Processing method and device for molecular skeleton transition, medium and electronic equipment

Abstract

Embodiments of the present application provide a molecular skeleton transition processing method, device, medium and electronic equipment. The molecular skeleton transition processing method comprises: generating an atom hidden vector corresponding to a reference drug molecule according to a connection graph structure corresponding to the reference drug molecule; performing atom occlusion processing on the atom hidden vector to obtain a skeleton hidden vector and a side chain hidden vector contained in the atom hidden vector; generating a target skeleton hidden vector having a specified transition degree with the skeleton hidden vector according to the spatial distribution of the skeleton hidden vector; and generating a transition drug molecule according to the target skeleton hidden vector and the side chain hidden vector. The technical solution of the embodiments of the present application can improve the novelty of the newly generated drug molecule.

Description

Technical Field

[0001] The present application relates to the field of computer and communication technology, and in particular to a method, device, medium and electronic equipment for processing molecular skeleton transitions. Background Art

[0002] Skeleton hopping is a crucial tool in drug design. Its primary purpose is to modify the structure of existing patented molecules, replace the local structure of complex natural products, and improve the pharmacokinetic properties of molecules by modifying their scaffolds. Skeleton hopping approaches in related technologies are primarily based on pharmacophore models and molecular shape similarity searches. However, these rule-based approaches struggle to transcend the design thinking of drug chemists, resulting in insufficient novelty in the resulting molecules. Summary of the Invention

[0003] The embodiments of the present application provide a method, device, medium and electronic device for processing molecular skeleton transitions, thereby improving the novelty of newly generated drug molecules at least to a certain extent.

[0004] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by practice of the present application.

[0005] According to one aspect of an embodiment of the present application, a method for processing molecular skeleton transitions is provided, comprising: generating an atomic latent vector corresponding to a reference drug molecule based on a connection graph structure corresponding to the reference drug molecule; performing atomic occlusion processing on the atomic latent vector to obtain the skeleton latent vector and the side chain latent vector contained in the atomic latent vector; generating a target skeleton latent vector having a specified transition degree with the skeleton latent vector based on the spatial distribution of the skeleton latent vector; and generating a drug molecule after the transition based on the target skeleton latent vector and the side chain latent vector.

[0006] According to one aspect of an embodiment of the present application, a device for processing molecular skeleton transitions is provided, comprising: a first generation unit, configured to generate an atomic latent vector corresponding to a reference drug molecule based on a connection graph structure corresponding to the reference drug molecule; a first processing unit, configured to perform atomic occlusion processing on the atomic latent vector to obtain a skeleton latent vector and a side chain latent vector contained in the atomic latent vector; a second generation unit, configured to generate a target skeleton latent vector having a specified transition degree with the skeleton latent vector based on the spatial distribution of the skeleton latent vector; and a third generation unit, configured to generate a drug molecule after the transition based on the target skeleton latent vector and the side chain latent vector.

[0007] In some embodiments of the present application, based on the aforementioned scheme, the nodes in the connection graph structure represent atoms in the reference drug molecules; the first generation unit is configured to: based on the node features and edge features contained in the connection graph structure, calculate the node information of each node in the connection graph structure through a graph encoder, the node features represent the atomic features in the reference drug molecules, and the edge features represent the features between the atoms in the reference drug molecules; generate the latent vector of each node according to the node information of each node and the node features of each node; generate the atomic latent vector corresponding to the reference drug molecule according to the latent vector of each node and the atoms contained in the reference drug molecule.

[0008] In some embodiments of the present application, based on the aforementioned solution, the first generating unit is configured to: calculate information between the first node and the second node in a subsequent hidden layer based on a node feature of a first node in the connection graph structure, a node feature of a second node in the connection graph structure, and edge information between the first node and other nodes among the neighboring nodes of the first node except the second node in a previous hidden layer, wherein the first node is any node in the connection graph structure, and the second node is a neighboring node of the first node in the connection graph structure;

[0009] According to the side information between the first node and the second node in the previous hidden layer and the information between the first node and the second node in the subsequent hidden layer, the side information between the first node and the second node in the subsequent hidden layer is calculated; wherein, the side information between the two nodes in the connection graph structure in the initial hidden layer is calculated based on the node features of one of the two nodes and the edge features between the two nodes; the side information corresponding to each node in all hidden layers is summed to obtain the node information of each node.

[0010] In some embodiments of the present application, based on the aforementioned scheme, the first processing unit is configured to: obtain a bit vector corresponding to the reference drug molecule, the length of the bit vector is the same as the number of atoms contained in the reference drug molecule, and the bit value corresponding to the backbone atom in the bit vector is a first value; based on the bit vector, select the backbone atom's latent vector and the side chain atom's latent vector from the atom latent vector corresponding to the reference drug molecule; generate the backbone latent vector through a multi-head attention mechanism and the backbone atom's latent vector, and generate the side chain latent vector based on the multi-head attention mechanism and the side chain atom's latent vector.

[0011] In some embodiments of the present application, based on the aforementioned scheme, the first processing unit is further configured to: obtain multiple skeleton clusters, the cluster centers of each skeleton cluster in the multiple skeleton clusters conform to the Gaussian mixture distribution; calculate the first distance between the skeleton latent vector and the cluster centers of each skeleton cluster, and select the target skeleton cluster to which the skeleton of the reference drug molecule belongs based on the first distance; determine the Gaussian mixture distribution to which the skeleton latent vector belongs based on the cluster center of the target skeleton cluster.

[0012] In some embodiments of the present application, based on the aforementioned scheme, the second generation unit is configured to: perform random sampling processing according to the specified transition degree and the target skeleton cluster to obtain an offset corresponding to the specified transition degree; add the skeleton latent vector and the offset corresponding to the specified transition degree to generate the target skeleton latent vector.

[0013] In some embodiments of the present application, based on the aforementioned scheme, the second generation unit is configured as: if the specified transition degree is the first transition degree, then a first offset is generated according to the product of the variance of the target skeleton cluster and the first vector obtained by random sampling, and the first offset is used as the offset corresponding to the first transition degree.

[0014] In some embodiments of the present application, based on the aforementioned scheme, the second generation unit is configured as: if the specified transition degree is the second transition degree, then selecting a first skeleton cluster from the multiple skeleton clusters whose cluster center is less than or equal to a first set value with respect to the target skeleton cluster; generating a second offset based on the product of the variance of the first skeleton cluster and the second vector obtained by random sampling, the cluster center of the target skeleton cluster and the cluster center of the first skeleton cluster, and using the second offset as the offset corresponding to the second transition degree.

[0015] In some embodiments of the present application, based on the aforementioned scheme, the second generation unit is configured as: if the specified transition degree is the third transition degree, then selecting a second skeleton cluster from the multiple skeleton clusters whose cluster center is greater than or equal to a second set value with the target skeleton cluster; generating a third offset based on the product of the variance of the second skeleton cluster and the third vector obtained by random sampling, the cluster center of the target skeleton cluster and the cluster center of the second skeleton cluster, and using the third offset as the offset corresponding to the third transition degree.

[0016] In some embodiments of the present application, based on the aforementioned scheme, the third generation unit is configured to: obtain the specified target and target activity value of the reference drug molecule; and generate the drug molecule after the transition based on the target skeleton latent vector, the side chain latent vector, the target of the reference drug molecule and the target activity value.

[0017] In some embodiments of the present application, based on the foregoing scheme, the device further comprises a second processing unit configured to: after generating the transitioned drug molecules, performing a molecular filtering process on the transitioned drug molecules according to the physical and chemical properties to obtain drug molecules with drug-like properties; obtaining a co-crystal structure corresponding to the reference drug molecule, and docking the drug molecules with drug-like properties into the co-crystal structure; removing drug molecules that do not match the co-crystal structure through the binding mode of the drug molecules with drug-like properties and the co-crystal structure to obtain screened drug molecules; and performing synthesis and verification of the compounds according to the docking of the screened drug molecules and the co-crystal structure.

[0018] In some embodiments of the present application, based on the foregoing scheme, the transitioned drug molecules are generated by a machine learning model; the device further comprises a third processing unit configured to: obtain a sample skeleton hidden vector corresponding to a sample molecule, and obtain a plurality of skeleton clusters, the cluster centers of each skeleton cluster in the plurality of skeleton clusters conforming to a Gaussian mixture distribution; calculate a second distance between the sample skeleton hidden vector of the sample molecule and the cluster center of each skeleton cluster, and select a skeleton cluster to which a sample skeleton of the sample molecule belongs according to the second distance; generate a distance-based cross-entropy loss according to the distance between the sample skeleton hidden vector and the cluster center of the skeleton cluster to which the sample skeleton belongs; generate a loss function of the machine learning model according to the cross-entropy loss and a prediction loss of the sample molecule by the machine learning model; and adjust parameters of the machine learning model based on the loss function.

[0019] In some embodiments of the present application, based on the foregoing scheme, the machine learning model comprises a decoder; and the third processing unit is further configured to: after obtaining the sample skeleton hidden vector and a sample side chain hidden vector corresponding to the sample molecule by the machine learning model, input the sample skeleton hidden vector, the sample side chain hidden vector, and a target molecule corresponding to the sample molecule into the decoder; and calculate the prediction loss according to an output of the decoder and the target molecule.

[0020] According to an aspect of some embodiments of the present application, there is provided a computer readable medium having stored thereon a computer program, which, when executed by a processor, implements the processing method for molecular skeleton transition as described in the above embodiments.

[0021] According to an aspect of some embodiments of the present application, there is provided an electronic device comprising: one or more processors; a storage device configured to store one or more programs, which, when executed by the one or more processors, cause the one or more processors to implement the processing method for molecular skeleton transition as described in the above embodiments.

[0022] According to an aspect of some embodiments of the present application, a computer program product or computer program is provided, which includes computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the processing method of molecular skeleton transition provided in various optional embodiments described above.

[0023] In the technical solutions provided in some embodiments of the present application, the atomic occlusion processing is performed on the atomic hidden vectors corresponding to the drug molecules to obtain the skeleton hidden vectors and the side chain hidden vectors, and then the target skeleton hidden vector with a specified transition degree is generated according to the spatial distribution of the skeleton hidden vectors, so that the generated target skeleton hidden vector can jump out of the design thinking stereotype of the medicinal chemist by mapping the skeleton hidden vector to the spatial distribution, and thus good novelty can be achieved, and meanwhile the technical solutions can be automatically executed by the device, reducing the labor and time costs.

[0024] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present application. BRIEF DESCRIPTION OF DRAWINGS

[0025] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application. It is clear that the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings. In the drawings:

[0026] Figure 1 A schematic diagram of an exemplary system architecture to which the technical solutions of the embodiments of the present application can be applied is shown;

[0027] Figure 2 A flowchart of a processing method of molecular skeleton transition according to an embodiment of the present application is shown;

[0028] Figure 3A A flowchart of generating atomic hidden vectors corresponding to reference drug molecules according to an embodiment of the present application is shown;

[0029] Figure 3B A flowchart of performing atomic occlusion processing on atomic hidden vectors corresponding to reference drug molecules according to an embodiment of the present application is shown;

[0030] Figure 3C A flowchart of generating a target skeleton hidden vector with a specified transition degree according to an embodiment of the present application is shown;

[0031] Figure 4A A schematic diagram of the structure of a machine learning model according to an embodiment of the present application is shown;

[0032] Figure 4B A schematic diagram of a processing process of a graph encoder according to an embodiment of the present application is shown;

[0033] Figure 5 A schematic diagram of an atomic shielding and readout function portion according to an embodiment of the present application is shown;

[0034] Figure 6 A schematic diagram of a distance representation method according to an embodiment of the present application is shown;

[0035] Figure 7 A schematic diagram of a processing process of a decoder according to an embodiment of the present application is shown;

[0036] Figure 8 A schematic diagram of a process for generating skeleton latent vectors and side chain latent vectors from a model according to one embodiment of the present application is shown;

[0037] Figure 9 A schematic diagram of a model decoding process according to an embodiment of the present application is shown;

[0038] Figure 10 A schematic diagram of a skeleton transition method according to an embodiment of the present application is shown;

[0039] Figure 11 A block diagram of a device for processing molecular skeleton transition according to an embodiment of the present application is shown;

[0040] Figure 12 A schematic diagram of the structure of a computer system suitable for implementing an electronic device according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0041] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be thorough and complete and will fully convey the concepts of the example embodiments to those skilled in the art.

[0042] In addition, described feature, structure or characteristic can be combined in one or more embodiments in any suitable manner.In the following description, many specific details are provided so as to provide a full understanding of the embodiments of the present application. However, it will be appreciated by those skilled in the art that the technical scheme of the present application can be put into practice without one or more of the specific details, or other methods, components, devices, steps etc. can be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring the various aspects of the application.

[0043] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0044] The flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, while others may be combined or partially combined. Therefore, the actual execution order may vary depending on the actual situation.

[0045] It should be noted that the term "plurality" used in this document refers to two or more. "And / or" describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A alone, A and B together, or B alone. The character " / " generally indicates an "or" relationship between the associated objects.

[0046] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also involves studying the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0047] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0048] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning through demonstration.

[0049] The solution provided in the embodiments of this application involves technologies such as artificial intelligence machine learning, specifically applying machine learning technology to the skeleton transition scheme of drug molecules.

[0050] Before introducing the molecular skeleton transition scheme of the present invention, we first introduce the processing schemes in the related art. The skeleton transition schemes proposed in the related art are mainly based on pharmacophore models, molecular shape-based searches, fingerprint-based chemical similarity searches, and machine learning algorithms.

[0051] Among them, the pharmacophore model simulates the active conformation of the ligand molecule through conformational search and molecular superposition, that is, retaining the molecular framework of the characteristic atoms required for activity. The most significant feature of the pharmacophore is the molecular interaction characteristics shared by a group of active molecules. Therefore, the pharmacophore does not represent a real molecule or a group of chemical groups, but is an abstract concept. Common pharmacophore features include: hydrogen bond acceptors and donors, positive and negative charge interactions, hydrophobic interactions, and aromatic ring interactions. If such pharmacophoric features can be transferred from one molecule to another, that is, the reference molecule and the test molecule have the same pharmacophore features, then backbone hopping can be achieved. A similar approach is the protein structure-based drug design method, which achieves backbone hopping by representing the interaction between the small molecule and the residues of the binding site in the protein as a vector and then searching the compound library for molecules with corresponding characteristic vectors.

[0052] Molecular shape-based searches primarily consider molecular volume in space to search for similarities, hoping to maintain binding to the target protein and achieve backbone replacement. This approach, like other search methods, faces significant challenges: long search times and limitations in existing chemical space, limiting the search to existing compound libraries. Furthermore, false positives are common, making it difficult to guarantee the activity of the resulting molecules.

[0053] With the development of artificial intelligence technology, especially its application in molecule generation, the ability of drug development has been accelerated. The biggest advantage of molecule generation is that it generates brand-new molecules, directly realizes the de novo design of drug molecules, and expands the existing molecular space. The molecule generation method based on AI algorithm proposed in related technologies pays too much attention to the reconstruction ability and legality of molecules, but it is difficult to meet the actual needs of pharmaceutical companies. For example, pharmaceutical companies hope to modify existing molecules and maintain the activity of molecules while breaking away from the existing structure. However, although the solutions in related technologies can meet the requirements in terms of activity maintenance, since they are all rule-based generation, it is difficult to break away from the design ideas of pharmaceutical experts, resulting in insufficient novelty of the newly generated molecules.

[0054] Based on the above problems, the embodiment of the present application provides a new solution for processing molecular skeleton transitions, which can ensure that the generated target skeleton latent vector can break away from the design thinking of pharmaceutical experts by mapping the skeleton latent vector to the spatial distribution, thereby achieving great novelty. At the same time, it can be automatically executed by the device, reducing manpower and time costs. The technical solution of the embodiment of the present application is described in detail below:

[0055] Figure 1 A schematic diagram shows an exemplary system architecture to which the technical solutions of the embodiments of the present application can be applied.

[0056] like Figure 1 As shown, the system architecture 100 may include a terminal 110 , a network 120 , and a server 130 . The terminal 110 and the server 130 are connected via the network 120 .

[0057] In one embodiment of the present application, the terminal 110 can be a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto. The network 120 can be a communication medium of various connection types that can provide a communication link between the terminal 110 and the server 130, for example, it can be a wired communication link, a wireless communication link or an optical fiber cable, etc., and the present application does not limit this. The server 130 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0058] It should be understood that Figure 1 The number of terminals 110, networks 120, and servers 130 in the embodiment is merely illustrative. Depending on the implementation requirements, there may be any number of terminals 110, networks 120, and servers 130. For example, the server 130 may be a server cluster consisting of multiple servers.

[0059] In one embodiment of the present application, a user may submit a reference drug molecule to server 130 via terminal 110 via network 120. The reference drug molecule is a molecule that requires skeleton transition processing, and the skeleton that requires transition may be identified. Identifying the skeleton that requires transition is not a necessary process. After obtaining the reference drug molecule, server 130 may convert the structure of the reference drug molecule into a connection graph structure, and then generate an atomic latent vector corresponding to the reference drug molecule based on the connection graph structure corresponding to the reference drug molecule.

[0060] After generating the atomic latent vector corresponding to the reference drug molecule, server 130 can perform atomic occlusion processing on the atomic latent vector to obtain the backbone latent vector and side chain latent vector contained in the atomic latent vector. To achieve backbone transition processing, a target backbone latent vector with a specified transition degree relative to the backbone latent vector can be generated based on the spatial distribution of the backbone latent vector. The transitioned drug molecule can then be generated based on the target backbone latent vector and the side chain latent vectors obtained above.

[0061] Because the technical solution of the embodiment of the present application enables the generated target skeleton latent vector to be mapped to the spatial distribution, it can break away from the design thinking pattern of pharmaceutical experts, thereby achieving great novelty, and can be automatically executed by the equipment, reducing manpower and time costs.

[0062] It should be noted that the molecular skeleton transition processing method provided in the embodiments of the present application is generally executed by the server 130, and accordingly, the molecular skeleton transition processing device is generally arranged in the server 130. However, in other embodiments of the present application, the terminal device can also have similar functions as the server, and thus execute the molecular skeleton transition processing scheme provided in the embodiments of the present application.

[0063] The implementation details of the technical scheme of the embodiments of the present application are described in detail as follows:

[0064] Figure 2 A flowchart of a molecular skeleton transition processing method according to one embodiment of the present application is shown, which can be executed by a device with a computing processing function, such as the server 130 shown in Figure 1 . Referring to Figure 2 , the molecular skeleton transition processing method at least includes steps S210 to S240, which are described in detail as follows:

[0065] In step S210, an atom hidden vector corresponding to a reference drug molecule is generated according to a connection graph structure corresponding to the reference drug molecule.

[0066] In one embodiment of the present application, the reference drug molecule is the molecule that needs to be processed by the skeleton transition, and the connection graph structure corresponding to the reference drug molecule is the connection graph structure converted from the structure of the reference drug molecule.

[0067] Alternatively, the connection graph structure corresponding to one drug molecule can be represented as G=(A, X, E), where A represents a connection matrix, X represents a node feature, and E represents an edge feature. In the connection graph structure, the node represents an atom in the drug molecule; the node feature is used to represent the atom feature in the drug molecule, which can include, for example: atomic weight, atomic charge number, atomic type, atomic valence state, whether the atom is in a ring, whether it is an atom in an aromatic ring, etc. The edge feature is used to represent the feature between the atoms in the drug molecule, which can include, for example: whether the edge is a single bond edge or a double bond edge, whether the edge is in a ring, whether the edge is in an aromatic ring, etc.

[0068] In one embodiment of the present application, after the connection graph structure corresponding to the reference drug molecule is generated, the atom hidden vector corresponding to the reference drug molecule can be generated according to the connection graph structure corresponding to the reference drug molecule, which can include, for example, the following steps S310a, S320a and S330a, which are described in detail as follows: Figure 3A

[0069] Step S310a, based on the node features and edge features contained in the connection graph structure, the node information of each node in the connection graph structure is calculated by a graph encoder. ​

[0070] In one embodiment of the present application, for the sake of convenience, the following is an example of a node v and a node w in a connection graph structure. It is assumed that the node feature of the node v can be expressed as x v , the node feature of node w can be expressed as x w , the edge feature between node v and node w can be expressed as e vw , the node information of node v can be expressed as m v , the hidden vector of node v can be expressed as h v Then the process of calculating the node information of each node in the connection graph structure through the graph encoder may include:

[0071] According to the node feature (denoted as x) of the first node in the connection graph structure (the first node is any node in the connection graph structure, such as node v) v ), the node feature (denoted as x) of the second node in the connection graph structure (the second node is the neighbor node of the first node in the connection graph structure, such as node w) w ), and the side information (denoted as ) between the first node and other nodes (such as node k) in the neighboring nodes of the first node except the second node in the previous hidden layer (assuming it is hidden layer t) ), calculate the information between the first node and the second node in the next hidden layer (assuming it is hidden layer t+1) (denoted as ).

[0072] Then, based on the edge information between the first node and the second node in the previous hidden layer And the information between the first node and the second node in the next hidden layer Calculate the side information between the first node and the second node in the next hidden layer (denoted as ); Among them, the edge information between two nodes in the connection graph structure in the initial hidden layer (for example, the edge information between node v and node w in the initial hidden layer can be expressed as ) is calculated based on the node feature of one of the two nodes and the edge feature between the two nodes. Based on the above calculations, the corresponding edge information of each node in all hidden layers can be summed to obtain the node information of each node.

[0073] In one embodiment of the present application, the aforementioned It can be obtained by the following formula 1:

[0074]

[0075] In the above formula 1, k∈{N(v)\w} indicates that node k is a node other than node w in the neighbor nodes N(v) of node v; ft (·) represents the aggregation process, which can be the process of combining the variables (i.e. x v 、x w and ) (similar to the cat(·) function below), or by replacing the variables (i.e. x v 、x w and ) are mapped to the same dimension and then summed or averaged. Alternatively, the variables (i.e. x v 、x w and ) are combined.

[0076] In one embodiment of the present application, the aforementioned It can be obtained by the following formula 2:

[0077]

[0078] In the above formula 2, g t (·) represents the update process, which can be a simple accumulation or averaging, or a GRU (Gated Recurrent Unit). If it is a GRU, then is the hidden layer input of GRU, is the actual input to the GRU.

[0079] In one embodiment of the present application, the aforementioned It can be obtained by the following formula 3:

[0080]

[0081] In Formula 3 above, τ(·) represents the ReLU (Rectified Linear Unit) function; W represents the parameter to be learned; and cat(·) represents concatenating two vectors to form a longer vector. For example, concatenating a 3-dimensional vector with a 5-dimensional vector yields an 8-dimensional vector.

[0082] In one embodiment of the present application, it is assumed that the node information of node v is m v , then the node information m v It can be obtained by the following formula 4:

[0083]

[0084] In the above formula 4, Represents the edge information between node k and node v in all hidden layers; k∈N(v) means that node k is a node in node v’s neighbor node N(v); Since there is edge information between neighbor nodes, m v In fact, it sums the side information of node v in all hidden layers.

[0085] Step S320a: Generate a latent vector for each node based on the node information and node features of each node.

[0086] In one embodiment of the present application, the node information m of the node v is obtained. v Afterwards, we can use the node information m of node v to v and node feature x of node v v Generate the hidden vector h of node v v , for example, can be obtained by the following formula 5:

[0087] h v =τ(W a ·cat(x v ,m v )) Formula 5

[0088] In the above formula 5, τ(·) represents the ReLU function; W a Indicates the parameters that need to be learned; cat(·) means concatenating two vectors to form a longer vector.

[0089] Step S330a: Generate an atomic latent vector corresponding to the reference drug molecule based on the latent vector of each node and the atoms contained in the reference drug molecule.

[0090] In one embodiment of the present application, after obtaining the latent vectors of each node, the atomic latent vectors corresponding to the reference drug molecules can be represented in a matrix manner, that is, the latent vectors of each node are arranged in a row and column manner in a matrix (such as an H matrix) to represent the atomic latent vectors corresponding to the reference drug molecules.

[0091] It should be noted that the aforementioned formulas 1 to 5 are merely examples. In other embodiments of the present application, the aforementioned formulas 1 to 5 may be appropriately modified (such as increasing the multiple, decreasing the multiple, increasing a certain value, decreasing a certain value, etc.) to obtain a new calculation formula.

[0092] Continue to refer to Figure 2 As shown, in step S220, atomic occlusion processing is performed on the atomic latent vector corresponding to the reference drug molecule to obtain the skeleton latent vector and the side chain latent vector contained in the atomic latent vector.

[0093] In one embodiment of the present application, after obtaining the atomic latent vector corresponding to the reference drug molecule, the process of performing atomic occlusion processing on the atomic latent vector corresponding to the reference drug molecule can be as follows: Figure 3B As shown, it includes step S310b, step S320b and step S330b, which are detailed as follows:

[0094] Step S310b: Obtain a bit vector corresponding to the reference drug molecule, where the length of the bit vector is the same as the number of atoms contained in the reference drug molecule, and the bit value corresponding to the skeleton atom in the bit vector is the first value.

[0095] Alternatively, in the aforementioned embodiment, the first value may be 1, for example, that is, if an atom belongs to a backbone atom, then the corresponding bit value in the bit vector is 1; if an atom belongs to a side chain atom, then the corresponding bit value in the bit vector is 0. In this case, the bit vector can be represented as the matrix S shown in the following formula 6: sca :

[0096]

[0097] In the above formula 6, i∈scaffold indicates that atom i belongs to the skeleton atom; Indicates that atom i does not belong to the backbone atoms.

[0098] In one embodiment of the present application, the aforementioned bit vector may be preset to indicate which atoms in the reference drug molecule are backbone atoms and which atoms are side chain atoms. Specifically, based on the backbone that needs to be transitioned (replaced) in the reference drug molecule, a structural search may be used to search and match the reference drug molecule to determine the backbone atoms and side chain atoms in the reference drug molecule, thereby obtaining the aforementioned bit vector.

[0099] In one embodiment of the present application, the aforementioned bit vector may be pre-set according to a set of backbone determination rules. The backbone determination rules may include various requirements, such as requirements for the number of heavy atoms in the backbone, requirements for the number of backbone rings, etc., which are not limited by the present application. Specifically, for a drug molecule, the backbone portion of the drug molecule can be automatically detected according to the backbone determination rules, and then the bit vector is generated based on the backbone portion and the portion other than the backbone portion (i.e., the side chain portion) in the drug molecule.

[0100] Step S320b: Selecting the implicit vectors of backbone atoms and side chain atoms from the atomic implicit vectors corresponding to the reference drug molecule based on the bit vector.

[0101] In one embodiment of the present application, it is assumed that the atomic latent vector corresponding to the reference drug molecule (ie, the latent vector of the original atom) is represented by H node , then the latent vector of the skeleton atom selected from the atomic latent vector corresponding to the reference drug molecule can be expressed as H node [S sca ], the latent vector of the side chain atom selected from the atomic latent vector corresponding to the reference drug molecule can be expressed as

[0102] Step S330b: Generate a skeleton latent vector through the multi-head attention mechanism and the latent vector of the skeleton atom, and generate a side chain latent vector based on the multi-head attention mechanism and the latent vector of the side chain atom.

[0103] In one embodiment of the present application, the multi-head attention mechanism is used to determine the score (i.e., weight) corresponding to the latent vector of each atom (skeleton atoms and side chain atoms), and then calculate the skeleton latent vector and the side chain latent vector based on this score.

[0104] Alternatively, assume that the atomic latent vector corresponding to the reference drug molecule (i.e., the latent vector of the original atom) is expressed as H node , then in one embodiment the skeleton latent vector Z sca The side chain hidden vector Z can be shown as the following formula 7: sc It can be expressed as the following formula 8:

[0105]

[0106]

[0107] In the above formulas 7 and 8, the softmax(·) function implements the function of the multi-head attention mechanism; W1 and W2 are both learnable parameters; Indicates H node The transpose of .

[0108] Of course, the aforementioned formulas 7 and 8 are merely examples. In other embodiments of the present application, the aforementioned formulas 7 and 8 may be appropriately modified (such as increasing the multiple, decreasing the multiple, increasing a certain value, decreasing a certain value, etc.) to obtain new calculation formulas.

[0109] Reference Figure 2 As shown, in step S230, a target skeleton latent vector having a specified transition degree with the skeleton latent vector is generated according to the spatial distribution of the skeleton latent vector.

[0110] In one embodiment of the present application, the spatial distribution of the skeleton latent vector can be a Gaussian mixture distribution, or a vMFM (von Mises-Fisher Mixture) distribution. The following is an example of a Gaussian mixture distribution:

[0111] In one embodiment of the present application, multiple skeleton clusters can be preset, and the cluster centers of each skeleton cluster in these multiple skeleton clusters conform to the Gaussian mixture distribution. Optionally, these multiple skeleton clusters can be obtained by clustering the skeletons of existing molecules using a skeleton clustering algorithm (i.e., a clustering algorithm), and the cluster centers of a skeleton cluster are fitted to the Gaussian mixture distribution, so that the skeleton latent vectors contained in such a skeleton cluster all belong to the spatial distribution corresponding to this Gaussian mixture distribution.

[0112] In this case, after obtaining the skeleton latent vector, the first distance between the skeleton latent vector and the cluster center of each skeleton cluster can be calculated, and then the target skeleton cluster to which the skeleton of the reference drug molecule belongs can be selected based on the first distance, and then the Gaussian mixture distribution to which the skeleton latent vector belongs can be determined based on the cluster center of the target skeleton cluster.

[0113] Specifically, the cluster center of the mth skeleton cluster can be expressed as (μ m ,σ m ), where μ m represents the center of the cluster, σ m In order to make a general statement, it is assumed that the skeleton latent vector corresponding to the i-th drug molecule is represented by Z sca,i , then the skeleton latent vector Z corresponding to the i-th drug molecule sca,i The distance d from the cluster center of the mth skeleton cluster i It can be shown as formula 9:

[0114]

[0115] The skeleton latent vector Z corresponding to the i-th drug molecule calculated based on the above formula 9 sca,i The distance d from the cluster center of the mth skeleton cluster i , we can select the skeleton latent vector Z corresponding to the cluster center and the i-th drug molecule sca,i The target skeleton cluster with the closest distance between them (denoted as c i ), and then based on the target skeleton cluster c i The cluster center (μ i ,σ i ) determines the Gaussian mixture distribution to which the skeleton cluster vector belongs.

[0116] When the spatial distribution of the skeleton latent vectors corresponding to the reference drug molecules is obtained and the target skeleton cluster is determined, the target skeleton latent vector with a specified transition degree between the skeleton latent vectors corresponding to the reference drug molecules can be generated according to requirements, which will be described in detail as follows:

[0117] As shown in FIG. 3, the process of generating the target skeleton latent vector with the specified transition degree according to one embodiment of the present application can include steps S310c and S320c, which will be described in detail as follows: Figure 3C

[0118] Step S310c, a random sampling process is performed according to the specified transition degree and the target skeleton cluster to obtain an offset corresponding to the specified transition degree.

[0119] Step S320c, the skeleton latent vector of the reference drug molecule is added to the offset corresponding to the specified transition degree to generate the target skeleton latent vector.

[0120] Optionally, the specified transition degree can be a transition within the same skeleton cluster, a transition between adjacent skeleton clusters, or a transition between distant skeleton clusters, which will be described as follows.

[0121] In one embodiment of the present application, if the specified transition degree is a first transition degree, a first offset can be generated according to the product of the variance of the target skeleton cluster and the first vector obtained by random sampling, and then the first offset is added to the skeleton latent vector corresponding to the reference drug molecule to generate the target skeleton latent vector. Optionally, the first transition degree can be a transition within the same skeleton cluster.

[0122] Specifically, assuming that the target skeleton cluster is represented as c i , and the skeleton latent vector corresponding to the reference drug molecule is represented as Z sca , the target skeleton latent vector with the first transition degree can be represented by the following formula 10:

[0123] Z new_sca = Z sca + σ 2 (c i ) × N(0, 1) Formula 10

[0124] In the above formula 10, Z new_sca represents the generated target skeleton latent vector; σ 2 (c i ) represents the variance of the Gaussian mixture distribution to which the cluster center of the target skeleton cluster c i conforms; and N(0, 1) represents a random sampling based on a distribution with a mean value of 0 and a standard deviation of 1.

[0125] ​In one embodiment of the present application, if the specified transition degree is a second transition degree, a first skeleton cluster whose cluster center is less than or equal to a first set value with respect to the target skeleton cluster can be selected from multiple skeleton clusters, and then a second offset is generated based on the product of the variance of the first skeleton cluster and the second vector obtained by random sampling, the cluster center of the target skeleton cluster, and the cluster center of the first skeleton cluster. The second offset is then added to the skeleton latent vector corresponding to the reference drug molecule to generate the target skeleton latent vector. Optionally, the second transition degree can be a transition in an adjacent skeleton cluster.

[0126] Specifically, assuming the target skeleton cluster is represented as c i , the first skeleton cluster is denoted as c j , the skeleton latent vector corresponding to the reference drug molecule is represented by Z sca , then the generated target skeleton latent vector with the second transition degree can be expressed by the following formula 11:

[0127]

[0128] In the above formula 11, Z new_sca represents the generated target skeleton latent vector; σ 2 (c j ) represents the first skeleton cluster c j The variance of the Gaussian mixture distribution that the cluster center conforms to; N(0,1) represents a random sampling based on a distribution with a mean of 0 and a standard deviation of 1; μ(c i ) represents the target skeleton cluster c i The center of μ(c j ) represents the first skeleton cluster c j The center of μ(c k ) represents the skeleton cluster c k The center of ; δ' represents the first set value; π(·) represents a multinomial matrix (multi-nominal) sampling; c j =π({c k |||μ(c i )-μ(c k )||≤δ',j≠i}) means finding a cluster with the skeleton cluster c i The skeleton cluster c whose cluster center is less than or equal to δ' j .

[0129] In one embodiment of the present application, if the specified transition degree is a third transition degree, a second skeleton cluster whose cluster center is greater than or equal to a second set value with respect to the target skeleton cluster can be selected from multiple skeleton clusters, and then a third offset is generated based on the product of the variance of the second skeleton cluster and the third vector obtained by random sampling, the cluster center of the target skeleton cluster, and the cluster center of the second skeleton cluster. The third offset is then added to the skeleton latent vector corresponding to the reference drug molecule to generate the target skeleton latent vector. Optionally, the third transition degree can be a transition in a more distant skeleton cluster.

[0130] Specifically, assuming the target skeleton cluster is represented as c i , the second skeleton cluster is denoted as c j , the skeleton latent vector corresponding to the reference drug molecule is represented by Z sca , then the generated target skeleton latent vector with the third transition degree can be expressed by the following formula 12:

[0131]

[0132] In the above formula 12, Z new_sca represents the generated target skeleton latent vector; σ 2 (c j ) represents the second skeleton cluster c j The variance of the Gaussian mixture distribution that the cluster center conforms to; N(0,1) represents a random sampling based on a distribution with a mean of 0 and a standard deviation of 1; μ(c i ) represents the target skeleton cluster c i The center of μ(c j ) represents the second skeleton cluster c j The center of μ(c k ) represents the skeleton cluster c k ; Δ represents the second set value; π(·) represents a multi-nominal matrix (multi-nominal) sampling; c j =π({c k |||μ(c i )-μ(c k )||≥Δ,j≠i}) means finding a cluster that matches the skeleton cluster c i The skeleton cluster c whose cluster center is greater than or equal to Δ j .

[0133] It should be noted that the aforementioned formulas 9 to 12 are merely examples. In other embodiments of the present application, the aforementioned formulas 9 to 12 can be appropriately modified (such as increasing the multiple, decreasing the multiple, increasing a certain value, decreasing a certain value, etc.) to obtain a new calculation formula.

[0134] Continue to refer to Figure 2As shown, in step S240, the drug molecule after the transition is generated according to the target skeleton latent vector and the side chain latent vector.

[0135] In one embodiment of the present application, the backbone latent vector in the reference drug molecule can be replaced by the target backbone latent vector and combined with the side chain latent vector to obtain the drug molecule after the transition.

[0136] In one embodiment of the present application, the target and target activity value of a specified reference drug molecule can also be obtained, and then the transitioned drug molecule can be generated based on the target backbone latent vector, side chain latent vector, target and target activity value of the reference drug molecule. The technical solution of this embodiment can limit the activity of the generated drug molecule by using the target and target activity value of the reference drug molecule.

[0137] In one embodiment of the present application, after generating the transitioned drug molecules, the generated drug molecules can be screened. Specifically, for example, the transitioned drug molecules can be filtered based on their physical and chemical properties to obtain drug-like drug molecules. The co-crystal structure corresponding to the reference drug molecule can then be obtained, and the drug-like drug molecules can be docked into the co-crystal structure. The binding mode between the drug-like drug molecules and the co-crystal structure can be used to remove drug molecules that do not match the co-crystal structure, thereby obtaining screened drug molecules. Compound synthesis and verification can then be performed based on the docking of the screened drug molecules with the co-crystal structure.

[0138] Alternatively, the co-crystal structure corresponding to the reference drug molecule may be a co-crystal structure of the reference drug molecule or a co-crystal structure of a compound of the same series as the reference drug molecule. The drug molecule that does not match the co-crystal structure may be a drug molecule whose configuration after docking is obviously unreasonable.

[0139] In one embodiment of the present application, the relevant processing in the above embodiment can be performed through a machine learning model. In the process of training the machine learning model, the technical solution of the embodiment of the present application proposes a solution for generating a loss function by using cross entropy loss and the prediction loss of the machine learning model for the sample molecules. The following describes how to obtain the cross entropy loss and prediction loss respectively:

[0140] In one embodiment of the present application, when calculating the cross entropy loss, the sample skeleton latent vector corresponding to the sample molecule can be obtained, and multiple skeleton clusters can be obtained at the same time (these multiple skeleton clusters can be the same as the multiple skeleton clusters used when processing the reference drug molecules mentioned above). The cluster centers of each skeleton cluster in these multiple skeleton clusters conform to the Gaussian mixture distribution. Then, the second distance between the sample skeleton latent vector of the sample molecule and the cluster centers of each skeleton cluster is calculated, and the skeleton cluster to which the sample skeleton of the sample molecule belongs is selected based on the second distance. Then, based on the distance between the sample skeleton latent vector and the cluster center of the skeleton cluster to which the sample skeleton belongs, a distance-based cross entropy loss is generated.

[0141] It should be noted that the method for obtaining the sample skeleton latent vector corresponding to the sample molecule is the same as the method for obtaining the skeleton latent vector corresponding to the reference drug molecule, and will not be repeated here. Furthermore, the formula for calculating the second distance between the sample skeleton latent vector of the sample molecule and the cluster center of each skeleton cluster can also be calculated using the aforementioned formula 9.

[0142] Optionally, the aforementioned formula 9 is used as an example for illustration (since the relevant calculation formulas and processing methods for the sample molecule and the reference drug molecule are the same, formula 9 can be used to calculate the distance between the skeleton latent vector corresponding to the reference drug molecule and the cluster center of the skeleton cluster, and can also be used to calculate the distance between the skeleton latent vector corresponding to the sample molecule and the cluster center of the skeleton cluster). Assume that the distance between the sample skeleton latent vector corresponding to the i-th drug molecule (which can be understood as the i-th sample molecule here) and the cluster center of the skeleton cluster to which the sample skeleton belongs is expressed as d i , then you can i On the basis of , a certain degree of flexibility is added to improve the accuracy of model training, as shown in formula 13:

[0143] d adj,i =d i +onehot(c i )×δd i Formula 13

[0144] In formula 13, d adj,i Indicates that in d i The distance after adding deflection on the basis of ; onehot(·) represents the function of one-hot encoding; c i Here, is used to indicate the skeleton cluster to which the sample skeleton of the i-th sample molecule belongs; δ indicates a parameter.

[0145] In one embodiment of the present application, after obtaining d adj,i Afterwards, the distance-based cross entropy loss L can be generated according to the following formula 11: cls :

[0146]

[0147] In formula 14, σ m represents the standard deviation of the Gaussian mixture distribution to which the cluster center of the mth skeleton cluster conforms.

[0148] In an embodiment of the present application, the machine learning model comprises a decoder. After obtaining the sample skeleton latent vector and the sample side chain latent vector corresponding to the sample molecule by the machine learning model, the sample skeleton latent vector, the sample side chain latent vector and the target molecule corresponding to the sample molecule are input into the decoder, and then the prediction loss of the machine learning model is calculated according to the output of the decoder and the target molecule.

[0149] It should be noted that the target molecule is a molecule expected to be generated after processing the sample molecule. The scheme for obtaining the sample skeleton latent vector and the sample side chain latent vector corresponding to the sample molecule is similar to the scheme for obtaining the skeleton latent vector and the side chain latent vector corresponding to the reference drug molecule, and will not be described again.

[0150] In an embodiment of the present application, after the cross-entropy loss and the prediction loss of the machine learning model are calculated, the loss function of the machine learning model can be generated according to the cross-entropy loss and the prediction loss of the machine learning model, and then the parameters of the machine learning model are adjusted based on the loss function. Alternatively, the loss function L of the machine learning model can be generated by the following formula 12:

[0151] L = L recon + βL cls Formula 15

[0152] In the above formula 15, L recon represents the prediction loss of the machine learning model; β represents a hyperparameter for adjusting the weight between the two losses.

[0153] The purpose of training the machine learning model is to minimize the above loss function L, wherein the cross-entropy loss L cls is set to ensure that each skeleton latent vector determined by the machine learning model is near the center of the skeleton cluster to which it belongs after the machine learning model is trained; and the prediction loss L recon is set to ensure that the machine learning model can find a better target skeleton latent vector, and thus ensure that a qualified drug molecule is obtained after the machine learning model is trained.

[0154] After training the machine learning model, the reference drug molecule can be processed to obtain a transitioned drug molecule based on the machine learning model. In order to facilitate the understanding of the technical solutions of the embodiments of the present application, the following will be combined with FIG. 3 to Figure 10The implementation details of the technical solution of the embodiment of the present application are described in detail:

[0155] like Figure 4A As shown, when processing molecular skeleton transitions through a machine learning model, the model structure can include the following parts: a graph encoder, a node masking and graph readout function, a Gaussian mixture distribution (GM) fitting process, and a decoder. Among them, the graph encoder is mainly used to generate the atomic latent vectors corresponding to the drug molecules; the atom masking and readout function part is mainly used to obtain the skeleton latent vectors and side chain latent vectors through atom masking processing; the Gaussian mixture distribution fitting process is used to realize the Gaussian mixture distribution of the skeleton latent vector to realize the processing of different transition degrees; the decoder is used to output the drug molecules obtained after the transition processing. The following is a detailed explanation of these parts:

[0156] In one embodiment of the present application, the graph encoder is primarily composed of a Directed Message Passing Neural Network (D-MPNN), a graph convolutional neural network. Graph convolutional neural networks operate directly on graph structures, including chemical structures. Unlike fingerprint representations, which assign a single fixed-length feature vector to a molecule, graph representations assign feature vectors to each bond and atom in a chemical structure.

[0157] In short, D-MPNN can be understood as a multi-step neural network, where each step is essentially a feedforward neural network that generates a set of hidden representations that serve as input to the next step. The core of D-MPNN is the message passing step, which exploits the local substructure of the molecular graph to update the hidden vector. After the message passing step, the hidden vectors from all edges are aggregated into a single fixed-length hidden vector, which is fed into the feedforward neural network to produce a prediction. Figure 4B As shown, each bond is represented by a pair of directed edges, Figure 4B In (a), the messages from the orange directional keys (i.e., 3→2 and 4→2 in (a)) are used to notify the hidden state update of the red directional key (i.e., 2→1 in (a)). In (b), the messages from the green directional key (i.e., 5→1 in (b)) are used to notify the hidden state update of the purple directional key (i.e., 1→2 in (b)). Figure 4B (c) shown shows the update function of the hidden representation of the red directed key in (a) (i.e., 2→1 in (a)), which is an iterative process that can be repeated multiple times (usually 5 times). Figure 4B The concat strategy shown in

[15] is a strategy in deep learning that can efficiently handle input samples of variable sizes.

[0158] Before inputting a drug molecule into the graph encoder, it can be converted into a connected graph structure with corresponding chemical bonds and atomic properties at its edges and nodes. The connected graph structure corresponding to such a drug molecule can be represented as G = (A, X, E), where A represents the connectivity matrix, X represents the node features, and E represents the edge features. In this connected graph structure, nodes represent atoms in the drug molecule. Node features represent atomic characteristics within the drug molecule, such as atomic weight, atomic charge, atomic type, atomic valence, whether the atom is in a ring, or whether it is an atom within an aromatic ring. Edge features represent the characteristics between atoms within the drug molecule, such as whether the edge has a single or double bond, whether the edge is in a ring, or whether it is within an aromatic ring. Based on this, the connected graph structure is input into the D-MPNN for processing, which can be specifically represented by Equations 1 through 5. Finally, the latent vectors for each node in the connected graph structure, i.e., the latent vectors for each atom in the drug molecule, are obtained. Furthermore, the atomic latent vectors corresponding to the drug molecules can be represented by a matrix, that is, the latent vectors of each node are arranged in a row-column manner in a matrix (such as an H matrix) to represent the atomic latent vectors corresponding to the drug molecules.

[0159] In one embodiment of the present application, Figure 5 As shown in the figure, the atomic occlusion and readout function is mainly used to obtain the latent vector representation of the skeleton and side chain, i.e., the skeleton latent vector and the side chain latent vector, by occluding and reading out the atoms after obtaining the latent vector representation of all atoms (i.e., the atomic latent vector corresponding to the drug molecule).

[0160] Optionally, the atomic occlusion process can be performed using a bit vector, the length of which is the same as the number of atoms contained in the drug molecule. Specifically, the bit vector can be expressed by the above formula 6.

[0161] In one embodiment of the present application, the readout function is to obtain the skeleton latent vector and the side chain latent vector, and a selective self-attention mechanism is used in the embodiment of the present application. Assume that the atomic latent vector of a drug molecule is H node , then the skeleton latent vector and the side chain latent vector can be calculated by the above formula 7 and formula 8 respectively.

[0162] In one embodiment of the present application, the Gaussian mixture distribution fitting process is mainly to realize the Gaussian mixture distribution of the skeleton latent vector to achieve the processing of different transition degrees. For the side chain, either Gaussian distribution fitting can be performed or no distribution assumption can be made. In an embodiment of the present application, in order to better keep the side chain unchanged, the side chain latent space can be processed using the AutoEncoder method without making the Gaussian distribution assumption.

[0163] In one embodiment of the present application, the skeletons of existing molecules can be pre-divided into M different skeleton clusters by a skeleton clustering algorithm. The existing molecules can be sample molecules used to train the machine learning model or molecules selected from a molecule library, and these molecules are not limited to drug molecules. In the skeleton latent space, it is expected that points of the same skeleton cluster can be close to each other and points of different skeleton clusters can be far away from each other. Therefore, the cluster centers of M latent spaces can be set: (μ m ,σ m ), where μ m represents the center of the cluster, σ m Represents the standard deviation. Then, the skeleton latent vector Z corresponding to the i-th drug molecule can be calculated by the above formula 9. sca,i The distance d from the cluster center of the mth skeleton cluster i At the same time, the distance d can be obtained by the above formula 13 i Add deflection, and the distance after adding deflection can be expressed as follows Figure 6 After calculating the distance, a distance-based cross entropy loss L can be calculated using the above formula 14. cls .

[0164] In one embodiment of the present application, the decoder can use a SMILES decoder, that is, the representation of the hidden layer is decoded into SMILES instead of a graph. SMILES can be understood as a spanning tree of a graph that is expanded according to a rule. Each drug molecule can have a corresponding canonical SMILES, so it is reasonable to use SMILES in the decoder. Figure 7 As shown in Figure 2, the decoder can follow the teacher forcing mode. The working principle of the teacher forcing mode is to use the real data (ground truth) of the training dataset as the input x(t+1) of the next moment at time t in the training process, instead of using the output of the model at the previous moment. Figure 7 701 is the input part of the real data, and 702 is the output part of the model.

[0165] In one embodiment of the present application, the final output of the encoder will be subjected to a reconstruction loss (i.e., prediction loss) with the correct answer (i.e., real data) to obtain L recon The loss function of the model consists of reconstruction loss and cross entropy loss, which can be specifically referred to in the above formula 15.

[0166] After training the model, it can be used to generate molecules. During the molecular generation process, a molecular input is required as a reference drug molecule. The reference drug molecule is the drug molecule that needs to undergo backbone replacement. At the same time, the backbone that needs to be replaced in the reference drug molecule can also be marked. After the structure of the reference drug molecule is converted into a connection graph structure and input into the model, the model can obtain the backbone latent vector and side chain latent vector corresponding to the reference drug molecule. Figure 8 As shown, the processing process is similar to the relevant content introduced in the above embodiment and will not be repeated here.

[0167] In one embodiment of the present application, after obtaining the skeleton latent vector, the process of molecule generation is slightly different from the process of model training. Specifically, in the process of molecule generation, when obtaining the target skeleton latent vector, resampling processing is required. The specific decoding process is as follows: Figure 9 As shown in Figure 2, the sidechain latent vector remains unchanged and is not sampled. Due to model training, the skeleton latent vector exhibits a Gaussian mixture distribution, which facilitates skeleton transition processing.

[0168] In one embodiment of the present application, according to the degree of hopping, the hopping modes can be divided into the following three types: scaffold crawling, scaffold hopping, and scaffold leaping. Figure 10 As shown, Scaffoldcrawling is the mildest transition, and the molecular changes after the transition are the least, mainly because the skeleton latent vector is sampled from the same skeleton cluster as the reference drug molecule ( Figure 10 The target skeleton latent vector corresponding to the new sampling point (i.e., the newly generated skeleton latent vector) can be expressed by the aforementioned formula 10.

[0169] Scaffold hopping is a large jump, and the molecular skeleton changes after the jump will be large. Specifically, the skeleton latent vector is sampled from the adjacent skeleton cluster of the reference drug molecule ( Figure 10 The target skeleton latent vector corresponding to the new sampling point (i.e., the newly generated skeleton latent vector) can be expressed by the aforementioned formula 11.

[0170] Scaffold leaping is the maximum leap, and the molecular skeleton changes after the leap will be very large. Specifically, the skeleton latent vector is sampled from a cluster far away from the skeleton cluster of the reference drug molecule ( Figure 10 The target skeleton latent vector corresponding to the new sampling point (i.e., the newly generated skeleton latent vector) can be expressed by the aforementioned formula 12.

[0171] Continue to refer to Figure 9 As shown in Figure 1, in addition to obtaining the side chain latent vectors and target backbone latent vectors, the model also needs to input an activity condition, such as the target of a reference drug molecule and the corresponding expected pIC50 value. Once these three vectors are obtained, the model can generate new transition drug molecules through the SMILES decoder.

[0172] In one embodiment of the present application, the drug molecules generated after the skeleton transition can be screened through two steps: the first step is the molecular filtering based on physical and chemical properties, the purpose of which is to ensure that the molecules subsequently evaluated are drug-like, for example, they can be screened by the Lipinsiki five rules, etc.; the second step is to prepare ligands for the drug-like molecules that meet the physical and chemical property requirements and enter the subsequent molecular docking step, the purpose of which is to select drug-like molecules that have a strong ability to bind to the target.

[0173] Specifically, the crystal structure of molecular docking can be searched from the PDB (protein data bank, protein data file) database. For example, the co-crystal structure of the reference drug molecule or its homologous compound can be selected. At the same time, it is necessary to ensure that the resolution is high and the protein structure near the binding pocket is intact. When docking, the protein is first prepared using molecular docking software, and then the molecule is docked back to the prepared crystal structure. The accuracy of the configuration is judged by the binding pattern. At the same time, the molecular binding pattern in the co-crystal structure is also used as a template for molecular docking to analyze whether the binding pattern of the AI-generated molecule is reasonable. The technical solution of this embodiment can remove molecules with obviously unreasonable configurations through virtual screening, and then perform molecular docking with higher precision on all the configurations retained in the previous step, and then use the 3D-CNN (Convolutional Neural Network, convolutional neural network) method to re-score the obtained binding pattern, select molecules with a 3D-CNN score of at least >0.8 (the value is only an example) and the binding pattern of the key action site is not lost, and perform compound synthesis and verification processing.

[0174] It should be noted that in the aforementioned embodiment, the graph encoder can also use a Dual-MPNN, and the SMILES decoder can be replaced with various natural language processing decoders, such as grammar-VAE (Variational Autoencoder), SD-VAE (Syntax Directed-VAE), and the decoding part of the Transformer.

[0175] The technical solutions of the above embodiments of the present application enable the generated target skeleton latent vector to be mapped to the spatial distribution to ensure that it can break away from the design thinking pattern of pharmaceutical experts, thereby achieving great novelty. At the same time, it can be automatically executed by the equipment, reducing manpower and time costs.

[0176] The following describes an embodiment of the device of the present application, which can be used to perform the method for processing molecular skeleton transitions in the above-mentioned embodiment of the present application. For details not disclosed in the embodiment of the device of the present application, please refer to the embodiment of the method for processing molecular skeleton transitions in the above-mentioned embodiment of the present application.

[0177] Figure 11 A block diagram of a molecular skeleton transition processing device according to an embodiment of the present application is shown. The molecular skeleton transition processing device can be set in a device with a computing processing function, such as Figure 1 Within the server 130 shown in .

[0178] Reference Figure 11 As shown, a molecular skeleton transition processing device 1100 according to an embodiment of the present application includes: a first generating unit 1102 , a first processing unit 1104 , a second generating unit 1106 and a third generating unit 1108 .

[0179] Among them, the first generation unit 1102 is configured to generate the atomic latent vector corresponding to the reference drug molecule based on the connection graph structure corresponding to the reference drug molecule; the first processing unit 1104 is configured to perform atomic occlusion processing on the atomic latent vector to obtain the skeleton latent vector and the side chain latent vector contained in the atomic latent vector; the second generation unit 1106 is configured to generate a target skeleton latent vector with a specified transition degree between the skeleton latent vector and the skeleton latent vector based on the spatial distribution of the skeleton latent vector; the third generation unit 1108 is configured to generate a drug molecule after the transition based on the target skeleton latent vector and the side chain latent vector.

[0180] In some embodiments of the present application, based on the foregoing scheme, the nodes in the connection graph structure represent atoms in the reference drug molecule; the first generation unit 1102 is configured to: calculate node information of each node in the connection graph structure by a graph encoder based on node features and edge features contained in the connection graph structure, the node features representing atom features in the reference drug molecule, and the edge features representing features between atoms in the reference drug molecule; generate a hidden vector of each node according to the node information of each node and the node features of each node; and generate an atom hidden vector corresponding to the reference drug molecule according to the hidden vector of each node and the atoms contained in the reference drug molecule.

[0181] In some embodiments of the present application, based on the foregoing scheme, the first generation unit 1102 is configured to: calculate information between a first node and a second node in a next hidden layer according to node features of the first node, node features of the second node in the connection graph structure, and edge information between other nodes of the first node except the second node and the first node in a previous hidden layer, the first node being any node in the connection graph structure and the second node being a neighbor node of the first node in the connection graph structure; calculate edge information between the first node and the second node in the next hidden layer according to the edge information between the first node and the second node in the previous hidden layer and the information between the first node and the second node in the next hidden layer; wherein the edge information between two nodes in the connection graph structure in an initial hidden layer is calculated according to node features of one of the two nodes and edge features between the two nodes; and summing the edge information corresponding to each node in all hidden layers to obtain node information of the node.

[0182] In some embodiments of the present application, based on the foregoing scheme, the first processing unit 1104 is configured to: obtain a bit vector corresponding to the reference drug molecule, the length of the bit vector being the same as the number of atoms contained in the reference drug molecule, and a bit value corresponding to a skeleton atom in the bit vector being a first value; select a hidden vector of a skeleton atom and a hidden vector of a side chain atom from the atom hidden vectors corresponding to the reference drug molecule based on the bit vector; generate the skeleton hidden vector by a multi-head attention mechanism and the hidden vector of the skeleton atom, and generate the side chain hidden vector according to the multi-head attention mechanism and the hidden vector of the side chain atom.

[0183] In some embodiments of the present application, based on the aforementioned scheme, the first processing unit 1104 is further configured to: obtain multiple skeleton clusters, the cluster centers of each skeleton cluster in the multiple skeleton clusters conform to the Gaussian mixture distribution; calculate the first distance between the skeleton latent vector and the cluster centers of each skeleton cluster, and select the target skeleton cluster to which the skeleton of the reference drug molecule belongs based on the first distance; determine the Gaussian mixture distribution to which the skeleton latent vector belongs based on the cluster center of the target skeleton cluster.

[0184] In some embodiments of the present application, based on the aforementioned scheme, the second generation unit 1106 is configured to: perform random sampling processing according to the specified transition degree and the target skeleton cluster to obtain an offset corresponding to the specified transition degree; add the skeleton latent vector and the offset corresponding to the specified transition degree to generate the target skeleton latent vector.

[0185] In some embodiments of the present application, based on the aforementioned scheme, the second generation unit 1106 is configured as: if the specified transition degree is the first transition degree, then a first offset is generated according to the product of the variance of the target skeleton cluster and the first vector obtained by random sampling, and the first offset is used as the offset corresponding to the first transition degree.

[0186] In some embodiments of the present application, based on the aforementioned scheme, the second generation unit 1106 is configured as: if the specified transition degree is the second transition degree, then selecting a first skeleton cluster from the multiple skeleton clusters whose cluster center is less than or equal to a first set value with respect to the target skeleton cluster; generating a second offset based on the product of the variance of the first skeleton cluster and the second vector obtained by random sampling, the cluster center of the target skeleton cluster and the cluster center of the first skeleton cluster, and using the second offset as the offset corresponding to the second transition degree.

[0187] In some embodiments of the present application, based on the above-mentioned scheme, the second generation unit 1106 is configured as: if the specified transition degree is the third transition degree, then select a second skeleton cluster from the multiple skeleton clusters whose cluster center is greater than or equal to a second set value with the target skeleton cluster; generate a third offset based on the product of the variance of the second skeleton cluster and the third vector obtained by random sampling, the cluster center of the target skeleton cluster and the cluster center of the second skeleton cluster, and use the third offset as the offset corresponding to the third transition degree.

[0188] In some embodiments of the present application, based on the aforementioned scheme, the third generation unit 1108 is configured to: obtain the specified target and target activity value of the reference drug molecule; and generate the drug molecule after the transition based on the target skeleton latent vector, the side chain latent vector, the target of the reference drug molecule and the target activity value.

[0189] In some embodiments of the present application, based on the foregoing scheme, the device 1100 further comprises a second processing unit configured to: after generating the transitioned drug molecules, performing a molecular filtering process on the transitioned drug molecules according to the physical and chemical properties to obtain drug-like drug molecules; obtaining a co-crystal structure corresponding to the reference drug molecules, and docking the drug-like drug molecules into the co-crystal structure; removing drug molecules that do not match the co-crystal structure through the binding mode of the drug-like drug molecules and the co-crystal structure to obtain screened drug molecules; and performing synthesis and verification of the compounds according to the docking of the screened drug molecules and the co-crystal structure.

[0190] In some embodiments of the present application, based on the foregoing scheme, the transitioned drug molecules are generated by a machine learning model; the device 1100 further comprises a third processing unit configured to: obtain a sample skeleton hidden vector corresponding to a sample molecule, and obtain a plurality of skeleton clusters, the cluster centers of each skeleton cluster in the plurality of skeleton clusters conforming to a Gaussian mixture distribution; calculate a second distance between the sample skeleton hidden vector of the sample molecule and the cluster center of each skeleton cluster, and select a skeleton cluster to which a sample skeleton of the sample molecule belongs according to the second distance; generate a distance-based cross-entropy loss according to the distance between the sample skeleton hidden vector and the cluster center of the sample skeleton belonging to the skeleton cluster; generate a loss function of the machine learning model according to the cross-entropy loss and a prediction loss of the sample molecule by the machine learning model; and adjust parameters of the machine learning model based on the loss function.

[0191] In some embodiments of the present application, based on the foregoing scheme, the machine learning model comprises a decoder; and the third processing unit is further configured to: after obtaining the sample skeleton hidden vector and the sample side chain hidden vector corresponding to the sample molecule by the machine learning model, input the sample skeleton hidden vector, the sample side chain hidden vector, and a target molecule corresponding to the sample molecule into the decoder; and calculate the prediction loss according to an output of the decoder and the target molecule.

[0192] Figure 12 A structural schematic diagram of a computer system of an electronic device suitable for implementing embodiments of the present application is shown.

[0193] It should be noted that, Figure 12 The computer system 1200 of the electronic device shown is only an example and should not impose any limitation on the functions and use range of embodiments of the present application.

[0194] As Figure 12As shown, the computer system 1200 includes a central processing unit (CPU) 1201, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 1202 or the program loaded from the storage part 1208 to the random access memory (RAM) 1203, such as executing the method described in the above embodiment. Various programs and data required for system operation are also stored in the RAM 1203. The CPU 1201, ROM 1202 and RAM 1203 are connected to each other via a bus 1204. An input / output (I / O) interface 1205 is also connected to the bus 1204.

[0195] The following components are connected to the I / O interface 1205: an input section 1206 including a keyboard, a mouse, and the like; an output section 1207 including devices such as a cathode ray tube (CRT), a liquid crystal display (LCD), and a speaker; a storage section 1208 including a hard disk; and a communication section 1209 including a network interface card such as a LAN (Local Area Network) card or a modem. The communication section 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to the I / O interface 1205 as needed. Removable media 1211, such as a magnetic disk, an optical disk, a magneto-optical disk, or a semiconductor memory, is installed in the drive 1210 as needed, so that computer programs read from the removable media can be installed in the storage section 1208 as needed.

[0196] In particular, according to an embodiment of the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present application includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 1209, and / or installed from a removable medium 1211. When the computer program is executed by the central processing unit (CPU) 1201, the various functions defined in the system of the present application are executed.

[0197] It should be noted that the computer-readable medium in the embodiments of the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination thereof. The computer-readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination thereof. More specific examples of the computer-readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disk read-only memory (Compact Disc Read-Only Memory, CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or apparatus. In the present application, the computer-readable signal medium can include a data signal carrying computer-readable computer programs in a baseband or as a part of a carrier wave. Such a propagated data signal can take various forms, including but not limited to an electromagnetic signal, an optical signal, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than the computer-readable storage medium, which can send, propagate or transmit programs for use by or in conjunction with an instruction execution system, device or apparatus. The computer programs contained in the computer-readable medium can be transmitted by any suitable medium, including but not limited to wireless, wired, or the like, or any suitable combination thereof.

[0198] The flowcharts and block diagrams in the drawings illustrate the possible implementation architectures, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In the flowcharts or block diagrams, each block can represent a module, a program segment or a part of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different orders than that shown in the drawings. For example, two blocks represented in succession can actually be executed substantially in parallel, and sometimes in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams or flowcharts, and the combination of blocks in the block diagrams or flowcharts, can be implemented by a dedicated hardware-based system for performing the specified functions or operations, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0199] The units involved in the embodiments described in this application may be implemented by software or hardware, and the units described may also be set in a processor. In some cases, the names of these units do not constitute limitations on the units themselves.

[0200] As another aspect, the present application further provides a computer-readable medium, which may be included in the electronic device described in the above embodiments, or may exist independently without being incorporated into the electronic device. The computer-readable medium carries one or more programs, and when the one or more programs are executed by the electronic device, the electronic device implements the method described in the above embodiments.

[0201] It should be noted that, although several modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiment of the application, the features and functions of two or more modules or units described above can be concretized in one module or unit. On the contrary, the features and functions of one module or unit described above can be further divided into multiple modules or units to be concretized.

[0202] Through the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solution according to the embodiments of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, and includes several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the embodiments of the present application.

[0203] Those skilled in the art will readily conceive of other embodiments of the present application after considering the specification and practicing the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present application that follow the general principles of this application and include common knowledge or customary techniques in the art that are not disclosed herein.

[0204] It should be understood that the present application is not limited to the exact structures described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A method for processing molecular skeleton transition, characterized in that: include: generating an atomic latent vector corresponding to the reference drug molecule according to the connection graph structure corresponding to the reference drug molecule; Performing atomic occlusion processing on the atomic latent vector to obtain a skeleton latent vector and a side chain latent vector contained in the atomic latent vector; generating a target skeleton latent vector having a specified transition degree with respect to the skeleton latent vector according to the spatial distribution of the skeleton latent vector; A drug molecule after transition is generated according to the target skeleton latent vector and the side chain latent vector.

2. The method for processing molecular skeleton transition according to claim 1, characterized in that: The nodes in the connected graph structure represent atoms in the reference drug molecule; Generating the atomic latent vector corresponding to the reference drug molecule according to the connection graph structure corresponding to the reference drug molecule includes: Based on the node features and edge features contained in the connection graph structure, calculating node information of each node in the connection graph structure by a graph encoder, the node features representing atomic features in the reference drug molecule, and the edge features representing features between atoms in the reference drug molecule; Generating a latent vector of each node according to the node information of each node and the node features of each node; According to the latent vectors of the nodes and the atoms contained in the reference drug molecule, an atomic latent vector corresponding to the reference drug molecule is generated.

3. The method for processing molecular skeleton transition according to claim 2, characterized in that: The step of calculating node information of each node in the connection graph structure by a graph encoder based on the node features and edge features contained in the connection graph structure includes: Calculating information between the first node and the second node in a subsequent hidden layer based on a node feature of a first node in the connection graph structure, a node feature of a second node in the connection graph structure, and edge information between the first node and other nodes among neighboring nodes of the first node except the second node in a previous hidden layer, wherein the first node is any node in the connection graph structure, and the second node is a neighboring node of the first node in the connection graph structure; Calculating the edge information between the first node and the second node in the subsequent hidden layer based on the edge information between the first node and the second node in the previous hidden layer and the information between the first node and the second node in the subsequent hidden layer; wherein the edge information between the two nodes in the connection graph structure in the initial hidden layer is calculated based on the node feature of one of the two nodes and the edge feature between the two nodes; The side information corresponding to each node in all hidden layers is summed to obtain the node information of each node.

4. The method for processing molecular skeleton transition according to claim 1, characterized in that: The performing atomic occlusion processing on the atomic latent vector to obtain the skeleton latent vector and the side chain latent vector contained in the atomic latent vector includes: Obtaining a bit vector corresponding to the reference drug molecule, where the length of the bit vector is the same as the number of atoms contained in the reference drug molecule, and the bit value corresponding to the backbone atom in the bit vector is a first value; Selecting implicit vectors of backbone atoms and implicit vectors of side chain atoms from implicit vectors of atoms corresponding to the reference drug molecule based on the bit vector; The skeleton latent vector is generated by a multi-head attention mechanism and the latent vector of the skeleton atom, and the side chain latent vector is generated according to the multi-head attention mechanism and the latent vector of the side chain atom.

5. The method for processing molecular skeleton transition according to claim 1, characterized in that: Before generating a target skeleton latent vector having a specified transition degree with the skeleton latent vector according to the spatial distribution of the skeleton latent vector, the method further includes: Acquire multiple skeleton clusters, wherein the cluster center of each skeleton cluster in the multiple skeleton clusters conforms to a Gaussian mixture distribution; calculating a first distance between the skeleton latent vector and the cluster center of each skeleton cluster, and selecting a target skeleton cluster to which the skeleton of the reference drug molecule belongs based on the first distance; The Gaussian mixture distribution to which the skeleton latent vector belongs is determined according to the cluster center of the target skeleton cluster.

6. The method for processing molecular skeleton transition according to claim 5, characterized in that: Generating a target skeleton latent vector having a specified transition degree with the skeleton latent vector according to the spatial distribution of the skeleton latent vector includes: Performing random sampling processing according to the specified transition degree and the target skeleton cluster to obtain an offset corresponding to the specified transition degree; The skeleton latent vector is added to the offset corresponding to the specified transition degree to generate the target skeleton latent vector.

7. The method for processing molecular skeleton transition according to claim 6, characterized in that: The performing random sampling processing according to the specified transition degree and the target skeleton cluster to obtain an offset corresponding to the specified transition degree includes: If the designated transition degree is the first transition degree, a first offset is generated according to the product of the variance of the target skeleton cluster and the first vector obtained by random sampling, and the first offset is used as the offset corresponding to the first transition degree.

8. The method for processing molecular skeleton transition according to claim 6, characterized in that: The performing random sampling processing according to the specified transition degree and the target skeleton cluster to obtain an offset corresponding to the specified transition degree includes: If the designated transition degree is a second transition degree, selecting a first skeleton cluster from the plurality of skeleton clusters whose cluster center with the target skeleton cluster is less than or equal to a first set value; A second offset is generated according to the product of the variance of the first skeleton cluster and the second vector obtained by random sampling, the cluster center of the target skeleton cluster and the cluster center of the first skeleton cluster, and the second offset is used as the offset corresponding to the second transition degree.

9. The method for processing molecular skeleton transition according to claim 6, characterized in that: The performing random sampling processing according to the specified transition degree and the target skeleton cluster to obtain an offset corresponding to the specified transition degree includes: If the designated transition degree is a third transition degree, selecting a second skeleton cluster from the plurality of skeleton clusters, the second skeleton cluster having a cluster center greater than or equal to a second set value with the target skeleton cluster; A third offset is generated based on the product of the variance of the second skeleton cluster and the third vector obtained by random sampling, the cluster center of the target skeleton cluster and the cluster center of the second skeleton cluster, and the third offset is used as the offset corresponding to the third transition degree.

10. The method for processing molecular skeleton transition according to claim 1, characterized in that: The step of generating a transitioned drug molecule according to the target skeleton latent vector and the side chain latent vector comprises: Obtaining the target and target activity value of the specified reference drug molecule; A drug molecule after transition is generated according to the target skeleton latent vector, the side chain latent vector, the target of the reference drug molecule and the target activity value.

11. The method for processing molecular skeleton transition according to claim 1, characterized in that: The method further comprises: After generating the transitioned drug molecules, molecular filtering is performed based on the physical and chemical properties of the transitioned drug molecules to obtain drug-like drug molecules; Obtaining a co-crystal structure corresponding to the reference drug molecule, and docking the drug-like drug molecule into the co-crystal structure; Based on the binding mode between the drug-like drug molecules and the co-crystal structure, drug molecules that do not match the co-crystal structure are removed to obtain screened drug molecules; Based on the docking of the screened drug molecules and the co-crystal structure, the compound is synthesized and verified.

12. The method for processing molecular skeleton transition according to any one of claims 1 to 11, characterized in that: Generating the drug molecule after the transition by a machine learning model; the method further comprises: Obtaining a sample skeleton latent vector corresponding to the sample molecule, and obtaining a plurality of skeleton clusters, wherein the cluster center of each skeleton cluster in the plurality of skeleton clusters conforms to a Gaussian mixture distribution; Calculating a second distance between the sample skeleton latent vector of the sample molecule and the cluster center of each skeleton cluster, and selecting the skeleton cluster to which the sample skeleton of the sample molecule belongs according to the second distance; Generate a distance-based cross entropy loss according to the distance between the sample skeleton latent vector and the cluster center of the skeleton cluster to which the sample skeleton belongs; generating a loss function of the machine learning model according to the cross entropy loss and the prediction loss of the machine learning model for the sample molecule; Parameters of the machine learning model are adjusted based on the loss function.

13. The method for processing molecular skeleton transition according to claim 12, characterized in that: The machine learning model includes a decoder; the method further includes: After obtaining a sample backbone latent vector and a sample side chain latent vector corresponding to the sample molecule through the machine learning model, inputting the sample backbone latent vector, the sample side chain latent vector, and the target molecule corresponding to the sample molecule into the decoder; The prediction loss is calculated based on the decoder output and the target molecule.

14. A molecular skeleton transition processing device, characterized in that: include: A first generating unit is configured to generate an atomic latent vector corresponding to the reference drug molecule according to a connection graph structure corresponding to the reference drug molecule; a first processing unit configured to perform atomic occlusion processing on the atomic latent vector to obtain a skeleton latent vector and a side chain latent vector contained in the atomic latent vector; a second generating unit configured to generate a target skeleton latent vector having a specified transition degree with the skeleton latent vector according to the spatial distribution of the skeleton latent vector; The third generating unit is configured to generate a drug molecule after the transition according to the target skeleton latent vector and the side chain latent vector.

15. A computer-readable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for processing molecular skeleton transition according to any one of claims 1 to 13 is implemented.

16. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement the method for processing molecular skeleton transitions as described in any one of claims 1 to 13.

Images