Multimodal drug virtual screening method

Abstract

The application provides a kind of multimodal drug virtual screening method, comprising the following steps: selecting pre-trained graph transformer GT model, based on public data set, fine-tuning graph transformer GT model;Using active learning strategy to select molecules, and carrying out Vina-GPU+ docking, using molecular docking score to fine-tune graph transformer GT model;Using the biological activity experimental data of target target to fine-tune the model;Reasoning on pre-screening compound library molecules;According to molecular score, sorting, according to the needs of selecting the first N molecules as candidate compounds.The application greatly reduces the number of molecules that need to be docked when performing virtual screening on a large-scale compound library, shortens the time required for virtual screening, improves the accuracy of virtual screening, and significantly reduces the cost and time of drug development.

Description

TECHNICAL FIELD

[0001] The present application relates to a multi-modal drug virtual screening method, belonging to the technical field of artificial intelligence drug discovery. BACKGROUND

[0002] Virtual screening is a method of screening and evaluating a large number of compounds using computer simulation technology. Compared with the traditional drug screening method, virtual screening can quickly evaluate the biological activity and pharmacological properties of thousands of compounds through simulation calculation, greatly shorten the drug development time, reduce the cost of drug development, and provide more possibilities and opportunities for drug discovery.

[0003] Modern drug design often faces large-scale virtual screening, that is, more virtual compound libraries need to be screened for promising compounds. Continuously improving the speed and accuracy of virtual screening methods will be the core problem of modern drug virtual screening for a long time.

[0004] Therefore, it is necessary to provide a multi-modal drug virtual screening method to solve the above problems. SUMMARY

[0005] The purpose of the present application is to provide a multi-modal drug virtual screening method, which can accelerate the virtual screening speed and improve the virtual screening accuracy.

[0006] To achieve the above purpose, the present application provides a multi-modal drug virtual screening method, comprising the following steps:

[0007] Step S100, selecting a pre-trained graph transformer GT model, and fine-tuning the graph transformer GT model based on a public dataset;

[0008] Step S200, selecting molecules using an active learning strategy, and performing Vina-GPU+ docking, and fine-tuning the graph transformer GT model using the molecular docking score;

[0009] Step S300, fine-tuning the model using biological activity experimental data of the target;

[0010] Step S400, reasoning on the pre-screening compound library molecules;

[0011] Step S500, sorting according to the molecular score, and selecting the top N molecules as candidate compounds as needed.

[0012] As a further improvement of the present invention, in step S100, a pre-trained graph transformer GT model is selected and fine-tuned based on a public dataset. First, Vina-GPU+ docking is used to obtain a score between the target and the small molecule in the dataset. The score result is then saved to the dock folder in the working directory, and the file name is entered into the configuration file as the dataset for fine-tuning the graph transformer GT model.

[0013] As a further improvement of the present invention, the parameters in the configuration file include:

[0014] dataset: configuration related to the dataset;

[0015] docking_arg: parameter information for Vina-GPU+ docking;

[0016] In dataset: configuration related to the dataset, more specific parameters are:

[0017] stage1_1: path to the public dataset used to fine-tune the GT model of the graph transformer.

[0018] screen_dataset: path information of the molecular library to be screened;

[0019] bioactivity_dataset.path: path information of the bioactivity experiment dataset;

[0020] bioactivity_dataset.assay_id: the specific experimental batch used for fine-tuning using the bioactivity assay dataset;

[0021] In the configuration of docking_arg: Vina-GPU+ docking parameter information, more specific parameters include:

[0022] receptor: the pathway where the docking receptor is located;

[0023] center_x, center_y, center_z: the center position of the docking box;

[0024] size_x, size_y, size_z: the size of the docking box;

[0025] thread: number of threads.

[0026] As a further improvement of the present invention, step S100 includes the following specific steps:

[0027] Step S101: Select data corresponding to a target in the public dataset and use Vina-GPU+ for docking;

[0028] Step S102, according to the docking results of Vina-GPU+ of the public data set, a data set for fine-tuning the graph transformer GT model is generated;

[0029] Step S103, load the pre-trained graph transformer GT model, and fine-tune the graph transformer GT model using mean absolute error MAE as the loss function.

[0030] As a further improvement of the application, in step S103, the formula of MAE is as follows:

[0031]

[0032] Wherein, y i and are the true value and the predicted value of the i-th sample, and n is the number of samples.

[0033] As a further improvement of the application, in step S200, it specifically includes the following steps:

[0034] Step S201, load the fine-tuned graph transformer GT model;

[0035] Step S202, use the graph transformer GT model to infer the pre-screening compound library molecules, and obtain the inference score corresponding to each SMILES in the pre-screening compound library;

[0036] Step S203, sort the inference scores in ascending order, and add the SMILES of the last 10% of the top 20% of the sorted results to the docking list;

[0037] Step S204, read the molecules in the SMILES list, and use Vina-GPU+ for docking;

[0038] Step S205, parse the docking results obtained in step S204, and construct a data set for the fine-tuning model according to the docking results;

[0039] Step S206, send the SMILES in the constructed data set to the model, the model encodes the SMILES, and calculates the predicted value Again with the TARGET value y i in the data set, and calculate the loss value using the MAE loss function.

[0040] As a further improvement of the application, step S204 specifically further includes the following steps:

[0041] Step S2042, the sub-process 1 reads the molecules in the SMILES list generated in step S203, generates a 3D conformation using pybel (the python API of openbabel), and generates a file named ID.pdbqt, where ID represents the index of the current molecule in the SMILES list;

[0042] Step S2043, the sub-process 1 writes the generated file name into the pipe, which will be read by the sub-process 2 and subjected to docking processing;

[0043] Step S2044, the sub-process 1 repeats steps S2042-S2043 until the SMILES list is read and all 3D structure files are generated, and then an end string is transmitted to the pipe, indicating that there is no molecule to be docked subsequently;

[0044] Step S2045, the sub-process 2 reads the file name in the pipe and generates a configuration file for docking according to the docking parameter variable in step S100 above;

[0045] Step S2046, the sub-process 2 calls Vina-GPU+ to perform docking according to the configuration file generated in step S2045, and extracts the docking scores in the docking result file into a dictionary with SMILES as the key and the docking score as the value;

[0046] Step S2047, steps S2045-S2046 are repeated until the end string in the pipe is read, indicating that there is no molecule subsequently, and a hash table storing the docking results is returned.

[0047] As a further improvement of the present application, in step S300, the model is fine-tuned using part of the biological activity experimental data of the current target, specifically including the following steps:

[0048] Step S301, according to the biological activity experimental data, each two constitutes a molecule pair, and is assigned as 0 and 1 according to the relative size, y label = 0, 1, and after processing, a ranking data set for fine-tuning the model is obtained;

[0049] Step S302, the SMILES molecule pairs in the constructed data set are sent into the model for prediction to obtain predicted scores and

[0050] Step S303, according to the predicted scores and

[0051] As a further improvement of the present application, in step S301, SMILES1 and SMILES2 constitute a molecular pair, and label is the size of the relative biological activity value of the molecular pair SMILES1 and SMILES2, when the biological activity value of SMILES1 is less than the biological activity value of SMILES2, Label is 1, otherwise, Label is 0.

[0052] As a further improvement of the present application, in step S303, a binary cross-entropy loss function is used:

[0053] loss = -y label *log(p)-(1-y label )*log(1-p),

[0054] wherein, σ is a hyperparameter, the model is trained.

[0055] The beneficial effects of the present application are: when the present application is used for virtual screening of large-scale compound library, the number of molecules that need to be docked is greatly reduced, the time required for virtual screening is shortened, the accuracy of virtual screening can be improved, and the cost and time of drug research and development are significantly reduced. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is the overall flowchart of the multi-modal drug virtual screening method of the present application.

[0057] Figure 2 is the flowchart of the disclosed data set fine-tuning graph transformer GT model in the multi-modal drug virtual screening method of the present application.

[0058] Figure 3 is an example of the configuration file of the multi-modal drug virtual screening method of the present application.

[0059] Figure 4 is the flowchart of the active learning fine-tuning model of the multi-modal drug virtual screening method of the present application.

[0060] Figure 5 is the flowchart of the molecular docking of the multi-modal drug virtual screening method of the present application. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be described in detail below in combination with the drawings and specific embodiments.

[0062] Here, it should be noted that, in order to avoid obscuring the present application due to unnecessary details, only structures and / or processing steps closely related to the scheme of the present application are shown in the drawings, and other details not closely related to the present application are omitted.

[0063] It is also to be noted that the terms "comprising", "comprises" or any other variant thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0064] As shown in Figures 1 to 5 The present application discloses a multi-modal drug virtual screening method, which is based on a pre-trained Graph Transformer (GT) model and uses a small amount of molecular docking information and high-quality biological activity experimental data to gradually fine-tune the Graph Transformer GT model, so as to significantly improve the precision and speed of drug virtual screening, and specifically includes the following steps:

[0065] Step S100, selecting a pre-trained Graph Transformer GT model, fine-tuning the Graph Transformer GT model based on a public dataset;

[0066] Step S200, using an active learning strategy to select molecules and performing Vina-GPU+ docking, fine-tuning the Graph Transformer GT model using molecular docking scores;

[0067] Step S300, fine-tuning the model using biological activity experimental data of the target;

[0068] Step S400, reasoning on the pre-screening compound library molecules;

[0069] Step S500, sorting according to the molecular scores, and selecting the top N molecules as candidate compounds as needed.

[0070] The steps S100-S500 will be described in detail below.

[0071] As shown in Figure 2 and Figure 3 In step S100, the pre-trained Graph Transformer GT model is selected, and the Graph Transformer GT model is fine-tuned based on a public dataset. First, the scoring results between the target and small molecules in the dataset are obtained using Vina-GPU+ docking, and then the scoring results are saved to the dock folder under the working directory. The file name is filled into the configuration file as the dataset for fine-tuning the Graph Transformer GT model.

[0072] As shown in Figure 3 Taking the configuration file in the figure as an example, the scoring result file name "aldh1_vina_gpu.csv" includes the following parameters:

[0073] dataset: configuration related to dataset;

[0074] docking_arg: parameter information of Vina-GPU+ docking;

[0075] In the dataset: configuration related to dataset, more specific parameters are:

[0076] stage1_1: public dataset path information used for fine-tuning the graph transformer GT model;

[0077] screen_dataset: path information of the molecule library that needs to be screened;

[0078] bioactivity_dataset.path: path information of the bioactivity experiment dataset;

[0079] bioactivity_dataset.assay_id: specific experiment batch for fine-tuning using the bioactivity experiment dataset;

[0080] In the docking_arg: parameter information of Vina-GPU+ docking, more specific parameters include:

[0081] receptor: path to the receptor.

[0082] center_x, center_y, center_z: center position of the docking box;

[0083] size_x, size_y, size_z: size of the docking box;

[0084] thread: number of threads;

[0085] After obtaining the scoring results of Vina-GPU+ on the public dataset, the pre-trained graph transformer GT model needs to be loaded and fine-tuned using MAE as the loss function.

[0086] Specifically, in step S100, a pre-trained graph transformer GT model is selected, and the graph transformer GT model is fine-tuned based on the public dataset. Step S100 includes the following specific steps:

[0087] Step S101, select a target corresponding data in the public dataset, and use Vina-GPU+ docking.

[0088] Step S102, according to the docking results of Vina-GPU+ on the public dataset, generate a dataset for fine-tuning the graph transformer GT model.

[0089] Step S103, load the pre-trained graph transformer GT model, fine-tune the graph transformer GT model using MAE as the loss function, and the formula of MAE is as follows:

[0090]

[0091] Where y i and are the true value and the predicted value of the i-th sample, and n is the number of samples. The smaller the value of the MAE loss function, the smaller the error of the model prediction, and at the same time, it also indicates that the fitting effect of the model on the same sample is better. When the model converges to a smaller value of MAE by updating the value of its weight parameters, the fine-tuning of the model is completed.

[0092] In step S200, in order to reduce the docking of molecules, a part of the screening data set is selected according to the active learning strategy, and Vina-GPU+ is used for docking. The docking score of the molecules is used to fine-tune the graph transformer GT model, and then the model learns the scoring rule of Vina-GPU+ corresponding to SMILES, so as to predict the scoring of the remaining molecules of Vina-GPU+.

[0093] The specific process is shown in Figure 3 , and specifically includes:

[0094] Step S201, load the fine-tuned graph transformer GT model;

[0095] Step S202, use the graph transformer GT model to reason the pre-screening compound library molecules, and get the reasoning score corresponding to each SMILES in the pre-screening compound library;

[0096] Step S203, sort the reasoning scores in ascending order (the lower the score, the higher the ranking), and add the SMILES of the last 10% of the top 20% of the sorted results to the docking list;

[0097] Step S204, read the molecules in the SMILES list, and use Vina-GPU+ for docking;

[0098] Step S205, parse the docking results obtained in step 2.4, and construct a data set for the fine-tuning model according to the docking results;

[0099] Step S206, send the SMILES in the constructed data set to the model, and the model encodes the SMILES to calculate the predicted value and compare it with the TARGET value y iThe loss value is calculated using the MAE (Mean Absolute Error) loss function. The model is fine-tuned when the MAE converges to a smaller value by updating the values of its weight parameters.

[0100] The specific process is shown in Figure 3 Further, in step S204, Vina-GPU+ is used to generate the 3D conformation of the small molecule required for docking, and the generation of the 3D conformation is faster than the docking. In order to shorten the time of this step, the task of generating the 3D conformation of the small molecule and the task of docking can be parallel. Step S2041, the main process starts two sub-processes, process 1 and process 2, which are used for molecular conformation generation and docking respectively. Step S204 further includes the following steps:

[0101] Step S2042, sub-process 1 reads the molecules in the SMILES list generated in step S203, generates 3d conformation using pybel (python api of openbabel), and the generated file is named ID.pdbqt, where ID represents the index of the current molecule in the SMILES list;

[0102] Step S2043, sub-process 1 writes the generated file name into the pipe, which will be read by sub-process 2 and processed for docking;

[0103] Step S2044, sub-process 1 loops step S2042~step S2043 until the SMILES list is read and all 3d structure files are generated, and then end string is transmitted to the pipe, indicating that there is no molecule to be docked subsequently;

[0104] Step S2045, sub-process 2 reads the file name in the pipe, and generates the configuration file for docking according to the docking parameter variable in step S100 above;

[0105] Step S2046, sub-process 2 calls Vina-GPU+, docks according to the configuration file generated in step S2045, and extracts the docking score in the docking result file and stores it in a dictionary with SMILES as the key and docking score as the value;

[0106] Step S2047, repeat step S2045~step S2046 until the end string in the pipe is read, indicating that there is no molecule subsequently, return the hash table storing the docking results, and the results are shown in Table 1:

[0107] Table 1 docking result example

[0108] KEY VALUE [H][C@]12CCCC[C@]11CCN(C)[C@H]2CC2=C1C=C(OC)C=C2 -5.6 CS(=O)(=O)C1=CC=C(C=C1)C1=C(C(=O)OC1)C1=CC=CC=C1 -5.7 CC1=C(C(=NO1)C1=CC=CC=C1)C1=CC=C(C=C1)S(N)(=O)=O -6 CCN(CC)CC1=C(O)C=CC(NC2=C3C=CC(Cl)=CC3=NC=C2)=C1 -5.4 ClC1=CC=C(CCC(CN2C=CN=C2)SC2=C(Cl)C=CC=C2Cl)C=C1 -5.1 CN1CCN(CC(=O)N2C3=CC=CC=C3C(=O)NC3=C2N=CC=C3)CC1 -6.8 [H][C@]12CC3=C(C(O)=C(O)C=C3)C3=CC=CC(CCN1C)=C23 -6.4 NC(=O)CCC\N=C(\C1=CC=C(Cl)C=C1)C1=C(O)C=CC(F)=C1 -6.8 CC1=NN=C2CN=C(C3=CC=CC=C3Cl)C3=C(C=CC(Cl)=C3)N12 -6.2 CCOC(=O)C=C(C)C=CC=C(C)C=CC1=C(C)C(C)=C(OC)C=C1C -5.2 NS(=O)(=O)C1=C(Cl)C=C2NC(NS(=O)(=O)C2=C1)C(Cl)Cl -5.4 O=C1N(CC2=CC=CC=C2)C2C[S+]3CCCC3C2N1CC1=CC=CC=C1 -5.5 CN(C)CCC[C@]1(OCC2=C1C=CC(=C2)C#N)C1=CC=C(F)C=C1 -5.7 [H][C@@]12CC3=CC=C(O)C=C3[C@@](C)(CCCCC1)[C@H]2N -6.9 CCC(=O)N(C1CCN(CC(O)C2=CC=CC=C2)CC1C)C1=CC=CC=C1 -5.9 [O-][N+](=O)C1=CC2=C(NC(=O)CN=C2C2=CC=CC=C2)C=C1 -5.8 CC(C)[C@H](N)C(=O)OCC(CO)OCN1C=NC2=C1NC(N)=NC2=O -4.7 [H]C(=O)[C@H](NC(C)=O)[C@@H](O)[C@H](O)[C@H](O)CO -6.1 CC(C)NC(=O)NS(=O)(=O)C1=C(NC2=CC=CC(C)=C2)C=CN=C1 -5.7 ClC1=CC=CC=C1C(N1C=CN=C1)(C1=CC=CC=C1)C1=CC=CC=C1 -5.3 OC(=O)COCCN1CCN(CC1)C(C1=CC=C=C=C1)C1=CC=C(C1)C=C1 -5.4

[0109] In step S300, the model is fine-tuned using part of the biological activity experimental data of the current target, which specifically includes the following steps:

[0110] Step S301, according to the biological activity experimental data, each two constitutes a molecular pair, and according to the relative size, it is valued as 0 and 1, (y label =0,1) After processing, the ranking data set used to fine-tune the model is obtained. The obtained data set is shown in Table 2:

[0111] Table 2 Ranking data set example

[0112]

[0113]

[0114] SMILES1 and SMILES2 constitute a molecular pair, and label is the relative biological activity value of the molecular pair SMILES1 and SMILES2. When the biological activity value of SMILES1 is less than that of SMILES2, Label is 1, otherwise, Label is 0;

[0115] Step S302, the SMILES molecular pair in the constructed data set is sent to the model for prediction, and the prediction score is obtained.

[0116] Step S303, according to the prediction score of the model, the binary cross-entropy loss function is used:

[0117] loss=-y label *log(p)-(1-y label )*log(1-p),

[0118] wherein, σ is a hyperparameter, the model is trained.

[0119] After step S400, the entire pre-screening compound library is inferred using the graph transformer GT model, and step S500, the results of the inference are sorted in ascending order, and the example results shown in Table 3 can be obtained:

[0120] Table 3 Model prediction affinity result example

[0121] SMILES SCORE O=c1cc(oc2c3ccccc3ccc12)c1ccccc1 -9.33903 c1cccc2c1c1c(c3cnccc3c(c1[nH]2)C)C -8.79049 Cc1cccc2c1ccc1c2C(=O)C(=O)c2c1occ2C -8.88783 c1ccc2c(c1)C(=O)c1c(C2=O)c(c(cc1)O)O -8.57223 c1ccc2c(c1C)ccc1c2C(=O)C(=O)c2c1occ2C -8.90757 COc1cc2c(ccnc2cc1)C(C1CC2CCN1CC2C=C)O -8.63086 c12c(oc(cc1=O)c1ccccc1)cc(c(c2)O)CC=C -8.60548 c1ccc(cc1)c1cc(=O)c2c(c(c(cc2o1)O)O)O -8.56208 Cc1coc2c1C(=O)C(=O)c1c2ccc2c1CCCC2(C)C -9.02078 c1cccc(c1)C(c1ccc(cc1)c1ccccc1)n1cncc1 -8.98971 c1(c2c(oc(=O)c1)cc1c(c2)CCC1)CN1CCCCC1 -8.81343 C1CC(=C2c3c(CCc4c2nccc4)cc(cc3)Cl)CCN1 -8.74149 c1cc(ccc1c1cc(=O)c2c(c(c(cc2o1)O)O)O)O -8.62634 Cc1cc(c2c(c1)C(=O)c1cc(cc(c1C2=O)O)O)O -8.58374 C[n+]1c2c3nc4ccccc4c3CCn2c(=O)c2ccccc12 -8.9788 c1(c(cc(cc1C)C)C)C(=O) / C=C / c1ccc(cc1)OC -8.80546 C(=O)(NCC(C1CCCCC1)O)c1ccc(cc1)c1ccccc1 -8.77593 C(=C\c1c2c(ncc1)cccc2) / C(=O)c1ccc(cc1)O -8.71441

[0122] According to the screening needs, the top N compounds in the ranking result are candidate compounds.

[0123] In summary, the present application greatly reduces the number of molecules that need to be docked when performing virtual screening on large compound libraries, shortens the time required for virtual screening, and improves the accuracy of virtual screening based on multi-modal technology and the use of biological activity experimental data, thereby significantly reducing the cost and time of drug development.

[0124] The above examples are only used to illustrate the technical solutions of the present application and not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application.

Claims

1. A multimodal drug virtual screening method, characterized in that: The following steps are involved: Step S100: Select a pre-trained graph transformer GT model and fine-tune the graph transformer GT model based on a public dataset. Specifically, the steps include: Step S101: Select data corresponding to a target in the public dataset and use Vina-GPU+ for docking; Step S102: Generate a dataset for fine-tuning the Graphics Transformer (GT) model based on the Vina-GPU+ docking results of the public dataset. Step S103: Load the pre-trained graph transformer GT model and use the mean absolute error (MAE) as the loss function to fine-tune the graph transformer GT model. Step S200: Use active learning strategy to select molecules, perform Vina-GPU+ docking, and use molecular docking scores to fine-tune the Graphics Converter GT model; specifically, the following steps are included: Step S201: Loading the fine-tuned graph transformer GT model; Step S202: Use the GT model to infer the molecules in the pre-screened compound library to obtain an inference score corresponding to each SMILES in the pre-screened compound library; Step S203: sort in ascending order according to the inference score, and add the SMILES of the last 10% of the molecules in the top 20% of the sorted results to the docking list; Step S204: read the molecules in the SMILES list and use Vina-GPU+ for docking; Step S205: Analyze the docking results obtained in step S204, and construct a data set for fine-tuning the model based on the docking results; Step S206: Send the SMILES in the constructed data set to the model, the model encodes the SMILES, and calculates the predicted value , and then compare it with the TARGET value in the data set Use MAE loss function to calculate the loss value; Step S300: fine-tuning the model using the biological activity experimental data of the target; specifically comprising: Step S301: According to the biological activity experimental data, every two molecules form a molecular pair, and are assigned values ​​of 0 and 1 according to their relative sizes. ,After the processing is completed, a sorted dataset is obtained for fine-tuning the model; Step S302: Send the SMILES molecules in the constructed data set into the model for prediction to obtain the prediction score. and ; Step S303: Based on the prediction score of the model and ; Step S400, inferring the molecules in the pre-screened compound library; Step S500: Sort by molecular scores and select the top N molecules as candidate compounds as needed.

2. The multimodal drug virtual screening method according to claim 1, characterized in that: In step S100, a pre-trained Graph Transformer GT model is selected and fine-tuned based on a public dataset. First, Vina-GPU+ docking is used to obtain scores between targets and small molecules in the dataset. The scores are then saved to a folder called "dock" in the working directory. The file name is entered into the configuration file as the dataset for fine-tuning the Graph Transformer GT model.

3. The multimodal drug virtual screening method according to claim 2, characterized in that: The parameters in the configuration file shown are: dataset: configuration related to the dataset; docking_arg: parameter information for Vina-GPU+ docking; In dataset: configuration related to the dataset, more specific parameters are: stage1_1: path to the public dataset used to fine-tune the GT model of the graph transformer. screen_dataset: path information of the molecular library to be screened; bioactivity_dataset.path: path information of the bioactivity experiment dataset; bioactivity_dataset.assay_id: the specific experimental batch used for fine-tuning using the bioactivity assay dataset; In the configuration of docking_arg: Vina-GPU+ docking parameter information, more specific parameters include: receptor: the pathway where the docking receptor is located; center_x, center_y, center_z: the center position of the docking box; size_x, size_y, size_z: the size of the docking box; thread: number of threads.

4. The multimodal drug virtual screening method according to claim 1, wherein: In step S103, the formula of MAE is as follows: ， in, and They are The true value and predicted value of the sample, is the number of samples.

5. The multimodal drug virtual screening method according to claim 1, wherein: Step S204 specifically further includes the following steps: Step S2042: Subprocess 1 reads the molecules in the SMILES list generated in step S203 and generates 3D conformations using pybel (OpenBabel's Python API). The generated file is named ID.pdbqt, where ID represents the index of the current molecule in the SMILES list. Step S2043: Sub-process 1 writes the generated file name into the pipeline, which is read by sub-process 2 and then connected to the pipeline. Step S2044: Subprocess 1 loops through steps S2042 to S2043 until all SMILES lists are read and all 3D structure files are generated. Then, an end string is passed into the pipeline, indicating that no more molecules need to be docked. Step S2045: Subprocess 2 reads the file name in the pipeline and generates a configuration file for docking based on the docking parameter variables in step S100 above. Step S2046: Sub-process 2 calls Vina-GPU+, performs docking according to the configuration file generated in step S2045, and extracts the docking score from the docking result file and stores it in a dictionary with SMILES as the key and the docking score as the value; Step S2047, repeating steps S2045 to S2046 until the end string in the pipeline is read, indicating that there are no subsequent molecules, and returning to the hash table storing the docking results.

6. The multimodal drug virtual screening method according to claim 1, wherein: In step S301, SMILES1 and SMILES2 form a molecular pair, and label is the relative bioactivity value of the molecular pair SMILES1 and SMILES2. When the bioactivity value of SMILES1 is less than the bioactivity value of SMILES2, Label is 1, otherwise, Label is 0.

7. The multimodal drug virtual screening method according to claim 6, characterized in that: In step S303, the binary cross entropy loss function is used: ， in, , is a hyperparameter, Train the model.

Images