Method, apparatus, device and computer program for reconstructing molecular structure

Abstract

A method, apparatus, device and readable storage medium for reconstructing a molecular structure, which relates to the field of machine learning, includes the steps of: acquiring structural data of a reference molecule (101), structurally decomposing the structural data of the reference molecule to acquire a data set of molecular fragments corresponding to the reference molecule (102), performing feature processing on the data set of molecular fragments to acquire candidate fragments for replacing shard fragments (103), and generating structural data of a reconstructed molecule based on the candidate fragments and side chain fragments (104).

Description

[Technical field]

[0001] [CROSS REFERENCE TO RELATED APPLICATIONS] This application claims priority to a Chinese patent application filed on March 10, 2021, bearing application number "202110260462.7" and titled "Method, apparatus, device and readable storage medium for reconstructing molecular structure," the entire contents of which are incorporated herein by reference.

[0002] The embodiments of the present application relate to the field of machine learning, and in particular to a method, an apparatus, a device, a storage medium, and a program product for reconstructing a molecular structure. [Background technology]

[0003] Molecular generation refers to generating a molecular structure that is different from the conventional molecular structure but has the same activity by modifying it based on the conventional molecular structure. For example, a reference molecule is a molecule that exhibits activity on a target target, and molecular generation generates molecules of other structures that exhibit similar activity on a target target based on the reference molecule.

[0004] In the prior art, molecular generation is performed using artificial intelligence (AI) methods, and AI molecular generation technology generally combines the ability of molecular reconstruction and the legitimacy of molecular reconstruction based on a reference molecule to generate molecules with new structures.

[0005] However, when the above method reconstructs molecules, it reconstructs them based on rules and is difficult to escape from the main concept of molecular structure design, so that the reconstructed molecules are unlikely to avoid the existing molecular structure, resulting in a low success rate of molecular reconstruction. Summary of the Invention [Problem to be solved by the invention]

[0006] According to the embodiments of the present application, there are provided a method, an apparatus, a device, a storage medium and a program product for reconstructing a molecular structure, which can improve the success rate and efficiency of molecular reconstruction. [Means for solving the problem]

[0007] In one aspect, there is provided a method for reconstructing a molecular structure implemented by a computing device, comprising: obtaining structural data for a reference molecule, the reference molecule being a molecule that exhibits activity at a target target; decomposing the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule, the molecular fragments including shard fragments of the reference molecule and side chain fragments corresponding to the shard fragments; performing feature analysis on the data set of molecular fragments to obtain candidate fragments for replacing the shard fragment; A method for reconstructing a molecular structure is provided, comprising: generating structural data for a reconstructed molecule based on the candidate fragments and the side chain fragments, wherein the reconstructed molecule exhibits activity at the target target.

[0008] In another aspect, an acquisition module for acquiring structural data of a reference molecule, the reference molecule being a molecule that exhibits activity at a target target; A decomposition module that decomposes the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule, the molecular fragments including shard fragments of the reference molecule and side chain fragments corresponding to the shard fragments; A generation module that performs feature analysis on the data set of molecular fragments to obtain candidate fragments for replacing the shard fragments; The generation module further generates structural data of a reconstituted molecule based on the candidate fragments and the side chain fragments, the reconstituted molecule exhibiting activity at the target target. An apparatus for reconstituting a molecular structure is provided.

[0009] In another aspect, there is provided a computer device including a processor and a memory, wherein the memory stores at least one program, and the at least one program is loaded and executed by the processor to realize a molecular structure reconstruction method described in any one of the embodiments of the present application.

[0010] In another aspect, there is provided a computer-readable storage medium having stored therein at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, code set or instruction set being loaded and executed by a processor to realize a method for reconstructing a molecular structure as described in any one of the embodiments of the present application.

[0011] In another aspect, a computer program product or computer program is provided, comprising computer instructions, the computer instructions being stored in a computer readable storage medium, a processor of a computing device reading the computer instructions from the computer readable storage medium, the processor executing the computer instructions to cause the computing device to perform the method for reconstructing a molecular structure according to any one of the above embodiments. Effect of the Invention

[0012] The beneficial effects of the technical means of the embodiments of the present application include at least: decomposing the structural data of the reference molecule to obtain shard fragments and side chain fragments, and then predicting a new molecular fragment structure for replacing the shard fragment using the side chain fragment as a structural condition to obtain a candidate fragment, i.e., cutting the molecule to learn the structural rules of the related fragments, and using the newly generated candidate fragment to replace the shard fragment in the original molecule to realize the molecular modification; since a large number of candidate fragments can be generated based on the structural conditions, the probability that the newly generated molecule is different from the previous molecular structure is improved, thereby improving the success rate of molecular reconstruction. [Brief description of the drawings]

[0013] [Figure 1] 1 is a flowchart of a method for reconstructing a molecular structure according to an exemplary embodiment of the present application. [Diagram 2] FIG. 2 is a schematic diagram of a scaffold decomposition process according to the embodiment shown in FIG. 1. [Diagram 3] FIG. 2 is a schematic diagram of the scaffold decomposition results according to the embodiment shown in FIG. 1. [Figure 4] FIG. 4 is a schematic diagram of the skeletal decomposition results according to the embodiment shown in FIG. 3. [Diagram 5] FIG. 4 is a schematic diagram of a shard decomposition result according to the embodiment shown in FIG. 3. [Figure 6] FIG. 2 is a schematic diagram of a model training process according to one exemplary embodiment of the present application. [Figure 7] FIG. 1 is a schematic diagram of a molecular reorganization process according to one exemplary embodiment of the present application. [Figure 8] 4 is a flowchart of a molecular structure reconstruction method according to another exemplary embodiment of the present application; [Figure 9] 4 is a flowchart of a molecular structure reconstruction method according to another exemplary embodiment of the present application; [Figure 10] FIG. 10 is a schematic diagram of a bonding scheme for molecular sites according to the embodiment shown in FIG. 9. [Figure 11]1 is an overall flowchart of a molecular reconstitution method according to one exemplary embodiment of the present application; [Figure 12] FIG. 1 is a block diagram of a molecular structure reconstruction device according to an exemplary embodiment of the present application. [Figure 13] FIG. 2 is a block diagram of a molecular structure reconstruction device according to another exemplary embodiment of the present application. [Figure 14] FIG. 1 is a schematic block diagram of a computing device according to an exemplary embodiment of the present application. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] First, the nouns related to the embodiments of the present application will be briefly described.

[0015] Molecular reconstruction refers to modifying a conventional molecular structure to generate a molecular structure that is different from the conventional molecular structure but has the same activity.For example, a reference molecule is a molecule that targets a target target, that is, the reference molecule shows activity on the target target, and molecular generation generates a molecule of another structure that shows similar activity on the target target based on the reference molecule.In the field of drugs as an example, showing activity on the target target refers to being able to show drug action on the target target.

[0016] In the prior art, AI molecular generation technology is a technology for generating new structural molecules. AI molecular generation technology needs to combine the ability of molecular reconstruction based on a reference molecule and the legality of molecular reconstruction, and mainly focuses on how to establish a latent space of molecules based on Gaussian distribution, so that the reconstructed molecule is very similar to the molecular structure of the target and cannot break away from the structural rules of the reference molecule.

[0017] An embodiment of the present application provides an AI molecular design method for performing fragment modification on a reference molecule, by decomposing the reference molecule to obtain shard fragments and corresponding side chain fragments that need to be modified and replaced, and then generating new fragments based on the shard fragments, and then joining the side chain fragments with the generated new fragments to obtain a reconstructed molecular structure.

[0018] In addition, in the embodiments of the present application, replacing shard fragments is described as an example, and in some embodiments, this can be achieved by replacing side chain fragments or by replacing side chain fragments and shard fragments, and the embodiments of the present application are not limited thereto.

[0019] In addition, information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.) and signals related to this application are all permitted by the user or fully permitted by each party, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions. For example, the acquisition of reference molecules related to this application is all obtained when fully permitted.

[0020] First, a molecular structure reconstruction method according to an embodiment of the present application will be described. Fig. 1 is a flowchart of a molecular structure reconstruction method according to one exemplary embodiment of the present application, and is described by taking as an example that the method is executed by a computer device (e.g., a terminal or a server). As shown in Fig. 1, the method includes the following steps 101 to 104.

[0021] In step 101, structural data for a reference molecule is obtained, the reference molecule being a molecule that exhibits activity at a target target.

[0022] In some embodiments, the reference molecule is a currently existing molecular structure. Preferably, the reference molecule is a protected molecular structure in a predefined molecular library, or the reference molecule is a molecule whose structure is to be simplified, or the reference molecule is a molecule whose pharmacokinetic properties are to be improved.

[0023] In this embodiment, the medical field is taken as an example, and the reference molecule is a molecule targeted to the target target, i.e., the reference molecule can exhibit activity on the target target, thereby exhibiting drug action on the target target. In the embodiment of the present application, a structural reconstruction needs to be performed based on the reference molecule, so as to generate a molecule that is structurally different from the reference molecule but has the same or similar activity, i.e., the reconstructed molecule is also a molecule that exhibits activity on the target target.

[0024] In some embodiments, the reference molecule is a molecular structure uploaded by a user through a molecular reconstruction interface, and the computer device obtains the reference molecule by drawing the molecular structure of the reference molecule in the molecular reconstruction interface, or the computer device obtains the reference molecule by selecting a stored molecular structure in the molecular reconstruction interface, or the computer device obtains the reference molecule by editing and uploading the chemical formula of the reference molecule in the molecular reconstruction interface. The computer device may be the terminal itself, or may be a server that receives the molecular structure uploaded by the terminal, and the embodiment is not limited thereto.

[0025] In the embodiment of the present application, the structural data of the reference molecule is realized as a chemical formula of the reference molecule, or the structural data of the reference molecule is realized as a patterned structural representation of the reference molecule, and the embodiment is not limited thereto.

[0026] In step 102, the structural data of the reference molecule is decomposed to obtain a data set of molecular fragments corresponding to the reference molecule.

[0027] The set of molecular fragments includes shard fragments of the reference molecule and side chain fragments that correspond to the shard fragments.

[0028] In some embodiments, the structural decomposition of the structural data of the reference molecule needs to be based on a predetermined decomposition rule.

[0029] That is, first, a predetermined decomposition rule is obtained, the predetermined decomposition rule includes at least one of a backbone decomposition rule and a rotatable bond decomposition rule. The structural data of the reference molecule is decomposed according to the predetermined decomposition rule to obtain a data set of at least one molecular fragment group corresponding to the reference molecule. The molecular fragment group includes one shard fragment and at least one side chain fragment, and the shard fragment is obtained by decomposition according to the predetermined decomposition rule, and after the shard fragment is deleted from the structure of the reference molecule, the remaining structure is a side chain fragment corresponding to the deleted shard fragment.

[0030] The predetermined decomposition rule includes at least one of two decomposition rules, that is, when a molecular fragment group is obtained by decomposing according to the skeleton decomposition rule, and b molecular fragment group is obtained by decomposing according to the rotatable bond decomposition rule, molecular reconstruction is performed by a molecular fragment group, or molecular reconstruction is performed by b molecular fragment group, or molecular reconstruction is performed by a+b molecular fragment group. Here, a and b are both positive integers. In some embodiments, when molecular reconstruction is performed by a+b molecular fragment group, the generalizability of the molecular reconstruction model is increased. The above two predetermined decomposition rules are described below respectively.

[0031] 1. Rules for decomposing the skeleton That is, the predetermined decomposition rules include skeletal decomposition rules, and when structurally decomposing the reference molecule, a skeletal structure that satisfies skeletal requirements is extracted from the structural data of the reference molecule, the skeletal structure is deleted from the structural data of the reference molecule to obtain a side chain structure corresponding to the skeletal structure, and a data set of at least one group of molecular fragments corresponding to the reference molecule is obtained by treating the skeletal structure as a shard fragment and the side chain structure as a side chain fragment.

[0032] When the reference molecule is decomposed according to the skeleton decomposition rules, the corresponding primary skeleton is first extracted from the structural data of the reference molecule, and the primary skeleton is obtained, and then the smaller secondary skeleton structure is further decomposed according to different cyclic structures. The shard fragment obtained by decomposition according to the skeleton decomposition rules may be the primary skeleton structure in the reference molecule, or may be the secondary skeleton structure obtained by decomposing the primary skeleton structure.

[0033] In summary, FIG. 2 shows a schematic diagram of a scaffold decomposition process according to one exemplary embodiment of the present application, in which, as shown in FIG. 2, a patterned structure of a reference molecule 210 is first subjected to scaffold decomposition to obtain a primary scaffold 220, and then one or more decompositions are performed based on the primary scaffold 220 to obtain smaller secondary scaffold structures 230.

[0034] After obtaining the primary skeleton 220 and each secondary skeleton 230, a skeleton structure that satisfies the skeleton requirement is selected from the primary skeleton 220 and the secondary skeleton 230 to be used for decomposing the reference molecule. In some embodiments, when there are multiple skeleton structures that satisfy the skeleton requirement, one of the skeleton structures that satisfies the skeleton requirement is deleted from the reference molecule, and then the reference molecule is crushed into one or more side chains, and the deleted skeleton structure and the side chain obtained by crushing are combined to obtain one molecular fragment group, that is, one molecular fragment group includes one skeleton structure and at least one side chain.

[0035] After performing the above process for each skeleton structure that satisfies the skeleton requirements, at least one molecular fragment group is obtained. In some embodiments, if the number of side chains obtained after the skeleton structure is deleted is at least two, the skeleton structure and the at least two side chains are taken as one molecular fragment group. Schematically, as shown in FIG. 3, after deleting the secondary skeleton 320 from the molecule 310, two side chains 330 including side chain 331 and side chain 332 are obtained. The secondary skeleton 320 and the two side chains 330 are taken as one molecular fragment group corresponding to the molecule 310.

[0036] In some embodiments, the dataset of molecular fragments (also referred to as "swarm data") includes a backbone structure and side chain fragments corresponding to the backbone structure. In other embodiments, the dataset of molecular fragments may include structural data and a backbone structure of a reference molecule, and the molecular reconstruction model can directly determine the side chain fragments based on the structural data and the backbone structure of the reference molecule.

[0037] The above skeletal requirements include at least one of a ring number requirement, a heavy atom number requirement, and a rotatable bond requirement, and in response to the skeletal requirement including the ring number requirement, a skeletal structure whose ring number is within the ring number requirement range is extracted from the reference molecule, in response to the skeletal requirement including the heavy atom number requirement, a skeletal structure whose heavy atom number is within the heavy atom number requirement range is extracted from the reference molecule, and in response to the skeletal requirement including the rotatable bond requirement, a skeletal structure whose rotatable bond number is within the requirement range is extracted from the reference molecule.

[0038] In general, the framework requirements specifically include the following requirements:

[0039] 1. The number of rings in the skeletal structure is 2 or more, i.e., the number of rings in the skeletal structure obtained by decomposing the reference molecule must be 2 or more; 2. The number of heavy atoms in the backbone structure is less than 20, i.e., the number of heavy atoms in the backbone structure obtained by decomposing the reference molecule must be less than 20; 3. The number of rotatable bonds in the backbone structure is less than 3, i.e., the number of rotatable bonds in the backbone structure obtained by decomposing the reference molecule must be less than 3.

[0040] 4 shows a schematic diagram of a backbone decomposition result according to one exemplary embodiment of the present application, and as shown in FIG. 4, after decomposing a reference molecule 400, the decomposition result includes at least two decomposition results, in a first decomposition result 410, the reference molecule 400 is decomposed to obtain a backbone structure 411 and a side chain 412, and in a second decomposition result 420, the reference molecule 400 is decomposed to obtain a backbone structure 421 and a side chain 422. The backbone structure 411 and the side chain 412 constitute one molecular fragment group, and the backbone structure 421 and the side chain 422 constitute one molecular fragment group.

[0041] II. Decomposition rules for rotatable bonds That is, the predetermined decomposition rule includes a decomposition rule for a rotatable bond, and when decomposing the structural data of the reference molecule, the structural data of the reference molecule is cut from the rotatable bond to obtain a shard fragment that satisfies the shard requirements, and at least one group of molecular fragments corresponding to the reference molecule is obtained based on the shard fragment.

[0042] When a reference molecule is cut from a rotatable bond, the cut shard must satisfy shard requirements. In some embodiments, the shard requirements include at least one of shard base requirements and shard skeleton requirements. When a reference molecule is cut from a rotatable bond, the cut portion is called a fragment, and each shard includes a scaffold. The shard base requirements are requirements for the entire shard, and the shard skeleton requirements are requirements for the skeleton included in the shard.

[0043] In some embodiments, the shard base requirements include at least one of a shard ring number requirement, a shard atom number requirement, a shard rotatable bond requirement, and a shard structure requirement, and the shard skeleton requirements include at least one of a skeleton atom number requirement and a skeleton rotatable bond requirement for the shard skeleton.

[0044] In general terms, the shard requirements specifically include the following requirements:

[0045] 1. The shard fragment contains only one ring; 2. The number of non-H atoms in the shard fragment is greater than 5 and less than 30, i.e., the number of non-hydrogen atoms in the shard fragment is between 5 and 30; 3. The size of the ring within the shard fragment is less than 8, i.e., the ring in the shard fragment is less than 8; 4. The number of rotatable bonds in the shard fragment is less than 9, i.e., the number of rotatable bonds in the shard fragment is less than 9; 5. A shard skeleton exists in the shard fragment, i.e., the shard fragment is a fragment that includes a skeleton; 6. The number of non-H atoms in the shard skeleton is less than 20, i.e., the number of non-hydrogen atoms in the skeleton of the shard fragment is less than 20; 7. The number of rotatable bonds in the shard skeleton is less than 3, i.e., the number of rotatable bonds in the skeleton of the shard fragment is less than 3; 8. The difference between the number of non-H atoms of the shard fragment and the number of non-H atoms of the shard skeleton is less than 10; 9. A shard fragment may not contain only one single ring structure.

[0046] 5 shows a schematic diagram of a shard decomposition result according to one exemplary embodiment of the present application, and as shown in FIG. 5, after decomposing the patterned structure of the reference molecule 500, the decomposition result includes at least four decomposition results, in which the first decomposition result decomposes the reference molecule 500 to obtain a shard fragment 511, in the second decomposition result decomposes the reference molecule 500 to obtain a shard fragment 521, in the third decomposition result decomposes the reference molecule 500 to obtain a shard fragment 531, and in the fourth decomposition result decomposes the reference molecule 500 to obtain a shard fragment 541. After removing the shard fragments in each decomposition method from the reference molecule, side chain fragments are obtained. Each decomposition method corresponds to one group of molecular fragments.

[0047] In step 103, feature analysis is performed on the molecular fragment dataset to obtain candidate fragments for replacing the shard fragment.

[0048] In some embodiments, a data set of molecular fragments is input to a molecular reconstruction model, and the molecular reconstruction model performs feature analysis on the data set of molecular fragments, and obtains candidate fragments for replacing shard fragments.Preferably, if there is at least one molecular fragment group for one reference molecule, the data set of at least one molecular fragment group is input sequentially to the molecular reconstruction model, and each molecular fragment group in the at least one molecular fragment group is analyzed sequentially by the molecular reconstruction model, and a plurality of candidate fragments corresponding to each molecular fragment group are obtained.

[0049] In some embodiments, the data set of molecular fragments is encoded based on the side chain fragments to generate fragment features, and the fragment features are decoded to generate candidate fragments for replacing the shard fragments.

[0050] The molecular reconstruction model includes an encoder and a decoder in a codec architecture, and first, based on the side chain fragment, the encoder encodes the shard fragment to obtain fragment features, then perturbs the fragment features according to a predetermined perturbation rule to obtain perturbation features, and decodes the perturbation features by the decoder to generate candidate fragments for replacing the shard fragments. The perturbation process is a process of increasing the diversity of the generated candidate fragments. That is, by perturbing, the candidate fragments are changed based on the initial shard fragments, and thus the candidate fragments are determined in the vicinity of the feature space corresponding to the initial shard fragments, thereby improving the diversity of the candidate fragments.

[0051] In the process of encoding to obtain fragment features, the side chain fragments are mapped into a feature space to obtain side chain spatial features, and the shard fragments are encoded by an encoder using the side chain spatial features as structural conditions to obtain fragment features. In some embodiments, after the shard fragments are encoded by an encoder, the encoding features are first obtained, and the side chain spatial features are used as structural conditions to map the encoding features into a feature space to obtain fragment features.

[0052] The encoder, decoder and feature space are obtained by pre-training, i.e., the mapping process of the shard fragments and side chain fragments in the feature space is realized based on a pre-trained molecular reconstruction model. The feature space is realized as a Gaussian space that satisfies the Gaussian distribution condition. The training and application of the molecular reconstruction model will be specifically described in the following examples.

[0053] In step 104, structural data for the reconstituted molecule is generated based on the candidate fragments and the side chain fragments.

[0054] In some embodiments, hydrogen sites in the candidate fragments and side chain fragments are joined to obtain structural data for the reconstituted molecule, which exhibits activity at the target target, i.e., the reconstituted molecule and the reference molecule exhibit activity at the same target target.

[0055] After joining the candidate fragments and the side chain fragments, the resulting reconstituted molecules must be subjected to further molecular structure screening, and the screening process is preset.

[0056] In addition, the candidate fragments, side chain fragments, shard fragments, and candidate molecules in this embodiment refer to structural data of candidate fragments, structural data of side chain fragments, structural data of shard fragments, and structural data of candidate molecules.

[0057] In conclusion, the molecular structure reconstruction method according to the embodiment of the present application decomposes the structural data of a reference molecule to obtain shard fragments and side chain fragments, and then predicts a new molecular fragment structure for replacing the shard fragment using the side chain fragment as a structural condition to obtain a candidate fragment, i.e., cuts the molecule to learn the structural rules of the related fragments, and uses the newly generated candidate fragment to replace the shard fragment in the original molecule to realize molecular modification. Since a large number of candidate fragments can be generated based on the structural conditions, the probability that the newly generated molecule is different from the previous molecular structure is improved, thereby improving the success rate of molecular reconstruction.

[0058] For the molecular reconstruction model in step 103 above, the training process and application process of the molecular reconstruction model are respectively described. Whether it is the sample molecule in the model training process or the reference molecule in the actual reconstruction, the molecular structure needs to be decomposed into shard fragments and side chain fragments by decomposing the molecular structure. The model processing process in the training process and application process are respectively described below.

[0059] (1) Training Process The training process includes a pre-training process and a fine-tuning process.

[0060] The pre-training process is mainly to enable the model to learn vector representations of molecules by learning molecular structures in public databases, that is, the pre-training process is the process in which the model learns to convert molecules into vector features.

[0061] In some embodiments, the candidate fragments are obtained as predicted by a molecular rearrangement model.

[0062] In summary, FIG. 6 shows a schematic diagram of a model training process according to one exemplary embodiment of the present application, taking a group of molecular fragments obtained by decomposing a sample molecule as an example. As shown in FIG. 6, the structural data of the sample molecule is decomposed to obtain shard fragments 601 and side chain fragments 602. The molecular reconstruction model 610 includes an encoder 611 and a decoder 612. The encoder 611 encodes the shard fragments 601 and the side chain fragments 602 to obtain a first fragment vector 621 corresponding to the shard fragment 601 and a second fragment vector 622 corresponding to the side chain fragment 602. The first fragment vector 621 and the second fragment vector 622 are decoded by the decoder 612, and then the codec is trained with the goal of restoring the shard fragments 601 and the side chain fragments 602.

[0063] The molecular reconstruction model according to the embodiment of the present application is improved based on Conditional Variational Autoencoders (CVAE). In the prior art, the training condition parameters of CVAE are quantized vectors, and in the embodiment of the present application, the molecule is decomposed into shards and side chains, and then the side chains are input into the model as a structural condition. Specifically, the side chain structure is mapped into an embedding space, and then the mapped structural embedding is input as a condition. The specific optimization conditions are shown in the following formula 1.

number

[0064] where f(·) denotes the Simplified Molecular Input Line Entry System (SMILES) encoder in the embodiments of the present application, and in some embodiments, the SMILES encoder and the CVAE encoder share parameters at the initial stage. f(·) maps the side chain y into the embedding space and performs the SMILES decoding calculation with the shard x. Relative entropy (Kullback Leibler, KL) denotes the divergence calculation. θ and Θ are model parameters, i.e., the parameters to be optimized in the training process. p θ (x|f(y)) denotes the conditional probability distribution conditional on f(y),

number

[0065] After inputting the structural data of the sample molecule into the molecular reconstruction model 610, the molecular reconstruction model 610 optimizes the above-mentioned first fragment vector 621 and second fragment vector 622, which are vector representations extracted from the molecular structure, by training model parameters based on the divergence between the output molecular fragments and the shard fragments of the sample molecule, i.e., optimizes the encoder, and also optimizes the process of converting from the vector representation to molecular fragments, i.e., optimizes the decoder.

[0066] It should be noted that the sample molecules applied to the pre-training process are molecules without target restrictions, or the sample molecules applied to the pre-training process are molecules to which no target information is attached.

[0067] In the fine-tuning process, intentional training is performed based on the target, i.e., a molecule showing activity on the target target is input and decomposed, and the model is further trained, so that the molecule reconstructed by the molecular reconstruction model 610 tends to be a molecule similar to the molecule showing activity on the target target. The adjustment method for the model parameters in the fine-tuning process is the same as the adjustment method for the model parameters in the pre-training process, but differs in that the dataset is switched from the public database to the molecule dataset showing activity on the target target.

[0068] (2) Application process In summary, Figure 7 shows a schematic diagram of the molecular reconstruction process of the present application. As shown in Figure 7, first, the structural data of the reference molecule 700 is input into the trained molecular reconstruction model 710, where the reference molecule 700 includes a shard fragment 721 and a side chain fragment 722. Before inputting into the molecular reconstruction model 710, the structural data of the reference molecule 700 is decomposed, or the reference molecule 700 is decomposed by the molecular reconstruction model 710, the shard fragment 721 in the reference molecule 700 is encoded by the encoder 711 to obtain a fragment vector 730 corresponding to the shard fragment 721, and the fragment vector 730 is decoded by the decoder 712 to obtain a candidate fragment 740 for replacing the shard fragment 721. In the process of the encoder 711 encoding the reference molecule 700, the side chain fragment 722 is used as a structural condition to constrain the feature mapping of the shard fragment 721 to obtain a fragment vector 730; in order to make a change in the fragment structure, the fragment vector 730 needs to be perturbed according to a predetermined perturbation rule, and a perturbed vector in the vicinity of the Gaussian space for the fragment vector 730 is determined; in summary, the feature value of the fragment vector 730 is adjusted within a predetermined perturbation range. After the side chain fragment 722 and the candidate fragment 740 are joined, the structural data of the reconstructed molecule 750 is obtained.

[0069] In some embodiments, the application process further includes a screening process of the candidate fragments 740 and a screening process of the reconstituted molecules 750, which are described in the subsequent embodiments.

[0070] The molecular reconstruction model can be further realized as a natural language processing-based model or a graph structure-based model, and the embodiments of the present application do not limit the specific realization method of the model.

[0071] It should be noted that in some embodiments, the perturbations applied to the fragment vectors are performed in the application process, and no perturbations are applied to the fragment vectors in the training process for the molecular reconstruction model.

[0072] In some embodiments, after the molecular reconstruction model predicts and obtains the candidate fragments, it is necessary to further screen the candidate fragments. Figure 8 is a flowchart of a molecular structure reconstruction method according to another exemplary embodiment of the present application, and is illustrated by taking as an example that the method is executed by a computer device. As shown in Figure 8, the method includes steps 801 to 805.

[0073] In step 801, structural data for a reference molecule is obtained, the reference molecule being a molecule that exhibits activity at a target target.

[0074] In this embodiment, the medical field is taken as an example, and the reference molecule is a molecule that targets a target target, i.e., the reference molecule can exhibit activity on the target target, thereby exhibiting drug action on the target target.

[0075] In step 802, the structural data of the reference molecule is decomposed to obtain a data set of molecular fragments corresponding to the reference molecule.

[0076] In some embodiments, the structural decomposition of the reference molecule must be based on predefined decomposition rules.

[0077] That is, firstly, obtain a predetermined decomposition rule, the predetermined decomposition rule includes at least one of a skeleton decomposition rule and a rotatable bond decomposition rule. The decomposition method corresponding to the predetermined decomposition rule has already been described in detail in step 102 above, and will not be described here.

[0078] In step 803, feature analysis is performed on the molecular fragment dataset to obtain candidate fragments for replacing the shard fragment.

[0079] After inputting a data set of a molecular fragment group into a molecular reconstruction model, an encoder in the molecular reconstruction model encodes shard fragments in the molecular fragment group using the features after encoding the side chain fragments as structural conditions to obtain an encoded vector, and after perturbing the encoded vector, a decoder in the molecular reconstruction model decodes the encoded vector to obtain candidate fragments for replacing the shard fragments.

[0080] In step 804, the candidate fragments are screened based on predefined screening rules.

[0081] In some embodiments, the predefined screening rules include at least one of a legality screening rule, a uniqueness screening rule, a heavy atom number screening rule, and a ring screening rule.

[0082] The legality screening rule refers to retaining candidate fragments that meet the SMILES legality and removing candidate fragments that do not meet the SMILES legality. The SMILES specification is a specification that clearly describes a molecular structure with an American Standard Code for Information Interchange (ASCII) string, i.e., the generated candidate fragments must meet the SMILES specification. In some embodiments, the SMILES specification specifies the grammar rules of a molecular structure, and generally, if five bonds are attached to a tetravalent nitrogen (N) site, the SMILES specification is not met.

[0083] The uniqueness screening rule refers to removing overlapping fragments among multiple candidate fragments, and ensuring that each candidate fragment does not overlap with other candidate fragments.In some embodiments, one or more candidate fragments can be obtained based on one molecular fragment group, so when multiple candidate fragments are obtained based on one molecular fragment group, the uniqueness screening rule is performed on the candidate fragments obtained by prediction based on one molecular fragment group, or is performed on the candidate fragments obtained by prediction based on all molecular fragment groups.

[0084] The heavy atom number screening rule refers to screening the generated candidate fragments by selecting a heavy atom number range based on the molecule that needs to be reconstructed.Generally, the heavy atom number range is determined based on the heavy atom number of the reference molecule itself.In outline, the reference molecule contains 12 heavy atoms, and in some embodiments, the heavy atom number range of the molecule that needs to be reconstructed is determined to be 9 to 15, and the candidate fragments are screened based on the requirements of the heavy atom number range, and the candidate fragments whose heavy atom number does not meet the heavy atom number range are filtered and deleted.

[0085] In some embodiments, it is necessary to further screen the candidate fragments according to ring screening rules. In general, after the above screening is completed, selectively screen the candidate fragments in the generated fragment set, and the screening includes retaining only the fragments that contain a ring or an aromatic ring. Preferably, the ring screening rules further include a requirement for the number of rings, or a requirement for the number of aromatic rings, etc.

[0086] In step 805, structural data of the reconstructed molecule is generated based on the remaining fragments and side chain fragments obtained after screening.

[0087] The reconstituted molecule is a newly generated molecule that exhibits activity on the target target. In some embodiments, since molecules with similar structures are more likely to exhibit activity on the same target, it is necessary to obtain the fragment similarity between the remaining fragments and the shard fragments, and the remaining fragments with the highest similarity, q (q is a positive integer), are retained to generate the reconstituted molecule together with the side chain fragments.

[0088] Fragment similarity refers to the similarity between the remaining fragment and the shard fragment in three-dimensional (3D) structure. That is, after obtaining the generated candidate fragments and screening to obtain a large amount of remaining fragments, it is necessary to ensure that the remaining fragments and the part that needs to be modified in the reference molecule (i.e., the shard fragment part) are as similar as possible, so that the molecule retains activity. In some embodiments, the cheminformatics software package RDKit tool or other 3D similarity calculation tool is used to calculate the 3D similarity between the candidate fragments (or the remaining fragments obtained after screening) and the reference fragments (i.e., the shard fragments in the reference molecule), and the q fragments with the highest similarity are extracted, and further joined to finally obtain the reconstructed molecule.

[0089] In conclusion, the molecular structure reconstruction method according to the embodiment of the present application involves decomposing the structure of a reference molecule to obtain shard fragments and side chain fragments, and then using the side chain fragments as structural conditions to predict new molecular fragment structures for replacing the shard fragments to obtain candidate fragments, i.e., cutting the molecule to learn the structural rules of the related fragments, and using the newly generated candidate fragments to replace the shard fragments in the original molecule to achieve molecular modification. Since a large number of candidate fragments can be generated based on the structural conditions, the probability that the newly generated molecule is different from the previous molecular structure is improved, thereby improving the success rate of molecular reconstruction.

[0090] The method of this embodiment screens candidate fragments to filter out and eliminate illegal or overlapping fragments, and also filters out and eliminates fragments that do not meet the heavy atom number requirements based on the heavy atom range requirements, thereby repeatedly screening to retain fragments that meet the basic requirements in the candidate fragments and eliminating fragments that do not meet the requirements, thus avoiding the problem of low molecular reconstruction efficiency caused by the fact that the number of subsequently generated reconstructed molecules is too large, and therefore molecules with illegal fragments cannot be filtered out.

[0091] The method of this embodiment performs 3D similarity calculations on the candidate fragments and shard fragments to ensure that the retained candidate fragments and the shard fragments that need to be modified in the reference molecule are as similar as possible in 3D structure, thus making them more likely to retain activity.

[0092] In some embodiments, when generating a reconstructed molecule based on the candidate fragments and side chain fragments, further molecular screening is required. Figure 9 is a flowchart of a molecular structure reconstruction method according to another exemplary embodiment of the present application, and is illustrated by taking as an example that the method is executed by a computer device. As shown in Figure 9, the method includes steps 901 to 905.

[0093] In step 901, structural data for a reference molecule is obtained, the reference molecule being a molecule that exhibits activity at a target target.

[0094] In this embodiment, the medical field is taken as an example, and the reference molecule is a molecule that targets a target target, i.e., the reference molecule can exhibit activity on the target target, thereby exhibiting drug action on the target target.

[0095] In step 902, structural data of the reference molecule is decomposed to obtain a data set of at least one molecular fragment corresponding to the reference molecule.

[0096] In some embodiments, the structural decomposition of the reference molecule must be based on predefined decomposition rules.

[0097] That is, firstly, obtain a predetermined decomposition rule, the predetermined decomposition rule includes at least one of a skeleton decomposition rule and a rotatable bond decomposition rule. The decomposition method corresponding to the predetermined decomposition rule has already been described in detail in the above step 302, and will not be described here.

[0098] In step 903, feature analysis is performed on the molecular fragment dataset to obtain candidate fragments for replacing the shard fragment.

[0099] After the group of molecular fragments is input into the molecular reconstruction model, an encoder in the molecular reconstruction model encodes the shard fragments in the group of molecular fragments using the side chain fragments as structural conditions to obtain an encoded vector, and a decoder in the molecular reconstruction model decodes the encoded vector using the side chain fragments as structural conditions to obtain candidate fragments for replacing the shard fragments.

[0100] In addition, although the present embodiment uses the candidate fragment as an example, in some embodiments, the candidate fragment can also be realized as a remaining fragment obtained after screening in combination with the technical means of the embodiment shown in FIG. 8, that is, the candidate fragment is first screened, and then the remaining fragment is joined to the subsequent side chain fragment, and the embodiments of the present application are not limited thereto.

[0101] In step 904, the candidate fragments and the side chain fragments are joined to obtain n candidate molecules (n is a positive integer).

[0102] In some embodiments, after deleting the parts of the reference molecule that need to be modified (e.g., the original backbone), the original parts connected to the backbone are joined to any H sites in the candidate fragments, and the SMILES validity is ensured to generate new molecules, the n candidate molecules.

[0103] In summary, FIG. 10 shows a schematic diagram of a molecular site bonding method according to one exemplary embodiment of the present application. As shown in FIG. 10, a reference molecule 1000 is decomposed to obtain a shard fragment 1010 and a side chain fragment 1020, a candidate fragment 1030 is generated by a molecular reconstruction model, and then the candidate fragment 1030 and the side chain fragment 1020 are bonded to obtain a bonding result 1040, which includes the bonding result between the side chain fragment 1020 and different bondable hydrogen bonds in the candidate fragment 1030.

[0104] In step 905, n candidate molecules are screened based on the molecular structure data in a predetermined molecular library to obtain structural data of the reconstituted molecule.

[0105] A reconstituted molecule is a newly generated molecule that exhibits activity at a desired target.

[0106] In some embodiments, n candidate molecules are screened based on molecular structure data in a predetermined molecular library to obtain m (0≦m≦n) candidate molecules that are not included in the predetermined molecular library, thereby determining structural data of the reconstituted molecule from the m candidate molecules.

[0107] In some embodiments, the conjugated candidate molecules are screened to avoid existing molecules in the predetermined molecular library, and molecules included in the predetermined molecular library are deleted, preferably, the predetermined molecular library includes a patent library or a set of patent scaffolds, in some embodiments, the predetermined molecular library may further be a molecular library set against a target target, and the embodiments of the present application do not limit the screening method.

[0108] In some embodiments, the screening process for candidate molecules further includes a screening process for druggable properties realized by a computer device. That is, screening for druggable properties is performed on the generated candidate molecules. Generally speaking, the specific screening process includes the following processes 1 to 2. 1. Screening is performed by a Mini Capsule Filter (MCF), mainly to remove candidate molecules containing active groups or toxic groups. 2. Screening for the physicochemical properties of candidate molecules is performed. For example, the candidate molecules need to meet conditions such as Molecular Weight (MW) ≤ 550, water solubility within the range [-5, 6], molecular topological polar surface area ≤ 120, number of rotatable bonds < 10, number of hydrogen bond acceptors within the range [0, 10], and number of hydrogen bond donors within the range [0, 5].

[0109] In some embodiments, when determining a reconstructed molecule from m candidate molecules, m molecular similarities corresponding to the m candidate molecules and a reference molecule are obtained respectively. The i-th (i is a positive integer) candidate molecule corresponds to the i-th molecular similarity, and the reconstructed molecule is determined from the k candidate molecules (0 < k < m) with the highest molecular similarities.

[0110] That is, after obtaining the filtered candidate molecules, in order to increase the probability of screening for active molecules, it is required that the candidate molecules and the reference molecule are as similar as possible, so that the activity is easy to maintain. Therefore, the 3D similarity between the candidate molecule and the reference molecule is calculated using the RDKit tool or other 3D similarity calculation tools, and the k candidate molecules with the highest similarity are extracted to perform the subsequent reconstructed molecule generation process.

[0111] In some embodiments, the screening for the candidate molecules further uses the protein pocket information to extract the main activity of the generated candidate molecules, and screens out and removes molecules that do not contain the activity of the reference molecule; then, uses a molecular activity prediction model to input the protein pocket information and the generated candidate molecules, output the PIC50 activity value, and sort the candidate molecules to achieve the screening purpose.

[0112] In view of the above, the molecular structure reconstruction method according to the embodiment of the present application decomposes the reference molecule to obtain shard fragments and side chain fragments, and then predicts a new molecular fragment structure for replacing the shard fragment using the side chain fragment as a structural condition to obtain a candidate fragment, that is, cuts the molecule according to the ring structure to learn the structural rules of the related fragments, and uses the newly generated candidate fragment to replace the shard fragment in the original molecule to realize molecular modification, so that the generated molecule has a high probability of breaking away from the structural rules of the previous molecular structure, thereby improving the success rate of molecular reconstruction.

[0113] The method of this embodiment generates candidate molecules, then screens the candidate molecules using multiple screening mechanisms, and determines a reconstituted molecule from the candidate molecules obtained after screening, thus avoiding the problem of the cumbersome artificial screening process caused by determining a reconstituted molecule from a large number of candidate molecules.

[0114] 11 is an overall flowchart of a molecular reconstitution method according to one exemplary embodiment of the present application, and as shown in FIG. 11, the process includes the following steps 1101 to 1112.

[0115] In step 1101, a pre-training process occurs.

[0116] The pre-training process refers to training the structure-conditioned variational autoencoder model on molecules in a public database, which records drug-like molecules in the SMILES format, so that the model can better learn the vector representation of molecules.

[0117] In the pre-training process, the sample molecule is decomposed to obtain backbone and side chain pairs by at least one of two decomposition methods, and the decomposed backbone and side chain pairs are input as training data into the structure-conditioned variational autoencoder model to complete the task of model pre-training.

[0118] In step 1102, a fine-tuning process occurs.

[0119] In the fine-tuning process, sample molecules that are active on the target target are input and decomposed, and the model is further trained (the training method is the same as the pre-training process, but the input dataset is switched from the public database to the dataset of molecules that are active on the target target), so that the generated molecules tend to be similar to the molecules that are active on the target target.

[0120] After the fine-tuning process is completed, the trained model can obtain candidate fragments by performing fragment prediction on the reference molecule.

[0121] In step 1103, a fragment size screening process occurs.

[0122] Preferably, the generated candidate fragment set is subjected to SMILES legalization and uniqueness processing to remove illegal and duplicate fragments, and then the generated candidate fragments are screened based on the selected heavy atom range, and candidate fragments outside the heavy atom range are removed.

[0123] In step 1104, a screening process for molecular ring number and structural characteristics is carried out.

[0124] Candidate molecules are screened based on the number of rings or aromatic rings in the molecule, or based on the properties of the molecular structure.

[0125] It should be noted that the above steps 1103 and 1104 are two parallel steps, and step 1103 may be executed first, or step 1104 may be executed first, or further, step 1103 and step 1104 may be executed simultaneously.

[0126] In step 1105, a fragment 3D similarity calculation process is performed.

[0127] After obtaining a large amount of candidate fragments, we require that the candidate fragments and the parts in the reference molecule that need to be modified are as similar as possible, and thus are more likely to retain activity. Therefore, we use the RDKit to calculate the 3D similarity between the generated fragments and the reference fragment, and extract the candidate fragments with the highest q similarities for further joining.

[0128] In step 1106, a molecular fragment joining process is performed.

[0129] After deleting the parts of the reference molecule that need to be modified (the original backbone), a new molecule is generated by joining the original parts connected to the original backbone with any H sites in the candidate fragments and ensuring their SMILES validity.

[0130] In step 1107, a filtering process is performed using a predetermined molecular library.

[0131] The conjugated candidate molecules are screened against a predetermined molecular library to remove molecules contained in the predetermined molecular library.

[0132] In step 1108, a screening process of the druggable property parameters is performed.

[0133] The generated candidate molecules are screened for druggability. In summary, the specific screening details include the following steps 1-2: 1. Screening is performed using a Mini Capsule Filter (MCF), mainly removing candidate molecules containing active or toxic groups; 2. Screening the physicochemical properties of the candidate molecules. For example, the candidate molecules must satisfy the following conditions: molecular weight MW≦550, water solubility in the range [-5,6], topological polar surface area of ​​the molecule ≦120, number of rotatable bonds<11, number of hydrogen bond acceptors in the range [0,11], number of hydrogen bond donors in the range [0,5], etc.

[0134] Note that the above steps 1107 and 1108 are two parallel steps.

[0135] In step 1109, a molecular 3D similarity calculation process is performed.

[0136] After obtaining the filtered candidate molecules, the candidate molecules and the reference molecules are required to be as similar as possible, so as to increase the probability of screening molecules that show activity, and thus are more likely to retain activity. Using the RDKit tool or other 3D similarity calculation tools, the 3D similarity between the candidate molecules and the reference molecules is calculated, and the top k candidate molecules are extracted to carry out the subsequent generation flow of the reconstituted molecule.

[0137] In step 1110, the main force screening and virtual screening process is performed.

[0138] The screening for the candidate molecules further uses the protein pocket information to extract the main effects of the generated candidate molecules, and screens out and eliminates molecules that do not contain the effects of the reference molecules. Then, a molecular activity prediction model is used to input the protein pocket information and the generated candidate molecules, output the PIC50 activity values, and sort the candidate molecules to achieve the screening purpose.

[0139] In step 1111, a virtual inspection process is performed.

[0140] In some embodiments, manual screening is performed on candidate molecules that remain after multiple rounds of screening.

[0141] In step 1112, the reconstituted preferred molecule is obtained.

[0142] Typically, several dozen candidate molecules are screened to select a preferred molecule.

[0143] In conclusion, the molecular structure reconstruction method according to the embodiment of the present application involves decomposing the structure of a reference molecule to obtain shard fragments and side chain fragments, and then using the side chain fragments as structural conditions to predict new molecular fragment structures for replacing the shard fragments to obtain candidate fragments, i.e., cutting the molecule to learn the structural rules of the related fragments, and using the newly generated candidate fragments to replace the shard fragments in the original molecule to achieve molecular modification. Since a large number of candidate fragments can be generated based on the structural conditions, the probability that the newly generated molecule is different from the previous molecular structure is improved, thereby improving the success rate of molecular reconstruction.

[0144] FIG. 12 is a schematic diagram of a molecular structure reconstruction device according to one exemplary embodiment of the present application. As shown in FIG. 12, the device includes: an acquisition module 1210 for acquiring structural data of a reference molecule, the reference molecule being a molecule that exhibits activity at a target target; a decomposition module 1220 for decomposing the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule, the molecular fragments including shard fragments of the reference molecule and side chain fragments corresponding to the shard fragments; A generation module 1230 that performs feature analysis on the data set of molecular fragments to obtain candidate fragments for replacing the shard fragments; The generation module 1230 further generates structural data for a reconstituted molecule based on the candidate fragments and the side chain fragments, wherein the reconstituted molecule exhibits activity at the target target.

[0145] In a preferred embodiment, the acquiring module 1210 further acquires a predetermined decomposition rule, the predetermined decomposition rule including at least one of a skeleton decomposition rule and a rotatable bond decomposition rule; The decomposition module 1220 further decomposes the structural data of the reference molecule according to a predefined decomposition rule to obtain a data set of at least one molecular fragment corresponding to the reference molecule.

[0146] In one preferred embodiment, the predetermined decomposition rules include decomposition rules for the skeleton; The decomposition module 1220 further extracts a skeletal structure that satisfies skeletal requirements from the structural data of the reference molecule, removes the skeletal structure from the structural data of the reference molecule to obtain a side chain structure corresponding to the skeletal structure, and obtains a dataset of the at least one group of molecular fragments corresponding to the reference molecule, with the skeletal structure as the shard fragment and the side chain structure as the side chain fragment.

[0147] In a preferred embodiment, the decomposition module 1220 further extracts a primary skeletal structure from the structural data of the reference molecule, where the primary skeletal structure is the maximum skeletal structure in the reference molecule, extracts a secondary skeletal structure from the primary skeletal structure, and determines the skeletal structure that satisfies the skeletal requirements from the primary skeletal structure and the secondary skeletal structure.

[0148] In one preferred embodiment, the framework requirement includes at least one of a ring number requirement, a heavy atom number requirement, and a rotatable bond requirement; The decomposition module 1220 is further configured to extract, in response to the skeleton requirement including the ring number requirement, from the structural data of the reference molecule, the skeleton structure having a ring number within the range of the ring number requirement; The decomposition module 1220 is further configured to extract, in response to the backbone requirement including the heavy atom number requirement, from the structural data of the reference molecule, the backbone structure whose number of heavy atoms is within the range of the heavy atom number requirement; The decomposition module 1220 is further responsive to the backbone requirements, including the rotatable bond requirements, to extract from the structural data of the reference molecule the backbone structures whose rotatable bond numbers are within the range of the rotatable bond number requirements.

[0149] In one preferred embodiment, the predetermined decomposition rules include decomposition rules for the rotatable bonds, The decomposition module 1220 further decomposes the structural data of the reference molecule from the rotatable bonds to obtain the shard fragments that satisfy shard requirements, removes the shard fragments from the reference molecule to obtain the side chain fragments, and obtains a dataset of at least one group of molecular fragments corresponding to the reference molecule based on the shard fragments and the side chain fragments.

[0150] In one preferred embodiment, each shard fragment comprises a shard skeleton, and the shard requirements comprise at least one of shard base requirements and shard skeleton requirements; The shard basic requirements include at least one of a shard ring number requirement, a shard atom number requirement, a shard rotatable bond requirement, and a shard structure requirement; The shard skeleton requirements include at least one of a skeleton atom number requirement and a skeleton rotatable bond requirement for the shard skeleton.

[0151] In one preferred embodiment, as shown in FIG. 13, the generating module 1230: A screening unit 1231 for screening the candidate fragments according to a predetermined screening rule, the predetermined screening rule including at least one of a legality screening rule, a uniqueness screening rule, a heavy atom number screening rule, and a ring screening rule; and a generating unit 1232 for generating structural data of the reconstructed molecule based on the remaining fragments obtained after screening and the side chain fragments.

[0152] In a preferred embodiment, the acquisition module 1210 further acquires a fragment similarity between the remaining fragment and the shard fragment; The generating unit 1232 further retains q (q is a positive integer) remaining fragments with the highest fragment similarity, and generates structural data of the reconstructed molecule together with the side chain fragments.

[0153] In one preferred embodiment, the generation module 1230 further inputs the dataset of the molecular fragment group into a molecular reconstruction model, and performs feature analysis on the dataset of the molecular fragment group by the molecular reconstruction model to output and obtain the candidate fragment for replacing the shard fragment.

[0154] In one preferred embodiment, the molecular reconstruction model comprises an encoder and a decoder in a codec architecture, The generation module 1230 further encodes the shard fragment using the encoder based on the side chain fragment to obtain fragment features, perturbs the fragment features using a predetermined perturbation rule to obtain perturbation features, and decodes the perturbation features using the decoder to generate the candidate fragment for replacing the shard fragment.

[0155] In a preferred embodiment, the generation module 1230 further maps the side chain fragment into a feature space to obtain a side chain space feature, uses the side chain space feature as a structural condition, and encodes the shard fragment by the encoder to obtain the fragment feature.

[0156] In a preferred embodiment, the generation module 1230 further encodes the shard fragment to obtain encoded features, takes the side chain space features as structural conditions, and maps the encoded features into the feature space to obtain fragment features.

[0157] In one preferred embodiment, the molecular reconstruction model comprises model parameters: The acquisition module 1210 acquires structural data of a sample molecule, the sample molecule being decomposable into at least one set of sample shards and sample side chains; The device comprises: The system further includes a training module 1240 that performs feature analysis on the structural data of the sample molecule using the molecular reconstruction model to output and obtain a reconstruction shard for replacing the sample shard, and adjusts model parameters based on the degree of difference between the sample shard and the reconstruction shard.

[0158] In one preferred embodiment, the training module 1240 further adjusts the model parameters according to the discreteness of the mapping in the feature space between the sample shard and the reconstructed shard, and the feature space is determined by the model parameters.

[0159] In one preferred embodiment, the generation module 1230 a joining unit 1233 that joins the hydrogen sites in the candidate fragment and the side chain fragment to obtain n candidate molecules (n is a positive integer), a screening unit 1231 that screens the n candidate molecules based on the molecular structure data in a predetermined molecular library to obtain the structure data of the reconstructed molecule.

[0160] In one preferred embodiment, the screening unit 1231 further screens the n candidate molecules based on the molecular structure data in the predetermined molecular library to obtain m candidate molecules (0 ≦ m ≦ n) that are not included in the predetermined molecular library, the generation module 1230 further includes a generation unit 1232 that determines the structure data of the reconstructed molecule from the m candidate molecules.

[0161] In one preferred embodiment, the generation unit 1232 further obtains m molecular similarities corresponding to the m candidate molecules to the reference molecule respectively, the i-th candidate molecule (i is a positive integer) corresponds to the i-th molecular similarity, and determines the structure data of the reconstructed molecule from k candidate molecules (0 < k < m) with the highest molecular similarities.

[0162] In view of the above, the molecular structure reconstruction device of the embodiment of the present application decomposes the structural data of a reference molecule to obtain shard fragments and side chain fragments, and then predicts a new molecular fragment structure for replacing the shard fragment using the side chain fragment as a structural condition to obtain a candidate fragment, i.e., cuts the molecule to learn the structural rules of the related fragments, and uses the newly generated candidate fragment to replace the shard fragment in the original molecule to realize molecular modification. Since the number of candidate fragments that can be generated based on the structural conditions is large, the probability that the newly generated molecule is different from the previous molecular structure is improved, thereby improving the success rate of molecular reconstruction.

[0163] It should be noted that the molecular structure reconstructing device according to the above embodiment is described by taking only the division of each of the above functional modules as an example, but in actual application, the above functions can be assigned to different functional modules to be completed as necessary, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the above-described functions. In addition, the molecular structure reconstructing device according to the above embodiment belongs to the same concept as the embodiment of the molecular structure reconstructing method, and the specific realization process thereof can be referred to the embodiment of the method, and will not be described again here.

[0164] FIG. 14 shows a schematic block diagram of a computer device according to an exemplary embodiment of the present application, which can be implemented as a server or a terminal. The computing device 1400 includes a central processing unit (CPU) 1401, a system memory 1404 including a random access memory (RAM) 1402 and a read only memory (ROM) 1403, and a system bus 1405 connecting the system memory 1404 to the central processing unit 1401. The computing device 1400 further includes a mass storage device 1406 for storing an operating system 1413, application programs 1414, and other program modules 1415.

[0165] The mass storage device 1406 is connected to the central processing unit 1401 via a mass storage controller (not shown) that is connected to the system bus 1405. The mass storage device 1406 and its associated computer readable media provide non-volatile storage for the computing device 1400. That is, the mass storage device 1406 may include a computer readable medium (not shown), such as a hard disk or a Compact Disc Read Only Memory (CD-ROM) drive.

[0166] Without loss of generality, computer readable media may include computer storage media and communication media. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes RAM, ROM, Erasable Programmable Read Only Memory (EPROM), Electrically Erasable Programmable Read Only Memory (EEPROM), Flash memory or other solid-state storage technology, CD-ROM, Digital Versatile Disc (DVD) or other optical storage, tape cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices. Of course, those skilled in the art will recognize that computer storage media is not limited to the above types. The system memory 1404 and mass storage device 1406 may be collectively referred to as memory.

[0167] According to various embodiments of the present application, the computing device 1400 may also be connected to a remote computer in a network, such as the Internet, for execution via the network, i.e., the computing device 1400 may be connected to a network 1412 via a network interface unit 1411 connected to the system bus 1405, or may be connected to another type of network or remote computer system (not shown) using the network interface unit 1411.

[0168] The memory further includes one or more programs, the one or more programs being stored in the memory and configured to be executed by the CPU.

[0169] According to an embodiment of the present application, there is further provided a computer device including a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by the processor to realize the method for reconstructing a molecular structure described in the embodiments of each of the above methods.

[0170] According to an embodiment of the present application, there is further provided a computer-readable storage medium storing at least one instruction, at least one program, code set or instruction set, which, when loaded and executed by a processor, realizes the method for reconstructing a molecular structure described in the embodiments of each of the above methods.

[0171] According to an embodiment of the present application, there is further provided a computer program product or a computer program comprising computer instructions, the computer instructions being stored in a computer readable storage medium, a processor of a computing device reading the computer instructions from the computer readable storage medium, and the processor executing the computer instructions to cause the computing device to perform the method for reconstructing a molecular structure according to any one of the above embodiments.

Claims

1. 1. A computational device implemented method for reconstructing a molecular structure, comprising: obtaining structural data for a reference molecule, the reference molecule being a molecule that exhibits activity at a target target; The computing device decomposes the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule, the molecular fragments including shard fragments of the reference molecule and side chain fragments corresponding to the shard fragments; performing feature analysis on the data set of molecular fragments by the computing device to obtain candidate fragments for replacing the shard fragment; generating structural data for a reconstituted molecule based on the candidate fragments and the side chain fragments, the reconstituted molecule exhibiting activity at the target target; The step of decomposing the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule includes: obtaining predetermined decomposition rules, the predetermined decomposition rules including skeletal decomposition rules; and decomposing the structural data of the reference molecule based on a predetermined decomposition rule to obtain a data set of at least one molecular fragment group corresponding to the reference molecule; The step of structurally decomposing the structural data of the reference molecule based on a predetermined decomposition rule to obtain a data set of at least one molecular fragment group corresponding to the reference molecule includes: Extracting a backbone structure that satisfies a backbone requirement from the structural data of the reference molecule; removing the backbone structure from the structural data of the reference molecule to obtain a side chain structure corresponding to the backbone structure; obtaining a dataset of the at least one molecular fragment corresponding to the reference molecule, the skeleton structures being the shard fragments and the side chain structures being the side chain fragments; The step of performing a feature analysis on the group of molecular fragments to obtain candidate fragments for replacing the shard fragments includes: inputting said data set of molecular fragments into a molecular reconstruction model; and performing feature analysis on the data set of the molecular fragments by the molecular reconstruction model to output and obtain the candidate fragments for replacing the shard fragments; The molecular reconstruction model includes an encoder and a decoder in a codec architecture; The step of performing feature analysis on the data set of the molecular fragment group by the molecular reconstruction model to output and obtain the candidate fragment for replacing the shard fragment, encoding the shard fragment by the encoder based on the side chain fragment to obtain fragment features; perturbing the fragment features according to a predefined perturbation rule to obtain perturbed features; and decoding, with the decoder, the perturbation features to generate the candidate fragment for replacing the shard fragment.

2. The step of extracting a scaffold structure satisfying a scaffold requirement from the reference molecule includes: extracting a primary backbone structure from structural data of the reference molecule, the primary backbone structure being a maximal backbone structure in the reference molecule; extracting a secondary skeletal structure from the primary skeletal structure; and determining the skeletal structure that satisfies the skeletal requirements from the primary skeletal structure and the secondary skeletal structure.

3. The framework requirement includes at least one of a number of rings requirement, a number of heavy atoms requirement, and a rotatable bond requirement; The step of extracting a scaffold structure satisfying a scaffold requirement from the reference molecule includes: extracting, in response to the skeleton requirement including the ring number requirement, from structural data of the reference molecule, the skeleton structure whose ring number is within the range of the ring number requirement; extracting, in response to the framework requirement including the heavy atom number requirement, from structural data of the reference molecule, the framework structure whose number of heavy atoms is within the range of the heavy atom number requirement; 2. The method of claim 1, comprising the step of: in response to the backbone requirements including the rotatable bond requirements, extracting from structural data of the reference molecule the backbone structures whose rotatable bond number is within the range of the rotatable bond number requirements.

4. 1. A computational device implemented method for reconstructing a molecular structure, comprising: obtaining structural data for a reference molecule, the reference molecule being a molecule that exhibits activity at a target target; The computing device decomposes the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule, the molecular fragments including shard fragments of the reference molecule and side chain fragments corresponding to the shard fragments; performing feature analysis on the data set of molecular fragments by the computing device to obtain candidate fragments for replacing the shard fragment; generating structural data for a reconstituted molecule based on the candidate fragments and the side chain fragments, the reconstituted molecule exhibiting activity at the target target; The step of decomposing the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule includes: obtaining a predefined decomposition rule, the predefined decomposition rule including a decomposition rule for a rotatable bond; and decomposing the structural data of the reference molecule based on a predetermined decomposition rule to obtain a data set of at least one molecular fragment group corresponding to the reference molecule; The step of structurally decomposing the structural data of the reference molecule based on a predetermined decomposition rule to obtain a data set of at least one molecular fragment group corresponding to the reference molecule includes: severing the structural data of the reference molecule from the rotatable bond to obtain the shard fragments that satisfy shard requirements; removing the shard fragment from the reference molecule to obtain the side chain fragment; obtaining a data set of at least one molecular fragment corresponding to the reference molecule based on the shard fragments and the side chain fragments; each shard fragment comprises a shard skeleton, and the shard requirements comprise at least one of shard base requirements and shard skeleton requirements; The shard basic requirement includes at least one of a shard ring number requirement, a shard atom number requirement, a shard rotatable bond requirement, and a shard structure requirement; the shard skeleton requirements include at least one of a skeleton atom number requirement and a skeleton rotatable bond requirement for the shard skeleton; The step of performing a feature analysis on the group of molecular fragments to obtain candidate fragments for replacing the shard fragments includes: inputting said data set of molecular fragments into a molecular reconstruction model; and performing feature analysis on the data set of the molecular fragments by the molecular reconstruction model to output and obtain the candidate fragments for replacing the shard fragments; The molecular reconstruction model includes an encoder and a decoder in a codec architecture; The step of performing feature analysis on the data set of the molecular fragment group by the molecular reconstruction model to output and obtain the candidate fragment for replacing the shard fragment, encoding the shard fragment by the encoder based on the side chain fragment to obtain fragment features; perturbing the fragment features according to a predefined perturbation rule to obtain perturbed features; and decoding, with the decoder, the perturbation features to generate the candidate fragment for replacing the shard fragment.

5. encoding the shard fragment by the encoder based on the side chain fragment to obtain fragment features, mapping the side chain fragments into a feature space to obtain side chain spatial features; The method of claim 1 or 4, further comprising: encoding the shard fragment with the encoder using the side chain spatial features as structural conditions to obtain the fragment features.

6. The step of encoding the shard fragment by the encoder using the side chain spatial feature as a structural condition to obtain the fragment feature includes: encoding the shard fragments to obtain encoded features; and mapping the encoding features to the feature space as structural conditions on the side chain spatial features to obtain fragment features.

7. The training process of the molecular reconstruction model executed by the computer device comprises: The computing device acquiring structural data of a sample molecule, the sample molecule being decomposable into at least one set of sample shards and sample side chains; The computing device performs a feature analysis on the structural data of the sample molecule using the molecular reconstruction model to output and obtain a reconstruction shard for replacing the sample shard; The method of claim 1 or 4, further comprising: the computing device adjusting model parameters of the molecular reconstitution model with the degree of dissimilarity between the sample shard and the reconstituted shard.

8. The step of adjusting the model parameters according to the degree of dissimilarity between the sample shard and the reconstructed shard includes:

8. The method of claim 7, comprising adjusting the model parameters at a discreteness of a mapping in a feature space between the sample shards and the reconstructed shards, the feature space being determined by the model parameters.

9. generating structural data of the reconstituted molecule based on the candidate fragments and the side chain fragments, screening the candidate fragments based on predefined screening rules, the predefined screening rules including at least one of a legality screening rule, a uniqueness screening rule, a heavy atom number screening rule, and a ring screening rule; and generating structural data for the reconstituted molecule based on the remaining fragments and the side chain fragments obtained after screening.

10. generating structural data of the reconstituted molecule based on the remaining fragments and the side chain fragments obtained after screening, obtaining a fragment similarity between the remaining fragment and the shard fragment; and retaining the q remaining fragments having the highest fragment similarity (q being a positive integer) and generating structural data for the reconstructed molecule together with the side chain fragments.

11. generating structural data of the reconstituted molecule based on the candidate fragments and the side chain fragments, joining hydrogen sites in the candidate fragments with the side chain fragments to obtain n candidate molecules (n is a positive integer); The method according to claim 1 or 4, further comprising a step of screening the n candidate molecules based on molecular structure data in a predetermined molecular library to obtain structural data of the reconstituted molecule.

12. The step of screening the n candidate molecules based on molecular structure data in the predetermined molecular library to obtain structural data of the reconstituted molecule includes: Screening the n candidate molecules based on molecular structure data in the predetermined molecular library to obtain m (0≦m≦n) candidate molecules that are not included in the predetermined molecular library; and determining structural data for the reconstituted molecule from the m candidate molecules.

13. The step of determining structural data of the reconstituted molecule from the m candidate molecules comprises: obtaining m molecular similarities of the m candidate molecules, each corresponding to the reference molecule, where an i-th candidate molecule corresponds to an i-th molecular similarity, where i is a positive integer; and determining structural data of the reconstituted molecule from the k (0<k<m) candidate molecules with the highest molecular similarity.

14. an acquisition module for acquiring structural data of a reference molecule, the reference molecule being a molecule that exhibits activity at a target target; a decomposition module that decomposes the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule, the molecular fragments including shard fragments of the reference molecule and side chain fragments corresponding to the shard fragments; A generation module that performs feature analysis on the data set of molecular fragments to obtain candidate fragments for replacing the shard fragments; The generation module further generates structural data for a reconstituted molecule based on the candidate fragments and the side chain fragments, the reconstituted molecule exhibiting activity at the target target; Decomposing the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule, Obtaining predetermined decomposition rules, the predetermined decomposition rules including skeletal decomposition rules; and and decomposing the structural data of the reference molecule based on a predetermined decomposition rule to obtain a data set of at least one molecular fragment group corresponding to the reference molecule; Decomposing the structural data of the reference molecule based on a predetermined decomposition rule to obtain a data set of at least one molecular fragment group corresponding to the reference molecule, Extracting a backbone structure that satisfies a backbone requirement from the structural data of the reference molecule; removing the backbone structure from structural data of the reference molecule to obtain a side chain structure corresponding to the backbone structure; obtaining a dataset of the at least one molecular fragment corresponding to the reference molecule, the skeleton structures being the shard fragments and the side chain structures being the side chain fragments; performing feature analysis on the group of molecular fragments to obtain candidate fragments for replacing the shard fragments; inputting the data set of molecular fragments into a molecular reconstruction model; Performing feature analysis on the data set of the molecular fragment group by the molecular reconstruction model to output and obtain the candidate fragment for replacing the shard fragment; The molecular reconstruction model includes an encoder and a decoder in a codec architecture; Performing feature analysis on the data set of the molecular fragment group by the molecular reconstruction model to output and obtain the candidate fragment for replacing the shard fragment; encoding the shard fragment by the encoder based on the side chain fragment to obtain fragment features; perturbing the fragment features according to a predefined perturbation rule to obtain perturbed features; and decoding the perturbation features by the decoder to generate the candidate fragments for replacing the shard fragments.

15. an acquisition module for acquiring structural data of a reference molecule, the reference molecule being a molecule that exhibits activity at a target target; a decomposition module that decomposes the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule, the molecular fragments including shard fragments of the reference molecule and side chain fragments corresponding to the shard fragments; A generation module that performs feature analysis on the data set of molecular fragments to obtain candidate fragments for replacing the shard fragments; The generation module further generates structural data for a reconstituted molecule based on the candidate fragments and the side chain fragments, the reconstituted molecule exhibiting activity at the target target; Decomposing the structural data of the reference molecule to obtain a data set of molecular fragments corresponding to the reference molecule, obtaining a predetermined decomposition rule, the predetermined decomposition rule including a decomposition rule for a rotatable bond; and decomposing the structural data of the reference molecule based on a predetermined decomposition rule to obtain a data set of at least one molecular fragment group corresponding to the reference molecule; Decomposing the structural data of the reference molecule based on a predetermined decomposition rule to obtain a data set of at least one molecular fragment group corresponding to the reference molecule, severing the structural data of the reference molecule from the rotatable bond to obtain the shard fragment that satisfies a shard requirement; removing the shard fragment from the reference molecule to obtain the side chain fragment; and obtaining a data set of at least one molecular fragment corresponding to the reference molecule based on the shard fragments and the side chain fragments; each shard fragment comprises a shard skeleton, and the shard requirements comprise at least one of shard base requirements and shard skeleton requirements; The shard basic requirement includes at least one of a shard ring number requirement, a shard atom number requirement, a shard rotatable bond requirement, and a shard structure requirement; the shard skeleton requirements include at least one of a skeleton atom number requirement and a skeleton rotatable bond requirement for the shard skeleton; performing feature analysis on the group of molecular fragments to obtain candidate fragments for replacing the shard fragments; inputting the data set of molecular fragments into a molecular reconstruction model; Performing feature analysis on the data set of the molecular fragment group by the molecular reconstruction model to output and obtain the candidate fragment for replacing the shard fragment; The molecular reconstruction model includes an encoder and a decoder in a codec architecture; Performing feature analysis on the data set of the molecular fragment group by the molecular reconstruction model to output and obtain the candidate fragment for replacing the shard fragment; encoding the shard fragment by the encoder based on the side chain fragment to obtain fragment features; perturbing the fragment features according to a predefined perturbation rule to obtain perturbed features; and decoding the perturbation features by the decoder to generate the candidate fragments for replacing the shard fragments.

16. A computing device including a processor and a memory, the memory storing at least one program, the at least one program being loaded and executed by the processor to implement the method of claim 1 or 4.

17. A computer program product which, when run on a computer, causes the computer to carry out the method according to claim 1 or 4.

Images