Method and apparatus for training a composite data prediction model

Abstract

The application provides a training method and device of a complex data prediction model, a storage medium and an electronic device, and the method comprises the following steps: acquiring first target point data and first ligand data; docking the first target point data and the first ligand data to obtain simulated first complex data; training an intermediate model based on the first target point data, the first ligand data and the first complex data; acquiring second target point data, second ligand data and second complex data; and adjusting the intermediate model based on the second target point data, the second ligand data and the second complex data to obtain a complex data prediction model. In the technical scheme provided in the application, the intermediate model is obtained by pre-training using simulation data, and then the intermediate model is fine-tuned using real experimental data to obtain the complex data prediction model, and the model performance of the complex data prediction model is relatively high.

Description

Technical Field

[0001] The present invention relates to the technical field of computer-aided drug design, and more particularly, to a method and apparatus for training a compound data prediction model. Background Art

[0002] Molecular docking is an important molecular modeling method widely used in computer-aided drug design. In drug design, the goal of molecular docking is to predict the binding conformation of a potential small molecule drug ligand with a receptor protein. Drug developers can use this conformation to analyze ligand-receptor interactions or use it as input to further predict ligand-receptor binding affinity using high-precision computational chemistry methods such as energy optimization and molecular simulation.

[0003] Drug (ligand)-target (receptor) affinity is an important indicator for judging whether a drug can be developed into a drug and an important basis for screening lead compounds. Affinity prediction can help quickly understand the mechanism of action of ligand-receptor binding, greatly shorten the drug development process, reduce many unnecessary biochemical experiments, and thus achieve more efficient drug screening.

[0004] In recent years, with the development of artificial intelligence (AI) technology, deep learning models have begun to be applied to drug R&D to shorten R&D cycles and reduce resource waste. However, deep learning models rely on the amount of experimental data, making it difficult to obtain high-performance deep learning models when experimental data is limited. Summary of the Invention

[0005] The present invention provides a training method, device, computer-readable storage medium and electronic device for a composite data prediction model to solve the technical problem of poor performance of deep learning models in the prior art.

[0006] According to a first aspect of the present invention, a method for training a composite data prediction model is provided, comprising:

[0007] Acquiring first target data and first ligand data;

[0008] docking the first target data and the first ligand data to obtain simulated first complex data;

[0009] Training an intermediate model based on the first target data, the first ligand data, and the first complex data;

[0010] Acquiring second target data, second ligand data, and second complex data, where the second complex data is experimental data corresponding to the second target data and the second ligand data;

[0011] Based on the second target data, the second ligand data and the second complex data, the intermediate model is adjusted to obtain a complex data prediction model.

[0012] Optionally, the training of an intermediate model based on the first target data, the first ligand data, and the first complex data includes:

[0013] Inputting the first target data and the first ligand data into an initial model to obtain first predicted data, wherein the first predicted data at least includes a predicted complex conformation;

[0014] determining a loss value based on the first predicted data and the first complex data, the first complex data including at least a simulated complex conformation;

[0015] The model parameters of the initial model are adjusted based on the loss value to train an intermediate model.

[0016] Optionally, inputting the first target data and the first ligand data into an initial model to obtain first prediction data includes:

[0017] Inputting the first target data into the target embedding layer of the initial model to obtain target atomic features and target atom pair features; inputting the first ligand data into the ligand embedding layer of the initial model to obtain ligand atomic features and ligand atom pair features;

[0018] Inputting the target atomic features and the target atom pair features into a first encoder to obtain encoded target atomic features and encoded target atom pair features; inputting the ligand atomic features and the ligand atom pair features into a second encoder to obtain encoded ligand atomic features and encoded ligand atom pair features;

[0019] Inputting the encoded target atomic features, the encoded target atom pair features, the encoded ligand atomic features, and the encoded ligand atom pair features into a decoder to obtain complex atomic features and complex atom pair features;

[0020] The complex atom features and the complex atom pair features are input into a predictor to obtain first prediction data.

[0021] Optionally, the first encoder, the second encoder, and the decoder each include a plurality of Transformer encoding layers, an outer product module, and a triangular attention module; when the input data of the first encoder, the second encoder, and the decoder are regarded as first atomic features and first atom pair features, and the output data are regarded as second atomic features and second atom pair features, the first encoder, the second encoder, and the decoder are configured to perform the following steps:

[0022] Inputting the first atomic feature into the first Transformer coding layer, using the first atomic pair feature as a bias item of the self-attention module in each Transformer coding layer, each self-attention module outputs an attention weight, and adding the attention weight output by each self-attention module and the first atomic pair feature as a new first atomic pair feature; the first Transformer coding layer outputs a new first atomic feature, and the new first atomic feature is input into the next Transformer coding layer until the last Transformer coding layer outputs the second atomic feature;

[0023] The second atomic feature is input into the outer product module, and the output data of the outer product module and the new first atom pair feature are input into the triangular attention module to obtain the second atom pair feature output by the triangular attention module.

[0024] Optionally, the first prediction data further includes a prediction score and / or a prediction energy item, and the prediction energy item is used to calculate the prediction score;

[0025] In the case where the first prediction data includes a prediction score, the first complex data also includes a docking score;

[0026] In the case where the first prediction data includes a predicted energy term, the first complex data further includes a simulated energy term, and the simulated energy term is used to calculate the docking score.

[0027] Optionally, inputting the first target data and the first ligand data into an initial model to obtain first prediction data includes:

[0028] applying first noise to the first ligand data to obtain third ligand data;

[0029] Inputting the first target data and the third ligand data into an initial model to obtain first predicted data and second noise;

[0030] The determining of the loss value based on the first prediction data and the first composite data includes:

[0031] A loss value is determined based on the first predicted data, the first composite data, the first noise, and the second noise.

[0032] Optionally, the method further includes:

[0033] Obtaining fixed data and obtaining data to be filtered corresponding to the fixed data;

[0034] Inputting the data to be screened and the fixed data into the composite data prediction model to obtain second predicted data;

[0035] Based on the second prediction data, a screening result is determined in the data to be screened.

[0036] According to a second aspect of the present invention, there is provided a training device for a composite data prediction model, comprising:

[0037] A first data acquisition module, used to acquire first target data and first ligand data;

[0038] a complex simulation module, configured to dock the first target data with the first ligand data to obtain simulated first complex data;

[0039] A pre-training module, configured to train an intermediate model based on the first target data, the first ligand data, and the first complex data;

[0040] A second data acquisition module is configured to acquire second target data, second ligand data, and second complex data, wherein the second complex data is experimental data corresponding to the second target data and the second ligand data;

[0041] A model adjustment module is used to adjust the intermediate model based on the second target data, the second ligand data and the second complex data to obtain a complex data prediction model.

[0042] According to a third aspect of the present invention, a computer-readable storage medium is provided, wherein the storage medium stores a computer program for executing the above-mentioned training method for the composite data prediction model.

[0043] According to a fourth aspect of the present invention, there is provided an electronic device, comprising:

[0044] processor;

[0045] a memory for storing instructions executable by the processor;

[0046] The processor is configured to read the executable instructions from the memory and execute the instructions to implement the above-mentioned training method for the composite data prediction model.

[0047] Compared with the prior art, the training method, device, computer-readable storage medium, and electronic device for the composite data prediction model provided by the present invention have at least the following beneficial effects:

[0048] The technical solution of the present invention, after obtaining the first target data and the first ligand data, docks the first target data and the first ligand data to obtain simulated first complex data, and then pre-trains based on the first target data, the first ligand data, and the first complex data to train an intermediate model. Further, real experimental data, namely the second target data, the second ligand data, and the second complex data, are obtained, and the intermediate model is adjusted using the second target data, the second ligand data, and the second complex data to obtain a complex data prediction model. In the technical solution provided by the present invention, pre-training is first performed using simulated data to obtain an intermediate model, and then the intermediate model is fine-tuned using real experimental data to obtain a complex data prediction model. The model performance of the complex data prediction model is relatively high. A high-performance complex data prediction model can be obtained based on a small amount of experimental data. The complex data prediction model is further used for virtual screening, which can avoid the intervention of human factors and has high screening efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the description of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 1 is a flow chart of a method for training a composite data prediction model provided by an exemplary embodiment of the present invention;

[0051] Figure 2 This is a schematic diagram of the model structure in the training method of the composite data prediction model provided by an exemplary embodiment of the present invention. Figure 1 ;

[0052] Figure 3 1 is a schematic diagram of a local model structure in a method for training a composite data prediction model provided by an exemplary embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the model structure in the training method of the composite data prediction model provided by an exemplary embodiment of the present invention. Figure 2 ;

[0054] Figure 51 is a schematic structural diagram of a training device for a composite data prediction model provided by an exemplary embodiment of the present invention;

[0055] Figure 6 is a structural diagram of an electronic device provided by an exemplary embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of these embodiments.

[0057] Exemplary Methods

[0058] Figure 1 FIG. 1 is a flow chart of a method for training a composite data prediction model provided by an exemplary embodiment of the present invention, which includes at least the following steps:

[0059] Step 11: Acquire first target data and first ligand data.

[0060] Taking protein targets as an example, a target refers to a region of a protein in the body that can bind to other substances. Drugs generally bind to this target, affecting the protein's structure and, in turn, its functional expression, thereby curing or alleviating symptoms. Targets are sometimes also referred to as receptors, meaning that the first target data can also be referred to as the first receptor data. A ligand refers to a substance that binds to a target, typically a small molecule or peptide.

[0061] In this step, the first target data can be the target protein conformation, which can be predicted by computer simulation, experimental data in public data sets, or experimental data obtained by self-experimentation. When predicting by computer simulation, the AlphaFold series of models can be used for prediction. The AlphaFold series of models can predict protein conformation based on protein sequence. When obtaining experimental data by self-experimentation, the target protein is generally combined with a specific active ligand to obtain a complex precipitate to analyze the conformation of the target protein. Of course, the pdbbind data set can also be used to obtain public experimental data.

[0062] In one possible implementation, the first target data is obtained using the pdbbind dataset. Since the conformations in this dataset are experimentally derived and more accurate than those derived from simulations, the conformation of the target protein is not altered in subsequent simulations.

[0063] In one possible implementation, the first target data is obtained through the pdbbind dataset and the user's own experimental data. The public target protein conformation is obtained through the pdbbind dataset. The user's own experimental data is obtained after conducting experiments on their own. Using the user's own experimental data as part of the first target data is conducive to improving model performance.

[0064] In this step, the first ligand data can be obtained through a public data set. Pubchem collects a large amount of publicly available small molecule data, so the first ligand data can be obtained through the Pubchem data set. Specifically, most of the known small molecule substances are stored in Pubchem, from which a part of small molecules are randomly selected to constitute a small molecule library to be docked, and the first ligand data is obtained. Furthermore, after obtaining the small molecule, the conformation of the small molecule can also be optimized, that is, RDKit (a software) is used to generate an initial conformation based on the obtained small molecule, and simulation software such as XTB is used to optimize the conformation to obtain the optimized small molecule conformation, i.e., the first ligand data.

[0065] Step 12: docking the first target data and the first ligand data to obtain simulated first complex data.

[0066] In this step, after obtaining the first target data and the first ligand data, the first target data and the first ligand data are docked to simulate first complex data. The complex is composed of multiple types of molecules. The first complex data is parameter data obtained after docking the first target data and the first ligand data. This parameter data can be used to conduct a preliminary assessment of the complex, such as determining the stability of the complex structure. Therefore, the first complex data is not actual experimental data, but simulated data, and the amount of data generated is relatively large.

[0067] Specifically, simulation software can be used to dock the first target data and the first ligand data, such as The glide module of the software (Schrödinger software) can include at least one of flexible docking, semi-flexible docking, and rigid docking. For example, the first target data and the first ligand data are subjected to flexible docking to obtain simulated first complex data. Alternatively, the first target data and the first ligand data are subjected to flexible docking and semi-flexible docking to obtain simulated first complex data.

[0068] For example, for each target protein, 50 small molecules are randomly selected from the small molecule library and the The software's Glide module docks 50 times, and up to 2,500 complex conformations can be obtained for each target protein (some small molecules and target protein pairs cannot simulate stable complex conformations). There are about 19,000 small molecules in the PubChem dataset. After this step, a large amount of simulation data (about 40 million) can be obtained, which is used for the subsequent model pre-training step.

[0069] In a possible implementation manner, the first complex data at least includes a simulated complex conformation.

[0070] In one possible implementation, the first complex data includes not only a simulated complex conformation but also a docking score and / or a simulated energy term, wherein the simulated energy term is used to calculate the docking score, and the docking score is used to evaluate the relevant intermolecular forces in the complex. Specifically, the first complex data may include a simulated complex conformation and a docking score; or the first complex data may include a simulated complex conformation and a simulated energy term; or the first complex data may include a simulated complex conformation, a docking score, and a simulated energy term.

[0071] Step 13: training an intermediate model based on the first target data, the first ligand data, and the first complex data.

[0072] In this step, after obtaining the first complex data simulated by the first target data and the first ligand data, model pre-training is performed using the first target data, the first ligand data and the first complex data to train an intermediate model.

[0073] In some embodiments, the first target data and the first ligand data are input into an initial model, which is a pre-built mathematical model. A loss value is determined based on the output data of the initial model and the first complex data. The model parameters of the initial model are adjusted based on the loss value to obtain a pre-trained intermediate model.

[0074] In some embodiments, step 13, training an intermediate model based on the first target data, the first ligand data, and the first complex data, comprises:

[0075] Step 131: Input the first target data and the first ligand data into an initial model to obtain first prediction data, wherein the first prediction data at least includes a predicted complex conformation.

[0076] In this step, after obtaining the first target data and the first ligand data, the first target data and the first ligand data are input into the initial model, which is a pre-built mathematical model, to obtain the first predicted data output by the initial model. The first predicted data at least includes the predicted conformation of the complex. The conformation refers to the arrangement of the atoms of a substance in three-dimensional space. The overall translation and rotation do not change the conformation of the substance.

[0077] In some embodiments, the first prediction data also includes: a prediction score and / or a prediction energy item, the prediction energy item is used to calculate the prediction score, and the prediction score is used to evaluate the interaction between the molecules of the complex. The higher the prediction score, the tighter the binding between the molecules of the complex.

[0078] In one possible implementation, the first target data and the first ligand data are input into an initial model to obtain a predicted complex conformation.

[0079] In one possible implementation, the first target data and the first ligand data are input into the initial model to obtain a predicted complex conformation and a predicted score.

[0080] In one possible implementation, the first target data and the first ligand data are input into the initial model to obtain a predicted complex conformation and a predicted energy term.

[0081] In one possible implementation, the first target data and the first ligand data are input into the initial model to obtain a predicted complex conformation, a predicted score, and a predicted energy term.

[0082] In some embodiments, when the first prediction data includes at least two items of data, the number of the initial models may be one or at least two.

[0083] In one possible implementation, the first target data and the first ligand data are input into an initial model. The initial model includes at least two predictors, each of which outputs one item of prediction data, wherein the predicted complex conformation, the predicted score, and the predicted energy term constitute three items of prediction data. Exemplarily, the first target data and the first ligand data are input into the initial model, the complex conformation predictor of the initial model outputs a predicted complex conformation, and the affinity predictor of the initial model outputs a predicted score.

[0084] In one possible implementation, the first target data and the first ligand data are input into at least two initial models, with each initial model outputting a prediction datum. For example, the first target data and the first ligand data are input into a first initial model, resulting in a predicted complex conformation output by the first initial model. The first target data and the first ligand data are input into a second initial model, resulting in a prediction score output by the second initial model.

[0085] The model structure of the above-mentioned initial model may include a target embedding layer, a ligand embedding layer, a first encoder, a second encoder, a decoder and a predictor, wherein the predictor can be adaptively adjusted according to the number of items of the predicted data. It should be noted that the model structure of the above-mentioned initial model can also be other structures, which is not specifically limited in this embodiment.

[0086] In some embodiments, step 131 inputs the first target data and the first ligand data into an initial model to obtain first prediction data, including:

[0087] Step 1311: input the first target data into the target embedding layer of the initial model to obtain target atomic features and target atom pair features; input the first ligand data into the ligand embedding layer of the initial model to obtain ligand atomic features and ligand atom pair features.

[0088] In this step, the embedding layer maps the original data into a dense representation. After the first target data is input into the target embedding layer of the initial model, densely represented target atomic features and target atom pair features are obtained. The target atomic features cover various information about the atoms at the corresponding positions in the target protein, while the target atom pair features cover the atomic combinations and bond information of the corresponding atom pairs in the target protein. The first ligand data is input into the ligand embedding layer of the initial model, which results in densely represented ligand atomic features and ligand atom pair features. The ligand atomic features cover various information about the atoms at the corresponding positions in the ligand, while the ligand atom pair features cover the atomic combinations and bond information of the corresponding atom pairs in the ligand.

[0089] Step 1312: input the target atomic features and the target atom pair features into a first encoder to obtain encoded target atomic features and encoded target atom pair features; input the ligand atomic features and the ligand atom pair features into a second encoder to obtain encoded ligand atomic features and encoded ligand atom pair features.

[0090] In this step, the first encoder is the target encoder, and the second encoder is the ligand encoder. The target encoder is used to model the target protein, allowing a full understanding of its data characteristics; the ligand encoder is used to model the ligand, allowing a full understanding of its data characteristics. The target atomic features and target atom pair features are input into the first encoder, and after encoding by the first encoder, the encoded target atomic features and encoded target atom pair features are obtained. The ligand atomic features and ligand atom pair features are input into the second encoder, and after encoding by the second encoder, the encoded ligand atomic features and encoded ligand atom pair features are obtained.

[0091] Step 1313: Input the encoded target atomic features, the encoded target atom pair features, the encoded ligand atomic features, and the encoded ligand atom pair features into a decoder to obtain complex atomic features and complex atom pair features.

[0092] In this step, the decoder is an interactive decoder, which is used to interact and decode the encoded information of the target protein and ligand. The encoded target atomic features, encoded target atom pair features, encoded ligand atomic features, and encoded ligand atom pair features are spliced ​​and input into the decoder to obtain complex atomic features and complex atom pair features. Complex atomic features include various information about atoms at corresponding positions in the complex, while complex atom pair features include information about the atomic combinations and bonds of corresponding atom pairs in the complex.

[0093] Step 1314: Input the complex atom features and the complex atom pair features into a predictor to obtain first prediction data.

[0094] In this step, after obtaining the complex atomic features and the complex atom pair features, the complex atomic features and the complex atom pair features are input into a predictor to obtain first prediction data. The first prediction data conforms to human language habits and is readable.

[0095] Specifically, the number of predictors can be at least one, that is, the number of predictors can be one or at least two. When the number of predictors is at least two, complex atom features and complex atom pair features are input into each predictor to obtain data output by each predictor. When the predictor is a complex conformation predictor, the predicted complex conformation is output; when the predictor is an affinity predictor, the predicted score is output; and when the predictor is an energy predictor, the predicted energy term is output.

[0096] For example, Figure 2As shown, the acquired first target data includes atomic information, interatomic distances, and bond information. The first target data is input into embedding layer 21, i.e., the target embedding layer, and the first ligand data is input into embedding layer 22, i.e., the ligand embedding layer. The output data of embedding layer 21, i.e., the target atomic features and target atom pair features, serve as input data to target encoder 23, while the output data of embedding layer 22, i.e., the ligand atomic features and ligand atom pair features, serve as input data to ligand encoder 24. The encoded target atomic features and encoded atom pair features output by target encoder 23 are concatenated with the encoded ligand atomic features and encoded ligand atom pair features output by ligand encoder 24 and then input into interactive decoder 25. Interactive decoder 25 outputs complex atomic features and complex atom pair features, which are then input into complex conformation predictor 26 to obtain a predicted complex conformation. The complex atomic features and complex atom pair features are then input into affinity predictor 27 to obtain a prediction score.

[0097] In some embodiments, the first encoder, the second encoder, and the decoder each include a plurality of Transformer encoding layers, an outer product module, and a triangular attention module; when the input data of the first encoder, the second encoder, and the decoder are regarded as first atomic features and first atom pair features, and the output data are regarded as second atomic features and second atom pair features, the first encoder, the second encoder, and the decoder are configured to perform the following steps:

[0098] Inputting the first atomic feature into the first Transformer coding layer, using the first atomic pair feature as a bias item of the self-attention module in each Transformer coding layer, each self-attention module outputs an attention weight, and adding the attention weight output by each self-attention module and the first atomic pair feature as a new first atomic pair feature; the first Transformer coding layer outputs a new first atomic feature, and the new first atomic feature is input into the next Transformer coding layer until the last Transformer coding layer outputs the second atomic feature;

[0099] The second atomic feature is input into the outer product module, and the output data of the outer product module and the new first atom pair feature are input into the triangular attention module to obtain the second atom pair feature output by the triangular attention module.

[0100] Specifically, for the first encoder, the first atomic feature of the above-mentioned input data is the target atomic feature, the first atom pair feature is the target atom pair feature, the second atomic feature of the above-mentioned output data is the encoded target atomic feature, and the second atom pair feature is the encoded target atom pair feature.

[0101] For the second encoder, the first atomic feature of the above-mentioned input data is the ligand atom feature, the first atom pair feature is the ligand atom pair feature, and the second atomic feature of the above-mentioned output data is the encoded ligand atom feature, and the second atom pair feature is the encoded ligand atom pair feature.

[0102] For the decoder, the first atomic feature of the above-mentioned input data is the concatenation result of the encoded target atomic feature and the encoded ligand atomic feature, the second atomic pair feature is the concatenation result of the encoded target atomic pair feature and the encoded ligand atomic pair feature, the output atomic feature is the complex atomic feature, and the output atomic pair feature is the complex atomic pair feature.

[0103] For example, the structures of the first encoder and the second encoder are as follows: Figure 3 As shown in the figure, 31 represents the input atomic feature, i.e., the first atomic feature; 32 represents the input atom pair feature, i.e., the second atomic feature; Transformer enconderlayer 33 represents the Transformer encoding layer; Outer Product 34 represents the outer product module; Triangular SelfAttention 35 represents the triangular attention module; 36 represents the output atomic feature, i.e., the second atomic feature; 37 represents the output atom pair feature, i.e., the second atom pair feature; and 38 represents the new second atomic feature. There are six Transformer enconderlayers 33.

[0104] Furthermore, the input data of the first Transformer enconder layer 33 is the first atomic feature 31, and the second atomic feature 32 serves as the bias item of the self-attention module in the first Transformer enconder layer. The self-attention module outputs an attention weight a, and the first Transformer enconder layer 33 outputs a new first atomic feature 3-1.

[0105] The new first atomic feature 3-1 is input into the second Transformer enconder layer 33, and the second atomic feature 32 is used as the bias item of the self-attention module in the second Transformer enconder layer 33. The self-attention module outputs an attention weight b, and the attention weight b is added to the attention weight a to obtain the attention weight c. The second Transformer enconder layer 33 outputs the new first atomic feature 3-2.

[0106] The new first atomic feature 3-2 is input into the third Transformer enconder layer 33. The second atomic feature 32 is used as the bias item of the self-attention module in the third Transformer enconder layer 33. The self-attention module outputs an attention weight d. The attention weight d is added to the attention weight c to obtain the attention weight e. The third Transformer enconder layer 33 outputs the new first atomic feature 3-3.

[0107] The new first atomic feature 3-3 is input into the fourth Transformer enconder layer 33, and the second atomic feature 32 is used as the bias item of the self-attention module in the fourth Transformer enconder layer 33. The self-attention module outputs an attention weight f. The attention weight f is added to the attention weight e to obtain the attention weight g. The fourth Transformer enconder layer 33 outputs the new first atomic feature 3-4.

[0108] The new first atomic feature 3-4 is input into the fifth Transformer enconder layer 33. The second atomic feature 32 is used as the bias item of the self-attention module in the fifth Transformer enconder layer 33. The self-attention module outputs an attention weight h. The attention weight h is added to the attention weight g to obtain the attention weight i. The fifth Transformer enconder layer 33 outputs the new first atomic feature 3-5.

[0109] The new first atomic feature 3-5 is input into the sixth Transformer enconder layer 33, and the second atomic feature 32 is used as the bias item of the self-attention module in the sixth Transformer enconder layer 33. The self-attention module outputs an attention weight j. The attention weight j is added to the attention weight i and the second atomic feature 32 to obtain a new second atomic feature 38. The sixth Transformer enconder layer 33 outputs the new first atomic feature, which is the second atomic feature.

[0110] In the above process, the first atom pair feature 32 is used as the bias item of the self-attention module in each Transformer enconder layer 33. That is, the first atom pair feature is introduced into each Transformer enconder layer 33, which is conducive to the Transformer enconder layer 33 to better learn the atomic features and atom pair features, and obtain output data with higher accuracy and stronger expressiveness.

[0111] Furthermore, the second atomic feature 36 is input into the outer product module, that is, the second atomic feature is used as the outer product as a bias term, and is input into the Triangular Self Attention 35 with the new first atomic pair feature 38, and the output data of the Triangular Self Attention 35 is used as the second atomic pair feature 37. Among them, the detailed structure of Triangular Self Attention 35 can refer to the four modules of triangle update using "outgoing" edges, triangle update using "incoming" edges, triangle self-attention around starting node, and triangle self-attention around ending node in AlphaFold2. The introduction of Transformer enconder layer 33 realizes better interaction between atomic features and atomic pair features, which is conducive to obtaining output data with higher accuracy and stronger expressiveness. For the decoder, Figure 3 In the code, replace Transformer enconder layer 33 with Transformer deconder layer.

[0112] Step 132 : determining a loss value based on the first predicted data and the first composite data, wherein the composite data includes at least a second composite data conformation.

[0113] In this step, after determining the first predicted data, a loss value is determined based on the first predicted data and the first complex data. Because the first predicted data at least includes the predicted complex conformation, the first complex data accordingly at least includes the simulated complex conformation. Therefore, the predicted complex conformation can be used as the predicted data, and the simulated complex conformation as the label data, and the loss value can be obtained by substituting them into a preset loss function.

[0114] In some embodiments, the first complex data further includes a docking score and / or a simulated energy term, wherein the simulated energy term is used to calculate the docking score. Specifically, when the first prediction data includes a predicted score, the first complex data includes the docking score; and when the first prediction data includes a predicted energy term, the first complex data includes the simulated energy term.

[0115] In one possible implementation, the first target data and the first ligand data are input into an initial model to obtain a predicted complex conformation, and a loss value is determined based on the predicted complex conformation and a simulated complex conformation corresponding to the first complex data.

[0116] In one possible implementation, the first target data and the first ligand data are input into an initial model to obtain a predicted complex conformation and a predicted score, and a penalty value is determined based on the predicted complex conformation, the predicted score, the simulated complex conformation corresponding to the first complex data, and the docking score.

[0117] In one possible implementation, first target data and first ligand data are input into an initial model to obtain a predicted complex conformation and a predicted energy term, and a loss value is determined based on the predicted complex conformation, the predicted energy term, and a simulated complex conformation and simulated energy term corresponding to the first complex data.

[0118] In one possible implementation, first target data and first ligand data are input into an initial model to obtain a predicted complex conformation, a predicted score, and a predicted energy term. A penalty value is determined based on the predicted complex conformation, predicted score, predicted energy term, and a simulated complex conformation, docking score, and simulated energy term corresponding to the first complex data.

[0119] Step 133: Adjust the model parameters of the initial model based on the loss value to train an intermediate model.

[0120] In this step, after determining the loss value, the model parameters of the initial model are adjusted using the loss value to train an intermediate model, which is the model obtained after pre-training.

[0121] In some embodiments, step 13, training an intermediate model based on the first target data, the first ligand data, and the first complex data, comprises:

[0122] Step 134 : applying first noise to the first ligand data to obtain third ligand data.

[0123] In this step, noise processing is performed on the first ligand data, that is, first noise is applied to the first ligand data to obtain third ligand data.

[0124] Step 135 : Input the first target data and the third ligand data into an initial model to obtain first prediction data and second noise.

[0125] In this step, the first target data and the third ligand data are input into the initial model to obtain first predicted data and second noise, wherein the first predicted data at least includes the predicted complex conformation and may also include a predicted score and / or a predicted energy term. The second noise is the noise data predicted by the initial model.

[0126] For example, Figure 4 As shown, the input ligand data in this case is the third ligand data. After the ligand atomic features and ligand atom pair features are input into the ligand encoder, noise prediction is performed based on the encoded ligand atomic features and encoded ligand atom pair features output by the ligand encoder. That is, the encoded ligand atomic features and encoded ligand atom pair features are input into the noise predictor to obtain the second noise. Simultaneously, a prediction score is obtained according to the affinity predictor, and a predicted complex conformation is obtained according to the complex conformation predictor.

[0127] Step 136 : determining a loss value based on the first predicted data, the first composite data, the first noise, and the second noise.

[0128] In this step, the first prediction data and the second noise are the prediction data of the initial model, and the first composite data and the first noise are the label data. Then, the loss value is accurately calculated based on the first prediction data, the first composite data, the first noise and the second noise.

[0129] Step 137: Adjust the model parameters of the initial model based on the loss value to train an intermediate model.

[0130] In this step, after determining the loss value, the model parameters of the initial model are adjusted according to the loss value to train an intermediate model. In this embodiment, the first ligand data is subjected to noise processing, and the noise is predicted to enhance the learning performance of the model and improve the model performance.

[0131] Step 14: Acquire second target data, second ligand data, and second complex data, where the second complex data is experimental data corresponding to the second target data and the second ligand data.

[0132] In this step, the second target data, the second ligand data, and the second complex data are obtained. The second complex data is the experimental data corresponding to the second target data and the second ligand data. Specifically, the pdbbind data set can be used to obtain real experimental data, or the real experimental data can be obtained using own experimental data. It should be noted that pdbbind is an experimentally derived complex data set, which contains both complex conformation data and affinity data. The complex conformation data includes the target protein conformation and the ligand small molecule conformation. Therefore, the first target data can be obtained through the pdbbind data set, and the second complex data can also be obtained through the pdbbind data set. Step 15, based on the second target data, the second ligand data and the second complex data, the intermediate model is adjusted to obtain a complex data prediction model.

[0133] In this step, after acquiring real experimental data, the intermediate model is adjusted using the real experimental data to obtain an adjusted complex data prediction model. This complex data prediction model is used to predict the complex data to obtain at least one complex data item, wherein the at least one predicted complex data item includes at least a predicted complex conformation and may also include predicted affinity data and / or predicted capacity items. Specifically, the second target data and the second ligand data are input into the intermediate model to obtain third predicted data. Based on the real data corresponding to the third predicted data and the second complex data, a loss value is determined; and based on this loss value, the model parameters of the intermediate model are adjusted to obtain the complex data prediction model.

[0134] In one possible implementation, the second target data and the second ligand data are input into an intermediate model to obtain a complex conformation predicted by the intermediate model. A loss value is determined based on the complex conformation predicted by the intermediate model and the actual complex conformation corresponding to the second complex data. Model parameters of the intermediate model are adjusted based on the loss value to obtain a complex data prediction model.

[0135] In one possible implementation, the second target data and the second ligand data are input into an intermediate model to obtain complex conformation and predicted affinity data predicted by the intermediate model. A loss value is determined based on the complex conformation and predicted affinity data predicted by the intermediate model and the actual complex conformation and affinity data corresponding to the second complex data. The model parameters of the intermediate model are adjusted based on this loss value to obtain a complex data prediction model. The intermediate model predicts complex conformation and affinity, and a loss value is constructed based on the difference between the actual conformation and affinity in the experimental data and the predicted results. The model parameters are then optimized to obtain a highly accurate complex data prediction model.

[0136] In one possible implementation, the second target data and the second ligand data are input into an intermediate model to obtain a complex conformation and energy term predicted by the intermediate model. A loss value is determined based on the complex conformation and energy term predicted by the intermediate model and the true complex conformation and true energy term corresponding to the second complex data. Model parameters of the intermediate model are adjusted based on the loss value to obtain a complex data prediction model.

[0137] In one possible implementation, the second target data and the second ligand data are input into an intermediate model to obtain a complex conformation, predicted affinity data, and energy term predicted by the intermediate model. A loss value is determined based on the complex conformation, predicted affinity data, and energy term predicted by the intermediate model, and the actual complex conformation, affinity data, and energy term corresponding to the second complex data. The model parameters of the intermediate model are adjusted based on this loss value to obtain a complex data prediction model. This complex data prediction model can be used for virtual screening, effectively shortening virtual screening time, improving virtual screening efficiency, and achieving high screening accuracy.

[0138] In some embodiments, the method further comprises:

[0139] Step 16: Obtain fixed data, and obtain the data to be screened corresponding to the fixed data.

[0140] Step 17: input the target data to be screened and the fixed data into the complex data prediction model to obtain second predicted data.

[0141] Step 18: Determine a screening result from the data to be screened based on the second prediction data.

[0142] In this example, a trained composite data prediction model is put into use online for data screening. Specifically, fixed data (predetermined data) is obtained, and data to be screened corresponding to the fixed data is determined. The data to be screened is the data subject to be screened. The data to be screened and the fixed data are input into the composite data prediction model to obtain second predicted data output by the composite data prediction model. The second predicted data is then used to determine the screening result within the data to be screened.

[0143] Among them, when the fixed data is fixed target data, the data to be screened is ligand data to be screened, and the screening result is target ligand data; when the fixed data is fixed ligand data, the data to be screened is target data to be screened, and the screening result is target target data; when the fixed data has no data, the data to be screened is ligand data to be screened and target data to be screened, and the screening result is target ligand data and target target data.

[0144] Furthermore, the second prediction data includes at least the complex conformation predicted by the complex data prediction model, and also includes an energy term and a docking score predicted by the complex data prediction model. Based on the docking scores included in the second prediction data or the docking scores calculated based on the energy terms included in the second prediction data, the candidate data are sorted from highest to lowest according to their docking scores, and the candidate data with the highest ranking are selected as the screening results, thereby achieving higher screening efficiency, shorter screening time, and higher accuracy.

[0145] In the above embodiment, after obtaining the first target data and the first ligand data, the first target data and the first ligand data are docked to obtain simulated first complex data, and then pre-training is performed based on the first target data, the first ligand data, and the first complex data to train an intermediate model. Then, real experimental data, namely the second target data, the second ligand data, and the second complex data, are obtained, and the intermediate model is adjusted using the second target data, the second ligand data, and the second complex data to obtain a complex data prediction model. In the technical solution provided by the present invention, pre-training is first performed using simulated data to obtain an intermediate model, and then the intermediate model is fine-tuned using real experimental data to obtain a complex data prediction model. The model performance of the complex data prediction model is relatively high, and a high-performance complex data prediction model is obtained based on a small amount of experimental data. Further, virtual screening is performed using the complex data prediction model, which can realize automated virtual screening, avoid human intervention, and achieve high screening efficiency and accuracy.

[0146] Exemplary devices

[0147] Based on the same concept as the method embodiment of the present invention, the embodiment of the present invention also provides a training device for a composite data prediction model.

[0148] Figure 5 A schematic diagram of a structure of a training device for a composite data prediction model provided by an exemplary embodiment of the present invention is shown, comprising:

[0149] A first data acquisition module 51 is used to acquire first target data and first ligand data;

[0150] a complex simulation module 52, configured to dock the first target data with the first ligand data to obtain simulated first complex data;

[0151] A pre-training module 53 is configured to train an intermediate model based on the first target data, the first ligand data, and the first complex data;

[0152] A second data acquisition module 54 is configured to acquire second target data, second ligand data, and second complex data, where the second complex data is experimental data corresponding to the second target data and the second ligand data;

[0153] The model adjustment module 55 is configured to adjust the intermediate model based on the second target data, the second ligand data, and the second complex data to obtain a complex data prediction model.

[0154] In an exemplary embodiment of the present invention, the pre-training module includes:

[0155] a prediction processing unit, configured to input the first target data and the first ligand data into an initial model to obtain first prediction data, wherein the first prediction data at least includes a predicted complex conformation;

[0156] a loss value determining unit, configured to determine a loss value based on the first predicted data and the first complex data, the complex data including at least a simulated complex conformation;

[0157] A model adjustment unit is used to adjust the model parameters of the initial model based on the loss value to train an intermediate model.

[0158] In an exemplary embodiment of the present invention, the prediction processing unit includes:

[0159] a first processing subunit, configured to input the first target data into the target embedding layer of the initial model to obtain target atomic features and target atom pair features; and input the first ligand data into the ligand embedding layer of the initial model to obtain ligand atomic features and ligand atom pair features;

[0160] a second processing subunit, configured to input the target atomic features and the target atom pair features into a first encoder to obtain encoded target atomic features and encoded target atom pair features; and input the ligand atomic features and the ligand atom pair features into a second encoder to obtain encoded ligand atomic features and encoded ligand atom pair features;

[0161] a third processing subunit, configured to input the encoded target atomic features, the encoded target atom pair features, the encoded ligand atomic features, and the encoded ligand atom pair features into a decoder to obtain complex atomic features and complex atom pair features;

[0162] The fourth processing subunit is configured to input the complex atom features and the complex atom pair features into a predictor to obtain first prediction data.

[0163] In an exemplary embodiment of the present invention, the first encoder, the second encoder, and the decoder each include a plurality of Transformer encoding layers, an outer product module, and a triangular attention module; when the input data of the first encoder, the second encoder, and the decoder are regarded as first atomic features and first atom pair features, and the output data are regarded as second atomic features and second atom pair features, the first encoder, the second encoder, and the decoder are configured to perform the following steps:

[0164] Inputting the first atomic feature into the first Transformer coding layer, using the first atomic pair feature as a bias item of the self-attention module in each Transformer coding layer, each self-attention module outputs an attention weight, and adding the attention weight output by each self-attention module and the first atomic pair feature as a new first atomic pair feature; the first Transformer coding layer outputs a new first atomic feature, and the new first atomic feature is input into the next Transformer coding layer until the last Transformer coding layer outputs the second atomic feature;

[0165] The second atomic feature is input into the outer product module, and the output data of the outer product module and the new first atom pair feature are input into the triangular attention module to obtain the second atom pair feature output by the triangular attention module.

[0166] In an exemplary embodiment of the present invention, the first prediction data further includes a prediction score and / or a prediction energy item, and the prediction energy item is used to calculate the prediction score;

[0167] In the case where the first prediction data includes a prediction score, the complex data further includes a docking score;

[0168] In the case where the first prediction data includes a predicted energy term, the complex data further includes a simulated energy term, and the simulated energy term is used to calculate the docking score.

[0169] In an exemplary embodiment of the present invention, the prediction processing unit is further configured to apply first noise to the first ligand data to obtain third ligand data; input the first target data and the third ligand data into an initial model to obtain first prediction data and second noise;

[0170] The loss value determining unit is further configured to determine a loss value based on the first prediction data, the first composite data, the first noise, and the second noise.

[0171] In an exemplary embodiment of the present invention, the apparatus further comprises:

[0172] The model application module is used to obtain fixed data and obtain data to be screened corresponding to the fixed data; input the data to be screened and the fixed data into the composite data prediction model to obtain second prediction data; and determine a screening result in the data to be screened based on the second prediction data.

[0173] Exemplary electronic devices

[0174] Figure 6 A block diagram of an electronic device according to an embodiment of the present invention is illustrated.

[0175] like Figure 6 As shown, the electronic device 60 includes one or more processors 61 and a memory 62 .

[0176] The processor 61 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 60 to perform desired functions.

[0177] The memory 62 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), a hard disk, or flash memory. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 61 may execute the program instructions to implement the composite data prediction model training methods and / or other desired functions described above in various embodiments of the present invention.

[0178] In one example, the electronic device 60 may further include an input device 63 and an output device 64 , and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0179] Of course, to simplify, Figure 6 Only some of the components related to the present invention in the electronic device 60 are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, the electronic device 60 may further include any other appropriate components according to specific application scenarios.

[0180] Exemplary computer program products and computer-readable storage media

[0181] In a sixth aspect, in addition to the above-mentioned methods and devices, an embodiment of the present invention may also be a computer program product, which includes computer program instructions, which, when executed by a processor, enable the processor to execute the steps of the training method of the composite data prediction model according to various embodiments of the present invention described in the above-mentioned "Exemplary Method" section of this specification.

[0182] The computer program product may be written in any combination of one or more programming languages ​​to implement the operations of embodiments of the present invention, including object-oriented programming languages ​​such as Java, C++, and conventional procedural programming languages ​​such as C or similar programming languages. The program code may be executed entirely on the user's computing device, partially on the user's computing device, as a stand-alone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0183] In addition, an embodiment of the present invention may also be a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, causes the processor to execute the steps of the method for training a composite data prediction model according to various embodiments of the present invention described in the above “Exemplary Method” section of this specification.

[0184] The computer-readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can, for example, include but is not limited to a system, device or component of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0185] The basic principles of the present invention have been described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, and effects mentioned in the present invention are merely illustrative and non-limiting, and should not be construed as necessarily possessed by each embodiment of the present invention. Furthermore, the specific details of the invention described above are provided for illustrative purposes and to facilitate understanding, and are not intended to be limiting. These details do not necessarily limit the present invention to being implemented using these specific details.

[0186] The block diagrams of the devices, apparatuses, equipment, and systems involved in the present invention are intended to be illustrative examples only and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As will be appreciated by those skilled in the art, these devices, apparatuses, equipment, and systems may be connected, arranged, or configured in any manner. Words such as "include," "comprise," "have," and the like are open-ended words, meaning "including but not limited to," and may be used interchangeably therewith. The words "or" and "and" used herein refer to the words "and / or" and may be used interchangeably therewith, unless the context clearly indicates otherwise. The word "such as" used herein refers to the phrase "such as but not limited to," and may be used interchangeably therewith.

[0187] It should also be noted that in the apparatus, device and method of the present invention, each component or each step can be decomposed and / or recombined, and such decomposition and / or recombination should be regarded as equivalent solutions of the present invention.

[0188] The above description of the invented aspects is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the present invention. Therefore, the present invention is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features of the invention herein.

[0189] The above description has been presented for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the present invention to the forms invented herein. Although a number of example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for training a composite data prediction model, characterized in that: include: Acquiring first target data and first ligand data, wherein the first target data is a target protein conformation; docking the first target data and the first ligand data to obtain simulated first complex data; Training an intermediate model based on the first target data, the first ligand data, and the first complex data; Acquiring second target data, second ligand data, and second complex data, where the second complex data is experimental data corresponding to the second target data and the second ligand data; Adjusting the intermediate model based on the second target data, the second ligand data, and the second complex data to obtain a complex data prediction model; The step of training an intermediate model based on the first target data, the first ligand data, and the first complex data includes: Inputting the first target data and the first ligand data into an initial model to obtain first predicted data, wherein the first predicted data at least includes a predicted complex conformation; determining a loss value based on the first predicted data and the first complex data, the first complex data including at least a simulated complex conformation; The model parameters of the initial model are adjusted based on the loss value to train an intermediate model.

2. The method according to claim 1, characterized in that The step of inputting the first target data and the first ligand data into an initial model to obtain first prediction data comprises: Inputting the first target data into the target embedding layer of the initial model to obtain target atomic features and target atom pair features; inputting the first ligand data into the ligand embedding layer of the initial model to obtain ligand atomic features and ligand atom pair features; Inputting the target atomic features and the target atom pair features into a first encoder to obtain encoded target atomic features and encoded target atom pair features; inputting the ligand atomic features and the ligand atom pair features into a second encoder to obtain encoded ligand atomic features and encoded ligand atom pair features; Inputting the encoded target atomic features, the encoded target atom pair features, the encoded ligand atomic features, and the encoded ligand atom pair features into a decoder to obtain complex atomic features and complex atom pair features; The complex atom features and the complex atom pair features are input into a predictor to obtain first prediction data.

3. The method according to claim 2, characterized in that The first encoder, the second encoder, and the decoder each include a plurality of Transformer encoding layers, an outer product module, and a triangular attention module; when the input data of the first encoder, the second encoder, and the decoder are regarded as first atomic features and first atom pair features, and the output data are regarded as second atomic features and second atom pair features, the first encoder, the second encoder, and the decoder are configured to perform the following steps: Inputting the first atomic feature into the first Transformer coding layer, using the first atomic pair feature as a bias item of the self-attention module in each Transformer coding layer, each self-attention module outputs an attention weight, and adding the attention weight output by each self-attention module and the first atomic pair feature as a new first atomic pair feature; the first Transformer coding layer outputs a new first atomic feature, and the new first atomic feature is input into the next Transformer coding layer until the last Transformer coding layer outputs the second atomic feature; The second atomic feature is input into the outer product module, and the output data of the outer product module and the new first atom pair feature are input into the triangular attention module to obtain the second atom pair feature output by the triangular attention module.

4. The method according to claim 1, wherein The first prediction data further includes a prediction score and / or a prediction energy item, wherein the prediction energy item is used to calculate the prediction score; In the case where the first prediction data includes a prediction score, the first complex data also includes a docking score; In the case where the first prediction data includes a predicted energy term, the first complex data further includes a simulated energy term, and the simulated energy term is used to calculate the docking score.

5. The method according to claim 1, wherein The step of inputting the first target data and the first ligand data into an initial model to obtain first prediction data comprises: applying first noise to the first ligand data to obtain third ligand data; Inputting the first target data and the third ligand data into an initial model to obtain first predicted data and second noise; The determining of the loss value based on the first prediction data and the first composite data includes: A loss value is determined based on the first predicted data, the first composite data, the first noise, and the second noise.

6. The method according to claim 1, characterized in that The method further comprises: Obtaining fixed data and obtaining data to be filtered corresponding to the fixed data; Inputting the data to be screened and the fixed data into the composite data prediction model to obtain second predicted data; Based on the second prediction data, a screening result is determined in the data to be screened.

7. A training device for a composite data prediction model, characterized in that: include: A first data acquisition module is used to acquire first target data and first ligand data, wherein the first target data is the conformation of the target protein; a complex simulation module, configured to dock the first target data with the first ligand data to obtain simulated first complex data; A pre-training module, configured to train an intermediate model based on the first target data, the first ligand data, and the first complex data; A second data acquisition module is configured to acquire second target data, second ligand data, and second complex data, wherein the second complex data is experimental data corresponding to the second target data and the second ligand data; a model adjustment module, configured to adjust the intermediate model based on the second target data, the second ligand data, and the second complex data to obtain a complex data prediction model; The pre-training module includes: a prediction processing unit, configured to input the first target data and the first ligand data into an initial model to obtain first prediction data, wherein the first prediction data at least includes a predicted complex conformation; a loss value determining unit, configured to determine a loss value based on the first prediction data and the first complex data, wherein the first complex data at least includes a simulated complex conformation; A model adjustment unit is used to adjust the model parameters of the initial model based on the loss value to train an intermediate model.

8. A computer-readable storage medium storing a computer program, wherein the computer program is used to execute the method for training a composite data prediction model according to any one of claims 1 to 6.

9. An electronic device, comprising: processor; a memory for storing instructions executable by the processor; The processor is configured to read the executable instructions from the memory and execute the instructions to implement the method for training a composite data prediction model according to any one of claims 1 to 6.

Images