Joint virtual screening method fusing global and local information

Abstract

The invention discloses a joint virtual screening method fusing global and local information, and belongs to the field of artificial intelligence drug discovery, and the method comprises the following steps: firstly, preparing an SMILES data set of active and inactive molecules of a target to be screened and an SMILES data set of random molecules; then, respectively training a global model and a local model of the GBDT by selecting a coding mode fusing substructures and materialization attribute information; predicting a to-be-screened data set by using the global model, and selecting TopK1 molecules with high prediction probability; predicting TopK1 molecules by using a local model, and selecting TopK2 molecules with a high product of the global model prediction probability and the local model prediction probability in the TopK1 molecules; and finally, training a regression model of the GBDT for predicting the IC50 value, and reordering the TopK2 molecules according to the predicted value. According to the method, the number of molecules predicted by the model is greatly reduced, the screening speed is high, the screening precision is high, the method is suitable for small-scale and super-large-scale virtual screening, and the cost and time of drug research and development are remarkably reduced.

Description

Technical Field

[0001] The present invention relates to a joint virtual screening method integrating global and local information, and belongs to the field of artificial intelligence drug discovery. Background Art

[0002] Virtual screening is a technology that uses computer simulation to efficiently screen a large number of compounds. Compared with traditional drug screening methods, its advantage is that it can evaluate the biological activity and pharmacological properties of tens of thousands of compounds through rapid computer simulation, thereby significantly shortening the drug development cycle, reducing R&D costs, and broadening the scope of drug exploration.

[0003] With the rapid increase in the number of molecules, which has exceeded 100 billion and will continue to grow at a high speed, it is particularly important to search for a larger space of drug candidate molecules. Ultra-large-scale virtual screening is therefore becoming more and more critical. Studies have shown that ultra-large-scale virtual screening can explore a vast candidate space, identify efficient candidate molecules and lead compounds, improve fault tolerance, and reduce false positive rates. Although the screening speed of molecular docking has been significantly improved compared to traditional experimental high-throughput screening, the existing molecular docking methods require a lot of time and resources when facing the screening of molecules that may reach billions or even tens of billions. We are in urgent need of more efficient virtual screening methods.

[0004] In recent years, the application of deep learning in the field of large-scale virtual screening has become a new solution. It is undeniable that deep learning models have achieved a series of remarkable achievements in the field of drug screening. However, there is a significant defect in deep learning models, that is, the model effect is usually positively correlated with the number of model parameters, while the model's inference speed is negatively correlated with the number of parameters. Therefore, it is difficult to find a deep learning model that has both extremely high screening accuracy and extremely fast screening speed. This problem is particularly prominent when facing ultra-large-scale virtual screening of billions or even tens of billions of molecules. Therefore, there is an urgent need for a new method that can take into account both screening accuracy and fast screening speed.

[0005] Gradient boosted decision tree (GBDT) is a method of ensemble learning in the field of machine learning, which is usually used to solve regression and classification problems. It builds a powerful model by connecting multiple decision trees in series. During the training process, each tree aims to correct the errors of the previous tree to gradually improve the performance of the model. GBDT is an excellent ligand-based screening method in the field of drug screening due to its excellent effect and extremely fast screening speed. The GBDT-based method provides an effective solution for ultra-large-scale virtual screening. Based on this, the present invention proposes a joint virtual screening method that integrates global and local information. Summary of the invention

[0006] In view of the defects and shortcomings of the above-mentioned prior art, the present invention proposes a joint virtual screening method that integrates global and local information, uses the global model of GBDT and the local model of GBDT for joint prediction, captures the global and local characteristics of the ligand molecule space respectively, integrates the two types of information for virtual screening, and then trains the IC50 regression model for reranking, so as to achieve the effect of improving the accuracy of virtual screening and ensuring the screening speed.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows: a joint virtual screening method integrating global and local information, comprising the following steps:

[0008] Step S100: preparing SMILES data sets of active and inactive molecules of the target to be screened and SMILES data sets of random molecules;

[0009] Step S200: Formulate a coding strategy for a SMILES expression;

[0010] Step S300: taking random molecules as negative samples and active molecules as positive samples, training a GBDT (gradient boosted decision tree) global model;

[0011] Step S400: Using active and inactive molecules as positive and negative samples to train the GBDT local model;

[0012] Step S500: Use the GBDT global model to predict the dataset to be screened and select the TopK according to the prediction probability. 1 molecular;

[0013] Step S600: Use the GBDT local model to predict the selected TopK 1 The numerator is the product of the prediction probability of the GBDT global model and the prediction probability of the GBDT local model. 1 The molecules are reordered and the TopK after sorting is selected 2 molecular;

[0014] Step S700: Use the IC50 activity data of positive and negative samples to train the GBDT regression model, and select the TopK 2 Rearrange the molecules to get the final result

[0015] As a further improvement of the present invention, the specific process of step S100 includes:

[0016] Select a target to be screened, prepare SMILES data sets of active and inactive molecules of the target to be screened and SMILES data sets of random molecules from public databases or other channels such as experiments, and select active and inactive molecules of the target to be screened and random molecules to meet the following requirements: all or part of the active and inactive molecules have IC50 activity data values, and the distribution of random molecules satisfies uniform distribution. Among them, IC50 (half maximal inhibitory concentration) refers to the half inhibitory concentration of the antagonist being measured. The smaller the IC50 value, the stronger the ligand activity.

[0017] As a further improvement of the present invention, the specific process of step S200 includes:

[0018] The encoding method combines Morgan fingerprint and physical and chemical properties. The physical and chemical properties can be the following attributes: Molecular weight, logP, TPSA, Total atom count, Bond count, Molecular refractivity, Positive Charge, Negative Charge, Formal Charge, Electronegative Atom Count, Halogen Count, NO Count, Sulfur Count, QED, HBA, HBD, Rotatable Bond Count, Ring Count, Aromatic Ring Count, fsp3, Heavy Atom Count, Aromatic proportion, Chiral Center Count, a total of 23 physical and chemical properties. Morgan fingerprint encoding uses multi-dimensional encoding, the encoding dimension is d, and Morgan fingerprint is expressed as M = [m 1 ,m 2 ,m 3 ,……,m d ], where m x Represents the substructure information of the molecule, x∈[1,d]; the physical and chemical properties are represented by P=[p 1 ,p 2 ,p 3 ,……,p 23 ], where p y Represents the value of the materialized attribute, y∈[1,23]; so the final encoding is represented as H=[m 1 ,m 2 ,m 3 ,……,m d ,p 1 ,p2 ,p 3 ,……,p 23 ].

[0019] Among them, the explanations of the above physical and chemical properties are as follows:

[0020] Molecular Weight: The relative molecular mass of a molecule.

[0021] LogP (oil-water partition coefficient): The logarithm of the octanol-water partition coefficient of a molecule, indicating the degree of hydrophobicity or lipophilicity of the molecule.

[0022] TPSA (Topological Polar Surface Area): The surface area of ​​polar atoms (usually nitrogen and oxygen) in a molecule, which is used to predict the solubility and permeability of the molecule.

[0023] Total Atom Count: The number of all atoms in a molecule.

[0024] Bond Count: The number of chemical bonds in a molecule.

[0025] Molecular Refractivity: Describes the ability of a molecule to scatter light and is related to the size and shape of the molecule.

[0026] Positive Charge: The total number of positive charges present in a molecule.

[0027] Negative Charge: The total number of negative charges present in a molecule.

[0028] Formal Charge: The sum of the formal charges of all atoms in a molecule.

[0029] Electronegative Atom Count: The number of atoms in a molecule that are more electronegative than hydrogen, such as oxygen, nitrogen, halogens, etc.

[0030] Halogen Count: The number of halogen atoms (fluorine, chlorine, bromine, iodine) in a molecule.

[0031] NO Count: The total number of nitrogen and oxygen atoms in a molecule.

[0032] Sulfur Count: The number of sulfur atoms in a molecule.

[0033] QED (Quantitative Estimate of Drug-likeness): A quality metric for evaluating a compound as a potential drug candidate.

[0034] HBA (Hydrogen Bond Acceptors): The number of atoms in a molecule that can accept hydrogen bonds, mainly nitrogen and oxygen atoms.

[0035] HBD (Hydrogen Bond Donors): The number of hydrogen atoms in a molecule that can form hydrogen bonds, usually connected to nitrogen and oxygen.

[0036] Rotatable Bond Count: The number of single bonds in a molecule that can rotate freely.

[0037] Ring Count: The number of ring structures in a molecule.

[0038] Aromatic Ring Count: The number of aromatic rings in a molecule.

[0039] fsp3 (Fraction of sp3-hybridized Carbon Atoms, the proportion of sp3 hybridized carbon atoms): the proportion of sp3 hybridized carbon atoms in all carbon atoms in a molecule.

[0040] Heavy Atom Count: The number of all atoms in a molecule except hydrogen.

[0041] Aromatic Proportion: The ratio of aromatic atoms (usually atoms participating in aromatic rings) to all heavy atoms in a molecule.

[0042] Chiral Center Count: The number of chiral centers in a molecule, that is, the number of atoms with stereoisomerism.

[0043] As a further improvement of the present invention, the specific process of step S300 includes:

[0044] Step S301: Select random molecules as negative samples and active molecules as positive samples to construct a data set. Active molecules with IC50 data labels are selected as high-quality molecules.

[0045] Step S302: 20% of the high-quality molecules in step S301 are extracted as the high-quality molecule validation set, and the remaining part of the active molecules after removing the high-quality molecule validation set is used as the remaining active molecule data set. The random molecules are divided into a random molecule training set and a random molecule validation set in a ratio of 8:2. Finally, the training set and validation set for hyperparameter search are obtained: the training set is composed of the random molecule training set and the remaining active molecule data set, and the validation set is composed of the random molecule validation set and the high-quality molecule validation set.

[0046] Step S303: Set the fixed hyperparameters of the GBDT global model, including the objective function, the type of boosting algorithm, and the number of iterations of GBDT. At the same time, set the evaluation indicators to monitor and evaluate the model performance. The model performs a binary classification task.

[0047] Step S304: Set the number of hyperparameter search rounds, the hyperparameters to be searched, and their corresponding search space. Use the training set and validation set obtained in step S302 to perform hyperparameter search, evaluate the model performance under different hyperparameter combinations on the validation set, and find the optimal hyperparameters.

[0048] Step S305: Using all random molecules and active molecules prepared in step S100 as training sets, and using the optimal hyperparameters obtained in step S304 to train a global binary classification model.

[0049] As a further improvement of the present invention, the specific process of step S400 includes:

[0050] Step S401: Select negative samples of inactive molecules and positive samples of active molecules of the target as data sets. Select molecules with IC50 data labels among active and inactive molecules as high-quality molecules.

[0051] Step S402: 20% of the high-quality molecules in step S401 are extracted as the high-quality molecule verification set. The remaining part of the active and inactive molecules minus the high-quality molecule verification set is called the residual molecule set. Finally, the training set and verification set for hyperparameter search are obtained: the training set is composed of the residual molecule set, and the verification set is composed of the high-quality molecule verification set.

[0052] Step S403, set the fixed hyperparameters of the GBDT local model, including the objective function, the type of boosting algorithm, and the number of iterations of GBDT. At the same time, set the evaluation index to monitor and evaluate the model performance. The model performs a binary classification task.

[0053] Step S404: Set the number of hyperparameter search rounds, the hyperparameters to be searched, and their corresponding search space. Use the training set and validation set obtained in step S402 to perform hyperparameter search, evaluate the model performance under different hyperparameter combinations on the validation set, and find the optimal hyperparameter.

[0054] Step S405: Using all active and inactive molecules based on the target prepared in step S100 as a training set, and using the optimal hyperparameters obtained in step S404 to train a local binary classification model.

[0055] As a further improvement of the present invention, the specific process of step S500 includes:

[0056] Use the global binary classification model to predict the data set to be screened, sort the prediction results from large to small, and select the TopK 1 of molecules.

[0057] As a further improvement of the present invention, the specific process of step S600 includes:

[0058] Use the local binary classification model to classify the TopK 1 The molecules are predicted, and the TopK 1 Rearrange the molecules from largest to smallest and select TopK 2 of molecules.

[0059] As a further improvement of the present invention, the specific process of step S700 includes:

[0060] Step S701: Select active and inactive molecules with IC50 experimental data as IC50 biological activity data sets.

[0061] Step S702: Divide the IC50 biological activity data set of step S701 into a training set and a validation set in a ratio of 8:2 for hyperparameter search.

[0062] Step S703: Set the fixed hyperparameters of the GBDT regression model, including the objective function, the type of boosting algorithm, and the number of iterations of GBDT. At the same time, specify the evaluation index to monitor and evaluate the model performance. The model performs the regression task, and the evaluation index is the mean square error MSE. The formula of MSE is as follows:

[0063]

[0064] Where: n is the number of samples, y i represents the true value of the i-th sample, It represents the predicted value of the i-th sample. The smaller the MSE value, the better the model fitting effect.

[0065] Step S704: Set the number of hyperparameter search rounds, the hyperparameters to be searched, and their corresponding search space. Use the training set and validation set obtained in step S702 to perform hyperparameter search, evaluate the model performance under different hyperparameter combinations on the validation set, and find the optimal hyperparameter.

[0066] Step S705: Using all active and inactive molecules with IC50 experimental data based on the target prepared in step S100 as a training set, the GBDT regression model for predicting IC50 values ​​is trained using the optimal hyperparameters obtained in step S704.

[0067] Step S706: Use the regression model to analyze the TopK 2 The molecule is predicted and the TopK 2 The molecules are sorted from small to large according to the regression prediction value to obtain the final sorting result. The top N molecules are selected as needed.

[0068] The present invention also provides a computer storage medium and an electronic device, wherein a computer program is stored on the computer storage medium, and when the computer program is executed by a processor, the steps of the method are implemented. The electronic device includes a memory and one or more processors, and the memory is used to store one or more programs; when the one or more programs are executed by the one or more processors, the method is implemented.

[0069] Compared with the prior art, the beneficial effects of the present invention are as follows: the molecular coding uses Morgan fingerprint coding and physicochemical property fusion coding, integrating the information of molecular substructure and physicochemical properties. The global model is first used for preliminary screening, which greatly reduces the number of molecules predicted by the model and reduces the time required for screening. The global model and the local model are used for joint prediction, and the global and local information are integrated to improve the screening accuracy. The regression model is used for re-ranking, and the information of IC50 experimental data is used to further improve the screening accuracy, which is suitable for small-scale and ultra-large-scale virtual screening, and significantly reduces the cost and time of drug development. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is the overall flow chart of the present invention.

[0071] Figure 2 This is the training flowchart of the global model.

[0072] Figure 3 Flowchart for hyperparameter search.

[0073] Figure 4 Flowchart for local model training.

[0074] Figure 5 Flowchart for regression model training. DETAILED DESCRIPTION

[0075] In order to better understand the content of the present invention, specific embodiments are described below with reference to the accompanying drawings.

[0076] The present invention is further described below in conjunction with the accompanying drawings. The following examples are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the scope of protection of the present invention. NMDA target is selected as an example target of the present method. Figure 1 , a joint virtual screening method integrating global and local information, comprising the following steps:

[0077] Step S100: preparing SMILES datasets of active and inactive molecules of NMDA targets and SMILES datasets of 1 million random molecules;

[0078] Step S200: Formulate a coding strategy for a SMILES expression;

[0079] Step S300: taking random molecules as negative samples and active molecules as positive samples, performing hyperparameter search to obtain optimal hyperparameters, and training the GBDT global model;

[0080] Step S400: Active and inactive molecules are used as positive and negative samples to perform hyperparameter search to obtain optimal hyperparameters and train the GBDT local model;

[0081] Step S500: Use the GBDT global model to predict the data set to be screened, and select the top 5000 molecules according to the prediction probability;

[0082] Step S600: Use the GBDT local model to predict the selected Top5000 molecules, re-rank the Top5000 molecules according to the product of the GBDT global model prediction probability and the GBDT local model prediction probability, and select the Top1000 molecules;

[0083] Step S700: Use the activity data of positive and negative sample IC50 to train the GBDT regression model, re-rank the selected Top1000 molecules according to the predicted values ​​of the GBDT regression model, and obtain the final result.

[0084] The following are the detailed execution steps of the above steps.

[0085] Executing step S100 includes the following specific steps:

[0086] Active and inactive molecules of NMDA targets were collected from the pubchem website, and 1 million random molecules were randomly sampled from the REAL database. A SMILES dataset of 10 million molecules was prepared for screening.

[0087] Execute step S200 to formulate a coding strategy for a SMILES expression, including the following specific steps:

[0088] The encoding method is a combination of Morgan fingerprint and physical and chemical properties. The physical and chemical properties can be the following attributes: Molecular weight, logP, TPSA, Total atom count, Bond count, Molecular refractivity, Positive Charge, Negative Charge, Formal Charge, Electronegative Atom Count, Halogen Count, NO Count, Sulfur Count, QED, HBA, HBD, Rotatable Bond Count, Ring Count, Aromatic Ring Count, fsp3, Heavy Atom Count, Aromatic proportion, Chiral Center Count, a total of 23 physical and chemical properties. Morgan fingerprint uses 1024-dimensional encoding. Morgan fingerprint is expressed as M = [m 1 ,m 2 ,m 3 ,……,m 1024 ], where m x Represents the substructure information of the molecule, x∈[1,1024]; the physical and chemical properties are represented by P=[p 1 ,p 2 ,p 3 ,……,p 23 ], where p y represents the value of the above materialized attribute, y∈[1,23]; so the final encoding is represented as H=[m 1 ,m 2 ,m 3 ,……,m 1024 ,p 1 ,p 2 ,p 3 ,……,p 23 ] is a 1047-dimensional encoding of . Morgan fingerprints and materialized properties can be calculated using Rdkit and Openbabel.

[0089] Among them, the explanations of the above physical and chemical properties are as follows:

[0090] Molecular Weight: The relative molecular mass of a molecule.

[0091] LogP (oil-water partition coefficient): The logarithm of the octanol-water partition coefficient of a molecule, indicating the degree of hydrophobicity or lipophilicity of the molecule.

[0092] TPSA (Topological Polar Surface Area): The surface area of ​​polar atoms (usually nitrogen and oxygen) in a molecule, which is used to predict the solubility and permeability of the molecule.

[0093] Total Atom Count: The number of all atoms in a molecule.

[0094] Bond Count: The number of chemical bonds in a molecule.

[0095] Molecular Refractivity: Describes the ability of a molecule to scatter light and is related to the size and shape of the molecule.

[0096] Positive Charge: The total number of positive charges present in a molecule.

[0097] Negative Charge: The total number of negative charges present in a molecule.

[0098] Formal Charge: The sum of the formal charges of all atoms in a molecule.

[0099] Electronegative Atom Count: The number of atoms in a molecule that are more electronegative than hydrogen, such as oxygen, nitrogen, halogens, etc.

[0100] Halogen Count: The number of halogen atoms (fluorine, chlorine, bromine, iodine) in a molecule.

[0101] NO Count: The total number of nitrogen and oxygen atoms in a molecule.

[0102] Sulfur Count: The number of sulfur atoms in a molecule.

[0103] QED (Quantitative Estimate of Drug-likeness): A quality metric for evaluating a compound as a potential drug candidate.

[0104] HBA (Hydrogen Bond Acceptors): The number of atoms in a molecule that can accept hydrogen bonds, mainly nitrogen and oxygen atoms. HBD (Hydrogen Bond Donors): The number of hydrogen atoms in a molecule that can form hydrogen bonds, usually connected to nitrogen and oxygen.

[0105] Rotatable Bond Count: The number of single bonds in a molecule that can rotate freely.

[0106] Ring Count: The number of ring structures in a molecule.

[0107] Aromatic Ring Count: The number of aromatic rings in a molecule.

[0108] fsp3 (Fraction of sp3-hybridized Carbon Atoms, the proportion of sp3 hybridized carbon atoms): the proportion of sp3 hybridized carbon atoms in all carbon atoms in a molecule.

[0109] Heavy Atom Count: The number of all atoms in a molecule except hydrogen.

[0110] Aromatic Proportion: The ratio of aromatic atoms (usually atoms participating in aromatic rings) to all heavy atoms in a molecule.

[0111] Chiral Center Count: The number of chiral centers in a molecule, that is, the number of atoms with stereoisomerism.

[0112] Combination Figure 2 , Figure 3 , step S300 includes the following specific steps:

[0113] Step S301: Select random molecules as negative samples and active molecules as positive samples to construct a data set. Active molecules with IC50 data labels are selected as high-quality molecules.

[0114] Step S302: 20% of the high-quality molecules in step S301 are extracted as the high-quality molecule validation set, and the remaining part of the active molecules minus the high-quality molecule validation set is used as the remaining active molecule data set. The random molecules are divided into a random molecule training set and a random molecule validation set in a ratio of 8:2. Finally, the training set and validation set for hyperparameter search are obtained: the training set is composed of the random molecule training set and the remaining active molecule data set, and the validation set is composed of the random molecule validation set and the high-quality molecule validation set.

[0115] Step S303: Select the Python package LightGBM to train the GBDT binary classification model. Select DART of GBDT (gradient boosted decision tree) as the model paradigm for training. The fixed hyperparameters of the decision tree are:

[0116]

[0117] "objective": "binary", indicating that GBDT is a binary classification model.

[0118] "boosting_type": "dart", which means using DART, an improved version of GBDT. DART introduces the concept of Dropout based on GBDT, that is, in each round of iteration, a part of the trees are randomly discarded, and then training is performed on the remaining trees. This method is similar to Dropout in neural networks and is used to prevent model overfitting.

[0119] "metric": "auc", indicating that the evaluation metric of the decision tree is AUC. AUC (Area Under Curve) is defined as the area under the ROC curve. It is often used to evaluate the classification effect of a binary classification model. As a numerical value, a classifier with a larger AUC has better effect. The number of iterations of GBDT is selected as 200.

[0120] Combination Figure 3 , step S304: perform hyperparameter search. The model effect of GBDT is deeply affected by the hyperparameters of the decision tree. This method performs hyperparameter search on the training set and validation set used for hyperparameter search, evaluates the evaluation indicators under different hyperparameters on the validation set, and finds the optimal hyperparameter. Set the number of hyperparameter search rounds to 300. Set the search hyperparameters and hyperparameter space. The hyperparameter search framework used is Ray Tune and the hyperparameter search toolkit Hyperopt. The following is an introduction to the tools:

[0121] Ray Tune: Ray Tune is an open source Python library for hyperparameter tuning of machine learning models. Ray Tune focuses specifically on automating and accelerating the hyperparameter optimization process during machine learning model training, providing an efficient and flexible way to find the optimal model parameters.

[0122] Hyperopt: Hyperopt is a Python library for optimizing hyperparameters of machine learning models. It has multiple built-in hyperparameter search strategies. The search algorithm we use is TPE (Tree of Parzen Estimators). TPE is a Bayesian optimization algorithm used to solve the global optimization problem of black box functions. The TPE algorithm can adaptively adjust the size of the parameter search space and find the global optimal solution in the fewest possible iterations. It is an excellent search algorithm. The following is the corresponding search space of the hyperparameter search.

[0123]

[0124] The following is an introduction to the hyperparameters:

[0125] learning_rate: defined as tune.uniform(0.05,0.1) means that the learning rate will be uniformly randomly selected from 0.05 to 0.1. The learning rate controls the step size of the model weight update. A smaller learning rate can make the model training more stable, but it may take more iterations to converge.

[0126] num_leaves: defined as tune.choice(list(range(20,501))) means that the maximum number of leaf nodes of the decision tree will be randomly selected between 20 and 500. Increasing the number of leaf nodes can improve the expressiveness of the model, but it may also lead to overfitting.

[0127] max_depth: defined as tune.choice(list(range(3,12))) means that the maximum depth of the decision tree will be randomly selected between 3 and 11. Limiting the maximum depth helps prevent overfitting.

[0128] min_child_samples: defined as tune.choice(list(range(15,50))) means that the minimum number of samples required to split a node will be randomly selected between 15 and 49. Increasing this value can reduce model complexity and avoid overfitting.

[0129] subsample: defined as tune.uniform(0.8,1.0) means that the random subsample ratio used in each tree training will be uniformly randomly selected from 0.8 to 1.0. This is a regularization strategy that can reduce overfitting.

[0130] colsample_bytree: Also defined as tune.uniform(0.8,1.0), which means that the feature subset ratio used in each tree training will also be uniformly randomly selected from 0.8 to 1.0. This is also to reduce overfitting.

[0131] lambda_l1 and lambda_l2: defined as tune.uniform(0,0.1), respectively, indicating that the L1 and L2 regularization strengths will be uniformly randomly selected from 0 to 0.1. These regularization terms help prevent overfitting.

[0132] bagging_freq: defined as tune.randint(1,8) means that the frequency of the bagging operation will be a random integer selected between 1 and 7. Bagging is an ensemble learning technique used to reduce model variance.

[0133] Step S305: Use all random molecules and active molecules prepared in step S100 as training sets, and use the optimal hyperparameters obtained in S304 to train the final global binary classification model. Use the Python package joblib to save the global binary classification model, and save it as global_model.joblib. The global model can extract the global information of the ligand space, identify the features that represent the global information in the molecular encoding, and can well screen out the suspected active molecules from the ultra-large-scale data set.

[0134] refer to Figure 4 , Figure 3 , step S400 includes the following specific steps:

[0135] Step S401: Select inactive molecule negative samples and active molecule positive samples as data sets. Select molecules with IC50 data labels among active and inactive molecules as high-quality molecules.

[0136] Step S402: 20% of the high-quality molecules in step S401 are extracted as the high-quality molecule verification set. The remaining part of the active and inactive molecules minus the high-quality molecule verification set is called the residual molecule set. Finally, the training set and verification set for hyperparameter search are obtained: the training set is composed of the residual molecule set, and the verification set is composed of the high-quality molecule verification set.

[0137] Step S403: Set the fixed hyperparameters of the GBDT local model, including the objective function, the type of boosting algorithm, and the number of iterations of GBDT. At the same time, set the evaluation indicators to monitor and evaluate the model performance. The model performs a binary classification task. The specific implementation method of this step is the same as step S303.

[0138] Step S404: Set the number of rounds of hyperparameter search, the hyperparameters to be searched, and their corresponding search space. Use the training set and validation set obtained in step S402 to perform hyperparameter search, evaluate the model performance under different hyperparameter combinations on the validation set, and find the optimal hyperparameter. The specific implementation method of this step is the same as step S304.

[0139] Step S405: Use all active and inactive molecules of the target prepared in step S100 as the training set, and use the optimal hyperparameters obtained in S404 to train the local binary classification model, and save it as local_model.joblib. The local model can extract local information in the molecular space, identify features that characterize local information in the molecular encoding, and identify active molecules in a data set that is similar to active samples and inactive samples.

[0140] Step S500 includes the following specific steps:

[0141] Use the global binary classification model to predict the dataset to be screened, save the results as global_result.csv file, sort the prediction results from large to small, select the top 5000 molecules, and save the results as global_select5k_result.csv. This step uses the global model for preliminary screening, reduces the number of model prediction samples, and increases the screening speed without reducing the screening accuracy.

[0142] Part of the global_select5k_result.csv file is shown below:

[0143] SMILES score_global Nc1cccc(C[P+](c2ccccc2)(c2ccccc2)c2ccccc2)c1 0.99249578 O=C(O)CCCC(=O)N1N=C(c2c(-c3ccc(Cl)cc3)c3ccccc3[nH]c2=O)CC1c1ccc(Br)cc1 0.99210174 Clc1ccc(C2NCC(c3ccccc3Cl)c3c2[nH]c2ccccc32)cc1 0.99197971 c1ccc(C2(c3ccccc3)CC2(c2ccccc2)c2ccccc2)cc1 0.99087373 c1ccc(C2Sc3sc4ccccc4[n+]3C2c2ccccc2)cc1 0.99014122

[0144] Step S600 includes the following steps:

[0145] Use the local binary classification model to predict the top 5000 molecules of S500, re-sort the top 5000 molecules from large to small according to the product of the prediction probability of the global binary classification model and the prediction probability of the local binary classification model, select the top 1000 molecules, and save the results as global_local_select1k_result.csv. This step can integrate global information and local information for secondary screening to ensure the accuracy of screening.

[0146] Part of the global_local_select1k_result.csv is shown below:

[0147] SMILES score_local*score_global Cc1nc(-c2cnn(C)c2-c2ccc(C(F)(F)F)cn2)c2c(N3CCC3)ncnn12 0.987228763 Clc1ccccc1C1Nc2c(ccc(Cl)c2Cl)C2C=CCC21 0.98612583 Cc1ccc2[nH]cc(C(=O)CN3CCC(Cc4ccccc4)CC3)c2c1 0.983618369 FC(F)(F)Oc1ccc2c(c1)C1C=CCC1C(c1ccc(Cl)cc1Cl)N2 0.982422218 O=C(c1cc2ccccc2[nH]1)N1CCC(Cc2ccccc2)CC1 0.981782345

[0148] Combination Figure 5 , Figure 3, execute step S700, the specific steps are as follows:

[0149] Step S701: Select active and inactive molecules with IC50 experimental data to construct an IC50 biological activity data set.

[0150] Step S702: Divide the IC50 biological activity data set of step S701 into a training set and a validation set in a ratio of 8:2 for hyperparameter search.

[0151] Step S703: Select the Python package LightGBM to train the GBDT regression model. The fixed hyperparameters of the decision tree are as follows:

[0152]

[0153] "objective": "regression", indicating that GBDT is a regression model.

[0154] "boosting_type":"dart", means using DART, an improved version of GBDT. DART introduces the concept of Dropout based on GBDT, that is, in each round of iteration, a part of the trees are randomly discarded, and then training is performed on the remaining trees. This method is similar to Dropout in neural networks and is used to prevent model overfitting.

[0155] "metric": "l2", L2 loss, also known as Mean Squared Error (MSE), is one of the commonly used evaluation indicators in regression tasks. It calculates the mean of the sum of squares of the difference between the predicted value and the true value. The formula of MSE is as follows:

[0156]

[0157] Where: n is the number of samples. i Represents the true value of the i-th sample. Represents the predicted value of the i-th sample. The smaller the MSE value, the better the model fitting effect. Finally, the number of iterations of GBDT is selected as 200.

[0158] Step S704: Set the number of hyperparameter search rounds, the hyperparameters to be searched, and their corresponding search spaces; use the training set and validation set obtained in step S702 to perform hyperparameter search, evaluate the model performance under different hyperparameter combinations on the validation set, and find the optimal hyperparameters. The specific implementation method of this step is the same as step S304.

[0159] Step S705: Using all active and inactive molecules with IC50 experimental data based on the target prepared in step S100 as a training set, the GBDT regression model for predicting IC50 values ​​is trained using the optimal hyperparameters obtained in step S704 and saved as regression_model.joblib.

[0160] Step S705: Use the regression model to predict the top 1000 molecules obtained in step S600, sort the top 1000 molecules from small to large according to the regression prediction value, and obtain the final sorting result, which is saved as regression_final.csv. Select the top N molecules for use as needed. This step integrates the information of the experimental IC50 data for sorting, which further improves the screening accuracy.

[0161] Part of the regression_final.csv file is shown below, where the IC50 is in nanomolar units:

[0162]

[0163]

[0164] The above specifically introduces a method and implementation of joint virtual screening that integrates global and local information provided by the present application. This article uses specific examples to elaborate on the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the method of the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A joint virtual screening method integrating global and local information, characterized in that: The following steps are involved: Step S100: preparing SMILES data sets of active and inactive molecules of the target to be screened and SMILES data sets of random molecules; Step S200: Formulate a coding strategy for a SMILES expression; Step S300: Use random molecules as negative samples and active molecules as positive samples to train the GBDT global model; Step S400: Using active and inactive molecules as positive and negative samples to train the GBDT local model; Step S500: Use the GBDT global model to predict the data set to be screened, and select the TopK1 molecule according to the prediction probability; Step S600: Use the GBDT local model to predict the selected TopK1 molecules, re-rank the TopK1 molecules according to the product of the GBDT global model prediction probability and the GBDT local model prediction probability, and select the sorted TopK2 molecules; Step S700: Use the IC50 activity data of positive and negative samples to train the GBDT regression model, and reorder the selected TopK2 molecules according to the predicted values ​​of the GBDT regression model to obtain the final result.

2. A joint virtual screening method integrating global and local information according to claim 1, characterized in that: The step S100 includes the following specific processes: Select a target to be screened, and prepare SMILES data sets of active and inactive molecules of the target to be screened and SMILES data sets of random molecules. The active and inactive molecules of the target to be screened and the random molecules are selected to meet the following requirements: all or part of the active and inactive molecules have IC50 activity data values, and the distribution of random molecules satisfies uniform distribution.

3. A joint virtual screening method integrating global and local information according to claim 1, characterized in that: The step S200 includes the following specific contents: The encoding adopts the encoding method combining Morgan fingerprint and physicochemical properties. The physicochemical properties adopt the following attributes: Molecular weight, logP, TPSA, Total atom count, Bond count, Molecular refractivity, Positive Charge, Negative Charge, Formal Charge, Electronegative Atom Count, Halogen Count, NO Count, Sulfur Count, QED, HBA, HBD, Rotatable Bond Count, Ring Count, Aromatic Ring Count, fsp3, Heavy Atom Count, Aromatic proportion, Chiral Center Count, a total of 23 physicochemical properties; Morgan fingerprint encoding adopts multi-dimensional encoding, the encoding dimension is d, and Morgan fingerprint is expressed as M=[m1,m2,m3,……,m d ], where m x Represents the substructure information of the molecule, x∈[1,d]; the physical and chemical properties are represented by P=[p1,p2,p3,……,p 23 ], where p y represents the value of the materialized attribute, y∈[1,23]; so the final encoding is represented by H=[m1,m2,m3,……,m d ,p1,p2,p3,……,p 23 ].

4. A joint virtual screening method integrating global and local information according to claim 1, characterized in that: The step S300 includes the following specific steps: Step S301: Select random molecules as negative samples and active molecules as positive samples, and construct a data set; select active molecules with IC50 data labels as high-quality molecules; Step S302: extract 20% of the high-quality molecules in step S301 as a high-quality molecule validation set, remove the remaining part of the active molecules from the high-quality molecule validation set as the remaining active molecule data set, divide the random molecules into a random molecule training set and a random molecule validation set in a ratio of 8:2, and finally obtain a training set and a validation set for hyperparameter search: the training set is composed of the random molecule training set and the remaining active molecule data set, and the validation set is composed of the random molecule validation set and the high-quality molecule validation set; Step S303: Set fixed hyperparameters of the GBDT global model, including the objective function, the type of boosting algorithm, and the number of iterations of GBDT; at the same time, specify evaluation indicators to monitor and evaluate model performance; the model performs a binary classification task; Step S304: Set the number of rounds of hyperparameter search, the hyperparameters to be searched, and their corresponding search space; use the training set and validation set obtained in step S302 to perform hyperparameter search, evaluate the model performance under different hyperparameter combinations on the validation set, and find the optimal hyperparameter; Step S305: Using all random molecules and active molecules prepared in step S100 as training sets, and using the optimal hyperparameters obtained in step S304 to train a global binary classification model.

5. A joint virtual screening method integrating global and local information according to claim 1, characterized in that: The step S400 includes the following specific steps: Step S401: Select inactive molecules as negative samples and active molecules as positive samples, construct a data set, and select molecules with IC50 data labels from active and inactive molecules as high-quality molecules; Step S402: 20% of the high-quality molecules in step S401 are extracted as a high-quality molecule verification set, and the remaining part of the active and inactive molecules minus the high-quality molecule verification set is called a residual molecule set, and finally a training set and a verification set for hyperparameter search are obtained: the training set is composed of the residual molecule set, and the verification set is composed of the high-quality molecule verification set; Step S403, setting fixed hyperparameters of the GBDT local model, including the objective function, the type of boosting algorithm, and the number of iterations of GBDT; at the same time, specifying evaluation indicators to monitor and evaluate model performance; the model performs a binary classification task; Step S404: Set the number of rounds of hyperparameter search, the hyperparameters to be searched, and their corresponding search space; use the training set and validation set obtained in step S402 to perform hyperparameter search, evaluate the model performance under different hyperparameter combinations on the validation set, and find the optimal hyperparameter; Step S405: Using all active and inactive molecules based on the target prepared in step S100 as a training set, and using the optimal hyperparameters obtained in step S404 to train a local binary classification model.

6. A joint virtual screening method integrating global and local information according to claim 1, characterized in that: The step S500 includes the following specific processes: Use the global binary classification model to predict the data set to be screened, sort the prediction results from large to small, and select the TopK1 molecules.

7. A joint virtual screening method integrating global and local information according to claim 1, characterized in that: The step S600 includes the following specific processes: The TopK1 molecules of step S500 are predicted using the local binary classification model, and the TopK1 molecules are reordered from large to small according to the product of the prediction probability of the global binary classification model and the prediction probability of the local binary classification model, and the TopK2 molecules are selected.

8. A joint virtual screening method integrating global and local information according to claim 1, characterized in that: The step S700 includes the following specific steps: Step S701: Select active and inactive molecules with IC50 experimental data to construct an IC50 bioactivity data set; Step S702: dividing the IC50 biological activity data set of step S701 into a training set and a validation set in a ratio of 8:2 for hyperparameter search; Step S703: Set the fixed hyperparameters of the GBDT regression model, including the objective function, the type of boosting algorithm, and the number of iterations of GBDT; at the same time, specify the evaluation index to monitor and evaluate the model performance; the model performs the regression task, and the evaluation index is the mean square error MSE. The formula of MSE is as follows: Where: n is the number of samples, y i represents the true value of the i-th sample, It represents the predicted value of the i-th sample. The smaller the MSE value is, the better the model fitting effect is. Step S704: Set the number of rounds of hyperparameter search, the hyperparameters to be searched, and their corresponding search space; use the training set and validation set obtained in step S702 to perform hyperparameter search, evaluate the model performance under different hyperparameter combinations on the validation set, and find the optimal hyperparameter; Step S705: using all active and inactive molecules with IC50 experimental data based on the target prepared in step S100 as a training set, and using the optimal hyperparameters obtained in step S704 to train a GBDT regression model for predicting IC50 values; Step S706: Use the regression model to predict the TopK2 molecules obtained in step S600, sort the TopK2 molecules in ascending order according to the regression prediction values, obtain the final sorting result, and select the top N molecules as needed.

9. A computer storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

10. An electronic device, characterized in that: The method comprises a memory and one or more processors, wherein the memory is used to store one or more programs; when the one or more programs are executed by the one or more processors, the method according to any one of claims 1 to 8 is implemented.

Images