A model training, identification method and system of non-covalent interaction

Abstract

This invention discloses a method and system for training and identifying non-covalent interaction models. The identification method includes: acquiring the spatial structure information of the molecular structure to be identified; inputting the spatial structure information of the molecular structure to be identified into an identification model based on a preset electron density; and generating the position coordinates and topological features of non-covalent interactions in the molecular structure to be identified through the identification model. The technical solution provided by this invention solves the problem of inaccurate identification of non-covalent interactions in molecular systems.

Description

Technical Field

[0001] This invention relates to the field of structural biology analysis technology, specifically to a method and system for training and identifying non-covalent interaction models. Background Technology

[0002] Non-covalent interactions (NCIs), unlike covalent bonds, do not involve shared electrons but rather more dispersed electromagnetic interactions between or within molecules. NCIs are crucial for maintaining the three-dimensional structures of macromolecules such as proteins and nucleic acids. Furthermore, NCIs participate in many biological processes in which macromolecules specifically but transiently bind together. These interactions also significantly influence drug design, crystallinity, and materials design, particularly self-assembly, as well as the synthesis of many organic molecules. Therefore, the recognition of NCIs in protein structures has extremely strong applications.

[0003] Existing technologies for NCI (Neural Cross-linked Indicator) identification have summarized several rules for determining whether a structure is an NCI. However, rule-based NCI identification has the limitation of being only qualitative and not quantitative, and its ability to identify NCI in complex environments is poor. Besides NCI structure identification, predicting NCI properties such as intensity is also an effective basis for accurate NCI type determination. Traditional methods based on molecular force fields are empirical and rely on rule refinement. While they can predict a few classic NCIs, such as hydrogen bonds and salt bonds, they lack prediction models for non-classical NCIs and cannot consider the influence of the surrounding environment on NCI intensity. Methods based on quantum chemical calculations of NCI intensity, although highly accurate, are only applicable to small molecular systems. In large molecular systems, their computational complexity and global perspective are insufficient, leading to the failure of these methods. Therefore, how to accurately identify NCIs is an urgent problem to be solved. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method and system for training and identifying non-covalent interaction models, thereby improving the accuracy of identifying non-covalent interactions in macromolecular systems.

[0005] According to the first aspect, a method for training a non-covalent interaction identification model, the method comprising:

[0006] Training samples are established based on the target molecular structure. The training samples include spatial structure information of the target molecular structure and corresponding labels. The labels include position labels and topological feature labels. The position labels are used to mark the spatial positions of non-covalent interactions in the target molecular structure, and the topological feature labels are used to mark the properties of non-covalent interactions in the target molecular structure.

[0007] Using the spatial structure information as input, an initial machine learning model is used to predict non-covalent interactions in the target molecular structure. The initial machine learning model is then corrected based on the error between the prediction results and the label to obtain an identification model. This identification model is used to identify the positions and properties of non-covalent interactions in the target molecular structure.

[0008] Optionally, the initial machine learning model is built using a convolutional neural network, and training samples are established based on the target molecular structure, including:

[0009] The target molecular structure is standardized and its attitude is corrected, and the coordinate information of a preset number of atoms in the corrected target molecular structure is stored in a three-dimensional matrix to obtain the spatial structure information.

[0010] The tags corresponding to the spatial structure information are generated based on the electron density.

[0011] Optionally, the spatial structure information can be mapped from a three-dimensional matrix to a four-dimensional matrix based on different types of chemical characteristics to obtain four-dimensional spatial structure information.

[0012] Optionally, generating the tag corresponding to the spatial structure information based on electron density includes:

[0013] The target molecular structure is latticized to obtain a lattice molecular structure;

[0014] Obtain at least one electron density lattice point information with a preset resolution, wherein the electron density lattice point information is used to store the electron density of each lattice point within the lattice molecular structure;

[0015] The saddle point position is determined based on the lattice position of the electron density saddle point in the electron density lattice information. The saddle point position is used to represent the spatial position of the non-covalent interaction in the target molecular structure.

[0016] Based on the atomic coordinate information, the attribution relationship corresponding to each saddle point position is determined. The attribution relationship is used to represent the matching relationship between non-covalent interactions and nearby atom pairs.

[0017] The saddle point location and its corresponding attribution relationship are used as the location labels of the spatial structure information;

[0018] The electron density topological features at the saddle point are generated using at least the electron density, first-order gradient, and Hessian matrix at the saddle point, and these electron density topological features are used as topological feature labels for the spatial structure information. The electron density topological features are used to describe the properties of non-covalent interactions.

[0019] Optionally, determining the saddle point location based on the grid point location of the electron density saddle point in the electron density grid information includes:

[0020] The electron density grid information is divided into multiple sub-information blocks of a first preset range size;

[0021] Obtain the ligand-receptor atom pairs in the sub-information block, and take the range covered by the midpoint of the connection line of the ligand-receptor atom pairs as the center and the preset length as the radius as the candidate range;

[0022] Calculate the reduced electron density gradient value of each grid point within the candidate range, and select the grid point with the smallest reduced electron density gradient value as the candidate saddle point;

[0023] Calculate the reduced electron density gradient values ​​of the adjacent lattice points of the candidate saddle point, and take the lattice point with the smallest reduced electron density gradient value among the candidate saddle point and the adjacent lattice points as the saddle point position.

[0024] Optionally, after calculating the reduced electron density gradient values ​​of the adjacent lattice points of the candidate saddle point and taking the lattice point with the smallest reduced electron density gradient value among the candidate saddle point and the adjacent lattice points as the saddle point location, the method further includes:

[0025] Discard saddle point locations where the eigenvalues ​​of the Hessian matrix of electron density do not satisfy λ1<λ2<0<λ3;

[0026] When a saddle point location corresponds to multiple ligand-acceptor atom pairs, the atom pair with the closest distance between the ligand and the acceptor is taken as the target atom pair marked at the saddle point location;

[0027] When the distance between any two saddle points is less than a preset distance, the saddle point with the smaller electron density value is discarded.

[0028] According to the second aspect, a method for identifying non-covalent interactions, the method comprising:

[0029] Obtain spatial structural information of the molecular structure to be identified;

[0030] The spatial structure information of the molecular structure to be identified is input into the identification model obtained by the training method described in the first aspect, and the identification model is used to generate the position coordinates, attribution relationships and topological features of the non-covalent interactions in the molecular structure to be identified.

[0031] According to a third aspect, a system for identifying non-covalent interactions, the system comprising:

[0032] The information acquisition module obtains the spatial structure information of the molecular structure to be identified;

[0033] The identification module is used to input the spatial structure information of the molecular structure to be identified into the identification model obtained according to the training method described in the first aspect, and to generate the position coordinates and topological features of non-covalent interactions in the molecular structure to be identified through the identification model.

[0034] According to the fourth aspect, an electronic device includes:

[0035] A memory and a processor are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the methods described in the first aspect and the second aspect, or any optional embodiment of the first aspect.

[0036] According to a fifth aspect, embodiments of the present invention provide a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions for causing the computer to perform the methods described in the first aspect and the second aspect, or any optional embodiment of the first aspect.

[0037] The technical solution of this invention has the following advantages:

[0038] This invention provides a method and system for training and identifying non-covalent interactions (NCIs). The method specifically includes: acquiring the three-dimensional spatial structure of a target molecule and using the position and topological features of NCIs within that structure as labels corresponding to the target molecule structure to create training samples. Then, a convolutional neural network algorithm is used to train a machine learning model capable of accurately identifying NCIs in macromolecule-ligand pairs, improving the accuracy and efficiency of NCI identification. During the creation of training samples, the spatial structure of the macromolecule-ligand pair is divided into multiple lattice points, and the electron density of each lattice point is calculated based on crystallographic theory to obtain more reliable electron density lattice point information. Based on the properties of electron density itself, it is known that the electron density saddle point is almost identical to the position of the non-covalent interaction. Then, the lattice point representing the electron density saddle point is identified through features such as the electron density gradient within each lattice point. The spatial coordinates of this lattice point are used as the spatial coordinates of the NCI, and the calculated topological features of the electron density at this lattice point are used as the topological features of the NCI, thereby accurately labeling a large number of target molecule structures. Furthermore, by employing electron density grid information at multiple resolutions, training samples under various resolution conditions are obtained, thereby achieving NCI analysis of the target molecular structure from both microscopic and macroscopic perspectives, further improving the accuracy and reliability of NCI recognition. In addition, based on the different atomic characteristics of atoms in the target molecular structure, the training samples are mapped from three-dimensional vectors to four-dimensional vectors, further enhancing the detail of the convolutional neural network training, thus further improving the reliability of the recognition results provided by the recognition model in this embodiment of the invention. Attached Figure Description

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

[0040] Figure 1 This is a schematic diagram illustrating the training steps of a non-covalent interaction recognition model according to an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the sample dimensionality-upgrading mapping structure for a training method of a non-covalent interaction recognition model according to an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of the CP point structure in a training method for a non-covalent interaction recognition model according to an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram illustrating the steps of a method for identifying non-covalent interactions according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a training system for a non-covalent interaction recognition model according to an embodiment of the present invention;

[0044] Figure 6 This is a schematic diagram of the structure of a non-covalent interaction recognition system according to an embodiment of the present invention;

[0045] Figure 7 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] The technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0048] like Figure 1 As shown in the figure, the training method for a non-covalent interaction recognition model provided by an embodiment of the present invention specifically includes the following steps:

[0049] Step S101: Establish training samples based on the target molecular structure. The training samples include spatial structural information of the target molecular structure and corresponding labels. These labels include positional labels and topological feature labels. Positional labels are used to identify the spatial locations of non-covalent interactions within the target molecular structure, while topological feature labels are used to identify the properties of these interactions. Specifically, in the field of biological structures, NCIs are weak interactions crucial for maintaining the three-dimensional structure of molecules such as proteins and nucleic acids. NCIs also participate in many biological processes in which molecules specifically but transiently bind to each other. These interactions significantly impact drug design, crystallinity, and material design. Therefore, accurately labeling NCIs within molecular systems (including between and within molecules) is crucial for the research and development design process. By establishing training samples based on electron density methods to train the recognition model, the training samples contain detailed NCI position coordinates and topological properties, improving the accuracy of the trained recognition model. By establishing a recognition model, it is possible to identify non-covalent interactions within and between molecules in the target molecular structure. In this embodiment of the invention, 10,000 protein complexes from the pdbbind dataset are used to construct samples, and 200,000 corresponding non-covalent interaction position labels at different resolutions and the topological feature labels of the electronic density topological features of each point are obtained.

[0050] Step S102: Using spatial structure information as input, the initial machine learning model is used to predict non-covalent interactions in the target molecular structure. The initial machine learning model is then corrected based on the error between the prediction results and the label to obtain an identification model. The identification model is used to identify the positions and properties of non-covalent interactions in the target molecular structure.

[0051] Specifically, in this embodiment of the invention, the non-covalent interaction (NCI) positions are saddle points of the electron density in the target molecular structure, represented by grid-like coordinates. To facilitate the processing of grid-like training samples, this embodiment employs a Convolutional Neural Network (CNN) to build the recognition model, making the training process more accurate and the resulting recognition model more reliable. Training samples are continuously input into the CNN, and the CNN model is corrected based on the error between the CNN's prediction results and pre-labeled tags. After training, a trained recognition model is obtained. Based on the saddle point positions and electron density topological features output by the model, it can be used to identify the NCI positions and attributes in any target molecular structure. The NCI attributes represented by electron density topological features include, but are not limited to, the presence, type, direction, intensity, and affiliation of the NCI. The NCI also generates an embedding vector representing the local features of the NCI in the fully connected layer of the CNN. Furthermore, using training samples obtained with electron densities at different resolutions, the CNN model can directly derive recognition results at different resolutions for the same target molecular structure's spatial structure information. The differences between the different recognition results facilitate further analysis of noise in the NCI by those skilled in the art.

[0052] Specifically, step S101 above includes the following steps:

[0053] Step S201: The target molecular structure is standardized and its orientation corrected, and a predetermined number of atomic coordinates from the corrected target molecular structure are stored in a three-dimensional matrix to obtain spatial structure information. Specifically, before constructing the training samples, to improve computational efficiency during training, the target molecular structure is standardized and its orientation corrected by rotating it to a standard orientation where the ligand atom is located at the origin (0, 0, 0) and the acceptor atom is located on the Z-axis. Then, a predetermined number of atomic coordinates are extracted from the target molecular structure (in this embodiment, the molecular structure used is a macromolecule, and 1000 atomic coordinates are extracted from each molecular structure). Storing these atomic coordinates in a three-dimensional matrix yields input data for training the recognition model, i.e., spatial structure information. Using this method, a large amount of spatial structure information can be generated from a large number of target molecular structures, thereby obtaining complete training samples.

[0054] Step S202: Generate labels corresponding to spatial structure information based on electron density. Specifically, the spatial structure information obtained in step S201 is the original input data used to train the model from the training samples. Each original input data needs to correspond to a desired output, i.e., a label for the training sample. In this embodiment of the invention, the labels are the position, affiliation, and topological features of NCIs in the target molecular structure. Compared with traditional methods, the labels generated based on electron density in this step are more accurate in labeling NCIs in the target molecular structure.

[0055] Specifically, in one embodiment, after step S201 described above, the following step is further included:

[0056] Step S203: Based on different types of chemical features, the spatial structure information is mapped from a three-dimensional matrix to a four-dimensional matrix to obtain four-dimensional spatial structure information. Specifically, in a preferred embodiment, considering that the target molecule structure has multiple chemical features, such as atomic type, aromaticity, pre-training features, etc., the numerical values ​​and their distribution patterns in the three-dimensional matrix differ under different feature conditions. Therefore, as... Figure 2 As shown, the original spatial structure information is mapped from a three-dimensional feature vector to a four-dimensional feature vector representing different atomic features based on atomic features, which increases the dimension of the training features and makes the training results more accurate.

[0057] Specifically, in one embodiment, step S202 above includes the following steps:

[0058] Step S301: The target molecular structure is latticized to obtain a lattice molecular structure. Specifically, traditional techniques mostly use a number of summarized rules for NCI identification, such as the distance and angle between two atoms. This method has the drawback of being qualitative but not quantitative, and its ability to identify NCI in complex environments is poor. Electron density represents the probability of finding electrons at specific positions around an atom or molecule. Due to the difference in the strength of the interaction between NCI and covalent bonds, the distribution of electron density has a certain regularity. This invention, through extensive experimental measurements, has found that the positions of electron density saddle points are very close to the positions of weak interactions. Using electron density saddle points to mark NCI can greatly improve the accuracy of NCI identification in the target molecular structure. The specific position of the electron density saddle point needs to be represented in space using spatial coordinates. Therefore, the target molecular structure first needs to be divided into a lattice form to obtain the electron density values ​​of each lattice point, thereby determining the position and specific coordinates of the electron density saddle point.

[0059] Step S302: Obtain at least one electron density grid information with a preset resolution. The electron density grid information is used to store the electron density of each grid point in the grid molecular structure.

[0060] Specifically, to obtain these electron density saddle points, it is first necessary to know the electron density lattice information representing the electron density distribution in the target molecular structure (a single molecule or a combination of multiple molecules), and the electron density value within each lattice point of the target molecular structure. Thus, the location of the electron density saddle points can be found using the electron density values ​​of each lattice point.

[0061] Methods for obtaining electron density lattice information include, but are not limited to, experimental and computational methods. Commonly used methods include: methods based on X-ray crystallography, methods based on electron microscopy, and methods based on quantum chemical calculations. Traditional X-ray crystallography can analyze the fine three-dimensional structure of molecules at the atomic or near-atomic level, thereby obtaining molecular structure information through experiments and then calculating electron density lattice information. Experimental electron densities obtained through electron microscopy are directly saved as spatial lattice files and can be directly read. In the absence of experimental conditions, electron density can be obtained based on quantum chemical calculations using the readily available atomic spatial coordinates in the target molecular structure. This method can divide the target molecular structure into multiple small parts, i.e., multiple small molecular systems, and calculate the electron density value of each lattice point in each small molecular system.

[0062] In the above methods, on the one hand, due to limitations, the experimental conditions cannot be achieved by all laboratories; on the other hand, the electron density calculated based on quantum chemistry focuses too much on small molecule systems and often ignores the connections between small molecules in macromolecular systems. Therefore, from a macroscopic perspective, the calculation results are often erroneous. Based on this, this invention proposes a method for calculating electron density using Fourier and inverse Fourier transforms based on crystallographic theory to obtain electron density lattice information. Specifically, firstly, atomic coordinate information is extracted from the spatial structure information obtained in step S201. Then, the structure factor is calculated using Fourier transform based on the atomic coordinate information, where the calculation formula is:

[0063]

[0064] Where r represents atomic coordinate information, x, y, and z represent the three components of atomic coordinate information, and f n Let represent the diffraction factor of n different elements, s represent the reciprocal space vector, h, k, and l represent the three components of the reciprocal space vector respectively, F(s) represent the structure factor of the s vector (including diffraction amplitude and phase information), and 2πi is an imaginary number.

[0065] The structure factor is used to calculate and generate electron density lattice information at a preset resolution. This preset resolution is obtained by summing the preset reciprocal space vectors; that is, by varying the summation range of h, k, and l, electron densities at different resolutions can be obtained. Higher resolution yields clearer electron density maps with more accurate representations of interatomic bonds; lower resolution results in blurrier electron density maps, but a more accurate representation of the overall molecular framework. The calculation formula is as follows:

[0066]

[0067] Where r represents atomic coordinates, s(h,k,l) ​​represents the reciprocal space vector, F(s) is the diffraction amplitude of the s vector, ρ(r) represents electron density lattice information, and Vcell represents the unit cell volume. The method for calculating electron density based on crystallography not only considers the macroscopic connections between atoms in macromolecular systems but also obtains electron density maps with varying clarity by adjusting different preset resolutions, thus enabling the acquisition of electron density maps under multiple resolution conditions. Then, using the number and coordinates of the lattice points in the target molecular structure obtained in step S301, and taking different x, y, and z values ​​according to different lattice point coordinates, the electron density map is divided into discrete forms, yielding electron density lattice information. Since the saddle point positions obtained from electron density lattice information at different resolutions are also different, analyzing and comparing the saddle point positions at high resolution and low resolution can achieve noise removal from both macroscopic and microscopic perspectives, making the identification results more accurate.

[0068] Step S303: Determine the saddle point position based on the lattice position of the electron density saddle point in the electron density lattice information. The saddle point position is used to represent the spatial position of non-covalent interactions in the target molecule structure. Specifically, by calculating the reduced electron density gradient of each lattice point in the electron density lattice information, the saddle point position can be determined by finding the lattice point with the minimum value of the reduced electron density gradient.

[0069] Step S304: Determine the attribution relationship corresponding to each saddle point position based on atomic coordinate information. The attribution relationship is used to determine the matching relationship between non-covalent interactions and nearby atomic pairs. Specifically, after the spatial coordinates of the saddle point positions are determined, the position coordinates of NCI in space are determined. However, to determine which specific atomic pair in the molecule NCI belongs to, it is still necessary to comprehensively analyze and judge based on atomic coordinate information, combined with the properties of NCI and the distances between each atomic pair.

[0070] Step S305: Use the saddle point location and its corresponding attribution relationship as the location label for spatial structure information;

[0071] Step S306: Generate electron density topological features at saddle point locations using at least the electron density, first-order gradient, and Hessian matrix. These electron density topological features serve as topological feature labels for spatial structure information, describing the properties of non-covalent interactions (NCIs). Specifically, NCIs can be categorized into electrostatic interactions, π-effects, van der Waals forces, and hydrophobic effects, each with different commonly used descriptive methods. For example, π-effects are typically described using electron density, while van der Waals forces are usually described using Lagrange kinetic energy density. Commonly used electron density topological features for describing NCI properties and / or types include, but are not limited to: electron density, Lagrange kinetic energy density, Hamiltonian kinetic energy density, potential energy density, energy density, electron density Laplace, and electrostatic potential. All of these commonly used electron density topological features can be calculated using at least the electron density, first-order gradient, and Hessian matrix at the saddle point locations. Subsequently, during the training process, all the obtained electron density topological features are used as topological feature labels for spatial structure information. Thus, when using machine learning models for identification, the identification results can output various topological feature values ​​for each NCI. Those skilled in the art can directly select any of the results according to actual needs, which greatly facilitates experimental research.

[0072] The saddle point positions and electron density topological features of the target molecular structure are obtained through steps S301-S306. This complete process can identify NCIs in the target molecular structure. However, the computational costs for each instance are substantial, both financially and in terms of time. For long-term and large-scale NCI identification, the cost becomes incalculable. Therefore, in steps S101-S102, a machine learning identification model is trained using the large amount of data samples established by the methods in steps S301-S306. This solves the problem of high long-term costs in subsequent identification processes.

[0073] Specifically, in one embodiment, step S303 above includes the following steps:

[0074] Step S401: Divide the electron density grid information into multiple sub-information blocks of a first preset size. Specifically, by searching for the saddle point position of the electron density in the electron density grid information obtained in step S1, the coordinates of the NCI can be obtained. First, the electron density grid information is divided into multiple sub-information blocks of a first preset size in the spatial coordinate system, and the saddle point position is searched in each sub-information block to ensure the accuracy of the search. In this embodiment of the invention, the following is adopted: The range is used to divide the electron density grid information.

[0075] Step S402: Obtain ligand-acceptor atom pairs in the sub-information block, and use the area covered by a circle centered at the midpoint of the connection line between the ligand-acceptor atom pairs and with a preset length as the radius as the candidate range. Specifically, based on the characteristics of the electron density saddle point itself, its position is closer to the middle position of the two atoms in the atom pair. Therefore, all ligand-acceptor atom pairs in the sub-information block are obtained, and the area covered by the circle centered at the midpoint of the connection line between the atom pairs and with a preset length (using the method described in this embodiment of the invention) is selected. The range covered by the radius is used as the candidate range.

[0076] Step S403: Calculate the reduced electron density gradient (RDG) value of each grid point within the candidate range, and select the grid point with the smallest reduced electron density gradient value as the candidate saddle point. Specifically, in the electron density distribution, combined with the properties of mathematical saddle points, the gradient of a saddle point is 0. Due to the characteristics of discrete space, there is no point with an absolutely zero gradient. Therefore, calculate the reduced electron density gradient (RDG) value of each grid point within the candidate range, and select the grid point with the smallest reduced electron density gradient value as the candidate saddle point.

[0077] Step S404: Calculate the reduced electron density gradient values ​​of adjacent lattice points of the candidate saddle point, and take the lattice point with the smallest reduced electron density gradient value among the candidate saddle point and its adjacent lattice points as the saddle point position. Specifically, if the candidate saddle point falls on the edge of the candidate range, this saddle point is likely not the smallest gradient point in the vicinity. Since points outside the edge are not compared, but are very close to the candidate saddle point, the points outside the edge are likely to be the electron density gradient saddle points of the target atom pair. Therefore, the reduced electron density gradient values ​​of adjacent lattice points of the candidate saddle point are calculated, and the lattice point with the smallest reduced electron density gradient value among the candidate saddle point and its adjacent lattice points is taken as the saddle point position, ensuring the accuracy of the saddle point position.

[0078] Specifically, in one embodiment, after step S404 described above, the following step is further included:

[0079] Step S405: Discard saddle point locations where the eigenvalues ​​of the electron density Hessian matrix do not satisfy λ1 < λ2 < 0 < λ3. Specifically, since electron density saddle points are not unique, their electron density can decrease in two perpendicular directions and increase in a third direction; or it can increase in two perpendicular directions and decrease in a third direction. To avoid confusion caused by simultaneous statistical analysis of multiple saddle points, such as... Figure 3As shown, in this embodiment of the invention, the electron density saddle point [(3, -1)CP point], where the electron density decreases in two spatial directions and increases in a third direction, is used to mark the NCI position. That is, the saddle point where the electron density is maximum in two spatial directions and minimum in the third direction. Therefore, it is required that the eigenvalues ​​of the Hessian matrix at this point be two positive and one negative. Thus, a preferred method is to further filter the saddle point positions to obtain the desired (3, -1)CP point, discarding saddle point positions where the eigenvalues ​​of the electron density Hessian matrix do not satisfy λ1 < λ2 < 0 < λ3.

[0080] Step S406: When a saddle point corresponds to multiple ligand-acceptor atom pairs, the atom pair with the closest distance between the ligand and acceptor is selected as the target atom pair for labeling the saddle point. Specifically, since a saddle point can only label the non-covalent interaction of one atom pair, in this embodiment of the invention, when a saddle point corresponds to multiple ligand-acceptor atom pairs, the atom pair with the closest distance between the ligand and acceptor is selected as the target atom pair for labeling the saddle point.

[0081] Step S407: When the distance between any two saddle points is less than a preset distance, discard the saddle point with the lower electron density value. Specifically, when the distance between two saddle points is too close, based on experience, it is judged that one of the saddle points is likely an error point. Therefore, in this embodiment of the invention, when the distance between any two saddle points is less than a preset distance, discard the saddle point with the lower electron density value. Afterwards, the finally obtained saddle point positions can be used as location labels for spatial structure information during model training.

[0082] By executing the above steps, this invention provides a method for training a non-covalent interaction (NCI) identification model. This method acquires the three-dimensional spatial structure of the target molecule and uses the position and topological features of the NCI within this structure as labels corresponding to the target molecule, creating training samples. Then, combined with a convolutional neural network algorithm, a machine learning model capable of accurately identifying NCIs in macromolecule-ligand pairs is trained, improving the accuracy and efficiency of NCI identification. In the process of creating training samples, the spatial structure of the macromolecule-ligand pair is divided into multiple lattice points, and the electron density of each lattice point is calculated based on crystallographic theory to obtain more reliable electron density lattice point information. Based on the properties of electron density itself, it is known that the electron density saddle point is almost identical to the position of the non-covalent interaction. Then, the lattice point representing the electron density saddle point is identified through features such as the electron density gradient within each lattice point. The spatial coordinates of this lattice point are used as the spatial coordinates of the NCI, and the calculated topological features of the electron density at this lattice point are used as the topological features of the NCI, thereby accurately labeling a large number of target molecule structures. Furthermore, by employing a lattice of multiple resolutions to divide the molecule-ligand structure, training samples under multiple resolution conditions are obtained, thereby achieving NCI analysis of the target molecule structure from both microscopic and macroscopic perspectives, further improving the accuracy and reliability of NCI recognition. In addition, based on the different atomic characteristics of atoms in the target molecule structure, the training samples are mapped from three-dimensional vectors to four-dimensional vectors, further improving the detail of the convolutional neural network training, thus further enhancing the reliability of the recognition results of the recognition model provided in this embodiment of the invention.

[0083] like Figure 4 As shown in the figure, this embodiment of the invention also provides a method for identifying non-covalent interactions, the specific steps of which are as follows:

[0084] Step S103: Obtain the spatial structure information of the molecular structure to be identified.

[0085] Step S104: Input the spatial structure information of the molecular structure to be identified into the identification model obtained in step S102, and generate the position coordinates, attribution relationships and topological features of non-covalent interactions in the molecular structure to be identified through the identification model.

[0086] Specifically, the spatial structure information of the molecular structure to be identified can be obtained directly through experiments or simulations, and the spatial structure information can be directly input into the identification model obtained in step S102. The identification model can immediately output the spatial position coordinates of all NCIs in the molecular structure and the electron density topological features representing the NCI attributes. The available electron density topological features can be selected according to the needs of the experimenter, which greatly improves the speed and convenience of identification.

[0087] By performing the above steps, the non-covalent interaction (NCI) identification method provided by this embodiment of the invention can identify any input target molecular structure by using a complete training model obtained through a large number of training samples. This method can accurately identify the NCI position, attribution relationship and topological features in the target molecular structure in a short time, thereby improving the identification efficiency and accuracy of NCI.

[0088] like Figure 5 As shown, this embodiment of the invention also provides a training system for a non-covalent interaction recognition model, the system comprising:

[0089] The sample creation module 101 establishes training samples based on the target molecular structure. The training samples include spatial structure information of the target molecular structure and corresponding labels. The labels include positional labels and topological feature labels. Positional labels are used to identify the spatial locations of non-covalent interactions within the target molecular structure, and topological feature labels are used to identify the properties of non-covalent interactions within the target molecular structure. For details, please refer to the relevant description of step S101 in the above method embodiments, which will not be repeated here.

[0090] Training module 102 uses spatial structure information as input to predict non-covalent interactions in the target molecular structure using an initial machine learning model. It then corrects the initial machine learning model based on the error between the prediction results and the labels to obtain an identification model. This identification model is used to identify the positions and properties of non-covalent interactions in the target molecular structure. For details, please refer to the relevant description of step S102 in the above method embodiments; it will not be repeated here.

[0091] The present invention provides a training system for a non-covalent interaction identification model, which is used to execute the training method for a non-covalent interaction identification model provided in the above embodiments. Its implementation method and principle are the same. For details, please refer to the relevant description of the above method embodiments, which will not be repeated here.

[0092] like Figure 6 As shown, this embodiment also provides a system for identifying non-covalent interactions, the system comprising:

[0093] The information acquisition module 103 acquires the spatial structure information of the molecular structure to be identified. For details, please refer to the relevant description of step S103 in the above method embodiments, which will not be repeated here.

[0094] The identification module 104 is used to input the spatial structure information of the molecular structure to be identified into the identification model obtained in step S102, and to generate the position coordinates, attribution relationships, and topological features of non-covalent interactions in the molecular structure to be identified through the identification model. For details, please refer to the relevant description of step S104 in the above method embodiments, which will not be repeated here.

[0095] The present invention provides a non-covalent interaction identification system for executing a non-covalent interaction identification method provided in the above embodiments. Its implementation and principle are the same, and for details, please refer to the relevant description of the above method embodiments, which will not be repeated here.

[0096] Figure 7 An electronic device according to an embodiment of the present invention is shown. The device includes a processor 901 and a memory 902, which can be connected via a bus or other means. Figure 7 Taking the example of a connection between China and Israel via a bus.

[0097] Processor 901 can be a Central Processing Unit (CPU). Processor 901 can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0098] The memory 902, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the above method embodiments. The processor 901 executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions, and modules stored in the memory 902, thereby implementing the methods in the above method embodiments.

[0099] The memory 902 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor 901, etc. Furthermore, the memory 902 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 902 may optionally include memory remotely located relative to the processor 901, and these remote memories may be connected to the processor 901 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0100] One or more modules are stored in memory 902 and, when executed by processor 901, perform the methods described in the above method embodiments.

[0101] The specific details of the aforementioned electronic device can be understood by referring to the relevant descriptions and effects in the above method embodiments, and will not be repeated here.

[0102] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The implemented program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0103] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A training method for a non-covalent interaction recognition model, characterized in that, The method includes: Training samples are established based on the target molecular structure. These training samples include spatial structure information of the target molecular structure and corresponding labels. The labels include positional labels and topological feature labels. Positional labels are used to identify the spatial locations of non-covalent interactions within the target molecular structure, and topological feature labels are used to identify the properties of non-covalent interactions within the target molecular structure. The initial machine learning model is built using a convolutional neural network. Establishing training samples based on the target molecular structure includes: standardizing and correcting the orientation of the target molecular structure, and storing a predetermined number of atomic coordinates in the corrected target molecular structure into a three-dimensional matrix to obtain the spatial structure information; generating labels corresponding to the spatial structure information based on electron density; generating labels corresponding to the spatial structure information based on electron density includes: latticizing the target molecular structure to obtain a lattice molecular structure; obtaining… At least one electron density lattice information of preset resolution is provided, which is used to store the electron density of each lattice point within the lattice molecular structure. Saddle point positions are determined based on the lattice positions of the electron density saddle points in the electron density lattice information, where the saddle point positions represent the spatial positions of non-covalent interactions within the target molecular structure. Assignment relationships are determined based on the atomic coordinate information, where the assignment relationships represent the matching relationships between non-covalent interactions and nearby atomic pairs. The saddle point positions and their corresponding assignment relationships are used as positional labels for the spatial structure information. At least the electron density, first-order gradient, and Hessian matrix of the saddle point positions are used to generate electron density topological features of the saddle point positions, and these topological features are used as topological feature labels for the spatial structure information, where the electron density topological features describe the properties of non-covalent interactions. Using the spatial structure information as input, an initial machine learning model is used to predict non-covalent interactions in the target molecular structure. The initial machine learning model is then corrected based on the error between the prediction results and the label to obtain an identification model. This identification model is used to identify the positions and properties of non-covalent interactions in the target molecular structure.

2. The method according to claim 1, characterized in that, Based on different types of chemical characteristics, the spatial structure information is mapped from a three-dimensional matrix to a four-dimensional matrix to obtain four-dimensional spatial structure information.

3. The method according to claim 1, characterized in that, Determining the saddle point location based on the grid point location of the electron density saddle point in the electron density grid information includes: The electron density grid information is divided into multiple sub-information blocks of a first preset range size; Obtain the ligand-receptor atom pairs in the sub-information block, and take the range covered by the midpoint of the connection line of the ligand-receptor atom pairs as the center and the preset length as the radius as the candidate range; Calculate the reduced electron density gradient value of each grid point within the candidate range, and select the grid point with the smallest reduced electron density gradient value as the candidate saddle point; Calculate the reduced electron density gradient values ​​of the adjacent lattice points of the candidate saddle point, and take the lattice point with the smallest reduced electron density gradient value among the candidate saddle point and the adjacent lattice points as the saddle point position.

4. The method according to claim 3, characterized in that, After calculating the reduced electron density gradient values ​​of the adjacent lattice points of the candidate saddle point and taking the lattice point with the smallest reduced electron density gradient value among the candidate saddle point and the adjacent lattice points as the saddle point location, the method further includes: The eigenvalues ​​of the Hessian matrix, which do not satisfy the given electron density, are not satisfied. The saddle point position; When a saddle point location corresponds to multiple ligand-acceptor atom pairs, the atom pair with the closest distance between the ligand and the acceptor is taken as the target atom pair marked at the saddle point location; When the distance between any two saddle points is less than a preset distance, the saddle point with the smaller electron density value is discarded.

5. A method for identifying non-covalent interactions, characterized in that, The method includes: Obtain spatial structural information of the molecular structure to be identified; The spatial structure information of the molecular structure to be identified is input into the identification model obtained by the method described in any one of claims 1-4, and the position coordinates, attribution relationships and topological features of the non-covalent interactions in the molecular structure to be identified are generated through the identification model.

6. A non-covalent interaction identification system, characterized in that, The system includes: The information acquisition module obtains the spatial structure information of the molecular structure to be identified; The identification module is used to input the spatial structure information of the molecular structure to be identified into the identification model obtained by the method as described in any one of claims 1-4, and to generate the position coordinates, attribution relationships and topological features of the non-covalent interactions in the molecular structure to be identified through the identification model.

7. An electronic device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to perform the method as described in any one of claims 1-5.

Images