Method for predicting drug binding sites based on interface geometry and deep learning model
Patent Information
- Application Number
- CN202611213420.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-11
- Publication Date
- 2026-11-03
AI Technical Summary
现有药物结合位点预测方法在靶标蛋白分子界面的刻画不够准确，无法准确体现疏水效应、溶剂排布和静电势与非静电势的耦合效应，导致预测准确性低。
A method based on interface geometry and deep learning models is adopted. The problem of minimizing solubility free energy is solved by interface variational model to obtain the multi-solubility interface features and curvature of proteins. The multi-scale isovariant graph neural network EGNN is used for prediction, and physical contributions such as electrostatics, van der Waals and interfacial tension are explicitly considered to construct an accurate interface point cloud structure.
It improves the accuracy of drug binding site prediction, eliminates errors caused by interface inaccuracies, and bases the prediction results on real physics, thus enhancing the interpretability of the physical mechanism and the continuity of the prediction results.
Smart Images
Figure CN123033562A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer-aided drug design technology, specifically relating to a method for predicting drug binding sites based on interface geometry and deep learning models. Background Technology
[0002] Drug binding site prediction is a computational biology task that uses computer programs to predict the binding region between drug molecules and proteins based on the sequence or structural information of target proteins. It has important guiding significance for virtual screening of drug molecules, lead compound optimization, target discovery, and antibody / peptide design.
[0003] Existing methods for predicting drug binding sites are mainly divided into two categories: computer-aided drug design (CADD) methods and artificial intelligence-driven drug design (AIDD) methods. However, existing methods generally suffer from a common fundamental problem: the characterization of target protein molecular interfaces is not accurate enough. Specifically, both CADD and AIDD methods largely construct protein interfaces based on simple geometric structures using atomic coordinates and empirical radii. These interfaces are directly determined empirically by the three-dimensional coordinates of biomolecules, and are essentially still atomic-scale geometric envelopes, rather than biomolecular dissolution interfaces that reflect the true physicochemical equilibrium state.
[0004] The above phenomena prevent this interface from reflecting the solvent arrangement and fluctuation phenomena caused by the hydrophobic effect in the local binding pocket. The hydrophobic effect is one of the important forces driving the binding of drug molecules to target proteins. On the other hand, it ignores the influence of the coupling effect between electrostatic potential and non-electrostatic potential at the interface on the interface position, resulting in a non-negligible deviation between the constructed interface and the real biomolecule-solvent interface, which in turn makes the accuracy of drug binding site prediction very low. Summary of the Invention
[0005] To address the aforementioned problems, this invention provides a method for predicting drug binding sites based on interface geometry and deep learning models.
[0006] To achieve the above objectives, this invention provides a drug binding site prediction method based on interface geometry and a deep learning model, comprising: Obtain the three-dimensional structural data of the target protein, as well as the atomic parameter properties of each atom in the target protein.
[0007] Based on the aforementioned three-dimensional structural data and atomic parameter properties, a variational model is used to solve the problem of minimizing the solvation free energy of the interface between the target protein and the solvent containing the target protein. This yields a first level set function for the target protein at the dehumidified interface and a second level set function for the protein at the wetted interface. Discrete interface point clouds are extracted from the protein at the wetted interface based on the second level set function. The coordinates of each point in the interface point cloud are substituted into the first level set function to obtain the multi-solution interface features corresponding to each point. Based on the second level set function, a differential discretization method is used to calculate the multi-scale curvature of each interface point in the interface point cloud. The multi-scale curvature is combined with the multi-solution interface features to obtain the geometric features of each interface point. Based on the atomic parameter properties and geometric features, the physicochemical characteristics of each interface point are determined. Using each interface point as a node, the graph structure of the interface point cloud is obtained.
[0008] Based on atomic parameter properties and geometric features, the physicochemical characteristics of each interface point are determined; with each interface point as a node, the graph structure of the interface point cloud is obtained; the graph structure and physicochemical characteristics are input into a pre-trained multi-scale isovariant graph neural network (EGNN) to obtain the prediction results of whether each interface point of the target protein is a drug binding site.
[0009] Preferably, the multi-scale equivariant graph neural network (EGNN) is an improved multi-scale equivariant graph neural network (EGNN). The improved EGNN replaces the input features based on a single atom type in the traditional EGNN model with a dual-path encoding structure that separates geometric and chemical features; it replaces the single-path serial stacked structure of the traditional EGNN model with a multi-branch parallel EGNN architecture with different numbers of layers and independent parameters, and the outputs of each branch are spliced and fused; the chemical features include atom type features, electrostatic potential, and van der Waals potential.
[0010] Preferably, the atomic parameter attributes include the position, type, charge, and van der Waals parameters of each atom; the multi-scale curvature includes mean curvature and Gaussian curvature; the calculation of the multi-scale curvature of each interface point in the interface point cloud using the differential discretization method specifically includes: firstly, smoothing the second level set function with different resolutions using Gaussian filtering to obtain the second level set function representation at multiple scales; then, calculating the mean curvature and Gaussian curvature at each scale using the differential discretization method to obtain the multi-scale curvature of each interface point.
[0011] Preferably, the multi-solubility interface features are used to characterize the coupling effect of electrostatic and non-electrostatic effects at the protein interface and the degree of local solvent fluctuation; substituting the coordinates of each point in the interface point cloud into the first level set function to obtain the multi-solubility interface features corresponding to each point specifically includes: obtaining the coordinates of each interface point, substituting the coordinates of each interface point into the first level set function, and calculating the function value as the multi-solubility interface feature of the interface point.
[0012] Preferably, the step of determining the physicochemical characteristics of each interface point based on atomic parameter attributes and geometric features, and obtaining the graph structure of the interface point cloud with each interface point as a node, specifically includes: calculating the atomic type characteristics, electrostatic potential, and van der Waals potential of each interface point based on atomic parameter attributes; concatenating the atomic type characteristics, electrostatic potential, van der Waals potential, and geometric features to obtain the physicochemical characteristics of each interface point; correcting the geodesic distance between each node based on the multi-scale curvature; and using the KNN algorithm to determine the neighboring nodes of each node and connecting each neighboring node based on the geodesic distance to obtain the graph structure of the interface point cloud.
[0013] Preferably, correcting the geodesic distance between nodes based on the multi-scale curvature specifically includes: correcting the Euclidean distance between nodes based on the multi-scale curvature to obtain the geodesic distance. The expression is as follows: in, Representative node and The Euclidean distance between them, H i and H j They are nodes and The average curvature, and They are nodes and Gaussian curvature.
[0014] Preferably, the step of calculating the atomic type feature of each interface point based on the atomic parameter attributes specifically includes: for each interface point, searching for K atoms in its vicinity, constructing Gaussian weights based on the distance between the atoms and the interface point and normalizing them, and then weighting and summing the atom type information through the normalized Gaussian weights to obtain the atomic type feature of the interface point.
[0015] Preferably, the improved multi-scale equivariant graph neural network (EGNN) is trained using a combined loss function. The training process specifically includes: comparing the predicted interface points output by the model with the corresponding real labels, and calculating the binary cross-entropy loss function and the smoothing loss function respectively; combining the two loss functions with a preset regularization coefficient and weighting them to obtain the total loss; with the goal of minimizing the total loss, calculating the gradient of each parameter of the network through the backpropagation algorithm, and updating the network parameters using the gradient descent strategy; iteratively executing the forward prediction, loss calculation, backpropagation and parameter update steps until the model converges.
[0016] This invention also provides a drug binding site prediction system based on interface geometry and deep learning models, comprising: The data acquisition module is used to acquire the three-dimensional structural data of the target protein, as well as the atomic parameter properties of each atom in the target protein.
[0017] The feature extraction module is used to solve the solvation free energy minimization problem of the interface between the target protein and the solvent using an interface variational model based on the three-dimensional structural data and atomic parameter properties, to obtain a first level set function of the target protein at the dehumidified interface and a second level set function of the protein at the wetted interface; to extract discrete interface point clouds of the protein at the wetted interface based on the second level set function; to substitute the coordinates of each point in the interface point cloud into the first level set function to obtain the multi-solution interface features corresponding to each point; to calculate the multi-scale curvature of each interface point in the interface point cloud using a differential discretization method based on the second level set function; to merge the multi-scale curvature with the multi-solution interface features to obtain the geometric features of each interface point; and to determine the physicochemical features of each interface point based on the atomic parameter properties and geometric features, and to obtain the graph structure of the interface point cloud with each interface point as a node.
[0018] The prediction module is used to determine the physicochemical characteristics of each interface point based on atomic parameter properties and geometric features; to obtain the graph structure of the interface point cloud with each interface point as a node; and to input the graph structure and physicochemical characteristics into a pre-trained multi-scale isovariant graph neural network (EGNN) to obtain the prediction result of whether each interface point of the target protein is a drug binding site.
[0019] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, is capable of executing any of the steps in the drug binding site prediction method based on interface geometry and deep learning models.
[0020] The drug binding site prediction method based on interface geometry and deep learning models provided by this invention has the following beneficial effects: This invention transforms the interface determination problem from a geometric problem into a physical problem. Through an interface variational model, it simultaneously incorporates the solubilization physical contributions of electrostatics, van der Waals forces, hydrophobicity, and interfacial tension into the free energy functional. The variational method is used to solve for the interface position that minimizes the total free energy. This position is the result of the equilibrium of all physical forces, and therefore represents a true thermodynamic equilibrium interface. This invention uses this physically accurate interface as the common foundation for all subsequent operations, such as point cloud extraction, curvature calculation, physicochemical feature assignment, and graph structure and prediction based on it. This eliminates the error propagation caused by inaccurate interfaces at the source, ensuring that the prediction results are based on a correct physical foundation. Attached Figure Description
[0021] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0022] Figure 1 This is a flowchart illustrating a drug binding site prediction method based on interface geometry and a deep learning model, according to an embodiment of the present invention. Figure 2 This is a module architecture diagram of a drug binding site prediction system based on interface geometry and deep learning models according to an embodiment of the present invention. Detailed Implementation
[0023] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.
[0024] The improved multi-scale equivariant graph neural network (EGNN) of this invention has the following structural improvements over the basic EGNN: it replaces the single-atom type input features with a dual-path encoding structure that separates geometric and chemical features; it replaces the single-path serial stacked structure with a multi-branch parallel architecture with different numbers of layers and independent parameters; it introduces geodesic distance as the edge weight of the graph structure, so that geometric constraints continuously participate in the calculation of message weights in each layer; and it replaces the single loss function with a classification loss function and a smoothing combined loss function.
[0025] Based on this, the present invention provides a drug binding site prediction method based on interface geometry and deep learning models, which is based on multi-scale interface geometry and isovariant graph neural networks. For example... Figure 1 As shown, it includes the following steps:
[0026] S1, Data Acquisition and Interface Processing The three-dimensional structure and sequence information of the protein were obtained from the RCSB PDB database; the protein structure was repaired and completed using pdbfixer; the force field parameters were matched one by one with the atoms in the protein using pdb2pqr, and the charge of each atom was calculated; then the above information was integrated into the same file, and the dewetting interface and the wetted interface and their corresponding level set function representations were calculated using the interface variational model, and the van der Waals interaction potential (van der Waals potential) and electrostatic potential near the interface were calculated.
[0027] The dehumidified interface (dry interface) and the wetted interface (wet interface) refer to the protein interfaces at different energy minima calculated using different initial interfaces based on an interface variational model. The initial interface required to calculate the dehumidified interface is a cuboid box that can encompass all protein atoms, and the shortest distance from the protein atoms to the surface of the cuboid box is not less than an adjustable distance parameter. The initial interface required to calculate the wetted interface is a relatively compact closure of the protein atoms. The wetted interface refers to the actual interface where the protein is in direct contact with water molecules in a real solution environment; the dehumidified interface refers to the theoretical interface calculated based on the interface variational model, used to characterize the interface position of the protein in a hydrophobic state (i.e., when water molecules are completely displaced).
[0028] The process of solving for the interface position using the interface variational model is as follows: The interface Γ of the protein-solvent system is represented as the zero-level set of a level set function φ, i.e. The solvation free energy functional F[Γ] of the system consists of interfacial energy terms, van der Waals interaction energy terms, and electrostatic energy terms:
[0029] + (x)dx+ [Γ]; Where γ is the interfacial tension coefficient. For the interface area, Solvent density, For the solvent region, Let x be the van der Waals interaction potential between the solute atom and the solvent molecule at spatial position x (described by the Lennard-Jones potential). [Γ] is the electrostatic energy term, which includes the coupling of electrostatic, polarization, and hydrophobic effects. It is obtained by solving the Poisson-Boltzmann equation and reflects the long-range interaction between solute charge distribution and solvent polarization effect.
[0030] Interface position is determined by the free energy functional. The variational solution is performed on the interface Γ, i.e., the solution is obtained. This yields the Euler-Lagrange equations that satisfy the free energy minimization condition. Their physical meaning is that at the equilibrium interface, interfacial tension, van der Waals force, and electrostatic force are in equilibrium. After discretizing these equations, the level set function φ is iteratively updated using the level set method until the free energy functional converges to its minimum value. The zero level set at this point represents the desired interface location.
[0031] Since the aforementioned free energy functional is a non-convex functional, it has multiple local minima. Using a cuboid encompassing all protein atoms as the initial interface for iterative solution, the free energy converges to the local minimum corresponding to the dehumidified state, yielding the first level set function. Using a compact closure of the protein atoms as the initial interface for iterative solution, the free energy converges to the local minimum corresponding to the wetted state, yielding the second level set function.
[0032] This invention employs mesh discretization and difference schemes to transform the solution of partial differential equations into a standard algebraic iteration. The entire computation process can be executed automatically by a program without manual intervention. The implicit representation of the level set method naturally supports interface topological changes (such as cavity merging or splitting), eliminating the need to handle complex mesh reconnection problems as with explicit interface methods. The global variation of the free energy functional with respect to the interface position ensures the physical self-consistency of the solution process, so that the final interface position simultaneously satisfies the equilibrium conditions of all physical forces.
[0033] S2, Multi-scale curvature calculation and point cloud extraction Three-dimensional level set function based on protein infiltration interface (The second level set function) is first represented at different resolutions using Gaussian filtering, serving as the basis for multi-scale curvature calculation. The formula for Gaussian filtering is as follows:
[0034] ; in, For resolution parameters, Representative at The second level set function after Gaussian filtering at the specified resolution.
[0035] Subsequently, based on the second level set function, a high-precision differential discretization method is used to calculate the mean curvature and Gaussian curvature: ; in, for Different resolution parameters Representative at Average curvature calculated at the specified resolution. Representative at The Gaussian curvature is calculated at the specified resolution, where adj represents the adjoint matrix. Subsequently, based on the second level set function of the wetting interface, its iso-level set surface is extracted as the interface point cloud.
[0036] in, Here, A is the interface thickness parameter, and A is a given horizontal set function value (usually taken as 0).
[0037] For each interface point in the point cloud Extracting multi-scale curvature (hereinafter referred to as ), van der Waals interaction potential (hereinafter referred to as ) and electrostatic potential (hereinafter referred to as ), which serves as the physicochemical information of the interface point.
[0038] Based on the relative distance between the interface point and the protein atom, the one-hot encoding of the atom type is weighted by distance and assigned to the interface point. Specifically, for each interface point... First, search for the K nearest neighbor set in the set of atoms. Definition point With atoms The Euclidean distance between them is:
[0039] Construct Gaussian weights based on distance: in, Here are the parameters for the Gaussian weights, used to control how quickly the Gaussian weights decay as the Euclidean distance increases. Normalization is then performed.
[0040] in, This is used to prevent oddities caused by an excessively small denominator.
[0041] Finally, the one-hot type information of the atoms is obtained by weighted summation. Mapped onto a point cloud, this serves as the atomic type feature of the point cloud (hereinafter referred to as...). ): S201, Calculation of information on multiple dissolved interfaces For each interface point Calculate the three-dimensional level set function corresponding to the dehumidified interface. The value (first level set function) reflects the local multi-dissolution interface characteristics at that point (hereinafter referred to as...). Hydrophobicity (HH) is used to characterize the hydrophobicity of a specific location on a protein surface, that is, how easily water molecules are "repelled" at that location. Its expression is as follows:
[0042] This value reflects the coupling of electrostatic and non-electrostatic effects, characterizes the fluctuation of local solvent near the interface, and contains geometric and chemical interaction information of the biomolecular interface.
[0043] S3. Feature extraction based on residual connection networks The extracted physicochemical information is divided into geometric features and chemical features, and then input into two feature extraction networks with the same structure but different parameters. The geometric features are as follows:
[0044] Here, Concat represents the vector concatenation function. It incorporates multi-scale curvature and multi-solution interface features; its chemical characteristics are:
[0045] It includes van der Waals interaction potential, electrostatic potential, atomic one-hot type mapping features, and multi-solution interface features.
[0046] The feature extraction network consists of a normalization layer, a linear layer, and a residual connection network. Let the input features be denoted as... The input features first pass through a standardization layer, then through a linear layer to increase dimensionality, and finally into a three-layer residual connection network. Each layer of the residual connection network consists of a "linear layer-activation layer-random deactivation layer-linear layer" architecture, with residual connections between layers. Specifically, the formula is as follows: first, the input features... Standardization process:
[0047] in, The mean, The standard deviation is then used. The features are then projected into a high-dimensional space using a linear mapping:
[0048] ; in, and Here are the network parameters for the linear layer. Based on this, a multi-layer residual structure is used for feature extraction. For the ... Layers, including:
[0049] in, , , , For the first lThe parameters of the linear layer in a residual network, For the LeakyReLU activation function: Finally, the output features are obtained through linear mapping: in, For output features and For the network parameters of the linear layer, This represents the total number of layers in the multi-layer residual structure.
[0050] S301, Construction of KNN graph based on geodesic distance Based on the extracted features, this embodiment of the invention further uses KNN to construct a graph structure of the protein interface point cloud. In the distance calculation of KNN, Euclidean distance is replaced by curvature-based distance. Approximate geodetic distance:
[0051] in, Representative interface points and The Euclidean distance between them, H i and H j They are nodes and The average curvature, and They are nodes and The Gaussian curvature. Then, according to the KNN calculation process, k nearest neighbors are selected for each point and connected to them to form a graph structure.
[0052] Protein surfaces are highly curved two-dimensional manifolds. Existing techniques use Euclidean distance to measure the spatial relationships of interface point clouds, essentially simplifying the distance problem on a curved surface into a straight-line distance problem in flat space, without considering the impact of surface curvature on the shortest path on the manifold. This leads to two typical geometric distortions: first, misclassifying points that are actually far apart along the curved surface on both sides of a pocket or groove as neighbors, causing incorrect connections across regions; second, on convex surfaces, incorrectly connecting two points that have taken a long path along the curved surface because they are close in straight line. These distortions are passed to the message passing process of graph neural networks, causing the network to propagate information along incorrect topological paths, resulting in systematic deviations in node feature updates. This invention corrects Euclidean distance to geodesic distance based on interface curvature, ensuring that the selection of node neighborhoods strictly follows the true manifold structure of the protein surface. The graph structure can accurately reflect the actual spatial relationships of interface points along the biomolecular surface, fundamentally eliminating geometric distortions and their cumulative effect on subsequent network propagation.
[0053] Applications of S4, Multi-scale Equivariant Graph Neural Networks After obtaining the graph structure of the protein point cloud, this embodiment of the invention proposes a multi-scale geodesic isotropic graph neural network (improved multi-scale isotropic graph neural network EGNN) based on the receptive field to implement the feature fusion layer of the network.
[0054] An improved multi-scale isotropic graph neural network (EGNN) is constructed by inputting the graph structure and the physicochemical features of each interface point. Geodesic distance serves as the edge weights of the graph structure, defining the message passing paths between nodes. The improved EGNN uses the physicochemical features of each interface point as the initial embeddings of each node, and performs message passing and node feature updates under the constraints of the message passing paths defined by the graph structure. The geodesic distance edge weights continuously participate in the message weight calculation of each layer as geometric constraints during message passing. After multiple rounds of message passing and node feature updates, a point-by-point classification discriminator outputs the predicted probability of each interface point being a drug binding site, thus obtaining the predicted drug binding site of the target protein. (A geodesic isotropic graph neural network module is described.) The single-layer formula can be expressed as:
[0055] ; in, Let i and j be the node features of the i-th and j-th interface points, respectively. The edge features connecting the i-th interface point and the j-th interface point, Let i be the three-dimensional coordinates of the i-th interface point. After constructing the graph using KNN, the first... All adjacent nodes of a given interface point and It is a multilayer perceptron. This represents the neighborhood aggregation module. In this embodiment of the invention, neighborhood aggregation is achieved through direct summation.
[0056] This invention uses Each geodesic isotropic map neural network module has a different number of layers and its parameters are independent. Features are fused from the interface point cloud. After passing through each geodesic isotropic map neural network module, the protein interface point cloud is concatenated and passed through a linear layer to obtain the final protein point cloud embedding. The specific formula is as follows:
[0057] ; ; in, This represents the number of layers in the geodesic isomorphic map neural network. With... As the number of layers increases, the receptive field of the model expands, thereby obtaining multi-scale information of the point cloud through stitching.
[0058] S401, Application of Discriminator and Loss Function For drug binding site prediction tasks based on interface geometry and deep learning models, this method employs a point-by-point binary classification discriminator to embed and decode protein point clouds into point-by-point binary labels. The decoder structure consists of a superposition of multiple "linear layer-activation layer" layers, as shown in the following formula:
[0059] in, The LeakyReLU activation function is used. These are the parameters for the linear layer.
[0060] The loss function of the overall model is: ; in, The coefficient of the regularization term, The cross-entropy loss function is: For smoothing loss function: in, and Representing different nodes, This indicates the first digit of the graph constructed using KNN. All adjacent nodes of a given interface point Defined as: in, This is a positive regularization parameter used to prevent singularities caused by excessively small denominators, and here and Let be the three-dimensional coordinates of the i-th interface point and the j-th interface point, respectively. The parameter for Gaussian weights is used to control the rate at which the Gaussian weights decay as the Euclidean distance increases.
[0061] The advantages of this invention are: 1. Regarding the issue of inaccurate characterization of interface locations. Existing technologies use unreasonable protein interface locations, failing to accurately reflect the local hydrophobic effects, curvature changes, and coupling of electrostatic and non-electrostatic effects at drug binding sites. This invention obtains high-precision biomolecular interfaces through an interface variational model, accurately characterizing protein interface locations and their geometric and chemical features, explicitly considering solvent-protein interactions; simultaneously, based on the extraction characteristics of interfaces in multiple soluble states (dehydrated state, wetted state), it reflects the coupling of electrostatic and non-electrostatic effects and characterizes the degree of local solvent fluctuations near the interface.
[0062] 2. Regarding the lack of interpretability in geometric information computation. Existing technologies require a large amount of geometric data and network parameters for learning when computing geometric information on the interface, lacking interpretability in an intuitive geometric sense. This invention, based on the interface level set function, uses a high-precision difference approximation to directly calculate the average curvature and Gaussian curvature, reducing the dependence on data volume and network complexity, and enhancing the interpretability of the algorithm in terms of physical mechanisms.
[0063] 3. Regarding the issue of inaccurate characterization of interface positional relationships. Existing technologies employ distance calculation methods that have significant errors compared to actual geodesic distances when characterizing positional relationships on interfaces. This invention uses curvature approximation to calculate geodesic distances, making the distance calculation closer to the actual path along the protein surface, thus improving the accuracy of characterizing interface positional relationships.
[0064] 4. Regarding the lack of physical interpretability of chemical features. Existing technologies require a large amount of data and network parameters for learning when extracting chemical information from protein interfaces, failing to directly reflect the solvent-protein interaction and lacking physical interpretability. This invention proposes a multi-scale, multi-modal prediction framework based on protein interface point clouds, combining geometric curvature, chemical features, and atomic type features to explicitly calculate van der Waals potential, electrostatic potential, and multi-solution-state interface features, giving each feature a clear physical meaning.
[0065] 5. Regarding the issue of spatial dispersion in prediction results. Existing technologies predict drug binding sites that are spatially dispersed, which does not conform to the characteristic of continuous distribution of common drug binding sites. This invention introduces a smoothing loss function as a regularization term into the loss function, forcing the prediction results of adjacent interface points to be similar, thus avoiding dispersed predicted binding sites and making the prediction results more consistent with the characteristics of common drug binding sites.
[0066] Furthermore, this invention utilizes multi-scale EGNN to fuse interface information from different receptive fields, thereby improving the feature recognition capability of local interface regions. Experimental results demonstrate an AUC-ROC value of 0.9 on the MaSIF-site dataset, surpassing existing best-in-class models such as dMaSIF. This achieves high-quality protein interface feature embeddings while accurately predicting drug binding sites, providing guidance for protein-protein interaction research, target discovery, and the design of proteins, antibodies, and peptides.
[0067] like Figure 2 As shown, the system includes a data acquisition module, a feature extraction module, a graph construction module, and a prediction module. These modules are connected sequentially and are used to complete data acquisition and preprocessing, interface point cloud feature extraction, graph structure construction, and site prediction, respectively. A detailed demonstration is shown below: Based on the same inventive concept, this invention also provides a drug binding site prediction system based on interface geometry and a deep learning model, comprising: The data acquisition module is used to acquire the three-dimensional structural data of the target protein, as well as the atomic parameter properties of each atom in the target protein.
[0068] The feature extraction module is used to solve the solvation free energy minimization problem of the interface between the target protein and the solvent using an interface variational model based on the three-dimensional structural data and atomic parameter properties, to obtain a first level set function of the target protein at the dehumidified interface and a second level set function of the protein at the wetted interface; to extract discrete interface point clouds of the protein at the wetted interface based on the second level set function; to substitute the coordinates of each point in the interface point cloud into the first level set function to obtain the multi-solution interface features corresponding to each point; to calculate the multi-scale curvature of each interface point in the interface point cloud using a differential discretization method based on the second level set function; to merge the multi-scale curvature with the multi-solution interface features to obtain the geometric features of each interface point; and to determine the physicochemical features of each interface point based on the atomic parameter properties and geometric features, and to obtain the graph structure of the interface point cloud with each interface point as a node.
[0069] The prediction module is used to determine the physicochemical characteristics of each interface point based on atomic parameter properties and geometric features; to obtain the graph structure of the interface point cloud with each interface point as a node; and to input the graph structure and physicochemical characteristics into a pre-trained multi-scale isovariant graph neural network (EGNN) to obtain the prediction result of whether each interface point of the target protein is a drug binding site.
[0070] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the drug binding site prediction method based on interface geometry and deep learning models provided above.
[0071] Specific limitations on the computational system for the drug binding site prediction method based on interface geometry and deep learning models can be found in the limitations outlined above, and will not be repeated here. Each module in the aforementioned drug binding site prediction system based on interface geometry and deep learning models can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, allowing the processor to call and execute the corresponding operations of each module.
[0072] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A method for predicting drug binding sites based on interface geometry and deep learning models, characterized in that, Includes the following steps: Obtain the three-dimensional structural data of the target protein, as well as the atomic parameter properties of each atom in the target protein; Based on the three-dimensional structural data and atomic parameter properties, the solvation free energy minimization problem of the interface between the target protein and the solvent in which the target protein is located is solved by an interface variational model to obtain the first level set function of the target protein at the dehumidified interface and the second level set function of the protein at the wetted interface; based on the second level set function, discrete interface point clouds of the protein at the wetted interface are extracted; the coordinates of each point in the interface point cloud are substituted into the first level set function to obtain the multi-solution interface features corresponding to each point; Based on the second level set function, the multi-scale curvature of each interface point in the interface point cloud is calculated using the differential discretization method; the multi-scale curvature is then combined with the multi-dissolution interface features to obtain the geometric features of each interface point. Based on atomic parameter properties and geometric characteristics, the physicochemical characteristics of each interface point are determined; Using each interface point as a node, the graph structure of the interface point cloud is obtained; the graph structure and physicochemical features are input into a pre-trained multi-scale isovariant graph neural network (EGNN) to obtain the prediction result of whether each interface point of the target protein is a drug binding site.
2. The drug binding site prediction method based on interface geometry and deep learning model according to claim 1, characterized in that, The multi-scale equivariant graph neural network EGNN is an improved multi-scale equivariant graph neural network EGNN. The improved multi-scale equivariant graph neural network EGNN replaces the input features based on a single atom type in the traditional EGNN model with a dual-path encoding structure that separates geometric features and chemical features; it replaces the single-path serial stacked structure of the traditional EGNN model with a multi-branch parallel EGNN architecture with different numbers of layers and independent parameters, and the outputs of each branch are spliced and fused; the chemical features include atom type features, electrostatic potential, and van der Waals potential.
3. The drug binding site prediction method based on interface geometry and deep learning model according to claim 1, characterized in that, The atomic parameter attributes include the position, type, charge, and van der Waals parameters of each atom; the multi-scale curvature includes mean curvature and Gaussian curvature; the calculation of the multi-scale curvature of each interface point in the interface point cloud using the differential discretization method specifically includes: firstly, smoothing the second level set function with different resolutions using Gaussian filtering to obtain the second level set function representation at multiple scales; then, calculating the mean curvature and Gaussian curvature at each scale using the differential discretization method to obtain the multi-scale curvature of each interface point.
4. The drug binding site prediction method based on interface geometry and deep learning model according to claim 1, characterized in that, The multi-solution interface features are used to characterize the coupling effect of electrostatic and non-electrostatic effects at the protein interface and the degree of local solvent fluctuation. Substituting the coordinates of each point in the interface point cloud into the first level set function, the multi-dissolution interface features corresponding to each point are obtained. Specifically, this includes: obtaining the coordinates of each interface point, substituting the coordinates of each interface point into the first level set function, and calculating the function value as the multi-dissolution interface feature of the interface point.
5. The drug binding site prediction method based on interface geometry and deep learning model according to claim 1, characterized in that, The physicochemical characteristics of each interface point are determined based on atomic parameter properties and geometric features; Using each interface point as a node, the graph structure of the interface point cloud is obtained, specifically including: calculating the atomic type features, electrostatic potential, and van der Waals potential of each interface point based on atomic parameter attributes; concatenating the atomic type features, electrostatic potential, van der Waals potential, and geometric features to obtain the physicochemical features of each interface point; correcting the geodesic distance between each node based on the multi-scale curvature; and using the KNN algorithm based on the geodesic distance to determine the neighboring nodes of each node and connect each neighboring node to obtain the graph structure of the interface point cloud.
6. The drug binding site prediction method based on interface geometry and deep learning model according to claim 5, characterized in that, Correcting the geodesic distance between nodes based on the multi-scale curvature specifically includes: correcting the Euclidean distance between nodes based on the multi-scale curvature to obtain the geodesic distance. The expression is as follows: in, Representative node and The Euclidean distance between them, H i and H j They are nodes and The average curvature, and They are nodes and Gaussian curvature.
7. The drug binding site prediction method based on interface geometry and deep learning model according to claim 5, characterized in that, The calculation of the atomic type feature of each interface point based on the atomic parameter attributes specifically includes: for each interface point, searching for the K nearest atoms, constructing Gaussian weights based on the distance between the atoms and the interface point and normalizing them, and then summing the atomic type information by weighting the normalized Gaussian weights to obtain the atomic type feature of the interface point.
8. The drug binding site prediction method based on interface geometry and deep learning model according to claim 2, characterized in that, The improved multi-scale equivariant graph neural network (EGNN) is trained using a combined loss function. The training process specifically includes: comparing the predicted interface points output by the model with the corresponding real labels, and calculating the binary cross-entropy loss function and the smoothing loss function respectively; combining the two loss functions with a preset regularization coefficient and weighting them to obtain the total loss; with the goal of minimizing the total loss, calculating the gradient of each parameter of the network through the backpropagation algorithm, and updating the network parameters using the gradient descent strategy; iteratively executing the forward prediction, loss calculation, backpropagation and parameter update steps until the model converges.
9. A drug binding site prediction system based on interface geometry and deep learning models, characterized in that, include: The data acquisition module is used to acquire the three-dimensional structural data of the target protein, as well as the atomic parameter properties of each atom in the target protein; The feature extraction module is used to solve the solvation free energy minimization problem of the interface between the target protein and the solvent using an interface variational model based on the three-dimensional structural data and atomic parameter properties, to obtain a first level set function of the target protein at the dehumidified interface and a second level set function of the protein at the wetted interface; to extract discrete interface point clouds of the protein at the wetted interface based on the second level set function; and to substitute the coordinates of each point in the interface point cloud into the first level set function to obtain the multi-solution interface features corresponding to each point. Based on the second level set function, the multi-scale curvature of each interface point in the interface point cloud is calculated using the differential discretization method; the multi-scale curvature is then combined with the multi-dissolution interface features to obtain the geometric features of each interface point. The prediction module is used to determine the physicochemical characteristics of each interface point based on atomic parameter properties and geometric features; Using each interface point as a node, the graph structure of the interface point cloud is obtained; the graph structure and physicochemical features are input into a pre-trained multi-scale isovariant graph neural network (EGNN) to obtain the prediction result of whether each interface point of the target protein is a drug binding site.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 8.