Methods, apparatus, media, and program product for drug screening

Abstract

The present application relates to a kind of method, equipment, medium and program product of drug screening.The method includes: for obtaining the plurality of chemical molecules of chemical molecule library, screening is carried out via first deep learning model, so as to obtain the plurality of candidate molecules satisfying predetermined condition;Determine the binding characterization data of each candidate molecule in the plurality of candidate molecules and specified target, so as to select candidate molecule in the plurality of candidate molecules based on the calculated binding characterization data;For candidate molecule, construct molecular feature map, the molecular feature map at least indicates the feature about the molecule, functional group and atom of candidate molecule;And determine the molecular characterization data of candidate molecule based on the constructed molecular feature map, so as to determine the target molecule for drug in candidate molecule based on molecular characterization data.The present application can give consideration to higher screening efficiency, and good versatility and accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to bio-information processing, and in particular, to a method, a computing device, a computer storage medium and a computer program product for drug screening. BACKGROUND

[0002] Existing methods for drug screening include, for example, high-throughput cell or molecular experiment-based screening drug methods, molecular docking-based virtual screening drug methods, and deep learning-based molecular screening methods.

[0003] In the high-throughput cell or molecular experiment-based screening drug method, the method is mainly based on high-throughput molecular level and cell level experiments combined with large-scale data analysis processing procedures, and therefore, it faces the problems of high investment, long cycle and low efficiency.

[0004] In the virtual screening drug method, special molecules for specific targets are screened mainly by calculating the degree of combination of drug molecules and targets in space. However, the molecular docking is too dependent on the selection of targets, and for most cases, the accurate target is not easy to obtain. Therefore, in the case of unknown targets or unknown targets, the molecular docking technology is difficult to play its advantages. Therefore, the virtual screening drug method has poor universality and accuracy

[0005] As for the deep learning-based molecular screening method, it has the technical advantages of high efficiency and good transferability, but the existing deep learning-based method cannot well represent the molecular properties, which leads to the unsatisfactory accuracy of molecular screening. In addition, the existing deep learning-based method has poor interpretability, and it is difficult to provide intuitive explanation for the results.

[0006] In summary, the traditional method for drug screening has the following shortcomings: it is difficult to balance high screening efficiency, good universality and accuracy. SUMMARY

[0007] The present application provides a method, a computing device, a computer storage medium and a computer program product for drug screening, which can balance high screening efficiency, good universality and accuracy.

[0008] According to a first aspect of the present application, there is provided a method for drug screening. The method comprises: screening, via a first deep learning model, for a plurality of chemical molecules of a chemical molecule library, so as to obtain a plurality of candidate molecules satisfying a predetermined condition; determining binding characterization data of each of the plurality of candidate molecules to a specified target, so as to select a candidate molecule from the plurality of candidate molecules based on the calculated binding characterization data; constructing, for the candidate molecule, a molecule feature map indicating at least features of molecules, functional groups, and atoms with respect to the candidate molecule; and determining, based on the constructed molecule feature map, molecule characterization data of the candidate molecule, so as to determine a target molecule for a drug from the candidate molecule based on the molecule characterization data.

[0009] According to a second aspect of the present application, there is also provided a computing device, comprising: a memory configured to store one or more computer programs; and a processor coupled to the memory and configured to execute the one or more programs to cause the apparatus to perform the method of the first aspect of the present application.

[0010] According to a third aspect of the present application, there is also provided a non-transitory computer-readable storage medium. The non-transitory computer-readable storage medium has stored thereon machine executable instructions which, when executed, cause a machine to perform the method of the first aspect of the present application.

[0011] According to a fourth aspect of the present application, there is also provided a computer program product. A computer program is stored thereon, which, when executed by a machine, implements the method of the first aspect of the present application.

[0012] In some embodiments, determining, based on the constructed molecule feature map, the molecule characterization data of the candidate molecule comprises: based on the constructed molecule feature map, extracting, via a heterogeneous graph attention network model, relationship information of nodes and edges of the molecule feature map, so as to aggregate the extracted relationship information of nodes and edges, thereby generating the molecule characterization data with respect to the candidate molecule.

[0013] In some embodiments, aggregating the extracted relationship information of nodes and edges, thereby generating the molecule characterization data with respect to the candidate molecule comprises: aggregating the relationship information of nodes and edges with respect to the molecule nodes, the two-dimensional graph atom nodes, and the three-dimensional graph atom nodes to the functional group nodes, for generating the molecule characterization data with respect to the candidate molecule.

[0014] In some embodiments, constructing the molecular feature graph for the candidate molecule comprises: taking the molecule of the candidate molecule, each functional group in the candidate molecule, each atom in a two-dimensional chemical structure of the candidate molecule, and each atom in a three-dimensional atomic space conformation of the candidate molecule as nodes in the molecular feature graph respectively, so as to construct a node set; associating each node in the node set with a mapping function of a corresponding node type in the set of node types; constructing a connection relationship between the molecule nodes, the functional group nodes, the two-dimensional graph atom nodes, and the three-dimensional graph atom nodes by edges, so as to form an edge set; and associating each edge in the edge set with a mapping function of a corresponding edge type in the set of edge types.

[0015] In some embodiments, determining the molecular representation data of the candidate molecule based on the constructed molecular feature graph comprises: initializing each node in the molecular feature graph as a vector, so as to generate an initial feature of the node, the dimensions of the initial features of different types of nodes being different; mapping the initial features of different nodes into the same dimension; and performing message passing and aggregation on the mapped initial features of the nodes, so as to obtain the molecular representation data.

[0016] In some embodiments, performing message passing and aggregation on the mapped initial features of the nodes, so as to obtain the molecular representation data comprises: calculating an attention coefficient of a current node in the molecular feature graph and other nodes connected to the current node; aggregating the features of the neighboring nodes of the current node for updating the feature of the current node; and determining the molecular representation data of the candidate molecule based on the updated feature of the node.

[0017] In some embodiments, determining the molecular representation data of the candidate molecule based on the updated feature of the node comprises: selecting the feature of the aggregated and updated functional group node; and generating the molecular representation data of the candidate molecule via a max-pooling operation on the feature of the aggregated and updated functional group node.

[0018] In some embodiments, selecting the candidate molecule from the plurality of candidates based on the calculated binding representation data comprises: determining the binding representation data of each candidate molecule to the specified target, so as to sort the candidate molecules based on the calculated binding representation data; and selecting a predetermined number of candidate molecules with high rankings as the candidate molecules based on the sorting result.

[0019] In some embodiments, the method for drug screening further comprises: in response to determining that the ratio of the depth between the mitochondrial fusion breakpoints and the depth of other regions is greater than a predetermined ratio threshold, determining that there is a variant site and a region; and screening the determined variant site and region based on a predetermined variant database, so as to obtain a variant site with pathogenicity.

[0020] In some embodiments, the determined target molecule is a target molecule for anti-aging

[0021] The summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. The summary is not intended to identify key features or essential features of the application, and is not intended to limit the scope of the application. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 A schematic diagram of a computing device for implementing a method for drug screening according to embodiments of the application is shown.

[0023] Figure 2 A flowchart of a method for implementing a method for drug screening according to embodiments of the application is shown.

[0024] Figure 3 A schematic diagram of a molecular profile according to embodiments of the application is shown.

[0025] Figure 4 A schematic diagram of a lateral comparison of a method for drug screening according to embodiments of the application with other methods is shown.

[0026] Figure 5 A flowchart of a method for constructing a molecular profile according to embodiments of the application is shown.

[0027] Figure 6 A flowchart of a method for determining molecular profiling data of a candidate molecule according to embodiments of the application is shown.

[0028] Figure 7 A schematic diagram of a method for determining molecular profiling data of a candidate molecule according to embodiments of the application is shown.

[0029] Figure 8 A flowchart of a method for molecular profiling data via node messaging and aggregation in order to obtain molecular profiling data according to embodiments of the application is shown.

[0030] Figure 9 A schematic diagram of a method for determining molecular profiling data of a candidate molecule according to further embodiments of the application is shown.

[0031] Figure 10 A block diagram of an electronic device suitable for use in implementing embodiments of the application is shown schematically.

[0032] Figure 11 A molecular schematic of a target molecule for a drug determined according to some embodiments of the application is shown.

[0033] Figure 12A schematic diagram of a molecule of a target molecule for a drug determined according to some embodiments of the present application is shown. Figure 13 A schematic diagram of a molecule of a target molecule for a drug determined according to some embodiments of the present application is shown.

[0034] Figure 14 A schematic diagram of a molecular profile constructed according to some embodiments of the present application is shown.

[0035] Figure 15 A schematic diagram of a molecular profile constructed according to some embodiments of the present application is shown.

[0036] In the various drawings, like or corresponding elements are denoted by like or corresponding reference numerals. DETAILED DESCRIPTION

[0037] Preferred embodiments of the present application will be described herein below with reference to the accompanying drawings. While the preferred embodiments of the present application are shown in the drawings, it is understood that the present application can be embodied in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0038] The term "comprising" and variations thereof as used herein are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements is not necessarily limited to those elements, but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Unless otherwise specified, the term "or" as used herein is intended to mean "and / or". The term "based on" means "based, at least in part, on". The terms "one example embodiment" and "an example embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one additional embodiment". The terms "a first", "a second", "the first", "the second", and the like, do not imply that there are only one of these items, but rather that these items are at least one of the items.

[0039] As described previously, the conventional method for drug screening has the disadvantage that it is difficult to balance high screening efficiency with good universality and accuracy.

[0040] To at least partially address one or more of the above-mentioned problems and other potential problems, example embodiments of the present application propose a method for drug screening. In the method, the candidate molecules screened by a deep learning model are re-screened by a method of molecular docking, on the one hand to further narrow down the range of candidate molecules, significantly improving the efficiency of screening, and on the other hand to provide more explainable basis for the entire screening process; and the present application obtains more modal information (molecular global information, molecular pharmacophore information, and atomic local information of the molecule) of the candidate molecules through the multi-view structure of the molecular feature map, so that the molecular representation data determined based on the molecular feature map can better represent the information of the candidate molecules, improving the generalization ability and screening accuracy of the properties of the candidate molecules. Thus, the present application can balance high screening efficiency, good versatility and accuracy.

[0041] Figure 1 A schematic diagram of a computing device for a method of drug screening according to an embodiment of the present application is shown.

[0042] Regarding the computing device 100, it is used for drug screening, for example. Specifically, the computing device 100 can screen, for a plurality of chemical molecules of a chemical molecule library, via a first deep learning model, to obtain a plurality of candidate molecules satisfying a predetermined condition; and determine binding representation data of each of the plurality of candidate molecules to a specified target, so as to select a candidate molecule from the plurality of candidate molecules based on the calculated binding representation data. In addition, the computing device 100 can construct, for the candidate molecule, a molecular feature map indicating at least features of molecules, functional groups, and atoms of the candidate molecule; and determine molecular representation data of the candidate molecule based on the constructed molecular feature map, so as to determine a target molecule for a drug in the candidate molecule based on the molecular representation data.

[0043] In some embodiments, the computing device 100 can have one or more processing units, including special-purpose processing units such as GPUs, FPGAs, and ASICs, and general-purpose processing units such as CPUs. In addition, one or more virtual machines can also be running on each computing device. The computing device 100 includes, for example: a candidate molecule screening unit 112, a binding representation data calculation and candidate molecule selection unit 114, a molecular feature map construction unit 116, and a target molecule determination unit 118. The above-mentioned candidate molecule screening unit 112, binding representation data calculation and candidate molecule selection unit 114, molecular feature map construction unit 116, and target molecule determination unit 118 can be configured on one or more computing devices 100.

[0044] As to the candidate molecule screening unit 112, it is configured to screen, via the first deep learning model, against a plurality of chemical molecules of an acquired chemical molecule library, so as to obtain a plurality of candidate molecules satisfying a predetermined condition.

[0045] As to the binding characterization data calculation and candidate molecule selection unit 114, it is configured to determine binding characterization data of each of the plurality of candidate molecules to a specified target, so as to select a candidate molecule from the plurality of candidates based on the calculated binding characterization data.

[0046] As to the molecular feature map construction unit 116, it is configured to construct, for the candidate molecule, a molecular feature map indicating at least features of molecules, functional groups, and atoms of the candidate molecule.

[0047] As to the target molecule determination unit 118, it is configured to determine, based on the constructed molecular feature map, molecular characterization data of the candidate molecule, so as to determine a target molecule for a drug from the candidate molecule based on the molecular characterization data.

[0048] The method for drug screening according to the embodiments of the present application will be described below in conjunction with Figure 2 、 Figure 3 A method for drug screening 200 according to the embodiments of the present application is shown in a flowchart. Figure 2 A method for drug screening 200 according to the embodiments of the present application is shown in a flowchart. Figure 3 A schematic diagram of a molecular feature map 300 according to the embodiments of the present application is shown. The method 200 may, for example, be performed at the electronic device 1000 described. Figure 10 The method 200 may, for example, also be performed at the computing device 100 described. It should be understood that the method 200 can also include additional actions not shown and / or can omit actions shown, without limitation in this respect. Figure 1 The method 200 may, for example, also be performed at the computing device 100 described. It should be understood that the method 200 can also include additional actions not shown and / or can omit actions shown, without limitation in this respect.

[0049] At step 202, the computing device 100 screens, via the first deep learning model, against a plurality of chemical molecules of an acquired chemical molecule library, so as to obtain a plurality of candidate molecules satisfying a predetermined condition.

[0050] For example, the computing device 100 first inputs all or part of the chemical molecules of the chemical molecule library into the trained first deep learning model for preliminary screening, so as to screen out candidate molecules satisfying the predetermined condition. The first deep learning model is, for example, trained via a large number of samples of candidate molecules having effective ingredients (e.g., effective for a certain disease, or effective for anti-aging).

[0051] As to the candidate molecules satisfying the predetermined condition, they are, for example, chemical molecules having effective ingredients. They are, for example, chemical molecules containing effective ingredients effective for delaying aging.

[0052] At step 204, the computing device 100 determines binding characterization data of each of the plurality of candidate molecules to the specified target, so as to select a candidate molecule from the plurality of candidates based on the calculated binding characterization data.

[0053] As to the method of calculating the binding characterization data (e.g., binding score) of the candidate molecule to the predetermined target, it may, for example, include that the computing device 100 acquires the target points in the specific pathway in the known literature; retrieves the spatial structure and the information of possible binding sites of the corresponding target points in the Protein DataBank (PDB, which is a database specially collecting the three-dimensional structure data of proteins and nucleic acids); and calculates the binding characterization data of the candidate molecule to the specified target based on the spatial structure and the information of possible binding sites. In some embodiments, the MGLTools and AutoDock Vina software for molecular docking may be used to calculate the binding characterization data of the candidate molecule to the specified target. It should be understood that AutoDock Vina is an existing molecular docking software which is used based on the MGLTools toolkit. The MGLTools toolkit includes AutoDock Tools and Python Molecular Viewer. AutoDock Tools is used to generate input files for Vina, and Python Molecular Viewer is used to view the results.

[0054] As to the method of selecting the candidate molecule, it may, for example, include that the binding characterization data of each of the candidate molecules to the specified target is determined, so as to sort the candidate molecules based on the calculated binding characterization data; and a predetermined number of candidate molecules with high ranking are selected as the candidate molecule based on the sorting result.

[0055] At step 206, the computing device 100 constructs a molecular feature map for the candidate molecule, the molecular feature map indicating at least the features of the molecule, functional groups, and atoms of the candidate molecule. It should be understood that the molecular feature map constructed by the present application indicates more information about the candidate molecule, and involves information of multiple types (rather than a single type) of molecules, functional groups, and atoms, thereby facilitating more accurate and comprehensive identification of the properties of the candidate molecule.

[0056] Regarding the molecular feature of the candidate molecule, it is, for example, a molecular view of the candidate molecule. In some embodiments, the molecular view is generated based on a SMILES sequence of the candidate molecule via a pre-trained model, for example, but not limited to, a MOLBERT model. MOLBERT uses a Bert model architecture to learn global representations of molecules by masking atoms or other characters in the predicted molecular sequence using the Simplified molecular input line entry specification (SMILES) sequence information of the molecule. It should be understood that the SMILES sequence information is in the form of a string representing the molecular structure. The SMILES encoding corresponds to the chemical structure. It uses the letters C, H, O, N, S, etc. to represent the corresponding atoms. The two adjacent atoms represent the single bond connection. The double bond connection between atoms is represented by “=”. The triple bond connection between atoms is represented by “#”. The brackets are used to represent the branched atoms, etc.

[0057] Regarding the functional group feature of the candidate molecule, it is, for example, a functional group view of the candidate molecule, which is used to indicate the properties of the functional group of the candidate molecule. It should be understood that RDKit has built-in multiple molecular fingerprint calculation methods. For example, the computing device 100 calculates the MACCS fingerprint of the functional group using the “rdkit.Chem.MACCSkeys.GenMACCSKeys()” function built-in RDKit, and obtains the characteristics of the functional group using the “factory.GetFeaturesForMol(m)” function built-in RDKit. Based on the calculated MACCS fingerprint of the functional group and the characteristics of the functional group, the functional group feature of the candidate molecule can be obtained.

[0058] Regarding the atomic feature of the candidate molecule, it includes at least one of a two-dimensional atomic view of the candidate molecule and a three-dimensional atomic view of the candidate molecule.

[0059] Regarding the two-dimensional atomic view of the candidate molecule, it is, for example, constructed according to the chemical structural formula of the candidate molecule. For example, the atoms of the candidate molecule are set as points, and the chemical bonds of the candidate molecule are set as edges, so as to construct the two-dimensional atomic view of the candidate molecule.

[0060] Regarding the three-dimensional atomic view of the candidate molecule, it is constructed, for example, by a molecular space conformation generation model (e.g., without limitation, a known GEODIFF model). For example, first, the computing device 100 utilizes the GEODIFF model to simulate the process of a molecule from a chaotic spatial geometry to a stable spatial conformation using a diffusion model, so as to obtain a three-dimensional molecular space conformation containing the three-dimensional coordinates of each atom in the molecule. Then, the atoms in the three-dimensional molecular space conformation are configured as nodes, and the Euclidean distance of the atoms in the three-dimensional molecular space conformation generated by the GEODIFF model is calculated. It is determined whether the calculated Euclidean distance is less than a predetermined value; if the calculated Euclidean distance is less than the predetermined value (e.g., the predetermined value is 4), an edge connection is configured between the atoms. After that, the configured nodes and edge connections generate the three-dimensional atomic view of the candidate molecule. Regarding the atomic representation in the three-dimensional atomic view of the candidate molecule, it is generated, for example, by pre-training of the Uni-Mol model. For example, the computing device 100 utilizes the Uni-Mol model to add fixed spatial position encoding, atom pair information, and molecular smiles sequence information into a Transformer architecture to output the spatial structure and atomic representation of the molecule. Since the spatial structure of the molecule is utilized in the training process, the atomic representation (512 dimensions) output by the Uni-Mol model contains the spatial conformation information of the molecule.

[0061] In some embodiments, the three-dimensional atomic view of the candidate molecule indicates a three-dimensional atomic space conformation of the candidate molecule.

[0062] Regarding the molecular feature map, in some embodiments, it includes: a node set, an edge set, a set of node types, and a set of edge types, the node set includes a molecule node, a functional group node, a two-dimensional graph atom node, and a three-dimensional graph atom node. As Figure 3As shown, the molecular feature map 300 is a heterogeneous topology graph about nodes and edges. The node on the top layer of the molecular feature map 300 is a molecule node 310. The molecule node 310 represents the whole molecule of the candidate molecule. The second layer of the molecular feature map 300 includes one or more functional group nodes 312. It should be understood that a functional group is an atom or group of atoms that determines the chemical properties of an organic compound. Functional groups include, for example, carbon-carbon double bond, carbon-carbon triple bond, hydroxyl group, carboxyl group, ether bond, aldehyde group, benzene ring, etc. The present application projects each of all functional groups in a candidate molecule into a functional group node. Each molecule node is connected to each functional group node by an edge (e.g., as shown by the label 314). If there is an atomic connection between functional groups, the functional group nodes with the atomic connection are connected by an edge (e.g., as shown by the label 316). Atomic nodes are divided into two types. For example, the left half of the third layer of the molecular feature map 300 indicates the two-dimensional chemical structure of the candidate molecule. Among them, the two-dimensional graph atomic node 318 indicates the atom in the two-dimensional chemical structure of the candidate molecule, and the edge (e.g., as shown by the label 320) connected to the two-dimensional graph atomic node indicates the chemical bond of the atom. The right half of the third layer of the molecular feature map 300 indicates the spatial conformation of the atom of the candidate molecule. The atom in the spatial conformation is a three-dimensional graph atomic node 324.

[0063] The following formula (1) shows the functional expression of the molecular feature map.

[0064]

[0065] In the above formula (1), V represents the node set of the molecular feature map. E represents the edge set of the molecular feature map. represents the set of node types of the molecular feature map. represents the set of edge types of the molecular feature map.

[0066] It should be understood that in the molecular feature map, there are four types of nodes, namely M, P, A1, and A2. Figure 14 The schematic diagram of the molecular feature map constructed according to some embodiments of the present application is shown. The following takes the molecular feature map constructed for the molecule "CC=CC(=O)N(CC)c1ccccc1C" as an example to specifically explain the structure of the molecular feature map and the construction method thereof. Figure 14 The molecular node M represents the whole molecule of the candidate molecule (as indicated by the label 1402 in FIG. 14B). There is only one molecule node in the molecular feature map of one candidate chemical molecule. The functional group node P represents each functional group as a node (as indicated by the label 1404 in FIG. 14B). The two-dimensional graph atomic node A1 represents the atom in the two-dimensional chemical structure of the candidate molecule (as indicated by the label 1406 in FIG. 14B). The three-dimensional graph atomic node A2 represents the atom in the spatial conformation of the candidate molecule (as indicated by the label 1408 in FIG. 14B). Figure 14 Figure 14 Figure 14 ​​A2 represents three-dimensional graph atom nodes, each atom represents a node (as indicated by label 1406 in FIG. 14B). Figure 14 A2 represents three-dimensional graph atom nodes, each atom represents a node (as indicated by label 1408 in FIG. 14B).

[0067] In some embodiments, in the molecular feature graph, there are six types of edges, which are used to indicate the connection relationship between the objects (i.e., nodes) connected at both ends of the edge. The edges are, for example, M-P, P-P, P-A1, P-A2, A1-A1, and A2-A2 edges respectively. Among them, M-P represents the connection relationship between the molecule node and the functional group node (as indicated by label 1412 in FIG. 14C). Figure 14 In some embodiments, the computing device 100 connects all functional group nodes (i.e., P) to the molecule node (i.e., M). P-P represents the connection relationship between the functional group nodes (as indicated by label 1414 in FIG. 14C). P-A1 represents the connection relationship between the two-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1416 in FIG. 14C). P-A2 represents the connection relationship between the three-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1418 in FIG. 14C). A1-A1 represents the chemical bond in the molecule of the candidate molecule (as indicated by label 1420 in FIG. 14C). A2-A2 represents the distance relationship between the atoms in the candidate molecule in space (as indicated by label 1422 in FIG. 14C). Figure 14 In some embodiments, the computing device 100 connects all functional group nodes (i.e., P) to the molecule node (i.e., M). P-P represents the connection relationship between the functional group nodes (as indicated by label 1414 in FIG. 14C). P-A1 represents the connection relationship between the two-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1416 in FIG. 14C). P-A2 represents the connection relationship between the three-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1418 in FIG. 14C). A1-A1 represents the chemical bond in the molecule of the candidate molecule (as indicated by label 1420 in FIG. 14C). A2-A2 represents the distance relationship between the atoms in the candidate molecule in space (as indicated by label 1422 in FIG. 14C). Figure 14 In some embodiments, the computing device 100 connects all functional group nodes (i.e., P) to the molecule node (i.e., M). P-P represents the connection relationship between the functional group nodes (as indicated by label 1414 in FIG. 14C). P-A1 represents the connection relationship between the two-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1416 in FIG. 14C). P-A2 represents the connection relationship between the three-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1418 in FIG. 14C). A1-A1 represents the chemical bond in the molecule of the candidate molecule (as indicated by label 1420 in FIG. 14C). A2-A2 represents the distance relationship between the atoms in the candidate molecule in space (as indicated by label 1422 in FIG. 14C). Figure 14 In some embodiments, the computing device 100 connects all functional group nodes (i.e., P) to the molecule node (i.e., M). P-P represents the connection relationship between the functional group nodes (as indicated by label 1414 in FIG. 14C). P-A1 represents the connection relationship between the two-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1416 in FIG. 14C). P-A2 represents the connection relationship between the three-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1418 in FIG. 14C). A1-A1 represents the chemical bond in the molecule of the candidate molecule (as indicated by label 1420 in FIG. 14C). A2-A2 represents the distance relationship between the atoms in the candidate molecule in space (as indicated by label 1422 in FIG. 14C). Figure 14 In some embodiments, the computing device 100 connects all functional group nodes (i.e., P) to the molecule node (i.e., M). P-P represents the connection relationship between the functional group nodes (as indicated by label 1414 in FIG. 14C). P-A1 represents the connection relationship between the two-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1416 in FIG. 14C). P-A2 represents the connection relationship between the three-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1418 in FIG. 14C). A1-A1 represents the chemical bond in the molecule of the candidate molecule (as indicated by label 1420 in FIG. 14C). A2-A2 represents the distance relationship between the atoms in the candidate molecule in space (as indicated by label 1422 in FIG. 14C). Figure 14 In some embodiments, the computing device 100 connects all functional group nodes (i.e., P) to the molecule node (i.e., M). P-P represents the connection relationship between the functional group nodes (as indicated by label 1414 in FIG. 14C). P-A1 represents the connection relationship between the two-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1416 in FIG. 14C). P-A2 represents the connection relationship between the three-dimensional graph atom node and the corresponding attributed functional group node (as indicated by label 1418 in FIG. 14C). A1-A1 represents the chemical bond in the molecule of the candidate molecule (as indicated by label 1420 in FIG. 14C). A2-A2 represents the distance relationship between the atoms in the candidate molecule in space (as indicated by label 1422 in FIG. 14C).

[0068] Regarding the method of constructing the molecular feature graph, it may, for example, include: the computing device 100 taking the molecule of the candidate molecule, each functional group in the candidate molecule, each atom in the two-dimensional chemical structure of the candidate molecule, and each atom in the three-dimensional atomic space conformation of the candidate molecule as nodes in the molecular feature graph respectively, so as to construct a node set; associating each node in the node set with a mapping function of the corresponding node type in the set of node types; constructing the connection relationship between the molecule node, the functional group node, the two-dimensional graph atom node, and the three-dimensional graph atom node by means of edges, so as to form an edge set; and associating each edge in the edge set with a mapping function of the corresponding edge type in the set of edge types. The specific implementation of the method 400 for constructing the molecular feature graph will be described below in detail. Figure 4 Here, the specific implementation of the method 400 for constructing the molecular feature graph will not be described again.

[0069] At step 208, the computing device 100 determines the molecular representation data of the candidate molecule based on the constructed molecular feature graph, so as to determine the target molecule for the drug in the candidate molecule based on the molecular representation data.

[0070] In some embodiments, the candidate molecule determined via the method 200 is, for example, O=C1C=2C=CC=CC2N(C=3OC(=C(C)CC4C5=C(OC(C)(C)C4)C=6C=CC=CC6N(C5=O)C)CC13)C; and the target molecule determined for the drug is, for example, Figure 11 as shown.

[0071] In other embodiments, the candidate molecule determined via the method 200 is, for example, O=C(C=CC1=CC=C(OC)C(=C1)COC=2C=CC=CC2C3CCCCC3)C4=CC=C(O)C=C4; and the target molecule determined for the drug is, for example, Figure 12 as shown. In still other embodiments, the candidate molecule determined via the method 200 is, for example, FC(F)(F)C1=CC=C(C=C1)CN2CCC(C3=NC=4C=C(C=CC4N3)C)C2; and the target molecule determined for the drug is, for example, Figure 13 as shown.

[0072] As to the method of generating the molecular representation data, in some embodiments, it includes, for example, the computing device 100 extracting, based on the constructed molecular feature graph, the relationship information of the nodes and edges of the molecular feature graph via the heterogeneous graph graph attention network model, so as to aggregate the extracted relationship information of the nodes and edges, thereby generating the molecular representation data about the candidate molecule. Specifically, for example, as shown in Figure 7 the computing device 100 aggregates the relationship information of the nodes and edges about the molecular nodes 720, the two-dimensional graph atom nodes 726, and the three-dimensional graph atom nodes 728 to the functional group nodes 724 for generating the molecular representation data 740 about the candidate molecule. Then, based on the molecular representation data 740, the property of the candidate molecule (e.g., determining whether it is effective for delaying aging) is judged via the trained second neural network model 742 (e.g., MLP network), so as to screen out the target molecule from the candidate molecules.

[0073] The specific implementation of the method of generating the molecular representation data will be described below in combination with Figure 5 The specific implementation of the method of generating the molecular representation data will be described below in combination with

[0074] To verify the effect of the present application, the present application selects the data of an anti-aging dataset, for example but not limited to the experimentally verified anti-aging molecular dataset provided by Wong F et al. The data of the anti-aging dataset used contains a total of 2353 different small molecule drugs, of which 45 molecules are positive. The 45 positive molecules represent their ability to eliminate senescent cells. In order to make the test results more representative, the present application divides the anti-aging dataset using the Scaffold dataset division method, wherein the proportions of the training set, the validation set and the test set are 7:1:2 respectively. In order to accurately compare and evaluate the drug screening effect of the method of the present application and other methods, the present application uses five-fold cross-validation, that is, the data is divided into 5 equal parts, and each time an experiment is taken for testing, the rest is used for training. The experiment is repeated 5 times to obtain the average value. Moreover, the present application uses the calculated average AUC (Area under Curve, the area under the Roc curve) as the evaluation standard. As a comparative method, for example, are: known D-MPNN-based method, FraGAT model-based method, CoMPT model-based method, ABT-MPNN model-based method, and PharmHGT model-based method. Figure 4 A transverse comparison diagram of the method for drug screening according to the embodiment of the present application and other methods is shown. Mark 410 indicates the average AUC according to the method of the present application for the test set. Mark 412 indicates the average AUC according to the D-MPNN model-based method for the test set. Mark 414 indicates the average AUC according to the FraGAT model-based method for the test set. Mark 416 indicates the average AUC according to the CoMPT model-based method for the test set. Mark 418 indicates the average AUC according to the ABT-MPNN model-based method for the test set. Mark 420 indicates the average AUC according to the PharmHGT model-based method for the test set. It can be seen from Figure 4 It can be seen that the method of the present application shows better performance relative to the comparative method, that is, even if the training set is only 70%, the average AUC on the test set still reaches 80%. Compared with the method based on other models (for example, D-MPNN), the effect is improved by 15%.

[0075] In the above method, the application selects the candidate molecule based on the binding representation data of the candidate molecule and the specified target selected by the first deep learning model, constructs a molecular feature graph including atoms, functional group views, and molecular representation for the candidate molecule, and determines the molecular representation data of the candidate molecule based on the molecular feature graph, thereby determining the target molecule for the drug. The application can further screen the candidate molecule screened by the deep learning model through the molecular docking method, which can further reduce the range of the candidate molecule, significantly improve the screening efficiency, and provide more explainable basis for the entire screening process. Moreover, the application obtains more modal information (molecular global information, molecular pharmacophore information, and atomic local information of the molecule) of the candidate molecule through the multi-view structure of the molecular feature graph, so that the molecular representation data determined based on the molecular feature graph can better represent the information of the candidate molecule, improve the generalization ability of the method, and improve the screening accuracy of the properties of the candidate molecule. Therefore, the application can balance the high screening efficiency and good universality and accuracy.

[0076] The following will combine Figure 5 A method of constructing a molecular feature graph according to an embodiment of the application is described. Figure 5 A flowchart of a method 500 of constructing a molecular feature graph according to an embodiment of the application is shown. It should be understood that the method 500 may, for example, be performed at Figure 10 The electronic device 1000 described. It can also be performed at Figure 1 The computing device 100 described. It should be understood that the method 500 can also include additional actions not shown and / or can omit actions shown, without limitation in this regard.

[0077] At step 502, the computing device 100 constructs a node set by taking the molecule of the candidate molecule, each functional group in the candidate molecule, each atom in the two-dimensional chemical structure of the candidate molecule, and each atom in the three-dimensional atomic space conformation of the candidate molecule as a node in the molecular feature graph, respectively. It should be understood that by simultaneously constructing the atomic nodes of the two-dimensional view and the three-dimensional view, the application can more accurately and stably identify the properties of the candidate molecule.

[0078] For example, as Figure 3 shown, the computing device 100 constructs a molecular node (e.g., shown by label 310) for the entire molecule of each candidate molecule; constructs each of the functional groups in the candidate molecule into a functional group node (e.g., shown by label 312); constructs the atoms of the two-dimensional chemical structure of the candidate molecule into a two-dimensional graph atom node (e.g., shown by label 318); and constructs the atoms of the atomic space conformation of the candidate molecule into a three-dimensional graph atom node (e.g., shown by label 324). ​

[0079] At step 504, the computing device 100 causes each node in the set of nodes to be associated with a mapping function of a corresponding node type in the set of node types.

[0080] It should be appreciated that each node in the molecular feature map belongs to the set of nodes. That is, v e V (where v represents a node and V represents the set of nodes). Each node in the molecular feature map is associated with its corresponding node type mapping function φ(v). That is, (where V represents the set of nodes, and φ represents the set of node types).

[0081] At step 506, the computing device 100 constructs connections between the molecular nodes, the functional group nodes, the two-dimensional graph atom nodes, and the three-dimensional graph atom nodes by edges to form a set of edges.

[0082] As to the way of connecting between the functional group nodes (i.e., P) and the molecular nodes (i.e., M), it for example includes that the computing device 100 connects all the functional group nodes (i.e., P) with the molecular nodes (i.e., M) by edges.

[0083] As to the way of connecting between the functional group nodes and the functional group nodes, it for example includes that if the computing device 100 determines that there is an atom connecting any two functional groups by a chemical bond, then the two functional group nodes are connected by an edge.

[0084] As to the way of connecting between the two-dimensional graph atom nodes and the functional group nodes, it for example includes that the computing device 100 connects the two-dimensional graph atom nodes with their corresponding attributed functional group nodes by edges.

[0085] As to the way of connecting between the two-dimensional graph atom nodes, it for example includes that if there is a chemical bond between the atoms corresponding to the two-dimensional graph atom nodes, then the two-dimensional graph atom nodes are connected by an edge (e.g., as shown in marker 320).

[0086] As to the way of connecting between the three-dimensional graph atom nodes and the functional group nodes, it for example includes that the computing device 100 connects the three-dimensional graph atom nodes with their corresponding attributed functional group nodes by edges. For example, as shown in Figure 3 if a functional group has a connection with an atom, then the corresponding functional group node and the two-dimensional graph atom node, or the three-dimensional graph atom node are connected by an edge (e.g., as shown in markers 322, 328).

[0087] As to the connection manner between the three-dimensional graph atomic nodes, it for example includes: if the computing device 100 determines that the spatial distance between the atoms corresponding to the three-dimensional graph atomic nodes is less than a set threshold, then connecting between the corresponding three-dimensional graph atomic nodes through an edge. For example, as shown in Figure 3 if the distance between the atoms corresponding to the three-dimensional graph atomic nodes is less than a set threshold (for example, but not limited to, 4), then the three-dimensional graph atomic nodes of the corresponding atoms are connected in the form of an edge (for example, as shown by the mark 326).

[0088] At step 508, the computing device 100 causes each edge in the edge set to be associated with a mapping function of a corresponding edge type in the edge type set.

[0089] Each edge in the molecular feature graph belongs to the edge set. That is, e ∈ E (where e represents an edge, and E represents an edge set). Each edge in the molecular feature graph is also associated with an edge type mapping function ψ(e) to which it belongs. That is, (where E represents an edge set, and the edge type set is represented by T). Figure 15 A schematic diagram of a molecular feature graph constructed according to another embodiment of the present application is shown. Figure 15 A schematic diagram of a constructed molecular feature graph constructed for a candidate molecule “COc1cc(O)cc(O)c1C(=O)c1ccccc1” is shown.

[0090] In the above scheme, the molecular feature graph constructed by the present application includes a multi-view model, which not only contains molecular global information, molecular pharmacophore information, and atomic information of the molecule, but also includes both structural atomic view and spatial atomic view in the atomic information. Thus, the constructed molecular feature graph can obtain more levels of information of the molecule and can better represent the molecular information.

[0091] The method for determining molecular representation data of a candidate molecule according to an embodiment of the present application will be described below in combination with Figure 6 and Figure 7 The method for determining molecular representation data of a candidate molecule according to an embodiment of the present application will be described below in combination with Figure 6 A flowchart of the method 600 for determining molecular representation data of a candidate molecule according to an embodiment of the present application is shown. Figure 7 A schematic diagram of the method for determining molecular representation data of a candidate molecule according to an embodiment of the present application is shown. It should be understood that the method 600 may, for example, be executed at the electronic device 1000 described above. It may, for example, also be executed at the computing device 100 described above. Figure 10 It should be understood that the method 600 may, for example, also include additional actions not shown and / or the actions shown may be omitted, and the scope of the present application is not limited in this respect. Figure 1 It should be understood that the method 600 may, for example, also include additional actions not shown and / or the actions shown may be omitted, and the scope of the present application is not limited in this respect.

[0092] At step 602, the computing device 100 initializes each node in the molecular feature graph as a vector to generate initial features of the node, and the dimensions of the initial features of different types of nodes are different.

[0093] For example, the computing device 100 initializes each node v i (1≤i≤N n ) in the candidate molecular feature graph as a corresponding node vector x i to generate initial features of the node. represents the initial features of the node. N n represents the number of all types of nodes in the candidate molecule. d k represents the initialization dimension of different types of nodes. The following Table 1 schematically shows the initial feature dimensions of different types of nodes.

[0094] Table 1

[0095]

[0096]

[0097] Regarding the method of generating the initial features of the node, for example, it includes that the computing device 100 uses different patterns to construct the corresponding node vector for different types of nodes.

[0098] Specifically, in some embodiments, for the molecular node (M), the computing device 100 initializes, for example, using the pre-trained model MOLBERT to form the corresponding node vector. It should be understood that the MOLBERT pre-trained model uses the Bert model architecture, and learns the global representation of the molecule by masking the atoms or other characters in the predicted molecular sequence using the SMILES sequence information of the molecule.

[0099] For the functional group node (P), the computing device 100 initializes, for example, using the RDKit software, using the molecular fingerprint and one-hot encoding representation. It should be understood that the molecular fragment is a group of connected atoms and can contain relevant functional groups. A series of tools for analyzing and operating molecular fragments are provided in the RDKit software. For example, the functional group library: “FunctionalGroups.txt” in the RDConfig.RDDataDir directory can be used to instantiate a parameterizer: FragmentCatalog.FragCatParams().

[0100] For two-dimensional graph atom nodes (A1), the computing device 100 constructs a corresponding node vector based on a plurality of atomic information (atomic type, atomic degree, atomic number, atomic formal charge, chirality, hybridization, whether there is an aromatic substance, mass, number of connected H atoms, etc.), all types of atomic information are encoded using one-hot, for example. Table 2 below schematically shows the feature dimensions corresponding to different atomic information.

[0101] Table 2

[0102]

[0103]

[0104] For three-dimensional graph atom nodes (A2), the computing device 100 generates a corresponding node vector using a Uni-Mol model. It should be understood that the node vector corresponding to the three-dimensional graph atom node generated via the Uni-Mol model not only contains information about the atom in the molecular sequence, but also contains distribution information about the atom in the molecular space.

[0105] At step 604, the computing device 100 maps the initial features of different nodes into the same dimension.

[0106] For example, the computing device 100 uses different multilayer perceptron (MLP) structures to map the initial features of different nodes into the same dimension. Figure 7 In the above description of the candidate molecule, the dimensions of the corresponding node initial features of different types of nodes (molecular nodes, functional group nodes, two-dimensional graph atom nodes, and three-dimensional graph atom nodes) in the molecular feature graph 710 are mapped into the same dimension.

[0107] At step 606, the computing device 100 performs message passing and aggregation on the initial features of the mapped nodes to obtain molecular representation data.

[0108] In some embodiments, the computing device 100 uses a Heterogeneous Graph Transformer (HGT) model to perform message passing and aggregation on the initial features of the mapped nodes to obtain molecular representation data. As shown in the above description of the candidate molecule, the HGT model is used to perform message passing and aggregation on the initial features of the mapped nodes to obtain molecular representation data. Figure 7 The message passing and aggregation process for the initial features of each node is schematically shown in the mark 720 in the above description of the candidate molecule. It should be understood that the molecular feature graph is a heterogeneous graph. The heterogeneous mutual attention of the HGT model is defined by decomposing each edge e = (s, t) based on the meta-relation triple. The following formula (2) schematically shows the function expression of the edge. Formula (3) schematically shows the expression of the meta-relation of the edge.

[0109] e = (s, t) (2)

[0110] <φ(s),ψ(e),φ(t)> (3)

[0111] In the above equations (2) and (3), s, t represent two nodes in the heterogeneous graph, respectively,

[0112] φ(s), φ(t) represent the mapping of the node type of the two nodes s, t, respectively. ψ(e) represents the mapping of the edge type of the edge e. As described above, the mapping of the node type has four kinds, which correspond to M, P, A1, A2, respectively. The mapping of the edge type has six kinds, which correspond to M-P, P-P,

[0113] P-A1, P-A2, A1-A1, A2-A2.

[0114] It should be understood that the Heterogeneous Graph Transformer (HGT) model includes three processes, which are: a heterogeneous mutual attention process, a heterogeneous message passing process, and an aggregation (or “aggregation”) update process. In some embodiments, obtaining the molecular representation data includes: the computing device 100 calculates the attention coefficient of the current node in the molecular feature graph and other nodes connected to the current node; aggregating the features of the neighboring nodes of the current node to update the features of the current node; and determining the molecular representation data about the candidate molecule based on the updated node features.

[0115] The following Table 3 schematically shows the main parameters of the model.

[0116] Table 3

[0117]

[0118] The specific embodiments of the method for obtaining molecular representation data via node message passing and aggregation will be described below, and will not be described here. Figure 8 Specific embodiments of the method for obtaining molecular representation data via node message passing and aggregation will be described below, and will not be described here.

[0119] It should be understood that in addition to the implementation of using the Transformer (HGT) model to pass and aggregate the initial features of the mapped nodes to obtain the molecular representation data. The present application can also replace the method of Heterogeneous Graph Transformer (HGT) model with HEAT, HAN, etc. algorithm. It should be understood that the algorithm HEAT uses a multi-channel trajectory prediction structure of a three-channel framework, considering the individual dynamics, interaction of nodes, and edge structure and features of each edge. The algorithm HAN uses different meta-paths to convert the heterogeneous graph into multiple homogeneous subgraphs, learns the node-level and semantic-level features in different word graphs based on the attention mechanism, and finally weights the vectors of each meta-path subgraph to obtain the global representation.

[0120] Figure 9 A schematic diagram of a method 900 for determining molecular characterization data for candidate molecules according to other embodiments of the present invention is shown. In other embodiments, such as Figure 9 As shown, the present invention can also use a homogeneous graph algorithm, such as a Graph Convolutional Network (GCN) model 910, to aggregate two-dimensional and three-dimensional graph atomic information for atomic layers in heterogeneous graphs of molecular feature maps. Then, the aggregated atomic information 912 is input into the heterogeneous graph network, while removing edges (A1-A1) between two-dimensional graph atomic nodes and edges (A2-A2) between three-dimensional graph atomic nodes. Afterwards, message passing and aggregation (as indicated by label 920) are performed on the initial features after removing atomic edges to obtain molecular characterization data. It should be understood that if local fragment information of the molecule is more useful, then by using GCN to pre-aggregate the atomic views, local fragment information that differs from functional group features can be obtained. This information, then passed to the functional groups via the heterogeneous graph, may more effectively characterize molecular properties.

[0121] The following will combine Figure 7 and Figure 8 A method for obtaining molecular characterization data via node message passing and aggregation according to embodiments of the present invention is described. Figure 8 A flowchart of a method 800 for obtaining molecular characterization data via node messaging and aggregation according to an embodiment of the present invention is shown. It should be understood that method 800 can, for example, be implemented in... Figure 10 The described electronic device is executed at point 1000. It can also be used in... Figure 1 The described computing device 100 performs the operation. It should be understood that method 800 may also include additional actions not shown and / or the actions shown may be omitted, and the scope of the invention is not limited in this respect.

[0122] At step 802, computing device 100 calculates the attention coefficients of the current node and other nodes connected to the current node in the molecular feature map. The following formulas (4) to (7) schematically illustrate the algorithm for calculating heterogeneous mutual attention coefficients.

[0123]

[0124] In formulas (4) to (7) above, Attention(s,e,t) represents the heterogeneous attention coefficient for (s,e,t). Here, s and t represent two nodes in the heterogeneous graph. e represents the edge connecting the two nodes s and t. N(t) represents all other nodes connected to node t. i represents the i-th head in the multi-head attention, and the value of i ranges from [1,h]. Hl-1 Hs represents the feature of node s at layer l-1. l-1 [s] represents the feature of node s at layer l-1. l-1 Ht represents the feature of node t at layer l-1. i (t) represents the feature of node t after mapping via the linear mapping parameters corresponding to the node type of the i-th head. i (s) represents the feature of node s after mapping via the linear mapping parameters corresponding to the node type of the i-th head. φ(s) represents the linear mapping corresponding to the node type of node s. φ(t) represents the linear mapping corresponding to the node type of node t. ψ(e) represents the mapping of the edge type of edge e. μ represents the coefficients corresponding to different edge types, and m is a constant. ATT-head i (s, e, t) represents the attention parameter of the i-th head. The above formula (6) indicates that the feature Hs of node s at layer l-1 is mapped to K (s) using the parameter l-1 (s, e, t). i (s) using the parameter Ht at layer l-1 is mapped to Q l-1 (t) using the parameter i (t). φ(e) represents the edge mapping parameter of edge e. This parameter is shared by different attention heads for the same type of edge.

[0125] At step 804, the computing device 100 aggregates the features of the neighboring nodes of the current node for updating the feature of the current node. For example, as shown in FIG. 7B, the features of the neighboring nodes around the functional group node 724, i.e., the features of the molecule node 722, the two-dimensional graph atom node 726, and the three-dimensional graph atom node 728, are aggregated in order to update the feature of the functional group node 724. It should be understood that the computing device 100 updates the feature of each node using the above-mentioned aggregation and updating method. Figure 7

[0126] With regard to the method of aggregating the features of the neighboring nodes of the current node for updating the feature of the current node, the following formulas (8) to (9) schematically show the algorithm of aggregating the features of the neighboring nodes of any node (i.e., the message passing process). It should be understood that the message passing includes a multi-head linear mapping process.

[0127]

[0128] In the above formulas (8) to (9), and ​These represent linear mappings. S and t represent nodes, e represents edges, and h represents the number of heads in the multi-head attention mechanism. H l-1 [s] represents the characteristics of node s at layer l-1.

[0129] The following formulas (10) to (11) schematically illustrate the algorithm for updating the characteristics of any of the nodes (i.e., the aggregation process). Formula (10) represents the algorithm for aggregating information from the different neighboring nodes of node t. Formula (11) represents the algorithm for updating the characteristics of node t using the aggregated information.

[0130]

[0131] In the above formulas (10) to (11), WA φ(t) Let represent the linear mapping parameters, σ represent the activation function, s and t represent nodes, and e represent edges. H l-1 [t] represents the characteristics of node t in layer l-1. This represents the characteristics of node t in layer l.

[0132] At step 806, computing device 100 determines molecular characterization data for candidate molecules based on the updated node features.

[0133] For example, after the computing device 100 obtains the updated features of each node, it selects the features of functional group nodes (i.e., P) to generate molecular characterization data for candidate molecules. This is because functional group nodes, as intermediate nodes connecting molecular nodes (i.e., M) and different atomic nodes (e.g., two-dimensional graph atomic node A1, three-dimensional graph atomic node A2), are more representative of the molecular characterization of candidate molecules, and pharmacophores play a significant role in candidate molecules. Therefore, this invention selects the features of all functional group nodes to characterize the final molecular characterization data of candidate molecules (e.g., obtaining...). Figure 7 Molecular characterization data of the candidate molecules indicated by marker 740.

[0134] Specifically, methods for generating molecular characterization data about candidate molecules based on the features of functional group nodes (i.e., P) include, for example, computing device 100 using max pooling operations to obtain molecular characterization data of candidate molecules.

[0135] The following formula (12) schematically illustrates the algorithm for generating molecular characterization data for candidate molecules.

[0136]

[0137] In the above formula (12), c i Representative molecular characterization data C={c1,c2,...,c d The i-th element in}. Np This represents the number of functional group nodes (i.e., P).

[0138] By adopting the above method, the present invention can not only learn the local features of nodes in the molecular feature map well, but also learn the global features of the molecule well. Therefore, it can obtain a good global characterization of candidate molecules.

[0139] Figure 10 A block diagram schematically illustrates an electronic device 1000 suitable for implementing embodiments of the present invention. The electronic device 1000 may be used to implement... Figure 2 , Figure 5 , Figure 6 , Figure 8 Methods 200, 500, 600, and 800 are shown. (For example...) Figure 10 As shown, the electronic device 1000 includes a central processing unit (i.e., CPU 1001), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (i.e., ROM 1002) or loaded from storage unit 1008 into random access memory (i.e., RAM 1003). The RAM 1003 may also store various programs and data required for the operation of the electronic device 1000. The CPU 1001, ROM 1002, and RAM 1003 are interconnected via bus 1004. An input / output interface (i.e., I / O interface 1008) is also connected to bus 1004.

[0140] Multiple components in electronic device 1000 are connected to I / O interface 1008, including: input unit 1006, output unit 1008, and storage unit 1008. CPU 1001 executes the various methods and processes described above, such as methods 200, 500, 600, and 800. For example, in some embodiments, methods 200, 500, 600, and 800 may be implemented as computer software programs stored in a machine-readable medium, such as storage unit 1008. In some embodiments, part or all of the computer program may be loaded and / or installed on electronic device 1000 via ROM 1002 and / or communication unit 1009. When the computer program is loaded into RAM 1003 and executed by CPU 1001, one or more operations of methods 200, 500, 600, and 800 described above may be performed. Alternatively, in other embodiments, CPU 1001 may be configured by any other suitable means (e.g., by means of firmware) to perform one or more actions of methods 200, 500, 600 and 800.

[0141] It is further noted that the present application can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for performing various aspects of the present application.

[0142] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or punched tape, a

[0143] The computer readable program instructions described herein can be downloaded to respective computing / processing devices from a computer readable storage medium or to an external computer or external storage device via a network, for example, the Internet, a local area network, a wide area network and / or a wireless network. The network can comprise copper transmission cables, optical transmission fibers, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. A network adapter card or network interface in each computing / processing device receives computer readable program instructions from the network and forwards the computer readable program instructions for storage in a computer readable storage medium within the respective computing / processing device.

[0144] Computer readable program instructions for carrying out operations of the present application can be assembler instructions, instruction-set-architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or either source code or object code written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate array (FPGA), or programmable logic array (PLA) can execute the computer readable program instructions by utilizing state information of the computer readable program instructions to configure the electronic circuitry in order to act in accord with the

[0145] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0146] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0147] The computer readable program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus or other device to produce a computer implemented process such that the instructions which execute on the computer, other programmable data processing apparatus, or other device implement the functions / acts specified in the flowchart and / or block diagram block or blocks.

[0148] The flow diagrams and the block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flow diagrams and the block diagrams can represent a module, segment, or portion of instructions, which comprises one or more executable instructions for implementing the specified logical functions (s). In some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession can in fact be executed substantially concurrently or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flow diagrams, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and

[0149] Embodiments of the application have been described above with the understanding that the explanations are illustrative and are not exhaustive, and are not limited to the disclosed embodiments. Many modifications and changes can be made by those skilled in the art without departing from the scope and spirit of the described embodiments. The selection of terms to be used in the description is intended to best explain the principles of the embodiments, practical application, or improvement over the technology in the market, or to enable other ordinary skilled persons in the art to understand the embodiments disclosed herein.

[0150] The above are only optional embodiments of the present application, and are not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for drug screening, characterized by, The method comprises: screening, via a first deep learning model, a plurality of chemical molecules in a chemical molecule library to obtain a plurality of candidate molecules satisfying a predetermined condition; determining binding characterization data of each candidate molecule in the plurality of candidate molecules to a specified target, so as to further select a candidate molecule based on the specified target from the plurality of candidate molecules based on the calculated binding characterization data; constructing a molecular feature graph for the candidate molecule, the molecular feature graph indicating at least features of molecules, functional groups, and atoms of the candidate molecule, wherein the molecular feature graph comprises a node set and an edge set, the node set comprising two-dimensional graph atom nodes and three-dimensional graph atom nodes, molecule nodes, and functional group nodes, the functional group nodes being intermediate nodes connecting the molecule nodes and the atom nodes; and determining molecular characterization data of the candidate molecule based on the constructed molecular feature graph, so as to determine a target molecule for a drug from the candidate molecule based on the molecular characterization data, wherein constructing the molecular feature graph comprises initializing each node in the molecular feature graph as a vector to generate initial features of the node, for the molecule node, the node vector being a molecular view generated via a pre-trained model using SMILES sequence information of the candidate molecule, for the functional group node, the node vector being represented using molecular fingerprints and one-hot encoding, for the two-dimensional graph atom node, the node vector being constructed based on a plurality of atomic information using one-hot encoding, and for the three-dimensional graph atom node, the node vector comprising information of the atom in the molecular sequence and also containing distribution information of the atom in the molecular space.

2. The method of claim 1, wherein, The molecular feature graph further comprises a set of node types and a set of edge types.

3. The method of claim 2, wherein, Determining the molecular characterization data of the candidate molecule based on the constructed molecular feature graph comprises: extracting, via a heterogeneous graph attention network model, relationship information of the nodes and edges of the molecular feature graph based on the constructed molecular feature graph, so as to aggregate the extracted relationship information of the nodes and edges to generate the molecular characterization data of the candidate molecule.

4. The method of claim 3, wherein, Aggregating the extracted relationship information of the nodes and edges to generate the molecular characterization data of the candidate molecule comprises: aggregating the relationship information of the nodes and edges related to the molecule nodes, the two-dimensional graph atom nodes, and the three-dimensional graph atom nodes to the functional group nodes for generating the molecular characterization data of the candidate molecule.

5. The method of claim 2, wherein, Constructing the molecular feature graph for the candidate molecule comprises: respectively taking the molecule of the candidate molecule, each functional group in the candidate molecule, each atom in the two-dimensional chemical structure of the candidate molecule, and each atom in the three-dimensional atomic space conformation of the candidate molecule as nodes in the molecular feature graph to construct the node set; associating each node in the node set with a mapping function of a corresponding node type in the set of node types; constructing connection relationships between the molecule nodes, the functional group nodes, the two-dimensional graph atom nodes, and the three-dimensional graph atom nodes by means of edges to form the edge set; and associating each edge in the edge set with a mapping function of a corresponding edge type in the set of edge types.

6. The method of claim 4, wherein, The dimensions of the initial features of different types of nodes are different, and based on the constructed molecular feature map, determining the molecular representation data of the candidate molecule comprises: mapping the initial features of different nodes into the same dimension; and performing message passing and aggregation on the initial features of the mapped nodes to obtain the molecular representation data.

7. The method of claim 4, wherein, Performing message passing and aggregation on the initial features of the mapped nodes to obtain the molecular representation data comprises: calculating attention coefficients of the current node in the molecular feature map and other nodes connected to the current node; aggregating the features of the neighboring nodes of the current node for updating the features of the current node; and determining the molecular representation data of the candidate molecule based on the updated features of the nodes.

8. The method of claim 7, wherein, Determining the molecular representation data of the candidate molecule based on the updated features of the nodes comprises: selecting the features of the aggregated and updated functional group nodes; and generating the molecular representation data of the candidate molecule via a max-pooling operation on the features of the aggregated and updated functional group nodes.

9. The method of claim 1, wherein, Selecting the candidate molecule from the plurality of candidates based on the calculated binding representation data comprises: determining the binding representation data of each candidate molecule to a specified target for ranking the candidate molecules based on the calculated binding representation data; and selecting a predetermined number of candidate molecules with high rankings as the candidate molecules based on the ranking results.

10. The method of claim 3, wherein, The determined target molecule is a target molecule for anti-aging.

11. A computing device comprising: at least one processing unit; at least one memory coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions, when executed by the at least one processing unit, cause the device to perform steps of the method according to any one of claims 1 to 10.

12. A computer-readable storage medium having stored thereon a computer program which, when executed by a machine, implements the method according to any one of claims 1 to 10.

13. A computer program product, characterised in that, A computer program comprising instructions which, when executed by a machine, perform the method according to any one of claims 1 to 10.

Images