Cyclic peptide medicine auxiliary screening method based on multi-channel feature fusion and machine learning model

Abstract

The invention belongs to the technical field of cyclopeptide drug screening, and particularly relates to a cyclopeptide drug auxiliary screening method based on multichannel feature fusion and a machine learning model. Comprising the following steps: collecting a plurality of cyclopeptide samples having a binding effect on the receptor protein as candidate samples, and additionally collecting a plurality of cyclopeptide samples having a target activity effect on the receptor protein as reference samples; respectively acquiring molecular fingerprints, two-dimensional structure information, physicochemical properties and ADMET properties of the reference sample and the candidate sample, and serially connecting and fusing to form a four-channel feature data set; and processing the candidate samples based on the reference samples by adopting a diclustering method or a machine learning method, and screening the candidate samples. According to the method, cyclic peptides are classified from multiple aspects of sequence-structure-biochemistry-toxicology, so that the problem of relatively low screening precision caused by single characteristics is avoided; corresponding auxiliary effects can be provided for new drug research and development and traditional drug effect optimization.

Description

Technical Field

[0001] The present invention belongs to the technical field of cyclic peptide drug screening, and specifically relates to a cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning models. Background Art

[0002] The information disclosed in this background technology section is only intended to enhance understanding of the overall background of the invention and should not necessarily be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art.

[0003] Cyclic peptide drugs have been widely used in the biopharmaceutical field due to their advantages such as strong specificity and low toxicity and side effects. In particular, in the fields of tumor treatment, immunomodulation, and anti-infection, many cyclic peptide drugs have achieved remarkable clinical effects. The development of cyclic peptide drugs often relies on screening from potential cyclic peptide libraries: a successful case is the RaPID technology pioneered by Professor Suga of the University of Tokyo, which significantly expanded the chemical space of traditional cyclic peptide libraries by introducing non-natural amino acids, enabling it to generate a large number of cyclic peptide molecules with high affinity and high specificity. This technology has been successfully used to screen cyclic peptide drugs against a variety of protein targets. However, how to select drugs with significant therapeutic effects from the huge cyclic peptide library has become a major problem in the field of drug research and development.

[0004] In recent years, machine learning (ML) has shown great potential as an auxiliary tool in the discovery and optimization of cyclic peptides. By analyzing existing cyclic peptide data, machine learning models can quickly predict the three-dimensional structure and activity properties of cyclic peptides, thereby shortening the screening cycle. In the field of cyclic peptide screening, machine learning technology has been widely used to analyze sequence-structure-function relationships. Graph neural network (GNN) models can be used to screen large cyclic peptides with high affinity for specific targets from a large cyclic peptide library. Cluster analysis can also complete the screening of cyclic peptides with functional activity potential from the perspective of classification problems. It can be seen that the introduction of artificial intelligence technology has optimized the design and screening process of cyclic peptide drugs. The Chinese patent application publication number CN 116343922 A, "A method for predicting peptides based on machine learning", discloses a method for predicting peptide drugs based on machine learning, which provides a reference for the subsequent application of machine learning technology in the development of peptide and cyclic peptide drugs.

[0005] However, existing technologies are all based on independent features or independent models. It is difficult for independent features to fully describe the overall characteristics of a cyclic peptide, which adds many uncertainties to the screening of cyclic peptide drugs. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention provides a cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model, which can significantly improve the accuracy while rapidly screening potential cyclic peptide drugs.

[0007] In order to achieve the above object, the technical solution of the present invention is:

[0008] In the first aspect, a cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model includes the following steps.

[0009] S1. Collect multiple cyclic peptide samples with a length of 8-15 amino acids as candidate samples, and also collect multiple cyclic peptide samples with target activity against receptor proteins as reference samples;

[0010] S2. Obtain the molecular fingerprint, two-dimensional structural information, physicochemical properties, and ADMET properties of the reference sample and candidate sample, a total of four characteristic parameters, and fuse them in series to form a four-channel reference sample characteristic dataset and a four-channel candidate sample characteristic dataset;

[0011] S3. Using a binary clustering method or a machine learning method, based on the four-channel reference sample feature data set, the four-channel candidate sample feature data set is processed to screen the candidate samples.

[0012] In the second aspect, a cyclic peptide drug-assisted screening device based on multi-channel feature fusion and machine learning model includes:

[0013] The first module is used to collect multiple cyclic peptide samples with a length of 8-15 amino acids as candidate samples, and also to collect multiple cyclic peptide samples with target activity against receptor proteins as reference samples;

[0014] The second module is used to obtain the molecular fingerprint, two-dimensional structural information, physicochemical properties and ADMET properties of the reference sample and candidate sample respectively, a total of four characteristic parameters, and fuse them in series to form a four-channel candidate reference feature dataset and a four-channel candidate sample feature dataset;

[0015] The third module is used to perform binary clustering on the four-channel reference sample feature dataset and the four-channel candidate sample feature dataset using multiple clustering models, and screen out candidate samples with similar distribution and close average distance to the reference samples;

[0016] The fourth module is used to collect multiple cyclic peptides that have no target activity against receptor proteins but have binding effects on receptor proteins from the same cyclic peptide library as the reference sample as negative samples, and to fuse them in series to form a four-channel negative sample feature data set;

[0017] The fifth module is used to train a model using the four-channel reference sample feature data set and the four-channel negative sample feature data set, and use the model to process the four-channel reference sample feature data set to screen candidate samples.

[0018] According to a third aspect, an electronic device includes:

[0019] Memory, used to store computer programs;

[0020] A processor is used to execute the computer program to implement the above-mentioned cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model.

[0021] The beneficial effects of the present invention are:

[0022] By using computer calculation and model prediction methods, the scope of potential cyclic peptide drugs is narrowed, the number of subsequent wet experiments is reduced, and the experimental difficulty is reduced; cyclic peptides are classified from multiple aspects such as sequence-structure-biochemistry-toxicology, avoiding the problem of low screening accuracy caused by single features; different classification models are used for cyclic peptide samples with sufficient or insufficient data sets, making this method applicable to a wider range of cyclic peptide drug screening, and can provide corresponding auxiliary effects for both new drug development and traditional drug efficacy optimization. Based on the method proposed in the present invention, 9 cyclic peptides with inhibitory potential against neutrophil elastase (NE) have been successfully obtained from a cyclic peptide library screened for targeting NE, and 7 of them have been confirmed to have significant NE inhibitory effects through activity verification. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings, 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.

[0024] Figure 1 Schematic diagram of the process of cyclic peptide drug-assisted screening method in Example 1.

[0025] Figure 2 Schematic diagram of the clustering results in Example 1.

[0026] Figure 3 This is a multi-channel feature flow chart in Example 1 and Example 2. DETAILED DESCRIPTION

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0028] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0029] A cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model includes the following steps.

[0030] S1. Collect multiple cyclic peptide samples with a length of 8-15 amino acids as candidate samples, and also collect multiple cyclic peptide samples with target activity against receptor proteins as reference samples;

[0031] S2. Obtain the molecular fingerprint, two-dimensional structural information, physicochemical properties, and ADMET properties of the reference sample and candidate sample, a total of four characteristic parameters, and fuse them in series to form a four-channel reference sample characteristic dataset and a four-channel candidate sample characteristic dataset;

[0032] S3. Using a binary clustering method or a machine learning method, based on the four-channel reference sample feature data set, the four-channel candidate sample feature data set is processed to screen the candidate samples.

[0033] Optionally, the sequence lengths of the cyclic peptides in the reference sample, the candidate sample, and the negative sample are all between 8 and 15 amino acids.

[0034] Optionally, the candidate sample is selected as a cyclic peptide that has a binding effect on a receptor protein; however, in order to expand the scope of selection, it is not limited to whether the cyclic peptide has a binding effect on a receptor protein. Cyclic peptides generated by computational software or cyclic peptides that do not have a binding effect on a specific receptor can be selected as candidate samples for screening.

[0035] Optionally, in S2, the molecular fingerprint refers to the binary ECFP molecular fingerprint; the two-dimensional structural information includes the number of node atoms and the number of boundaries; the physical and chemical property information includes: molecular weight, heavy atom weight, aromatic heavy atom weight, rotatable bonds, number of hydrogen bond acceptors, number of hydrogen bond donors, number of rings, number of aromatic rings, molar refractive index, topological polar surface area, lipid-water partition coefficient, acid dissociation constant and water solubility, a total of 13 items; when using graph neural networks to classify and predict biological molecules, the two-dimensional molecular graph of the biological molecules will be analyzed, and the information such as the number of rings, nodes, boundaries, etc. will be digitized and then handed over to the neural network model for learning. Although the graph neural network was not selected in this scheme, the information in the two-dimensional molecular graph was extracted separately as one of the feature channels to describe the macrocyclic peptide. Since it only includes information in the two-dimensional structure, it is named "two-dimensional structural information".

[0036] Optionally, in S2, the methods for obtaining molecular fingerprints, two-dimensional structural information, physicochemical properties, and ADMET properties include:

[0037] The cyclic peptide sequence was rewritten into SMILES string format using the PepSMI tool of the NovoPro online platform and fine-tuned according to the positions of natural amino acids and modified residues in the cyclic peptide;

[0038] The RDkit tool was used to collect sequence information and two-dimensional structural information of cyclic peptide samples. The sequence information was one item: a binary ECFP molecular fingerprint with a radius of 2 and a length of 2048 bits; the structural information included two items: the number of node atoms and the number of boundaries, as two-dimensional structural information; the physicochemical property information included 13 items: molecular weight, heavy atom weight, aromatic heavy atom weight, rotatable bond, number of hydrogen bond acceptors, number of hydrogen bond donors, number of rings, number of aromatic rings, molar refractive index, topological polar surface area, lipid-water partition coefficient, acid dissociation constant and water solubility; the ADMET properties included 34 items: permeability, intestinal absorption rate, transport protein catalysis rate, transport protein inhibition rate, oral bioavailability, blood-brain barrier passage rate, plasma protein binding rate, and the like. The results were as follows: synthesis rate, catalytic rate and inhibition rate of CYP1A2 metabolizing enzyme, catalytic rate and inhibition rate of CYP2C19 metabolizing enzyme, catalytic rate and inhibition rate of CYP2C9 metabolizing enzyme, catalytic rate and inhibition rate of CYP2D6 metabolizing enzyme, catalytic rate and inhibition rate of CYP3A4 metabolizing enzyme, half-life, acute oral toxicity in rodents, acute oral toxicity in humans, carcinogenicity, AMES mutagenicity, hepatotoxicity, hERG inhibition, androgen receptor activity, estrogen receptor activity, aromatase inhibition, peroxisome proliferator-activated receptor activity, aryl hydrocarbon receptor activity, mitochondrial membrane potential disturbance, p53 pathway activity, antioxidant response element signaling pathway activity, heat shock response signaling pathway activity and ATPase family type 5 domain protein activity.

[0039] The HelixADMET baseline model of the PaddleHelix online platform was used to collect the physicochemical properties and ADMET properties of the cyclic peptide samples. All indicators in the physicochemical properties and ADMET properties were maintained in the range of [0,1]. The closer the value of a characteristic indicator is to 1, the stronger the indicator of the cyclic peptide is, and conversely, the closer it is to 0, the weaker the indicator of the cyclic peptide is.

[0040] Optionally, in S2, serial fusion means saving the data of all channels into the same file, which is then read uniformly by the model for classification and screening.

[0041] Optionally, in S2, the four collected feature parameters are standardized and normalized respectively by the StandardScaler tool and the MinMaxScaler tool, so that the feature parameter values ​​are kept in the interval [0, 1] under the same dimension; finally, they are connected in series and fused to form a four-channel feature data set.

[0042] Optionally, in S3, when the number of reference samples is not greater than 50, a binary clustering method is used to screen candidate samples; when the number of reference samples is greater than 50, a method combining a four-channel negative sample feature dataset and a training model is used to screen candidate samples.

[0043] Optionally, in S3, when the number of reference samples is not greater than 50, a binary clustering method is used to screen candidate samples; including: using multiple clustering models to perform binary clustering respectively to screen candidate samples with similar distribution and close average distance to the reference samples.

[0044] Optionally, the dichotomous clustering model includes K-means clustering, hierarchical clustering, and DBSCAN clustering.

[0045] Optionally, during the clustering process, the characteristic data of the candidate samples are marked with two labels, 0 and 1, respectively. Samples marked as 0 or 1 in all three clustering models are extracted, and all samples with different labels in the three models are screened out. The data of the candidate samples with two labels of 0 and 1 are distributed in the same plane as the reference samples. The reference samples are used to evaluate the two categories of candidate samples 0 and 1. "1" belongs to the category that has an active effect on the receptor protein, and "0" belongs to the category that has no active effect on the receptor protein. Finally, the candidate samples with more similar distribution to the reference sample data in space and closer average distance are cyclic peptides with active efficacy on the receptor protein, and are developed as candidate drugs.

[0046] Optionally, in S3, when the number of reference samples is greater than 50, a method combining the four-channel negative sample feature dataset and the training model is used to screen candidate samples; including:

[0047] The four-channel reference sample feature data set is used as the positive sample, and cyclic peptides with no active efficacy on the receptor protein, binding effect on the receptor protein and sequence length of 8-15 amino acids, which are not less than 10 times the number of the positive samples, are selected from the cyclic peptide library with the same source as the positive sample as the negative sample; 70-80% of the sample data sets are taken from the four-channel positive data set and the four-channel negative data set respectively as the training data set, and the classification model is used for training to learn the feature information of the positive samples and negative samples; after the training is completed, the classification effect of the model is tested with the remaining sample data (i.e., test samples) in the positive data set and the negative data set respectively, and according to the classification results of the test model, a classification model with an accuracy of >90% is screened and obtained for the classification of candidate samples.

[0048] Optionally, the accuracy is the ratio of the number of correctly predicted samples in all test samples to the total number of test samples:

[0049]

[0050] Optionally, in S3, when training the model, one of four classification models, namely, decision tree, random forest, support vector machine and BP neural network, is used for training until the accuracy of the model is not less than 90%.

[0051] Optionally, in S3, multiple classification models are used to process the four-channel candidate sample feature data set respectively, and candidate samples whose classification results are all positive in the multiple classification models are selected.

[0052] A cyclic peptide drug-assisted screening device based on multi-channel feature fusion and machine learning model, comprising:

[0053] The first module is used to collect multiple cyclic peptide samples with a length of 8-15 amino acids as candidate samples, and also to collect multiple cyclic peptide samples with target activity against receptor proteins as reference samples;

[0054] The second module is used to obtain the molecular fingerprint, two-dimensional structural information, physicochemical properties and ADMET properties of the reference sample and candidate sample respectively, a total of four characteristic parameters, and fuse them in series to form a four-channel candidate reference feature dataset and a four-channel candidate sample feature dataset;

[0055] The third module is used to perform binary clustering on the four-channel reference sample feature dataset and the four-channel candidate sample feature dataset using multiple clustering models, and screen out candidate samples with similar distribution and close average distance to the reference samples;

[0056] The fourth module is used to collect multiple cyclic peptides that have no target activity against receptor proteins but have binding effects on receptor proteins from the same cyclic peptide library as the reference sample as negative samples, and to fuse them in series to form a four-channel negative sample feature data set;

[0057] The fifth module is used to train a model using the four-channel reference sample feature data set and the four-channel negative sample feature data set, and use the model to process the four-channel reference sample feature data set to screen candidate samples.

[0058] An electronic device comprising: a memory for storing a computer program;

[0059] A processor is used to execute the computer program to implement the above-mentioned cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model.

[0060] Example 1

[0061] A cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model, such as Figure 1 As shown; the purpose of the active efficacy in this embodiment refers to the neutrophil elastase inhibitory effect; the screening method comprises the following steps:

[0062] S1. 33 cyclic peptide neutrophil elastase inhibitors from the public literature were compiled as reference samples, and their basic information is shown in Table 1.

[0063] Table 1

[0064]

[0065] IC 50 : Half inhibitory concentration, which represents the inhibitor concentration required to inhibit NE protein activity by 50%.

[0066] Cyclic peptides that can target and bind to neutrophil elastase were screened using mRNA display technology. The variable amino acids used in the RaPID system are 8-12 natural amino acids. Screening was performed using magnetic beads with immobilized neutrophil elastase to obtain protein-ligand complexes that bind to NE. Six rounds of screening were repeated, and after completion of the screening, sequencing was performed to obtain a D-type cyclic peptide library (NE-6D) with a library capacity of 11,318,487. These cyclic peptides are all composed of 8-15 amino acids. The frequency of occurrence of the same cyclic peptide sequence in the cyclic peptide library is called the sequence enrichment. The sequences in the cyclic peptide library were arranged from high to low according to the richness, and the top 100 cyclic peptides with the highest enrichment were selected as candidate samples. The original sequences of the candidate samples are shown in Table 2 (the enrichment arrangement order in the table is: the richness of the left column is higher than that of the right column, and the enrichment of the upper sample in the same column is higher than that of the lower sample).

[0067] Table 2

[0068]

[0069] S2. Use the PepSMI tool to rewrite the candidate cyclic peptide sequences into SMILES string format for subsequent experiments; use RDkit and HelixADMET tools to calculate the molecular fingerprints, two-dimensional structural information, physicochemical properties, and ADMET properties of 100 candidate cyclic peptides and reference samples;

[0070] Among them, molecular fingerprint refers to the binary ECFP molecular fingerprint;

[0071] The 2D structural information includes the number of atoms at nodes and the number of boundaries;

[0072] The physicochemical property information includes: molecular weight, heavy atom weight, aromatic heavy atom weight, rotatable bonds, number of hydrogen bond acceptors, number of hydrogen bond donors, number of rings, number of aromatic rings, molar refractive index, topological polar surface area, lipid-water partition coefficient, acid dissociation constant and water solubility, a total of 13 items;

[0073] ADMET properties include 34 items: permeability, intestinal absorption rate, transporter catalytic rate, transporter inhibition rate, oral bioavailability, blood-brain barrier passage rate, plasma protein binding rate, CYP1A2 metabolizing enzyme catalytic rate and inhibition rate, CYP2C19 metabolizing enzyme catalytic rate and inhibition rate, CYP2C9 metabolizing enzyme catalytic rate and inhibition rate, CYP2D6 metabolizing enzyme catalytic rate and inhibition rate, CYP3A4 metabolizing enzyme catalytic rate and inhibition rate, half-life, acute oral toxicity in rodents, acute oral toxicity in humans, carcinogenicity, AMES mutagenicity, hepatotoxicity, hERG inhibition, androgen receptor activity, estrogen receptor activity, aromatase inhibition, peroxisome proliferator-activated receptor activity, aryl hydrocarbon receptor activity, mitochondrial membrane potential interference, p53 pathway activity, antioxidant response element signaling pathway activity, heat shock response signaling pathway activity and ATPase family type 5 domain protein activity.

[0074] Construct a four-channel candidate sample feature dataset (5000 feature data) and a four-channel reference sample feature dataset (1650 feature data). Use the data preprocessing process to standardize, normalize and reduce the dimension of the dataset so that the feature data are kept in the interval of [0, 1] under the same dimension. The parameters are set as follows:

[0075] Data standardization (Z-Score normalization) uses: StandardScaler (with_mean = True, with_std = True);

[0076] Data normalization (linear scaling) uses: MinMaxScaler(feature_range=(0,1));

[0077] Linear dimensionality reduction uses: PCA (n_components=2).

[0078] S3. Since the number of reference samples is less than 50, the two-clustering method is used for screening, including the following steps:

[0079] The candidate cyclic peptide data were input into the K-means clustering, hierarchical clustering, and DBSCAN clustering models in the sklearn.cluster tool for binary clustering analysis, and the parameters were set:

[0080] K-means clustering was performed using: KMeans(init = 'k-means++', n_clusters = 2, max_iter = 500);

[0081] Hierarchical clustering was performed using: AgglomerativeClustering(n_clusters=2, compute_distances=True);

[0082] DBSCAN clustering uses: DBSCAN (eps=1.5, min_samples=1).

[0083] Given two clustering results, we mark them with 0 and 1 respectively. In the three models, there are 40 candidate samples that are all marked as 0, and 37 candidate samples that are all marked as 1. The remaining 23 candidate samples are eliminated. The two marked candidate samples are input together with the reference samples into a two-dimensional plane for display, as shown in the figure. Figure 2 As shown, it is observed that the candidate sample labeled 1 is closer to the reference sample in distribution and distance, so 1 is labeled as a candidate positive sample with active inhibitory potential for NE, and 0 is labeled as a candidate negative sample with no active inhibitory potential for NE.

[0084] The 37 candidate positive samples marked as 1 and the 40 candidate negative samples marked as 0 were sorted from high to low according to their enrichment in the macrocyclic peptide library. Nine samples with the highest enrichment were selected from the candidate positive samples for chemical synthesis, and four samples with the highest enrichment were selected from the candidate negative samples for chemical synthesis, as shown in Table 3, and their inhibitory effects on neutrophil elastase were determined.

[0085] Table 3

[0086]

[0087]

[0088] Finally, 7 of the 9 candidate positive samples showed significant inhibitory effects on neutrophil elastase, with an accuracy of 78%; none of the 4 candidate negative samples showed any inhibitory effect on neutrophil elastase, with an accuracy of 100%. Previously, synthetic verification was performed using only enrichment order, and the screening accuracy of the experiment was only 12.5% ​​(only 1 of the 8 synthetic cyclic peptides showed an inhibitory effect).

[0089] Example 2

[0090] In this embodiment, the purpose of screening cyclic peptide drugs is to have anti-Gram-negative bacteria activity; the screening method includes the following steps:

[0091] S1. From the CyclicPepedia antimicrobial peptide database, 100 cyclic peptides with significant inhibitory effects on Gram-negative bacteria were selected as positive samples, and no less than 1,000 cyclic peptides with no active effects on the target and a sequence length of 8-15 amino acids were selected as negative samples, all of which were composed of 8-15 amino acids; at the same time, 9 cyclic peptides with significant inhibitory effects on Gram-negative bacteria (inhibitory samples) and 9 cyclic peptides without inhibitory effects on Gram-negative bacteria (non-inhibitory samples) as shown in Table 4 were randomly selected, and their active efficacy labels were erased to serve as 18 candidate cyclic peptide samples.

[0092] Table 4

[0093]

[0094] S2. Calculate the molecular fingerprints, two-dimensional structural information, physicochemical properties, and ADMET properties of 100 positive samples using RDkit and HelixADMET tools.

[0095] Among them, molecular fingerprint refers to the binary ECFP molecular fingerprint;

[0096] The 2D structural information includes the number of atoms at nodes and the number of boundaries;

[0097] The physicochemical property information includes: molecular weight, heavy atom weight, aromatic heavy atom weight, rotatable bonds, number of hydrogen bond acceptors, number of hydrogen bond donors, number of rings, number of aromatic rings, molar refractive index, topological polar surface area, lipid-water partition coefficient, acid dissociation constant and water solubility, a total of 13 items;

[0098] ADMET properties include 34 items: permeability, intestinal absorption rate, transporter catalytic rate, transporter inhibition rate, oral bioavailability, blood-brain barrier passage rate, plasma protein binding rate, CYP1A2 metabolizing enzyme catalytic rate and inhibition rate, CYP2C19 metabolizing enzyme catalytic rate and inhibition rate, CYP2C9 metabolizing enzyme catalytic rate and inhibition rate, CYP2D6 metabolizing enzyme catalytic rate and inhibition rate, CYP3A4 metabolizing enzyme catalytic rate and inhibition rate, half-life, acute oral toxicity in rodents, acute oral toxicity in humans, carcinogenicity, AMES mutagenicity, hepatotoxicity, hERG inhibition, androgen receptor activity, estrogen receptor activity, aromatase inhibition, peroxisome proliferator-activated receptor activity, aryl hydrocarbon receptor activity, mitochondrial membrane potential interference, p53 pathway activity, antioxidant response element signaling pathway activity, heat shock response signaling pathway activity, and ATPase family type 5 domain protein activity;

[0099] The data preprocessing step is used to standardize, normalize and reduce the dimension of the data set so that the feature data are kept in the interval [0, 1] under the same dimension. The parameter settings are:

[0100] StandardScaler(with_mean=True,with_std=True);

[0101] MinMaxScaler(feature_range=(0,1));

[0102] PCA (n_components=2).

[0103] The final series fusion constitutes a four-channel positive sample dataset.

[0104] The same method is used to obtain a four-channel candidate sample dataset.

[0105] S3. Since this example has a lot of reference data (more than 50), a training model is selected; including:

[0106] 1000 cyclic peptides without Gram-negative bacteria inhibitory efficacy were randomly selected from the CyclicPepedia antimicrobial peptide database as negative samples. They were all composed of 8-15 amino acids. The four-channel negative sample dataset was obtained using the same method as S2. The four-channel positive sample dataset and the four-channel negative sample dataset included a total of 1100 training samples.

[0107] 80% of the sample data from the four-channel positive sample data set and the four-channel negative sample data set were taken as training data sets. A total of 880 training set samples were labeled with positive and negative labels and input into four training models in the sklearn tool: decision tree (DT), random forest (RF), support vector machine (SVM) and BP neural network (BP) to learn the feature information of positive and negative samples;

[0108] After model training is completed, 20% of the sample data that did not participate in the training in the four-channel positive sample data set and the four-channel negative sample data set are used as the test set, totaling 220 test set samples. The test set samples are not labeled and randomly shuffled and input into the model for classification testing. A five-fold cross-validation is used to evaluate the model and output a classification report. Set the parameters:

[0109] Random Forest (RF) uses: RandomForestClassifier(n_estimators=100,random_state=42);

[0110] Decision Tree:

[0111] DecisionTreeClassifier(random_state=42);

[0112] Support vector machine: 'SVM': SVC (random_state = 42);

[0113] The BP neural network uses: 'BP NeuralNetwork':MLPClassifier(random_state=42,max_iter=1000);

[0114] Cross validation uses: kf = KFold (n_splits = 5, shuffle = True, random_state = 42).

[0115] The output classification report is shown in Table 2:

[0116]

[0117]

[0118] The results showed that the classification effect of the test samples in the random forest model was the best, reaching 92.90%, and all four classification models were able to accurately screen the cyclic peptides against Gram-negative bacteria in the CyclicPepedia database.

[0119] The selected sample data set obtained in S2 was processed using the model in Example 2, and the intersection of the screening results of different models was further taken. The results were as follows: Among the 18 candidate samples in Table 4, CP00038, CP00418, CP00788, CP00998, CP01044, CP01057, and CP01505 were predicted to be cyclic peptides with inhibitory efficacy against Gram-negative bacteria, and the rest were predicted to have no inhibitory efficacy, which matched the original labels. The prediction accuracy of antibacterial samples was 77.78%, and the prediction accuracy of non-antibacterial samples was 100%.

[0120] The above examples all illustrate that multi-channel feature fusion and machine learning models can assist in the efficient and accurate screening of cyclic peptide drugs. It should be understood that the specific embodiments described herein are merely illustrative of the present invention. Researchers skilled in the art of the present invention may make corresponding modifications to the described specific embodiments without departing from the scope of the present invention.

Claims

1. A cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model, characterized in that: The steps are as follows: S1. Collect multiple cyclic peptide samples with a length of 8-15 amino acids as candidate samples, and also collect multiple cyclic peptide samples with target activity against receptor proteins as reference samples; S2. Obtain the molecular fingerprint, two-dimensional structural information, physicochemical properties, and ADMET properties of the reference sample and candidate sample, a total of four characteristic parameters, and fuse them in series to form a four-channel reference sample characteristic dataset and a four-channel candidate sample characteristic dataset; S3. Using a binary clustering method or a machine learning method, based on the four-channel reference sample feature data set, the four-channel candidate sample feature data set is processed to screen the candidate samples.

2. The cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model according to claim 1, characterized in that: The sequence lengths of the cyclic peptides in the reference samples, candidate samples, and negative samples were all between 8 and 15 amino acids.

3. The cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model according to claim 1, characterized in that: In S2, the molecular fingerprint refers to the binary ECFP molecular fingerprint; the two-dimensional structural information includes the number of node atoms and the number of edges; the physical and chemical property information includes: molecular weight, heavy atom weight, aromatic heavy atom weight, rotatable bonds, number of hydrogen bond acceptors, number of hydrogen bond donors, number of rings, number of aromatic rings, molar refractive index, topological polar surface area, lipid-water partition coefficient, acid dissociation constant, and water solubility; ADMET properties include: permeability, intestinal absorption rate, transporter catalytic rate, transporter inhibition rate, oral bioavailability, blood-brain barrier passage rate, plasma protein binding rate, CYP1A2 metabolizing enzyme catalytic rate and inhibition rate, CYP2C19 metabolizing enzyme catalytic rate and inhibition rate, CYP2C9 metabolizing enzyme catalytic rate and inhibition rate, CYP2D6 metabolizing enzyme catalytic rate and inhibition rate, CYP3A4 metabolizing enzyme catalytic rate and inhibition rate, half-life, acute oral toxicity in rodents, acute oral toxicity in humans, carcinogenicity, AMES mutagenicity, hepatotoxicity, hERG inhibition, androgen receptor activity, estrogen receptor activity, aromatase inhibition, peroxisome proliferator-activated receptor activity, aryl hydrocarbon receptor activity, mitochondrial membrane potential interference, p53 pathway activity, antioxidant response element signaling pathway activity, heat shock response signaling pathway activity, and ATPase family type 5 domain protein activity; Alternatively, in S2, methods for obtaining molecular fingerprints, two-dimensional structural information, physicochemical properties, and ADMET properties include: The cyclic peptide sequence was rewritten into SMILES string format using the PepSMI tool of the NovoPro online platform and fine-tuned according to the positions of natural amino acids and modified residues in the cyclic peptide; The RDkit tool was used to collect the sequence information and two-dimensional structure information of the cyclic peptide samples, and the HelixADMET baseline model of the PaddleHelix online platform was used to collect the physicochemical properties and ADMET properties of the cyclic peptide samples; The obtained features are standardized and normalized by the StandardScaler tool and the MinMaxScaler tool respectively, so that the feature parameter values ​​are kept in the [0, 1] interval under the same dimension; finally, they are connected in series and fused to form a four-channel feature data set.

4. The cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model according to claim 1, characterized in that: In S3, when the number of reference samples is not greater than 50, a binary clustering method is used to screen candidate samples; this includes: using multiple clustering models to perform binary clustering to screen out candidate samples that have similar distributions and close average distances to the reference samples.

5. The cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model according to claim 4, characterized in that: The two-clustering models include K-means clustering, hierarchical clustering, and DBSCAN clustering; Alternatively, during the clustering process, the characteristic data of the candidate samples are marked with two labels, 0 and 1, respectively. Samples marked as 0 or 1 in all three clustering models are extracted, and all samples with different labels in the three models are screened out. The data of the candidate samples with the two labels of 0 and 1 are distributed in the same plane as the reference samples. The reference samples are used to evaluate the two categories of candidate samples 0 and 1. "1" belongs to the category with active effect on the receptor protein, and "0" belongs to the category with no active effect on the receptor protein. Finally, the candidate samples with similar distribution and close average distance to the reference sample data in space are regarded as cyclic peptides with active efficacy on the target receptor protein and are developed as candidate drugs.

6. The cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model according to claim 1, characterized in that: In S3, when the number of reference samples is greater than 50, a method of combining a four-channel negative sample feature data set and a training model is used to screen candidate samples; the method includes: using the four-channel reference sample feature data set as a positive sample, selecting no less than 10 times the number of positive samples from a cyclic peptide library with the same source as the positive sample, and having no active effect on the receptor protein and a sequence length of 8-15 amino acids as a negative sample; taking 70-80% of the sample data sets from the four-channel positive data set and the four-channel negative data set as training data sets, respectively, using a classification model for training, and learning the characteristic information of positive samples and negative samples; after the training is completed, the classification effect of the model is tested with the remaining sample data in the positive data set and the negative data set, and according to the classification results of the test model, a classification model with an accuracy of >90% is screened and obtained for the classification of candidate samples.

7. The cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model according to claim 6, characterized in that: In S3, the model training method includes training one of the four classification models: decision tree, random forest, support vector machine, and BP neural network; Alternatively, in S3, multiple classification models are used to process the four-channel candidate sample feature dataset, and candidate samples whose classification results are positive in multiple classification models are selected.

8. The cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model according to claim 7, characterized in that: Multiple classification models are used to process the four-channel candidate sample feature dataset, and candidate samples whose classification results are positive in multiple classification models are selected.

9. A cyclic peptide drug-assisted screening device based on multi-channel feature fusion and machine learning model, characterized in that: include: The first module is used to collect multiple cyclic peptide samples that have binding effects on receptor proteins as candidate samples, and also to collect multiple cyclic peptide samples that have target active effects on receptor proteins as reference samples; The second module is used to obtain the molecular fingerprint, two-dimensional structural information, physicochemical properties and ADMET properties of the reference sample and candidate sample respectively, a total of four characteristic parameters, and fuse them in series to form a four-channel candidate reference feature dataset and a four-channel candidate sample feature dataset; The third module is used to perform binary clustering on the four-channel reference sample feature dataset and the four-channel candidate sample feature dataset using multiple clustering models, and screen out candidate samples with similar distribution and close average distance to the reference samples; The fourth module is used to collect multiple cyclic peptides that have no target activity against receptor proteins (but have binding effects on receptor proteins) from the same cyclic peptide library as the reference sample as negative samples, and fused them in series to form a four-channel negative sample feature data set; The fifth module is used to train a model using the four-channel reference sample feature data set and the four-channel negative sample feature data set, and use the model to process the four-channel reference sample feature data set to screen candidate samples.

10. An electronic device, characterized in that: include : Memory, used to store computer programs; A processor, configured to execute the computer program to implement the cyclic peptide drug-assisted screening method based on multi-channel feature fusion and machine learning model as described in any one of claims 1 to 8.

Images