A method for predicting unknown small molecule ligand targets and its application
Through Bayesian Gaussian mixed model and three-dimensional coordinate conversion technology, the interaction modes of protein and small molecule ligand are identified, and the problems of low true positive and high false positives in the prediction of traditional Chinese medicine targets in the prior art are solved, achieving high accuracy target prediction has promoted the research and development of new Chinese medicine drugs.
Patent Information
- Application Number
- CN202111245650.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2021-10-06
- Filing Date
- 2021-10-26
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2041-10-26
AI Technical Summary
When predicting traditional Chinese medicine targets, the existing molecular simulation docking technology has a low true positive rate and a high false positive rate, which leads to insufficient understanding of the mechanism of drug action and the difficulty of rational use of drugs in clinical practice.
The Bayesian Gaussian mixed model is used to combine three-dimensional coordinate transformation and clustering analysis to identify the interaction modes of proteins and small molecule ligands, reduce the false positive rate, and improve the accuracy of target prediction.
It effectively reduces the false positive rate, improves the accuracy of target prediction, and hits up to 80%, providing new methods and tools for the research and development of new Chinese medicines.
Smart Images

Figure CN113963744B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computational toxicology, and in particular to determining whether the protein fragment is highly likely to be the target of the small molecule atom based on computer simulation of the interaction between a ligand molecule (e.g., a small molecule ligand atom of a traditional Chinese medicine) and a target receptor (e.g., a protein fragment). Background Art
[0002] "Software for molecular docking: a review. DOI: 10.1007 / s12551-016-0247-1" introduces: using computer simulation of the binding mode of proteins with small molecule ligands to predict the targets of small molecule ligands. Molecular docking is one of the important methods of molecular simulation. Its essence is the recognition process between two molecules, and its process involves spatial matching and energy matching between molecules. As more and more protein structures are experimentally determined by X-ray crystallography or nuclear magnetic resonance (NMR) spectroscopy, molecular docking is increasingly used as a tool for drug discovery. For proteins with unknown structures, docking with homology model targets is also possible. Through docking strategies, the druggability of compounds and their specificity for specific targets can be calculated for further lead optimization processes. Molecular docking programs implement the scarch algorithm, in which the conformations of the ligand are recursively calculated until they converge to the minimum energy. Finally, the affinity scoring function ΔG [U total in kcal / mol] is used to rank the candidate conformations as the sum of electrostatic energy and van der Waals energy. As a result, the true positive rate of targets predicted by molecular docking is generally 5% to 10%, and the false positive rate exceeds 90%, which seriously hinders people's understanding of the drug's mechanism of action and is not conducive to rational clinical drug use.
[0003] Although there are many docking programs for conformational searching and bound conformation prediction, the scoring functions are not accurate and need further improvement. Although each docking strategy has shortcomings, active reinvestigation is ongoing.
[0004] Protein-ligand docking is the study of how to predict the stable spatial structure of protein-ligand complexes through computational techniques. It is currently an important tool in network pharmacology and an active research field that has been the subject of important developments in the past decade. This is well described by the increasing number of available protein-ligand docking software programs, the increasing sophistication of their latest applications, and the increasing number of users. At the same time, protein-ligand docking can be formulated as a search algorithm associated with an accurate scoring function, the goal of which is to determine the conformation of the ligand and receptor with the lowest energy.
[0005] At present, the targets of most traditional Chinese medicines are unclear. Target determination involves three elements: protein, traditional Chinese medicine components, and protein-traditional Chinese medicine component interactions. Traditional target prediction algorithms only consider the lowest energy conformation, without considering the three-dimensional structural characteristics of protein-traditional Chinese medicine component complexes and protein-traditional Chinese medicine component interactions in real space. Target information and interaction information between traditional Chinese medicine components and targets are ignored. The accuracy of the docking binding site in the docking process of protein-ligand docking will directly affect the size of the sampling space when the receptor and ligand are docked. However, there is currently no unified, high-precision binding site prediction method. As a result, the true positive rate of identifying traditional Chinese medicine targets using different algorithms is generally 5%-10%, and the false positive rate exceeds 90%, which seriously hinders people's understanding of the mechanism of action of traditional Chinese medicine and is not conducive to rational clinical drug use. Summary of the invention
[0006] In view of the gaps and deficiencies in the current research on Chinese medicine targets, based on the wide application of network pharmacology in the prediction of Chinese medicine targets, taking the prediction of Chinese medicine targets as the starting point, we conducted in-depth research on the spatial structure of proteins and the spatial distribution characteristics of protein ligands and constructed machine learning models. We developed a Chinese medicine target prediction method that can effectively reduce false positives, reveal the mechanism of action of Chinese medicine, and provide new methods and tools for the research and development of new Chinese medicines.
[0007] The technical solution of the present invention is: a method for predicting unknown small molecule ligand targets, characterized in that it comprises the following steps:
[0008] Step 1. Classify the ligand-protein fragment complexes that have interactions and the same ligand atom type and the same protein fragment type into one category, transform the three-dimensional coordinates of this type of ligand-protein fragment complexes so that each protein fragment in this category can be placed in the same three-dimensional space coordinate system, and then superimpose this type of ligand-protein fragment complex based on the protein fragment structure to construct a superposition model of this category, and use the Bayesian Gaussian mixture model to cluster the ligand atoms distributed around the protein fragments in each type of superposition model, delete the class with less than 20 ligand atoms in the clustering results, and use 95% confidence ellipsoids to represent the spatial distribution of the ligand atoms that have not been deleted in each type of superposition model:
[0009]
[0010] Store the spatial distribution μ x , μ y , μ z , σ x , σ y , σ z , where μ x , μ y , μ zRefers to the average value of the x, y, and z axis coordinates of all ligand atoms that have not been deleted in various superposition models, σ x , σ y , σ z Refers to the standard deviation of the x, y, and z axis coordinates of all ligand atoms that have not been deleted in various superposition models. D(x, y, z) is the distance between the ligand atom and the 95% confidence ellipsoid, and x, y, z are the coordinates of the ligand atom: LIGAND: (Xligand, Yligand, Zligand);
[0011] Step 2. Input the simulated binding conformation data of the protein fragment and the unknown small molecule ligand, i.e., the file data containing the three-dimensional structure of the protein-unknown small molecule ligand. The format of the file data is .pdb, which is a standard file format for the spatial structure of the protein and the unknown small molecule ligand, and contains the coordinate information of each atom in the three-dimensional space coordinate system. The simulated binding conformation data is generated by docking the protein fragment and the unknown small molecule ligand using molecular simulation docking software;
[0012] Step 3. Process the binding conformation data in step 2: Use SYBYL atom type to mark each atom of the unknown small molecule ligand, split the protein into protein fragments, each protein fragment is an amino acid residue, and three heavy atoms are taken from the covalently linked heavy atoms of each amino acid residue for arrangement and combination. Each arrangement and combination of three heavy atoms is a protein fragment classification. Each protein fragment classification is named according to the name of the amino acid residue and the arrangement order of the three heavy atoms, so that the arrangement of the three heavy atoms in each protein fragment is different, or the type of heavy atoms is different, and each of the three heavy atoms is one of C, N, O, and S;
[0013] Step 4. Calculate the atomic distance between the first atom in the sequence of the three heavy atoms of the protein fragment and each atom of the unknown small molecule ligand. The protein fragment and the unknown small molecule ligand atom are called interaction pairs. The number of ligand-protein fragment complexes in each group with the same ligand atom type and the same protein fragment type is counted in all interaction pairs and recorded as M. The ligand-protein fragment complexes in the M interaction pairs of each group are transformed into three-dimensional coordinates so that each protein fragment of the same type can be placed in the same three-dimensional space coordinate system. According to the atomic types of the protein fragment and the unknown small molecule ligand in the M interaction pairs of each group, the spatial distribution corresponding to the type in step 1 is searched, and the value is extracted: μ x , σ x , μ y , μ z , σ y , σ z, based on the coordinates of the converted unknown small molecule ligand atoms of each group, the distance D(x, y, z) between the unknown small molecule ligand atoms of each group and the 95% confidence ellipsoid obtained in step 1 is calculated according to formula 1, and the total number of ligand-protein fragment complexes satisfying D(x, y, z) ≤ 1 in each group is counted as N, and the hit rate of the interaction pattern of each group is calculated: P = N / M;
[0014] Step 5. If N≥600 and P≥0.8 in each group in step 4, the protein fragments in the group are targets of the unknown small molecule ligand.
[0015] Preferably, the covalently linked heavy atoms of the amino acid residues include: heavy atoms in the peptide bond between two amino acids bound to each other; oxygen atoms in the terminal carboxyl group of an amino acid; side chain atoms in a disulfide bridge between two adjacent cysteines; heavy atoms remaining after the amide group of the side chain of asparagine, lysine, glutamine or alanine reacts with the α-carboxyl group of other amino acids; heavy atoms in the peptide bond formed by the pyrrolidine ring of proline with other amino acids; amide bonds formed by the carboxyl group on the side chain of glutamic acid or aspartic acid with other amino acids, and heavy atoms in the amide bond formed between the α-carboxyl group of each amino acid and the side chain amino group.
[0016] Preferably, step 1 includes the following steps:
[0017] S1. Sample data set construction: PDB files from the PDB protein structure database from January 1995 to the present, the PDB files were screened using the following criteria: (1) proteins from humans; (2) proteins with crystal structures (3) Crystal structures without antibodies; (4) Crystal structures without mutations; (5) Crystal structures with at least one ligand having a heavy atom number greater than 5 and a molecular weight of ≥80Da and <800Da; (6) Crystal structures with at least one polypeptide having an amino acid residue number greater than 30; (7) Crystal structures with at least one ligand molecule having a relative accessible surface area of <0.6 between the complex state and the isolated state; the resulting PDB files are constructed into a ligand-protein fragment complex dataset;
[0018] The Reduce program in pdb-tools was used to add hydrogen to the ligand-protein fragment complex selected by conditions (1) to (6), and then dr-sasa was used to calculate the accessible surface area.
[0019] Preferably, the ligand-protein fragment complex having interaction and the same ligand atom type and the same protein fragment type refers to the ligand atom having the same SYBYL atom type label, the three heavy atoms covalently bonded in the protein fragment being the same and in the same order, and the first atom in the arrangement order of the three heavy atoms of the protein fragment being the same as the ligand atom. The interatomic distances are calculated using the pairwise_dist function in PyMOL: the ligand atom identities are entered and the maximum distance is set to The pairwise_dist function will output The distances of all protein atoms within the range and their relative ligand atoms.
[0020] Preferably, transforming the three-dimensional coordinates of the ligand-protein fragment complex comprises:
[0021] The ligand atom in the interacting ligand-protein fragment complex is set as LIGAND, and the three covalently connected heavy atoms in the protein fragment are set as ATOM_1, ATOM_2, and ATOM_3:
[0022] Then translate the coordinates of the four atoms until the coordinates of ATOM_2 are at the origin of the coordinate system (0,0,0);
[0023] Rotate the four atoms around the x-axis until the coordinates of ATOM_1 are in the xz plane and z>0;
[0024] Then rotate the four atoms around the y-axis until the coordinate of ATOM_1 is on the z-axis (z>0);
[0025] Then rotate the four atoms around the z-axis until the coordinates of ATOM_3 are in the xz plane;
[0026] The final coordinates of the 4 atoms are: ATOM_1: (0,0,Zatom_1), ATOM_2: (0,0,0), ATOM_3: (Xatom_3,0,Zatom_3), and LIGAND: (Xligand,Yligand,Zligand).
[0027] Preferably, if the number of ligand atoms in the superposition model in step 1 is less than 100, the superposition model is deleted.
[0028] From the above, the method for predicting unknown small molecule ligand targets is applied in predicting the targets of the main components of traditional Chinese medicine.
[0029] Based on the above, the method for predicting unknown small molecule ligand targets is applied in predicting the targets of the main components of Nux vomica.
[0030] From the above, the method for predicting unknown small molecule ligand targets is applied in predicting the targets of the main components of Astragalus.
[0031] The beneficial effects of the present invention are:
[0032] Strychnine is a commonly used Chinese medicine in clinical practice. It has an analgesic effect and can be used to treat many common tumors such as liver cancer, lung cancer, breast cancer, and colorectal cancer. Brucine, Strychnine, and icajine are the main components of Strychnine. The existing technology (molecular docking technology) was used to predict the targets of the three main components of Strychnine, and 200 targets were found. The analysis showed that more than 100 of the targets had no clinical significance, that is, the false positive rate was at least 50%.
[0033] The prediction method of the present invention was used to further calculate the interaction patterns between the three main components of Strychnos nux vomica and 200 targets. Under the conditions of N≥600 and P≥0.8, it was found that the three main components of Strychnos nux vomica had 23, 19 and 9 targets, respectively. By searching relevant academic papers, we found that these targets were all related to tumor treatment and analgesia, indicating that the target prediction method designed by the present invention has an effective hit rate of 80%, reducing the false positive rate of molecular simulation docking results.
[0034] Target prediction for the main components of other traditional Chinese medicines such as astragalus also reduced the false positive rate.
[0035] Traditional methods usually use computers to simulate the binding mode between proteins and small molecule ligands to predict the targets of small molecule ligands. Molecular simulation docking is one of the important methods of molecular simulation. Its essence is the recognition process between two molecules, which involves spatial matching and energy matching between molecules.
[0036] Pagadala mentioned in the paper "Software for molecular docking: a review" (DOI: 10.1007 / s12551-016-0247-1) that traditional methods do not take into account the characteristic information of the interaction between proteins and small molecules in the real world, resulting in the true positive rate of targets predicted by molecular docking being generally 5% to 10%, and the false positive rate exceeding 90%, which seriously hinders people's understanding of the mechanism of drug action and is also not conducive to rational clinical drug use.
[0037] The present invention proposes a new method for predicting unknown small molecule ligand targets. The method is based on the structure of protein and compound small molecule interactions in the real world, and uses the Bayesian Gaussian mixture model (BGMM) to identify the interaction pattern between protein and compound small molecule. Therefore, on the basis of molecular docking prediction of small molecule ligand targets, the protein and compound small molecule interaction pattern is further used to screen targets, which ultimately greatly reduces the false positive rate of target prediction and improves the accuracy of prediction results (~75%). BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1This is a comparison chart of SYBYL atom types and their meanings;
[0039] Figure 2 The diagram is the interaction pattern between typical protein fragments and small molecule ligand atoms obtained by fitting with Bayesian Gaussian mixture model (BGMM);
[0040] Figure 3 This is a diagram of the PDB file format;
[0041] Figure 4 This is the protein conformation analysis result table;
[0042] Figure 5 The process and result diagram for predicting the target of traditional Chinese medicine Strychnos nux vomica;
[0043] Figure 6 These are the atomic names of the 20 amino acids. DETAILED DESCRIPTION
[0044] The present invention is further described in detail below in conjunction with specific examples. Unless otherwise specified, the experimental methods described in the examples are performed according to conventional molecular biology experimental methods.
[0045] Example 1
[0046] 1. Construction of target prediction method:
[0047] S1. Sample dataset construction
[0048] Download 176,773 PDB files from January 1995 to April 2021 (or until now, then the number of PDB files will increase) from the PDB protein structure database (Protein Data Bank, PDB for short). PDB was created by the Brookhaven National Laboratory in the United States in 1971 and is currently the most important database for collecting the three-dimensional structures of biological macromolecules (proteins, nucleic acids and sugars). The data therein are three-dimensional structural data of biological macromolecules such as proteins, polysaccharides, nucleic acids, viruses, etc. determined by experimental means such as X-ray single crystal diffraction, nuclear magnetic resonance, and electron diffraction. The PDB database stores data in the .pdb file format, and each molecule has an independent .pdb format file (called the PDB file of the compound).
[0049] These PDB files were screened using the following criteria: (1) proteins from human origin; (2) proteins with crystal structures (3) Crystal structures without antibodies; (4) Crystal structures without mutations; (5) Crystal structures with at least one ligand with a heavy atom number greater than 5 and a molecular weight of ≥80Da and <800Da; (6) Crystal structures with at least one peptide with an amino acid residue number greater than 30; (7) Crystal structures with at least one ligand molecule with a relative accessible surface area of <0.6 between the complex state and the isolated state. Finally, 9063 PDB files were obtained and constructed into a ligand-protein fragment complex data set.
[0050] The Reduce program in pdb-tools is used to add hydrogen. After adding hydrogen to the ligand-protein fragment complex selected by conditions (1) to (6), dr-sasa is used to calculate the accessible surface area.
[0051] Reduce is a program that adds hydrogen to molecular structure files in the Protein Data Bank (PDB). 3+ , Met methyl, Asn (asparagine) and Gln (glutamine) side chain amides, and the His ring (histidine) orientations with hydrogens added according to standardized geometry. dr-sassa is a software for calculating the solvent accessible surface area (SASA) of biomolecules using the Shrake & Rupley algorithm. The Shrake-Rupley algorithm is a numerical method that draws a grid of points equidistant from every atom in a molecule and uses the number of these solvent accessible points to determine the surface area.
[0052] Identification of atomic interactions between S2 protein fragments and small molecule ligands
[0053] S2-1 Label each atom of the ligand with the SYBYL atom type, i.e., each atom of the small molecule ligand ( Figure 1 SYBYL atom types and their meanings).
[0054] S2-2 splits the protein into protein fragments, each protein fragment is an amino acid residue, and 3 heavy atoms are taken from the covalently linked heavy atoms of each amino acid residue (each atom is one of C, N, O, and S) for arrangement and combination. Each arrangement and combination of 3 heavy atoms is a protein fragment classification. Each protein fragment classification is named according to the amino acid residue name and the arrangement order of the three heavy atoms, so that the arrangement of the 3 heavy atoms in each protein fragment is different, or the type of heavy atoms is different, that is, the heavy atom of each amino acid residue can be arranged at any position in the protein fragment (the purpose of disassembly is to make every atom have the possibility of becoming ATOM_1). The heavy atom is one of C, N, O, and S.
[0055] Proteins in nature are formed by spatial folding of peptide chains formed by dehydration condensation of 20 standard amino acids. By calculating the order of atoms in the 20 standard amino acids, 712 types of protein fragments can ultimately be determined.
[0056] For example Figure 6 The amino acid residue Gly is -N-CA-C=OO-, and 3 heavy atoms are taken out in sequence to combine to obtain 4 amino acid fragments: N-CA-C; CA-C=O; CA-CO; O=CO; and then 3 heavy atoms are taken out in reverse order to combine to obtain 4 amino acid fragments: C-CA-N; O=C-CA; OC-CA; OC=O; and then, considering that the amino acid residues are connected to each other, 6 more protein fragments will be added. There are 348 protein fragments for 20 amino acids without considering the amino acid connection and the case where there is an extra oxygen atom in the carboxyl group.
[0057] Since there is sometimes an extra oxygen atom (OXT) in the terminal carboxyl group, this situation increases 4×20=80 protein fragment types. When the peptide bond between two amino acids is taken into account, there are 6 protein fragment types for each amino acid, namely CN-CA, OCN, NCO, NC-CA, CA-NC, CA-CN, so there are 6×20=120 protein fragment types. There is a disulfide bridge between two adjacent cysteines (Cys), which increases 2 protein fragment types, namely SG-SG-CB, CB-SG-SG. The side chains of asparagine (Asn), lysine (Lys), glutamine (Gln), and alanine (Arg) are terminated by an amide group and react with the α-carboxyl groups of other amino acids, respectively, to produce 4×3=12 protein fragment types, namely NCC, NC-CA, CNC. The pyrrolidine ring of proline (Pro) forms peptide bonds with other amino acids (2+20=22 protein fragments). As for glutamic acid (Glu) and aspartic acid (Asp), the carboxyl groups on their side chains form amide bonds with other amino acids, resulting in 4×2+3×20=68 protein fragments. On the other hand, when the amide bonds formed between the α-carboxyl group of each amino acid and the amino group of the side chain are taken into account, each amino acid has 3 types of protein fragments (3×20=60). Finally, 348+80+120+2+12+22+68+60=712 types of protein fragments were determined. Figure 6The atom names of 20 amino acids are displayed, among which C, CA, N, O are backbone atoms, and CB, CG, CG1, CG2, CB, CD, CD1, CD2, CE, CE1, CE2, CZ, CZ2, CZ3, CH2, SG, SD, ND1, ND2, NE, NE1, NE2, NX, NH1, NH2, OG, OE1, OE2, OH, OD1, OD2 are side chain atoms.
[0058] In the above definitions, the first letter of CA, CB, CG, CG1, CG2, CB, CD, CD1, CD2, CE, CE1, CE2, CZ, CZ2, CZ3, CH2, SG, SD, ND1, ND2, NE, NE1, NE2, NX, NH1, NH2, OG, OE1, OE2, OH, OD1, OD2 indicates the symbol of the heavy atom, and the second letter indicates the position of the amino acid. A corresponds to Figure 6 α in the equation, B corresponds to Figure 6 β, G corresponds to Figure 6 γ in , D corresponds to Figure 6 δ, E in corresponds to Figure 6 ε, Z in Figure 6 ζ in , H corresponds to Figure 6 η in, X corresponds to Figure 6 In ξ, the third letter indicates the number of heavy atoms.
[0059] S2-3 calculates the atomic distance between the first atom in the order of the three heavy atoms of each protein fragment and each ligand atom (small molecule ligand atom). The protein fragment and the ligand atom (small molecule ligand atom) are considered to be an interaction pair. The interatomic distance is calculated by the pairwise_dist function in PyMOL: enter the ligand atom identifier and set the maximum distance to (parameter max_dist = 5), the pairwise_dist function will output The distances of all protein atoms within the range and their relative ligand atoms.
[0060] S2-4: selecting ligand-protein fragment complexes with the same protein fragment type from the interacting ligand-protein fragment complexes and dividing them into multiple primary groups, further grouping them according to the SYBYL atom type, selecting ligand-protein fragment complexes with the same small molecule ligand atom (ligand atom) type from the primary groups and dividing them into multiple secondary groups, and counting the number of interactions between protein fragments and small molecule atoms in each secondary group as M;
[0061] Identification of interaction patterns between S3 protein fragments and small molecule ligand atoms
[0062] S3-1 Set the ligand atoms in the interacting ligand-protein fragment complex to LIGAND (e.g. Figure 2 The three covalently linked heavy atoms in the classified protein fragments are set as ATOM_1, ATOM_2 and ATOM_3 in sequence (e.g. Figure 2 The three nodes in the ligand-protein fragment complex are connected by lines to transform the three-dimensional coordinates of the interacting ligand-protein fragment complex so that the protein fragments of each ligand-protein fragment complex can be placed in the same three-dimensional space coordinate system (called the protein fragment standard coordinate system):
[0063] S3-1a translates the coordinates of the four atoms until the coordinates of ATOM_2 are at the origin, (0,0,0);
[0064] S3-1b rotates the four atoms around the x-axis until the coordinates of ATOM_1 are in the xz plane and z>0;
[0065] S3-1c The four atoms rotate around the y-axis until the coordinate of ATOM_1 is on the z-axis (z>0);
[0066] S3-1d The 4 atoms rotate around the z-axis until the coordinates of ATOM_3 are in the xz plane.
[0067] The final coordinates of the 4 atoms are: ATOM_1: (0,0,Zatom_1), ATOM_2: (0,0,0), ATOM_3: (Xatom_3,0,Zatom_3), and LIGAND: (Xligand,Yligand,Zligand).
[0068] S3-2 selects ligand-protein fragment complexes of the same protein fragment type from the ligand-protein fragment complexes with interactions, and superimposes them in the same three-dimensional space coordinate system. At this time, many small molecule ligand atoms (ligand atoms) are distributed around the superimposed protein fragments.
[0069] S3-3 is further grouped according to the SYBYL atom type. The ligand-protein fragment complexes with the same type of small molecule ligand atoms (ligand atoms) are selected from the ligand-protein fragment complexes with the same type of protein fragments obtained in the previous step, and are superimposed in the same three-dimensional space coordinate system so that the same type of small molecule ligand atoms (ligand atoms) are distributed around the same type of protein fragments, forming the following: Figure 2 When the number of small molecule ligand atoms (ligand atoms) in any superposition model is less than 100, the superposition model is deleted.
[0070] S3-4 Bayesian Gaussian mixture model is an advanced clustering analysis algorithm that assumes that the data distribution is composed of a combination of several Gaussian distributions. The Bayesian Gaussian mixture model (BGMM) is used to cluster the small molecule ligand atoms (ligand atoms) distributed around the protein fragments in each superposition model retained in the previous step, and the classes with less than 20 ligand atoms in the clustering results are deleted. The remaining classes are defined as the interaction patterns between protein fragments and small molecule ligand atoms (ligand atoms) (such as Figure 2 The high-density distribution area of the ligand atomic coordinates is shown).
[0071] The Bayesian Gaussian mixture model (BGMM) uses the BGMM in the Python open source third-party library scikit-learn (sklearn). The BayesianGaussianMixture in scikit-learn is used to cluster ligands with the same SYBYL atom type around protein fragments with the same type with 15 components. The specific code is as follows:
[0072]
[0073] That is, define a function named clustering with the following parameters: input data x, clustering method method, number of clusters n (default n = 15). This function converts the input data x into an array form x1, and when the clustering method is specified as "BGMM", calls the Bayesian Gaussian mixture model in sklean to cluster x1, and outputs the category label array of each sample after clustering.
[0074] The source code of sklearn is open source and available. The code of Bayesian-Gaussian mixture model can be found at: https: / / github.com / scikit-learn / scikit-learn / blob / main / sklearn / mixture / _bayesian_mixture.py
[0075] “https: / / blog.csdn.net / mago2015 / article / details / 82749732” provides an explanation of the Bayesian Gaussian mixture model.
[0076] Since the data after clustering of small molecule ligand atoms (ligand atoms) in the three-dimensional space x, y, z axis directions (superimposed model coordinate system) all present Gaussian distribution, according to the definition of Gaussian distribution and its mathematical characteristics, a 95% confidence ellipsoid is used to represent the spatial distribution of the interaction pattern. For each interaction distribution, there is:
[0077]
[0078] Among them, μ x is the average x-axis coordinate of all ligand atoms not deleted in the interaction, σ x is the standard deviation of the x-axis coordinates of all ligand atoms not deleted in the interaction, μ y , μ z , σ y , σ z The definition of is similar. D(x,y,z) is the distance between the small molecule ligand atom and the 95% confidence ellipsoid. When the ligand atom is inside the 95% confidence ellipsoid, D(x,y,z)<1; when the ligand atom is outside the 95% confidence ellipsoid, D(x,y,z)>1; when the ligand atom is located on the surface of the 95% confidence ellipsoid, D(x,y,z)=1. Therefore, when D(x,y,z)≤1, the ligand atom is considered to have a significant interaction with the protein fragment.
[0079] When finding the target of an unknown small molecule, calculate the interaction between the protein fragment in the binding conformation and each atom of the unknown small molecule. If N pairs of interactions conform to the known interaction pattern, then the hit rate of the interaction pattern is P = N / M. When N ≥ 600 and P ≥ 0.8, it can be considered that the protein is highly likely to be the target of the small molecule atom.
[0080] 2. Use of target prediction methods:
[0081] T1 inputs the protein to be analyzed-small molecule ligand
[0082] Input the simulated binding conformation data of the protein fragment and the unknown small molecule ligand (the simulated binding conformation data is generated by docking the protein fragment and the unknown small molecule ligand using molecular simulation docking software), that is, the file data containing the three-dimensional structure of the protein-small molecule ligand. The file data format is .pdb (called PDB file). This file format is the standard file format for representing the spatial structure of proteins and small molecule ligands in the field of bioinformatics. It contains the coordinate (x, y, z) information of each atom in the three-dimensional space coordinate system. For the description of the PDB file format, see Figure 3 .
[0083] The .pdb format is just a file format that can store the three-dimensional structure of chemical molecules. Specifically, the standard pdb file conforms to the format described at http: / / blog.sciencenet.cn / blog-548663-895916.html.
[0084] T2 Target Analysis
[0085] For each input protein and unknown small molecule ligand file data, calculate the atomic distance between each atom of the protein fragment and the unknown small molecule ligand according to its PDB file according to step S2-3, and calculate the atomic distance between each atom of the protein fragment and the unknown small molecule ligand according to step S2-4. The number of interactions between the protein fragment and the unknown small molecule ligand atoms in the secondary group is counted in all interaction pairs. The ligand-protein fragment complex in the M interaction pairs is transformed into three-dimensional coordinates according to the method of step S3-1. According to the numerical value in the spatial distribution corresponding to the M interactions: μ x , σ x , μ y , μ z , σ y , σ z The transformed coordinates are calculated according to formula 1 D(x, y, z), and the total number of ligand-protein fragment complexes satisfying D(x, y, z) ≤ 1 in each secondary group is counted as N, that is, there are N pairs of interactions that conform to the known interaction pattern, and the hit rate of the interaction pattern is P = N / M (the same below), and the calculation results are output.
[0086] T3 analysis results
[0087] The analysis results are in Excel form (such as Figure 4 The protein conformation analysis results are shown in the table), including the name of the small molecule ligand, the name of the protein, the number of interaction pattern hits (N), and the interaction pattern hit rate (P). When N ≥ 600 and P ≥ 0.8 in each secondary group, the protein can be considered to be the target of the unknown small molecule ligand. Otherwise, the protein is considered not to be the target of the unknown small molecule ligand.
[0088] Effect verification:
[0089] The traditional Chinese medicine Astragaloside has a significant anti-cancer effect. The main components are Astragaloside IV, Formononetin, and Calycosin-7-glucoside. First, the molecular docking technology (existing technology) was used to predict the targets of the three main components of Astragaloside. It was found that the three components had 6872, 7526, and 7826 targets (binding energy ≤ 3.0 kJ / mol), respectively. It is speculated that the false positive rate of the molecular simulation docking result is at least 99.85%.
[0090] The interaction patterns between the three main components of Astragalus and 70 targets were further calculated. Under the conditions of N≥600 and P≥0.8, it was found that the three main components of Astragalus had 3, 4 and 1 targets, respectively, as shown in Table 1. By searching relevant academic papers (see Table 1), it was found that 2, 4 and 0 of these targets were related to anti-cancer effects, respectively, indicating that the target prediction method designed by the present invention effectively reduces the false positive rate of molecular simulation docking results.
[0091] Strychnine is a commonly used Chinese medicine in clinical practice. It has analgesic effects and can be used to treat a variety of common tumors such as liver cancer, lung cancer, breast cancer, and colorectal cancer. Brucine, Strychnine, and icajine are the main components of Strychnine. First, molecular docking technology (existing technology) was used to predict the targets of the three main components of Strychnine, and 8073, 7062, and 7425 targets (binding energy ≤-3.0 kJ / mol) were found for the three components, respectively. According to common sense in this field, the number of targets of a single drug small molecule is generally less than 10. It can be inferred that the false positive rate of the molecular simulation docking result is at least:
[0092] 1-(10 / 7062)*100%=98.86%.
[0093] Table 1 Prediction results of main component targets of Astragalus and supporting literature
[0094]
[0095] according to Figure 5 The process, i.e., steps T1 to T3 in Example 1, was used to further calculate the interaction patterns between the three main components of Nux vomica and 200 targets. The results are shown in Table 1. The number of interactions between protein fragments and small molecule atoms is M pairs, of which N pairs of interactions conform to the known interaction patterns, and the hit rate of the interaction pattern is P = N / M ( Figure 4 ). Under the conditions of N ≥ 600, P ≥ 0.8, as shown in Table 1, it was found that the three main components of Strychnos nux vomica had 23, 19 and 9 targets, respectively. By searching relevant academic papers, it was found that 8, 4 and 6 of these targets were related to tumor treatment and analgesia, respectively ( Figure 4 ), indicating that the target prediction method designed by the present invention effectively reduces the false positive rate of molecular simulation docking results.
[0096] Table 2 Prediction results of targets of main components of Nux Vomica and supporting literature
[0097]
[0098]
Claims
1. A method for predicting unknown small molecule ligand targets, characterized in that: The following steps are involved: Step 1. Classify the ligand-protein fragment complexes that have interactions and the same ligand atom type and the same protein fragment type into one category, transform the three-dimensional coordinates of this type of ligand-protein fragment complexes so that each protein fragment in this category can be placed in the same three-dimensional space coordinate system, and then superimpose this type of ligand-protein fragment complex based on the protein fragment structure to construct a superposition model of this category, and use the Bayesian Gaussian mixture model to cluster the ligand atoms distributed around the protein fragments in each type of superposition model, delete the class with less than 20 ligand atoms in the clustering results, and use 95% confidence ellipsoids to represent the spatial distribution of the ligand atoms that have not been deleted in each type of superposition model: Store the spatial distribution μ x , μ y , μ z , σ x , σ y , σ z , where μ x , μ y , μ z Refers to the average value of the x, y, and z axis coordinates of all ligand atoms that have not been deleted in various superposition models, σ x , σ y , σ z Refers to the standard deviation of the x, y, and z axis coordinates of all ligand atoms that have not been deleted in various superposition models. D(x, y, z) is the distance between the ligand atom and the 95% confidence ellipsoid, and x, y, z are the coordinates of the ligand atom: LIGAND: (Xligand, Yligand, Zligand); Step 2. Input the simulated binding conformation data of the protein fragment and the unknown small molecule ligand, i.e., the file data containing the three-dimensional structure of the protein-unknown small molecule ligand. The format of the file data is .pdb, which is a standard file format for the spatial structure of the protein and the unknown small molecule ligand, and contains the coordinate information of each atom in the three-dimensional space coordinate system. The simulated binding conformation data is generated by docking the protein fragment and the unknown small molecule ligand using molecular simulation docking software; Step 3. Process the binding conformation data in step 2: use SYBYL atom type to mark each atom of the unknown small molecule ligand, split the protein into protein fragments, each protein fragment is an amino acid residue, and the amino acid residue is the remaining part of the amino acids connected by peptide bonds after the 20 standard amino acids in nature are combined with each other and dehydrated. Take out 3 heavy atoms from the covalently connected heavy atoms of each amino acid residue and arrange and combine them. Each arrangement and combination of 3 heavy atoms is a protein fragment classification. Each protein fragment classification is named according to the name of the amino acid residue and the arrangement order of the three heavy atoms, so that the arrangement of the 3 heavy atoms in each protein fragment is different, or the type of heavy atoms is different, and each of the 3 heavy atoms is one of C, N, O, and S; Step 4. Calculate the atomic distance between the first atom in the sequence of the three heavy atoms of the protein fragment and each atom of the unknown small molecule ligand. The protein fragment and the unknown small molecule ligand atom are called interaction pairs. The number of ligand-protein fragment complexes in each group with the same ligand atom type and the same protein fragment type is counted in all interaction pairs and recorded as M. The ligand-protein fragment complexes in the M interaction pairs of each group are transformed into three-dimensional coordinates so that each protein fragment of the same type can be placed in the same three-dimensional space coordinate system. According to the atomic types of the protein fragment and the unknown small molecule ligand in the M interaction pairs of each group, the spatial distribution corresponding to the type in step 1 is searched, and the value is extracted: μ x , σ x , μ y , μ z , σ y , σ z , based on the coordinates of the converted unknown small molecule ligand atoms of each group, the distance D(x, y, z) between the unknown small molecule ligand atoms of each group and the 95% confidence ellipsoid obtained in step 1 is calculated according to formula 1, and the total number of ligand-protein fragment complexes satisfying D(x, y, z) ≤ 1 in each group is counted as N, and the hit rate of the interaction pattern of each group is calculated: P = N / M; Step 5. If N≥600 and P≥0.8 in each group in step 4, the protein fragments in the group are targets of the unknown small molecule ligand.
2. The method for predicting unknown small molecule ligand targets according to claim 1, characterized in that: The covalently linked heavy atoms of the amino acid residues include: heavy atoms in the peptide bond between two amino acids that are bound to each other; oxygen atoms in the terminal carboxyl group of an amino acid; heavy atoms in the side chain of a disulfide bridge between two adjacent cysteines; heavy atoms remaining after the amide group of the side chain of asparagine, lysine, glutamine or alanine reacts with the α-carboxyl group of other amino acids; heavy atoms in the peptide bond formed by the pyrrolidine ring of proline and other amino acids; heavy atoms in the amide bond formed by the carboxyl group on the side chain of glutamic acid or aspartic acid and other amino acids, and heavy atoms in the amide bond formed between the α-carboxyl group of each amino acid and the side chain amino group.
3. The method for predicting unknown small molecule ligand targets according to claim 1, characterized in that: Steps before step 1 include: S1. Sample data set construction: PDB files from the PDB protein structure database from January 1995 to the present were screened using the following criteria: (1) proteins from humans; (2) (3) Crystal structures without antibodies; (4) Crystal structures without mutations; (5) Crystal structures with at least one ligand having a heavy atom number greater than 5 and a molecular weight of ≥80Da and <800Da; (6) Crystal structures with at least one polypeptide having an amino acid residue number greater than 30; (7) Crystal structures with at least one ligand molecule having a relative accessible surface area of <0.6 between the complex state and the isolated state; the resulting PDB files are constructed into a ligand-protein fragment complex dataset; The Reduce program in pdb-tools was used to add hydrogen to the ligand-protein fragment complex selected by conditions (1) to (6), and then dr-sasa was used to calculate the accessible surface area.
4. The method for predicting unknown small molecule ligand targets according to claim 3, characterized in that: The ligand-protein fragment complex having interaction and the same ligand atom type and the same protein fragment type refers to the ligand atom having the same SYBYL atom type label, the three heavy atoms covalently connected in the protein fragment being the same and in the same order, and the first atom in the arrangement order of the three heavy atoms in the protein fragment being the same as the ligand atom. The interatomic distances are calculated using the pairwise_dist function in PyMOL: the ligand atom identities are entered and the maximum distance is set to The pairwise_dist function will output The distances of all protein atoms within the range and their relative ligand atoms.
5. The method for predicting unknown small molecule ligand targets according to claim 4, characterized in that: The transformation of the three-dimensional coordinates of the ligand-protein fragment complex includes: The ligand atom in the interacting ligand-protein fragment complex is set as LIGAND, and the three covalently connected heavy atoms in the classified protein fragments are set as ATOM_1, ATOM_2, and ATOM_3: Then translate the coordinates of the four atoms until the coordinates of ATOM_2 are at the origin of the coordinate system (0,0,0); Rotate the four atoms around the x-axis until the coordinates of ATOM_1 are in the xz plane and z>0; Then rotate the four atoms around the y-axis until the coordinate of ATOM_1 is on the z-axis (z>0); Then rotate the four atoms around the z-axis until the coordinates of ATOM_3 are in the xz plane; The final coordinates of the 4 atoms are: ATOM_1: (0,0,Zatom_1), ATOM_2: (0,0,0), ATOM_3: (Xatom_3,0,Zatom_3), and LIGAND: (Xligand,Yligand,Zligand).
6. The method for unknown small molecule ligand target according to claim 4, characterized in that: If the number of ligand atoms in the superposition model described in step 1 is less than 100, the superposition model is deleted.
7. Use of the method for predicting unknown small molecule ligand targets as described in any one of claims 1 to 6 in predicting the targets of the main components of traditional Chinese medicine.
8. Use of the method for predicting unknown small molecule ligand targets as described in any one of claims 1 to 6 in predicting the targets of the main components of Nux vomica.
9. Use of the method for predicting unknown small molecule ligand targets as described in any one of claims 1 to 6 in predicting the targets of the main components of Astragalus.
