Small molecule virtual screening method and system based on graph convolutional neural network

Abstract

The invention relates to a small molecule virtual screening method and system based on a graph convolutional neural network, and the method comprises the steps: firstly obtaining the data of an active compound and a bait compound of a target spot, carrying out the preprocessing of the obtained data, constructing a graph convolutional neural network model, carrying out the training and evaluation of the model through a preprocessed data set, and obtaining a small molecule virtual screening result. And then scoring the small molecule compound data obtained in the compound database by using the trained graph convolutional neural network model, retaining the small molecule compound with high prediction score, and finally obtaining the known drug molecule data of the target spot. And performing compound similarity search on the small molecule compounds with high prediction scores and known drug molecules to obtain the small molecule compounds with similarity with the known drug molecules. Compared with the prior art, the method has the advantages that virtual screening is utilized, the drug research and development process is accelerated, the experiment cost is reduced, compound similarity search is combined on the basis, the screening accuracy is further improved, and the screening result range is narrowed.

Description

Technical Field

[0001] The present invention relates to the field of computer-assisted medicine, and in particular to a small molecule virtual screening method and system based on graph convolutional neural network. Background Art

[0002] The design and development of drugs is a very complex process involving knowledge from various cutting-edge fields. It is considered to be an interdisciplinary work that requires a lot of time and manpower, and is characterized by high risk, high investment, and long cycle. With the emergence of problems such as the diversity of diseases and drug resistance, the demand for new drugs is also increasing, but the research and development of new drugs faces many problems such as long research and development cycle, high research and development costs, and low research and development success rate.

[0003] In traditional drug screening, drug development begins with the identification of disease targets of interest, followed by high throughput screening (HTS) experiments to screen out compounds in a database of small molecule compounds. The lead compound then needs to be optimized to increase its efficacy and other desired properties, such as solubility, and to eliminate toxicity and off-target effects. After these preclinical studies are completed, potential candidate drugs must pass a series of clinical trials before they can become approved drugs. Generally speaking, it takes billions of dollars and ten to fifteen years for a new drug to be successfully marketed from the initial development to the final launch. However, under the premise of high human and material investment and long time consumption, the success rate of small molecule drugs being able to be marketed is only 6.2%, so the risk of drug development failure is relatively high.

[0004] In order to further improve the efficiency and success rate of drug development, various computational methods are widely used in the process of drug development. Chinese patent CN115862768B discloses an optimization method for large-scale virtual screening of drugs. The multi-stage screening combined with consensus scoring can flexibly adjust the number of small molecules screened out at each stage to adjust the time and screening effect. It only uses a small amount of computing resources and has a higher early enrichment ability. Compared with traditional screening methods, this method can significantly speed up the speed of virtual screening while maintaining a higher early enrichment ability, meeting the actual needs of virtual screening of large compound databases, but the range of screening results after scoring screening is still large. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a small molecule virtual screening method and system based on graph convolutional neural network, which utilizes virtual screening to accelerate the process of drug development and reduce experimental costs, and on this basis combines compound similarity search to further improve screening accuracy and narrow the range of screening results.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] A small molecule virtual screening method based on graph convolutional neural network, comprising the following steps:

[0008] Step S1, obtaining data of active compounds and decoy compounds of target points, and preprocessing the obtained data;

[0009] Step S2, constructing a graph convolutional neural network model and using the preprocessed dataset to train and evaluate the model;

[0010] Step S3, obtaining small molecule compound data from the compound database, scoring the small molecule compound data using the trained graph convolutional neural network model, and retaining small molecule compounds with high prediction scores;

[0011] Step S4, obtaining known drug molecule data of the target, performing compound similarity search between small molecule compounds with high prediction scores and known drug molecules, and obtaining small molecule compounds that have similarities with known drug molecules.

[0012] Furthermore, in step S1, the pre-processing of the acquired data includes balancing the molecular weight, oil-water partition coefficient and formal charge of the active compound and the bait compound.

[0013] Furthermore, the balancing comprises the following steps:

[0014] Compare the molecular weight distribution of the active compound data set and the decoy compound data set, and remove compounds with too small or too large molecular weight in the decoy compound data set;

[0015] Compare the oil-water partition coefficient distributions of the active compound data set and the bait compound data set, and remove compounds in the bait compound data set with oil-water partition coefficient values ​​that are too small or too large;

[0016] The formal charge distributions of the active compound dataset and the decoy compound dataset are compared, and the charged molecules in the decoy compound dataset are neutralized to become neutral molecules.

[0017] Furthermore, after preprocessing the acquired data, the data is divided into a training set and a test set. The training set is used to train the graph convolutional neural network model, and the test set is used to evaluate the graph convolutional neural network model.

[0018] Furthermore, in step S2, the number of samples to be processed in batches and the initial learning rate are set before model training, and the Adam optimizer (Adaptive Moment Estimation) is used during model training to speed up the optimization of model parameters.

[0019] Furthermore, in step S2, Matthews correlation coefficient, accuracy, recall rate and F1 value are selected as model evaluation indicators to evaluate the model performance.

[0020] Furthermore, in step S3, the small molecule compound data obtained from the compound database is characterized, and the data set containing the molecular information is converted into a feature matrix and input into a graph convolutional neural network model for analysis.

[0021] Furthermore, the analysis and scoring criteria of the graph convolutional neural network model include the inhibitory ability of small molecules on target sites.

[0022] Furthermore, in step S4, the compound similarity search includes:

[0023] Calculate the molecular fingerprint of the small molecule compound to be verified;

[0024] Calculate the molecular fingerprints of the collected drug molecules;

[0025] The similarity between small molecule compounds and known drug molecules is calculated based on molecular fingerprints, and small molecule compounds with similarities are output.

[0026] A small molecule virtual screening system based on graph convolutional neural network, comprising:

[0027] An input module obtains data on active compounds and decoy compounds of target sites and pre-processes the obtained data;

[0028] The training module builds a graph convolutional neural network model and uses the preprocessed dataset to train and evaluate the model;

[0029] The screening module obtains small molecule compound data from the compound database, scores the small molecule compound data using the trained graph convolutional neural network model, and retains small molecule compounds with high prediction scores;

[0030] The comparison module obtains the known drug molecule data of the target, performs compound similarity search between the small molecule compounds with high prediction scores and the known drug molecules, and obtains the small molecule compounds that are similar to the known drug molecules;

[0031] The output module outputs small molecule compounds that are similar to known drug molecules.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] 1. The present invention uses virtual screening to screen out small molecule compounds that are similar to known drug molecules from a large number of compound databases through computer simulation and prediction, which accelerates the process of drug development and greatly reduces experimental costs;

[0034] 2. After the present invention uses the graph convolutional neural network model to screen small molecule compounds, it further performs a compound similarity search on the screening results. By verifying whether the screening results are similar to the known drug molecules of the target, the screening accuracy can be further improved and the scope of the screening results can be narrowed, thereby providing support for subsequent drug research and development work. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0036] 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 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 should fall within the scope of protection of the present invention.

[0037] The embodiment of the present invention discloses a small molecule virtual screening method based on a graph convolutional neural network, comprising the following steps:

[0038] Step 1, obtaining the data of active compounds and decoy compounds of the target, and preprocessing the obtained data;

[0039] Step 2: Build a graph convolutional neural network model and use the preprocessed dataset to train and evaluate the model;

[0040] Step 3, obtaining small molecule compound data from the compound database, scoring the small molecule compound data using the trained graph convolutional neural network model, and retaining small molecule compounds with high prediction scores;

[0041] Step 4, obtaining the known drug molecule data of the target, performing compound similarity search between the small molecule compounds with high prediction scores and the known drug molecules, and obtaining small molecule compounds that are similar to the known drug molecules.

[0042] Furthermore, the processing methods in step 1 are as follows:

[0043] Compare the molecular weight distribution of the active compound data set and the decoy compound data set, and remove compounds with too small or too large molecular weight in the decoy compound data set;

[0044] Compare the oil-water partition coefficient distributions of the active compound data set and the bait compound data set, and remove compounds in the bait compound data set with oil-water partition coefficient values ​​that are too small or too large;

[0045] The formal charge distributions of the active compound dataset and the decoy compound dataset are compared, and the charged molecules in the decoy compound dataset are neutralized to become neutral molecules.

[0046] Furthermore, the steps of using the graph convolutional neural network model to screen small molecule compounds in step 3 are as follows:

[0047] Step 31, collecting data of small molecule compounds to be screened;

[0048] Step 32, performing characterization processing on the data collected in step 31, converting the data set containing molecular information into a feature matrix for analysis by the graph convolutional neural network model;

[0049] Step 33, screening the processed small molecule compounds, the model outputs a prediction result for each small molecule, and the prediction result indicates the inhibitory ability of the small molecule on the target.

[0050] In this example, EGFR was used as the target, and a total of 542 active compounds targeting EGFR were obtained. Each active compound had 50 corresponding bait compounds. These bait compounds had similar physicochemical properties to the active compounds, but had different two-dimensional topological structures.

[0051] Before model training, the number of samples processed in batches was set to 128, and the initial learning rate was set to 0.001. During model training, the Adam optimizer was used to speed up the optimization of model parameters.

[0052] In the model evaluation stage, the embodiment of the present invention first trains a graph convolutional neural network model based on the EGFR target, using the Matthews correlation coefficient (MCC), accuracy, recall, and F1-Score as evaluation indicators of the model. By adjusting different Batch_sizes and learning rates, the model can achieve better performance. The experimental results are shown in Table 1.

[0053] Table 1 Batch_size and learning rate experimental results

[0054]

[0055]

[0056] From the results in Table 1, we can see that the model works best when Batch_size is 128 and the learning rate is 0.001. Since there is an imbalance between the amount of active compound data and the amount of decoy compound data when training the model, the MCC indicator is more convincing in evaluating the model when the amount of data is unbalanced. When Batch_size is 128 and the learning rate is 0.001, the MCC of the model reaches 0.9117.

[0057] After the model is trained, this embodiment uses the graph convolutional neural network model to screen a large number of small molecules from the ZINC database, predict the inhibitory ability of these small molecules on the EGFR target, and finally take out the 10 small molecules with the highest prediction results. The experimental results are shown in Table 2. In the table, SMILES (Simplified molecular input line entry system) is the structural text string representation of the small molecule compound after screening, ZINCID is the ID number of the small molecule compound in the ZINC database, and POS is the model scoring score.

[0058] Table 2 Information about the top 10 small molecules

[0059]

[0060]

[0061] After screening the small molecules, this embodiment further performs compound similarity search on these small molecules to verify whether they have similarity with known drug molecules of EGFR target. Here, the screened small molecules are sequentially subjected to compound similarity search, firstly calculating the molecular fingerprint of the small molecule to be verified, then calculating the molecular fingerprint of the known drug molecules of the collected EGFR target, and then calculating the similarity based on the molecular fingerprint, outputting the drug molecule with the highest similarity to the molecule to be verified, and the experimental results are shown in Table 3.

[0062] Table 3 Compound similarity search results

[0063]

[0064] As shown in Table 3, 4 of the 10 small molecules in the screening results have similarities with known drug molecules targeting EGFR.

[0065] In summary, the embodiment of the present invention uses a graph convolutional neural network model to perform virtual screening of large-scale small molecule compounds. Compared with other methods, this method searches for compound similarity between the screening results and drug molecules on this basis to verify whether they are similar to known drug molecules of EGFR targets, further improving the screening accuracy and narrowing the range of screening results.

[0066] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A small molecule virtual screening method based on graph convolutional neural network, characterized in that: The following steps are involved: Step S1, obtaining data of active compounds and decoy compounds of target points, and preprocessing the obtained data; Step S2, constructing a graph convolutional neural network model and using the preprocessed dataset to train and evaluate the model; Step S3, obtaining small molecule compound data from the compound database, scoring the small molecule compound data using the trained graph convolutional neural network model, and retaining small molecule compounds with high prediction scores; Step S4, obtaining known drug molecule data of the target, performing compound similarity search between small molecule compounds with high prediction scores and known drug molecules, and outputting small molecule compounds that have similarities with known drug molecules.

2. A small molecule virtual screening method based on graph convolutional neural network according to claim 1, characterized in that: In the step S1, the pre-processing of the acquired data includes balancing the molecular weight, oil-water partition coefficient and formal charge of the active compound and the bait compound.

3. A small molecule virtual screening method based on graph convolutional neural network according to claim 2, characterized in that: The balancing comprises the following steps: Compare the molecular weight distribution of the active compound data set and the decoy compound data set, and remove compounds with too small or too large molecular weight in the decoy compound data set; Compare the oil-water partition coefficient distributions of the active compound data set and the bait compound data set, and remove compounds in the bait compound data set with oil-water partition coefficient values ​​that are too small or too large; The formal charge distributions of the active compound dataset and the decoy compound dataset are compared, and the charged molecules in the decoy compound dataset are neutralized to become neutral molecules.

4. The small molecule virtual screening method based on graph convolutional neural network according to claim 1, characterized in that: After preprocessing the acquired data, the data is divided into a training set and a test set. The training set is used to train the graph convolutional neural network model, and the test set is used to evaluate the graph convolutional neural network model.

5. The small molecule virtual screening method based on graph convolutional neural network according to claim 1, characterized in that: In step S2, the number of samples to be processed in batches and the initial learning rate are set before model training, and the Adam optimizer is used during model training to speed up the optimization of model parameters.

6. The small molecule virtual screening method based on graph convolutional neural network according to claim 1, characterized in that: In step S2, Matthews correlation coefficient, accuracy, recall rate and F1 value are selected as model evaluation indicators to evaluate the model performance.

7. The small molecule virtual screening method based on graph convolutional neural network according to claim 1, characterized in that: In step S3, the small molecule compound data obtained from the compound database is characterized, the data containing the molecular information is converted into a feature matrix, and input into the graph convolutional neural network model for analysis.

8. A small molecule virtual screening method based on graph convolutional neural network according to claim 7, characterized in that: The analysis and scoring criteria of the graph convolutional neural network model include the inhibitory ability of small molecules on target sites.

9. The small molecule virtual screening method based on graph convolutional neural network according to claim 1, characterized in that: In step S4, the compound similarity search includes: Calculate the molecular fingerprint of the small molecule compound to be verified; Calculate the molecular fingerprints of the collected drug molecules; The similarity between small molecule compounds and known drug molecules is calculated based on molecular fingerprints, and small molecule compounds with similarities are output.

10. A small molecule virtual screening system based on graph convolutional neural network, characterized in that: include: An input module obtains data on active compounds and decoy compounds of target sites and pre-processes the obtained data; The training module builds a graph convolutional neural network model and uses the preprocessed dataset to train and evaluate the model; The screening module obtains small molecule compound data from the compound database, scores the small molecule compound data using the trained graph convolutional neural network model, and retains small molecule compounds with high prediction scores; The comparison module obtains the known drug molecule data of the target, performs compound similarity search between the small molecule compounds with high prediction scores and the known drug molecules, and obtains the small molecule compounds that are similar to the known drug molecules; The output module outputs small molecule compounds that are similar to known drug molecules.

Images