Drug molecule backbone replacement and screening method based on deep transfer learning model

Abstract

This invention discloses a method for drug molecule scaffold replacement and screening based on a deep transfer learning model, belonging to the field of computer-aided drug development. The specific steps include: acquiring source and target domain datasets; extracting the molecular scaffolds of corresponding compounds in the source and target domains, as well as the compounds to be screened; pre-training a graph neural network model based on the source domain dataset; inputting the target domain dataset into the pre-trained network model, fine-tuning the network model parameters to obtain new parameters and the model; obtaining small molecule scaffolds based on comprehensive scoring; and replacing the molecular scaffolds to obtain a series of novel molecules. Based on D-MPNN and FNN, this invention introduces a transfer learning approach, combining a dual upstream and downstream strategy of drug target upstream and downstream and transfer learning task upstream and downstream. It extracts and fine-tunes the feature values ​​of the pre-trained network, achieving better training results on small sample datasets, and helping to establish novel molecular scaffold screening and replacement models for upstream and downstream targets with small sample sizes.

Description

Technical Field

[0001] This invention relates to the field of computer-aided drug development, and in particular to a method for drug molecule scaffold replacement and screening based on a deep transfer learning model. Background Technology

[0002] Drug development is a lengthy, costly, and low-success-rate process. The clinical failure rate from obtaining clinical trial approval to drug market launch exceeds 90%, and the cost of advancing a new therapeutic to market approval is estimated to be over $2 billion, with an average timeframe of 15 years. Early computer-aided drug design methods primarily used quantitative structure-activity relationship (QSAR) models, which utilize regression models to find connections between molecular descriptions and biological activities. Meanwhile, machine learning methods (such as support vector machines, random forests, and decision trees) have been applied to drug discovery tasks. However, due to limited feature extraction, overall prediction performance has been poor. Furthermore, the traditional concept of "scaffold transition" emphasizes two key components: the new compound has a different scaffold structure relative to the template compound but possesses similar biological activities. The principle of scaffold transition is that structurally similar molecules also have similar activities. However, traditional scaffold transition methods suffer from insufficient correlation between structural similarity and activity results, high difficulty and cost in synthesizing new scaffolds, and low scaffold novelty.

[0003] With the rapid advancement of artificial intelligence (AI) technology, AI-enabled drug development can significantly shorten the drug development cycle, substantially reduce enterprise time and R&D costs, and improve the success rate of drug development. Therefore, a new track of enterprise development based on AI + drug development is gradually emerging and developing. In recent years, the progress of deep learning, an emerging AI technology, has accelerated and improved the drug discovery process, achieving remarkable results in applications such as molecular property prediction, virtual screening, retrosynthetic analysis, and molecular generation models. Compared with traditional "shallow" machine learning methods, deep learning uses deep neural networks with multiple hidden layers, which can represent and learn more complex knowledge. Currently, introducing deep learning methods based on message-passing neural networks into backbone activity prediction is an effective way to solve the problems of traditional backbone transition methods. Chinese patent application number 201910818097X discloses a deep learning-based drug molecule generation method. This method learns the semantics and features of known drug-like molecules to obtain a pre-trained model, and then uses this pre-trained model to learn the features of a drug molecule for a specific target in the transfer training stage to obtain a transfer training model, generating molecular data. The performance of this deep learning method still largely depends on the size of the training data. Within a finite number of hypothesis classes, larger sample sizes tend to produce more accurate models. However, at each stage of the drug discovery process, the size of labeled data is typically small, ranging from tens of thousands to hundreds of thousands, due to the time-consuming and expensive experiments required to generate it. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method for drug molecule scaffold replacement and screening based on a deep transfer learning model. By combining a dual upstream and downstream strategy of drug target upstream and downstream and transfer learning task upstream and downstream, it solves the problems of insufficient accuracy, low convergence, and difficult training caused by the limitations of training deep learning models on small sample datasets.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0006] A method for drug molecule scaffold replacement and screening based on a deep transfer learning model includes the following steps:

[0007] Step 1: Obtain the source domain dataset and the target domain dataset;

[0008] Step 2: Convert the compounds to be screened, as well as the corresponding compounds in the source domain dataset and target domain dataset obtained in Step 1, into compound skeletons respectively.

[0009] Step 3: Input the simplified molecular linear input text of the compound skeleton corresponding to the source domain dataset obtained in Step 2 and the bioactivity data into the graph neural network model for pre-training to obtain the network model;

[0010] Step 4: Input the simplified molecular linear input canonical format text of the target domain dataset corresponding to the compound skeleton and the bioactivity data obtained in Step 2 into the pre-trained network model obtained in Step 3, and fine-tune the parameters in the network model to obtain new parameters and model.

[0011] Step 5: Based on the deep learning network model for the target domain dataset obtained in Step 4, perform comprehensive skeletal screening of the compounds to be screened, obtain a comprehensive molecular skeletal score, and screen out the molecular skeletal structures.

[0012] Step six: Replace the selected molecular skeletons into the target small molecule compound to obtain a series of new molecules with skeleton replacements.

[0013] Furthermore, the target domain dataset is a small sample dataset, while the source domain dataset is a large sample dataset related to the target domain task; the source domain dataset and the target domain dataset can be obtained from public databases and public literature such as Pubchem, ChEMBL, and BindingDB.

[0014] Furthermore, step two specifically includes: using the Murcko skeleton clustering compound library based in the open-source package RDkit to extract the skeleton of source domain compounds, target domain compounds, and compounds to be screened.

[0015] Furthermore, the graph neural network model includes a directed message passing neural network module and a feedforward neural network module. The directed message passing neural network module is used to extract molecular features, and the feedforward neural network module is used to classify and predict molecular properties.

[0016] Furthermore, before inputting the simplified molecular linear input text (i.e., SMILE text) of the compound skeleton corresponding to the source domain dataset into the graph neural network model, the open-source package RDkit is used to calculate the SMILE text of the corresponding compound in the input source domain dataset to obtain the atomic features x of the corresponding small molecule compound. v and chemical bond characteristics e vw , as the initial input features of the graph neural network model;

[0017] In the directional message passing neural network module, the initial hidden layer is initialized before message passing:

[0018]

[0019] Where τ is the ReLU activation function, W i For learnable matrix parameters, cat(x) v e vw This involves combining atomic features and chemical bond features into a corresponding matrix;

[0020] Then, at each step t of the directed message passing, the hidden layer for the atomic features of the compound is... and transmitting information and the hidden layer of chemical bond characteristics and transmitting information Update;

[0021]

[0022]

[0023]

[0024] h v =τ(W a cat(x v m v ))

[0025] Then, the hidden layers of the atoms are summed to obtain the eigenvector h of the molecule:

[0026]

[0027] Finally passed Compound properties are predicted, where f(.) is a feedforward neural network module, which leads to a pre-trained graph neural network model.

[0028] Furthermore, the specific content of fine-tuning the parameters in the pre-trained network model in step four to obtain new parameters and model includes: randomly initializing the output layer of the pre-trained graph neural network model, keeping the framework of other layers unchanged, inputting simplified molecular linear input text and bioactivity data of the target domain dataset corresponding to the compound skeleton in a standardized format, and fine-tuning the parameters of each layer of the network model to obtain a new prediction model for the target domain small sample association task.

[0029] Furthermore, the fine-tuning of the parameters of each layer of the network model includes creating a new neural network, namely the target model, and then transferring all model designs and parameters from the model trained based on the source domain dataset, except for the final feedforward neural network module, to the target model; then adding an output layer to the target model, the output layer structure being the same as the output layer structure of the model trained based on the source domain dataset but with randomly updated initial parameters; then training the target model using the skeleton corresponding to the target domain dataset compounds obtained in step two; training the output layer from scratch, fine-tuning the parameters of other layers, and finally obtaining a prediction model for the target domain small-sample association task.

[0030] Furthermore, the specific content of obtaining the molecular skeleton comprehensive score in step five includes: firstly, based on the deep learning network prediction model for the target domain small sample association task obtained in step four, the activity prediction of the library of compounds to be screened is performed to obtain the predicted bioactivity value; the average of the normalized molecular docking score, drug-likeness score, and synthesis difficulty score of the top 20% of small molecule compound skeletons ranked by predicted bioactivity value is taken as the final molecular skeleton comprehensive score; then, the top 10% of small molecule skeletons are obtained based on the molecular skeleton comprehensive score.

[0031] Furthermore, the library of compounds to be screened is the DrugBank database.

[0032] Furthermore, the specific details of taking the average of the normalized molecular docking score, drug-likeness score, and synthesis difficulty score as the final comprehensive score of the molecular skeleton include:

[0033] Autodock 4.2.6 software was used to optimize the structure and energy of the molecular backbone and target protein domains. Then, global molecular docking technology was used to obtain multiple conformations with the highest scores, and the top 5 dominant conformations were selected. i (i = 1, 2, 3, 4, 5), calculate the average score, and use it as the molecular docking score.

[0034] The drug-likeness score Q of the molecular skeleton was calculated using the open-source package RDkit. score ∈[0,1];

[0035] The synthesis difficulty score (SA) of the molecular skeleton was calculated using the open-source package RDkit. score ∈[1,10];

[0036] The normalization algorithm makes (D) score Q score SA score )∈[0,1];

[0037] The normalization algorithm is as follows:

[0038]

[0039] Where, x=(D score Q score SA score ),max(x) andmin(x) are the highest and lowest scores for the corresponding indicators in the skeletons of the top 20% of small molecule compounds in the predicted bioactivity value ranking, respectively;

[0040] The overall score of the molecular skeleton is:

[0041]

[0042] Where y is the overall score of the molecular skeleton, and y∈[0,1]; x' D It is a normalized molecular docking score, x' Q It is a normalized drug-likeness score, x' SA It is a normalized score for the difficulty of synthesis.

[0043] The technical solution of the present invention can produce the following technical effects:

[0044] The drug molecule scaffold screening and replacement method based on a deep transfer learning model provided in this invention combines a dual upstream-downstream strategy of drug target upstream and downstream and transfer learning task upstream and downstream. This helps to establish novel molecular scaffold screening and replacement models for upstream and downstream targets with small sample sizes, solving the problem of poor molecular property prediction performance on small sample datasets. By introducing transfer learning, extracting and fine-tuning the feature values ​​of the pre-trained network, better training results are achieved on small sample datasets. This provides an application approach to address the problems of high difficulty and cost in synthesizing new scaffolds and low scaffold novelty. The model trained by this invention can be widely applied to the screening and replacement of drug molecule scaffolds for targeted drugs targeting anti-tumor signaling pathways and antiviral targets, helping to provide novel drug molecule scaffolds for the design of target molecules with scarce data. Attached Figure Description

[0045] Figure 1 This presents the algorithm framework and computational flow of the drug molecule backbone replacement and screening method based on a deep transfer learning model of the present invention.

[0046] Figure 2 These are some potential molecular skeletons screened out in the embodiments of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to specific embodiments and corresponding drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0048] like Figure 1 As shown, the overall process of the drug molecule backbone replacement and screening method based on deep transfer learning model described in this invention is as follows:

[0049] S1. Obtain the corresponding compounds from the source domain dataset and the target domain dataset.

[0050] The target domain dataset is a small sample dataset, while the source domain dataset is a large sample dataset; the source domain is the dataset associated with the target domain task. Both source and target domain data were obtained from public databases and literature such as Pubchem, ChEMBL, and BindingDB. Drug molecule bioactivity labels were defined using binary variables (active = 1, inactive = 0), where IC... 50 >10μM is defined as inactive, IC 50 <10μM is defined as active.

[0051] S2. Convert the compounds to be screened and the corresponding compounds in the obtained source domain dataset and target domain dataset into compound skeletons.

[0052] The Murcko skeleton clustering compound library, based on the open-source package RDkit, is used to extract the skeletons of corresponding compounds in the source and target domain datasets, as well as compounds to be screened, obtaining the source domain, target domain, and compound skeletons required for subsequent calculations. This method removes redundant side chains from compounds and expresses only the skeletal structure, facilitating compound expression at the skeletal level. The obtained molecular skeleton datasets corresponding to the source and target domains serve as the training set for training the model.

[0053] S3. Input the simplified molecular linear input canonical format (i.e., SMILEs) text of the source domain dataset obtained in step two, along with bioactivity data, into a graph neural network model for pre-training to obtain the network model.

[0054] A graph neural network model is established, comprising a directed message-passing neural network (D-MPNN) module and a feedforward neural network (FNN) module. The D-MPNN module is used to extract and learn molecular features, while the FNN module is used to classify and predict molecular properties. The message-passing neural network takes the smiley strings of molecules in the training set as input text and internally converts them into graphs for molecular representation using the open-source package RDkit. Simultaneously, after the D-MPNN module extracts and learns molecular features, the FNN module uses these learned features to construct a classifier and predict molecular properties, ultimately resulting in a pre-trained network model.

[0055] Specifically, by using the open-source package RDkit to calculate the input text of small molecule compound smiles, the atomic features x of the corresponding small molecule compound are obtained. v and chemical bond characteristics e vw , and serve as the initial input features for the model.

[0056] In the directional message passing neural network module, the initial hidden layer is initialized before message passing:

[0057]

[0058] Where τ is the ReLU activation function, W i For learnable matrix parameters, cat(x) v e vw This involves combining atomic features and chemical bond features into a corresponding matrix;

[0059] Then, at each step t of the directed message passing, the hidden layer for the atomic features of the compound is... and transmitting information and the hidden layer of chemical bond characteristics and transmitting information Update.

[0060]

[0061]

[0062]

[0063] h v =τ(W a cat(x v m v ))

[0064] Then, the hidden layers of the atoms are summed to obtain the eigenvector h of the molecule:

[0065]

[0066] Finally, through compound property prediction is performed, where f(.) is a feedforward neural network module, and then a pre-trained graph neural network model is obtained.

[0067] S4. Input the smiles text of the compound skeleton corresponding to the target domain dataset and the bioactivity data of the corresponding target into the pre-trained network model obtained above, and fine-tune the parameters in the network model to obtain new parameters and a model.

[0068] The fine-tuning of the parameters in the network model refers to creating a new neural network, that is, the target model; in the target model, except for the FNN module finally used to predict compound properties, all model designs and their parameters in the model trained based on the source domain dataset are migrated to the target model; further, an output layer is added to the target model, and its structure is the same as that of the output layer of the model trained based on the source domain, but the initial parameters are randomly updated. Then use the compound skeleton corresponding to the target domain to train the target model; the output layer is trained from scratch, while the parameters of other layers are fine-tuned. Finally, a deep learning network model for the target domain dataset is obtained, which is the prediction model for the target domain small sample association task dataset.

[0069] S5. Based on the deep learning network model for the target domain dataset obtained in step S4, perform a comprehensive screening of the skeletons of the compounds to be screened, obtain a comprehensive score of the molecular skeletons, and then screen out the molecular skeletons.

[0070] First, perform activity prediction on the DrugBank database based on the above deep learning network prediction model for the target domain dataset to obtain predicted bioactivity values; the prediction method preferably uses categorical variables, and the obtained predicted bioactivity value is a numerical value of 0 < score < 1, representing the probability value that the skeleton is active (Label = 1).

[0071] Then, take the average of the normalized molecular docking scores, druggability scores, and synthesis difficulty scores of the small molecule compound skeletons ranked in the top 20% of the predicted bioactivity values as the final comprehensive score of the molecular skeletons; then obtain the top 10% of the small molecule skeletons based on the comprehensive score of the molecular skeletons.

[0072] Among them, the specific content of obtaining the comprehensive score of the molecular skeletons includes:

[0073] Use the Autodock4.2.6 software to optimize the structural energy of the molecular skeleton and the target domain target protein, then obtain multiple conformations with the highest scores through global molecular docking technology, and then take the top 5 dominant conformations D i(i = 1, 2, 3, 4, 5), calculate the average score, and use it as the molecular docking score.

[0074] The drug-likeness score Q of the molecular skeleton was calculated using the open-source package RDkit. score ∈[0,1];

[0075] The synthesis difficulty score (SA) of the molecular skeleton was calculated using the open-source package RDkit. score ∈[1,10];

[0076] The normalization algorithm makes (D) score Q score SA score )∈[0,1];

[0077] The normalization algorithm is as follows:

[0078]

[0079] Where, x=(D score Q score SA score ),max(x) andmin(x) are the highest and lowest scores for the corresponding indicators in the skeletons of the top 20% of small molecule compounds in the predicted bioactivity value ranking, respectively;

[0080] The overall score of the molecular skeleton is:

[0081]

[0082] Where y is the overall score of the molecular skeleton, and y∈[0,1]; x' D It is a normalized molecular docking score, x' Q It is a normalized drug-likeness score, x' SA It is a normalized synthesis difficulty score. In this embodiment, some of the screened skeletons are as follows: Figure 2 As shown.

[0083] S6 involves replacing the selected molecular skeleton with the target small molecule compound to obtain a series of novel molecules with different skeletons.

[0084] To verify the effectiveness of this invention, this embodiment uses two sets of related task datasets as input: one where the target domain target is B1 and the source domain target is A1, which are downstream targets of the target domain target B1 in the anti-tumor signaling pathway; and the other where the target domain target is B2 and the source domain target is A2, which are downstream targets of the target domain target B2 in the anti-tumor signaling pathway. These datasets were used as targets for anti-tumor signaling pathways and antiviral targets, respectively. The data input information for each target set is shown in Table 1.

[0085] Table 1 Input Information for Each Target Set

[0086]

[0087] Based on the drug molecule backbone screening and replacement method based on deep transfer learning model provided in this embodiment and the information of the two sets of related task datasets mentioned above, the classification performance of this method on the two sets of datasets is compared with that of different algorithms. The comparison results are shown in Table 2.

[0088] Table 2 compares the classification performance of the method in this embodiment with that of different algorithms on two datasets.

[0089]

[0090]

[0091] As shown in Table 2, the performance results of the prediction model in this embodiment demonstrate that the drug molecule scaffold screening and replacement method based on deep transfer learning provided by this invention significantly improves the prediction performance of the upstream / downstream target drug molecule scaffold screening model for small sample data by combining a dual upstream / downstream strategy of drug target upstream and downstream and transfer learning task upstream and downstream, thus solving the problem of poor molecular property prediction performance for small sample datasets. Furthermore, this embodiment illustrates that the model trained by this invention exhibits excellent prediction performance in targeted drug molecule scaffold screening and replacement tasks for anti-tumor signaling pathway targets, antiviral targets, and other targets, helping to provide novel drug molecule scaffolds for target molecule design with scarce data.

[0092] The preferred embodiments of the present invention have been described in detail above, but they do not constitute any limitation on the present invention. Any equivalent substitutions or modifications made by those skilled in the art to the technical solutions and content disclosed in the present invention without departing from the scope of the present invention shall be deemed to have remained within the protection scope of the present invention.

Claims

1. A method for drug molecule scaffold replacement and screening based on a deep transfer learning model, characterized in that, The method comprises the following steps: Step one, obtaining a source domain data set and a target domain data set; Step two, converting the to-be-screened compound and the corresponding compounds of the source domain data set and the target domain data set obtained in step one into compound skeletons; Step three, inputting the simplified molecular linear input specification format text and biological activity data of the source domain data set corresponding compound skeleton obtained in step two into a network model pre-trained by a graph neural network model; The graph neural network model comprises a directed message passing neural network module and a feedforward neural network module, the directed message passing neural network module is used for extracting molecular features, and the feedforward neural network module is used for completing classification and prediction of molecular properties; Step four, inputting the simplified molecular linear input specification format text and biological activity data of the target domain data set corresponding compound skeleton obtained in step two into the pre-trained network model obtained in step three, and fine-tuning the parameters in the network model to obtain new parameters and a model; Before the source domain data set corresponding to the simplified molecular linear input specification format text of the compound skeleton, that is, the smiles text input into the graph neural network model, the open source package RDkit is used to calculate the input source domain data set corresponding to the compound smiles text, and the atomic features x of the corresponding small molecule compound are obtained v and bond features e vw as the initial input features of the graph neural network model; In the directed message passing neural network module, the initial hidden layer is calculated before message passing: where τ is the ReLU activation function, W i is a learnable matrix parameter, cat(x v ,e vw ) is the combination of atomic features and bond features into the corresponding matrix; Then at each step t of the directed message passing, the hidden layer of compound atomic features and passing information and the hidden layer of bond features and passing information is updated; h v = τ(W a cat(x v , m v )); Then, the hidden layer of the atom is summed to obtain a feature vector h of the molecule: h = ∑ v∈G h v ; Final pass performing compound property prediction, wherein f(.) is a feedforward neural network module, and the pre-trained graph neural network model is obtained; Step five, performing skeleton comprehensive screening on the to-be-screened compound based on the deep learning network model for the target domain data set obtained in step four to obtain a molecular skeleton comprehensive score and screen out a molecular skeleton; Step six, replacing the screened molecular skeleton into a target small molecule compound to obtain a series of skeleton-replaced new molecules.

2. The method for drug molecule scaffold hopping and screening based on deep transfer learning model according to claim 1, characterized in that, The target domain data set is a small sample data set, and the source domain data set is a large sample data set of a target domain related task; the source domain data set and the target domain data set are obtained from public databases and public literature such as Pubchem, ChEMBL and BindingDB.

3. The method for drug molecule scaffold hopping and screening based on deep transfer learning model according to claim 2, characterized in that, Step two specifically comprises: using the Murcko skeleton-based compound library in the open source package RDkit to extract the skeletons of the source domain compounds, the target domain compounds and the to-be-screened compound. 4.The method of claim 1, wherein the method is characterized by, The specific content of fine-tuning the parameters in the pre-trained network model in step four to obtain new parameters and a model comprises: randomly initializing the output layer of the pre-trained graph neural network model, keeping the framework of other layers unchanged, inputting the simplified molecular linear input specification format text and biological activity data of the target domain data set corresponding compound skeleton, fine-tuning the parameters of each layer of the network model, and obtaining a new prediction model for the target domain small sample related task. 5.The method of claim 1, wherein the method is characterized by, The fine-tuning of the parameters of each layer of the network model comprises creating a new neural network, i.e., a target model, and then migrating all model designs and parameters of the model trained based on the source domain dataset, except for the final feedforward neural network module, to the target model; then adding an output layer to the target model, which has the same structure as the output layer of the model trained based on the source domain dataset but with randomly updated initial parameters; then training the target model using the target domain dataset compounds obtained in step two; the output layer is trained from scratch, and the parameters of other layers are fine-tuned, and finally a prediction model for the target domain small sample correlation task is obtained. 6.The method of claim 1, wherein the method is characterized by, The specific content of obtaining the comprehensive molecular scaffold score in step five comprises: firstly, predicting the activity of the compound library to be screened based on the deep learning network prediction model for the target domain small sample correlation task obtained in step four to obtain a predicted biological activity value; the average of the normalized molecular docking score, drug-likeness score and synthesis difficulty score of the top 20% small molecule compound scaffolds is taken as the final comprehensive molecular scaffold score; and then the top 10% small molecule scaffolds are obtained based on the comprehensive molecular scaffold score.

7. The method of claim 6, wherein the method is characterized by, The compound library to be screened is the DrugBank database.

8. The method of drug molecule scaffold hopping and screening based on deep transfer learning model according to any one of claims 6-7, characterized in that, The specific content of taking the average of the normalized molecular docking score, drug-likeness score and synthesis difficulty score as the final comprehensive molecular scaffold score comprises: The Autodock 4.2.6 software is used to optimize the structure energy of the molecular skeleton and the target protein of the target domain, and then the global molecular docking technology is used to obtain a plurality of highest-score conformations, and then the top 5 advantage conformations D i , i = 1, 2, 3, 4, 5, calculate the average score as the molecular docking score Calculate druglikeness score Q of the molecular skeleton by open source package RDkit score ∈ [0, 1]; Scoring of synthetic accessibility of the molecular scaffold by open source package RDkit, SA score ∈ [1, 10]; (D score , Q score , SA score ) ∈ [0, 1]; The normalization algorithm is: where x = (D score , Q score , SA score ), max(x) and min(x) are the highest and lowest scores of the corresponding index in the top 20% of the small molecule compounds ranked by the predicted biological activity value, respectively. The comprehensive molecular scaffold score is: where y is the molecular scaffold composite score and y e [0, 1]; x' D is the normalized molecular docking score, x' Q is the normalized drug-likeness score, x' SA is the normalized synthetic difficulty score.

Images