A high-throughput virtual screening method based on deep active learning

Abstract

The present invention discloses a high-throughput virtual screening method based on deep active learning, in which each ASCII character of a SMILES string of a compound molecule is converted into a corresponding number as its starting encoding vector, and then a deep learning model is used to automatically learn the implicit features of the compound in the SMILES string. In combination with active learning technology, the amount of data required for annotation for model training is greatly reduced. Only 0.1% to 0.5% of the samples in the entire compound library need to be annotated, which can greatly reduce the time and training cost of annotating data sets and model training. In addition, the present invention adopts a deep learning model as a virtual screening engine, and the virtual screening accuracy is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of drug research and development, and specifically relates to a high-throughput virtual screening method based on deep active learning. Background Art

[0002] Drug R&D is characterized by high investment, high risk, and long cycle. Generally speaking, a drug R&D cycle is more than 10 years, with R&D investment in the hundreds of millions of dollars, and it is showing an upward trend year by year. With the emergence of on-demand commercial libraries, the number of compounds available for virtual screening and analysis has exploded, with several libraries exceeding 1 billion compounds. As screening libraries become larger and more diverse, the application of physics-based screening methods, where each molecule in the library must be independently enumerated and evaluated, is becoming increasingly costly.

[0003] The significant progress of deep learning in recent years has provided huge opportunities for drug development. Deep learning technology has also made important progress in the field of biomedicine. For example, AlphaFold2 can predict the tertiary structure of proteins with atomic precision. The pharmaceutical industry has begun to use deep learning technology to accelerate drug development and reduce R&D costs.

[0004] Previous studies have shown that deep learning technology has advantages over traditional machine learning methods in optimizing synthesis routes, predicting the pharmacokinetic properties of drugs, predicting drug targets, virtual screening, etc. However, deep learning is relatively data-dependent and requires a large amount of data supply to optimize a large number of parameters. Summary of the invention

[0005] In view of the deficiencies described in the above-mentioned prior art, the present invention provides a high-throughput virtual screening method based on deep active learning, which combines active learning technology to reduce the amount of labeled data required for model training, and can train a deep learning model with a smaller number of labeled samples, thereby reducing the time and training cost of labeling data sets and model training.

[0006] The technical solution adopted by the present invention is:

[0007] A high-throughput virtual screening method based on deep active learning, comprising:

[0008] Get Screening Learning Library I:

[0009] A large number of known compound molecules characterized by SMILES are downloaded from a large compound database and stored in the screening learning library I; the large compound database can be ZINC or other compound databases containing SMILES, each compound molecule has SMILES and a corresponding ID, and each unique ID corresponds to a unique SMILES.

[0010] Get the annotated molecule library:

[0011] Randomly screen compound molecules from the screening learning library I, use DOCK molecular docking software to dock and score each compound molecule with a given target, and store it in the annotated molecule library; the number of screened molecules depends on the processing capacity of the equipment used. Preferably, 0.1% to 0.5% of the compound molecules are randomly screened from the screening learning library I. The percentage is based on the total amount of the screening database I.

[0012] The labeled molecule library is divided into a total training set and a test set with a division ratio of 8:2 or 9:1.

[0013] Construct a deep learning model that can predict molecular docking scores: the input is a 128-dimensional vector and the output is a prediction score.

[0014] The deep learning model can be any deep learning model that predicts molecular docking scores.

[0015] Preferably, a deep convolutional neural network model is selected, including 3 convolutional Block units and 1 multi-layer perceptron (MLP) unit; each convolutional Block unit is composed of 5 Conv1d, 2 BatchNorm1d layers and 2 PReLU nonlinear activation layers;

[0016] The MLP unit consists of 3 fully connected layers (Linear), 2 Dropout layers, and 2 PReLU nonlinear activation layers.

[0017] Training a deep learning model:

[0018] K-fold cross validation is used to randomly divide the total training set into K equal parts, one of which is a validation set, and the remaining K-1 parts are training subsets, and the training subsets are combined into a training set; K different combinations of training sets and validation sets are obtained; the training set and validation set of each combination are subjected to data preprocessing and then input into the deep learning model in the form of embedded vectors in batches for training and evaluation, to obtain K trained deep learning models; preferably, 5-fold cross validation is used.

[0019] The test set is used to evaluate the predictions of the deep learning model:

[0020] After data preprocessing, the compound molecules in the test set were input into K trained deep learning models in batches in the form of embedding vectors, and the average of the K predicted docking scores of each compound molecule was taken as the final predicted docking score of the compound molecule; the dimension of the embedding vector was 128.

[0021] Calculate the root mean square error and Pearson coefficient of the final predicted docking score of each compound molecule and the annotation score of the compound molecule; when the root mean square error <0.7 and the Pearson coefficient >0.6, it means that the trained K deep learning models meet the basic requirements; whether the model parameters need to be adjusted should be modified according to the actual situation. The root mean square error <0.7 and the Pearson coefficient >0.6 are only the minimum standards, and better results can be obtained according to actual requirements.

[0022] Active Learning:

[0023] Predict and score the remaining compound molecules in the screening learning library I;

[0024] Update the total training set;

[0025] According to the predicted scores, the compound molecules for updating the total training set are screened according to the set screening rules and stored in the selected training set;

[0026] The selected training set was scored and annotated using DOCK molecular docking software and then merged with the original total training set to form an updated total training set;

[0027] Update training;

[0028] Update the total training set and train the K trained deep learning models again to obtain K deep learning models after active learning;

[0029] The test set performs prediction evaluation on the deep learning model after active learning;

[0030] Use the test set to evaluate the K deep learning models after active learning. If the evaluation result is suitable, stop active learning and obtain the final deep learning model. Otherwise, continue active learning.

[0031] As a preferred solution of the present invention, the training process for any combination in a total training set is:

[0032] Before entering the deep learning model, each data list belonging to the same batch in the training set is converted into an embedding vector; the dimension of the embedding vector is 128.

[0033] Use torch.nn.Embedding to convert all the lists of numbers in the current batch into embedding vectors;

[0034] Input the embedding vectors of the same batch into the deep learning model to train and obtain the predicted docking score of the current batch; calculate the loss function of the predicted docking score of the current batch and the corresponding annotation score and perform gradient backpropagation to update the deep learning model parameters;

[0035] After completing one epoch of training, a trained deep learning model is obtained;

[0036] Validation set for deep learning model prediction evaluation;

[0037] After data preprocessing, the validation set is input into the deep learning model in batches in the form of embedded vectors to obtain the predicted docking scores of the validation set. The root mean square error and Pearson coefficient are calculated by the predicted docking scores and the corresponding annotation scores to determine the training effect of the current deep learning model.

[0038] When the root mean square error is <0.7 and the Pearson coefficient is >0.6, the deep learning model is used as the deep learning model obtained by the current combined training.

[0039] As a preferred embodiment of the present invention, the screening rule adopted after the prediction and scoring is double screening, firstly, the remaining compound molecules in the screening learning library I are screened based on the final predicted docking score and stored in the candidate training set; then, the candidate training set is screened based on the prediction score variance.

[0040] Preferably, a set amount of compound molecules with good final predicted docking scores are selected as the training set to be selected; the smaller the final predicted score, the better; the set amount can be based on the total amount of the screening learning library I, preferably the top 5% of the compound molecules with good final predicted docking scores are selected;

[0041] A set amount of compound molecules with a large predicted score variance is selected from the candidate training set as the selected training set, wherein the larger the predicted score variance, the better; the set amount may also be based on the total amount of the screening learning library I, preferably selecting the top 0.1% of compound molecules with a large predicted score variance.

[0042] As a preferred solution of the present invention, the test set performs prediction evaluation on the deep learning model after active learning as follows:

[0043] After data preprocessing, the compound molecules in the test set are input into the K deep learning models after active learning in batches in the form of embedding vectors, and the average of the K predicted docking scores of each compound molecule is taken as the final predicted docking score of the compound molecule;

[0044] And calculate the root mean square error and Pearson coefficient between the final predicted docking score of each compound molecule and the annotation score of the compound molecule;

[0045] When the root mean square error is <0.7 and the Pearson coefficient is >0.6, it means that the K deep learning models after active learning meet the basic requirements, and active learning is stopped according to the set conditions.

[0046] As a preferred solution of the present invention, the data preprocessing includes:

[0047] Convert the SMILES of a compound molecule into a list of numbers:

[0048] The characters of the SMILES string of the compound molecule are converted into corresponding numbers one by one according to the dictionary, and then the SMILES of the compound molecule is converted into a number list of a set length; only the maximum length part is retained for those exceeding the maximum length of the number list, and the part less than the maximum length of the number list is filled.

[0049] Collection type conversion;

[0050] The converted data set is first converted to the TensorDataset type, and then converted to the DataLoader type. When input into the deep learning model, Dataloader is used to input it in batches.

[0051] The present invention converts each ASCII character of the SMILES string of the compound molecule into a corresponding number as its starting encoding vector, and then uses a deep learning model to automatically learn the implicit features of the compound in the SMILES string, and combines active learning technology to greatly reduce the amount of data required for model training. Only 0.1% to 0.5% of the samples in the entire compound library need to be annotated, which can greatly reduce the time and training cost of annotating data sets and model training. In addition, the present invention uses a deep learning model as a virtual screening engine for high accuracy virtual screening. The convolutional neural network model used in the present invention can be replaced with other deep learning models, and the method can also be adapted to other virtual screening tasks. The present invention is input into the deep learning model in batches, and high-throughput virtual screening can be quickly realized. From modeling training to virtual screening, all docking scoring and screening of hundreds of millions of compound libraries can be completed within two days. Compared with the scoring speed of 1 to 20s / piece of traditional DOCK software, this method is 6 orders of magnitude higher than it in prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. 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 creative work.

[0053] Figure 1 It is a flow chart of the present invention.

[0054] Figure 2 This is an overview of the convolutional neural network model of the present invention.

[0055] Figure 3This is a composition diagram of the convolutional Block unit in the convolutional neural network of the present invention.

[0056] Figure 4 It is a composition diagram of the MLP unit in the convolutional neural network of the present invention. DETAILED DESCRIPTION

[0057] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the 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 the present invention.

[0058] A high-throughput virtual screening method based on deep active learning, such as Figure 1 As shown, including:

[0059] Get Screening Learning Library I:

[0060] A large number of known drug-like molecules characterized by SMILES are screened from a large compound database and stored in the screening learning library I; in this embodiment, taking one target as an example, 400 million compound molecules are obtained from ZINC.

[0061] Get the annotated molecule library:

[0062] 0.1% of the compound molecules are randomly screened from the screening learning library I, that is, 400,000 compound molecules are randomly screened, and each compound molecule is docked with the given target using the DOCK molecular docking software, scored and annotated, and stored in the annotated molecule library; of course, the number of selected compound molecules can also be 0.2%, 0.3%, 0.4% or 0.5%, etc., and the number of screened molecules is related to the processing capacity of the equipment used.

[0063] The DOCK molecular docking software used may be professional software such as Glide, GOLD, MOE Dock, Ledock, Surflex-Dock, LigandFit, etc., and Ledock is used in the example.

[0064] The labeled molecule library is divided into a data set, which is divided into a total training set and a test set, with a division ratio of 8:2 or 9:1; 8:2 is selected in this embodiment.

[0065] Constructing a deep learning model that can predict molecular docking scores: the deep learning model input is a 128-dimensional vector, and the output is a prediction score;

[0066] This embodiment constructs a deep convolutional neural network model, such as Figure 2As shown, it includes 3 convolutional Block units and 1 multi-layer perceptron (MLP) unit; Figure 3 As shown, each convolutional Block unit consists of 5 Conv1d, 2 BatchNorm1d layers and 2 PReLU nonlinear activation layers;

[0067] like Figure 4 As shown in Figure 1, the multi-layer perceptron (MLP) unit consists of 3 fully connected layers (Linear), 2 Dropout layers, and 2 PReLU nonlinear activation layers, and the Dropout parameter in the Dropout layer is set to 0.5.

[0068] And the code implementations of Conv1d, BatchNorm1d, PReLU and Linear layers all come from the Torch.nn library.

[0069] Training a deep learning model:

[0070] In order to improve the prediction accuracy of the model, this application uses 5-fold cross-validation to randomly divide the total training set into 5 equal parts, one of which is a validation set, and the other 4 are training subsets, and each training subset is combined into a training set; 5 different combinations of training sets and validation sets are obtained, which are used to train the constructed deep convolutional neural network model respectively;

[0071] The training set and validation set of each combination are first preprocessed, that is, the SMILES of the compound molecules in the training set and validation set are converted into digital lists:

[0072] The characters of the SMILES string of the compound molecule are converted into corresponding numbers one by one according to a fixed dictionary, and then the SMILES of the compound molecule is converted into a digital list of a set length; the maximum length of the digital list is set to 100, and the size of the conversion dictionary is 64, that is, it contains 64 ASCII characters that can appear in SMILES. Only the maximum length part is retained if it exceeds the maximum length of the digital list, and the part less than the maximum length of the digital list is filled.

[0073] Convert the converted training set and validation set to the DataLoader type:

[0074] Specifically, it is converted to torch.utils.data.TensorDataset type first, and then converted to torch.utils.data.DataLoader type. The BatchSize parameter of Dataloader is set to 2560, shuffle is set to True, and num_worker is set to 4.

[0075] The training set after data preprocessing is input into the deep learning model in sequence according to the batch size using Dataloader for training, and a batch of a training set is first converted into a corresponding embedding vector using torch.nn.Embedding for each number in all the number lists in the batch. The embedding vector dimension is set to 128, and the character length of the compound molecule is 100. Then, the initial embedding vector size corresponding to the compound molecule is (100, 128), and the size of the entire batch is (2560, 100, 128).

[0076] The embedding vectors of the same batch are passed through three convolutional Block units and multi-layer perceptron MLP units in turn, and finally the prediction score of the entire batch is obtained, which is (2560,1); the predicted docking score of the current batch and the corresponding annotation score are calculated by loss function and gradient backpropagation, and the deep learning model parameters are updated; in this embodiment, the maximum Epoch number of training is set to 100, BatchSize is set to 2560, the initial learning rate is 0.001, the optimizer uses Adam, and the loss function uses torch.nn.MSELoss.

[0077] The training process for other batches is the same. When the entire training set is trained once, one epoch of training is completed.

[0078] Validation set for deep learning model prediction evaluation;

[0079] After data preprocessing, the validation set is input into the deep learning model in batches in the form of embedded vectors to obtain the predicted docking scores of the validation set. The root mean square error and Pearson coefficient are calculated by the predicted docking scores and the corresponding annotation scores to determine the training effect of the current deep learning model.

[0080] When the root mean square error is <0.7 and the Pearson coefficient is >0.6, it indicates that the prediction performance of the current deep learning model has met the basic requirements, and the deep learning model is used as the deep learning model obtained by the current combined training.

[0081] The test set is used to evaluate the predictions of the deep learning model:

[0082] After data preprocessing, the compound molecules in the test set were input into five trained deep learning models in batches in the form of embedded vectors, and the average of the five predicted docking scores of each compound molecule was taken as the final predicted docking score of the compound molecule.

[0083] Calculate the root mean square error and Pearson coefficient of the final predicted docking score of each compound molecule and the annotated score of the compound molecule; when the root mean square error <0.7 and the Pearson coefficient>0.6, it means that the 5 deep learning models trained meet the basic requirements; whether the model parameters need to be adjusted, according to the actual changes, the root mean square error <0.7 and the Pearson coefficient>0.6 are only the minimum standards, and better results can be trained according to the actual requirements. In this embodiment, after one round of training, the root mean square error of the test set is 0.3699, and the Pearson coefficient is 0.8028.

[0084] Active Learning:

[0085] Predict and score the remaining compound molecules in the screening learning library I. The prediction and scoring process is the same as the training process, which also requires data preprocessing and vector conversion and input into 5 deep learning models in batches;

[0086] Update the total training set;

[0087] According to the predicted scores, the compound molecules for updating the total training set are screened according to the set screening rules and stored in the selected training set;

[0088] The screening rule adopted after the prediction and scoring is double screening. First, the remaining compound molecules in the screening learning library I are screened based on the final predicted docking score and stored in the candidate training set; then, the candidate training set is screened based on the predicted score variance.

[0089] In this embodiment, the top 5% of compound molecules with good final predicted docking scores are selected as the candidate training set; because the smaller the final predicted score, the better, and the 5% here is based on the total amount of the screening learning library I;

[0090] Then, the top 0.1% of compound molecules with the largest predicted score variance are selected from the candidate training set as the selected training set. The 0.1% here is based on the total amount of the screening learning library I; and the larger the predicted score variance, the better. For the selected training set, the Ledock molecular docking software is used to score and annotate it, and then it is merged with the original total training set to form an updated total training set;

[0091] Update training;

[0092] The total training set is updated to train the five trained deep learning models again. The training process is the same as the training process of the total training set, and five deep learning models after active learning are obtained.

[0093] The test set performs prediction evaluation on the deep learning model after active learning;

[0094] The five deep learning models after active learning were evaluated using the test set. The compound molecules in the test set were preprocessed and then input into the five deep learning models after active learning in the form of batch embedded vectors. The average of the five predicted docking scores of each compound molecule was taken as the final predicted docking score of the compound molecule.

[0095] And calculate the root mean square error and Pearson coefficient between the final predicted docking score of each compound molecule and the annotation score of the compound molecule;

[0096] When the root mean square error <0.7 and the Pearson coefficient>0.6, it means that the five deep learning models after active learning meet the requirements, and choose whether to stop active learning or continue active learning in sequence according to the actual situation. Generally, active learning is performed up to 2 times, and active learning is stopped if the evaluation result is appropriate to obtain the final deep learning model. The five final deep learning models can quickly predict and score compound molecules and targets. In this embodiment, the root mean square error of the deep learning model obtained after one active learning on the test set is 0.3933, and the Pearson coefficient is 0.8986, indicating that the model prediction performance is excellent and has a good virtual screening effect.

[0097] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0098] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A high-throughput virtual screening method based on deep active learning, characterized in that: include: Get Screening Learning Library I: Obtain compound molecules characterized by SMILES from the compound database and store them in screening learning library I; The compound molecules in the screening learning library I are all characterized by SMILES; Get the annotated molecule library: Randomly screen compound molecules from the screening learning library I, use molecular docking software to dock and score each compound molecule with a given target, and store it in the annotated molecule library; The labeled molecule library is divided into a total training set and a test set; Constructing and training deep learning models: K-fold cross validation is used to randomly divide the total training set into K equal parts, one of which is a validation set, and the remaining K-1 parts are training subsets. The training subsets are combined into a training set; K different combinations of training sets and validation sets are obtained; the training set and validation set of each combination are preprocessed and then input into the deep learning model in the form of embedded vectors in batches for training and evaluation, and K trained deep learning models are obtained; The test set performs predictive evaluation on the deep learning model, and active learning is performed if the evaluation results are suitable; Active Learning: Predict and score the remaining compound molecules in the screening learning library I; Update the total training set; According to the predicted scores, the compound molecules for updating the total training set are screened according to the set screening rules and stored in the selected training set; The selected training set is scored and annotated using molecular docking software and then merged with the original total training set to form an updated total training set; Update training; Update the total training set and train the K trained deep learning models again to obtain K deep learning models after active learning; The test set performs prediction evaluation on the deep learning model after active learning; Use the test set to evaluate the K deep learning models after active learning. If the evaluation result is suitable, stop active learning and obtain the final deep learning model. Otherwise, continue active learning.

2. The high-throughput virtual screening method based on deep active learning according to claim 1, characterized in that: The test set performs predictive evaluation on the deep learning model including: After data preprocessing, the compound molecules in the test set are input into K trained deep learning models in batches in the form of embedded vectors, and the average of the K predicted docking scores of each compound molecule is used as the final predicted docking score of the compound molecule; The root mean square error and Pearson coefficient of the final predicted docking score of each compound molecule and the annotation score of the compound molecule were calculated; when the root mean square error <0.7 and the Pearson coefficient >0.6, it means that the trained K deep learning models meet the basic requirements.

3. The high-throughput virtual screening method based on deep active learning according to claim 1, characterized in that: The deep learning model is a deep convolutional neural network model, including 3 convolutional Block units and 1 multi-layer perceptron unit; each convolutional Block unit consists of 5 Conv1d, 2 BatchNorm1d layers and 2 PReLU nonlinear activation layers; The multi-layer perceptron unit consists of 3 fully connected layers, 2 Dropout layers, and 2 PReLU nonlinear activation layers.

4. The high-throughput virtual screening method based on deep active learning according to claim 1, characterized in that: The training process of a combination is: Each data list belonging to the same batch in the training set is converted into an embedding vector before being input into the deep learning model; Input the embedding vectors of the same batch into the deep learning model to train and obtain the predicted docking score of the current batch; calculate the loss function of the predicted docking score of the current batch and the corresponding annotation score and perform gradient backpropagation to update the deep learning model parameters; After completing one epoch of training, a trained deep learning model is obtained; Validation set for deep learning model prediction evaluation; After data preprocessing, the validation set is input into the deep learning model in batches in the form of embedded vectors to obtain the predicted docking scores of the validation set. The root mean square error and Pearson coefficient are calculated by the predicted docking scores and the corresponding annotation scores to determine the training effect of the current deep learning model. When the root mean square error is <0.7 and the Pearson coefficient is >0.6, it indicates that the prediction performance of the current deep learning model has met the basic requirements, and the deep learning model is used as the deep learning model obtained by the current combined training.

5. The high-throughput virtual screening method based on deep active learning according to claim 1, characterized in that: The screening rule is double screening, firstly, the remaining compound molecules in the screening learning library I are screened based on the final predicted docking score and stored in the candidate training set; and then the candidate training set is screened based on the predicted score variance.

6. The high-throughput virtual screening method based on deep active learning according to claim 5, characterized in that: Selecting a set amount of compound molecules with good final predicted docking scores as a candidate training set; the smaller the final predicted score, the better; A set amount of compound molecules with large prediction score variance are selected from the candidate training set as the selected training set, and the larger the prediction score variance, the better.

7. The high-throughput virtual screening method based on deep active learning according to claim 1, characterized in that: The test set performs prediction evaluation on the deep learning model after active learning as follows: After data preprocessing, the compound molecules in the test set are input into the K deep learning models after active learning in batches in the form of embedding vectors, and the average of the K predicted docking scores of each compound molecule is taken as the final predicted docking score of the compound molecule; And calculate the root mean square error and Pearson coefficient between the final predicted docking score of each compound molecule and the annotation score of the compound molecule; When the root mean square error is <0.7 and the Pearson coefficient is >0.6, it means that the K deep learning models after active learning meet the basic requirements, and active learning is stopped according to the set conditions.

8. The high-throughput virtual screening method based on deep active learning according to any one of claims 1 to 7, characterized in that: The data preprocessing includes: Convert the SMILES of a compound molecule into a list of numbers: Convert the characters of the SMILES string of the compound molecule into corresponding numbers one by one according to the dictionary, and then convert the SMILES of the compound molecule into a number list of a set length; Collection type conversion; Convert the converted dataset to the TensorDataset type first, and then to the DataLoader type.

9. The high-throughput virtual screening method based on deep active learning according to claim 1, characterized in that: The dimension of the embedding vector is 128.

10. The high-throughput virtual screening method based on deep active learning according to claim 1, characterized in that: 0.1% to 0.5% of compound molecules were randomly screened from the screening learning library I.

Images