Energy prediction method, device, apparatus, and computer readable storage medium
By combining molecular force field algorithms and graph neural network models, the problem of low computational efficiency in quantum mechanics has been solved, enabling efficient and accurate prediction of the energy of intermolecular interactions in drugs, thereby improving the efficiency of new drug discovery and reducing R&D costs.
Patent Information
- Application Number
- CN202310151787.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-14
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-02-14
AI Technical Summary
Existing technologies cannot effectively utilize quantum mechanical methods to calculate the interaction energy between small drug molecules and proteins, resulting in low computational efficiency and making it impossible to complete large-scale drug molecule screening.
The interaction energy between the compound and the protein was initially calculated using a molecular force field algorithm. Then, a trained graph neural network model was used to predict the energy difference. Finally, the interaction energy was determined by combining the results of molecular force field and quantum mechanics.
While improving computational efficiency, it ensures high accuracy in energy prediction, enabling faster screening of target drug compounds, improving the efficiency of new drug discovery, and reducing R&D costs.
Smart Images

Figure CN116978484B_ABST
Abstract
Description
Technical Field
[0001] This application relates to data processing technology, and more particularly to an energy prediction method, apparatus, device, and computer-readable storage medium. Background Technology
[0002] Currently, in the field of drug discovery, computational chemistry methods are typically used to calculate the interaction energy between small drug molecules and large protein molecules to determine their binding affinity. Quantum mechanics is currently the most accurate computational chemistry method, but also the most computationally complex. However, it's not currently possible to directly use quantum mechanics to calculate the interaction energy between small drug molecules and proteins because proteins are large molecules with a large number of atoms, significantly increasing the computational complexity of quantum mechanics, which is beyond the capabilities of current computing power. To address this issue, the industry has developed the Fragment Molecular Orbital (FMO) method. In FMO, proteins are divided into multiple fragments at a similar amino acid granularity, and then quantum mechanics is used to calculate the interaction energy between these protein fragments and the small molecule. In drug molecule screening processes, hundreds of thousands or even millions of molecules may need to be screened; however, the computational speed of FMO is limited, making large-scale calculations impossible. Summary of the Invention
[0003] This application provides an energy prediction method, apparatus, and computer-readable storage medium that can improve computational efficiency while achieving the accuracy of quantum mechanical calculations.
[0004] The technical solution of this application embodiment is implemented as follows: This application provides an energy prediction method, the method comprising: Obtain the compound atomic information of the compound to be predicted and the protein atomic information of the protein to be predicted; The intermolecular reference interaction energy of the compound to be predicted and the protein to be predicted is determined based on a preset molecular force field algorithm, the atomic information of the compound, and the atomic information of the protein. Obtain the trained graph neural network model; The trained graph neural network model is used to predict the atomic information of the compound and the atomic information of the protein to obtain the energy difference value, which is the difference between the intermolecular interaction energy determined based on quantum mechanics and the intermolecular interaction energy determined based on the molecular force field. The target intermolecular interaction energy between the compound to be predicted and the protein to be predicted is determined based on the intermolecular reference interaction energy and the energy difference.
[0005] This application provides an energy prediction device, including:
[0006] The first acquisition module is used to acquire the compound atomic information of the compound to be predicted and the protein atomic information of the protein to be predicted. The first determination model determines the intermolecular reference interaction energy of the compound to be predicted and the protein to be predicted based on a preset molecular force field algorithm, the atomic information of the compound, and the atomic information of the protein. The second acquisition module is used to acquire the trained graph neural network model; The first prediction module is used to predict the atomic information of the compound and the atomic information of the protein using the trained graph neural network model to obtain an energy difference value. The energy difference value is the difference between the intermolecular interaction energy determined based on quantum mechanics and the intermolecular interaction energy determined based on the molecular force field. The second determining module is used to determine the target intermolecular interaction energy between the compound to be predicted and the protein to be predicted based on the intermolecular reference interaction energy and the energy difference.
[0007] This application provides an electronic device, the electronic device comprising: Memory is used to store executable instructions for a computer; The processor, when executing computer-executable instructions stored in the memory, implements the energy prediction method provided in the embodiments of this application.
[0008] This application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the energy prediction method provided in this application.
[0009] This application provides a computer program product, including computer-executable instructions, which, when executed by a processor, implement the energy prediction method provided in this application.
[0010] The embodiments of this application have the following beneficial effects: After obtaining the compound atomic information of the compound to be predicted and the protein atomic information of the protein to be predicted, the intermolecular reference interaction energy of the compound and the protein to be predicted is first determined using a computationally efficient molecular force field algorithm. Then, a trained graph neural network model is used to predict the compound atomic information and the protein atomic information to obtain an energy difference. The energy difference is the difference between the intermolecular interaction energy determined based on quantum mechanics and the intermolecular interaction energy determined based on molecular force field. Then, the target intermolecular interaction energy of the compound and the protein to be predicted is determined based on the intermolecular reference interaction energy and the energy difference. The target intermolecular interaction energy obtained at this time is equivalent to the interaction energy determined using quantum mechanics. Therefore, it can improve computational efficiency while ensuring higher accuracy in energy prediction results. Attached Figure Description
[0011] Figure 1 This is a schematic diagram of the new drug development process in related technologies; Figure 2 This is a schematic diagram of the energy prediction system architecture provided in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application; Figure 4A This is a schematic diagram illustrating an implementation flow of the energy prediction method provided in an embodiment of this application; Figure 4B This is a schematic diagram illustrating the process of determining the energy of intermolecular reference interactions provided in an embodiment of this application; Figure 4C This is a schematic diagram of the implementation process for determining the energy difference provided in the embodiments of this application; Figure 5A This is a schematic diagram of an implementation process for training a preset graph neural network model provided in an embodiment of this application; Figure 5B This is a schematic diagram illustrating the data construction implementation process provided in the embodiments of this application; Figure 5C This is a schematic diagram of another implementation process for training a preset graph neural network model provided in an embodiment of this application; Figure 5D This is a schematic diagram illustrating the implementation process of testing a trained graph neural network model provided in an embodiment of this application; Figure 6 This is a schematic diagram of another implementation process of the energy prediction method provided in the embodiments of this application; Figure 7 This is a schematic diagram of the DeepFMO model provided in the embodiments of this application; Figure 8 This is a schematic diagram of the prediction results of the energy prediction method provided in this application embodiment on the same distribution test set; Figure 9 This is a schematic diagram illustrating the prediction results of the energy prediction method provided in this application on different distributed test sets. Detailed Implementation
[0012] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0013] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.
[0014] In the following description, the terms "first, second, third" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.
[0015] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.
[0016] Before providing a further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application will be explained, and the nouns and terms involved in the embodiments of this application shall be interpreted as follows.
[0017] 1) Graph Neural Network (GNN) refers to a general term for algorithms that use neural networks to learn graph-structured data, extract and discover features and patterns in graph-structured data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation. 2) Graph Convolution Networks (GCNs) extend convolution operations from traditional data (images or grids) to graph data. The key is learning a function that generates node representation vectors by aggregating its own features and those of its neighbors. 3) Graph Attention Network (GAT): A novel neural network architecture based on graph structure data, which utilizes hidden self-attention layers to address the shortcomings of previous methods based on graph convolution or its approximations. 4) The loss function is a function that maps the values of a random event or its related random variables to non-negative real numbers to represent the "risk" or "loss" of that random event; 5) Proteins are essential components of all cells and tissues in the human body. All vital components of the body require the participation of proteins. 6) Amino acids are organic compounds containing a basic amino group and an acidic carboxyl group. Amino acids are the basic building blocks of proteins, giving proteins their specific molecular structure and enabling their biochemical activity.
[0018] The energy prediction method provided in this application can be applied to new drug development and materials development processes. To better understand the energy prediction method provided in this application, the new drug development process will be described first. Figure 1 This is a schematic diagram of the new drug development process in related technologies, such as... Figure 1 As shown, the process includes: Step 001: Target identification and verification.
[0019] In step 001, the proteins corresponding to the diseases targeted by the newly developed drug compounds are first identified. Then, the drug compounds are bound to the proteins, and the drug compounds that bind to the proteins are identified as candidate drug compounds.
[0020] Step 002, compound screening and lead discovery.
[0021] In this step, the intermolecular interaction energy between each candidate drug compound and the protein is determined. The determined energies are then used to analyze the molecular properties and the binding ability of the molecules to the protein pocket, thereby screening out the target drug compounds from the candidate drug compounds.
[0022] In related technologies, the FMO algorithm is used to determine the intermolecular interaction energy between each candidate drug compound and a protein. Within the FMO method, quantum mechanics is used to calculate the interaction energy between protein fragments and small molecules. In drug molecule screening processes, hundreds of thousands or even millions of molecules often need to be screened. The computational efficiency limitations of FMO can severely impact drug development efficiency. Step 003, preclinical research.
[0023] Step 004, Clinical Trial.
[0024] This application provides an energy prediction method, apparatus, device, and computer-readable storage medium, which can solve the problem of low computational efficiency when calculating the intermolecular interaction energy of proteins and compounds using quantum mechanics in related technologies. The following describes exemplary applications of the electronic device provided in this application. The device provided in this application can be implemented as a terminal or as a server. The following will describe exemplary applications when the device is implemented as a server.
[0025] See Figure 2 , Figure 2 This is a schematic diagram of the architecture of the energy prediction system 100 provided in the embodiments of this application, as shown below. Figure 2 As shown, the energy prediction system 100 includes a terminal 200, a network 300, and a server 400. The terminal 200 is connected to the server 400 through the network 300, which can be a wide area network, a local area network, or a combination of both.
[0026] Terminal 200 is used to identify the proteins corresponding to the diseases targeted by newly developed drug compounds. Then, it binds the drug compounds to the proteins, identifying the drug compounds that bind to the proteins as candidate drug compounds. Terminal 200 sends an energy prediction request to server 400, carrying information on at least one candidate drug compound and the protein to be predicted. Server 400 receives the energy prediction request and obtains the information on at least one candidate drug compound and the protein to be predicted. Server 400 uses a preset molecular force field algorithm to determine the intermolecular reference interaction energy between each candidate drug compound and the protein to be predicted. Server 400 then uses a trained graph neural network model to predict the candidate drug compound and protein information, obtaining an energy difference. The energy difference is the difference between the intermolecular interaction energy determined based on quantum mechanics and the intermolecular interaction energy determined based on molecular force field. Based on the intermolecular reference interaction energy and the energy difference, the target intermolecular interaction energy between the drug compound and the protein to be predicted is determined. The target intermolecular interaction energy obtained in this way is equivalent to the interaction energy determined using quantum mechanics, thus improving computational efficiency while ensuring higher accuracy in energy prediction results. Server 400 sends an energy prediction response carrying the energy of the target interactions between each intermolecular element to terminal 200. Terminal 200 determines the target drug compound from at least one candidate drug compound based on the energy of the target interactions between each intermolecular element, thereby using the target drug compound for clinical research, which can improve the efficiency of new candidate drug discovery and reduce R&D costs.
[0027] In some embodiments, server 400 may be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. Terminal 200 may be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, smart TV, in-vehicle terminal, etc., but is not limited thereto. Terminals and servers can be directly or indirectly connected via wired or wireless communication, which is not limited in this embodiment.
[0028] See Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device can be a terminal or a server. Figure 3 In this example, we will use server 400 as an example to illustrate the concept. Figure 3 The server 400 shown includes at least one processor 410, memory 450, at least one network interface 420, and a user interface 430. The various components in server 400 are coupled together via a bus system 440. It is understood that the bus system 440 is used to implement communication between these components. In addition to a data bus, the bus system 440 also includes a power bus, a control bus, and a status signal bus. However, for clarity, ... Figure 3 The general labeled all buses as Bus System 440.
[0029] Processor 410 can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or any conventional processor, etc.
[0030] User interface 430 includes one or more output devices 431 that enable the presentation of media content, including one or more speakers and / or one or more visual displays. User interface 430 also includes one or more input devices 432, including user interface components that facilitate user input, such as a keyboard, mouse, microphone, touch screen display, camera, other input buttons and controls.
[0031] The memory 450 may be removable, non-removable, or a combination thereof. Exemplary hardware devices include solid-state storage, hard disk drives, optical disk drives, etc. The memory 450 may optionally include one or more storage devices physically located away from the processor 410.
[0032] The memory 450 may include volatile memory or non-volatile memory, or both. The non-volatile memory may be read-only memory (ROM), and the volatile memory may be random access memory (RAM). The memory 450 described in this application embodiment is intended to include any suitable type of memory.
[0033] In some embodiments, memory 450 is capable of storing data to support various operations, examples of which include programs, modules, and data structures or subsets or supersets thereof, as illustrated below.
[0034] Operating system 451 includes system programs for handling various basic system services and performing hardware-related tasks, such as the framework layer, core library layer, driver layer, etc., for implementing various basic business functions and handling hardware-based tasks; The network communication module 452 is used to reach other electronic devices via one or more (wired or wireless) network interfaces 420, exemplary network interfaces 420 including: Bluetooth, WiFi, and Universal Serial Bus (USB), etc. Presentation module 453 is configured to enable the presentation of information (e.g., a user interface for operating peripheral devices and displaying content and information) via one or more output devices 431 associated with user interface 430 (e.g., a display screen, a speaker, etc.). The input processing module 454 is used to detect and translate one or more user inputs or interactions from one or more input devices 432.
[0035] In some embodiments, the apparatus provided in this application can be implemented in software. Figure 3 An energy prediction device 455 stored in memory 450 is shown. This device can be software in the form of programs and plug-ins, and includes the following software modules: a first acquisition module 4551, a first determination module 4552, a second acquisition module 4553, a first prediction module 4554, and a second determination module 4555. These modules are logically connected and can therefore be arbitrarily combined or further separated according to their implemented functions. The functions of each module will be described below.
[0036] In other embodiments, the apparatus provided in this application can be implemented in hardware. As an example, the apparatus provided in this application can be a processor in the form of a hardware decoding processor, which is programmed to execute the energy prediction method provided in this application. For example, the processor in the form of a hardware decoding processor can use one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.
[0037] The energy prediction method provided in this application will be described in conjunction with exemplary applications and implementations of the server provided in the embodiments of this application.
[0038] To better understand the energy prediction method provided in the embodiments of this application, we will first explain artificial intelligence, its various branches, and the application fields involved in the method provided in the embodiments of this application.
[0039] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.
[0040] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly include computer vision, speech processing, natural language processing, and machine learning / deep learning. The energy prediction method provided in this application primarily relates to machine learning, which will be described below.
[0041] Key technologies in speech technology include Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and voiceprint recognition. Enabling computers to hear, see, speak, and feel is the future direction of human-computer interaction, with speech emerging as one of the most promising methods.
[0042] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, and inductive learning.
[0043] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, and smart customer service. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.
[0044] The energy prediction method provided in the embodiments of this application will be described below. As mentioned above, the electronic device implementing the energy prediction method of the embodiments of this application can be a terminal, a server, or a combination of both. Therefore, the executing entity of each step will not be described again below.
[0045] See Figure 4A , Figure 4A This is a schematic diagram of an implementation flow of the energy prediction method provided in this application embodiment, which will be combined with... Figure 4A The steps shown are explained.
[0046] In step 101, the compound atomic information of the compound to be predicted and the protein atomic information of the protein to be predicted are obtained.
[0047] In some embodiments, the compound atomic information of the compound to be predicted includes the atom identifier, number of atoms, valence of each atom, and three-dimensional coordinate information of each atom. The protein atomic information of the protein to be predicted includes the atom identifier, number of atoms, valence of each atom, and three-dimensional coordinate information of each atom.
[0048] The atom identifiers, number of atoms, and valence of each atom in the compound to be predicted can be obtained based on the chemical formula of the compound. The three-dimensional coordinates of each atom in the compound to be predicted can be determined according to the molecular structure of the compound, with the position of a certain atom in the compound as the origin. Similarly, the atom identifiers, number of atoms, and valence of each atom in the protein to be predicted can be obtained based on the chemical formula of the protein to be predicted. The three-dimensional coordinates of each atom in the protein to be predicted can be determined according to the molecular structure of the protein, with the position of a certain atom in the protein as the origin.
[0049] In step 102, the intermolecular reference interaction energy of the compound to be predicted and the protein to be predicted is determined based on a preset molecular force field algorithm, the atomic information of the compound, and the atomic information of the protein.
[0050] In some embodiments, before step 102, the protein to be predicted is first segmented based on the protein atom information of the protein to be predicted to obtain multiple protein fragments to be predicted, and the formal charge between each protein fragment to be predicted and the compound to be predicted is determined. Each protein fragment to be predicted and the compound to be predicted are then paired to obtain multiple pairing data, wherein the pairing data includes the fragment atom information of the protein fragment to be predicted, the compound atom information of the compound to be predicted, and the formal charge between the protein fragment to be predicted and the compound to be predicted.
[0051] Since amino acids are the basic building blocks of proteins, the segmentation of the predicted protein is based on its atomic information, using amino acids as the segmentation granularity. Hydrogen groups are then added to both ends of the segmented results to obtain the predicted protein fragments. The formal charge between the predicted protein fragment and the predicted compound represents the charge after their binding. The formal charge can be 0 or a positive or negative integer, such as +1, -1, +2, -2, etc. The formal charge between the predicted protein fragment and the predicted compound can be determined using the GetFormalCharge function provided in the Cheminformatics and Machine Learning Toolkit.
[0052] Based on this, in some embodiments, see Figure 4B , Figure 4A Step 102 can be implemented through steps 1021 to 1022 as described below, in conjunction with the following... Figure 4B Please provide an explanation.
[0053] In step 1021, based on a preset molecular force field algorithm and each pairing data to be predicted, the intermolecular reference interaction energy between each protein fragment to be predicted and the compound to be predicted is determined.
[0054] Intermolecular forces refer to the attractive or repulsive forces between molecular entities (or between groups within the same molecular entity), other than those caused by chemical bonding, interactions between ions (or ionic groups), or electrostatic interactions with neutral molecules. The intermolecular interaction energy can be determined based on the intermolecular forces and the distance between molecules. In step 1021, a preset molecular force field algorithm is used to determine the molecular reference interaction energy between the protein fragment to be predicted and the compound to be predicted. Compared to determining the intermolecular interaction energy between the protein fragment to be predicted and the compound to be predicted using quantum mechanics-based methods, the computational efficiency is greatly improved, but the computational accuracy is lower than that of quantum mechanics-based algorithms. In some embodiments, the preset molecular force field algorithm can be the xTB-IFF force field algorithm, or other molecular force field algorithms. The xTB-IFF force field algorithm is based on wavefunction input information and can consider the actual electronic structure, including items such as outer charge, orbital interactions, and three-body dispersion correction, which are not typically considered in simple molecular force fields. It can provide a good template for understanding intermolecular interactions. Furthermore, the xTB-IFF force field algorithm is fast, providing results for large systems with thousands of atoms within minutes, while quantum mechanical calculation algorithms take a day or even longer, thus greatly improving computational efficiency.
[0055] In step 1022, the intermolecular reference interaction energies between each protein fragment to be predicted and the compound to be predicted are accumulated to obtain the intermolecular reference interaction energies between the compound to be predicted and the protein to be predicted.
[0056] In some embodiments, each pair of data to be predicted determines an intermolecular reference interaction energy between a protein fragment to be predicted and the compound to be predicted. In step 1022, the intermolecular reference interaction energies between each protein fragment to be predicted and the compound to be predicted are summed to obtain the intermolecular reference interaction energy between the compound to be predicted and the protein to be predicted.
[0057] Through steps 1021 to 1022, the molecular force field algorithm can determine the intermolecular reference interaction energy of the compound and protein to be predicted in a short time, which can improve the computational efficiency of energy prediction.
[0058] See also Figure 4A The following will be an explanation following step 102 above.
[0059] In step 103, the trained graph neural network model is obtained.
[0060] In some embodiments, the graph neural network model can be a graph convolution network (GCN), a graph attention network (GAN), an equivariant graph neural network (EGNN), etc. In this embodiment, the graphs to be processed are those of small molecules of compounds and large molecules of proteins with geometric features. In this graph, in addition to some intrinsic features of atoms, the geometric feature of the three-dimensional coordinates of each atom in space also needs to be considered. Unlike general features, these geometric features often possess some symmetry and equivariance. Equivariant graph neural network models can model these equivariant symmetry features well. Therefore, in this embodiment, a trained graph neural network model is used as an example of a trained equivariant graph neural network model for illustration. Equivariant graph neural network models do not require computation of expensive high-order representations in intermediate layers for rotations, translations, reflections, permutations, etc.
[0061] In step 104, the trained graph neural network model is used to predict the atomic information of the compound and the atomic information of the protein to obtain the energy difference.
[0062] The trained graph neural network model is used to predict the energy difference between the intermolecular interaction energy determined by the molecular force field algorithm and the intermolecular interaction energy determined by quantum mechanics.
[0063] In some embodiments, see Figure 4C , Figure 4A Step 103 shown can be implemented through steps S1041 to S1042 as described below. Figure 4C Please provide an explanation.
[0064] In step 1041, the trained graph neural network model is used to perform prediction processing on each of the protein fragments to be predicted and the compounds to be predicted, so as to obtain the fragment energy difference between each protein fragment to be predicted and the compounds to be predicted.
[0065] In some embodiments, the trained graph neural network model first determines the embedding representation vector of each node in the graph data consisting of the protein fragment to be predicted and the compound to be predicted. Then, by aggregating information from other nodes around each node, the network layers in the trained graph neural network model learn and update the embedding representation vector of each node. Finally, the output layer in the trained graph neural network model determines the fragment energy difference between the protein fragment to be predicted and the compound to be predicted based on the embedding representation vector of each node.
[0066] In step 1042, the fragment energy differences between each protein fragment to be predicted and the compound to be predicted are accumulated to obtain the energy difference value.
[0067] In step 104, the energy difference between each protein fragment and compound to be predicted is determined by the trained graph neural network model. The energy differences of each fragment are then accumulated to obtain the energy difference between the intermolecular interaction energy determined by quantum mechanics and the intermolecular interaction energy determined by molecular force field between the protein and the compound to be predicted. This provides a data basis for obtaining the intermolecular interaction energy determined by quantum mechanics in the subsequent process.
[0068] See also Figure 4A The following will be an explanation following step 104 above.
[0069] In step 105, the target intermolecular interaction energy between the compound to be predicted and the protein to be predicted is determined based on the intermolecular reference interaction energy and the energy difference.
[0070] In some embodiments, the intermolecular reference interaction energy and the energy difference are added together to obtain the target intermolecular interaction energy for the compound and protein to be predicted. Since the energy difference characterizes the difference between the intermolecular interaction energy determined based on quantum mechanics and the intermolecular interaction energy determined based on molecular force fields, the target intermolecular interaction energy obtained by adding the intermolecular reference interaction energy obtained using the molecular force field algorithm and the energy difference is equivalent to the intermolecular interaction energy determined using quantum mechanics, thereby improving the accuracy of the intermolecular interaction energy.
[0071] In the energy prediction method provided in this application embodiment, after obtaining the compound atomic information of the compound to be predicted and the protein atomic information of the protein to be predicted, the intermolecular reference interaction energy of the compound to be predicted and the protein to be predicted is first determined using a computationally efficient molecular force field algorithm. Then, a trained graph neural network model is used to predict the compound atomic information and the protein atomic information to obtain an energy difference. The energy difference is the difference between the intermolecular interaction energy determined based on quantum mechanics and the intermolecular interaction energy determined based on molecular force field. Then, the intermolecular target interaction energy of the compound to be predicted and the protein to be predicted is determined based on the intermolecular reference interaction energy and the energy difference. The intermolecular target interaction energy obtained at this time can be equivalent to the intermolecular interaction energy determined using quantum mechanics. Therefore, it can improve computational efficiency while ensuring a higher accuracy energy prediction result.
[0072] In some embodiments, prior to step 103, it can be done by Figure 5A Steps 201 to 203 shown demonstrate the training of a preset graph neural network model to obtain a trained graph neural network model. The following steps, in conjunction with... Figure 5A Each step is explained.
[0073] In step 201, a preset graph neural network model is obtained, and a training dataset is obtained.
[0074] The training dataset includes multiple training pairs and label information for each pair. The training pairs include atomic information of training compounds, atomic information of training protein fragments, and formal charges between the training compounds and training protein fragments. The label information for the training pairs is the actual energy difference between the intermolecular interaction energy between the training compounds and training proteins determined using a preset molecular force field algorithm and the intermolecular interaction energy determined using quantum mechanics.
[0075] In step 202, the preset graph neural network model is used to perform prediction processing on each training pair data to obtain the prediction energy difference corresponding to each training pair data.
[0076] In some embodiments, the graph neural network model first determines the embedding representation vector of each node in the graph data consisting of training protein fragments and training compounds in the training pairing data. Then, the network layers in the graph neural network model learn and update the embedding representation vector of each node by aggregating information from other nodes around each node. Finally, the output layer in the graph neural network model determines the predicted energy difference corresponding to the training pairing data based on the embedding representation vector of each node.
[0077] In step 203, the preset graph neural network model is trained using the label information of each training pair data and the difference in predicted energy corresponding to each training pair data to obtain a trained graph neural network model.
[0078] In some embodiments, based on a preset loss function, the loss value is determined using the predicted energy differences determined by the graph neural network model and their corresponding label information. In some embodiments, the loss function can be a mean squared error function, a minimum variance function, etc. The loss value is fed back to the graph neural network model through back-forward to determine the gradient of each network layer in the graph neural network model, and the parameters of the graph neural network model are updated using the Adaptive Moment Estimation (Adam) algorithm until the training termination condition is met, resulting in a trained graph neural network model. The training termination condition can be reaching a preset number of training iterations, the loss value being less than a preset loss threshold, or the difference between two consecutive loss values being less than a preset difference threshold.
[0079] The training process of the graph neural network model is completed through steps 201 to 203 above, resulting in a trained graph neural network model. This trained graph neural network model can then be used to predict the difference between the intermolecular interaction energy determined based on quantum mechanics and the intermolecular interaction energy determined based on the molecular force field.
[0080] In some embodiments, prior to step 201, it can be done by Figure 5B Steps 301 to 306 shown illustrate obtaining the training dataset, validation dataset, first test dataset, and second test dataset. The following will combine... Figure 5B Please provide an explanation.
[0081] In step 301, multiple protein data are retrieved from the protein database.
[0082] The protein database can be an FMO database or a PDB database. The protein data includes initial protein information and the initial compound information corresponding to that initial protein information. In this embodiment, an FMO database is used as an example. An FMO database is a web-based public database used to store, manage, and share computational datasets of fragment molecular orbitals describing complex interactions between biomacromolecules. Each item in an FMO database contains background data, such as how the data was compiled, the total energy of each molecular system, the inter-fragment interaction energy, and the decomposition analysis values of the interaction energy.
[0083] In step 302, each initial protein information is segmented to obtain multiple initial protein fragments, and the formal charge between each initial protein fragment and the initial compound information is determined.
[0084] In some embodiments, each initial protein information is segmented at the amino acid level to obtain multiple segmentation results. Hydrogen groups are then added to the segmentation results to obtain multiple initial protein fragments. A single protein molecule can be segmented into dozens or even hundreds of protein fragments. The formal charge between each initial protein fragment and the initial compound information can be determined using preset functions provided in cheminformatics and machine learning toolkits, such as the GetFormalCharge function.
[0085] In step 303, each initial protein fragment and the initial compound data are paired to obtain multiple initial pairing data.
[0086] Here, multiple initial protein fragments obtained by segmenting the initial protein data are paired with the corresponding initial compound data to obtain multiple initial pairing data. Each initial pairing data includes the atomic information of an initial protein fragment, the atomic information of an initial compound, and the formal charge between the initial protein fragment and the initial compound.
[0087] In step 304, the tag information for each initial pairing is determined.
[0088] In some embodiments, the first intermolecular interaction energy of each initial protein fragment and the initial compound information is first determined based on a molecular force field algorithm. For example, the xTB-IFF force field algorithm can be used to determine the first intermolecular interaction energy of each initial protein fragment and the initial compound information. Then, a quantum mechanical algorithm is used to determine the second intermolecular interaction energy of each initial protein fragment and the initial compound information, and the energy difference between each second intermolecular interaction energy and the corresponding first intermolecular interaction energy is determined as the tag information of each initial pairing data.
[0089] In step 305, an initial dataset is constructed using the initial pairing data and the label information of the initial pairing data.
[0090] Here, the initial dataset includes multiple initial pairings and the label information corresponding to each initial pairing.
[0091] In step 306, the initial dataset is segmented to obtain a training dataset, a first test dataset, a second test dataset, and a validation dataset.
[0092] In some embodiments, a training dataset is first randomly segmented from the initial dataset according to a preset training data ratio to obtain a first residual dataset. The training data ratio can be a real number between 0 and 1, generally greater than 0.5, such as 0.8 or 0.75. The first residual dataset consists of the label information corresponding to the other initial paired data and other paired data in the initial dataset excluding the training dataset. Then, based on the classification information of each protein in the training dataset, a first test dataset and a second test dataset are segmented from the first residual dataset to obtain a second residual dataset. The classification information of the proteins in the first test dataset exists in the training dataset, meaning the first test dataset and the training dataset are distributed in the same way. The classification information of the proteins in the second test dataset does not exist in the training dataset, meaning the second test dataset and the training dataset are not distributed in the same way. The second residual dataset is then determined as the validation dataset.
[0093] Through steps 301 to 306 above, the dataset construction process is completed. First, the initial protein information obtained from the protein database is divided into multiple initial protein fragments according to the amino acid segmentation granularity. The initial protein fragments and initial compound information are then used to construct an initial dataset. The initial dataset is then divided into a training dataset, a first test dataset, a second test dataset, and a validation dataset. This not only provides the necessary data for subsequent model training, but also ensures the accuracy of the test results by dividing the first test dataset into one with the same distribution as the training dataset and the second test dataset into one with a different distribution.
[0094] In some embodiments, such as Figure 5C As shown, after step 203, the performance metrics of the trained graph neural network obtained in step 203 can be evaluated through steps 204 to 207 below. If the preset evaluation threshold is not reached, the graph neural network model continues to be trained. The following combines... Figure 5C Each step is explained.
[0095] In step 204, the trained graph neural network model is used to perform prediction processing on the verification pair data in the verification dataset to obtain the verification prediction energy difference of the verification pair data.
[0096] The implementation process of this step is similar to that of step 104 above. In practical applications, the implementation process of step 104 can be referred to.
[0097] In step 205, the evaluation index value of the trained graph neural network model is determined based on the verification prediction energy difference and the label information of the verification pairing data.
[0098] In some embodiments, the difference between the validation prediction energy difference corresponding to each validation pair data and the difference between the label information of each validation pair data can be determined first, and then the average difference value can be determined based on each difference value, and the average difference value can be determined as the evaluation index value.
[0099] In some embodiments, the evaluation metrics for a graph neural network model can also be accuracy, error rate, etc.
[0100] In step 206, when the evaluation index value is less than a preset index threshold, new training data is acquired.
[0101] In some embodiments, obtaining new training data can be done by re-acquiring it from the partitioned training dataset, or by obtaining initial protein data from a protein database, and then segmenting and pairing the initial protein data to obtain new training data. In this application embodiment, the method of obtaining new training data is not limited.
[0102] In step 207, the graph neural network model is trained again using new training data until the evaluation index value of the trained graph neural network model reaches the index threshold.
[0103] The process of continuing to train the graph neural network model using new training data is similar to the implementation process of steps 202 and 203. After continuing to train the graph neural network model using new training data and obtaining a new trained graph neural network model, the evaluation index value of the newly trained graph neural network model is determined again using the validation dataset. When the evaluation index value is determined to be greater than or equal to the index threshold, it indicates that the prediction performance of the trained graph neural network model has reached the target and there is no need to continue training the graph neural network model.
[0104] Through steps 204 to 207 above, the prediction performance of the trained graph neural network model is verified using the validation dataset. If the evaluation index value of the trained graph neural network model does not reach the index threshold, new training data is obtained to continue training the graph neural network model until a graph neural network model with an evaluation index value reaching the index threshold is obtained, thereby improving the prediction accuracy of the trained graph neural network model.
[0105] In some embodiments, it can be achieved through Figure 5D Steps 401 to 405, as shown, test the trained graph neural network model using the first and second test datasets. The following is a combination of... Figure 5D Each step is explained.
[0106] In step 401, the trained graph neural network model is used to perform prediction processing on the first test pair data in the first test dataset to obtain the first prediction energy difference of the first test pair data.
[0107] The implementation process of step 401 is similar to that of step 104 above. When implementing step 401, you can refer to the implementation process of step 104.
[0108] In step 402, based on the first predicted energy difference and the label information of the first test pair data, the first test result corresponding to the first test dataset is determined.
[0109] In some embodiments, based on the first predicted energy difference and the label information of the first test pairing data, a first correlation coefficient between the first predicted energy difference and the label information is determined, and this first correlation coefficient is used as the first test result. In implementation, the covariance and standard deviation of each first predicted energy difference and its corresponding label information are first determined, and the quotient of the covariance and standard deviation is used as the first correlation coefficient. This first correlation coefficient is a real number between -1 and 1, where a first correlation coefficient greater than 0 indicates a positive correlation between the first predicted energy difference and its corresponding label information; a first correlation coefficient less than 0 indicates a negative correlation between the first predicted energy difference and its corresponding label information. The larger the absolute value of the correlation coefficient, the stronger the correlation.
[0110] In step 403, the trained graph neural network model is used to predict the second test pair data in the second test dataset to obtain the second predicted energy difference of the second test pair data.
[0111] The implementation process of step 403 is similar to that of step 104 above. When implementing step 401, you can refer to the implementation process of step 104.
[0112] In step 404, based on the second predicted energy difference and the label information of the second test pair data, the second test result corresponding to the second test dataset is determined.
[0113] Similar to step 402, based on the second predicted energy difference of the second test pair data and the label information of the second test pair data, a second correlation coefficient between the second predicted energy difference and the label information is determined, and the second correlation coefficient is determined as the second test result corresponding to the second test dataset.
[0114] In step 405, the first test result and the second test result are output.
[0115] In some embodiments, the first test result and the second test result can be displayed directly on the device itself, or the first test result and the second test result can be sent to the terminal device that requested the test.
[0116] By using steps 401 to 405 above, the trained graph neural network model is tested using a first test dataset with the same distribution as the training dataset and a second test dataset with a different distribution than the training dataset. This avoids overfitting of the trained graph neural network, improves the generalization ability of the trained graph neural network model, and ensures that the trained graph neural network model can produce accurate prediction results for various protein classifications.
[0117] This application provides an energy prediction method, applied to... Figure 2 The energy prediction system shown Figure 6 This is a schematic diagram of another implementation flow of the energy prediction method provided in the embodiments of this application. The following is combined with... Figure 6 The energy prediction method provided in the embodiments of this application will be described.
[0118] In step 501, the terminal identifies the protein to be predicted through target identification.
[0119] Here, we will use the energy prediction method provided in this application as an example in the context of new drug development. Since protein malfunctions can lead to diseases in most organ systems, proteins are also the most common drug targets today. The protein to be predicted can be a protein molecule that can affect or control a certain disease, i.e., a target protein. The binding of drug molecules to target proteins can achieve the effect of treating the disease. Target identification is used to identify the target protein of a certain disease. Target identification requires long-term and extensive physiological and biochemical experiments. If a protein has an inhibitory effect on the disease, that protein must be activated and identified as a target protein.
[0120] In step 502, the terminal identifies a lead compound from a variety of candidate compounds.
[0121] To identify a lead compound, clue compounds are first needed. These can be compounds obtained through purification of the active ingredient from a natural drug or existing drug compounds. Compounds with structures similar to the clue compounds are then identified as candidate compounds. When identifying the lead compound from multiple candidate compounds, each candidate compound can be mixed with the protein to be predicted in a test tube solution, and the binding of the candidate compound to the protein is tested. The candidate compound that binds to the protein in a predetermined manner is identified as the lead compound. Alternatively, a computer simulation screening process can be used to identify the lead compound from multiple candidate compounds.
[0122] In step 503, the terminal sends an energy prediction request to the server.
[0123] In some embodiments, the energy prediction request carries the protein to be predicted and multiple lead compounds identified through target site recognition.
[0124] In step 504, the server obtains the protein to be predicted and multiple lead compounds obtained through target identification.
[0125] In some embodiments, the server receives an energy prediction request, parses the energy prediction request, and obtains the protein to be predicted and multiple lead compounds.
[0126] In step 505, the server determines the intermolecular target interaction energy between each lead compound and the protein to be predicted.
[0127] In some embodiments, the server acquires the compound atomic information of each lead compound and the protein atomic information of the protein to be predicted, and determines the intermolecular reference interaction energy between each lead compound and the protein to be predicted based on a preset molecular force field algorithm, the compound atomic information, and the protein atomic information. Then, it uses a trained graph neural network model to perform prediction processing on the compound atomic information and the protein atomic information to obtain the energy differences. Finally, it determines the target intermolecular interaction energy between each lead compound and the protein to be predicted based on the intermolecular reference interaction energy and the corresponding energy differences.
[0128] In step 506, the server determines at least one target compound from the plurality of lead compounds based on the intermolecular target interaction energy between the candidate compounds and the protein to be predicted.
[0129] In some embodiments, the server can determine the lead compound corresponding to the minimum intermolecular target interaction energy as the target compound, or it can determine the lead compound whose intermolecular target interaction energy is less than a preset energy threshold as the target compound.
[0130] In step 507, the server sends an energy prediction response to the terminal.
[0131] The energy prediction response carries at least one target compound.
[0132] In step 508, the terminal obtains at least one target compound and uses the target compound for preclinical studies.
[0133] In some embodiments, preclinical pharmacokinetic studies can be conducted using the target compound to quantitatively investigate its absorption, distribution, metabolism, and excretion in vivo, and to determine the changes in blood drug concentration over time, thereby determining the stability of the target compound in vivo. Additionally, pharmacological and toxicological studies are required to determine the efficacy and safety of the target compound, thus ensuring the safety of clinical trials. After completing preclinical studies and confirming that the conditions for clinical trials are met, clinical trials are conducted.
[0134] In the energy prediction method provided in this application embodiment, after the terminal identifies the protein to be predicted and multiple lead compounds, it sends an energy prediction request to the server, requesting prediction of the target intermolecular interaction energies between each lead compound and the protein to be predicted. The server first uses a computationally efficient molecular force field algorithm to determine the reference intermolecular interaction energy between the compound to be predicted and the protein to be predicted. Then, it uses a trained graph neural network model to predict the atomic information of the compound and the atomic information of the protein to obtain the energy difference. Finally, based on the reference intermolecular interaction energy and the energy difference, it determines the target intermolecular interaction energy between the compound to be predicted and the protein to be predicted. The target intermolecular interaction energy obtained at this time is equivalent to the interaction energy determined using quantum mechanics. Therefore, it can improve computational efficiency while ensuring higher accuracy in energy prediction results. The server determines the target compound based on each target intermolecular interaction energy and returns the energy prediction response carrying the target compound to the terminal, thereby enabling the terminal to use the target compound for preclinical analysis and clinical experiments. Since the efficiency of the server in determining the target compound is greatly improved, the efficiency of drug development is improved and the cost of drug development is reduced.
[0135] The following will describe an exemplary application of the embodiments of this application in a real-world application scenario.
[0136] To address the problem of slow calculation speed in existing methods for intermolecular interaction energy, this application provides an algorithm for predicting intermolecular interaction energy based on an isovariant graph neural network.
[0137] The energy prediction method provided in this application includes three steps: dataset construction, model training, and model prediction, which are described in detail below.
[0138] In the dataset construction step, a training dataset needs to be built. In some embodiments, the training dataset can be built based on the FMO database. The FMO database is a database obtained by performing FMO calculations using quantum mechanics software, and the data in it can be divided into seven categories, including X-ray All Entries and COVID-19. In this embodiment, data from X-ray All Entries is selected to construct the dataset required for the energy prediction method provided in this embodiment. When constructing the dataset, the data in X-ray All Entries needs to be cut and divided.
[0139] Data Segmentation: The X-ray All Entries dataset currently contains 3669 PDB data points. For each PDB, preprocessing is performed by segmenting the proteins within the PDB according to amino acid granularity and adding hydrogen atoms at both ends to form a fragment. Each fragment forms a data pair with a small molecule within the PDB. Each data segment contains atomic type information and atomic 3D coordinate information for both the protein fragment and the small molecule. Additionally, the formal charge between the protein fragment and the small molecule is calculated using RDKit software. Finally, the (atomic type, atomic 3D coordinates, formal charge) from each data pair is used as input data. After segmenting all the PDB data, nearly 200,000 data points are generated.
[0140] Data partitioning: The 200,000 data points obtained after data splitting are divided into training, validation, and test sets for the DeepFMO algorithm. After partitioning, the training set contains 170,000 data points, and the validation set contains 10,000 data points. In this embodiment, two test sets are constructed: the first test set has an identical distribution to the training set (identically distributed) and contains 5,000 data points; the second test set has an out-of-distribution distribution and contains 15,000 data points. Using these two test sets helps prevent model overfitting, thereby better validating the performance of the DeepFMO algorithm.
[0141] The energy prediction method provided in this application embodiment can utilize... Figure 7 The DeepFMO model implementation shown is as follows: Figure 7As shown, the model includes a molecular force field module 701 and an equivariant neural network model 702. The molecular force field module 701 is used to determine the interaction energy E_xTB-IFF between the protein fragment and the small molecule. The equivariant neural network model 702 is used to predict the difference between the intermolecular interaction energy E_FMO calculated by the FMO method and E_xTB-IFF determined by the molecular force field module 701, denoted as E_ML. Finally, the sum of E_ML and E_xTB-IFF (E_ML + E_xTB-IFF) is used as the predicted value of E_FMO.
[0142] During model training, the model loss is determined based on the E_ML prediction values obtained from forward computation and the true labels. The model loss is then used for back-forward computation to determine the gradient of each network layer, and the Adaptive Moment Estimation (Adam) algorithm is used to update the neural network model parameters. In this embodiment, the loss function used during model training is the Mean Squared Error (MSE), and the model's training hyperparameters are shown in Table 1.
[0143] Table 1. Hyperparameters for DeepFMO algorithm training Layer Number 8 Learning rate 0.002 Batch size 32 Weight decay 0.000001 Epochs 100 Dropout 0.1 During the model prediction process, the trained DeepFMO model can be used to predict new intermolecular interaction energies. Information such as the three-dimensional coordinates of protein fragments and small molecules is input into the trained DeepFMO model. Specifically, E_xTB-IFF is determined using the molecular force field module 701, and E_ML is predicted using the isovariant neural network model 702. The sum of the two yields the final predicted value E_FMO(E_xTB-IFF+E_ML).
[0144] In the new drug development process, calculating the binding affinity between proteins and small molecules requires calculating the intermolecular interaction energy. The existing computational method, FMO, is a quantum mechanics-based method with high accuracy, but it is computationally intensive, making the process very time-consuming and unsuitable for large-scale calculations or large-scale screening of drug molecules.
[0145] The goal of the energy prediction method provided in this application is to significantly improve the calculation speed of intermolecular interaction energies and reduce computation time, while maintaining prediction accuracy similar to that of the FMO calculation method. In the future, it can be deployed on a cloud computing platform for large-scale calculations of small molecules to complete drug molecule screening. To better verify the effectiveness of this invention, the DeepFMO algorithm was tested for both accuracy and speed.
[0146] To assess algorithm accuracy, two test sets were constructed: one with the same distribution as the training set and the other with a different distribution. The DeepFMO algorithm was then tested. The test results are as follows: Figure 8 and Figure 9 As shown, in Figure 8 and Figure 9 The horizontal axis represents the predicted values of the DeepFMO algorithm, and the vertical axis represents the label values of the FMO calculation method. The correlation between the two is calculated to evaluate the accuracy of the DeepFMO algorithm. On the same-distribution test set, the correlation between the predicted values of the DeepFMO algorithm and the label values of the FMO calculation method is Pearson R = 0.9947, which is very high. On test sets with different distributions, the correlation between the predicted values of the DeepFMO algorithm and the label values of the FMO calculation method is Pearson R = 0.9890, slightly lower than the result on the same-distribution test set, which is expected, but the correlation value is still very high. Overall, the accuracy of the DeepFMO algorithm on both test sets is similar to that of the FMO calculation method.
[0147] In terms of algorithm speed, the computation time of the DeepFMO algorithm and the FMO calculation method were compared. The calculations were performed on the same CPU server, processing multiple datasets, and the average computation time was calculated. The FMO calculation method used the GAMESS software. The results are shown in Table 2. As can be seen from the results in Table 2, the DeepFMO calculation speed significantly exceeds that of the traditional FMO calculation method, with a speedup of 3528 times.
[0148] Table 2 Comparison of computation time between DeepFMO algorithm and traditional FMO calculation method method Calculation time (seconds) Traditional FMO calculation methods 88200 DeepFMO algorithm 25 acceleration ratio 3528 As shown in Tables 1 and 2, the DeepFMO algorithm has similar computational accuracy to the traditional FMO calculation method, while its computational speed can significantly exceed that of the traditional FMO calculation method.
[0149] The following description continues to illustrate the exemplary structure of the energy prediction device 455 provided in the embodiments of this application as a software module. In some embodiments, such as... Figure 3 As shown, the software module stored in the energy prediction device 455 in the memory 450 may include: The first acquisition module 4551 is used to acquire the compound atomic information of the compound to be predicted and the protein atomic information of the protein to be predicted. The first determination model 4552 determines the intermolecular reference interaction energy of the compound to be predicted and the protein to be predicted based on a preset molecular force field algorithm, the atomic information of the compound, and the atomic information of the protein. The second acquisition module 4553 is used to acquire the trained graph neural network model; The first prediction module 4554 is used to predict the atomic information of the compound and the atomic information of the protein using the trained graph neural network model to obtain an energy difference value. The energy difference value is the difference between the intermolecular interaction energy determined based on quantum mechanics and the intermolecular interaction energy determined based on the molecular force field. The second determining module 4555 is used to determine the target intermolecular interaction energy between the compound to be predicted and the protein to be predicted based on the intermolecular reference interaction energy and the energy difference.
[0150] In some embodiments, the device further includes: The third acquisition module is used to acquire a preset graph neural network model and acquire a training dataset. The training dataset includes multiple training pair data and label information for each training pair data. The training pair data includes atomic information of training compounds, atomic information of training protein fragments, and formal charge between training compounds and training protein fragments. The second prediction module is used to perform prediction processing on each training pair data using the preset graph neural network model to obtain the prediction energy difference corresponding to each training pair data. The first training module is used to train the preset graph neural network model by using the label information of each training pair data and the difference in predicted energy corresponding to each training pair data, so as to obtain a trained graph neural network model.
[0151] In some embodiments, the device further includes: The fourth acquisition module is used to acquire multiple protein data from a protein database, wherein the protein data includes initial protein information and initial compound information corresponding to the initial protein information; The first segmentation module is used to segment each initial protein information to obtain multiple initial protein fragments and determine the formal charge between each initial protein fragment and the initial compound information. The first pairing module is used to pair each initial protein fragment with the initial compound information to obtain multiple initial pairing data. The third determining module is used to determine the label information of each initial pairing data; The data construction module is used to construct an initial dataset using the initial pairing data and the label information of the initial pairing data; The second segmentation module is used to segment the initial dataset to obtain a training dataset, a first test dataset, a second test dataset, and a validation dataset.
[0152] In some embodiments, the third determining module is further configured to: The first intermolecular interaction energy of each initial protein fragment and the initial compound information is determined based on the molecular force field algorithm; The second intermolecular interaction energy of each initial protein fragment and the initial compound information is determined based on a quantum mechanical algorithm. The energy difference between each second intermolecular interaction energy and the corresponding first intermolecular interaction energy is determined as the tag information of each initial pairing data.
[0153] In some embodiments, the second segmentation module is further configured to: According to a preset training data ratio, a training dataset is randomly segmented from the initial dataset to obtain the first remaining dataset; Based on the classification information of each protein in the training dataset, a first test dataset and a second test dataset are segmented from the first remaining dataset to obtain a second remaining dataset; wherein, the classification information of the proteins in the first test dataset exists in the training dataset, and the classification information of the proteins in the second test dataset does not exist in the training dataset. The second remaining dataset is selected as the validation dataset.
[0154] In some embodiments, the device further includes: The third prediction module is used to perform prediction processing on the verification pair data in the verification dataset using the trained graph neural network model to obtain the verification prediction energy difference of the verification pair data. The fourth determining module is used to determine the performance evaluation value of the trained graph neural network model based on the verification prediction energy difference and the label information of the verification pairing data. The fifth acquisition module is used to acquire new training data when the performance evaluation value is less than a preset evaluation threshold. The second training module is used to continue training the graph neural network model using new training data until the performance evaluation value of the trained graph neural network model reaches the evaluation threshold.
[0155] In some embodiments, the device further includes: The fourth prediction module is used to perform prediction processing on the first test pair data in the first test dataset using the trained graph neural network model to obtain the first prediction energy difference of the first test pair data. The fifth determining module is used to determine the first test result corresponding to the first test dataset based on the first predicted energy difference and the label information of the first test pairing data; The fifth prediction module is used to perform prediction processing on the second test pair data in the second test dataset using the trained graph neural network model to obtain the second predicted energy difference value of the second test pair data. The sixth determining module is used to determine the second test result corresponding to the second test dataset based on the second predicted energy difference and the label information of the second test pairing data; The output module is used to output the first test result and the second test result.
[0156] In some embodiments, the device further includes: The third segmentation module is used to segment the protein to be predicted based on the protein atom information of the protein to be predicted, and obtain multiple protein fragments to be predicted. The seventh determining module is used to determine the formal charge between each protein fragment to be predicted and the compound to be predicted; The second pairing module is used to pair each of the protein fragments to be predicted with the compounds to be predicted, thereby obtaining multiple pairs of prediction data.
[0157] In some embodiments, the first determining module is further configured to: Based on a preset molecular force field algorithm and each pairing data to be predicted, the intermolecular reference interaction energy between each protein fragment to be predicted and the compound to be predicted is determined. The intermolecular reference interaction energies between each protein fragment to be predicted and the compound to be predicted are summed to obtain the intermolecular reference interaction energies between the compound to be predicted and the protein to be predicted.
[0158] In some embodiments, the first prediction module is further configured to: The trained graph neural network model is used to predict each protein fragment and compound to be predicted, and the fragment energy difference between each protein fragment and compound to be predicted is obtained. The energy differences between the fragments of the protein to be predicted and the compound to be predicted are summed to obtain the energy difference value.
[0159] In some embodiments, the device further includes: The sixth acquisition module is used to acquire the protein to be predicted and multiple lead compounds obtained through target identification; The eighth determining module is used to determine the intermolecular target interaction energy between each lead compound and the protein to be predicted; The ninth determining module is used to determine at least one target compound from the plurality of lead compounds based on the intermolecular target interaction energy between the lead compounds and the protein to be predicted.
[0160] This application provides a computer program product including computer-executable instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer-executable instructions from the computer-readable storage medium and executes the instructions, causing the electronic device to perform the energy prediction method described in this application.
[0161] This application provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are executed by a processor, they cause the processor to execute the energy prediction method provided in this application, for example, as shown in the figure. Figure 4A , Figure 6 The energy prediction method is shown.
[0162] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disk, or CD-ROM; or it may be a variety of devices including one or any combination of the above-mentioned memories.
[0163] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.
[0164] As an example, computer-executable instructions may, but do not necessarily, correspond to files in a file system. They may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., a file that stores one or more modules, subroutines, or code sections).
[0165] As an example, computer-executable instructions can be deployed to execute on a single electronic device, or on multiple electronic devices located at one location, or on multiple electronic devices distributed across multiple locations and interconnected via a communication network.
[0166] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of this application are included within the scope of protection of this application.
Claims
1. An energy prediction method, characterized in that, The method includes: Obtain the compound atomic information of the compound to be predicted and the protein atomic information of the protein to be predicted; The intermolecular reference interaction energy of the compound to be predicted and the protein to be predicted is determined based on a preset molecular force field algorithm, the atomic information of the compound, and the atomic information of the protein. Obtain the trained graph neural network model; The trained graph neural network model is used to predict the atomic information of the compound and the atomic information of the protein to obtain the energy difference value, which is the difference between the intermolecular interaction energy determined based on quantum mechanics and the intermolecular interaction energy determined based on the molecular force field. The target intermolecular interaction energy between the compound to be predicted and the protein to be predicted is determined based on the intermolecular reference interaction energy and the energy difference.
2. The method according to claim 1, characterized in that, The method further includes: Obtain a preset graph neural network model and a training dataset. The training dataset includes multiple training pairs and label information for each training pair. The training pairs include atomic information of training compounds, atomic information of training protein fragments, and formal charges between training compounds and training protein fragments. The preset graph neural network model is used to perform prediction processing on each training pair data to obtain the prediction energy difference corresponding to each training pair data. The preset graph neural network model is trained using the label information of each training pair and the difference in predicted energy corresponding to each training pair, to obtain a trained graph neural network model.
3. The method according to claim 2, characterized in that, The method further includes: Multiple protein data entries are obtained from a protein database, including initial protein information and initial compound information corresponding to the initial protein information. Each initial protein information is segmented to obtain multiple initial protein fragments, and the formal charge between each initial protein fragment and the initial compound information is determined. Each initial protein fragment and the initial compound information are paired to obtain multiple initial pairing data; Determine the label information for each initial pairing; An initial dataset is constructed using the initial pairing data and the label information of each initial pairing data. The initial dataset is split to obtain a training dataset, a first test dataset, a second test dataset, and a validation dataset.
4. The method according to claim 3, characterized in that, The process of determining the label information for each initial pairing includes: The first intermolecular interaction energy of each initial protein fragment and the initial compound information is determined based on the molecular force field algorithm; The second intermolecular interaction energy of each initial protein fragment and the initial compound information is determined based on a quantum mechanical algorithm. The energy difference between each second intermolecular interaction energy and the corresponding first intermolecular interaction energy is determined as the tag information of each initial pairing data.
5. The method according to claim 3, characterized in that, The initial dataset is segmented to obtain a training dataset, a first test dataset, a second test dataset, and a validation dataset, including: According to a preset training data ratio, a training dataset is randomly segmented from the initial dataset to obtain the first remaining dataset; Based on the classification information of each protein in the training dataset, a first test dataset and a second test dataset are segmented from the first remaining dataset to obtain a second remaining dataset; wherein, the classification information of the proteins in the first test dataset exists in the training dataset, and the classification information of the proteins in the second test dataset does not exist in the training dataset. The second remaining dataset is selected as the validation dataset.
6. The method according to any one of claims 3 to 5, characterized in that, The method further includes: The trained graph neural network model is used to predict the validation pair data in the validation dataset to obtain the validation prediction energy difference of the validation pair data. Based on the verification prediction energy difference and the label information of the verification pairing data, the performance evaluation value of the trained graph neural network model is determined; When the performance evaluation value is less than the preset evaluation threshold, new training data is obtained; The graph neural network model is trained again using new training data until the performance evaluation value of the trained graph neural network model reaches the evaluation threshold.
7. The method according to any one of claims 3 to 5, characterized in that, The method further includes: The trained graph neural network model is used to predict the first test pair data in the first test dataset to obtain the first predicted energy difference of the first test pair data. Based on the first predicted energy difference and the label information of the first test pair data, the first test result corresponding to the first test dataset is determined; The trained graph neural network model is used to predict the second test pairing data in the second test dataset to obtain the second predicted energy difference of the second test pairing data. Based on the second predicted energy difference and the label information of the second test pair data, the second test result corresponding to the second test dataset is determined; Output the first test result and the second test result.
8. The method according to any one of claims 1 to 5, characterized in that, The method further includes: Based on the protein atom information of the protein to be predicted, the protein to be predicted is segmented to obtain multiple protein fragments to be predicted. Determine the formal charge between each protein fragment to be predicted and the compound to be predicted; Each of the protein fragments to be predicted and the compounds to be predicted are paired to obtain multiple prediction pairing data.
9. The method according to claim 8, characterized in that, The determination of the intermolecular reference interaction energy between the compound to be predicted and the protein to be predicted, based on a preset molecular force field algorithm, the compound's atomic information, and the protein's atomic information, includes: Based on a preset molecular force field algorithm and each pairing data to be predicted, the intermolecular reference interaction energy between each protein fragment to be predicted and the compound to be predicted is determined. The intermolecular reference interaction energies between each protein fragment to be predicted and the compound to be predicted are summed to obtain the intermolecular reference interaction energies between the compound to be predicted and the protein to be predicted.
10. The method according to claim 8, characterized in that, The step of using the trained graph neural network model to predict the atomic information of the compound and the atomic information of the protein to obtain the energy difference includes: The trained graph neural network model is used to predict each protein fragment and compound to be predicted, and the fragment energy difference between each protein fragment and compound to be predicted is obtained. The energy differences between the fragments of the protein to be predicted and the compound to be predicted are summed to obtain the energy difference value.
11. The method according to any one of claims 1 to 5, characterized in that, The method further includes: Acquire the target protein and multiple lead compounds obtained through target identification; Determine the target intermolecular interaction energies between each lead compound and the protein to be predicted; Based on the intermolecular target interaction energy between the lead compounds and the protein to be predicted, at least one target compound is determined from the plurality of lead compounds.
12. An energy prediction device, characterized in that, The device includes: The first acquisition module is used to acquire the compound atomic information of the compound to be predicted and the protein atomic information of the protein to be predicted. The first determination model determines the intermolecular reference interaction energy of the compound to be predicted and the protein to be predicted based on a preset molecular force field algorithm, the atomic information of the compound, and the atomic information of the protein. The second acquisition module is used to acquire the trained graph neural network model; The first prediction module is used to predict the atomic information of the compound and the atomic information of the protein using the trained graph neural network model to obtain an energy difference value. The energy difference value is the difference between the intermolecular interaction energy determined based on quantum mechanics and the intermolecular interaction energy determined based on the molecular force field. The second determining module is used to determine the target intermolecular interaction energy between the compound to be predicted and the protein to be predicted based on the intermolecular reference interaction energy and the energy difference.
13. An electronic device, characterized in that, The electronic device includes: Memory is used to store executable instructions for a computer; A processor, when executing computer-executable instructions stored in the memory, implements the energy prediction method according to any one of claims 1 to 11.
14. A computer-readable storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed by a processor, they implement the energy prediction method according to any one of claims 1 to 11.
15. A computer program product comprising computer-executable instructions, characterized in that, When the computer-executable instructions are executed by a processor, they implement the energy prediction method according to any one of claims 1 to 11.
Citation Information
Patent Citations
Method for predicting interaction and affinity of compound and protein based on end-to-end learning
CN113744799A
Drug and protein affinity prediction method and system based on graph neural network
CN115631805A
