A method for predicting polypeptides based on machine learning

Abstract

The application discloses a method for predicting polypeptides based on machine learning, comprising: predicting the binding site of PEX5 of the receptor; constructing a polypeptide sequence and determining the initial state of the polypeptide; generating the space required for polypeptide and PEX5 binding site collision simulation; performing polypeptide and PEX5 binding site collision simulation; performing data preprocessing; learning the known classified polypeptides through sklearn and outputting the combination of neural networks with high accuracy; reconstructing the neural network from sklearn through the neural network of Pytorch, learning the known classified polypeptides, predicting the unknown classified polypeptides, and outputting the confidence. The application can make accurate prediction of polypeptides by using various computer simulation methods such as molecular simulation and combining machine learning, can sufficiently reduce the order of magnitude of wet experiments, and thus reduce the cost and improve the efficiency. The method can also be used for screening the interaction of other functional polypeptides, drug polypeptides and corresponding target points.

Description

Technical Field

[0001] This invention belongs to the field of biomedical big data processing technology, and relates to a method for predicting peptides based on machine learning. Background Technology

[0002] Since the advent of insulin nearly a century ago, over 80 peptide drugs have entered the market. Drug development, including peptide drugs, involves the identification of lead compounds. Because current knowledge is insufficient to directly synthesize the target drug (peptide) without using existing models, the traditional development of effective peptide drugs relies on screening from potential peptide libraries, which is also a crucial step in the entire traditional peptide drug development process. However, similar to other drug development, the number of effective drugs is far fewer than the vast library. Therefore, traditional drug development has always suffered from the following pain points: long development time, high cost, and high failure rate. In today's world where time is of the essence, there is an urgent need for a low-cost, rapid screening method.

[0003] In recent years, artificial intelligence (AI) has gradually become a popular tool in assisting drug discovery through natural language processing, deep learning, machine learning, and image recognition. This technology primarily screens compounds by extracting valuable information from vast amounts of data in medicinal chemistry and biology. One example is in the development of monopeptide drugs; the 2022 Nature Communications Vol. 13, No. 6325 article, "Combining mass spectrometry and machine learning to discover bioactive peptides," disclosed a technique for discovering lead peptides, predicting hundreds of candidate peptides by combining mass spectrometry and machine learning. However, this method still relies heavily on extensive experimentation. This limitation is particularly pronounced when experimental subjects are limited or drugs are expensive. Therefore, this paper aims to propose a strategy to minimize the involvement of traditional experiments.

[0004] Chinese patent application CN114639440A discloses a machine learning-based prediction method for acylhomoserine lactone synthase, comprising the following steps: (1) collecting publicly available and verified acylhomoserine lactone synthase data to construct a protein dataset of acylhomoserine lactone synthase; (2) extracting feature vectors from the protein dataset using a protein sequence-based feature extraction algorithm; (3) extracting the optimal feature combination from the extracted feature vectors; (4) selecting several machine learning algorithms, establishing models based on the optimal feature combination, and conducting a five-fold cross-validation experiment to compare and evaluate the prediction performance of each model to obtain the optimal machine learning algorithm; constructing a prediction model for acylhomoserine lactone synthase based on the optimal machine learning algorithm and the optimal feature combination; (5) predicting possible acylhomoserine lactone synthases from high-throughput protein data to be predicted based on the established prediction model.

[0005] However, no specific methods have been publicly disclosed for applying machine learning to the development of peptide drugs. Summary of the Invention

[0006] This invention provides a method for predicting peptides based on machine learning. This method uses various computer simulation methods, such as molecular simulation, combined with machine learning to identify peptides, quickly predict potential peptide drugs, and accelerate the research and development of peptide drugs.

[0007] A method for predicting peptides based on machine learning, comprising:

[0008] Step 1: Predict the PEX5 binding site of the receptor using the PyMOL alignment method.

[0009] Step 1 is performed by AutoDock, mimicking the binding of the peptide to the receptor protein PEX5, and includes:

[0010] 1.1 Prepare the structure of receptor PEX5 and then determine the binding site of receptor PEX5;

[0011] 1.2 Impart the correct charge and polar hydrogen bond to the PEX5 binding site of the receptor.

[0012] Step 2: Construct the peptide sequence using PyMOL's fab function and determine the initial state of the peptide.

[0013] Step 3: Use the software autogrid4.2.6 to generate the space required for the collision simulation of the peptide binding site with PEX5.

[0014] Step 4: Use autodock4.2.6 to simulate the collision between the generated peptide and the PEX5 binding site.

[0015] Step 5 involves data preprocessing, including: one-hot encoding of the peptide sequence using AAindex, extracting statistically inconsistencies in the simulation results using box plots, and encoding the collision-related residues on the receptor using one-hot encoding.

[0016] Specifically, it includes:

[0017] 1. Constructing a dataset

[0018] The data contains a total of 16,000,000 (20x20x20x2000) independent data points, consisting of three parts: polypeptide sequences, energy data, and residues involved in the combination.

[0019] 1.1 Obtaining peptide sequences: Based on the physical and chemical properties of amino acids, multiple amino acids are described by multiple features and written as one-hot codes using AAindex.py.

[0020] The characteristics are: the volume of the amino acid, the electron-ion interaction potential of the amino acid, the hydrophobicity of the amino acid, the α-helix strength of the amino acid, and the pi-helix strength of the amino acid.

[0021] 1.2 Obtaining Energy Data: Based on 2000 binding events of the peptide, the energy and score for each binding event were extracted.

[0022] The energy data can be selected based on the specific peptide, including but not limited to: binding free energy, inhibition constant, final intermolecular energy, van der Waals + hydrogen bond + dissolution energy, electrostatic energy, total internal energy, torsional free energy, and AutoDock score.

[0023] 1.3 Residue Combination Based on the residue combinations obtained in step 4, the residue combinations of each peptide binding are statistically analyzed and encoded using one-hot encoding.

[0024] 2. Feature Matrix Optimization

[0025] Step 6: Learn the known classification of peptides using sklearn's MLPClassifier and output a combination of neural networks with high accuracy.

[0026] Data prediction employs a simple prediction neural network using sklearn: the training set is fed into the sklearn neural network, the model uses MLPClassifier, the activation function is tanh, and the decision function is SGD. Cross-validation is performed using a cross-validation function, and the neural network is generated by randomly generating the learning rate and alpha. The high-accuracy neural network is applied and validated using a test set. After validation, the composition of the high-accuracy neural network is output through "clf.n_features_in_", "clf.coefs_", and "clf.n_layers_".

[0027] Step 7: Reconstruct the neural network from sklearn using PyTorch's neural network, learn the known classification of peptides, predict the unknown classification of peptides, and output the confidence level.

[0028] The prediction was performed using a simple predictive neural network based on sklearn, and then reproduced in PyTorch using an nn model. Overfitting was avoided by adding dropout.

[0029] The following parameters are mainly considered when making predictions:

[0030] 1. The energy of peptide-receptor binding, including separation constant, electrostatic energy, estimated binding free energy, final intermolecular energy, final total internal energy, and torsional free energy;

[0031] 2. Structural data on peptide-receptor binding were analyzed using the InterfaceResidues script in PyMol software to determine the binding of residues during peptide-receptor binding.

[0032] Step 6 uses sklearn, a simple and fast iterative neural network generator that can generate excellent neural networks much faster than PyTorch. However, this approach carries the risk of overfitting, which can be significantly reduced by adding dropout to PyTorch in step 7.

[0033] The advantages of this invention are: by employing various computer simulation methods, such as molecular simulation, combined with machine learning, it achieves accurate prediction of the signal peptide at the C-terminus of the peroxisome, significantly reducing the order of magnitude of wet experiments, thereby lowering costs and improving efficiency. This method can also be used to screen the interactions between other functional peptides, drug peptides, and their corresponding targets, improving R&D efficiency and reducing R&D costs. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a method for predicting peroxisome signal peptides based on machine learning.

[0035] Figure 2 This is a computer simulation result of peptide binding to receptor using AutoDock, as shown in the example.

[0036] Figure 3 The example shows the final total internal energy of a portion of the peptide in 100 collisions.

[0037] Figure 4 The diagram shows the results of each candidate peptide exhibiting different energies at interacting residues in the example.

[0038] Figure 5 The results of screening 8,000 polypeptide drugs using the method of this invention are shown as an example. Detailed Implementation

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] The specific steps of the method of the present invention are as follows: Figure 1 As shown, it includes two main parts: data preparation and machine learning, which will be explained in detail below.

[0041] I. Data Preparation

[0042] 1. The binding site of the receptor PEX5 was predicted using PyMOL alignment. This was achieved by simulating the binding mechanism of the peptide to the receptor protein PEX5 using AutoDock, including:

[0043] 1.1 Prepare the structure of receptor PEX5 and then determine the binding site of receptor PEX5.

[0044] Based on the method disclosed in the paper "Predication of Aggregation-Prone Regions in Structured Proteins" published in the journal Molecular Biology, Volume 380, Issue 2, 2008, pages 425-436, the protein structures of the receptor protein PEX5 with signal peptide sites were selected: 3CV0, 3CVL, 3CVN, 3CVP, and 3CVQ. These protein structures were then compared with the structure of PEX5 in Arabidopsis thaliana using the "alignment" function of PyMOL. The cubic space is the binding site for the target receptor PEX5.

[0045] 1.2 Impart the correct charge and polar hydrogen bond to the PEX5 binding site of the receptor.

[0046] Use the "Add Kollman Charge" function in AutoDock Tools to calculate and correct the receptor.

[0047] The charge on the PEX5 binding site is used to obtain the receptor PEX5 binding site with the correct charge.

[0048] 2. Construct peptide sequences using PyMOL's fab function and determine the initial state of the peptide (position, charge, hydrogen bonds).

[0049] Specifically, it includes:

[0050] 2.1 Peptide generation: The "fab" function of PyMOL is used to generate peptides based on their amino acid sequences.

[0051] 2.2 Determine the initial state of peptide binding (position, charge, hydrogen bonds)

[0052] The polypeptide bound to the aligned 3CV0 protein structure is moved outward away from the receptor PEX5 binding site. Using this polypeptide as a reference, align it with the polypeptide generated in 1.2 (alignment method is the same as 1.1), and use autogrid_prepare.py to assign charge and polar hydrogen bonds to the aligned polypeptide.

[0053] 3. Use the software autogrid4.2.6 to generate the space required for the collision simulation of the peptide binding site with PEX5.

[0054] Confirm the spatial parameters required for the simulation and draw the space required for the simulation. Based on the binding site confirmed in 1.1, select the point (-10.425, 2.787, 0.13) in the space, and set the spatial parameters according to the length, width and height (57, 30, 36) and spatial density of 0.375. Use the software autogrid4.2.6 to generate the space required for the simulation.

[0055] 4. Use autodock4.2.6 to simulate the collision between the generated peptide and the PEX5 binding site.

[0056] Specifically, it includes:

[0057] 4.1 Obtain the best combination record

[0058] Using autodock_prepare.py, with the receptor being the PEX5 binding site and the ligand being a peptide, a maximum of 2,500,000 simulations were set. The Lamarckian genetic algorithm was selected as the algorithm for selecting the results. The simulation was performed using autodock 4.2.6, and the best 2,000 bindings of the peptide were recorded.

[0059] 4.2 Converting the best peptide binding records into spatial data of peptides

[0060] Based on the simulation records obtained in 4.1, the spatial data of peptide binding is saved using write_dlg_to_pdb.py.

[0061] 4.3 Calculate the residues involved in the PEX5 binding site based on the peptide spatial data. Based on the peptide spatial data obtained in 4.2, use the pymol script InterfaceResidue.py to obtain the residue combinations involved when the PEX5 binding site binds to the peptide 2000 times.

[0062] II. Machine Learning

[0063] Machine learning consists of two parts: data preprocessing and data prediction, including:

[0064] A. Data Preprocessing

[0065] This includes: one-hot encoding of peptide sequences using AAindex, identifying statistically inconsistencies in simulation results using box plots, and encoding collision-related residues on the receptor using one-hot encoding.

[0066] 5. Learn from known classifications of peptides using sklearn's MLPClassifier and output a combination of neural networks with high accuracy.

[0067] 5.1 Building the dataset

[0068] The data contains a total of 16,000,000 (20x20x20x2000) independent data points, consisting of three parts: polypeptide sequences, energy data, and residues involved in the combination.

[0069] Obtaining peptide sequences: Based on the physical and chemical properties of amino acids, we described 20 amino acids using five characteristics (including: amino acid volume, amino acid electron-ion interaction potential, amino acid hydrophobicity, amino acid α-helix strength, and amino acid pi-helix strength), and compiled them into one-hot codes using AAindex.py. One peptide consists of 15 one-hot codes (3x5).

[0070] Obtaining Simulated Data (Energy Data): Based on the 2000 binding events of the peptide in section 4.1, the energy and score for each binding event are extracted. Each peptide binding event contains 8 simulation-related features: using sklearn's standard scaling function, box plots are generated for the 2000 peptide binding events, and values ​​exceeding the upper or lower bounds are removed. The data for each feature is also standardized and scaled.

[0071] Residue combination identification: Based on the residue combinations obtained in 4.3, statistics are performed to encode the residue combinations that the peptide binds to each time using one-hot encoding.

[0072] 5.2 Feature Matrix Optimization

[0073] The feature matrix from 2.1 was analyzed by PCA to obtain a simplified feature matrix.

[0074] B. Data Prediction:

[0075] 6. Reconstruct the neural network from sklearn using PyTorch's neural network, learn from peptides of known classification, predict peptides of unknown classification, and output the confidence level.

[0076] 6.1 Data Partitioning

[0077] The data from section 2.1 were divided into known classification data and unknown classification data. The classification of the known classification data was verified through experiments, resulting in a total of 300 peptide combinations. These were then randomly divided into a training set (240 peptides) and a test set (60 peptides) at a ratio of 4:1. The remaining 7700 peptides were used as the prediction set.

[0078] 6.2 Simple Predictive Neural Networks using sklearn

[0079] The training set was fed into a neural network in sklearn, using the MLPClassifier model with tanh activation and SGD decision function. Cross-validation was performed using a cross-validation function, and a neural network was generated by randomly generating the learning rate and alpha. The high-accuracy neural network was applied and validated using a test set. After validation, the composition of the high-accuracy neural network was output through "clf.n_features_in_", "clf.coefs_", and "clf.n_layers_".

[0080] Based on the simple prediction neural network in sklearn, this method is reproduced in PyTorch using the nn model, and overfitting is avoided by adding dropout.

[0081] The function is trained using cross-validation and then validated on a test set. The validation scoring rules include accuracy and confidence: accuracy is calculated using an accuracy function, and confidence is calculated using an information entropy function.

[0082] Models with both accuracy and confidence of 80% were used for subsequent predictions. The prediction set was used to predict whether the peptides were good or bad, and the confidence level was used for evaluation.

[0083] In the prediction results, peptides rated as good and with high confidence were considered potential peptide targets in S4. For short-chain peptides with 7 or fewer amino acids, AutoDock software simulation was used. The search method was the Lamarckian 4 genetic algorithm. For peptides with more than 7 amino acids, Brownian dynamics (BD) software simulation was used to place the peptide near the active site region, considering only the case of contact with the active site.

[0084] The prediction results include two parts:

[0085] 1. The energy of peptide-receptor binding, including separation constant, electrostatic energy, estimated binding free energy, final intermolecular energy, final total internal energy, and torsional free energy;

[0086] 2. Structural data on peptide-receptor binding were analyzed using the InterfaceResidues script in PyMol software to determine the binding of residues during peptide-receptor binding.

[0087] The present invention will be further explained below in conjunction with the prediction of peroxisome signal peptides.

[0088] I. Determination of the active site

[0089] Drug binding to a target can be considered a thermodynamic random collision. Specifically, other parts of the target compete with the active site. During drug screening, binding to inactive sites can obscure the specificity of the results; only binding at specific locations can have an effective effect on the target. Therefore, it is first necessary to identify the active region of the target. For cases where the binding of a potential drug to the target is known, at least three existing crystal structures will be used, aligned using PyMOL's alignment tool, and then the effective active site region of the target and potential drug will be predicted using PyMOL's InterfaceResidues script. If the binding situation between the drug and the target is completely unknown, a small number of drugs will be randomly selected to bind to the target, and the effective drugs (up to three) will be crystallized to obtain crystal structures. For targets where it is difficult to obtain crystal structures, PyRosetta and Alpha Fold will be used for prediction, and Lightdock will be used for rapid binding prediction of the drug. Sites with high scores will be selected as potential active sites, and the regions with the highest overlap of these active sites will be selected as the active sites for the next step of target prediction.

[0090] II. Computer Simulation Settings

[0091] The crystal structure of drugs is generated with computer assistance. Specifically, for biological drugs, such as DNA, RNA, or small peptide chains, their crystal structures are synthesized using PyMOL combined with Avogadro. For other chemical drugs, the crystal structure is generated using ChemSpider, or the two-dimensional structure is first drawn using ChemDraw, and then the energy of the structure is minimized by UCSF Chimera to obtain the most stable three-dimensional structure.

[0092] like Figure 2 As shown, the binding of peptides to receptors was simulated using AutoDock. Receptors are represented in black, interacting residues are represented by dark, thick stick structures, and candidate peptides are represented by light-colored, thin stick structures.

[0093] 2.1 For small molecule drugs, computer simulations were performed using Autodock Simulation.

[0094] Here, a small molecule is defined as a molecule that, after the root of the drug molecule has been identified, rotates no more than 70 times. The root is defined as the fixed part of the ligand, and the rotatable "branches" sprout from the root. The root can be an atom that is selected manually, or it can be automatically set to the atom that is considered to be located at the center of the molecule in the most "balanced" overall structure (i.e., the atom with the fewest final rotations).

[0095] The polar hydrogen atoms and Gaussian energies of the ligands and acceptors were assigned via autodock. The search method was a Lamarckian 4 genetic algorithm. All simulations were run three times, 2,500,000 times each, and the top 100 results were reported.

[0096] 2.2. For macromolecular drugs, computer simulations rely on Brownian motion simulations.

[0097] Here, a macromolecule is defined as a molecule that has undergone more than 70 rotations after the root of the drug molecule has been identified. Examples include peptide chains with more than 7 amino acids and double-stranded deoxyribonucleotides with more than 10 base pairs. The Brownian Dynamicssimulation is implemented using Gromacs' Brownian dynamics algorithm. The polar hydrogen atoms and Gaussian energies of the ligand and receptor are assigned via Vega ZZ. Potential drug molecules are placed near the active site region (by human selection), and only cases involving contact with the active site are considered. All simulations were performed three times, 100,000 times each, and the top 100 results are reported.

[0098] 2.3. Results Analysis

[0099] The results include: separation constant (Kd), electrostatic energy, estimated binding free energy, final intermolecular energy, final total internal energy, and torsional free energy. The results also include the binding status of the drug molecule to the target, and further analysis using PyMOL's InterfaceResidue method reveals which amino acid residues are involved in the binding.

[0100] Figure 3 The results show the final total internal energy of some peptides in 100 collisions. One-way ANOVA statistics of the final total internal energy of known good / good peptides show significant differences.

[0101] like Figure 4 As shown, a cluster map displays the final total internal energy representation for a portion of the peptides that interact with certain residues on the receptor, with different gray levels representing different energies.

[0102] III. Drug Sequence

[0103] The sequence composition of drugs is compiled into unique heat data and used as data for machine learning. The amino acids required for peptide drugs are compiled using AAindex, while other drugs, such as DNA and RNA, are compiled according to A, T, C, G, U. Small molecules are taken as monomers, first converted into two-dimensional structures, and then the structures are aligned and compiled according to image processing methods.

[0104] IV. Machine Learning for Drug Prediction

[0105] 4.1 How to process data

[0106] The autodock simulation generated 100 datasets for the drug, each including energy data (electrostatic energy, estimated binding free energy, final intermolecular energy, final total internal energy, and torsional free energy), drug sequence, and target residue involvement. Data from the dry experiment were reorganized into a two-dimensional array to allow machine learning to understand the data composition.

[0107] Data used for machine learning needs to be preprocessed, and box plots are used for analysis. All data must not exceed the upper and lower bounds of the box plot (e.g., ...). Figure 4 (Box plot). Several energies were then subjected to cluster analysis, and the energies were processed using a scaling method that showed the most obvious classification in the reduced dendrogram. Data features were then analyzed by PCA, retaining only the effective components.

[0108] 4.2. The computational model used, how to use the model, and how to optimize the model.

[0109] The machine learning models are derived from scikit-learn. These models include Bayesian regression, logistic regression, neural networks, gradient boosting classifiers, SVC (Sequential Dynamic Classification), and linear SVC. Each learning iteration is performed 50 times. Only when the machine learning results meet the benchmark, or when the best learning model is selected at the end of the 50 iterations, is that model used to predict unknown drugs.

[0110] Formal model training employed a PyTorch-based multilayer-perceptron classification model, with parameters derived from previous sklearn tests. To prevent overfitting, dropout was applied between different layouts. The information entropy of the experiments was used to evaluate the classification reliability.

[0111] 4.3. How to train and test the model

[0112] Good / binding-compatible drugs are represented by 1, and bad / non-binding-compatible drugs are represented by 0. For all samples, 80% of the amino acids are randomly selected as the training group for machine learning, and 20% of the samples are used as the test group to observe machine learning performance. The 80% training group is used for learning after being split by cross-validation. To quantify the performance of machine learning, loss curves and ROC curves are plotted, and the performance of machine learning is calculated as accuracy.

[0113] 4.4. Predictive Drug Evaluation

[0114] Machine learning's drug prediction results comprise two parts: drug classification and the reliability of the classification results. Drugs that meet expectations and have high reliability are selected.

[0115] This example screens 8000 peptide drugs. The screening results are as follows: Figure 5 As shown in the diagram, each line represents a potential candidate, with different colors indicating whether a candidate is good or bad: light white represents a good candidate, and black represents a bad candidate. The certainty of the computer's prediction is quantified by information entropy; the lower the information entropy, the more certain the computer is about the expected outcome.

[0116] Sixteen experimental results were selected here. Among them, there were 7 good results (pure white) and 9 bad / uncertain results (gray or black). Of the 7 good results, the computer was very certain about 4 of them because their information entropy was very low, between 0 and 0.04; the computer was very uncertain about 3 of them because their information entropy was very wide.

[0117] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for predicting peptides based on machine learning, comprising: Step 1: Predict the PEX5 binding site of the receptor using the PyMOL alignment method. Step 2: Construct the peptide sequence using PyMOL's fab function and determine the initial state of the peptide. Step 3: Use the software Autogrid 4.2.6 to generate the space required for the collision simulation of the peptide binding site with PEX5; Step 4: Simulate the collision between the generated peptide and the PEX5 binding site using autodock4.2.6; Step 5, perform data preprocessing, including: one-hot encoding of peptide sequences using AAindex, extracting parts of the simulation results that do not conform to statistics using box plots, and encoding the residues on the receptor that participate in collisions using one-hot encoding; Step 6: Learn the known classification of peptides using sklearn's MLPClassifier and output a combination of neural networks with high accuracy. The learning process for peptides with known classifications includes: feeding the training set into a neural network using sklearn, employing the MLPClassifier model with tanh as the activation function and sgd as the decision function; performing cross-validation using a cross-validation function, generating a neural network by randomly generating the learning rate and alpha, applying a high-accuracy neural network, and validating it using a test set; after validation, the composition of the high-accuracy neural network is output through "clf.n_features_in_", "clf.coefs_", and "clf.n_layers_". Step 7: Reconstruct the neural network from sklearn using PyTorch's neural network, learn the known classification of peptides, predict the unknown classification of peptides, and output the confidence level. When making predictions based on the simple prediction neural network in sklearn, the results are reproduced in PyTorch based on the nn model. Overfitting is avoided by adding dropout. The prediction in step 7 considers the following parameters: 9.1 The energy of peptide-receptor binding, including separation constant, electrostatic energy, estimated binding free energy, final intermolecular energy, final total internal energy, and torsional free energy; 9.2 Structural data on peptide-receptor binding were analyzed using the InterfaceResidues script in PyMol software to determine the binding status of residues when the peptide binds to the receptor.

2. The method for predicting peptides based on machine learning according to claim 1, characterized in that, Step 1 is performed by AutoDock, mimicking the binding of the peptide to the receptor protein PEX5, and includes: 1.1 Prepare the structure of receptor PEX5 and then determine the binding site of receptor PEX5; 1.2 Impart the correct charge and polar hydrogen bond to the PEX5 binding site of the receptor.

3. The method for predicting peptides based on machine learning according to claim 1, characterized in that, In step 6, the known classified polypeptides are independent data sets consisting of 16,000,000 (20x20x20x2000) consisting of polypeptide sequence, energy data, and residues involved in combination.

4. The method for predicting peptides based on machine learning according to claim 3, characterized in that, The polypeptide sequence was obtained by describing multiple amino acids using various features based on their physical and chemical properties, and then writing them as one-hot codes using AAindex.py.

5. The method for predicting peptides based on machine learning according to claim 4, characterized in that, The characteristics are: the volume of the amino acid, the electron-ion interaction potential of the amino acid, the hydrophobicity of the amino acid, the α-helix strength of the amino acid, and the pi-helix strength of the amino acid.

6. The method for predicting peptides based on machine learning according to claim 3, characterized in that, The energy data includes binding free energy, inhibition constant, final intermolecular energy, van der Waals + hydrogen bond + dissolution energy, electrostatic energy, total internal energy, torsional free energy, and AutoDock score.

7. The method for predicting peptides based on machine learning according to claim 3, characterized in that, The residues involved in the combination are statistically analyzed based on the residue combinations obtained in step 4, and the residue combinations that the peptide binds to each time are encoded using one-hot encoding.

Images