Drug molecule discovery method, device, medium and equipment

Abstract

The embodiment of the invention discloses a drug molecule discovery method and device, a medium and equipment, and the method comprises the steps: extracting an entity of an input text, recognizing the intention of the input text, and scheduling one or more processes in drug molecule discovery processes related to the entity according to an intention recognition result, so that a user only needs to give the input text, and the user experience is improved. The subsequent operation of the drug molecule discovery process can automatically schedule one or more processes in the drug molecule discovery processes related to the entity in the input text according to the intention result of the input text, so that a user is prevented from manually importing and exporting data between different processes, and the processing efficiency of drug molecule discovery is improved.

Description

Technical Field

[0001] This application relates to the medical field, specifically to a method, apparatus, medium, and device for drug molecule discovery. Background Technology

[0002] Currently, with the rapid development of artificial intelligence technology, its application in drug discovery is becoming increasingly profound, evolving from a tool to assist in single tasks to an end-to-end automated workflow. Current drug molecule discovery workflows typically involve multiple stages, including target identification (TI), hit identification (HI), hit-to-lead (H2L) optimization, and lead optimization (LO). Executing these stages requires complex biochemical computing environments, and a workflow is often fragmented into multiple offline steps requiring manual intervention. These multi-step operations necessitate researchers manually importing and exporting data between different workflows, resulting in low processing efficiency in drug molecule discovery. Summary of the Invention

[0003] This application provides a method, apparatus, medium, and device for drug molecule discovery, which can automatically execute at least one step in the drug molecule discovery process, avoiding manual import and export of data between different steps and improving the processing efficiency of drug molecule discovery.

[0004] To address the aforementioned technical problems, the embodiments of this application provide the following technical solutions: A method for drug molecule discovery, characterized by comprising: Extract entities from the input text and perform intent recognition; based on the results of the intent recognition, schedule the drug molecule discovery process related to the entities. The drug molecule discovery process includes one or more of the following processes: Target identification: Extract disease entity names from input text, and determine the target targets corresponding to the disease entity names and the target structure of the target targets; Lead compound discovery: Binding pocket detection is performed on the target structure to determine the ligand binding pocket in the target structure and to obtain lead compounds that match the ligand binding pocket; From lead compound to lead compound selection: The lead compound is derivatized and expanded to obtain candidate derivatized compounds, and a lead compound is selected from each candidate derivatized compound based on the comprehensive score of each candidate derivatized compound. Lead compound optimization: The trained molecular generation model is used to generate molecules from the lead compound, resulting in multiple candidate molecules. The comprehensive optimization score and docking score of each candidate molecule are obtained, and a drug discovery molecule ranking list and rationale are generated.

[0005] In one feasible implementation, the step of scheduling the drug molecule discovery process associated with the entity based on the result of the intent recognition includes: When the result of the intent recognition is to run a single-process task, the target process related to the entity is scheduled and run from the target identification, lead compound discovery, lead compound screening to lead compound optimization and lead compound optimization in the drug molecule discovery process. When the result of the intent recognition is to run an end-to-end task, at least two processes in the drug molecule discovery process are run sequentially: target identification, lead compound discovery, lead compound screening to lead compound selection, and lead compound optimization.

[0006] In one feasible implementation, the process of sequentially running at least two steps in the drug molecule discovery pipeline—target identification, lead compound discovery, lead compound screening to lead compound selection, and lead compound optimization—includes: When the target node of the current process is detected, the first execution result of the previous running node of the target node is obtained; The first execution result is standardized and verified to output a prompt on whether to continue executing subsequent nodes based on the first execution result; When a continue execution instruction based on the first execution result is received, the subsequent nodes of the target node continue to run using the first execution result as input data. In one feasible embodiment, the drug molecule discovery method further includes: When a parameter correction instruction based on the first execution result is received, the first parameter carried by the parameter correction instruction is obtained; The first parameter is input to the previous running node of the target node, so that the previous running node of the target node re-runs according to the first parameter to obtain the second execution result; The second execution result is standardized and verified, and a prompt is output indicating whether to continue executing subsequent nodes based on the second execution result; If no continue execution instruction based on the second execution result is received, the execution of subsequent nodes of the target node is prohibited.

[0007] In one feasible embodiment, the drug molecule discovery method further includes: When a manual editing instruction based on the first execution result is received, the second parameter carried by the manual editing instruction is obtained; The second parameter is used as input data to continue running subsequent nodes of the target node.

[0008] In one feasible embodiment, the drug molecule discovery method further includes: When the current process is detected to have completed, the second execution result of the current process is obtained; Display the second execution result and output a prompt asking whether to continue the subsequent process based on the second execution result; When a continue execution instruction based on the second execution result is received, the subsequent processes of the current process continue to run.

[0009] In one feasible implementation, the steps of generating molecules from a lead compound using a trained molecular generation model to obtain multiple candidate molecules, acquiring a comprehensive optimization score and docking score for each candidate molecule, and generating a drug discovery molecule ranking list and rationale include: Based on the reward score, the trained molecular generation model performs reinforcement learning on each of the lead compounds, and selects multiple candidate molecules from the newly generated optimized molecules. The trained molecular ranking model is used to reason about the candidate molecules based on the comprehensive optimization score and docking score to obtain the ranking sequence and reasoning of drug discovery molecules, and each drug discovery molecule and reasoning is displayed in sequence according to the ranking sequence.

[0010] In one feasible implementation, the calculation process for the reward score includes: The affinity of each optimized molecule is evaluated using a trained affinity assessment model to obtain an affinity score for each optimized molecule. Obtain a quantitative drug-likeness score and a synthesis accessibility score for each of the optimized molecules; Calculate the structural similarity between each optimized molecule and the lead compound to obtain the structural similarity of each optimized molecule; A reward score for each optimized molecule is obtained by combining its affinity score, quantitative drug-likeness score, synthetic accessibility score, and structural similarity score.

[0011] In one feasible implementation, the step of fusing the affinity score, quantitative drug-likeness score, synthetic accessibility score, and structural similarity of each optimized molecule to obtain a reward score for each optimized molecule includes: The difference between the maximum synthetic accessibility score and the synthetic accessibility score of each candidate molecule is determined to obtain the synthetic inaccessibility score of each candidate molecule. The affinity score, quantitative drug-likeness score, synthetic inaccessibility score, and structural similarity score of each candidate molecule are weighted and summed to obtain the reward score for each candidate molecule.

[0012] In one feasible implementation, before selecting a lead compound from each of the candidate derivatized compounds based on a comprehensive score, the method further includes: The affinity of each candidate derivative compound is evaluated using a trained affinity assessment model to obtain an affinity score for each candidate derivative compound. Quantitative drug-likeness scores and synthetic accessibility scores were obtained for each of the candidate derivative compounds; The toxicity risk of each candidate derivative compound is assessed using a trained toxicity risk assessment model to obtain a toxicity risk score for each candidate derivative compound. A comprehensive score for each candidate derivative compound is obtained by combining its affinity score, quantitative drug-likeness score, synthetic accessibility score, and toxicity risk score.

[0013] In one feasible implementation, the step of fusing the affinity score, quantitative drug-likeness score, synthetic accessibility score, and toxicity risk score of each candidate derivative compound to obtain a comprehensive score for each candidate derivative compound includes: For each of the candidate derivative compounds, the corresponding affinity score, quantitative drug-likeness score, and synthetic accessibility score are weighted and summed to obtain the first calculation result for each candidate derivative compound; The difference between the first calculated result and the corresponding toxicity risk score for each candidate derivative compound is determined to obtain a comprehensive score for each candidate derivative compound.

[0014] In one feasible implementation, the step of determining the target point corresponding to the disease entity name includes: The disease name and disease entity name are mapped to the corresponding disease code. The literature corresponding to the disease code is retrieved and the literature is input into the trained target screening model so as to select candidate targets from the literature through the target screening model. The authority score of each candidate target is determined based on the document type of the document in which each candidate target is located. Determine the similarity distance between each candidate target and the disease code, and determine the relevance score for each candidate target; Based on the number of documents in which each candidate target is located, a literature basis score is determined for each candidate target. The authority score, relevance score, and literature reference score of each candidate target are weighted and summed to obtain the confidence score of each candidate target. Based on the confidence score, target targets are selected from each of the candidate targets.

[0015] In one feasible implementation, the step of obtaining a spur compound that matches the ligand binding pocket includes: The trained molecular generation model is used to generate molecules at the ligand binding pocket to obtain the first number of generated ligand molecules. The affinity of a second number of ligand molecules in a preset molecule set to the ligand binding pocket is evaluated using a trained affinity assessment model, and an affinity score is obtained for each of the second number of ligand molecules. A third number of ligand molecules are selected from the second number of ligand molecules according to the affinity score; The first number of ligand molecules and the third number of ligand molecules are subjected to deduplication to obtain a fourth number of ligand molecules; Determine the confidence score for each of the fourth number of ligand molecules, and screen for lead compounds that match the ligand binding pocket from the fourth number of ligand molecules according to the confidence score.

[0016] In one feasible implementation, the step of running a drug molecule discovery process related to the entity based on the result of the intent recognition includes: Based on the entity and the result of the intent recognition, a standardized task instruction is generated; After the task instructions are standardized, a tool executor that is uniformly encapsulated into a preset format interface is invoked so that the corresponding tool executor can perform task processing. The processing result of the tool executor executing the task instruction is obtained and standardized. The calling interface of at least one tool included in the standardized task instruction is encapsulated into a preset format interface; In one feasible implementation, the step of standardizing the task instructions and then calling a tool executor that is uniformly encapsulated into a preset format interface, so that the corresponding tool executor can perform task processing, includes: The task instructions are serialized into a task request in a preset format; When the task request requires the use of the local first tool executor, the preset encapsulated interface is invoked to send the task request to the first tool executor so that the first tool executor can perform task processing. When the task request requires the use of a remote second tool executor, a preset encapsulation interface is invoked and the task request is sent to the second tool executor through a preset secure network channel so that the second tool executor can perform task processing.

[0017] A drug molecule discovery device, comprising: A supervisory agent is used to extract entities from input text and perform intent recognition, and to schedule drug molecule discovery operation units related to the entities based on the results of the intent recognition. The drug molecule discovery operational unit includes one or more of the following units: Target identification unit: used to extract disease entity names from input text, and determine the target targets corresponding to the disease entity names and the target structure of the target targets; Lead compound discovery unit: used to detect binding pockets in target structures, determine ligand binding pockets in the target structures, and obtain lead compounds that match the ligand binding pockets; The lead compound to leader compound screening unit is used to derivatize and expand lead compounds to obtain candidate derivatized compounds, and to screen leader compounds from each candidate derivatized compound based on a comprehensive score of each candidate derivatized compound. Lead compound optimization unit: used to generate molecules from lead compounds using a trained molecular generation model, obtain multiple candidate molecules, acquire the comprehensive optimization score and docking score of each candidate molecule, and generate a drug discovery molecule ranking list and rationale.

[0018] A computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute the above-described drug molecule discovery method.

[0019] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described drug molecule discovery method.

[0020] This application's implementation extracts entities from input text and performs intent recognition. Based on the intent recognition results, it schedules a drug molecule discovery process related to the entities. The drug molecule discovery process includes one or more of the following steps: Target identification: Extracting disease entity names from input text and determining the target targets and their structures corresponding to the disease entity names; Lead compound discovery: Detecting binding pockets in the target structures to identify ligand binding pockets and obtaining lead compounds that match the ligand binding pockets; From lead compounds to lead compounds: Extending the lead compounds to obtain candidate derivative compounds, and selecting lead compounds from each candidate derivative compound based on their comprehensive score; Lead compound optimization: Generating molecules from the lead compounds using a trained molecular generation model to obtain multiple candidate molecules, obtaining the comprehensive optimization score and docking score for each candidate molecule, and generating a drug discovery molecule ranking list and rationale.

[0021] In this way, by extracting entities from the input text and recognizing the intent of the input text, one or more processes in the drug molecule discovery workflow related to the entities can be scheduled according to the intent result. This allows users to simply provide the input text, and the subsequent operation of the drug molecule discovery workflow will automatically schedule one or more processes in the drug molecule discovery workflow related to the entities in the input text according to the intent result of the input text. This avoids users manually importing and exporting data between different processes and improves the processing efficiency of drug molecule discovery.

[0022] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the description, claims and drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 A schematic diagram of a drug molecule discovery system provided in an embodiment of this application.

[0025] Figure 2 This is a flowchart illustrating the drug molecule discovery method provided in the embodiments of this application.

[0026] Figure 3 This is a schematic diagram illustrating the execution of different processes by entities based on different intentions and results, provided for the implementation of this application.

[0027] Figure 4 This is a flowchart illustrating the manual review process during the implementation of the present application.

[0028] Figure 5 This is a flowchart illustrating the target identification process provided in the embodiments of this application.

[0029] Figure 6 A schematic flowchart illustrating the lead compound discovery process provided for embodiments of this application.

[0030] Figure 7 This is a schematic flowchart illustrating the process from identifying a lead compound to screening for a precursor compound, as provided in the embodiments of this application.

[0031] Figure 8 A schematic flowchart illustrating the lead compound optimization process provided for embodiments of this application.

[0032] Figure 9 This is a schematic diagram of the structure of the drug molecule discovery device provided in the embodiments of this application.

[0033] Figure 10 A schematic diagram of the structure of a computer device provided for an embodiment of this application. Detailed Implementation

[0034] To enable those skilled in the art to better understand the solutions of this application, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0035] It is understood that in the specific implementation of this application, data such as input text is involved. When the implementation of this application is applied to a specific product or technology, permission or consent from the subject is required, and the collection, use and processing of related data must comply with relevant laws, regulations and standards.

[0036] Furthermore, when the embodiments of this application require separate permission or separate consent to obtain input text and related data, the necessary input text and related data for the normal operation of the embodiments of this application shall be obtained only after obtaining separate permission or separate consent to obtain input text and related data.

[0037] Before providing a further detailed description of the embodiments of this application, the terms and concepts used in the embodiments of this application are explained, and the terms and concepts used in the embodiments of this application shall be interpreted as follows: Drug molecule discovery is a systematic research process that starts with disease mechanisms, screening, designing, and optimizing candidate molecules that can specifically interact with disease-related biological targets (such as proteins and genes) to ultimately obtain compounds with potential drug value. It is the core starting point of new drug development and directly determines the success rate of subsequent preclinical and clinical trials. Its core objective is to find seed compounds (Hit) and optimize them into lead compounds (Lead), laying the foundation for subsequent drug development.

[0038] A target is a specific biomolecule that allows a drug molecule to exert its therapeutic effect. It is the core bridge connecting disease mechanisms and drug design, and the starting point for drug discovery. Essentially, it is a biomolecule directly related to the occurrence and development of a disease, and its abnormal function (such as overactivation or deficiency) is a key trigger for the disease.

[0039] A ligand-binding pocket is a binding site between a drug molecule and its target. It is a specific three-dimensional cavity structure on or inside the surface of a target macromolecule (such as a protein). It serves as the physical space for the specific binding of the drug molecule (ligand) to the target and is the core basis of structure-based drug design (SBDD). Essentially, it is a "depression" formed by the spatial arrangement of amino acid residues of the target protein. Its size, shape, and chemical properties (such as hydrophobicity and charge) are complementary to those of the ligand molecule, ensuring the specificity of binding.

[0040] Lead compounds are small molecule compounds that, after optimization, possess both strong target-binding activity and good drug-like properties. They represent a crucial transition from drug discovery to clinical candidates and are the core focus of subsequent preclinical and clinical trials. Essentially derived from lead compounds (hit), they require a balance maintained through multi-drug evaluation; they do not need to be perfect, but must meet the basic requirements for clinical translation.

[0041] Please see Figure 1 , Figure 1 This is a schematic diagram of a drug discovery system provided in an embodiment of this application. It includes a server 140, an internet connection 130, a gateway 120, and a client 110, etc.

[0042] Client 110 includes, but is not limited to, mobile phones, tablets, pre-configured desktop computers, laptops, tablet computers, and other electronic devices. Client 110 can communicate with the Internet 130 via wired or wireless means to exchange data.

[0043] Server 140 refers to a computer system capable of providing data processing services (such as drug molecule discovery) to client 110, and includes a database for storing relevant operations performed by client 110. Compared to a regular client 110, server 140 has higher requirements in terms of stability, security, and performance. Server 140 can be a single high-performance computer in a network platform, a cluster of multiple high-performance computers, a portion of a single high-performance computer (e.g., a virtual machine), or a combination of portions of multiple high-performance computers (e.g., virtual machines), etc., and the server stores pre-defined timing logic.

[0044] Gateway 120, also known as an internetwork connector or protocol converter, is a computer system or device that acts as a translator, enabling network interconnection at the transport layer. It bridges the gap between two systems using different communication protocols, data formats, languages, or even completely different architectures. Gateways can also provide filtering and security functions. In this application, the gateway serves as a pre-defined secure network channel; input text sent by client 110 to server 140 is transmitted to the corresponding server 140 via gateway 120.

[0045] The drug discovery method of this application can be implemented on server 140.

[0046] The server extracts entities from the input text and performs intent recognition. Based on the results of the intent recognition, it schedules a drug molecule discovery process related to the entities. The drug molecule discovery process includes one or more of the following steps: Target identification: Extracting disease entity names from the input text and determining the target targets and their structures corresponding to the disease entity names; Lead compound discovery: Detecting binding pockets in the target structures to identify ligand binding pockets and obtaining lead compounds that match the ligand binding pockets; From lead compounds to lead compounds: Extending the lead compounds to obtain candidate derivative compounds, and selecting lead compounds from each candidate derivative compound based on their comprehensive score; Lead compound optimization: Generating molecules from the lead compounds using a trained molecular generation model to obtain multiple candidate molecules, obtaining the comprehensive optimization score and docking score for each candidate molecule, and generating a drug discovery molecule ranking list and rationale.

[0047] It should be noted that, Figure 1The schematic diagram of the drug molecule discovery system shown is merely an example. The drug molecule discovery system and scenario described in this application are intended to more clearly illustrate the technical solutions of this application and do not constitute a limitation on the technical solutions provided in this application. As those skilled in the art will know, with the evolution of drug discovery and the emergence of new business scenarios, the technical solutions provided in this application are also applicable to similar technical problems.

[0048] In this embodiment, the description will be from the perspective of a drug molecule discovery device, which can be integrated into a computer device that has a storage unit and a microprocessor and computing capabilities. The computer device can be a server, and in this embodiment, the computer device will be described as a server.

[0049] Please see Figure 2 , Figure 2 This is a schematic flowchart of the drug molecule discovery method provided in this application. The drug molecule discovery method is applied to a server and includes: In step 201, entities are extracted from the input text and intent recognition is performed. Based on the results of intent recognition, the drug molecule discovery process related to the entities is scheduled.

[0050] To translate user requests (natural language) regarding a specific disease into executable system tasks, the system first extracts entities from the input text (such as disease entity names and target information) and identifies user intent (such as performing knowledge retrieval queries, running single-process tasks, or running end-to-end tasks). This provides a basis for subsequent automatic scheduling of execution units to execute the corresponding processes, addressing the disconnect between user needs and system execution in related technologies. Entities are core concepts in text that carry key information. Entity extraction is the technical process of identifying, extracting, and classifying these concepts from unstructured text. By recognizing user intent, the system determines the true purpose of the user's input text, thus transforming the user's natural language requests into executable system tasks.

[0051] Specifically, in one embodiment, the Large Language Model (LLM) built into the Supervisor Agent is used to extract entities from the user input text and perform intent recognition, as exemplified below: The input contains keywords such as "query," "literature," "latest research," and "what" as entities, and does not involve specific computational tasks or drug generation instructions. The intent is identified as: performing a knowledge retrieval query targeting these entities. The result of intent identification is to run a knowledge retrieval query task, utilizing LLM's own general biomedical parameterized knowledge and the mounted domain knowledge database for retrieval enhancement. The retrieval results are then directly output and fed back to the user by a supervised agent.

[0052] For example, if the input text is "query common targets of lung cancer", then LLM integrates general biomedical parameterization knowledge with the attached databases (such as abstract databases and activity databases) and outputs the search results (such as "common targets of lung cancer include XXX").

[0053] The drug discovery process includes target identification, lead compound discovery, screening of lead compounds to lead compounds, and lead compound optimization. This application's embodiments construct a reusable, automated closed-loop process for each step of the drug discovery process. Each step is constructed as a skill graph based on a directed acyclic graph (DAG), with each graph running as an independent state machine and supporting breakpoint resumption. Nodes in the DAG represent encapsulated atomic manipulation tools (e.g., "performing protein structure data query" or "running pocket ligand docking scoring"), and edges represent the direction of data flow and dependencies (i.e., the output of the preceding node is automatically formatted and used as the input of the subsequent node). The supervisory agent can run the corresponding skill graph based on the intent recognition results and can perform dynamic routing, instantiation invocation, and cross-stage state management.

[0054] In some implementations, a drug molecule discovery process associated with the entity is run based on the result of the intent recognition, including: (1) When the result of the intent recognition is to run a single-process task, the target process related to the entity is scheduled from the target identification, lead compound discovery, lead compound to lead compound screening and lead compound optimization in the drug molecule discovery process, and the target process is run; (2) When the result of the intent recognition is to run an end-to-end task, at least two processes in the drug molecule discovery process, namely target identification, lead compound discovery, lead compound to lead compound screening and lead compound optimization, are run in sequence.

[0055] If the result of intent recognition is to run a single-process task, the supervisory agent schedules the target process that the user expects to execute from the target identification, lead compound discovery, lead compound screening to lead compound optimization, and lead compound optimization of the drug molecule discovery process, and runs the target process.

[0056] If the result of intent recognition is to run an end-to-end task, the supervisory agent will execute at least two of the following processes in sequence: target identification, lead compound discovery, lead compound screening to lead compound selection, and lead compound optimization. During execution, the "output" of each skill graph will be automatically passed to the next graph as "input".

[0057] In some implementations, a drug molecule discovery process associated with the entity is run based on the result of the intent recognition, including: (1) Based on the entity and the result of the intent recognition, generate task instructions; (2) After standardizing the task instructions, call the tool executor that is uniformly encapsulated into a preset format interface so that the corresponding tool executor can perform task processing; (3) Obtain the processing results of the task instructions executed by the tool executor and perform standardization processing.

[0058] To achieve unified scheduling of multiple processes, task instructions need to be standardized to generate standardized task instructions in a unified format, including task type, target stage, input parameters, and tool requirements. These standardized task instructions are then distributed to various tool executors. Specifically... Task instructions are serialized into requests in a preset format (such as MCP requests) and distributed via the encapsulated MCP interface to tool executors (including local processing modules and / or external tool computing platforms) that are uniformly encapsulated with preset format interfaces, so that the corresponding tool executor can be invoked to process the request. If the tool executor includes an external tool computing platform, it is distributed to the external tool computing platform through a secure network channel. Asynchronous tasks and data format conversions can be performed simultaneously to ensure stable process operation.

[0059] Specifically, each process execution task requires a complex biochemical computing environment, and the models or computing tools called are often heterogeneous. Therefore, it is necessary to pre-encapsulate the calling interface of at least one tool executor included in the standardized task instructions into a preset format interface. Furthermore, since the formats returned by local processing modules and remote processing tools differ significantly, an adapter needs to be set up. After obtaining the processing results of the tool executor's task instructions, the task processing results need to be standardized, and a forced format conversion is performed before returning them to the supervisory agent. This ensures that the input and output data within the drug molecule discovery process conform to the predefined format (Schema), enabling the supervisory agent's large language model to read and understand with minimal barriers, preventing process interruptions due to format errors. This embodiment ensures the stability and process continuity of heterogeneous tool calls through serialized requests, asynchronous state processing, and data format conversion. It efficiently connects user needs with tool execution, avoiding the process breakage problems caused by incompatible tool interfaces and ambiguous instructions in traditional processes, while also reducing system maintenance and expansion costs, ensuring the efficiency, stability, and flexibility of the drug molecule discovery process.

[0060] On the one hand, this application does not directly deploy and maintain all complex biochemical computing environments locally, but instead abstracts the remote or independent computing resources required by various tool executors into standardized function call interfaces through the Model Context Protocol (MCP). In this embodiment, computing resources include, but are not limited to, three categories: (1) basic data services: including database query services such as UniProt (protein sequence), RCSB PDB (protein crystal structure), ChEMBL (activity data), PubMed literature database, ClinicalTrials clinical trial database, and Ensembl database; (2) standard computing tool services: including scientific computing tools such as fpocket / DeepSite, RDKit, molecular docking tools, chemical information processing / conversion, and REINVENT tool library; (3) dedicated AI model services: including molecular generation models, affinity prediction models, and pocket-structure-based molecular generation models. The computing resources used in this application are illustrative examples and do not constitute a limitation on the technical solutions provided in the embodiments of this application. Secondly, the native interfaces of different tool executors (such as REST API, CLI command line, and gRPC) are uniformly encapsulated into standardized interfaces in JSON-RPC format through the Model Context Protocol (MCP), forming a unified tool abstraction layer. In this way, the upper-layer supervisory agent does not need to care about the underlying tool executor implementation of task execution, achieving decoupling between upper-layer intent and lower-layer task execution instructions. It only needs to call the standardized interface, which greatly improves the system's scalability and tool compatibility.

[0061] On the other hand, for time-consuming tasks (such as high-throughput virtual filtering), the current process adopts an asynchronous non-blocking mode for computation. After sending the MCP request, a task tag (such as TaskID) is obtained immediately, and the task status is polled through a heartbeat mechanism (QUEUED->PROCESSING->COMPLETED) until the final result is obtained, thereby avoiding system blocking caused by long connections.

[0062] In some implementations, the task instructions are standardized and then invoked by a tool executor that is uniformly encapsulated into a preset format interface, so that the corresponding tool executor can perform task processing, including the following steps: (1.1) Serialize the task instructions into a task request in a preset format; (1.2) When the task request requires the use of the local first tool executor, the preset encapsulation interface is called to send the task request to the first tool executor so that the first tool executor can perform task processing; (1.3) When the task request requires the use of a second tool executor in a different location, the preset encapsulation interface is invoked and the task request is sent to the second tool executor through a preset secure network channel so that the second tool executor can perform task processing.

[0063] The preset format can be MCP format, which is converted into a task request MCP Request by serializing the task instructions. When the tool executor required by the task request is the local first tool executor, the preset encapsulation interface (i.e., the MCP interface) is called to send the task request directly to the local first tool executor so that the first tool executor can perform task processing. If the tool executor required by the task request is the external (i.e., remote) second tool executor, in order to improve the security of network transmission, while calling the preset encapsulation interface (i.e., the MCP interface), it is also necessary to send the task request directly to the remote second tool executor through a secure network channel so that the second tool executor can perform task processing.

[0064] Specifically, the supervisory agent determines the initial process to be executed by extracting key parameters from the input text.

[0065] Regarding the target identification process: the extracted entities include "disease entity name" but lack "target" or "protein structure" information.

[0066] For the discovery process of lead compounds: the extracted entities include "target name" or "PDB ID / file", and the goal is "find molecules" or "screen".

[0067] For lead compound optimization processes: the extracted entities include specific "small molecule structures (SMILES / SDF format)" or "lists of lead compounds," and the instructions contain optimization descriptions such as "improving affinity" and "improving drug-likeness." In some implementations, the intelligent supervisor, based on the intent recognition results and the extracted entities, first determines at least one process in the drug molecule discovery process related to the entity that needs to be scheduled, and generates task instructions. However, the process of running the drug molecule discovery process related to the entity does not directly call tools to complete the task. Instead, it uses a hierarchical tool search skill graph driven by LLM to first decide the required tool category at a macro level, and then precisely select the specific tool executor under the selected category. Thus, after the LLM recognizes the intent of the input text, it dynamically routes to various skill graphs to call various tool executors to run the drug molecule discovery process related to the entity. Specifically, the supervisory agent can determine the starting point of the task in the required execution process by checking the data modality of the input text (disease text, protein structure file, or small molecule string), thereby accurately routing to the corresponding skill graph.

[0068] If the result of intent identification is to run an end-to-end task, at least two of the following processes are executed in sequence: target identification, lead compound discovery, lead compound screening to lead compound selection, and lead compound optimization.

[0069] For example, if the input text is "design a new drug for non-small cell lung cancer", then the process needs to run sequentially: target identification, determination of targets for non-small cell lung cancer, discovery of lead compounds, identification of lead compounds matching the ligand binding pocket in the target, screening of lead compounds to lead compounds, determination of qualified lead compounds, and optimization of lead compounds to obtain the final drug discovery molecule. In other words, the four processes are run in sequence.

[0070] For example, if the input text is "identify lead compounds for non-small cell lung cancer", then the target identification process needs to be run sequentially to identify the target for non-small cell lung cancer, and the lead compound discovery process needs to be run to identify lead compounds that match the ligand binding pockets in the target. That is, the target identification process and the lead compound discovery process are run in sequence.

[0071] By employing a dual-driven process scheduling logic based on intent and entities, combined with a skill graph, the system achieves precise routing of single-process tasks, avoiding invalid process calls; it also supports the orderly connection of end-to-end tasks without requiring manual intervention in data transmission; and the breakpoint resume capability of the DAG state machine ensures process robustness. Overall, it improves the scheduling accuracy, execution efficiency, and operational flexibility of the drug molecule discovery process, while simultaneously addressing the dual requirements of rapid response for single tasks and full-process automation, significantly reducing manual operation costs and ensuring efficient matching and scientific rigor from user needs to process execution.

[0072] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating the execution of different processes by entities based on different intentions and results, provided for the implementation of this application. Figure 3 This demonstrates the overall logic of drug research, with its core being the flexible matching of fully automated processes, independent execution of individual stages, and knowledge-based retrieval based on the user's natural language needs. The specific logic is as follows: Starting with the user's natural language request (e.g., "Help me complete the entire drug molecule discovery process for non-small cell lung cancer" or "Optimize only a lead compound"), the workflow analyzes the user's request, extracts entities (e.g., disease, target), and identifies the user's objective (whether to run an end-to-end task or a single-process task). If the request intent is unclear or the core data is insufficient, the workflow ends directly. If the intent is clear, it determines whether the user needs to execute an end-to-end task or a single-process task. Based on the determination result, it executes in two modes: Single-process task (specific stage mode) only executes a single stage specified by the user (any one of the following: target discovery, lead compound discovery, lead to lead compound, or lead compound optimization). The end-to-end task first determines the two, three, or full processes to be executed (in this embodiment, the full process is taken as an example). Following the standard logic of drug discovery, the four stages of target discovery, lead compound discovery, lead to lead compound, and lead compound optimization are executed sequentially. Regardless of the mode selected, the entire workflow will end after the corresponding process is completed.

[0073] This enables drug research workflows to support both end-to-end fully automated execution from target to lead compound, as well as independent operation of individual stages, flexibly matching different R&D needs of users and improving the adaptability and efficiency of the drug discovery process.

[0074] In some implementations, for an end-to-end task, the process of sequentially running at least two steps in the drug molecule discovery pipeline—target identification, lead compound discovery, lead compound screening to lead compound selection, and lead compound optimization—includes the following steps: (1.1) When the target node of the current process is detected, obtain the first execution result of the previous running node of the target node; (1.2) Standardize and verify the first execution result, and output a prompt on whether to continue executing subsequent nodes based on the first execution result; (1.3) When a continue running instruction based on the first execution result is received, the subsequent nodes of the target node continue to run with the first execution result as input data.

[0075] Specifically, for end-to-end task execution, human interaction nodes are embedded in the critical decision paths of each process (such as after the generation of stage products). These human interaction nodes can be set at the process's start node, end node, and certain core critical nodes. Specifically, for the case where the human interaction node is set at the process's start node: during the process of running according to the DAG, if the process reaches the target node that serves as the human interaction node in the current process, the first execution result of the target node's preceding node is obtained, and this first execution result is standardized and verified. Furthermore, by rendering the first execution result to the front end (such as the client), a prompt is output asking whether to continue executing subsequent nodes based on the first execution result. When a continue execution instruction based on the first execution result is received, it indicates that the user has approved the first execution result, meaning that the subsequent nodes in the DAG will continue to run using the first execution result as input data.

[0076] In this way, in the multi-process connection of end-to-end drug discovery, by setting the human interaction node at the starting node, when the process reaches the target node that is the starting node, the first execution result is obtained and rendered on the front end. The human review process is introduced to confirm whether it meets the process objectives. This not only seamlessly integrates the human review node into the automated process, but also makes the key decisions of each process results controllable and traceable throughout the process. It effectively avoids the flow of unqualified results from the previous process into the current process, and balances the automation efficiency of drug discovery with the rigor of scientific decision-making.

[0077] In some implementations, the method further includes: (1.1) When a parameter correction instruction based on the first execution result is received, the first parameter carried by the parameter correction instruction is obtained; (1.2) Input the first parameter to the previous running node of the target node so that the previous running node of the target node re-runs according to the first parameter to obtain the second execution result; (1.3) Standardize and verify the second execution result, and output a prompt indicating whether to continue executing subsequent nodes based on the second execution result; (1.4) When no continue execution instruction based on the second execution result is received, the execution of subsequent nodes of the target node shall be prohibited.

[0078] If a parameter correction instruction based on the first execution result is received, it indicates that the user is dissatisfied with the first execution result and will provide a correction parameter (i.e., the first parameter). By inputting the first parameter to the previous running node of the target node, the previous running node of the target node will re-run according to the first parameter, obtaining the second execution result after re-run. The second execution result is then standardized and verified, and a prompt is output asking whether to continue executing subsequent nodes based on the second execution result. If a continue execution instruction based on the second execution result is received, it indicates that the user accepts the second execution result, and the subsequent nodes of the target node can continue running. If no continue execution instruction based on the second execution result is received, it indicates that the second execution result still does not meet the user's requirements. Since the user has already corrected the parameter using the parameter correction instruction, the running path can be determined to be invalid, and the running of subsequent nodes of the target node is prohibited.

[0079] In some embodiments, the method further includes: (1.1) When a manual editing instruction based on the first execution result is received, the second parameter carried by the manual editing instruction is obtained; (1.2) Continue running the subsequent nodes of the target node using the second parameter as input data.

[0080] If a manual editing instruction based on the first execution result is received, it means that the user is not satisfied with the first execution result. The user can directly modify the first execution result by manually editing it, that is, obtain the second parameter carried by the manual editing instruction, and use the second parameter as input data to continue running the subsequent nodes of the target node.

[0081] In some implementations, the method further includes: (1.1) When the current process is detected to have completed, obtain the second execution result of the current process; (1.2) Display the second execution result and output a prompt asking whether to continue the subsequent process based on the second execution result; (1.3) When a continue execution instruction based on the second execution result is received, the subsequent process of the current process continues to run.

[0082] Specifically, for cases where the human interaction node is set at the end point of the process: During task processing according to the DAG, if the current process is detected to have completed, it indicates that the current process has reached the end point. The second execution result of the current process is then obtained and rendered to the front end (e.g., the client). The second execution result is then displayed to the user, along with a prompt indicating whether the user should continue executing subsequent processes based on the second execution result, indicating whether the user has approved the second execution result. When a continue instruction based on the second execution result is received, it indicates that the user has approved the second execution result, and the subsequent processes in the DAG can continue.

[0083] By placing the human interaction node at the process termination point, and obtaining and rendering the second execution result on the front end after the current process is completed, a user review process is introduced to confirm whether it meets the process objectives. This seamlessly integrates into the automated process while ensuring controllability and full traceability of key decisions in each process, effectively preventing unqualified results from the previous process from flowing into subsequent processes, thus balancing the automation efficiency of drug discovery with the rigor of scientific decision-making. Furthermore, if the instruction received by the human interaction node is to stop execution, it means re-executing the previous node or process. After the user inputs text at the human interaction node, parameter correction or manual editing instructions can be executed. The supervisory agent will automatically roll back to the starting state of the current process based on the user's input text and recalculate using the new parameters, or directly use the revised data input by the user to overwrite the output data of the previous node or process; if the path is determined to be invalid, the task can be terminated. This multi-turn interactive dialogue mechanism ensures that users can correct, backtrack, and control the automated process of the same task in real time, rather than passively accepting results.

[0084] Please see Figure 4 , Figure 4 This is a flowchart illustrating the manual review process during the implementation of the present application. Figure 4 This demonstrates the execution logic of a specific stage in the drug discovery process. The core is a closed-loop process that combines data verification, human-computer interaction, core calculations, and result integration to achieve a balance between automation and manual control within a single stage. The specific steps and logic are as follows: Starting with the output of the previous round and historical information (i.e., the current process input data, the output results of the previous process, and historical interactive dialogue information): The preprocessing module first determines the required parameters based on the current process requirements, and then verifies whether the input data (previous round output, historical information) meets the parameter compliance standards. This is equivalent to data detection and standardization processing and verification to avoid starting the calculation with incorrect or non-compliant input. The human-computer interaction node (new user input data, or confirmation to continue execution) synchronizes the current interaction information to the user through a supervisory agent. The supervisory agent outputs a prompt asking whether to continue execution of subsequent nodes based on the first execution result, allowing the user to control the process before the core calculation and avoid invalid computations. The core calculation logic executes the core tasks of this stage (such as target screening, molecular docking, and other key operations in drug discovery). This is the core working link of this stage. Specifically, it serializes task instructions into MCP Requests and distributes them through a secure network channel via a packaged MCP interface to a uniformly packaged processing module and / or external tool computing platform to obtain the second execution result. Upon receiving the second execution result, the post-processing module sets up a set of adapters. After obtaining the second execution result, it needs to standardize it, forcibly performing a format conversion before returning it to the supervisory agent to display the second execution result and output a prompt asking whether to continue execution of subsequent processes based on the second execution result. The human-computer interaction node (confirmation of whether to continue execution) is used to ask the user whether to continue the subsequent process, thereby allowing the user to control the pace of the process and avoid automatically executing steps that do not meet the requirements. Within a single stage, the process achieves both automated and standardized execution (pre-processing verification, core calculations) and, through two human-computer interaction nodes, allows the user to control the validity of input and the start and stop of the process, thus balancing the efficiency of drug discovery with the controllability of key decisions.

[0085] The following sections will detail each step of the drug molecule discovery process, which includes one or more of the following steps: In step 202, target identification involves extracting disease entity names from the input text and determining the target points corresponding to the disease entity names, as well as the target point structure of the target points.

[0086] Specifically, for the target identification process, the disease entity names extracted by the supervised agent are standardized (e.g., mapped to MeSH ID or ICD encoding) to verify whether they fall within the scope of druggable research, excluding diseases without druggable targets. Unstructured text, such as research abstracts, associated genes, and clinical trial status related to the disease entity names, is obtained by parallel access to literature search tools (PubMed / Google Scholar / BioRxiv) and clinical trial data search tools (ClinicalTrials.gov) via the MCP protocol. LLM is then used to reason through the unstructured text, extracting a list of potential targets, selecting target targets from the list, and obtaining the target structure of the target targets.

[0087] Specifically, the target PDB file, i.e. the target structure of the target site, is obtained through automated chained lookup (Ensembl / UniProt API maps the target name to a unique UniProt ID; RCSB PDB API is called to retrieve the corresponding protein 3D structure data based on the UniProt ID); the PDBFixer tool is called to repair missing side chains, add hydrogen atoms, and remove water molecules in the PDB file to ensure that the structure meets the input specifications for subsequent computational simulations.

[0088] Finally, a comprehensive analysis report containing the target information is generated, and the path to the cleaned and preprocessed protein structure file is output. For single-process tasks involving target identification, the file is available for download; for end-to-end tasks involving target identification, the target information and its structure output from the target identification process can also be directly passed to the next process as input.

[0089] In some implementations, determining the target point corresponding to the disease entity name includes: (1) Map the disease entity name to the corresponding disease code, retrieve the literature corresponding to the disease code, and input the literature into the trained target screening model so as to screen out candidate targets from the literature through the target screening model; (2) Determine the authority score of each candidate target based on the document type of the document in which each candidate target is located; (3) Determine the similarity distance between each candidate target and the disease code, and determine the relevance score of each candidate target; (4) Determine the literature basis score for each candidate target based on the number of documents in the literature in which each candidate target is located; (5) The authority score, relevance score and literature basis score of each candidate target are weighted and summed to obtain the confidence score of each candidate target; (6) Select target points from each candidate target point according to the confidence score.

[0090] The supervisory agent serializes task instructions into MCP Requests, and uses medical knowledge graph tools to map the input disease entity names to corresponding disease codes (such as standard MeSH IDs or ICD codes) to verify whether they fall within the scope of druggable research and exclude diseases without druggable targets. It also retrieves literature related to the disease entity name, such as research abstracts, associated genes, and clinical trial status, through the MCP protocol, and inputs the obtained unstructured text into a trained target screening model (such as an LLM model) to extract a list of potential targets using its reasoning ability.

[0091] The specific method for selecting target targets from the target list can be as follows: generate a confidence score based on the number of documents in the literature containing each candidate target in the target list (determine the literature basis score), the type of the literature containing it (determine the authority score), and the relevance (determine the relevance score), and select the top K high-potential targets according to the confidence score as target targets.

[0092] Specifically, the literature reference score is determined based on the number of literature entries in databases (PubMed, ClinicalTrials, etc.) where each candidate target and disease co-occur, and the number of registered clinical trials; the more entries, the higher the score. The authority score is determined based on the impact factor of the literature source or the stage of the clinical trial (Phase I, II, and III). For example, evidence from *Nature* or clinical trials in Phase III have higher weight than those from ordinary journals or preclinical studies. The relevance score is determined based on the semantic similarity vector distance between the "disease mechanism description" and the "target function description" in the literature abstract, calculated using LLM or text embedding models; the closer the distance, the higher the relevance score.

[0093] The confidence score for each candidate target is determined by weighting and summing the literature reference score, authority score, and relevance score.

[0094] In this way, by standardizing and mapping disease entity names to disease codes, and combining the MCP protocol to efficiently pull the latest literature and clinical trial data from multiple sources, a three-dimensional objective scoring system is constructed based on the dimensions of quantity, authority, and relevance. By generating confidence scores through weighted summation, target points are screened. This not only avoids the problems of human subjectivity and literature dispersion in traditional target point screening, but also significantly improves the accuracy, reliability, and traceability of target point screening through standardized processes, multi-dimensional quantitative evaluation, and authoritative data support, laying a high-quality target foundation for subsequent drug discovery processes.

[0095] Please see Figure 5 , Figure 5This is a flowchart illustrating the target identification process provided in this application. Starting with the user's natural language research needs for a specific disease, core information, namely the disease entity name, is extracted through intent recognition combined with entity recognition, serving as the core input to the process. Scientific literature and clinical trial data related to the disease are obtained by calling literature databases (such as PubMed) and clinical trial databases (such as ClinicalTrials), providing data support for subsequent target screening. The "disease-target" database / literature resources are called to first rename / standardize the target names; then, LLM (Large Language Model) is used to assess and rank candidate targets, selecting high-potential targets; the target assessment information is summarized, outputting a clear target name and assessment results. The target name is queried in a free genome database (such as Ensembl) to obtain the corresponding UniProt ID; the UniProt ID is queried in the UniProt database to obtain the PDB ID; and the PDB ID is queried in the PDB database to obtain the target's three-dimensional structural information. The above data is integrated and summarized using the large model to generate a target discovery report, while also supporting the download of the target's protein three-dimensional structure file, thus ending the process.

[0096] Starting from the user's natural language needs, through multi-source data linkage, LLM intelligent screening, and database chain connection, it automates the entire process from disease, candidate target, high potential target to target 3D structure. It not only solves the problems of scattered data and inefficient manual screening in traditional target discovery, but also directly outputs target reports and structure files, providing a foundation for subsequent processes such as lead compound discovery.

[0097] In step 203, the lead compound discovery involves: probing the binding pocket of the target structure to identify the ligand binding pocket in the target structure and obtaining a lead compound that matches the ligand binding pocket.

[0098] The binding pocket detection includes two detection methods: first, if the PDB file contains a reference ligand, its center is used as the pocket center; second, if there is no reference ligand, a protein pocket detection tool (such as fpocket, DeepSite, etc.) is called to scan the surface geometry and hydrophobicity distribution of the target, sort them by druggability, and select the optimal ligand binding pocket. The parallel dual-stream screening strategy involves the system simultaneously launching two parallel and complementary sub-workflows to maximize the probability of discovering potential lead compounds.

[0099] In some embodiments, obtaining a spur compound that matches the ligand binding pocket includes: (1) Molecular generation is performed at the ligand binding pocket using the trained molecular generation model to obtain the first number of ligand molecules; (2) The affinity of the second number of ligand molecules in the preset molecular set to the ligand binding pocket is evaluated by the trained affinity evaluation model to obtain the affinity score of each of the second number of ligand molecules; (3) Select a third number of ligand molecules from the second number of ligand molecules according to the affinity score; (4) The first number of ligand molecules and the third number of ligand molecules are subjected to deduplication to obtain a fourth number of ligand molecules; (5) Determine the confidence score of each of the fourth number of ligand molecules, and screen out the lead compounds that match the ligand binding pocket from the fourth number of ligand molecules according to the confidence score.

[0100] The first number of ligand molecules are generated through workflow A, which is a structure-based drug design process. Molecular generation based on diffusion model adopts DiffSBDD algorithm, based on SE(3)-equivariant geometric diffusion model (i.e. trained molecular generation model), and models the molecular generation process as a 3D spatial denoising process in the target pocket, so as to realize the generation of ligand molecules that are highly complementary to the geometric structure of the target pocket.

[0101] To avoid generating a large number of redundant molecules with highly similar structures, the system uses Tanimoto similarity coefficient (which calculates the distance between molecular fingerprints) to cluster the generated ligand molecules, ensuring the diversity of the lead compounds screened from the ligand molecules in chemical space.

[0102] Docking verification involves using molecular docking tools (such as docking tools) to perform molecular docking on the screened lead compounds, calculating the binding free energy, and verifying the binding potential of the generated ligand molecules to the target site.

[0103] The third number of ligand molecules were generated via workflow B, which is a high-throughput virtual screening (HTVS). Library screening retrieves millions of readily available ligand molecules (i.e., the second number of ligand molecules) from a large, commercially available database integrated with the system.

[0104] We employed LigUnity, a self-developed general-purpose deep learning model for drug-target interactions, to rapidly score the affinity between ligand molecules and targets. This model has been trained on massive amounts of biological data to generalize its application, accurately predicting the affinity between unseen targets and ligand molecules. The screening speed is several orders of magnitude faster than traditional docking tools, and it can effectively identify nonlinear interactions that are difficult for traditional physical models to capture.

[0105] The high-scoring candidate lead compounds from workflows A and B are merged (i.e., duplicates are removed from the first and third number of ligand molecules). Structural deduplication and multi-dimensional weighted sorting are performed, comprehensively considering diffusion tensor imaging (DTI) prediction scores, docking scores, drug-likeness, and synthetic accessibility. A list of N high-potential lead compounds (i.e., the fourth number of ligand molecules) is output, including their 2D / 3D structures, predicted affinity values, origin (generation / screening), and corresponding confidence assessments. Based on the confidence scores, N lead compounds matching the ligand binding pocket are selected from the fourth number of ligand molecules in the list.

[0106] By employing two workflows, we not only address the issues of low efficiency and insufficient molecular diversity in the screening of lead compounds in existing technologies (simply screening or designing), but also consider both drug-likeness and synthetic feasibility in the generation of ligand molecules. Through complementary dual-pathway and multi-dimensional sorting, we maximize the probability of discovering effective molecules, ensuring the binding compatibility between molecules and targets while avoiding the problem of highly active but druggable molecules in advance, thus providing high-quality lead compounds for subsequent optimization.

[0107] Please see Figure 6 , Figure 6 This is a flowchart illustrating the lead compound discovery process provided for embodiments of this application. Through two workflows—structure-based drug design (SBDD) and high-throughput virtual screening (HTVS)—lead compounds that fit the target binding pocket are efficiently screened starting from the target's three-dimensional structure. The specific logic is as follows: Using the target's 3D structure file corresponding to the PDB ID as input, the active pocket prediction tool first identifies the ligand binding sites on the target, clarifying the pocket's coordinates, size, and corresponding amino acid sequence. This is the target region for subsequent molecular design / screening. Starting from the target pocket, two complementary workflows run simultaneously: Left side: SBDD workflow (structure-based drug design) generates new molecules from scratch to fit ligand binding pockets (covering novel structures not included in existing libraries). Molecules are generated from scratch within identified target pockets; a set of candidate ligand molecules is constructed and then clustered by molecular skeleton (to avoid structural duplication), with random sampling of each skeleton class; a sample set of clustered lead compounds is obtained and proceeds to subsequent docking evaluation.

[0108] Right side: HTVS workflow (High-throughput Virtual Screening) screens molecules for matching pockets from an existing large molecular library (covering known synthesizable molecules). It calls upon a large ligand library and uses affinity prediction tools for rapid initial screening, identifying the top Y range of candidate ligand molecules. These initially screened candidate ligand molecules are then clustered by backbone and randomly sampled to obtain a sample set of clustered lead compounds, which are then used for subsequent docking evaluation.

[0109] Both workflows obtain candidate ligand molecule sets, which undergo the same precise evaluation process: molecular docking scoring of the candidate ligand molecule sets (simulating the binding fit between ligand molecules and target pockets); sorting by scoring results, selecting the top K highly fitted ligand molecules, and finally outputting a set of lead compounds.

[0110] By employing a dual workflow of SBDD (de novo structure generation) and HTVS (high-throughput virtual screening), the potential of novel molecules is covered while also taking into account the syntheticity of existing molecules. Furthermore, the layered screening approach of "skeleton clustering + docking scoring" significantly improves the fit between lead compounds and target pockets, providing high-quality lead compounds for subsequent optimization.

[0111] In step 204, from seed compounds to lead compounds: the seed compounds are derivatized and expanded to obtain candidate derivatized compounds, and a lead compound is selected from each candidate derivatized compound based on the comprehensive score of each candidate derivatized compound.

[0112] Among them, two derivation strategies are executed in parallel, using the lead compounds input by the user or the lead compounds output by the lead compound discovery process as seeds: R-group substitution exploration: keeping the molecular skeleton unchanged, using the REINVENT tool library to enumerate and replace side chain groups to improve binding affinity; skeleton transition: keeping the key pharmacophore unchanged, replacing the core skeleton to improve physicochemical properties.

[0113] Two-stage automated filtering: (1) Primary filtering: Using the RDKit tool, molecules containing reactive groups, prone to false positives, or with significant toxicity risks are removed according to PAINS (panasonic assay interfering compounds) rules and the Brenk structure (unwanted substructure) alarm list; (2) Advanced filtering: Using prediction models such as the ADMET-AI model, the absorption, distribution, metabolism, excretion and toxicity of molecules are quantitatively predicted, and molecules with low oral bioavailability or high risk of cardiotoxicity (hERG inhibition) are filtered out to obtain candidate derivative compounds. Based on the comprehensive score of each candidate derivative compound, lead compounds are selected from each candidate derivative compound.

[0114] The selected lead compounds undergo structural optimization, eliminating significant defects and achieving an optimal balance between affinity and physicochemical properties. Their optimization pathways relative to the original lead compounds (such as "side chain modification" or "skeleton replacement") are clearly defined. This addresses the shortcomings of existing H2L technologies, which rely on artificial chemical modifications, resulting in low efficiency, easy neglect of druggability and toxicity, and a lack of standardized screening logic. It enables precise optimization and quality control of lead compounds by expanding chemical space coverage through chemical extension, eliminating defective molecules through two-stage filtration, and balancing activity, druggability, and safety through multi-objective scoring, ultimately producing lead compounds that meet clinical development requirements and shortening the optimization cycle.

[0115] In some embodiments, before selecting a lead compound from each of the candidate derivatized compounds based on a comprehensive score, the method further includes: (1) The affinity of each candidate derivative compound is evaluated by the trained affinity evaluation model to obtain an affinity score for each candidate derivative compound; (2) Obtain a quantitative drug-likeness score and a synthetic accessibility score for each of the candidate derivative compounds; (3) The toxicity risk of each candidate derivative compound is assessed by the trained toxicity risk assessment model to obtain a toxicity risk score for each candidate derivative compound; (4) The affinity score, quantitative drug-likeness score, synthetic accessibility score and toxicity risk score of each candidate derivative compound are combined to obtain the comprehensive score of each candidate derivative compound.

[0116] The comprehensive score for each candidate derivative compound is obtained through a comprehensive evaluation of four dimensions: affinity, quantitative drug-likeness, synthetic accessibility, and toxicity. The affinity score is obtained using a trained affinity assessment model (e.g., the DTI model); the toxicity risk score is obtained by evaluating multiple toxicities using a trained toxicity risk assessment model (e.g., the ADMET-AI model); and the quantitative drug-likeness and synthetic accessibility scores are calculated using the RDKit tool. The comprehensive score for each candidate derivative compound is obtained by integrating its affinity, quantitative drug-likeness, synthetic accessibility, and toxicity risk scores.

[0117] Therefore, by utilizing computational resources such as the DTI model, RDKit tool, and ADMET-AI model, candidate derivative compounds are professionally and accurately evaluated from four core dimensions of drug development: affinity (drug activity), quantitative drug-likeness (drug potential), synthetic accessibility (manufacturing feasibility), and toxicity risk (safety). The scores of each dimension are then integrated to form a comprehensive score. This approach avoids the problems of traditional screening that focus on a single indicator, resulting in compounds with strong activity but difficult to drug or toxic. Furthermore, the use of professional tools and models ensures the scientific rigor of the evaluation. Ultimately, it can screen out high-quality lead compounds that balance activity, drug potential, manufacturability, and safety, significantly improving the quality of lead compounds and reducing the ineffective costs and risks of subsequent research and development.

[0118] In some embodiments, an overall score for each candidate derivative compound is obtained by combining its affinity score, quantitative drug-likeness score, synthetic accessibility score, and toxicity risk score, including: (1.1) For each of the candidate derivative compounds, the corresponding affinity score, quantitative drug-likeness score and synthetic accessibility score are weighted and summed to obtain the first calculation result for each of the candidate derivative compounds; (1.2) Determine the difference between the first calculation result of each candidate derivative compound and the corresponding toxicity risk score to obtain the comprehensive score of each candidate derivative compound.

[0119] The process of calculating the affinity score, quantitative drug-likeness score, synthetic accessibility score, and toxicity risk score of the fusion candidate derivative compounds can be referenced by the following formula: ; in, The affinity score predicted by the DTI model. For quantitative drug property scoring, For the accessibility score of synthesis, A toxicity risk score is assigned to each type of toxicity. The sum of the toxicity risk scores for all toxicities, i.e., the toxicity risk score of the candidate derived compound. As a deduction item The weighting of affinity score The weighting of quantitative drug property scores, The weights for the toxicity risk scores of candidate derivative compounds.

[0120] The first calculated result for each candidate derivative compound is obtained by weighted summation of the corresponding affinity score, quantitative drug-likeness score, and synthetic accessibility score. Then, the first calculation result and the corresponding toxicity risk score of each candidate derivative compound are calculated. The difference is used to obtain a comprehensive score for each candidate derived compound, i.e. .

[0121] In this way, through the formulaic quantitative calculation of positive index weighted summation and cumulative deduction of toxicity risk, the core priorities of affinity (activity), quantitative drug-like properties (drug potential), and synthetic accessibility (manufacturing feasibility) are highlighted by the weight setting, while multiple toxicity risks are forcibly included in the evaluation as deduction items. This effectively avoids the biases caused by single-dimensional evaluation, such as emphasizing activity over safety or ignoring synthetic feasibility. It can accurately screen out comprehensive high-quality candidate derivative compounds that take into account activity, drug potential, manufacturability and low toxicity, greatly improve the scientific nature of the comprehensive score and the reliability of the screening results, provide clear quantitative basis for the selection of subsequent lead compounds, and reduce the cost of ineffective research and development.

[0122] Please see Figure 7 , Figure 7 This is a schematic flowchart illustrating the process from lead compound to pilot compound selection provided in this application. Based on a set of lead compounds including at least one lead compound, derivation and expansion are performed through R-group substitution exploration (substitution with R groups) and skeletal transition sampling, amplifying by a specific fold to obtain candidate derived compounds, i.e., the amplified set of lead compounds. Primary filtering is performed through three operations: structural standardization, filtering molecules containing toxic / reactive / undesirable chemical groups, and filtering unsuitable molecules according to filtering rules (such as REOS rules, Rule of Excluded Substances). Advanced filtering is then performed based on dimensions such as aromaticity ratio, intestinal permeability (caco2), CYP inhibition, and toxicity (such as potassium ion channel gene hERG, AMES test, drug-induced liver injury (DILI), and clinical toxicity (clintox), achieving a comprehensive evaluation of drug-likeness and ADMET characteristics. A comprehensive penalty score is calculated for the filtered candidate derived compounds, and the top K (e.g., K≤5) high-quality molecules are selected based on their scores, ultimately yielding the pilot compound. It not only covers more potential effective structures, but also, through rigorous structure and property screening, ultimately outputs high-quality lead compounds that balance activity, drug potential and safety, in SMILES or SDF format, laying the foundation for subsequent optimization.

[0123] In step 205, lead compound optimization: the lead compound is generated by the trained molecular generation model to obtain multiple candidate molecules, the comprehensive optimization score and docking score of each candidate molecule are obtained, and a drug discovery molecule ranking list and rationale are generated.

[0124] Starting with the lead compound or seed compound input by the user and the lead compound output from the lead compound screening process, the trained molecular generation model (such as the molecular generation model of the REINVENT4 framework) is invoked. The algorithm based on policy gradient optimizes the structure of high-quality lead compounds to generate new molecules during the reinforcement learning process. Thousands of new molecules can be generated iteratively by adding atoms, deleting bonds, modifying rings, etc. Then, based on the binding free energy re-score, L candidate molecules with the highest total score (e.g., L=10) are selected. Based on the comprehensive optimization score and docking score of each candidate molecule, the final ranking list of drug discovery molecules and the reasons are generated.

[0125] In some implementations, the steps of generating molecules from lead compounds using a trained molecular generation model to obtain multiple candidate molecules, acquiring a comprehensive optimization score and docking score for each candidate molecule, and generating a drug discovery molecule ranking list and rationale include: (1) Based on the reward score, each of the lead compounds is guided to perform reinforcement learning, and the newly generated optimized molecules are screened to obtain multiple candidate molecules; (2) The trained molecular ranking model is used to reason about the candidate molecules based on the comprehensive optimization score and docking score to obtain the ranking sequence and reason of the drug discovery molecules, and each drug discovery molecule and reason is displayed in sequence according to the ranking sequence.

[0126] The reward score is determined by a pre-defined composite reward function. Calculations are performed to guide the optimization direction of each lead compound during the reinforcement learning process, and then the reward function is selected from the thousands of optimized molecules generated by the optimization. Highest total score Ten molecules (e.g., the first 10) are selected as candidate molecules for further physical verification. During verification, binding free energy rescoring is performed using both molecular docking and MM-GBSA (molecular mechanics / generalized Born surface area) methods to obtain the MM-GBSA binding energy score, i.e., the binding free energy rescoring score. Docking evaluation is a physical simulation calculation that uses molecular docking tools (such as AutoDock Vina, Glide, or SBDD) to simulate the three-dimensional conformation of the small molecule binding to the protein pocket and calculate the Gibbs free energy. The lower the energy (the larger the negative value), the more stable the binding. This is a more accurate but more time-consuming physical scoring method than the previous DTI model prediction.

[0127] For the selected candidate molecules, a trained molecular ranking model (such as LLM) integrates its expertise learned from a vast amount of scientific literature to comprehensively weigh and judge the overall optimization score of each candidate molecule (including docking score, MM-GBSA binding energy, ADMET prediction details, and synthetic route feasibility) against the docking score. The trained molecular ranking model, by reading all the above numerical values ​​and scientific literature background, outputs a comprehensive ranking and rationale through logical reasoning, thus obtaining the ranking sequence and rationale for the drug discovery molecules, and displaying the ranking sequence and rationale for the drug discovery molecules.

[0128] In this way, a composite reward function guides the optimization direction of each lead compound during reinforcement learning, screening high-potential candidate molecules from the newly generated optimized molecules. Reliable scores are then obtained through precise physical verification such as molecular docking and MM-GBSA binding energy reassessment. Finally, a trained molecular ranking model (such as LLM) integrates multi-dimensional core data (docking score, binding free energy, ADMET details, synthetic feasibility) with a wealth of literature expertise for comprehensive evaluation and logical reasoning. This approach achieves multi-layered checks from initial screening to precise verification, improving the scientific rigor and reliability of molecular assessment. Furthermore, through intelligent integration and interpretable reasoning, it clarifies the molecular ranking criteria, effectively avoiding the one-sidedness of pure data quantification. This provides a molecular list and decision support with both high performance and feasibility for subsequent drug development, reducing R&D risks and trial-and-error costs.

[0129] In some implementations, the calculation process for reward points includes: (1.1) The affinity of each optimized molecule is evaluated by the trained affinity evaluation model to obtain the affinity score of each optimized molecule; (1.2) Obtain the quantitative drug-likeness score and synthesis accessibility score for each of the optimized molecules; (1.3) Calculate the structural similarity between each of the optimized molecules and the lead compound to obtain the structural similarity of each of the optimized molecules; (1.4) The affinity score, quantitative drug-likeness score, synthetic accessibility score and structural similarity of each optimized molecule are combined to obtain the reward score of each optimized molecule.

[0130] Each optimized molecule was evaluated from four dimensions: affinity, quantitative drug-likeness, synthetic accessibility, and structural similarity to the lead compound. Affinity score, quantitative drug-likeness score, synthetic accessibility score, and structural similarity score were obtained. By combining the affinity score, quantitative drug-likeness score, synthetic accessibility score, and structural similarity score of each optimized molecule, a reward score was obtained for each optimized molecule.

[0131] Therefore, by comprehensively evaluating and integrating four dimensions—affinity (molecular activity), quantitative drug-likeness (drug potential), synthetic accessibility (manufacturing feasibility), and structural similarity with lead compounds (effective structural continuity)—a reward score is obtained. This approach covers the core attributes of drug development and ensures, through structural similarity indicators, that molecule optimization is based on the effective structure of lead compounds, avoiding the blind generation of molecules that deviate from the activity base. Based on the reward score of each optimized molecule, high-value candidate molecules possessing activity, drug potential, manufacturability, and structural rationality can be ultimately screened, laying a solid foundation for subsequent precise validation and ranking, and improving the efficiency and quality of the drug discovery process.

[0132] In some implementations, a reward score for each optimized molecule is obtained by fusing its affinity score, quantitative drug-likeness score, synthetic accessibility score, and structural similarity score, including: (1.1) Determine the difference between the maximum synthesis accessibility score and the synthesis accessibility score of each candidate molecule to obtain the synthesis inaccessibility score of each optimized molecule; (1.2) The affinity score, quantitative drug-likeness score, synthetic inaccessibility score and structural similarity score of each candidate molecule are weighted and summed to obtain the reward score of each optimized candidate molecule.

[0133] Among them, the composite reward function The formula is as follows: ; in, The affinity score predicted by the trained affinity assessment model (such as the DTI model). This is a quantitative drug property score, with a value range of (0-1), to ensure that the molecule possesses good physicochemical properties. A synthetic accessibility score is assigned, ranging from 1 to 10, to ensure that the molecule is easy to synthesize. To score the structural similarity with the original lead compound and prevent the generated structure from deviating too far. The weighting of affinity score The weighting of quantitative drug property scores, Synthesis inaccessibility (i.e.) The weight of ) The weights for the structural similarity score are calculated. A reward score for each optimized molecule is obtained by weighting and summing the affinity score, quantitative drug-likeness score, synthetic inaccessibility score, and structural similarity score.

[0134] In this way, by transforming the synthesis accessibility score into synthesis inaccessibility (10- This approach uses a unified evaluation metric, combined with affinity (activity), quantitative drug-likeness (drug potential), and structural similarity to the lead compound (effective structural continuity) to construct a weighted summation composite reward function. This approach considers the core attributes of drug development while flexibly adapting to different development priorities through weight settings. Furthermore, the structural similarity index prevents optimized molecules from deviating from the effective structure of the lead compound during the reinforcement learning optimization process. Based on the reward score, high-quality candidate molecules with high activity, good drug potential, easy synthesis, and reasonable structure are ultimately selected from the optimized molecules, avoiding the generation of ineffective or unviable molecules. This lays a high-quality foundation for subsequent accurate validation and ranking, improving the efficiency of the drug discovery process and the practical value of candidate molecules.

[0135] Please see Figure 8 , Figure 8 This is a flowchart illustrating the lead compound optimization process provided in the embodiments of this application. The input is a set of lead compounds, including at least one lead compound, which can be a user-input set or a set of lead compounds obtained from the previous screening process. Through reinforcement learning techniques, multi-dimensional attribute optimization is performed on the lead compounds based on a composite reward function, focusing on improving core attributes such as drug-likeness (drug potential) and syntheticability (production feasibility), making the optimized molecules more aligned with the actual needs of drug development. Weighted scores are calculated based on the optimized molecular attributes, and the top K samples (sample leads) are selected to initially identify high-potential candidate molecules. Docking calculations are performed on the candidate molecules to obtain a docking score for each candidate molecule's binding pocket to the target, evaluating the binding suitability between the candidate molecule and the target. The corresponding attribute scores (optimization stage scores) and docking scores of the optimized molecules are input into a Large Language Model (LLM) for comprehensive analysis. The LLM integrates multi-dimensional data and combines its learned drug development expertise to comprehensively rank the candidate molecules. The final set of proposed lead compounds is output, representing high-quality candidate molecules after multiple rounds of optimization, validation, and intelligent decision-making. By enhancing the core molecular properties of lead compounds through reinforcement learning, ensuring binding compatibility through docking verification, and leveraging the comprehensive decision-making capabilities of LLM to avoid the limitations of a single indicator, high-quality lead compounds that balance drug-likeness, syntheticity, and target binding affinity are ultimately obtained, providing more reliable candidate molecules for subsequent preclinical studies.

[0136] Please see Figure 9 , Figure 9This is a schematic diagram of the structure of a drug molecule discovery device provided in an embodiment of this application. The drug discovery device is applied to a server and may include a supervisory agent 901, a target identification unit 902, a lead compound discovery unit 903, a lead compound screening and lead compound selection unit 904, and a lead compound optimization unit 905, etc.

[0137] Supervisory agent 901 is used to extract entities from input text and perform intent recognition, and to schedule drug molecule discovery operation units related to the entities based on the results of the intent recognition. The drug molecule discovery operational unit includes one or more of the following units: Target identification unit 902: used to extract disease entity names from input text, and determine the target targets corresponding to the disease entity names and the target structure of the target targets; Lead compound discovery unit 903: used to detect binding pockets in target structures, determine ligand binding pockets in target structures, and obtain lead compounds that match the ligand binding pockets; Lead compound to lead compound screening unit 904: used to derivatize and expand the lead compound to obtain candidate derivatized compounds, and to screen a lead compound from each candidate derivatized compound based on a comprehensive score of each candidate derivatized compound; Lead compound optimization unit 905: Generates molecules from lead compounds using a trained molecular generation model, obtains multiple candidate molecules, acquires the comprehensive optimization score and docking score for each candidate molecule, and generates a drug discovery molecule ranking list and rationale.

[0138] The Supervisor Agent, built upon the Large Language Model (LLM), is an AI component with top-level coordination, decision-making, and oversight capabilities. Its core role in multi-agent systems or complex task processes is to achieve efficient task execution and risk control through goal decomposition, resource scheduling, behavior monitoring, and result integration. It does not directly perform specific business operations but acts as a "manager" or "scheduling hub," addressing issues such as multi-module collaboration, dynamic environment adaptation, and complex process governance.

[0139] In the intelligent scheduling of tasks in the drug discovery process, a supervisory agent acts as the overall control center of the system. This agent maintains a global state machine. Additionally, the supervisory agent is responsible for parsing the natural language input (i.e., input text) to perform intent recognition. Based on the intent recognition results, it schedules the drug discovery execution units associated with specific entities. Each drug discovery execution unit, once scheduled, executes its corresponding drug discovery process. The specific implementation process of each unit can be found in the previous implementation methods for drug discovery, and will not be repeated here.

[0140] As described above, the embodiments of this application extract entities from the input text and identify the intent of the input text. Based on the intent result, one or more processes in the drug molecule discovery process related to the entities are scheduled. This allows users to simply provide the input text, and the subsequent operation of the drug molecule discovery process will automatically schedule one or more processes in the drug molecule discovery process related to the entities in the input text based on the intent result of the input text. This avoids users manually importing and exporting data between different processes and improves the processing efficiency of drug molecule discovery.

[0141] Figure 10 This is a partial structural block diagram of a computer device 1000 for implementing embodiments of this application. The computer device 1000 can vary considerably due to differences in configuration or performance, and may include a central processing unit (CPU) 422, a memory 432, an application program storage 442, and a storage medium 430 (e.g., one or more mass storage devices), where data 444 is stored. Furthermore, the CPU 422 may be configured to communicate with the storage medium 430, and the CPU 422 may execute drug molecule discovery methods stored in the storage medium 430.

[0142] Computer device 1000 may also include one or more power supplies 425, one or more wired or wireless network interfaces 450, one or more input / output interfaces 458, and / or one or more operating systems 441, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, etc.

[0143] This application also provides a computer-readable storage medium for storing multiple instructions, which are adapted for a processor to load and execute the drug molecule discovery methods of the foregoing embodiments. The specific execution process has been described in detail in the above methods and will not be repeated here.

[0144] This application also provides a computer program product comprising a computer program. A processor of a computer device reads and executes the computer program, causing the computer device to perform the drug molecule discovery method described above.

Claims

1. A method for drug molecule discovery, characterized in that, include: Extract entities from the input text and perform intent recognition; based on the results of the intent recognition, schedule the drug molecule discovery process related to the entities. The drug molecule discovery process includes one or more of the following processes: Target identification: Extract disease entity names from input text, and determine the target targets corresponding to the disease entity names and the target structure of the target targets; Lead compound discovery: Binding pocket detection is performed on the target structure to determine the ligand binding pocket in the target structure and to obtain lead compounds that match the ligand binding pocket; From lead compounds to lead compounds: lead compounds are derivatized and expanded to obtain candidate derivatized compounds, and lead compounds are selected from each candidate derivatized compound based on the comprehensive score of each candidate derivatized compound; Lead compound optimization: The trained molecular generation model is used to generate molecules from the lead compound, resulting in multiple candidate molecules. The comprehensive optimization score and docking score of each candidate molecule are obtained, and a drug discovery molecule ranking list and rationale are generated.

2. The drug molecule discovery method according to claim 1, characterized in that, The drug molecule discovery process, which is scheduled based on the result of the intent recognition and associated with the entity, includes: When the result of the intent recognition is to run a single-process task, the target process related to the entity is scheduled and run from the target identification, lead compound discovery, lead compound screening to lead compound optimization and lead compound optimization in the drug molecule discovery process. When the result of the intent recognition is to run an end-to-end task, at least two processes in the drug molecule discovery process are run sequentially: target identification, lead compound discovery, lead compound screening to lead compound selection, and lead compound optimization.

3. The drug molecule discovery method according to claim 2, characterized in that, The process of sequentially running at least two steps in the drug molecule discovery workflow, namely target identification, lead compound discovery, lead compound screening to lead compound selection, and lead compound optimization, includes: When the target node of the current process is detected, the first execution result of the previous running node of the target node is obtained; The first execution result is standardized and verified, and a prompt is output asking whether to continue executing subsequent nodes based on the first execution result. When a continue execution instruction based on the first execution result is received, the subsequent nodes of the target node continue to run using the first execution result as input data.

4. The drug molecule discovery method according to claim 3, characterized in that, The method further includes: When a parameter correction instruction based on the first execution result is received, the first parameter carried by the parameter correction instruction is obtained; The first parameter is input to the previous running node of the target node, so that the previous running node of the target node re-runs according to the first parameter to obtain the second execution result; The second execution result is standardized and verified, and a prompt is output indicating whether to continue executing subsequent nodes based on the second execution result; If no continue execution instruction based on the second execution result is received, the execution of subsequent nodes of the target node shall be prohibited.

5. The drug molecule discovery method according to claim 3, characterized in that, The method further includes: When a manual editing instruction based on the first execution result is received, the second parameter carried by the manual editing instruction is obtained; The second parameter is used as input data to continue running subsequent nodes of the target node.

6. The drug molecule discovery method according to claim 3, characterized in that, The method further includes: When the current process is detected to have completed, the second execution result of the current process is obtained; Display the second execution result and output a prompt asking whether to continue the subsequent process based on the second execution result; When a continue execution instruction is received based on the second execution result, the subsequent processes of the current process continue to run.

7. The drug molecule discovery method according to claim 1, characterized in that, The steps of generating molecules from lead compounds using a trained molecular generation model to obtain multiple candidate molecules, acquiring a comprehensive optimization score and docking score for each candidate molecule, and generating a drug discovery molecule ranking list and rationale include: Based on the reward score, the trained molecular generation model performs reinforcement learning on each of the lead compounds, and selects multiple candidate molecules from the newly generated optimized molecules. The trained molecular ranking model is used to reason about the candidate molecules based on the comprehensive optimization score and docking score to obtain the ranking sequence and reasoning of drug discovery molecules, and each drug discovery molecule and reasoning is displayed in sequence according to the ranking sequence.

8. The drug molecule discovery method according to claim 7, characterized in that, The calculation process for the reward points includes: The affinity of each optimized molecule is evaluated using the trained affinity assessment model to obtain an affinity score for each optimized molecule. Obtain a quantitative drug-likeness score and a synthesis accessibility score for each of the optimized molecules; Calculate the structural similarity between each optimized molecule and the lead compound to obtain the structural similarity of each optimized molecule; The affinity score, quantitative drug-likeness score, synthetic accessibility score, and structural similarity score of each candidate molecule are combined to obtain a reward score for each optimized molecule.

9. The drug molecule discovery method according to claim 8, characterized in that, The reward score for each optimized molecule is obtained by fusing its affinity score, quantitative drug-likeness score, synthetic accessibility score, and structural similarity score, including: The difference between the maximum synthesis accessibility score and the synthesis accessibility score of each optimized molecule is determined to obtain the synthesis inaccessibility score of each optimized molecule. The affinity score, quantitative drug-likeness score, synthetic inaccessibility score, and structural similarity score of each optimized molecule are weighted and summed to obtain the reward score for each optimized molecule.

10. The method for drug molecule discovery according to claim 1, characterized in that, Before selecting a lead compound from each of the candidate derived compounds based on a comprehensive score, the method further includes: The affinity of each candidate derivative compound is evaluated using a trained affinity assessment model to obtain an affinity score for each candidate derivative compound. Quantitative drug-likeness scores and synthetic accessibility scores were obtained for each of the candidate derivative compounds; The toxicity risk of each candidate derivative compound is assessed using a trained toxicity risk assessment model to obtain a toxicity risk score for each candidate derivative compound. A comprehensive score for each candidate derivative compound is obtained by combining its affinity score, quantitative drug-likeness score, synthetic accessibility score, and toxicity risk score.

11. The drug molecule discovery method according to claim 10, characterized in that, The affinity score, quantitative drug-likeness score, synthetic accessibility score, and toxicity risk score of each candidate derivative compound are integrated to obtain a comprehensive score for each candidate derivative compound, including: For each of the candidate derivative compounds, the corresponding affinity score, quantitative drug-likeness score, and synthetic accessibility score are weighted and summed to obtain the first calculation result for each candidate derivative compound; The difference between the first calculated result and the corresponding toxicity risk score for each candidate derivative compound is determined to obtain a comprehensive score for each candidate derivative compound.

12. The drug molecule discovery method according to claim 1, characterized in that, The process of determining the target point corresponding to the disease entity name includes: The disease entity name is mapped to the corresponding disease code, the literature corresponding to the disease code is retrieved, and the literature is input into the trained target screening model so as to select candidate targets from the literature through the target screening model. The authority score of each candidate target is determined based on the document type of the document in which each candidate target is located. Determine the similarity distance between each candidate target and the disease code, and determine the relevance score for each candidate target; Based on the number of documents in which each candidate target is located, a literature basis score is determined for each candidate target. The authority score, relevance score, and literature reference score of each candidate target are weighted and summed to obtain the confidence score of each candidate target. Based on the confidence score, target targets are selected from each of the candidate targets.

13. The method for drug molecule discovery according to any one of claims 1 to 12, characterized in that, The acquisition of a spur compound that matches the ligand binding pocket includes: The trained molecular generation model is used to generate molecules at the ligand binding pocket to obtain a first number of ligand molecules. The affinity of a second number of ligand molecules in a preset molecule set to the ligand binding pocket is evaluated using a trained affinity assessment model, and an affinity score is obtained for each of the second number of ligand molecules. A third number of ligand molecules are selected from the second number of ligand molecules based on the affinity score. The first number of ligand molecules and the third number of ligand molecules are subjected to deduplication to obtain a fourth number of ligand molecules; A confidence score is determined for each of the fourth number of ligand molecules, and a seed compound matching the ligand binding pocket is screened from the fourth number of ligand molecules according to the confidence score.

14. The drug molecule discovery method according to claim 1, characterized in that, The process of running a drug molecule discovery workflow related to the entity based on the result of the intent recognition includes: Based on the entity and the result of the intent recognition, a task instruction is generated; After the task instructions are standardized, a tool executor that is uniformly encapsulated into a preset format interface is invoked so that the corresponding tool executor can perform task processing. Obtain the processing result of the task instruction executed by the tool executor and perform standardization processing.

15. The drug molecule discovery method according to claim 14, characterized in that, The step of standardizing the task instructions and then calling a tool executor that is uniformly encapsulated into a preset format interface, so that the corresponding tool executor can perform task processing, includes: The task instructions are serialized into a task request in a preset format; When the task request requires the use of the local first tool executor, the preset encapsulated interface is invoked to send the task request to the first tool executor so that the first tool executor can perform task processing. When the task request requires the use of a remote second tool executor, a preset encapsulation interface is invoked and the task request is sent to the second tool executor through a preset secure network channel so that the second tool executor can perform task processing.

16. A drug molecule discovery device, characterized in that, include: A supervisory agent is used to extract entities from input text and perform intent recognition, and to schedule drug molecule discovery operation units related to the entities based on the results of the intent recognition. The drug molecule discovery operation unit includes one or more of the following units: Target identification unit: used to extract disease entity names from input text, and determine the target targets corresponding to the disease entity names and the target structure of the target targets; Lead compound discovery unit: used to detect binding pockets in the target structure, determine the ligand binding pockets in the target structure, and obtain lead compounds that match the ligand binding pockets; Lead compound to lead compound screening unit: used to derivatize and expand the lead compound to obtain candidate derivatized compounds, and to screen lead compounds from each candidate derivatized compound based on the comprehensive score of each candidate derivatized compound; Lead compound optimization unit: Generates molecules from lead compounds using a trained molecular generation model, obtains multiple candidate molecules, acquires the comprehensive optimization score and docking score for each candidate molecule, and generates a drug discovery molecule ranking list and rationale.

17. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of instructions adapted for loading by a processor to execute the drug molecule discovery method according to any one of claims 1 to 15.

18. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the drug molecule discovery method according to any one of claims 1 to 15.

Images