Protein ligand binding affinity prediction method and system based on multi-scale topology and storage medium

Abstract

The invention provides a protein ligand binding affinity prediction method and system based on multi-scale topology and a storage medium, and the method comprises the steps: obtaining original information of a protein-ligand compound, and extracting specific element information; carrying out protein atom screening; performing permutation and combination on elements in the screened specific protein atom set and the screened specific ligand atom set, and generating a plurality of protein-ligand element combination pairing modes; constructing a distance matrix corresponding to each element combination pairing mode, and performing multi-scale filtering; performing spectral analysis on all the sub-topological structures to generate multi-scale topological features; the multi-scale topological features are input into a preset neural network model for prediction, and a combined affinity prediction result is obtained; according to the method, the prediction accuracy, robustness and generalization ability under the background of multiple molecular structures are remarkably improved, and the problem that a traditional model is insufficient in adaptability when facing different molecular types and scale changes is effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence and drug design, and more particularly, to a protein ligand binding affinity prediction method and system based on multi-scale topology and a storage medium. BACKGROUND

[0002] With the development of artificial intelligence technology, computer-aided drug design has become an important tool in new drug research and development. Protein-ligand binding affinity prediction, as one of the key problems in this field, is directly related to the screening efficiency and clinical conversion potential of drug molecules. Although traditional experimental methods are accurate, they have problems such as high cost, long cycle, and low throughput, so efficient and accurate computational methods are increasingly concerned.

[0003] The current mainstream affinity prediction methods mainly include traditional machine learning and deep learning strategies based on feature extraction, and representative algorithms include machine learning methods based on Ricci curvature and deep convolution methods based on topology. Machine learning methods based on Ricci curvature mainly include two kinds based on Forman Ricci curvature and based on Ollivier Ricci curvature. They use different Ricci curvatures to characterize the local structural features of protein-ligand complexes at different scales, and design a combination model of protein-ligand complex features constructed by different Ricci curvatures and gradient boosting decision trees to predict the binding affinity of protein-ligand complexes. Deep convolution methods based on topology project biomolecules into multi-channel image-like representations suitable for convolutional neural networks. Specifically, the three-dimensional structure information of protein-ligand complexes is represented as one-dimensional topological invariants as the features of the complex, and the features are combined with deep convolutional neural networks to predict the binding affinity of the complex.

[0004] However, the existing technology generally has the following defects: 1) Limited feature extraction capability: existing methods rely heavily on prior-defined molecular descriptors or local structural information, which is difficult to fully reflect the geometric and topological features of molecules, especially in complex protein-ligand interactions, where local information is insufficient to characterize global structural properties; 2) Ignoring molecular topological properties: most methods lack modeling of the topological structure evolution of protein-ligand complexes at different scales, making it difficult to capture key spatial configurations and high-order structural relationships; 3) Insufficient model generalization: traditional models have limited ability to learn multi-scale features, making it difficult to adapt to feature changes at different scales or different molecular types, affecting the stability and robustness of protein-ligand binding affinity prediction. SUMMARY

[0005] The application provides a protein-ligand binding affinity prediction method and system based on multi-scale topology and a storage medium, which can more accurately model and characterize the protein-ligand interaction mode.

[0006] To solve the above technical problems, the technical scheme of the application is as follows: A protein-ligand binding affinity prediction method based on multi-scale topology comprises the following steps: S1: Obtain the original information of a protein-ligand complex and extract specific element information to obtain a protein specific atom set and a ligand specific atom set; S2: Take the ligand specific atom set as a reference, filter the protein atoms in the protein specific atom set with a distance not exceeding a preset distance threshold, and obtain a filtered protein specific atom set; S3: Arrange and combine the elements in the filtered protein specific atom set and the ligand specific atom set, respectively, and generate several protein-ligand element combination pairing modes; S4: Construct a distance matrix for each element combination pairing mode, and uniformly set multiple filter thresholds in a preset filter interval to generate several sub-topological structures represented by adjacency matrices; S5: Calculate the graph Laplacian matrix of all sub-topological structures and extract the corresponding statistical information to obtain multi-scale topological features; S6: Input the multi-scale topological features into a preset neural network model for prediction to obtain the binding affinity prediction result of the protein-ligand complex.

[0007] Preferably, in step S1, the original information of the protein-ligand complex at least includes atomic elements of the protein and the ligand, and three-dimensional coordinate information corresponding to each atom. The protein specific atom set includes atoms of four specific elements of carbon, nitrogen, oxygen and sulfur; and the ligand specific atom set includes atoms of nine specific elements of carbon, nitrogen, oxygen, sulfur, phosphorus, fluorine, chlorine, bromine and iodine.

[0008] Preferably, in step S2, the distance is specifically the Euclidean distance.

[0009] Preferably, in step S3, the element combination of the protein includes:

[0010] Wherein, represents the element combination set of the protein. The element combination of the ligand includes:

[0011] wherein, represents the element combination set of the ligand; According to the element combination set of the protein and the element combination set of the ligand , a protein-ligand element combination pairing mode is generated

[0012] Preferably, in the step S4, for each of the element combination pairing mode, a corresponding distance matrix is constructed , denoted as:

[0013] wherein, represents the element in the i-th row and j-th column of the distance matrix ; represents the Euclidean distance between the i-th and j-th atoms; and respectively represent the three-dimensional coordinate information of the i-th and j-th atoms; Preferably, in the step S3, a plurality of filter thresholds are uniformly set in the preset filter interval, denoted as: wherein, is the k-th filter threshold, is a positive integer; The adjacency matrix of the sub-topology generated by the k-th filter threshold is denoted as:

[0014] wherein, is the element in the i-th row and j-th column of the adjacency matrix .

[0015] Preferably, in the step S5, the graph Laplacian matrix is calculated for all sub-topologies, and the eigenvalues of each graph Laplacian matrix are calculated; the eigenvalues include zero eigenvalues and non-zero eigenvalues . Further, the corresponding statistical information is extracted, denoted as wherein, represents the harmonic spectrum information, i.e. the number of zero eigenvalues ; ~ represents the non-harmonic spectrum information; The ~​​​​ Specifically, , , , ,

[0016] wherein sum, mean, std, min and max represent summation function, mean function, standard deviation function, minimum value function and maximum value function, respectively; The statistical quantity information corresponding to all sub-topological structures is spliced to obtain continuous statistics, and a flattening operation is further performed to obtain the multi-scale topological features.

[0017] Preferably, in the step S6, the preset neural network model is specifically a trained gradient boosting decision tree model.

[0018] Preferably, the trained gradient boosting decision tree model is obtained according to the following steps: obtain a training data set; construct an original gradient boosting decision tree model; input the training data set into the original gradient boosting decision tree model, and perform iterative training through forward stepwise addition modeling and gradient boosting strategy, and obtain the trained gradient boosting decision tree model after training.

[0019] The application also provides a protein ligand binding affinity prediction system based on multi-scale topology, which applies the above method and comprises: a specific element extraction module: used for obtaining original information of a protein-ligand complex and extracting specific element information to obtain a protein specific atom set and a ligand specific atom set; a protein atom screening module: used for screening protein atoms in the protein specific atom set that are within a preset distance threshold based on the ligand specific atom set to obtain a screened protein specific atom set; an element combination pairing module: used for respectively arranging and combining elements in the screened protein specific atom set and the ligand specific atom set, and generating several protein-ligand element combination pairing modes; a multi-scale filtering module: used for constructing a distance matrix corresponding to each of the element combination pairing modes, and uniformly setting a plurality of filtering thresholds in a preset filtering interval to generate several sub-topological structures represented by adjacency matrices; a spectrum analysis module: used for calculating graph Laplacian matrices of all sub-topological structures and extracting corresponding statistical quantity information to obtain multi-scale topological features; ​The prediction module is used for inputting the multi-scale topological features into a preset neural network model for prediction to obtain a binding affinity prediction result of the protein-ligand complex.

[0020] The application further provides a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps in the method.

[0021] Compared with the prior art, the application has the beneficial effects that: The application provides a protein-ligand binding affinity prediction method and system based on multi-scale topology and a storage medium. First, original information of a protein-ligand complex is obtained, and specific element information is extracted to obtain a protein specific atom set and a ligand specific atom set. Then, the ligand specific atom set is taken as a reference to screen protein atoms in the protein specific atom set that are not more than a preset distance threshold, and a screened protein specific atom set is obtained. Next, elements in the screened protein specific atom set and the ligand specific atom set are arranged and combined respectively, and a plurality of protein-ligand element combination pairing modes are generated. Then, a distance matrix corresponding to each element combination pairing mode is constructed, a plurality of filter thresholds are uniformly set in a preset filter interval, and a plurality of sub-topological structures represented by adjacency matrices are generated. Next, graph Laplacian matrices of all the sub-topological structures are calculated, and corresponding statistical information is extracted to obtain multi-scale topological features. Finally, the multi-scale topological features are input into a preset neural network model for prediction to obtain a binding affinity prediction result of the protein-ligand complex. The application has the following beneficial effects: 1) Stronger structure evolution modeling capability: the application combines the persistent homology method and the spectral theory, generates a series of nested subgraphs (sub-topological structures) through a multi-scale filtering process, analyzes the Laplacian spectral features of each subgraph, describes the topological evolution law of the protein-ligand complex in the spatial structure change process, realizes quantitative modeling of key topological invariant features, significantly enhances the capture capability of geometric structures and topological properties in the scale change of the protein-ligand complex, enables the model to sensitively identify the topological patterns that stably exist in the structure evolution process, and improves the modeling accuracy of complex geometric configuration changes. 2) Feature expression is more comprehensive: the present application constructs the topological substructure sequence of protein-ligand complex under multi-scale filtering value based on persistent homology theory, and conducts eigenvalue analysis on each substructure by combining with spectrum theory, so as to comprehensively and systematically extract persistent statistics; the present application not only finely depicts the local interaction characteristics at atomic level, but also fuses the topological structure evolution information of the whole complex; compared with the traditional method relying on static molecular descriptors, the present application realizes expansion from 'local static characteristics' to'multi-scale geometric-topological fusion characteristics' in structure representation, and greatly improves the integrity and expressiveness of the characteristics; 3) The model generalization and prediction performance are better: the multi-scale geometric-topological features extracted by the present application have good structure stability and expression consistency, can effectively retain the key structure information of protein-ligand complex under different spatial scales and element combinations, and improve the learning ability of the model to complex structure and functional relationship by combining high-dimensional protein-ligand complex multi-scale topological features with a neural network model; In general, the present application significantly improves the prediction accuracy, robustness and generalization ability in various molecular structure backgrounds, and effectively overcomes the insufficient adaptability problem of the traditional model when facing different molecular types and scale changes. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A flow chart of a protein-ligand binding affinity prediction method based on multi-scale topology provided in embodiment 1.

[0023] Figure 2 A whole framework diagram of a protein-ligand binding affinity prediction method based on multi-scale topology provided in embodiment 2.

[0024] Figure 3 A structure diagram of a protein-ligand binding affinity prediction system based on multi-scale topology provided in embodiment 3. DETAILED DESCRIPTION

[0025] The drawings are only used for illustrative description, and cannot be understood as limitation to the present application; In order to better illustrate the present embodiment, some components in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0026] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0027] Embodiment 1 As Figure 1As shown, the embodiment provides a protein-ligand binding affinity prediction method based on multi-scale topology, comprising the following steps: S1: Obtain the original information of the protein-ligand complex, and extract the specific element information to obtain the protein specific atom set and the ligand specific atom set; S2: Take the ligand specific atom set as the reference, and screen the protein atoms in the protein specific atom set with a distance not exceeding a preset distance threshold to obtain the screened protein specific atom set; S3: Arrange and combine the elements in the screened protein specific atom set and the ligand specific atom set, respectively, and generate several protein-ligand element combination pairing modes; S4: Construct the distance matrix corresponding to each element combination pairing mode, and uniformly set multiple filtering thresholds within a preset filtering interval to generate several sub-topological structures represented by adjacency matrices; S5: Calculate the graph Laplacian matrix for all sub-topological structures, and extract the corresponding statistical information to obtain the multi-scale topological features; S6: Input the multi-scale topological features into a preset neural network model for prediction to obtain the binding affinity prediction result of the protein-ligand complex.

[0028] In the specific implementation process, first, the original information of the protein-ligand complex is obtained, and the specific element information is extracted to obtain the protein specific atom set and the ligand specific atom set; Then, taking the ligand specific atom set as the reference, the protein atoms in the protein specific atom set with a distance not exceeding a preset distance threshold are screened to obtain the screened protein specific atom set; Secondly, arrange and combine the elements in the screened protein specific atom set and the ligand specific atom set, respectively, and generate several protein-ligand element combination pairing modes; Then, construct the distance matrix corresponding to each element combination pairing mode, and uniformly set multiple filtering thresholds within a preset filtering interval to generate several sub-topological structures represented by adjacency matrices; Then, calculate the graph Laplacian matrix for all sub-topological structures, and extract the corresponding statistical information to obtain the multi-scale topological features; Finally, input the multi-scale topological features into a preset neural network model for prediction to obtain the binding affinity prediction result of the protein-ligand complex; The method realizes multi-scale, high-dimensional and structure-sensitive molecular characterization by constructing the topological substructure of the protein-ligand complex at different filtering scales and extracting the geometric and topological features, and then effectively establishes the mapping relationship between the features and the binding affinity by combining a machine learning model, thereby improving the accuracy and stability of the prediction and assisting efficient and accurate drug screening and molecular design.

[0029] Embodiment 2 The embodiment provides a protein-ligand binding affinity prediction method based on multi-scale topology, comprising the following steps: S1: obtaining original information of a protein-ligand complex and extracting specific element information to obtain a protein specific atom set and a ligand specific atom set; S2: taking the ligand specific atom set as a reference, screening protein atoms in the protein specific atom set with a distance not exceeding a preset distance threshold to obtain a screened protein specific atom set; S3: respectively arranging and combining elements in the screened protein specific atom set and the ligand specific atom set, and generating several protein-ligand element combination pairing modes; S4: constructing a distance matrix corresponding to each element combination pairing mode, and uniformly setting multiple filtering thresholds in a preset filtering interval to generate several sub-topological structures represented by adjacency matrices; S5: calculating graph Laplacian matrices for all sub-topological structures and extracting corresponding statistical information to obtain multi-scale topological features; S6: inputting the multi-scale topological features into a preset neural network model for prediction to obtain a binding affinity prediction result of the protein-ligand complex; In the step S1, the original information of the protein-ligand complex at least includes atomic elements of the protein and the ligand, and three-dimensional coordinate information corresponding to each atom; The protein specific atom set includes atoms of four specific elements of carbon, nitrogen, oxygen and sulfur; and the ligand specific atom set includes atoms of nine specific elements of carbon, nitrogen, oxygen, sulfur, phosphorus, fluorine, chlorine, bromine and iodine; In the step S2, the distance is specifically a Euclidean distance; In the step S3, the element combination of the protein includes:

[0030] Among them, represents an element combination set of the protein; The element combination of the ligand includes:

[0031] Among them, Represents the set of elements that combine ligands; Based on the set of elements of the protein The set of element combinations of the ligands ,generate Protein-ligand element pairing methods; In step S4, for each of the element combination pairing methods, a corresponding distance matrix is ​​constructed. , represented as:

[0032] in, Distance matrix The element in the i-th row and j-th column; This represents the Euclidean distance between the i-th and j-th atoms; and These represent the three-dimensional coordinate information of the i-th and j-th atoms, respectively; Uniformly set within the preset filtration range Each filtering threshold is represented as: ,in, For the k-th filtering threshold, It is a positive integer; Pass the k-th filtering threshold The adjacency matrix of the generated sub-topology Represented as:

[0033] in, Adjacency matrix The element in the i-th row and j-th column; In step S5, the graph Laplacian matrix is ​​calculated for all sub-topologies, and the eigenvalues ​​of each graph Laplacian matrix are calculated; the eigenvalues ​​include the zero eigenvalue. Non-zero eigenvalues ; Further extract the relevant statistical information, denoted as ,in, Represents harmonic spectrum information, i.e., zero eigenvalues Quantity; ~ Represents dissonant spectral information; The aforementioned ~ Specifically: , , , ,

[0034] wherein sum, mean, std, min and max represent summation function, mean function, standard deviation function, minimum function and maximum function, respectively; all the statistical information corresponding to the sub-topological structures are spliced to obtain continuous statistics, and a flattening operation is further performed to obtain the multi-scale topological features; In the step S6, the preset neural network model is specifically a trained gradient boosting decision tree model; The trained gradient boosting decision tree model is obtained according to the following steps: obtain a training data set; construct an original gradient boosting decision tree model; input the training data set into the original gradient boosting decision tree model, and perform iterative training through forward stepwise addition modeling and gradient boosting strategy, and obtain the trained gradient boosting decision tree model after training is completed.

[0035] In the specific implementation process, as shown in Figure 2 , first, the atomic elements and their three-dimensional coordinate information of the original protein-ligand complex are input into the element extractor to extract representative specific elements and coordinate information; specifically, for protein macromolecules, four specific elements of carbon, nitrogen, oxygen and sulfur are extracted; for ligand small molecules, nine specific elements of carbon, nitrogen, oxygen, sulfur, phosphorus, fluorine, chlorine, bromine and iodine are extracted; the protein specific atom set and the ligand specific atom set are respectively denoted as and ; Then, the extracted protein-ligand specific elements and their coordinate information are input into the protein atom filter to filter the protein atoms within a distance of 20Å from the ligand; specifically, for any given two sets and , the distance of element to set is defined as , wherein represents the Euclidean distance; in this embodiment, the given hyperparameter 20Å represents the preset distance threshold, and the filter will find the atoms in the protein specific atom set whose distance to the ligand specific atom set does not exceed the threshold 20Å, which is mathematically represented as , , that is, the final protein specific atom set obtained through the filter; After that, the obtained ligand atom set and the protein atom set obtained after filtering The element combination division is respectively performed to generate the atom sets corresponding to the protein element combination and the ligand element combination. Specifically, the element combination of the protein includes:

[0036] wherein, represents the element combination set of the protein; The element combination of the ligand includes:

[0037] wherein, represents the element combination set of the ligand; According to the element combination set of the protein and the element combination set of the ligand , the following is generated element combination pairing modes of the protein-ligand; For each element combination pairing mode, a distance matrix corresponding thereto is constructed, which is represented as:

[0038] wherein, represents the element in the ith row and the jth column of the distance matrix ; and represents the Euclidean distance between the ith atom and the jth atom. and represent the three-dimensional coordinate information of the ith atom and the jth atom, respectively. Finally, 143 distance matrices are generated. According to the distance matrix , the sub-complex structures generated by each element combination pairing mode are further filtered. Specifically, the filtering interval is set as [2 Å, 12 Å], and 50 filtering thresholds are uniformly selected in the filtering interval to generate 50 topological structures at different scales, which are represented as: wherein, is the kth filtering threshold. The adjacency matrix of the sub-topology structure generated by the kth filtering threshold is represented as:

[0039] wherein, is the element in the ith row and the jth column of the adjacency matrix ; and that is, when the distance between the atoms of the protein and the ligand does not exceed the filtering threshold 1 if the protein-ligand complex is filtered out, otherwise 0; after multi-scale filtering, a total of sub-topologies and their corresponding adjacency matrices are generated; spectrum analysis is then performed; based on the adjacency matrices of the series of sub-structures obtained in the above steps, a series of graph Laplacian matrices of each protein-ligand complex is calculated, denoted as the eigenvalues of each graph Laplacian matrix are calculated, and the eigenvalues of each Laplacian matrix are denoted as the harmonic spectrum and the inharmonic spectrum information thereof are respectively counted and integrated into a 6-dimensional feature vector , referred to as the statistics of the Laplacian matrix, wherein represents the harmonic spectrum information, ~ represents the inharmonic spectrum information; Specifically, represents the harmonic spectrum information, i.e., the number of zero eigenvalues , among the eigenvalues of each Laplacian matrix, if there are u zero eigenvalues , then , which is used to represent the topological invariant of the sub-structure; the inharmonic spectrum corresponds to the non-zero eigenvalues , and the sum, mean, standard deviation, extreme value, etc. of the inharmonic spectrum are counted to describe the geometric and topological characteristics of the sub-structure, i.e.,

[0040]

[0041]

[0042]

[0043]

[0044] wherein sum, mean, std, min and max represent the sum function, the mean function, the standard deviation function, the minimum value function and the maximum value function, respectively; the statistics information corresponding to all sub-topologies is spliced to obtain the continuous statistics with a dimension of (6, 50, 143), and further flattened to obtain a one-dimensional feature vector with a length of 42900, i.e., the multi-scale topological feature; Finally, the multi-scale topological feature is input into a preset neural network model for prediction to obtain the binding affinity prediction result of the protein-ligand complex; in this embodiment, the preset neural network model is specifically a trained gradient boosting decision tree model;​ Specifically, the trained gradient boosting decision tree model is obtained according to the following steps: Obtain a training data set, including PDBbind-v2007, PDBbind-v2013 and PDBbind-v2016 data sets; Construct an original gradient boosting decision tree model; Input the training data set into the original gradient boosting decision tree model, and perform iterative training through forward stepwise addition modeling and gradient boosting strategy, and obtain the trained gradient boosting decision tree model after training is completed; In this embodiment, the gradient boosting decision tree adopts different parameter settings for different data sets, and the specific parameter settings are shown in Table 1; Table 1 Parameters of gradient boosting decision tree model corresponding to different data sets

[0045] After training is completed, the prediction effect of the method and other methods (FPRC, Topologynet and OnionNet) is compared and tested on the three data sets, and the brief introduction of the three other methods is as follows: 1) FPRC (Forman persistent Ricci curvature, Forman persistent Ricci curvature): The Forman Ricci curvature is used to describe the local geometric structure features of the protein-ligand complex at different scales, and combined with the machine learning model to predict the binding affinity of the protein-ligand complex; 2) Topologynet (Topology based deep convolutional and multi-task neural networks, topology based deep convolutional and multi-task neural networks): The three-dimensional structure information of the protein-ligand complex is represented as one-dimensional topological invariant as the feature of the complex, and the feature is combined with the deep convolutional neural network to predict the binding affinity of the complex; 3) OnionNet (a multiple-layer intermolecular-contact-based convolutional neural network, a multiple-layer intermolecular-contact-based convolutional neural network): The local and non-local feature information generated by the non-rotation element pair between the ligand and protein atoms based on specific contact is combined with the deep convolutional neural network to predict the binding affinity of the protein-ligand complex; The prediction effects of the four methods on different data sets are shown in Table 2, where R represents the correlation coefficient of the predicted affinity value and the true affinity value, and RMSE represents the root mean square error of the predicted affinity value and the true affinity value. The larger the corresponding number in the table R is, the better the implementation effect of the corresponding method is, and the smaller the corresponding number in the table RMSE is, the better the implementation effect of the corresponding method is; Table 2 Comparison of prediction effects of different models on three data sets

[0046] As can be seen from Table 2, the prediction performance of the present method on the three data sets is leading, thereby verifying the effectiveness of the multi-scale topological algorithm proposed in the present embodiment, which can make the features of the protein-ligand complex more comprehensive and more robust; Unlike the method of extracting features only for the protein and the ligand respectively, the present method models the atomic level interaction relationship between the protein and the ligand from the perspective of the whole protein-ligand complex. Specifically, the present method selects four representative elements in the protein and nine common elements in the ligand, performs spatial screening on the protein atoms by setting a distance threshold to obtain a set of atoms that may participate in binding, and constructs 143 pairing modes between 11 types of protein element combinations and 13 types of ligand element combinations. On this basis, the filtering mechanism of the persistent homology theory is used to perform multi-scale filtering operation on each group of graph structures, and a nested substructure sequence of the protein-ligand complex at multiple spatial scales is constructed. Further, a distance matrix between atoms is established and converted into a graph structure represented by an adjacency matrix. Then, combined with the spectral analysis method, the eigenvalues of the graph Laplacian matrix at each scale are calculated, and the harmonic and non-harmonic spectral statistics including the number of zero eigenvalues (topological invariant), the maximum eigenvalue, the minimum eigenvalue, the mean, the standard deviation and the sum are extracted to constitute the multi-scale topological features of the protein-ligand complex. The present method not only retains the geometric structure information of the molecule at the local and global levels, but also systematically captures the topological evolution characteristics of the protein-ligand complex at different spatial scales, significantly enhances the information expression ability of the features in terms of spatial configuration, structural evolution and high-order relationship, and thus more comprehensively reflects the binding mechanism; In addition, the method extracts multi-scale geometric topological features of protein-ligand complexes at different spatial scales and atomic combinations as input, combines a gradient boosting decision tree model, and constructs a protein-ligand binding affinity prediction framework; the high-dimensional topological features of the protein-ligand complex are trained and predicted by introducing the gradient boosting decision tree model, which effectively mines the deep mapping relationship between topological evolution and geometric structure information and binding affinity by taking advantage of its nonlinear regression modeling and feature selection; experiments show that the effective fusion of multi-scale topological features of protein-ligand complexes and the gradient boosting decision tree model significantly improves the accuracy, robustness and generalization ability of the model under different structural types and scale conditions, and realizes the collaborative optimization of high-dimensional feature expression, topological sensitivity and prediction effect.

[0047] Embodiment 3 As shown in Figure 3 The embodiment provides a protein ligand binding affinity prediction system based on multi-scale topology, which applies the method described in embodiment 1 or 2, comprising: A specific element extraction module 301 is configured to obtain original information of a protein-ligand complex and extract specific element information to obtain a protein specific atom set and a ligand specific atom set. A protein atom screening module 302 is configured to screen protein atoms in the protein specific atom set that are within a preset distance threshold based on the ligand specific atom set to obtain a screened protein specific atom set. An element combination pairing module 303 is configured to arrange and combine elements in the screened protein specific atom set and the ligand specific atom set, respectively, and generate several protein-ligand element combination pairing modes. A multi-scale filtering module 304 is configured to construct a distance matrix corresponding to each of the element combination pairing modes and uniformly set multiple filtering thresholds within a preset filtering interval to generate several sub-topological structures represented by adjacency matrices. A spectrum analysis module 305 is configured to calculate graph Laplacian matrices for all sub-topological structures and extract corresponding statistical information to obtain multi-scale topological features. A prediction module 306 is configured to input the multi-scale topological features into a preset neural network model for prediction to obtain a binding affinity prediction result of the protein-ligand complex.

[0048] In the specific implementation process, first, the specific element extraction module 301 obtains the original information of the protein-ligand complex and extracts the specific element information to obtain the protein specific atom set and the ligand specific atom set. Then, the protein atom screening module 302 screens the protein specific atom set based on the ligand specific atom set, and obtains a screened protein specific atom set, wherein the protein atom in the protein specific atom set is screened if the distance between the protein atom and the ligand atom is not more than a preset distance threshold value; Secondly, the element combination pairing module 303 respectively arranges and combines the elements in the screened protein specific atom set and the ligand specific atom set, and generates several protein-ligand element combination pairing modes; After that, the multi-scale filtering module 304 constructs a distance matrix corresponding to each element combination pairing mode, and uniformly sets a plurality of filtering thresholds in a preset filtering interval, and generates several sub-topological structures represented by adjacency matrices; After that, the spectrum analysis module 305 calculates the graph Laplacian matrix of all sub-topological structures, and extracts the corresponding statistical information, and obtains the multi-scale topological features; Finally, the prediction module 306 inputs the multi-scale topological features into a preset neural network model for prediction, and obtains the binding affinity prediction result of the protein-ligand complex; The system realizes multi-scale, high-dimensional and structure-sensitive molecular representation by constructing the topological sub-structure of the protein-ligand complex under different filtering scales and extracting its geometric and topological features, and then effectively establishes the mapping relationship between the features and the binding affinity by combining the machine learning model, thereby improving the accuracy and stability of the prediction, and assisting efficient and accurate drug screening and molecular design.

[0049] The same or similar reference signs correspond to the same or similar components; The terms describing the positional relationship in the drawings are only used for illustrative description, and cannot be understood as a limitation to the present application; Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the implementation modes of the present application. For those skilled in the art, on the basis of the above description, other different forms of changes or modifications can also be made. Here, all the implementation modes are not required or can not be exhausted. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the claims of the present application.

Claims

1. A method for predicting protein ligand binding affinity based on multiscale topology, characterized in that, The method comprises the following steps: S1: obtaining original information of a protein-ligand complex and extracting specific element information to obtain a protein specific atom set and a ligand specific atom set; S2: taking the ligand specific atom set as a reference, screening protein atoms in the protein specific atom set that are not more than a preset distance threshold, and obtaining a screened protein specific atom set; S3: respectively arranging and combining elements in the screened protein specific atom set and the ligand specific atom set, and generating a plurality of protein-ligand element combination pairing modes; S4: constructing a distance matrix corresponding to each element combination pairing mode, uniformly setting a plurality of filtering thresholds in a preset filtering interval, and generating a plurality of sub-topological structures represented by adjacency matrices; S5: calculating graph Laplacian matrices of all sub-topological structures and extracting corresponding statistical information to obtain multi-scale topological features; S6: inputting the multi-scale topological features into a preset neural network model for prediction to obtain a binding affinity prediction result of the protein-ligand complex.

2. The method of claim 1, wherein the method is based on a multi-scale topological protein ligand binding affinity prediction. In the step S1, the original information of the protein-ligand complex at least includes atomic elements of the protein and the ligand, and three-dimensional coordinate information corresponding to each atom; The protein specific atom set includes atoms of four specific elements of carbon, nitrogen, oxygen and sulfur; and the ligand specific atom set includes atoms of nine specific elements of carbon, nitrogen, oxygen, sulfur, phosphorus, fluorine, chlorine, bromine and iodine.

3. The method of claim 1, wherein the method is based on multiscale topology. In the step S2, the distance is specifically a Euclidean distance.

4. The method of claim 2, wherein the method is based on a multi-scale topological protein ligand binding affinity prediction. In the step S3, the element combination of the protein includes: wherein, represents a set of elemental combinations of proteins; The element combination of the ligand includes: wherein, represents a set of elemental combinations of ligands; According to the elemental combination set of the protein and the elemental combination set of the ligand , generate a protein-ligand elemental combination pairing mode.

5. The method of claim 4, wherein the method is based on a multi-scale topological protein ligand binding affinity prediction method. In the step S4, for each of the element combination pairing modes, a corresponding distance matrix is constructed is expressed as: wherein, denotes a distance matrix the element in the i-th row and j-th column of denotes the Euclidean distance between the i-th and j-th atoms; and denotes the three-dimensional coordinate information of the i-th and j-th atoms, respectively; Uniformly set a filtering threshold value in the preset filtering interval, denoted as: wherein, wherein, is the kth filtering threshold value, is a positive integer; by the kth filtering threshold adjacency matrix of the generated sub-topology is represented as: wherein is an adjacency matrix the element in the ith row and jth column of 6. The method of claim 5, wherein the method is based on a multi-scale topological protein ligand binding affinity prediction method. The step S5 calculates the graph Laplacian matrix of all sub-topologies, and calculates the eigenvalue of each graph Laplacian matrix; the eigenvalue includes zero eigenvalue and non-zero eigenvalue ​ Further, the corresponding statistical information is extracted, denoted as wherein, represents the harmonic spectral information, i.e. the number of zero eigenvalues ; ~ represents the non-harmonic spectral information; The described Specifically:​ , , , , Wherein, sum, mean, std, min and max represent summation function, mean function, standard deviation function, minimum value function and maximum value function, respectively; statistical information of all sub-topologies perform splicing to obtain continuous statistics, and further perform flattening operation to obtain the multi-scale topological features.

7. The method of predicting protein-ligand binding affinity based on multiscale topology according to any one of claims 1-6, characterized in that, In the step S6, the preset neural network model is specifically a trained gradient boosting decision tree model.

8. The method of claim 7, wherein the method is based on a multi-scale topological protein ligand binding affinity prediction method. The trained gradient boosting decision tree model is obtained according to the following steps: Obtain a training data set; Construct an original gradient boosting decision tree model; Input the training data set into the original gradient boosting decision tree model, and perform iterative training through forward stepwise addition modeling and gradient boosting strategy, and obtain the trained gradient boosting decision tree model after training.

9. A multiscale topology-based protein ligand binding affinity prediction system, applying the method of any one of claims 1 to 8, characterized in that, It comprises: A specific element extraction module for obtaining original information of a protein-ligand complex and extracting specific element information to obtain a protein specific atom set and a ligand specific atom set; A protein atom screening module for taking the ligand specific atom set as a reference, screening protein atoms in the protein specific atom set that are not more than a preset distance threshold, and obtaining a screened protein specific atom set; An element combination pairing module for respectively arranging and combining elements in the screened protein specific atom set and the ligand specific atom set, and generating a plurality of protein-ligand element combination pairing modes; A multi-scale filtering module is configured to construct a distance matrix corresponding to each of the element combination pairs, and set multiple filtering thresholds uniformly within a preset filtering interval to generate several sub-topological structures represented by adjacency matrices; A spectrum analysis module is configured to calculate a graph Laplacian matrix for all the sub-topological structures, and extract corresponding statistical information to obtain multi-scale topological features; A prediction module is configured to input the multi-scale topological features into a preset neural network model for prediction to obtain a binding affinity prediction result of the protein-ligand complex.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by a processor to implement the steps in the method of any one of claims 1-8.

Images