Lead compound optimization method, apparatus, electronic device, and storage medium

Abstract

The application relates to a lead compound optimization method, device, electronic device and storage medium, wherein the method comprises the following steps: inputting protein small molecule mask training pair data into a lead compound optimization model to be trained; calling a plurality of network layers to predict training sample data of a lead compound, so as to obtain information of newly generated atoms matched with a protein pocket; based on the information of the newly generated atoms, a loss value of the lead compound optimization model to be trained is calculated; the lead compound optimization model to be trained whose loss value is within a preset threshold range is determined as a target lead compound optimization model; inputting protein pocket data to be optimized and molecular fragment data to be optimized into the target lead compound optimization model, so as to obtain an optimized lead compound. The application solves the problem that a traditional lead compound optimization method is limited to an active compound space expressed by a compound library prepared in advance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of neural networks, and in particular to a lead compound optimization method and device, an electronic device, and a storage medium. BACKGROUND

[0002] Drug design is a huge consuming process. The emergence of computer-aided drug design (CADD) brings drug discovery into the stage of rational design. Generally speaking, a drug goes through three stages of hit, lead, and drug candidate before it is finally put into clinical trials. As the starting point of drug design, the activity of hit compounds is not high. However, the method of lead compound optimization can transform hit compounds into lead compounds with better drug properties or even candidate drugs. The methods of lead compound optimization mainly include linker design, fragment replacement, skeleton transition, and side chain optimization.

[0003] In the existing CCAD method, the strategy of lead compound optimization is mainly realized by the method of library search. First-in-class drugs, that is, new drugs, refer to drugs that use a completely new and unique mechanism of action to treat a certain disease; compared with the high failure rate of first-in-class drugs, me-too drugs (specifically drugs with their own intellectual property rights) do not need to confirm the target because the target has been proven to be drugable, so the risk of developing me-too drugs is relatively low; however, the computer-aided drug design method for me-too drug design has not been widely applied, and the development of me-too drugs is still based on the experience and intuition of chemists. The traditional CADD lead compound optimization method is limited by the rough definition of the similarity between compounds and the limited active compound space expressed by the prepared compound library.

[0004] In view of the problem in the related art that the traditional lead compound optimization method is limited by the active compound space expressed by the prepared compound library, there is currently no effective solution. SUMMARY

[0005] In this embodiment, a lead compound optimization method, device, electronic device, and storage medium are provided to solve the problem in the related art that the traditional lead compound optimization method is limited by the active compound space expressed by the prepared compound library.

[0006] In a first aspect, a lead compound optimization method is provided in this embodiment, and the method comprises:

[0007] inputting preset training sample data of a lead compound into a lead compound optimization model to be trained, the lead compound optimization model to be trained comprising a plurality of network layers, the training sample data of the lead compound comprising protein pocket data and protein-small molecule mask training pair data;

[0008] calling the plurality of network layers to predict the training sample data of the lead compound, to obtain information of newly generated atoms matched with the protein pocket;

[0009] based on the information of the newly generated atoms, calculating a loss value of the lead compound optimization model to be trained;

[0010] when the loss value is within a preset threshold range, determining that the lead compound optimization model to be trained is a target lead compound optimization model;

[0011] inputting protein pocket data to be optimized and molecular fragment data to be optimized into the target lead compound optimization model, to generate a molecular structure corresponding to the protein pocket data to be optimized;

[0012] generating an optimized lead compound according to the molecular structure and the protein pocket data to be optimized.

[0013] In some embodiments, before the inputting of the preset training sample data of the lead compound into the lead compound optimization model to be trained, the method comprises:

[0014] masking protein structure information files of the protein pocket by using a self-supervised learning strategy and a task strategy of drug design, to obtain protein-small molecule mask training pair data.

[0015] In some embodiments, the calling of the plurality of network layers to predict the training sample data of the lead compound, to obtain information of newly generated atoms matched with the protein pocket, comprises:

[0016] extracting topological surface features in the protein pocket based on an interaction feature extraction network in the network layers;

[0017] processing the topological surface features by a protein surface grid method, to obtain protein surface feature vectors;

[0018] predicting the protein surface feature vectors based on an atom component generation network layer in the network layers, to predict the information of newly generated atoms matched with the protein pocket.

[0019] In some embodiments, when the loss value is within the preset threshold range, the determination that the lead compound optimization model to be trained is the target lead compound optimization model further comprises:

[0020] When the loss value is not within the preset threshold range, a self-supervised learning strategy and a drug design task strategy are adopted to perform mask processing on the protein structure information file of the protein pocket again to obtain expanded training sample data;

[0021] The expanded training sample data is input into the to-be-trained lead compound optimization model for training until the loss value is within the preset range.

[0022] In some embodiments, the inputting of the to-be-optimized protein pocket data and the to-be-optimized molecular fragment data into the target lead compound optimization model to generate a molecular structure corresponding to the to-be-optimized protein pocket comprises:

[0023] A plurality of network layers are called to perform atomic processing on the to-be-optimized protein pocket data and the to-be-optimized molecular fragment data to obtain information of a predicted newly generated atom matched with the to-be-optimized protein pocket;

[0024] Based on the information of the predicted newly generated atom, a three-dimensional molecular structure corresponding to the to-be-optimized protein pocket is generated, and the three-dimensional molecular structure is used to bind with the to-be-optimized protein pocket to obtain an optimized lead compound.

[0025] In some embodiments, when the loss value is within the preset threshold range, the to-be-trained lead compound optimization model is determined as a target lead compound optimization model, which comprises:

[0026] A ligand atom within a preset distance threshold range from the protein pocket is obtained;

[0027] Starting from the ligand atom, each atom in the protein pocket molecule is traversed, and a target training trajectory is generated;

[0028] When the coincidence degree of the newly generated atom trajectory and the target training trajectory is within a preset range, the to-be-trained lead compound optimization model is determined as a target lead compound optimization model.

[0029] In some embodiments, the information of the newly generated atom comprises an atom growth point, an atom position, an atom type, and an atom bonding relationship of the newly generated atom.

[0030] In a second aspect, a lead compound optimization device is provided in the present embodiment, and the device comprises a training module and an optimization module.

[0031] The training module is configured to input preset training sample data of a lead compound into a lead compound optimization model to be trained, the lead compound optimization model to be trained comprises a plurality of network layers, the training sample data of the lead compound comprises protein pocket data and protein-small-molecule mask training pair data; the plurality of network layers are called to predict the training sample data of the lead compound to obtain information of newly generated atoms matched with the protein pocket; a loss value of the lead compound optimization model to be trained is calculated based on the information of the newly generated atoms; and when the loss value is within a preset threshold range, the lead compound optimization model to be trained is determined as a target lead compound optimization model.

[0032] The optimization module is configured to input protein pocket data to be optimized and molecular fragment data to be optimized into the target lead compound optimization model to generate a molecular structure corresponding to the protein pocket data to be optimized; and generate an optimized lead compound based on the molecular structure and the protein pocket data to be optimized.

[0033] In a third aspect, an electronic device is provided in the embodiment, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the lead compound optimization method of the first aspect when executing the computer program.

[0034] In a fourth aspect, a storage medium is provided in the embodiment, which stores a computer program executable by a processor to implement the lead compound optimization method of the first aspect.

[0035] Compared with related technologies, the lead compound optimization method, device, electronic device, and storage medium provided in the embodiment, by performing mask on a molecule to obtain a protein-small-molecule mask training pair, inputting the protein-small-molecule mask training pair data as training sample data into a lead compound optimization model to be trained to train, predicting information of a plurality of newly generated atoms matched with a protein pocket, and calculating a loss value of the training model, when the loss value is within a preset threshold range, determining the training model as a target model; and then inputting protein pocket data to be optimized and molecular data to be optimized into the target model to obtain a molecular structure corresponding to the protein pocket to be optimized, and then combining the protein pocket to be optimized to obtain an optimized lead compound, the method of performing mask on a molecule to obtain a protein-small-molecule mask training pair and then predicting new atomic information is beneficial to improve the matching degree of a protein pocket and an atom, and is not limited to the active compound space expressed by an existing compound library.

[0036] The details of one or more embodiments of the present application are presented in the following drawings and description to make other features, objects and advantages of the present application more apparent. BRIEF DESCRIPTION OF DRAWINGS

[0037] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0038] Figure 1 is a hardware structure block diagram of the end of the lead compound optimization method of the present embodiment;

[0039] Figure 2 is a flowchart of the lead compound optimization method of the present embodiment;

[0040] Figure 3 is a flowchart of the lead compound optimization method of the present embodiment;

[0041] Figure 4 is a schematic diagram of the prediction atom generation of the present embodiment;

[0042] Figure 5A is a schematic diagram of the original molecule of the model inference of the present embodiment;

[0043] Figure 5B is a state diagram of the original molecule of the model inference of the present embodiment;

[0044] Figure 5C is a schematic diagram of the analysis molecule of the model inference of the present embodiment;

[0045] Figure 5D is a schematic diagram of the molecule to be optimized of the model inference of the present embodiment;

[0046] Figure 5E1 is a schematic diagram of the target molecule of the model inference of the present embodiment;

[0047] Figure 5E2 is a schematic diagram of another target molecule of the model inference of the present embodiment;

[0048] Figure 5E3 is a schematic diagram of another target molecule of the model inference of the present embodiment;

[0049] Figure 5E4 is a schematic diagram of another target molecule of the model inference of the present embodiment;

[0050] Figure 5F is a schematic diagram of the performance comparison of the model inference of the present embodiment;

[0051] Figure 5G is a schematic diagram of another performance comparison of the model inference of the present embodiment;

[0052] Figure 6is a structural block diagram of a lead compound optimization device of the embodiment. DETAILED DESCRIPTION

[0053] In order to more clearly understand the objects, technical solutions and advantages of the present application, the present application will be described and illustrated below in conjunction with the accompanying drawings and embodiments.

[0054] Unless otherwise defined, technical terms or scientific terms used in the present application shall have the general meaning understood by one of ordinary skill in the art to which the present application belongs. In the present application, "one", "a", "an", "the", "these" and similar words do not represent a quantitative limitation, and they can be singular or plural. In the present application, the terms "include", "contain", "have" and any variants thereof are intended to cover non-exclusive inclusion; for example, a process, method and system, product or device containing a series of steps or modules (units) are not limited to the listed steps or modules (units), but can include steps or modules (units) not listed, or can include other steps or modules (units) inherent to the process, method, product or device. In the present application, the terms "connected", "connected", "coupled" and the like do not limit to physical or mechanical connection, but can include electrical connection, whether direct or indirect. In the present application, "multiple" means two or more. The association between the associated objects described by "and / or" can represent three relationships, for example, "A and / or B" can represent three cases: A exists alone, A and B exist together, and B exists alone. Generally, the character " / " represents an "or" relationship between the objects before and after. In the present application, the terms "first", "second", "third" and the like are only used to distinguish similar objects, and do not represent a specific order of the objects.

[0055] The method embodiments provided in the present embodiment can be executed in a terminal, a computer or a similar computing device. For example, the method embodiments are executed on a terminal, Figure 1 is a hardware structural block diagram of a terminal of a lead compound optimization method of the embodiment. As shown in Figure 1 , the terminal can include one or more (only one is shown in Figure 1 ) processor 102 and memory 104 for storing data, wherein the processor 102 can include but not limited to processing devices such as microprocessor MCU or programmable logic device FPGA. The above terminal can also include transmission device 106 for communication function and input / output device 108. Those skilled in the art can understand that Figure 1 The structure shown is only schematic, which does not limit the structure of the above terminal. For example, the terminal can include more or less components than those shown in Figure 1 , or have a structure different from Figure 1The different configurations are shown.

[0056] The memory 104 can be used to store computer programs, such as software programs of application software and modules, such as the computer program corresponding to the lead compound optimization method in the present embodiment. The processor 102 executes various functional applications and data processing, i.e., implements the method described above, by running the computer program stored in the memory 104. The memory 104 can include a high-speed random access memory and can also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory 104 can further include memories remotely disposed relative to the processor 102, which can be connected to the terminal through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0057] The transmission device 106 is used to receive or send data via a network. The network includes a wireless network provided by a communication provider of the terminal. In one example, the transmission device 106 includes a network interface controller (NIC) that can be connected to other network devices through a base station to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module used to communicate with the Internet in a wireless manner.

[0058] In the present embodiment, a lead compound optimization method is provided, Figure 2 is a flowchart of the lead compound optimization method of the present embodiment, as shown in Figure 2 The flowchart includes the following steps:

[0059] In step S210, the preset training sample data of the lead compound is input into a lead compound optimization model to be trained. The lead compound model to be trained includes a plurality of network layers, and the training sample data of the lead compound includes protein pocket data and protein-small molecule mask training pair data.

[0060] Specifically, the processor inputs the preset training sample data of the lead compound composed of the protein pocket data and the protein-small molecule mask training pair data into the lead compound optimization model to be trained, and specifically inputs into the network layer of the lead compound model to be trained. The protein-small molecule mask training pair data is obtained according to three enhanced training mask strategies and four task-dependent mask strategies.

[0061] In step S220, a plurality of network layers are called to predict the training sample data of the lead compound to obtain information of newly generated atoms matched with the protein pocket.

[0062] Specifically, the processor calls a plurality of recurrently connected network layers to perform atom-by-atom prediction on the training sample data of the lead compound, and then obtains information of a newly generated atom matched with the protein pocket.

[0063] In step S230, a loss value of the lead compound optimization model to be trained is calculated based on the information of the newly generated atom.

[0064] Specifically, the processor calculates the loss value of the lead compound optimization model to be trained based on the information of the newly generated atom and the information of the last newly generated atom.

[0065] In step S240, when the loss value is within a preset threshold range, the lead compound optimization model to be trained is determined as a target lead compound optimization model.

[0066] Specifically, when the processor determines that the loss value of the lead compound optimization model to be trained is within the preset threshold range, the lead compound optimization model to be trained is determined as the target lead compound optimization model.

[0067] In step S250, the protein pocket data to be optimized and the molecular fragment data to be optimized are input into the target lead compound optimization model to generate a molecular structure corresponding to the protein pocket data to be optimized.

[0068] Specifically, after obtaining the target lead compound optimization model, the protein pocket data to be optimized and the molecular fragment data to be optimized are input into the target lead compound optimization model to generate a molecular structure corresponding to the protein pocket to be optimized.

[0069] In step S260, an optimized lead compound is generated according to the molecular structure and the protein pocket data to be optimized.

[0070] Specifically, after generating the molecular structure corresponding to the protein pocket to be optimized, an optimized lead compound is generated according to the molecular structure and the protein pocket data to be optimized.

[0071] By the above steps, the protein small molecule mask training pair is obtained by masking the molecules of the protein pocket, and the protein small molecule mask training pair data is input as training sample data into the to-be-trained lead compound optimization model for training, information of a plurality of newly generated atoms matched with the protein pocket is predicted, and a loss value of the to-be-trained model is calculated. When the loss value is within a preset threshold range, the to-be-trained model is determined as a target model; then the to-be-optimized protein pocket data and the to-be-optimized molecule data are input into the target model to obtain a molecular structure corresponding to the to-be-optimized protein pocket, which is then combined with the to-be-optimized protein pocket to obtain an optimized lead compound. The method of obtaining the protein small molecule mask training pair by masking the molecules, and then predicting the new atom information, is beneficial to improving the matching degree of the protein pocket and the atom, and further improves the activity of the lead compound without being limited by the active compound space expressed by the existing compound library.

[0072] In some embodiments, step S201 is included before step S210.

[0073] In step S201, the protein structure information file of the protein pocket is masked by using a self-supervised learning strategy and a task strategy of drug design to obtain protein small molecule mask training pair data.

[0074] Specifically, by the above steps, three enhanced training mask strategies and four task-dependent mask strategies are introduced. The first three are random mask, spatial mask and topological mask, which are used as a general self-supervised learning strategy. The last four strategies are fragment mask, linker mask, skeleton mask and side chain mask, which correspond to four specific tasks in drug design, i.e., determining the drug action target, discovering the ligand, designing the drug molecule and optimizing the drug molecule. The processor masks the protein structure file, i.e., the protein PDB file, of the protein pocket according to the three enhanced training mask strategies and the four task-dependent mask strategies, covers some parts between the protein and the small molecule, and then obtains the protein small molecule mask training pair data. By masking the protein pocket, the model can focus on the information of other parts, which is beneficial to further improve the prediction accuracy of the to-be-trained lead compound optimization model.

[0075] In some embodiments, step S220 includes steps S221 to S223.

[0076] In step S221, the topological surface features in the protein pocket are extracted based on an interaction feature extraction network in the network layer.

[0077] Specifically, the processor extracts topological surface features in the protein pocket based on an interacting feature extraction network in the recurrent connected network layer. The interacting feature extraction network (IFE-Net) is a deep learning model used to predict protein-small molecule interactions. The model is based on two core components: a feature extractor and a predictor. The feature extractor is used to extract interaction features from proteins and small molecules, while the predictor uses these features to make interaction predictions.

[0078] In step S222, the topological surface features are processed by a protein surface gridding method to obtain a protein surface feature vector.

[0079] Specifically, the processor processes the topological surface features in the protein pocket by a protein surface gridding method to obtain a protein surface feature vector. The protein surface gridding is a method for describing the geometric structure of the protein surface. The surface of the protein is composed of atoms, which occupy a certain position in space. By gridding these atoms, i.e., converting them into a two-dimensional or three-dimensional grid with fixed intervals, the distance between atoms on the protein surface and the description of the surface shape can be easily calculated.

[0080] In step S223, the protein surface feature vector is predicted based on an atomic component generation network in the network layer to obtain information about newly generated atoms that match the protein pocket.

[0081] Specifically, the processor processes the protein surface feature vector based on an atomic component generation network in the network layer to obtain information about newly generated atoms that match the protein pocket. The atomic component generation network (ACGN) is a deep learning model used to predict protein-small molecule interactions, i.e., converting the protein-small molecule interaction problem into a sequence generation task. Specifically, ACGN represents small molecules as a series of atomic components and proteins as a series of residues. Then, the model uses an encoder and decoder structure to encode the atomic component sequence of the small molecule and the residue sequence of the protein, and generates the interaction result through the decoder.

[0082] Through the above steps, the processor extracts the topological surface features in the protein pocket based on the interaction feature extraction network in the cyclically connected network layer, and then processes the topological surface features in the protein pocket through a protein surface grid method to obtain a protein surface feature vector. The processor generates a network layer based on the atomic components in the network layer, and performs prediction processing on the protein surface feature vector to predict information of newly generated atoms matched with the protein pocket. The processor performs atomic-by-atomic prediction processing on the training sample data of the lead compound by calling multiple cyclically connected network layers, and then obtains the information of the newly generated atoms matched with the protein pocket, which is conducive to improving the prediction accuracy of the newly generated atoms and further improving the accuracy of the trained lead compound optimization model.

[0083] In some embodiments, step S240 includes steps S241-S242.

[0084] Step S241, when the loss value is not within the preset threshold range, the self-supervised learning strategy and the task strategy of drug design are used to perform mask processing on the protein structure information file of the protein pocket again to obtain the expanded training sample data.

[0085] Specifically, when the processor determines that the loss value of the trained lead compound optimization model is not within the preset threshold range, the processor re-sets the mask strategy and performs mask processing on the protein structure information file of the protein pocket again to obtain the expanded training sample data.

[0086] Step S242, input the expanded training sample data into the trained lead compound optimization model for training until the loss value is within the preset range.

[0087] Specifically, the processor re-enters the expanded training sample data into the trained lead compound optimization model for training until the loss value is within the preset range.

[0088] Through the above steps, when the loss value of the trained lead compound optimization model is not within the preset range, it indicates that the trained lead compound optimization model has not converged, at this time, the expanded training sample data needs to be re-input into the trained lead compound optimization model for continuous training until the loss value is within the preset range, i.e., until the model converges. By determining whether the loss value of the trained lead compound optimization model is within the preset range, it is determined whether to re-train the trained lead compound optimization model until the model converges, thereby improving the prediction accuracy of the trained lead compound optimization model.

[0089] In some embodiments, step S240 further includes steps S243-S245.

[0090] Step S243, obtaining a ligand atom with a distance to the protein pocket within a preset distance threshold range.

[0091] Specifically, the processor obtains the ligand atom closest to the protein pocket.

[0092] Step S244, starting from the ligand atom, traversing each atom in the protein pocket molecule, and generating a target training trajectory.

[0093] Specifically, the processor starts from the ligand atom, traverses each atom in the protein pocket molecule using a breadth-first search algorithm, and generates a target training trajectory.

[0094] Step S245, when the coincidence degree of the newly generated atom trajectory and the target training trajectory is within a preset range, determining that the to-be-trained lead compound optimization model is the target lead compound optimization model.

[0095] Specifically, when the coincidence degree of the newly generated atom trajectory and the target training trajectory is within a preset range, the processor determines that the to-be-trained lead compound optimization model is the target lead compound optimization model.

[0096] Through the above steps, when training the to-be-trained lead compound optimization model, starting from the ligand atom closest to the protein pocket, then using a breadth-first search algorithm to traverse each atom in the molecule, obtaining a generated trajectory, and taking the recovered generated trajectory as the training target. By taking the recovered generated trajectory as the training target, it is beneficial to further improve the prediction accuracy of the to-be-trained lead compound optimization model, further improve the fit degree of the predicted newly generated atom and the protein pocket, and further optimize the lead compound.

[0097] In some embodiments, step S250 includes steps S251-S252.

[0098] Step S251, calling a plurality of network layers to perform atomic processing on the to-be-optimized protein pocket data and the to-be-optimized molecular fragment data, to obtain information of a predicted newly generated atom matched with the to-be-optimized protein pocket.

[0099] Step S252, based on the information of the predicted newly generated atom, generating a molecular three-dimensional structure corresponding to the to-be-optimized protein pocket, the molecular three-dimensional structure being used to bind with the to-be-optimized protein pocket, and thereby obtaining an optimized lead compound.

[0100] Specifically, the processor calls a plurality of network layers to perform atomic processing on the to-be-optimized protein pocket data and the to-be-optimized molecular fragment data, to obtain information of a predicted newly generated atom matched with the to-be-optimized protein pocket; and based on the information of the predicted newly generated atom, a molecular three-dimensional structure corresponding to the to-be-optimized protein pocket is generated, the molecular three-dimensional structure is used to combine with the to-be-optimized protein pocket, and then a lead compound after optimization is obtained. Through the above steps, the binding effect of the predicted newly generated atom and the protein pocket is further improved, and the activity of the lead compound after optimization is further improved.

[0101] In some embodiments, the information of the newly generated atom includes: an atom growth point, an atom position, an atom type, and an atom bond connection relationship of the newly generated atom.

[0102] The present embodiment will be described and explained below through specific embodiments.

[0103] Lead compound optimization is an efficient method. Lead optimization refers to transforming a molecule with low activity into a molecule with high activity and good drug properties through structural modification. The drug activity can be further improved by further structural exploration of a drug that has activity on a certain target but is not suitable for drug development. However, at present, since drug design is a very complex and expert experience-dependent process, it is unrealistic to completely entrust it to the recently developed deep learning-based molecule generation method. Drug discovery still relies on expert selection and modification of compounds, and the lead compound optimization model of the present embodiment can be a good starting point for expert knowledge to learn from a large number of historical cases how to further modify the structure to enhance drug properties.

[0104] Figure 3 is a flowchart of the lead compound optimization method of the present embodiment. As shown in Figure 3 , the lead compound optimization method includes the following steps:

[0105] Step S310, training a lead compound optimization model.

[0106] Specifically, a model training sample is obtained, wherein the model training sample comprises a protein pocket and a molecule fragment to be optimized. The model training sample is input into a lead compound optimization model to be trained for the protein pocket, the lead compound optimization model to be trained comprising a plurality of cyclically connected network layers, each network layer comprising an interaction feature extraction network and an atomic component generation network. The plurality of cyclically connected network layers are called to perform atomic-by-atomic prediction processing on the protein pocket and the molecule to be optimized, to obtain a growth point of a newly generated atom corresponding to the protein pocket, an atomic type, a relative position of the predicted atom, and a bond connection relationship between the predicted atom and the already generated atom. When a certain network layer is used to process the protein pocket, a feature extraction network layer of the network layer is called to extract atomic features in the protein pocket, and then an interaction extraction network layer of the network layer is called to perform vector representation processing on the atomic features, to obtain an atomic feature vector; then, an atomic component generation layer of a previous network layer of the network layer is called to process the atomic feature vector, to obtain a predicted atom growth point, a predicted atom position, a predicted atom type, and a predicted atom bond connection relationship corresponding to the atom of the protein pocket. The atomic position prediction module adopts a density mixture neural network (MDN), and the atomic growth point prediction, the atomic type prediction, and the bond type prediction module directly fit a Dirac distribution.

[0107] Further specifically, the atomic component generation layer can comprise a growth point prediction component, an atomic geometry prediction component, and an atomic prediction component. After obtaining the atomic feature vector of the newly generated atom, the atomic growth point prediction layer is called to select a predicted atom growth starting point for the next network layer from a plurality of already generated atoms according to a probability value of the already generated atom. A specific formula for determining which already generated atom is the starting point is as follows:

[0108]

[0109]

[0110]

[0111] wherein, is a vector in the latent space, and σ is a sigmoid function for outputting a probability value between 0 and 1; represents a probability that the i-th atom is a possible growth point, VP represents a vector perception machine, and MLP represents a multi-layer perception machine.

[0112] Further specifically, the probability distribution of the spatial position is fitted by using the MDN, and a new generated atom is predicted by sampling from the probability distribution; a specific formula for predicting the relative position of the newly generated atom is as follows:

[0113]

[0114]

[0115]

[0116] wherein, is the decomposition quantity of the predicted coordinate, is the combination coefficient of the predicted coordinate decomposition quantity, is the Cartesian coordinate of the growth point atom, and VP represents a vector perception machine.

[0117] Further specifically, the atom type prediction layer processes the atom feature vector according to the atom growth point selected by the previous network layer, and predicts the predicted atom type of the newly generated atom, and the specific formula is as follows:

[0118]

[0119] wherein, is the information after geometric message passing, S i is the predicted atom type, and VP represents a vector perception machine.

[0120] Further specifically, the bond type prediction layer is called, and then the atom feature vector is processed according to the atom growth point selected by the previous network layer, and the predicted atom bonding relationship between the newly generated atom and the atom generated by the previous network layer is predicted, wherein the formula for generating the predicted atom bonding relationship is:

[0121]

[0122]

[0123]

[0124] wherein, b″ ij is the bond type between the generated atom i and the existing atom j, and VP represents a vector perception machine.

[0125] After obtaining the growth point of the newly generated atom, the atom type, the relative position of the predicted atom, and the bonding relationship between the predicted atom and the generated atom, a loss value of the lead compound optimization model in the training protein pocket can be calculated based on the growth point of the newly generated atom, the atom type, the relative position of the predicted atom, and the bonding relationship between the predicted atom and the generated atom. The loss function is composed of four terms, including an atom growth point loss value function, a position loss value function, a type loss value function, and a bond type loss value function. During training, the ligand atom closest to the protein pocket is taken as the starting point, and then a breadth-first search algorithm is used to traverse each atom in the molecule to obtain a generation trajectory, so as to restore the trajectory as the training target. At the same time, the Teacher-Force strategy is used to enhance the convergence speed of the model.

[0126] Further specifically, seven mask strategies are used to mask the training data for training. The first three are random mask, spatial mask, and topological mask, which are used as general self-supervised learning strategies; the last four strategies are fragment mask, linker mask, skeleton mask, and side chain mask, which correspond to four specific tasks in drug design. Specifically, the random mask discards atoms in the molecule with a certain probability; the spatial mask discards atoms within a sphere centered at a certain atom with a distance threshold as the radius; the topological mask discards atoms in the molecule from a starting atom in a given number according to the topological sequence determined by breadth-first search; the fragment mask refers to masking functional groups in the molecule, which can be defined by (Matched-Molecular Pair Analysis, MMPA); the linker mask refers to masking linkers in the molecule, which is not limited to linkers connecting two fragments, including PROTAC (Protein Degradation Targeting Chimera) and linkers connecting three or more fragments, wherein PROTAC is a hybrid bifunctional small molecule that can induce selective intracellular proteolysis; the skeleton mask is a mask for the skeleton in the molecule, which is generally defined by the Murcko skeleton; and the side chain mask is a mask for the side chain in the molecule.

[0127] After obtaining the predicted atom growth point, the atom growth point loss value can be calculated based on the predicted atom growth point and the labeled atom growth point of the protein pocket. The calculation formula is as follows:

[0128]

[0129] wherein, represents the loss value of the predicted growth point atom; represents the loss value of the predicted atom geometry; represents the type and bond type loss of the predicted atom.

[0130] After the loss value of the to-be-trained lead compound optimization model in the protein pocket is calculated, it can be judged whether the loss value is within the preset range. If the loss value is not within the preset range, it indicates that the to-be-trained lead compound optimization model in the protein pocket has not converged, at which time the to-be-trained lead compound optimization model in the protein pocket can be continuously trained until the model converges. If the loss value is within the preset range, it indicates that the to-be-trained lead compound optimization model in the protein pocket has converged, at which time the trained to-be-trained lead compound optimization model in the protein pocket can be used as the final lead compound optimization model in the protein pocket, that is, the model with the smallest loss value obtained by training is used as the trained model.

[0131] Further specifically, referring to Figure 4 , Figure 4 is a schematic diagram of the predicted atom generated by the embodiment. The to-be-optimized molecule and the protein pocket are model training samples. After the model training samples are input into the to-be-trained lead compound optimization model, the position of the first optimization atom is predicted according to the growth point prediction component-Attachment Point Selection. Exemplarily, the growth point of the predicted newly generated atom is an atom in the to-be-optimized ligand in the protein pocket, and the position of the first optimization atom is predicted according to the atom. According to the atom geometry prediction component-Conformation Prediction, the KNN graph is constructed based on the position of the new atom to realize message passing of the structure of the protein pocket and the to-be-optimized molecule, and the category of the newly grown atom is predicted using the information. According to the atom prediction component-Atom Prediction, the newly generated atom is added to the to-be-optimized ligand, and the bond relationship between the newly generated atom and the atom in the to-be-optimized molecule is predicted using the current geometric state. The above process is trained in a loop until the molecular size of all the growth points of the atoms is less than or equal to the preset molecular size, that is, the optimized molecule is obtained. In order to capture long-range interactions, the protein pocket is characterized as a triangular face according to a program for calculating the depth of protein residues, that is, the MSMS program. Meanwhile, the small molecule is represented in the form of a molecular graph, and in order to capture the interaction between the pocket and the small molecule, the pocket graph and the small molecule graph are expanded into the form of a geometric graph.

[0132] Step S320, processing the to-be-optimized lead compound.

[0133] Specifically, the protein pocket to be processed and the molecule to be optimized are obtained, the topological surface of the protein pocket is obtained, and the protein pocket to be processed and the molecule to be optimized are input into the loss value minimum lead compound optimization model, a plurality of cyclically connected network layers are called to process the protein pocket to be processed and the molecule to be optimized atom by atom, and the corresponding atom growth point, atom position, atom type and atom bonding relationship of the protein pocket to be processed are predicted; based on the atom growth point, the atom position, the atom type and the atom bonding relationship, a three-dimensional structure of the molecule corresponding to the protein pocket to be processed is generated; and the optimized lead compound is obtained by combining the protein pocket to be processed and the three-dimensional structure of the molecule.

[0134] Further specifically, the lead compound optimization model is applied to model inference. First, a target is selected, for example, the target is the target Adrb1 of the G protein-coupled receptor, and the PDB structure file of the protein target is obtained, and the position of the protein pocket is determined; after analyzing the protein pocket, the protein pocket and the mask molecule pair are obtained by using the mask strategy, and then the three-dimensional structure of the molecule corresponding to the protein pocket is generated according to the lead compound optimization model, and then the generated three-dimensional structure of the molecule is combined with the residues of the protein pocket, and then the optimized lead compound is generated, and the optimized lead compound is compared with the original target molecule to obtain the performance of the lead compound optimization model.

[0135] Figure 5A is a schematic diagram of the original molecule of the model inference of the embodiment, and Figure 5A the target of the G protein-coupled receptor (GPCR) is taken as the original molecule; Figure 5B is a state diagram of the original molecule of the model inference of the embodiment, Figure 5B specifically indicating the state of the original molecule in the protein pocket; Figure 5C is a schematic diagram of the analysis molecule of the model inference of the embodiment, Figure 5C specifically indicating the type of the residues of the original molecule and the protein pocket, and the like, wherein W100, F110 and the like respectively indicate the residue type of the protein pocket; Figure 5D is a schematic diagram of the molecule to be optimized of the model inference of the embodiment, and specifically indicates the molecule to be optimized obtained by the mask strategy; Figure 5E1is a schematic diagram of a target molecule of model inference of the embodiment; the molecule to be optimized is input into the lead compound optimization model to generate a target molecule, the target molecule is combined with residues of a protein pocket, and then a lead compound molecule after the target molecule is combined with the residues of the protein pocket is obtained, wherein the residues of the protein pocket are W100, L94, T412, F389, W386, F390, S197 and F198; wherein the value of QED (Quantum-chemical Expansion of Drug-like, a method for evaluating the similarity of drug molecules) of the lead compound molecule is 0.6855, the value of SA (Surface Area, surface area) is 0.8200, the value of Energy (energy) is -9.4737, the value of 2D is 0.4290, and the value of 3D is 0.7329; Figure 5E2 is a schematic diagram of another target molecule of model inference of the embodiment; the molecule to be optimized is input into the lead compound optimization model to generate a target molecule, the target molecule is combined with residues of a protein pocket, and then a lead compound molecule after the target molecule is combined with the residues of the protein pocket is obtained, wherein, referring to Figure 5E2 , the residues of the protein pocket are W100, D114, T412, W386, F390, S197 and F198; wherein the value of QED of the lead compound molecule is 0.7311, the value of SA is 0.7300, the value of Energy is -10.7952, the value of 2D is 0.4923, and the value of 3D is 0.7887; Figure 5E3 is a schematic diagram of another target molecule of model inference of the embodiment; the molecule to be optimized is input into the lead compound optimization model to generate a target molecule, the target molecule is combined with residues of a protein pocket, and then a lead compound molecule after the target molecule is combined with the residues of the protein pocket is obtained, wherein, referring to Figure 5E3 , the residues of the protein pocket are W100, Y408, T412, W386, F390 and F198; wherein the value of QED of the lead compound molecule is 0.8231, the value of SA is 0.8300, the value of Energy is -8.3681, the value of 2D is 0.4895, and the value of 3D is 0.7342; Figure 5E4 is a schematic diagram of another target molecule of model inference of the embodiment; the molecule to be optimized is input into the lead compound optimization model to generate a target molecule, the target molecule is combined with residues of a protein pocket, and then a lead compound molecule after the target molecule is combined with the residues of the protein pocket is obtained, wherein, referring to Figure 5E4The residues of the protein pocket are W100, F110, Y408, D114, S197, T412, W386, F389 and F198. The value of QED of the lead compound molecule is 0.7874, the value of SA is 0.6700, the value of Energy is -8.4128, the value of 2D is -10.4177, and the value of 3D is 0.7446. Figure 5F is a diagram of a performance control of the model inference of the embodiment, Figure 5G is a diagram of another performance control of the model inference of the embodiment. The energy of the lead compound molecule is calculated according to the obtained target molecule and the protein pocket, and the energy difference between the molecule optimized by the lead compound optimization model and the target molecule is obtained. Wherein, Figure 5F The horizontal coordinate Scoring Energy of represents the molecular binding energy, and the vertical coordinate Density represents the density. -10.7 represents the binding energy of the target molecule, and 23.634% represents the part whose value of the binding energy of the target molecule and the protein pocket obtained according to the lead compound optimization model is lower than -10.7. The lower the energy of the molecular binding energy, the higher the activity of the lead compound. Figure 5G The horizontal coordinate Docking Energy of represents the docking energy, and the vertical coordinate Density represents the density. -11.9 represents the docking energy of the target molecule, and 43.6364% represents the part whose docking energy is lower than that of the original molecule and the protein pocket. The lower the docking energy, the higher the activity of the lead compound. Therefore, from Figure 5F and Figure 5G It can be obtained that the activity of the lead compound obtained by the trained lead compound optimization model is higher within a certain range, and therefore the problem that the traditional lead compound optimization method in the related art is limited to the active compound space expressed by the compound library prepared in advance can be solved.

[0136] It should be noted that the steps shown in the above flow or the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions.

[0137] In the embodiment, a lead compound optimization device is also provided, which is used to implement the above embodiments and preferred embodiments, and will not be described again. The terms "module", "unit", "sub-unit" and the like used below can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware, or a combination of software and hardware is also possible and is conceived.

[0138] Figure 6 is a structural block diagram of the lead compound optimization device of the embodiment, as Figure 6As shown, the device comprises a training module 10 and an optimization module 20.

[0139] The training module 10 is configured to input preset training sample data of a lead compound into a lead compound optimization model to be trained, the lead compound model to be trained comprises a plurality of network layers, the training sample data of the lead compound comprises protein pocket data and protein-small molecule mask training pair data; the plurality of network layers are called to predict the training sample data of the lead compound to obtain information of newly generated atoms matched with the protein pocket; a loss value of the lead compound optimization model to be trained is calculated based on the information of the newly generated atoms; when the loss value is within a preset threshold range, the lead compound optimization model to be trained is determined as a target lead compound optimization model.

[0140] The optimization module 20 is configured to input protein pocket data to be optimized and molecular fragment data to be optimized into the target lead compound optimization model to generate a molecular structure corresponding to the protein pocket data to be optimized; and generate an optimized lead compound according to the molecular structure and the protein pocket data to be optimized.

[0141] It should be noted that each of the above modules can be a functional module or a program module, and can be implemented by software or hardware. For the modules implemented by hardware, each of the above modules can be located in the same processor; or each of the above modules can be located in different processors in any combination.

[0142] In this embodiment, an electronic device is also provided, which comprises a memory and a processor, the memory stores a computer program, and the processor is configured to execute the computer program to perform the steps in any of the above method embodiments.

[0143] Optionally, the electronic device can further comprise a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0144] Optionally, in this embodiment, the processor can be configured to execute the following steps through the computer program:

[0145] S1, input preset training sample data of a lead compound into a lead compound optimization model to be trained, the lead compound model to be trained comprises a plurality of network layers, and the training sample data of the lead compound comprises protein pocket data and protein-small molecule mask training pair data.

[0146] S2, call the plurality of network layers to predict the training sample data of the lead compound to obtain information of newly generated atoms matched with the protein pocket.

[0147] S3, calculate a loss value of the to-be-trained lead compound optimization model based on the information of the newly generated atom.

[0148] S4, when the loss value is within a preset threshold range, determine the to-be-trained lead compound optimization model as a target lead compound optimization model.

[0149] S5, input the to-be-optimized protein pocket data and the to-be-optimized molecular fragment data into the target lead compound optimization model to generate a molecular structure corresponding to the to-be-optimized protein pocket data.

[0150] S6, generate an optimized lead compound according to the molecular structure and the to-be-optimized protein pocket data.

[0151] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation manners, which will not be described herein again.

[0152] In addition, in combination with the lead compound optimization method provided in the above embodiments, a storage medium can also be provided in this embodiment to realize. The storage medium has a computer program stored thereon; the computer program is executed by a processor to realize any one of the lead compound methods in the above embodiments.

[0153] It should be understood that the specific embodiments described herein are only used to explain this application, but not to limit it. According to the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of this application.

[0154] Obviously, the drawings are only some examples or embodiments of the present application, and those skilled in the art can also apply the present application to other similar situations without creative labor. In addition, it can be understood that although the work done in this development process may be complex and long, for those skilled in the art, some design, manufacture or production changes according to the technical content disclosed in this application are only routine technical means, and should not be regarded as insufficient disclosure of the present application.

[0155] The term "embodiment" in this application refers to the specific features, structures or characteristics described in combination with the embodiments, which can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily mean the same embodiment, nor does it mean independence or alternative to other embodiments. It can be clearly or implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments without conflict.

[0156] The above embodiments only express several implementation ways of the present application, and the description is specific and detailed, but it should not be understood as a limitation to the patent protection scope. It should be pointed out that, for ordinary skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A lead compound optimization method characterized by, The method comprises: inputting preset training sample data of a lead compound into a lead compound optimization model to be trained, the lead compound optimization model to be trained comprising a plurality of network layers, the training sample data of the lead compound comprising protein pocket data and protein-small molecule mask training pair data; the protein-small molecule mask training pair data is subjected to mask processing on a protein structure information file of the protein pocket by using a self-supervised learning strategy and a drug design task strategy to obtain; calling the plurality of network layers to predict the training sample data of the lead compound to obtain information of newly generated atoms matched with the protein pocket; calculating a loss value of the lead compound optimization model to be trained based on the information of the newly generated atoms; when the loss value is within a preset threshold range, determining that the lead compound optimization model to be trained is a target lead compound optimization model; inputting protein pocket data to be optimized and molecular fragment data to be optimized into the target lead compound optimization model, calling the plurality of network layers to perform atom processing on the protein pocket data to be optimized and the molecular fragment data to be optimized to obtain information of predicted newly generated atoms matched with the protein pocket to be optimized, generating a molecular three-dimensional structure corresponding to the protein pocket to be optimized based on the information of the predicted newly generated atoms, and combining residues of the molecular three-dimensional structure and the protein pocket data to be optimized to generate an optimized lead compound.

2. The lead compound optimization method of claim 1, wherein, The calling the plurality of network layers to predict the training sample data of the lead compound to obtain information of newly generated atoms matched with the protein pocket comprises: extracting topological surface features in the protein pocket based on an interaction feature extraction network in the network layer; processing the topological surface features by a protein surface grid method to obtain protein surface feature vectors; performing prediction processing on the protein surface feature vectors by an atom component generation network layer in the network layer to predict information of newly generated atoms matched with the protein pocket.

3. The lead compound optimization method of claim 1, wherein, When the loss value is not within the preset threshold range, the protein structure information file of the protein pocket is subjected to mask processing again by using the self-supervised learning strategy and the drug design task strategy to obtain expanded training sample data; inputting the expanded training sample data into the lead compound optimization model to be trained for training until the loss value is within the preset range. The calling the plurality of network layers to predict the training sample data of the lead compound to obtain information of newly generated atoms matched with the protein pocket comprises:

4. The lead compound optimization method of claim 1, wherein, obtaining ligand atoms having a distance within a preset distance threshold range from the protein pocket; traversing each atom in a protein pocket molecule from the ligand atom as a starting point and generating a target training trajectory; ​ When the coincidence degree of the trajectory of the newly generated atom and the target training trajectory is within a preset range, the to-be-trained lead compound optimization model is determined as a target lead compound optimization model.

5. The lead compound optimization method of claim 1, wherein, The information of the newly generated atom includes an atom growth point, an atom position, an atom type, and an atom bonding relationship of the newly generated atom.

6. A lead compound optimization apparatus characterized by, The device comprises a training module and an optimization module. The training module is configured to input preset training sample data of a lead compound into a to-be-trained lead compound optimization model, the to-be-trained lead compound model comprising a plurality of network layers, the training sample data of the lead compound comprising protein pocket data and protein-small molecule mask training pair data; the protein-small molecule mask training pair data is obtained by performing mask processing on a protein structure information file of the protein pocket by using a self-supervised learning strategy and a task strategy of drug design; a plurality of network layers are called to predict the training sample data of the lead compound, to obtain information of a newly generated atom matched with the protein pocket; a loss value of the to-be-trained lead compound optimization model is calculated based on the information of the newly generated atom; when the loss value is within a preset threshold range, the to-be-trained lead compound optimization model is determined as a target lead compound optimization model. The optimization module is configured to input to-be-optimized protein pocket data and to-be-optimized molecular fragment data into the target lead compound optimization model, call a plurality of network layers to perform atom processing on the to-be-optimized protein pocket data and the to-be-optimized molecular fragment data, and obtain information of a predicted newly generated atom matched with the to-be-optimized protein pocket; based on the information of the predicted newly generated atom, a molecular three-dimensional structure corresponding to the to-be-optimized protein pocket is generated; and a lead compound after optimization is generated by combining the molecular three-dimensional structure and residues of the to-be-optimized protein pocket data. 7.An electronic device comprising a memory and a processor, the electronic device characterized by, The memory stores a computer program, and the processor is configured to run the computer program to execute the lead compound optimization method in any one of claims 1 to 5.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the lead compound optimization method in any one of claims 1 to 5.

Images