods and simulation methods from the perspective of technical routes. Traditional methods include artificial synthesis verification, high-throughput, high-content screening, etc.; simulation methods include computer-aided drug design and artificial intelligence-assisted drug design. Traditional screening of antimicrobial cyclic peptide drugs generally requires a long period of wet method experiments, which is time-consuming and economically costly and has great risks. Therefore, the use of simulation calculations to carry out preliminary drug screening has gradually become the mainstream direction of this type of research. The core principle of computer-assisted antimicrobial cyclic peptide drug design is to build models based on molecular or quantum mechanics in computational chemistry. The advantages of this method are strong interpretability, the ability to analyze principles, summarize laws, and expand applications.

[0005] Existing antimicrobial cyclic peptide screening generally uses the one-dimensional linear Fasta sequence or two-dimensional SMILES graph structure of the peptide to predict antimicrobial activity. However, the above method ignores the complexity of the compound structure in real life and cannot effectively identify stereochemical properties such as chirality, resulting in the inability to accurately screen and classify antimicrobial cyclic peptides.

[0006] Therefore, the present invention provides an antibacterial cyclic peptide screening system and method based on a multimodal cross-attention mechanism to solve the above problems. Summary of the Invention

[0007] (1) Technical problems solved

[0008] The present invention provides an antibacterial cyclic peptide screening system and method based on a multimodal cross-attention mechanism, aiming to solve the problems raised in the background technology.

[0009] (2) Technical solution

[0010] To achieve the above objectives, the present invention provides the following technical solutions: an antimicrobial cyclic peptide screening system based on a multimodal cross-attention mechanism, including an antimicrobial cyclic peptide screening system with a multimodal cross-attention mechanism, wherein the antimicrobial cyclic peptide screening system with a multimodal cross-attention mechanism comprises:

[0011] A multimodal cross-attention screening model for screening antimicrobial cyclic peptides;

[0012] A data input module, wherein the data of the data input model includes one-dimensional sequence, two-dimensional structure and three-dimensional structure. The data sources in the data input model include but are not limited to APD3, DRAMP and UniProt public databases, which are used to transmit data of antimicrobial cyclic peptides;

[0013] A multimodal feature extraction module is used to train a multimodal cross-attention screening model to extract different features of antimicrobial cyclic peptides and map the feature vectors of different modalities to a unified dimension based on a multi-layer perceptron;

[0014] A cross-attention fusion module, which includes an input layer and an attention calculation model for calculating attention weights and fusion;

[0015] An isomorphic twin network model that compares the similarity of cyclic peptide samples and candidate cyclic peptides in a random screening space based on a contrastive learning module;

[0016] The classification and screening module is used to output the antibacterial activity probability score of the candidate cyclic peptides.

[0017] As a preferred technical solution of the present application, the one-dimensional sequence, two-dimensional structure and three-dimensional structure are respectively the amino acid linear sequence of the cyclic peptide, the two-dimensional molecular graph representation of the cyclic peptide and the three-dimensional spatial conformation of the cyclic peptide, and the data in the public database is obtained based on keyword screening to obtain antibacterial cyclic peptide data.

[0018] As an optimal technical solution of the present application, the multimodal feature extraction module trains feature extraction of a multimodal cross-attention screening model based on a pre-trained model. The multimodal feature extraction module includes one-dimensional sequence features, two-dimensional graph features, and three-dimensional structural features to realize multimodal feature extraction.

[0019] As a preferred technical solution of the present application, the one-dimensional sequence features, two-dimensional graph features and three-dimensional structural features respectively use the ESM2 model to extract sequence embedding vectors, the GROVER model to extract the molecular graph embedding of the SMILES graph, and the Uni-Mol model to extract the PDB space conformation embedding, and map the feature vectors of the three modalities based on a multi-layer perceptron.

[0020] As a preferred technical solution of the present application, the input layer is used to transmit the aligned trimodal feature vector to the cross attention fusion module. The trimodal feature is represented as F, S, and P. The attention calculation is a two-layer attention fusion. The two-layer attention fusion is:

[0021] First-layer fusion: With F as the key and value, denoted as K and V respectively, and S as the query, denoted as Q, attention weights are calculated and fused;

[0022] Second layer fusion: The fused features are used as new keys and values, P is Q, and secondary attention fusion is performed.

[0023] As a preferred technical solution of this application, the attention weight calculation method is:

[0024]

[0025] Among them, Attention is the attention weight, Q, K, V are query, key and value respectively, d k is the characteristic dimension of the key vector K, QK T is the dot product of the query and the key, used to calculate the similarity between the query and each key, Used to scale the dot product results, Softmax is used to normalize the similarity matrix into a probability distribution, Softmax = 1.

[0026] As a preferred technical solution of the present application, the comparative learning module is used to input the target cyclic peptide and the candidate cyclic peptides in the random screening space, and output the binary classification probability based on the multi-layer perceptron, compare the antibacterial activity similarity between the candidate cyclic peptide and the target cyclic peptide, score the binary classification probability based on the classification and screening module, and screen and classify the cyclic peptides based on the set cyclic peptide threshold.

[0027] The antimicrobial cyclic peptide screening method based on the multimodal cross-attention mechanism is applied to the above-mentioned antimicrobial cyclic peptide screening system based on the multimodal cross-attention mechanism, and includes the following specific steps:

[0028] S1: Construct a multimodal cross-attention screening model. Based on the data input module, a large number of cyclic peptide amino acid linear sequences, cyclic peptide two-dimensional molecular graph representations, and cyclic peptide three-dimensional spatial conformation data are processed and transmitted to the pre-training model. Based on the pre-training model, the multimodal feature extraction module is trained to extract different multimodal features, label different cyclic peptides as antimicrobial peptides, antibacterial peptides, antibacterial cyclic peptides, and antibacterial cyclic peptides, and generate a random screening space.

[0029] S2: In the fine-tuning stage, the non-antimicrobial cyclic peptide data is randomly downsampled to balance the ratio of positive and negative samples. Then, the multimodal feature extraction module is used to extract the trimodal features of the target cyclic peptide. Based on the attention calculation, the attention weight of the cyclic peptide is obtained to capture the features in the cyclic peptide that play a key role in antibacterial activity;

[0030] S3: Map the trimodal features to a unified dimension based on the multi-layer perceptron and fuse the trimodal features, finally achieving model convergence based on the multi-layer perceptron;

[0031] S4: Based on the multi-layer perceptron, the antibacterial activity similarity between the candidate cyclic peptide and the target cyclic peptide is compared, and the binary classification probability is output. The binary classification is divided into similarity / dissimilarity. The binary classification probability is scored based on the classification and screening module, and the set cyclic peptide threshold is compared with the binary classification probability score to screen and classify the cyclic peptide.

[0032] (3) Beneficial effects

[0033] Based on the multimodal feature extraction module and the cross-attention fusion module, by integrating one-dimensional sequence, two-dimensional structure and three-dimensional conformational information, the system can fully capture the stereochemical characteristics of cyclic peptides, avoiding the limitations of a single modality, thereby further improving the accuracy of antimicrobial cyclic peptide screening based on three-dimensional conformational features. Moreover, through the setting of the cross-attention fusion module, the weights of different modalities are dynamically allocated, making the key active areas of antimicrobial cyclic peptides more obvious, thereby enabling the multimodal cross-attention screening model to more accurately identify the relevant features of the antimicrobial activity of antimicrobial cyclic peptides, making the screening of antimicrobial cyclic peptides more accurate.

[0034] By integrating the three-dimensional geometric space structure based on PDB files, the existing pre-trained large models of ESM2, GROVER and Uni-Mol were used to extract the multimodal features of compounds. Since the representation vector dimensions of different models are different, the multi-layer perceptron was used to align the vectors, and then the cross-attention mechanism was used to fuse the multimodal features. Finally, the model convergence was achieved based on the multi-layer perceptron, and multi-classification results were output. The strategy of fusing the three-dimensional representations was adopted to carry out the antibacterial cyclic peptide screening research, which effectively improved the effect of antibacterial cyclic peptide screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a system block diagram of an antimicrobial cyclic peptide screening system and method based on a multimodal cross-attention mechanism;

[0036] Figure 2 This is a flowchart of the antibacterial cyclic peptide screening system and method based on the multimodal cross-attention mechanism;

[0037] Figure 3 Schematic diagram of the structural diversity of cyclic peptides. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] The present invention provides an antimicrobial cyclic peptide screening system and method based on a multimodal cross-attention mechanism, such as Figure 1-Figure 3 As shown, the antimicrobial cyclic peptide screening system based on the multimodal cross-attention mechanism includes an antimicrobial cyclic peptide screening system with a multimodal cross-attention mechanism, and the antimicrobial cyclic peptide screening system with a multimodal cross-attention mechanism includes:

[0040] A multimodal cross-attention screening model for screening antimicrobial cyclic peptides;

[0041] Data input module, the data of the data input model includes one-dimensional sequence, two-dimensional structure and three-dimensional structure. The data sources in the data input model include but are not limited to APD3, DRAMP and UniProt public databases, which are used to transmit data of antibacterial cyclic peptides;

[0042] A multimodal feature extraction module is used to train a multimodal cross-attention screening model to extract different features of antimicrobial cyclic peptides and map the feature vectors of different modalities to a unified dimension based on a multi-layer perceptron;

[0043] Cross-attention fusion module, which includes an input layer and an attention calculation model, and is used to calculate attention weights and fusion;

[0044] An isomorphic twin network model that compares the similarity of cyclic peptide samples and candidate cyclic peptides in a random screening space based on a contrastive learning module;

[0045] The classification and screening module is used to output the antibacterial activity probability score of the candidate cyclic peptides.

[0046] Furthermore, the one-dimensional sequence, two-dimensional structure and three-dimensional structure are the amino acid linear sequence of the cyclic peptide, the two-dimensional molecular graph representation of the cyclic peptide and the three-dimensional spatial conformation of the cyclic peptide, respectively. The data in the public database are obtained based on keyword screening to obtain antibacterial cyclic peptide data. The strategy of integrating the three-dimensional representations is adopted to carry out antibacterial cyclic peptide screening research, which effectively improves the effect of antibacterial cyclic peptide screening.

[0047] Furthermore, the multimodal feature extraction module trains the feature extraction of the multimodal cross-attention screening model based on the pre-trained model. The multimodal feature extraction module includes one-dimensional sequence features, two-dimensional graph features, and three-dimensional structural features to realize multimodal feature extraction and dynamically assign weights of different modalities, making the key active areas of the antimicrobial cyclic peptide more obvious, thereby enabling the multimodal cross-attention screening model to more accurately identify the relevant features of the antimicrobial activity of the antimicrobial cyclic peptide, making the screening of antimicrobial cyclic peptides more accurate;

[0048] Pre-training module: Utilizes massive data from public databases to quickly learn the common features of cyclic peptides.

[0049] Furthermore, the ESM2 model is used to extract sequence embedding vectors for one-dimensional sequence features, the GROVER model is used to extract molecular graph embeddings for SMILES graphs, and the Uni-Mol model is used to extract PDB space conformation embeddings for three-dimensional structural features. The feature vectors of the three modalities are mapped based on a multi-layer perceptron. Since the representation vector dimensions of different models are different, a multi-layer perceptron is used to align the vectors, which facilitates the subsequent use of the cross-attention mechanism for multimodal feature fusion.

[0050] Furthermore, the input layer is used to transmit the aligned trimodal feature vector to the cross attention fusion module. The trimodal features are represented as F, S, and P. The attention calculation is a two-layer attention fusion. The two-layer attention fusion is:

[0051] First-layer fusion: With F as the key and value, represented as K and V respectively, S as the query, represented as Q, the attention weights are calculated and fused;

[0052] Second layer fusion: The fused features are used as new keys and values, P is Q, and secondary attention fusion is performed.

[0053] Furthermore, the attention weight is calculated as follows:

[0054]

[0055] Among them, Attention is the attention weight, Q, K, V are query, key and value respectively, d k is the characteristic dimension of the key vector K, QK T is the dot product of the query and the key, used to calculate the similarity between the query and each key, Used to scale the dot product results, Softmax is used to normalize the similarity matrix into a probability distribution, Softmax = 1.

[0056] Furthermore, the contrastive learning module is used to input the target cyclic peptide and candidate cyclic peptides in the random screening space, and output the binary classification probability based on the multi-layer perceptron, compare the antibacterial activity similarity between the candidate cyclic peptide and the target cyclic peptide, score the binary classification probability based on the classification and screening module, and screen and classify the cyclic peptides based on the set cyclic peptide threshold. The contrastive learning module is adapted to specific tasks to reduce dependence on experimental data.

[0057] like Figure 3 As shown in a, it is the amino acid sequence information of a certain polypeptide. By using the cysteine ​​disulfide bond to form a ring, the following can be generated at the same time: Figure 3 b and c are two 2D structured cyclic peptide compounds. Therefore, in this case, the model cannot effectively characterize the structure by only considering the one-dimensional sequence data. Similarly, Figure 3 d shows the two-dimensional structure of a certain antibacterial cyclic peptide, but in the three-dimensional structure prediction, the following can be generated at the same time Figure 3 There are two spatial conformations, e and f. According to the universal law that structure determines properties, when using artificial intelligence models to screen antibacterial cyclic peptides, sufficient structural characterization is required to effectively predict properties;

[0058] against Figure 3 Scenario, this system assumes that cyclic peptide 1 is the target cyclic peptide, and cyclic peptide 2 is a compound randomly extracted from the screening space. If the system calculates the similarity and the output is a high probability similarity, it is considered that cyclic peptide 2 and cyclic peptide 1 have the same antibacterial activity with a high probability, otherwise not. This system uses a three-layer perceptron to finally output the binary classification probability. Secondly, for the multimodal cross-attention fusion module, the input layer includes a Fasta sequence based on one-dimensional linear amino acids, a two-dimensional graph structure based on SMILES, and a three-dimensional geometric space structure based on PDB files. The existing pre-trained large models of ESM2, GROVER and Uni-Mol are used to extract the multimodal features of the compound. Since the representation vector dimensions of different models are different, a multi-layer perceptron is used to align the vectors, and then the cross-attention mechanism is used to fuse the multimodal features. Finally, the model convergence is achieved based on the multi-layer perceptron, and the multi-classification results are output.

[0059] Among them, the cross-attention fusion module performs feature fusion. First, input 1 is used as K and V in the attention formula, and input 2 is used as Q to complete the first-layer fusion of inputs 1 and 2. The fused representation vector is used as the new K and V and the Q of input 3 for the second-layer fusion, and finally the fused representation is input.

[0060] The antimicrobial cyclic peptide screening method based on the multimodal cross-attention mechanism is applied to the above-mentioned antimicrobial cyclic peptide screening system based on the multimodal cross-attention mechanism, and includes the following specific steps:

[0061] S1: Construct a multimodal cross-attention screening model. Based on the data input module, a large number of cyclic peptide amino acid linear sequences, cyclic peptide two-dimensional molecular graph representations, and cyclic peptide three-dimensional spatial conformation data are processed and transmitted to the pre-training model. Based on the pre-training model, the multimodal feature extraction module is trained to extract different multimodal features, label different cyclic peptides as antimicrobial peptides, antibacterial peptides, antibacterial cyclic peptides, and antibacterial cyclic peptides, and generate a random screening space.

[0062] S2: In the fine-tuning stage, the non-antimicrobial cyclic peptide data is randomly downsampled to balance the ratio of positive and negative samples. Then, the multimodal feature extraction module is used to extract the trimodal features of the target cyclic peptide. Based on the attention calculation, the attention weight of the cyclic peptide is obtained to capture the features in the cyclic peptide that play a key role in antibacterial activity;

[0063] S3: Map the trimodal features to a unified dimension based on the multi-layer perceptron and fuse the trimodal features, finally achieving model convergence based on the multi-layer perceptron;

[0064] S4: Based on the multi-layer perceptron, the antibacterial activity similarity between the candidate cyclic peptide and the target cyclic peptide is compared, and the binary classification probability is output. The binary classification is divided into similarity / dissimilarity. The binary classification probability is scored based on the classification and screening module, and the set cyclic peptide threshold is compared with the binary classification probability score to screen and classify the cyclic peptide.

[0065] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. Antimicrobial cyclic peptide screening system based on multimodal cross-attention mechanism, characterized by: An antimicrobial cyclic peptide screening system including a multimodal cross-attention mechanism, wherein the antimicrobial cyclic peptide screening system including a multimodal cross-attention mechanism comprises: A multimodal cross-attention screening model for screening antimicrobial cyclic peptides; A data input module, wherein the data of the data input model includes one-dimensional sequence, two-dimensional structure and three-dimensional structure. The data sources in the data input model include but are not limited to APD3, DRAMP and UniProt public databases, which are used to transmit data of antimicrobial cyclic peptides; A multimodal feature extraction module is used to train a multimodal cross-attention screening model to extract different features of antimicrobial cyclic peptides and map the feature vectors of different modalities to a unified dimension based on a multi-layer perceptron; A cross-attention fusion module, which includes an input layer and an attention calculation model for calculating attention weights and fusion; An isomorphic twin network model that compares the similarity of cyclic peptide samples and candidate cyclic peptides in a random screening space based on a contrastive learning module; The classification and screening module is used to output the antibacterial activity probability score of the candidate cyclic peptides.

2. The antimicrobial cyclic peptide screening system based on multimodal cross-attention mechanism according to claim 1, characterized in that: The one-dimensional sequence, two-dimensional structure and three-dimensional structure are respectively the amino acid linear sequence of the cyclic peptide, the two-dimensional molecular graph representation of the cyclic peptide and the three-dimensional spatial conformation of the cyclic peptide. The data in the public database is obtained based on keyword screening to obtain antibacterial cyclic peptide data.

3. The antimicrobial cyclic peptide screening system based on multimodal cross-attention mechanism according to claim 1, characterized in that: The multimodal feature extraction module trains feature extraction of a multimodal cross-attention screening model based on a pre-trained model. The multimodal feature extraction module includes one-dimensional sequence features, two-dimensional graph features, and three-dimensional structural features to achieve multimodal feature extraction.

4. The antimicrobial cyclic peptide screening system based on multimodal cross-attention mechanism according to claim 3, characterized in that: The one-dimensional sequence features, two-dimensional graph features and three-dimensional structural features are respectively extracted using the ESM2 model to extract sequence embedding vectors, the GROVER model to extract the molecular graph embedding of the SMILES graph and the Uni-Mol model to extract the PDB space conformation embedding, and the feature vectors of the three modalities are mapped based on a multi-layer perceptron.

5. The antimicrobial cyclic peptide screening system based on multimodal cross-attention mechanism according to claim 1, characterized in that: The input layer is used to transmit the aligned trimodal feature vector to the cross attention fusion module. The trimodal features are represented as F, S, and P. The attention calculation is a two-layer attention fusion. The two-layer attention fusion is: First-layer fusion: With F as the key and value, denoted as K and V respectively, and S as the query, denoted as Q, attention weights are calculated and fused; Second layer fusion: The fused features are used as new keys and values, P is Q, and secondary attention fusion is performed.

6. The antimicrobial cyclic peptide screening system based on multimodal cross-attention mechanism according to claim 5, characterized in that: The attention weight is calculated as follows: Among them, Attention is the attention weight, Q, K, V are query, key and value respectively, d k is the characteristic dimension of the key vector K, QK T is the dot product of the query and the key, used to calculate the similarity between the query and each key, Used to scale the dot product results, Softmax is used to normalize the similarity matrix into a probability distribution, Softmax = 1.

7. The antimicrobial cyclic peptide screening system based on multimodal cross-attention mechanism according to claim 1, characterized in that: The comparative learning module is used to input the target cyclic peptide and the candidate cyclic peptides in the random screening space, and output the binary classification probability based on the multi-layer perceptron, compare the antibacterial activity similarity between the candidate cyclic peptide and the target cyclic peptide, score the binary classification probability based on the classification and screening module, and screen and classify the cyclic peptides based on the set cyclic peptide threshold.

8. An antimicrobial cyclic peptide screening method based on a multimodal cross-attention mechanism, applied to the antimicrobial cyclic peptide screening system based on a multimodal cross-attention mechanism according to claims 1-7, characterized in that: The specific steps include: S1: Construct a multimodal cross-attention screening model. Based on the data input module, a large number of cyclic peptide amino acid linear sequences, cyclic peptide two-dimensional molecular graph representations, and cyclic peptide three-dimensional spatial conformation data are processed and transmitted to the pre-training model. Based on the pre-training model, the multimodal feature extraction module is trained to extract different multimodal features, label different cyclic peptides as antimicrobial peptides, antibacterial peptides, antibacterial cyclic peptides, and antibacterial cyclic peptides, and generate a random screening space. S2: In the fine-tuning stage, the non-antimicrobial cyclic peptide data is randomly downsampled to balance the ratio of positive and negative samples. Then, the multimodal feature extraction module is used to extract the trimodal features of the target cyclic peptide. Based on the attention calculation, the attention weight of the cyclic peptide is obtained to capture the features in the cyclic peptide that play a key role in antibacterial activity; S3: Map the trimodal features to a unified dimension based on the multi-layer perceptron and fuse the trimodal features, finally achieving model convergence based on the multi-layer perceptron; S4: Based on the multi-layer perceptron, the antibacterial activity similarity between the candidate cyclic peptide and the target cyclic peptide is compared, and the binary classification probability is output. The binary classification is divided into similarity / dissimilarity. The binary classification probability is scored based on the classification and screening module, and the set cyclic peptide threshold is compared with the binary classification probability score to screen and classify the cyclic peptide.

Images