Target point determination method, electronic device, storage medium, and program product

Abstract

This application provides a target determination method, electronic device, storage medium, and program product. The method includes: receiving biomolecular information of an unknown biomolecule; performing contextual parsing on the biomolecular information to obtain multiple valid text segments, where each text segment represents key relevance information supporting the determination of the unknown biomolecule's target; concatenating the multiple valid text segments using a reordering module to obtain valid biological text corresponding to the unknown biomolecule; performing reasoning processing on the valid biological text based on a preset logic to obtain an initial reasoning result corresponding to the unknown biomolecule; and verifying the initial reasoning result to obtain a target reasoning result corresponding to the unknown biomolecule, where the target reasoning result indicates the potential drug target attributes of the unknown biomolecule. This method aims to improve the reliability of the reasoning results.

Description

Technical Field

[0001] This application relates to the biomedical field, and more particularly to a target determination method, electronic device, storage medium, and program product. Background Technology

[0002] In the biomedical field, it is necessary to analyze the potential drug target properties of corresponding biomolecules based on biomolecular information.

[0003] In existing technologies, the analysis task of unknown biomolecules is usually broken down into several sub-tasks. Then, for each sub-task, corresponding biomedical databases, computing tools, and literature analysis methods are called to obtain the required intermediate results and integrate them. Finally, based on preset artificial rules or preset algorithm scheduling, the potential drug targets of the biomolecule are inferred.

[0004] Since each subtask relies on independent databases, tools, or analysis methods, the intermediate results output by different subtasks often lack unified semantic associations and logical connections, which can easily lead to a break in the reasoning chain, resulting in low reliability of the final potential drug target reasoning results. Summary of the Invention

[0005] This application provides a target determination method, electronic device, storage medium, and program product to improve the reliability of inference results.

[0006] In a first aspect, embodiments of this application provide a target determination method, including:

[0007] Receive biomolecular information from unknown biomolecules;

[0008] The biomolecule information is subjected to context parsing to obtain multiple valid text segments, which are key relevant information supporting the determination of the target of the unknown biomolecule.

[0009] The reordering module concatenates the multiple valid text segments to obtain the valid biological text corresponding to the unknown biomolecule.

[0010] Based on a pre-defined thinking logic, the effective biological text is processed by reasoning to obtain the initial reasoning result corresponding to the unknown biomolecule. The pre-defined thinking logic is used to guide the reasoning process of the pre-defined reasoning model.

[0011] The initial inference result is verified to obtain the target inference result corresponding to the unknown biomolecule. The target inference result is used to indicate the potential drug target properties of the unknown biomolecule.

[0012] In one possible implementation,

[0013] The biomolecular information is subjected to context parsing to obtain multiple valid text segments, including:

[0014] A smart agent template is obtained, which is used to provide a framework for task analysis, intent analysis, query rethinking, and combined query for the determination of the target of the unknown biomolecule;

[0015] Based on the intelligent agent template, the biomolecular information is subjected to context parsing to obtain the multiple valid text segments;

[0016] The task analysis is used to extract key elements from biomolecular information and form a task chain by logically sorting them according to basic features, functional inference, and target potential assessment; the intent analysis is used to determine the task intent based on the task chain, focusing on exploring the potential of biomolecules as therapeutic targets for specific diseases, and obtaining intent analysis results; the query rethinking is used to generate and optimize a query set based on the intent analysis results; and the combined query is used to combine the query set according to logical association and priority, and initiate a search in the biomedical database to obtain multi-source related information.

[0017] In one possible implementation, the reordering module includes a rule-driven module or an algorithm-driven module; it concatenates the multiple valid text segments to obtain the valid biological text corresponding to the unknown biomolecule, including:

[0018] The sorting sequence number corresponding to each valid text segment is determined through rule-driven or algorithm-driven modules.

[0019] Based on the sorting number corresponding to each valid text segment, the multiple valid text segments are concatenated to obtain the valid biological text;

[0020] The rule-driven module determines the key inference dimensions corresponding to each valid text segment based on the dimension importance ranking rule, and determines the corresponding ranking number based on the key inference dimensions corresponding to each valid text segment; the algorithm-driven module determines the weight corresponding to each valid text segment, and determines the corresponding ranking number based on the weight corresponding to each valid text segment.

[0021] In one possible implementation,

[0022] Based on a pre-defined logic, the valid biological text is subjected to reasoning processing to obtain the initial reasoning result corresponding to the unknown biomolecule, including:

[0023] A reasoning logic model is obtained, which is a pre-trained model obtained by training an initial logic model with domain knowledge of the target set. The domain knowledge of the target set includes known molecular information of multiple known biomolecules and their corresponding potential drug target attributes.

[0024] The reasoning logic model is used to process the preset thinking logic and the effective biological text to obtain the initial reasoning result.

[0025] In one possible implementation,

[0026] Obtain the reasoning logic model, including;

[0027] The target analysis logic for biomolecules and multiple biological knowledge documents are obtained. The target analysis logic is used to standardize and organize gene and protein information related to biopharmaceutical targets in the biological knowledge documents.

[0028] Through the target analysis logic, the multiple biological knowledge documents are analyzed and processed to obtain multiple known biological texts corresponding to multiple known biomolecules, as well as the known drug target attributes corresponding to each known biomolecule.

[0029] The initial logic model is trained using the known biological text corresponding to each known biomolecule and the known drug target attributes to obtain the inference logic model.

[0030] In one possible implementation,

[0031] The target analysis logic includes standardized reconstructed information, confidence-adjusted cases, and public medical question-and-answer and targeted literature question-and-answer. For any given biological knowledge document, the target analysis logic analyzes and processes the biological knowledge document to obtain the known biological text and drug target attributes of the known biomolecules corresponding to the biological knowledge document, including:

[0032] A biomedical pre-trained model is used to extract multiple known text segments and known target types from the biological knowledge literature, and then the multiple known text segments are concatenated into the known biological text, which includes known biological sequences, known functions, and known expression data of known biomolecules.

[0033] Based on the standardized reconstruction information, the identified multiple known text segments are mapped into a structured data format;

[0034] Based on the confidence adjustment case, obtain the known preliminary confidence level corresponding to the molecular target association relationship between the known biological text and the known target type;

[0035] Using the reasoning logic provided by the aforementioned question-and-answer chain and targeted literature analysis, the initial confidence level corresponding to the known biological text is corrected to obtain the known target confidence level;

[0036] The known target type, known target confidence level, and the literature source of the biological knowledge literature are determined as the attributes of the known drug target.

[0037] In one possible implementation, the target analysis logic for obtaining biomolecules includes:

[0038] Obtain a preset synthesis template, which includes standardized reconstruction rules, confidence assessment rules, and medical question-and-answer generation rules for biomolecular analysis;

[0039] Using a large language model, initial analysis logic is generated based on the biological knowledge literature and the preset synthesis template;

[0040] The initial analysis logic is validated through cross-evaluation using a large language model and expert calibration to obtain the validation results.

[0041] If the verification result is successful, the initial analysis logic is determined as the target analysis logic.

[0042] Secondly, embodiments of this application provide a target determination device, including a receiving module, a parsing processing module, a splicing processing module, an inference processing module, and a verification processing module, wherein...

[0043] The receiving module is used to receive biomolecular information of unknown biomolecules;

[0044] The parsing and processing module is used to perform context parsing processing on the biomolecule information to obtain multiple valid text segments, which are key relevant information supporting the determination of the target of the unknown biomolecule.

[0045] The splicing processing module is used to splice the multiple valid text segments to obtain the valid biological text corresponding to the unknown biomolecule.

[0046] The reasoning processing module is used to perform reasoning processing on the effective biological text based on a preset thinking logic to obtain the initial reasoning result corresponding to the unknown biomolecule. The preset thinking logic is used to guide the reasoning process of the preset reasoning model.

[0047] The verification processing module is used to verify the initial inference result to obtain the target inference result corresponding to the unknown biomolecule. The target inference result is used to indicate the potential drug target attributes of the unknown biomolecule.

[0048] In one possible implementation, the parsing processing module is specifically used for:

[0049] A smart agent template is obtained, which is used to provide a framework for task analysis, intent analysis, query rethinking, and combined query for the determination of the target of the unknown biomolecule;

[0050] Based on the intelligent agent template, the biomolecular information is subjected to context parsing to obtain the multiple valid text segments;

[0051] The task analysis is used to extract key elements from biomolecular information and form a task chain by logically sorting them according to basic features, functional inference, and target potential assessment; the intent analysis is used to determine the task intent based on the task chain, focusing on exploring the potential of biomolecules as therapeutic targets for specific diseases, and obtaining intent analysis results; the query rethinking is used to generate and optimize a query set based on the intent analysis results; and the combined query is used to combine the query set according to logical association and priority, and initiate a search in the biomedical database to obtain multi-source related information.

[0052] In one possible implementation, the reordering module includes a rule-driven module or an algorithm-driven module; the concatenation processing module is specifically used for:

[0053] The sorting sequence number corresponding to each valid text segment is determined through rule-driven or algorithm-driven modules.

[0054] Based on the sorting number corresponding to each valid text segment, the multiple valid text segments are concatenated to obtain the valid biological text;

[0055] The rule-driven module determines the key inference dimensions corresponding to each valid text segment based on the dimension importance ranking rule, and determines the corresponding ranking number based on the key inference dimensions corresponding to each valid text segment; the algorithm-driven module determines the weight corresponding to each valid text segment, and determines the corresponding ranking number based on the weight corresponding to each valid text segment.

[0056] In one possible implementation, the inference processing module is specifically used for:

[0057] A reasoning logic model is obtained, which is a pre-trained model obtained by training an initial logic model with domain knowledge of the target set. The domain knowledge of the target set includes known molecular information of multiple known biomolecules and their corresponding potential drug target attributes.

[0058] The reasoning logic model is used to process the preset thinking logic and the effective biological text to obtain the initial reasoning result.

[0059] In one possible implementation, the inference processing module is specifically used for;

[0060] The target analysis logic for biomolecules and multiple biological knowledge documents are obtained. The target analysis logic is used to standardize and organize gene and protein information related to biopharmaceutical targets in the biological knowledge documents.

[0061] Through the target analysis logic, the multiple biological knowledge documents are analyzed and processed to obtain multiple known biological texts corresponding to multiple known biomolecules, as well as the known drug target attributes corresponding to each known biomolecule.

[0062] The initial logic model is trained using the known biological text corresponding to each known biomolecule and the known drug target attributes to obtain the inference logic model.

[0063] In one possible implementation, the target analysis logic includes standardized reconstructed information, confidence-adjusted cases, and public medical question answering and targeted literature question answering; for any biological knowledge literature; the reasoning processing module is specifically used for:

[0064] A biomedical pre-trained model is used to extract multiple known text segments and known target types from the biological knowledge literature, and then the multiple known text segments are concatenated into the known biological text, which includes known biological sequences, known functions, and known expression data of known biomolecules.

[0065] Based on the standardized reconstruction information, the identified multiple known text segments are mapped into a structured data format;

[0066] Based on the confidence adjustment case, obtain the known preliminary confidence level corresponding to the molecular target association relationship between the known biological text and the known target type;

[0067] Using the reasoning logic provided by the aforementioned question-and-answer chain and targeted literature analysis, the initial confidence level corresponding to the known biological text is corrected to obtain the known target confidence level;

[0068] The known target type, known target confidence level, and the literature source of the biological knowledge literature are determined as the attributes of the known drug target.

[0069] In one possible implementation, the inference processing module is specifically used for:

[0070] Obtain a preset synthesis template, which includes standardized reconstruction rules, confidence assessment rules, and medical question-and-answer generation rules for biomolecular analysis;

[0071] Using a large language model, initial analysis logic is generated based on the biological knowledge literature and the preset synthesis template;

[0072] The initial analysis logic is validated through cross-evaluation using a large language model and expert calibration to obtain the validation results.

[0073] If the verification result is successful, the initial analysis logic is determined as the target analysis logic.

[0074] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0075] The memory stores computer-executed instructions;

[0076] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0077] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0078] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0079] The target determination method, electronic device, storage medium, and program product provided in this application can receive biomolecular information of unknown biomolecules, perform context parsing on the biomolecular information to obtain multiple text segments, and concatenate these multiple text segments to obtain valid biological text corresponding to the unknown biomolecule. Based on preset reasoning logic, the valid biological text is used for reasoning to obtain an initial reasoning result. The initial reasoning result is then verified to obtain the target reasoning result corresponding to the unknown biomolecule. Reasoning is directly performed on the integrated valid biological text, eliminating the need to break down the reasoning into multiple isolated subtasks, and the verification process ensures the reliability of the results, thereby improving the accuracy and credibility of the reasoning results for the potential drug target attributes of unknown biomolecules. Attached Figure Description

[0080] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0081] Figure 1 A schematic diagram illustrating the application scenarios provided in the embodiments of this application;

[0082] Figure 2 A flowchart illustrating the target determination method provided in this application embodiment;

[0083] Figure 3 A schematic diagram of the architecture of a target determination method provided in an embodiment of this application;

[0084] Figure 4 A flowchart illustrating the determination of a reasoning logic model is provided for an embodiment of this application;

[0085] Figure 5 A schematic diagram of the architecture for determining target analysis logic provided in an embodiment of this application;

[0086] Figure 6 A schematic diagram of an architecture for training a reasoning logic model provided in an embodiment of this application;

[0087] Figure 7 This is a schematic diagram of the architecture of a target determination device provided in an embodiment of this application;

[0088] Figure 8 This is a schematic diagram of the structure of an electronic device provided in this application.

[0089] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0090] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0091] Figure 1 This is a schematic diagram illustrating an application scenario provided in an embodiment of this application. Please refer to [link / reference]. Figure 1 As shown, this application scenario includes a biomedical database 101 and a processing device 102, and the application scenario can be used for drug target discovery and evaluation in the biomedical field.

[0092] Biomedical databases can store known molecular information about known biomolecules. For example, databases such as the Kyoto Encyclopedia of Genes and Genomes (KEGG) and GenePath contain information related to gene function, protein interactions, and disease pathways.

[0093] The processing device 102 can obtain known molecular information and drug target attributes of known biomolecules from the biomedical database 101. The processing device 102 can then analyze the potential drug target attributes of unknown biomolecules using the known molecular information and their corresponding drug target attributes.

[0094] In existing technologies, the analysis of unknown biomolecules can be decomposed into several sub-tasks. Then, for each sub-task, corresponding biomedical databases, computational tools, and literature analysis methods are invoked to obtain the required intermediate results, which are then integrated. By integrating the intermediate results from each sub-task, potential drug targets are inferred based on pre-set manual rules or algorithms. However, because each sub-task relies on independent databases, tools, or analysis methods, the intermediate results often lack unified semantic connections and logical coherence, leading to broken inference chains and low reliability of the inference results.

[0095] The target determination method provided in this application can receive biomolecular information of an unknown biomolecule, perform context parsing on the biomolecular information to obtain multiple text segments, which are key relevance information supporting target determination; concatenate the multiple text segments to obtain valid biological text corresponding to the unknown biomolecule; perform reasoning processing on the valid biological text based on preset thinking logic to obtain an initial reasoning result corresponding to the unknown biomolecule; and perform verification processing on the initial reasoning result to obtain a target reasoning result corresponding to the unknown biomolecule, which is used to indicate the potential drug target attributes of the unknown biomolecule.

[0096] The above execution process can extract valid biological text from the biomolecular information of unknown biomolecules, and then directly perform reasoning processing on the valid biological text based on the preset thinking logic to determine the initial reasoning result corresponding to the unknown biomolecule. There is no need to split the reasoning into multiple sub-tasks for separate processing, which can improve the reliability of the reasoning result.

[0097] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0098] Figure 2 This is a flowchart illustrating the target determination method provided in an embodiment of this application. Please refer to... Figure 2 The method may include:

[0099] S201, Receive biomolecular information from unknown biomolecules.

[0100] The execution entity in this application embodiment can be a processing device or a target determination device disposed in the processing device. The target determination device can be implemented by software or by a combination of software and hardware.

[0101] Unknown biomolecules are biomolecules to be deduced, and their biomolecular information can include corresponding core attributes based on the molecular type (such as protein, nucleic acid, etc.). For example, biomolecular information can include molecular structure, sequence information, expression site, known interacting molecules, etc.

[0102] For example, if the unknown biomolecule is an unknown protein, its biomolecular information usually includes the protein sequence (such as amino acid sequence), sequence homology information, potential structural domains (such as enzyme active sites, receptor binding regions), and known tissue expression profiles (such as whether it is highly expressed in diseased tissues).

[0103] S202. Perform context parsing on the biomolecular information to obtain multiple valid text segments.

[0104] Among the input biomolecular information, there may be text segments that are irrelevant to target identification. Multiple valid text segments related to drug target attributes can be extracted from the biomolecular information, and then irrelevant text segments can be eliminated.

[0105] Effective text segments can provide crucial relevant information to support the identification of targets for unknown biomolecules. For example, effective text segments can include the biological sequence of the unknown biomolecule, potential interacting proteins, and the signaling pathway to which it belongs.

[0106] In some possible embodiments, multiple valid text segments can be identified from biomolecular information using multiple preset keywords. Specifically, biomolecular information may include multiple molecular segments, and multiple biological keywords can be used to match each molecular segment to determine the relevance of each segment. Molecular segments with a relevance greater than a preset threshold are identified as valid text segments.

[0107] Multiple biological keywords can include gene ID, amino acid sequence, three-dimensional structure, subcellular localization, coding gene, homologous protein, catalytic activity, receptor binding, signaling pathway, interaction, etc.

[0108] In other embodiments, an agent template can be obtained; based on the agent template, contextual parsing processing is performed on the biomolecular information to obtain multiple valid text segments. The agent template provides a framework for task analysis, intent extraction, query rethinking, and combinatorial decomposition to identify targets of unknown biomolecules.

[0109] To meet the urgent need to integrate heterogeneous biological knowledge to achieve precise target discovery, the "Agentic Dynamic Context Module" can intelligently retrieve and prioritize relevant evidence.

[0110] Considering that unprocessed raw query statements often lack specificity in the biomedical field, the "Agent Template" first analyzes the raw query through four refinement steps: Task Analysis, Intent Extraction, Semantic Rewriting, and Combinatorial Decomposition.

[0111] Task analysis can break down the target identification of unknown biomolecules and clarify the core sub-tasks. For example, key elements (such as whether the protein sequence contains a special motif, its protein family, potential disease association, etc.) can be extracted from biomolecular information and sorted logically according to "basic characteristics, functional inference, and target potential assessment" to form a task chain.

[0112] Intent analysis can be used to determine the results based on the task chain. For example, if an unknown protein originates from a lung cancer sample, the intent may focus on exploring its potential as a therapeutic target for lung cancer; if it comes from the field of metabolic diseases, the focus may shift to analyzing its regulatory role in metabolic pathways to determine whether it is a target for metabolic diseases.

[0113] Query rewriting, also known as query rethinking, involves generating an initial query based on intent analysis results, then reflecting on and optimizing it (e.g., checking if all intent dimensions are covered, adjusting keywords, and revising retrieval logic) to obtain an optimized query set. This optimized query set can include multiple initial text segments and the corresponding query results for each initial text segment.

[0114] Combinatorial decomposition, also known as combined query, combines optimized query sets according to logical associations and priorities, and initiates a search in a database / literature repository to obtain multi-source association information between the initial text segments in the optimized query set. For example, queries such as "sequence homology," "disease phenotype association," and "known interacting molecules" can be executed simultaneously, and the results can be weighted, with homology results between initial text segments having a higher weight than indirect literature mentions.

[0115] Finally, the K most valuable pieces of information are selected through filtering and ranking, which constitute the multiple valid text segments. A dual-scoring re-ranking mechanism can be used to select the K most valuable pieces of information, combining syntactic relevance and semantic similarity. The top-1 contexts are then selected, and the semantic similarity is determined using the following formula:

[0116]

[0117] in, Represents the first in the optimized query set An initial text segment (i.e., candidate options). This represents the user's query. The score of the entire formula is obtained by weighted summation of the left and right parts. It is a key balancing weight used to adjust the importance of the two parts.

[0118] To achieve discrete hit matching, Representing the An initial text segment The number of candidate keywords in the candidate keyword set yes The One candidate keyword. Indicates from the query The set of query keywords extracted from it. It's a statistical function representing the intersection of the candidate keyword set and the query keyword set; this part is essentially statistical. How many keywords are there? Exact match.

[0119] For cosine similarity, It is the first An initial text segment semantic vectors, This is to query Through parameters The model is transformed into query vectors. The formula calculates the dot product of the two vectors and divides it by the product of their magnitudes to obtain their semantic similarity. The two scores are then weighted and added together to obtain... (i.e., semantic similarity), and the top K effective text segments are obtained by sorting them by semantic similarity.

[0120] In this application, biomolecular information is processed through contextual analysis using an intelligent agent template to obtain multiple text segments. This allows for the precise selection of key information that is highly relevant to the core intent of target identification, thereby improving the accuracy and reliability of subsequent inferences based on these text segments.

[0121] S203. The reordering module concatenates multiple valid text segments to obtain the valid biological text corresponding to the unknown biomolecule.

[0122] In some embodiments, a rule-driven module or an algorithm-driven module is used to determine the sorting number corresponding to each text segment; based on the sorting number corresponding to each text segment, multiple text segments are concatenated to obtain valid biological text.

[0123] The rule-driven module is based on the dimension importance ranking rules formulated by human-preset key dimensions of reasoning, without the need for complex algorithms.

[0124] For example, key dimensions for reasoning can include disease-associated evidence, molecular functional information, and basic structural features. The order of importance for dimensions is: disease-associated evidence > molecular functional information, molecular functional information > basic structural features, and if a valid text segment belongs to a higher priority dimension, its ranking number is higher.

[0125] Algorithm-driven modules automatically evaluate the weights of text segments based on machine learning or pre-trained models, making them suitable for complex biomolecular information scenarios. For example, pre-trained models in the biomedical field (such as BioBERT) can be used to semantically encode effective text segments, and the semantic similarity between each effective text segment and the drug target attributes can be calculated. The higher the similarity, the greater the weight and the higher the ranking number. Alternatively, classification models trained on historical target data can be used to directly predict the contribution of effective text segments to the inference results. The greater the contribution, the greater the weight and the higher the ranking number.

[0126] Below, using a rule-driven module as an example, we will determine the valid biological text:

[0127] There are three valid text segments: text segment A, "Unknown protein X is associated with lung cancer cell proliferation"; text segment B, "The isoelectric point of unknown protein X is 7.2"; and text segment C, "Unknown protein X can bind to known target protein Y". The key inference dimensions corresponding to each text segment are determined: text segment A represents disease-related evidence, text segment B represents "basic structural features", and text segment C represents "molecular interaction information".

[0128] Based on the dimensional importance ranking rule of disease association evidence > molecular interaction information > basic structural features, text segments are prioritized: text segment A is ranked number 1; text segment C is ranked number 2; and text segment B is ranked number 3. The text segments are then concatenated sequentially according to their ranking numbers, with logical connectors added to avoid semantic abruptness. For example: First, unknown protein X is associated with lung cancer cell proliferation; second, unknown protein X can bind to known target protein Y; furthermore, the isoelectric point of unknown protein X is 7.2. The output of valid biological text checks for duplicate information or logical contradictions in the concatenated text (if no issues are found), and finally outputs the aforementioned coherent text as input data for subsequent target determination.

[0129] In this application, a rule-driven module or an algorithm-driven module is used to determine the sorting number corresponding to each valid text segment, and multiple valid text segments are spliced ​​together to obtain valid biological text. This can extract key information from the biomolecular information of unknown biomolecules that is conducive to target determination, thereby improving the accuracy of subsequent target determination.

[0130] S204. Based on the preset thinking logic, perform reasoning processing on the valid biological text to obtain the initial reasoning result corresponding to the unknown biomolecule.

[0131] In some embodiments, a reasoning logic model can be obtained; through the reasoning logic model, the preset thinking logic and valid biological text are reasoned to obtain an initial reasoning result.

[0132] The reasoning logic model can be a pre-trained model obtained by training the initial logic model with knowledge of the target set domain. The initial logic model learns specific knowledge of the biopharmaceutical target domain to acquire the ability to mine drug targets from biological information, thus obtaining the reasoning logic model.

[0133] The initial logical model can be a model architecture based on a large language model, combined with multiple Target Modules to build a training network model.

[0134] During the training of the initial logical model, supervised fine-tuning (SFT) can be performed using LoRA (low-rank adaptation). LoRA is an efficient model fine-tuning method that adjusts model parameters by introducing a low-rank matrix, enabling the model to quickly adapt to knowledge of biopharmaceutical targets without significantly increasing computational cost.

[0135] The target adaptation module can contain multiple structural units of multi-head attention and feedforward networks. The multi-head attention mechanism allows the model to learn information from different representation subspaces, capturing complex relationships between biomolecules; the feedforward network further processes and transforms the attention-processed information. The representation subspace is a mapping of the knowledge dimensions of the biopharmaceutical target domain onto a high-dimensional vector space, with each subspace corresponding to a specific biological relational topic, avoiding the problem that a single dimension cannot cover complex target attributes.

[0136] By progressively adapting multiple target modules, the model performs in-depth processing of the input domain knowledge, enabling it to understand the semantics, logic, and correlations of the biopharmaceutical target domain. This lays the foundation for target identification through a reasoning logic model.

[0137] Pre-defined thinking logic can be a thinking template, which can include multiple key elements to provide guidance for reasoning in the reasoning logic model.

[0138] Several key elements include Role, Preferences, Goals, Constraints, Skills, and Template Format.

[0139] Role elements: Clearly define the role of the reasoning logic model in the reasoning (e.g., a biopharmaceutical target research expert) to ensure that the reasoning conforms to the cognitive and behavioral logic of the professional identity.

[0140] Preference elements: Used to set the preference tendencies of reasoning, such as giving priority to a certain type of pathway or a certain type of molecular interaction, to narrow the scope of reasoning and improve its relevance.

[0141] The target element defines the goal of the reasoning, that is, the core problem to be solved. For example, "Based on valid biological text, infer whether protein X is a potential drug target", "Based on valid biological text, determine whether protein X is strongly associated with known diseases (such as lung cancer)", "Analyze whether protein X participates in a specific signaling pathway (such as the EGFR pathway) and assess its feasibility as a target in that pathway".

[0142] Constraints are used to clarify the biological rules and data sources that must be followed in reasoning, ensuring the rationality and rigor of the reasoning.

[0143] Skill elements endow the reasoning logic model with the skills required for reasoning, such as the ability to analyze protein sequences and identify functional domains, and to give higher weight to proteins with similar structures to known small molecule inhibitors when assessing target potential.

[0144] The template format elements define the format of the thinking template, making the elements organized in an orderly manner and facilitating the analysis and utilization of the reasoning logic model.

[0145] In the reasoning process, the role element is "expert in biopharmaceutical target research", and the target element is "to analyze whether protein X of an unknown biomolecule participates in a specific signaling pathway and to assess its feasibility as a target in that pathway".

[0146] When a valid biological text is retrieved containing a passage about "protein X being expressed in cardiac tissue," constraint elements (such as "excluding key cardiac proteins") will require the model to carefully assess the safety risks of using it as a target. However, when texts stating "protein X is highly expressed in lung cancer cells" and "protein X is associated with cell proliferation" are found simultaneously, preference elements (such as "prioritizing disease-specific expression and function") will guide the model to assign higher weights to that text.

[0147] The skills elements provide an understanding of professional concepts in the biological field (such as 'kinase domain' and 'apoptosis pathway'), which can ensure that the model can understand the professional concepts in the text and make logical connections.

[0148] Ultimately, the combined effect of all key elements guides the model to generate a clear and evidence-supported initial inference, following a template format of "from molecular characteristics to functional inference, and then to target evaluation." For example, the initial inference might be: "Based on sequence analysis, protein X contains a tyrosine kinase domain; literature indicates that it is overexpressed in NSCLC (non-small cell lung cancer) and associated with poor prognosis; the MAPK pathway it participates in is a mature anti-cancer target pathway. In summary, protein X is a high-potential candidate target for lung cancer treatment."

[0149] In this application, a reasoning logic model can be used to process the preset thinking logic and effective biological text to obtain the initial reasoning results, which can reduce the workload of manual logic sorting, shorten the target determination cycle, and improve processing efficiency.

[0150] S205. Verify the initial reasoning results to obtain the target reasoning results corresponding to the unknown biomolecules.

[0151] After obtaining the initial reasoning result, during the verification process, a self-critic stage is used. The task requirements and reasoning logic set in the pre-defined thinking logic serve as reference standards for judging the reasonableness of the initial reasoning result. If the reasoning process deviates from the reasonable path set in the pre-defined thinking logic, or if the reasoning result does not meet the expected task objective in the template, the system will be triggered to re-examine and correct the reasoning process. If it meets the objective, the initial reasoning result is determined as the target reasoning result.

[0152] The results of target inference can be used to indicate the potential drug target attributes of unknown biomolecules. For example, potential drug target attributes include whether they have target potential, target type, confidence level, supporting evidence, etc.

[0153] Among them, target types can include enzyme targets, receptor targets, ion channel targets, transporter targets, etc.

[0154] The target determination method provided in this application can receive biomolecular information of unknown biomolecules, perform context parsing on the biomolecular information to obtain multiple valid text segments, and concatenate these multiple valid text segments to obtain valid biological text corresponding to the unknown biomolecule. Based on a preset reasoning logic, the valid biological text is used for inference processing to obtain an initial inference result. The initial inference result is then verified to obtain the target inference result corresponding to the unknown biomolecule. Inference is directly performed on the integrated valid biological text without splitting the inference into multiple isolated subtasks, and the reliability of the result is ensured through a verification process, thereby improving the accuracy and credibility of the inference results for the potential drug target attributes of unknown biomolecules.

[0155] Figure 3 This is a schematic diagram of the architecture of a target determination method provided in an embodiment of this application. Please refer to... Figure 3 It can receive biomolecular information from unknown biomolecules. Based on an intelligent agent template, it performs task analysis, intent extraction, query rethinking, and combination decomposition on the biomolecular information to obtain multiple valid text segments (i.e., Top-K contexts). The context is then filtered through a reordering module (i.e., a rule-driven or algorithm-driven module), and then the pre-defined reasoning logic is used to drive the reasoning logic model to output initial reasoning results. Simultaneously, a self-evaluation mechanism is used to self-verify the initial reasoning results to ensure accuracy.

[0156] Figure 4 This is a flowchart illustrating a process for determining a reasoning logic model, as provided in an embodiment of this application. Please refer to [link / reference]. Figure 4 The method includes:

[0157] S401, the target analysis logic for obtaining biomolecules, and multiple biological knowledge literature.

[0158] Multiple biological knowledge documents can be retrieved from the database, including bioinformatics documents such as KEGG and GenePath.

[0159] The target analysis logic is used to standardize and organize gene and protein information related to biopharmaceutical targets in biological knowledge literature. The target analysis logic can include standardized reconstruction information of biomolecular information, knowledge-driven confidence adjustment cases, and public medical question-and-answer and targeted literature question-and-answer with thought chain.

[0160] Standardized Reconstruction Info: This involves standardizing and organizing information related to genes and proteins associated with biopharmaceutical targets (such as the function, expression level, and subcellular localization of the protein encoded by gene ALDH1A1 in the example). Specifically, the following operations are performed:

[0161] Gene information standardization: Gene names are mapped to NCBI Gene standard IDs, gene sequences are uniformly converted to FASTA format, and structural domains are marked using the sequence annotation rules of the Ensembl database;

[0162] ② Protein information standardization: Protein function descriptions are uniformly labeled according to the GO (Gene Ontology) glossary, expression level data are mapped to the [0,1] interval using the Z-score normalization algorithm, and subcellular localization information is encoded according to the classification standards of the UniProt database (e.g., cytoplasm=1, cell membrane=2, cell nucleus=3).

[0163] ③ Heterogeneous data fusion: Standardized gene and protein information is encapsulated in JSON format, with fields including "standard ID-sequence information-functional annotation-expression data-subcellular localization-data source". This transforms scattered and heterogeneous biological data into structured information in a unified format that facilitates subsequent calculation and analysis, providing basic data support for target attribute analysis.

[0164] Knowledge-Driven Confidence Cases: These cases are selected from authoritative literature and professional databases by leveraging expertise in biological fields such as pathway regulation and molecular interactions, combined with expert scoring rules.

[0165] For example, according to expert scoring rules, disease-related proteins are classified into... Positive samples (e.g., those validated as effective targets in preclinical in vitro and in vivo experiments, and supported by more than two independent studies, scored 10); Weakly positive samples (e.g., those with target potential verified by in vitro experiments, supported by 1 literature, and scored 7). Neutral samples (e.g., those related to disease but for which target function has not been validated, scored as 3) Negative samples (e.g., those experimentally verified to have no target potential or a serious off-target risk, scored as 0) are used to determine the confidence level of proteins as potential drug targets. Then, the "input features - target attributes - confidence score" of each sample are adapted to the large model for training in a QA format (e.g., question: "What is the confidence score and target attribute of protein X?", answer: "Score 7, weak positive sample, potential drug target"). This allows the large model to learn the distribution pattern of the confidence score, providing reliable reference samples for subsequent target attribute inference and assisting in the prediction of potential drug targets.

[0166] Public Medical Q&A with Think CoT: To prevent catastrophic forgetting in the model, a combination of QA and Think CoT is used. Based on publicly available datasets, public medical Q&A data with a think chain is synthesized. The data can be divided into training, validation, and test sets in an 8:1:1 ratio. The think chain is a breakdown and display of the reasoning process for answering questions.

[0167] For example, starting from the question "Is protein Y a potential target for liver cancer?", the reasoning steps are as follows: Step 1: The function of protein Y is to regulate hepatocyte proliferation (based on GO database ID: XX); Step 2: Abnormal hepatocyte proliferation is the core pathological mechanism of liver cancer (based on KEGG pathway XX); Step 3: In vitro experiments show that protein Y inhibitors can inhibit the proliferative activity of HepG2 liver cancer cells (based on PubMed literature XX); Step 4: Combining function, pathway, and experimental evidence, protein Y is determined to be a potential target for liver cancer.

[0168] By incorporating knowledge from vertical domains, the model can learn more logical reasoning methods that align with medical understanding, while also preventing the forgetting of existing knowledge.

[0169] Targeted Literature QA: Literature knowledge contains rich theoretical basis for reasoning about potential proteins, thus synthesizing a large amount of information with protein reasoning logic. Specifically, the model is guided by the target protein mining heuristics synthesized by the PriorSyn module to generate question-answer pairs around research literature on specific biopharmaceutical targets (such as experimental design and conclusion interpretation in a certain literature about target discovery), focusing on targeted knowledge mining from target research literature.

[0170] For example, regarding the experimental content of "validation of protein Z as a target for lung cancer" in the literature, a question-and-answer pair is generated: Question: "What are the key experimental methods in the literature for validating protein Z as a target for lung cancer?" Answer: "1. Immunoprecipitation assay to verify the interaction between protein Z and EGFR, a key pathway protein in lung cancer; 2. In a nude mouse xenograft model, knockdown of protein Z reduced tumor volume by 60%; 3. Clinical sample detection showed that the expression level of protein Z in lung cancer tissue was 3.2 times that in normal tissue."

[0171] Standardized reconstructed information provides basic structured biological data; knowledge-driven confidence cases provide highly credible reference examples; public medical Q&A with thought chains supplements logical reasoning knowledge in the general medical field; targeted literature Q&A focuses on the specific knowledge of target research literature.

[0172] Standardized reconstructed information provides basic structured data, knowledge-driven confidence cases provide highly credible references, public medical Q&A with thought chains supplements general medical reasoning logic, and targeted literature Q&A focuses on the specific knowledge of the literature.

[0173] S402. Through target analysis logic, multiple biological knowledge documents are analyzed and processed to obtain multiple known biological texts corresponding to multiple known biomolecules, as well as the known drug target attributes corresponding to each known biomolecule.

[0174] Based on the target analysis logic, the following specific steps can be taken to more accurately extract known biological texts and known drug target attributes of known biomolecules from any biological knowledge literature:

[0175] ① A pre-trained biomedical model is used to perform text segmentation and entity recognition on biological knowledge literature, extracting multiple known text segments such as genes, proteins, diseases, and pathways; ② Combined with standardized reconstruction information, the identified multiple known text segments are mapped into structured data formats (e.g., gene entities are associated with NCBI standard IDs, and disease entities are associated with ICD-11 codes); ③ Referring to knowledge-driven confidence cases, the confidence scores of molecular target associations between multiple known text segments are initially scored to obtain the known preliminary confidence scores; ④ The known preliminary confidence scores are verified and corrected using the reasoning logic provided by public medical question-and-answer platforms with thought chains and targeted literature question-and-answer platforms (e.g., supplementing pathway association evidence not mentioned in the literature to improve the credibility of the associations); ⑤ Finally, the known biological texts of known biomolecules (including standardized known biological sequences, known functions, and known expression data) and the corresponding known drug target attributes (including known target types, confidence scores, and supporting evidence) are determined.

[0176] Among them, the known biological text and known drug target attributes corresponding to each known biomolecule can generate target set domain knowledge. The target set domain knowledge is stored in CSV format, and the fields can include "Molecular ID - Standardized molecular information - Drug target attribute - Confidence score - Supporting evidence source".

[0177] In this application, biological knowledge literature is analyzed and processed through target analysis logic to obtain target set domain knowledge. This allows for more accurate extraction of known molecular information of known biomolecules and known drug target attributes from biological knowledge literature, providing higher quality and more targeted samples for model training. This improves the accuracy and reliability of the model trained based on the target set domain knowledge in tasks such as biological drug target analysis.

[0178] The target analysis logic provides important support for extracting known biological texts of known biomolecules and known drug target attributes from biological knowledge literature. The execution process of the target analysis logic is explained in detail below.

[0179] Specifically, a preset synthesis template can be obtained; an initial analysis logic can be generated based on biological knowledge literature and the preset synthesis template through a large language model; the initial analysis logic can be verified through cross-evaluation of the large language model and expert calibration to obtain the verification result; when the verification result is that the verification is passed, the initial analysis logic is determined as the target analysis logic.

[0180] The target analysis logic belongs to a bioinformatics synthesis workflow, which can generate, verify and output valuable analysis logic from biological knowledge literature based on a large language model (LLM).

[0181] Biological knowledge literature and preset synthesis templates can be input into the large language model. The large language model processes the biological knowledge literature with the help of the preset synthesis templates, and then determines the target analysis logic.

[0182] The preset composition template is a preset workflow. The preset composition template can include Role, Task, Action, Critic and Confidence, and Format.

[0183] The Task is used to define the core task, such as "extracting and optimizing preliminary scientific clues from {literature} for target protein discovery";

[0184] Actions guide the specific operational process, such as "Please break down the task step by step and generate a priority tree-like workflow";

[0185] The self-checking and confidence level (Critic) specifies the requirements for result evaluation and output, such as "make key decisions based on {literature} and {previous} knowledge and provide confidence level";

[0186] The output sample (Format) provides specific query examples, such as "Query example: Analyze the overlap between high-frequency mutation sites; confidence level: 0.92";

[0187] For the role, please refer to the relevant explanations in the pre-defined thinking logic.

[0188] Please see Figure 5 , Figure 5 This is a schematic diagram of an architecture for determining target analysis logic provided in an embodiment of this application. When biological knowledge literature and preset synthesis templates are input into the Large Language Model (LLM), they need to be formatted according to a specific format (FormatInput) so that they can be understood and processed by the LLM. Subsequently, the LLM performs synthesis analysis on the input data, leveraging its language understanding and knowledge integration capabilities to uncover the logic and relationships behind the biological data. Finally, the LLM generates a series of initial analysis logics corresponding to biological analysis, covering high-frequency data overlap analysis, transcriptome data integration and screening, and identification of core proteins in protein-protein interactions, which can cover the workflow of multiple tasks such as biological target discovery and protein function analysis.

[0189] Subsequently, the initial analysis logic generated by the large language model is critically validated. This critical validation is iteratively performed through cross-calibration of the large model and human experts, following optimized high-confidence heuristics.

[0190] Specifically, the initial analysis logic is first analyzed using a cross-model approach to determine the judgment confidence level. Then, it is calibrated by human experts to determine the calibration confidence level. The weighted sum of the judgment confidence level and the calibration confidence level is used to determine the current confidence level. The processed initial analysis logic and the current confidence level are then fed into the large language model for regeneration until the confidence level is greater than or equal to a preset confidence level. The iterative scoring formula is as follows:

[0191]

[0192] in, At the current confidence level, This indicates the confidence level of the cross-model evaluation. It reflects the calibration confidence level of human experts. The first weight for judging confidence level, To calibrate the second weight of the confidence score, .

[0193] Cross-modeling can utilize multiple different large language models to cross-evaluate the generation steps from their respective perspectives. Due to the differences in training data and model structure among different large language models, it can discover flaws in the generation of a single model (such as logical inconsistencies in steps, incorrect data associations, etc.).

[0194] Expert calibration involves human experts in the biological field. These experts, leveraging their expertise (such as a deep understanding of pathways and molecular mechanisms), assess the scientific validity and feasibility of the LLM generation steps, correcting errors caused by "knowledge blind spots" or "statistical biases."

[0195] By combining feedback from cross-disciplinary large models and human expert calibration, the generated initial analysis logic is revised, filtered, and sorted to ensure the scientific rationality of the steps.

[0196] If the verification finds problems with the generated initial analysis logic (such as errors or redundancy in some steps), the current confidence level is lower than the preset confidence level. The initial analysis logic and the current confidence level are then fed back to the large language model to drive the large language model to regenerate a more accurate target analysis logic.

[0197] When the verification result is successful (i.e., the current confidence level is greater than or equal to the preset confidence level), the initial analysis logic is determined as the target analysis logic. The final output is the verified and calibrated target analysis logic.

[0198] In this application, an initial analysis logic is generated based on biological knowledge literature and a preset synthesis template through a large language model. The initial analysis logic is then verified. When the verification result is passed, it is determined as the target analysis logic. This realizes an automated and scientific process for the target analysis logic from generation to verification, and improves the accuracy and reliability of the target analysis logic in tasks such as biological knowledge literature analysis and subsequent research on biological drug targets.

[0199] S403. The initial logic model is trained using the known biological text corresponding to each known biomolecule and the known drug target attributes to obtain the reasoning logic model.

[0200] After inputting the target set of domain knowledge (i.e., known biological texts corresponding to each known biomolecule and known drug target attributes), the initial logical model, based on a large language model-based architecture, is trained using SFT LoRA fine-tuning and the internal computational logic of the TargetModule (target adaptation module). The following section will combine... Figure 6 Detailed explanation.

[0201] SFT: Supervised fine-tuning addresses the domain adaptation gap in biological target discovery using general language models.

[0202] Optimization objective: Adjust the model parameters based on minimizing the loss, as shown in the following formula:

[0203]

[0204] in, This is used to instruct the initial logical model to undergo SFT fine-tuning based on the domain knowledge of the target set to minimize the loss. Used to indicate training data and sample labels From the domain knowledge of the target set Sampling is performed during the process. The log value used to indicate the distribution of the conditional probability function of the initial logical model. Used to indicate the balance between two losses, Used to indicate model parameters for fine-tuning learning. Used to indicate the parameters of a pre-trained model. Representation model regularization, Known biological text used to indicate the i-th known biomolecule. Used to indicate the known drug target properties of the i-th known biomolecule. Target set domain knowledge.

[0205] To meet the need for domain-specific adaptation of the Transformer architecture in the field of biological target discovery, while minimizing computational overhead, the Target Module employs a structured low-rank decomposition strategy. This module is implemented by enhancing query / key / value projection operations in the Transformer layer through additive low-rank parameterization.

[0206] Figure 6 This is a schematic diagram of an architecture for training a reasoning logic model, provided as an embodiment of this application. Please refer to [link / reference]. Figure 6 In the input phase, the target set domain knowledge is input into the initial logical model, which is a basic model based on a large language model.

[0207] During the fine-tuning phase, the initial logistic model is supervisedly fine-tuned (SFT) using LoRA (Low-Rank Adaptation). LoRA adjusts some parameters of the initial logistic model by introducing a low-rank matrix without significantly increasing the number of model parameters, allowing the initial logistic model to quickly learn target set domain knowledge while retaining its general capabilities. During fine-tuning, multiple Target Modules work collaboratively to process the input target set domain knowledge, helping the model gradually master the knowledge patterns and reasoning logic within the domain.

[0208] The target adaptation module, as the core inference unit of the model, follows the classic computational paradigm of Transformer-type models, including pre-processing of the attention mechanism and alternating computation of multi-head attention and feedforward networks.

[0209] During the preprocessing of the attention mechanism, the input feature vector is processed by the query matrix. Key matrix Sum matrix The linear transformations of these three weight matrices generate query vectors, key vectors, and value vectors, respectively. These transformations allow the vectors to interact within a "query-match-value" logic.

[0210] When multi-head attention and feedforward networks are used alternately for computation, firstly, attention scores (which can be understood as the degree of correlation between different information fragments) are calculated based on the Query and Key. Then, a weighted information representation is obtained by combining the Value, allowing the model to capture complex relationships between information (such as interactions between biomolecules, regulatory relationships within pathways, etc.) from the perspective of multiple attention heads. Afterward, the output of multi-head attention is further nonlinearly transformed and feature extracted through the feedforward network to enhance the model's ability to express complex patterns. The structures of multi-head attention and feedforward networks are stacked in multiple layers, enabling the model to deeply process domain knowledge of the biological target set and gradually extract key features and reasoning logic related to drug targets.

[0211] After inputting domain knowledge into the target set, fine-tuning with LoRA, and deep processing by the target adaptation module, the final training yields a reasoning logic model. This model is capable of performing tasks such as knowledge reasoning and target discovery in the field of biopharmaceutical targets.

[0212] The target determination method provided in this application can train an initial logic model using domain knowledge of the target set to obtain a reasoning logic model. This can improve the accuracy and professionalism of the model in tasks such as domain knowledge reasoning and target mining of biopharmaceutical targets, enabling the model to more accurately analyze the attributes of potential drug targets from biomolecular information and provide more reliable support for drug development and other work.

[0213] Below is a corresponding example provided in this application: The user inputs biomolecular information as "Let's see if there is a relationship between MAPK1 protein and Alzheimer's disease, and whether it can be used as a drug target?".

[0214] Through task analysis and intent analysis, the core task can be identified as "target potential assessment," and the user's deeper intent is "exploratory judgment requiring comprehensive evidence." Next, query rewriting transforms the ambiguous "relationship" into structured search keywords, such as "MAPK1 protein basic characteristics (Log2FoldChange, PPI), Alzheimer's disease KEGG pathway, similarity of MAPK1 to known neurological disease targets (blastp, deep_purpose)." Then, the system enters the relevant context retrieval and enhancement stage. Using the rewritten query, the system retrieves a large amount of candidate information from professional databases (such as literature), including a general functional description of MAPK1, its research in a certain cancer, and multiple research data related to Alzheimer's disease.

[0215] The re-ranking module scores and sorts the data based on its relevance to "target assessment," prioritizing the most relevant Top-K (e.g., Top-1) contextual information. For example, it might prioritize evidence stating that "MAPK1 is significantly upregulated in the brain tissue of Alzheimer's patients (Log2FoldChange=1.5) and has a strong interaction with the known target protein APP (PPI composite score=0.95)," while filtering out other less relevant information, thereby improving the signal-to-noise ratio of the model input.

[0216] Next, based on a pre-defined thinking logic, the reasoning instructions are assembled using a thinking template. Simply inputting queries and evidence into the reasoning model is far from sufficient; it must be guided into a specific "role" and "thinking framework." The pre-defined thinking logic can be: "Role: Set as 'a rigorous drug target evaluation expert.' Goal: Clearly defined as 'Assessing the likelihood of MAPK1 becoming a therapeutic target for Alzheimer's disease and outputting a probability value between 0 and 1.' Preferences: Instructing 'Prioritizing experimentally validated data (such as PPI, Log2FoldChange) and comprehensively calculating feature similarity (such as blastp, SVM model scores)'. Constraints: Stipulating that 'probability judgments must be based on provided contextual evidence, avoiding reliance on outdated internal knowledge.' Skills: Requires 'skills in multi-source information fusion and probability assessment'. Template Format: Limiting the output to 'Probability: [Probability Value]'."

[0217] The reasoning logic model, based on its general knowledge gained through pre-training on massive amounts of biomedical knowledge and its professional capabilities gained through fine-tuning on specific target mining data (through SFT LoRA fine-tuning and target adaptation modules), begins its deduction. It reviews the evidence like an expert: "Upregulation of MAPK1 expression (Log2FoldChange=1.5) suggests its active participation in disease progression; strong interaction with the core protein APP (PPI=0.95) indicates its location in a key pathway; it has high sequence similarity to known neural targets (high blastp score)..." The model comprehensively weighs the weights of this evidence and ultimately calculates an initial probability, such as "probability: 0.76".

[0218] Finally, to overcome the randomness of single-inference iterations, the ensemble learning mechanism—the self-critic stage—begins. Since large model generation inherently involves randomness, a single output may be unstable. Therefore, the system repeats the "query-retrieval-template reasoning" process multiple times (e.g., 5 times). It might obtain a set of results: [0.76, 0.72, 0.80, 0.77, 0.75]. Subsequently, the system integrates these results to obtain a final, more robust probability score: 0.76. This process is equivalent to multiple independent evaluations, synthesizing all opinions to arrive at the final conclusion, ensuring the high quality and reliability of the derivation results.

[0219] It is worth noting that the above embodiments are merely one way to determine the properties of known drug targets. If the input is "to see if MAPK1 protein is related to Alzheimer's disease, whether it can be used as a drug target, and what the target type is?", the obtained properties of known drug targets may also include the target type.

[0220] Figure 7 This is a schematic diagram of the architecture of a target determination device provided in an embodiment of this application. Please refer to... Figure 7 The target determination device 700 includes a receiving module 701, a parsing processing module 702, a splicing processing module 703, an inference processing module 704, and a verification processing module 705, wherein...

[0221] The receiving module 701 is used to receive biomolecular information of unknown biomolecules;

[0222] The parsing and processing module 702 is used to perform context parsing processing on biomolecular information to obtain multiple valid text segments. The valid text segments are key relevant information that supports the determination of the target of unknown biomolecules.

[0223] The splicing processing module 703 is used to splice multiple valid text segments through the reordering module to obtain the valid biological text corresponding to the unknown biomolecule.

[0224] The reasoning processing module 704 is used to perform reasoning processing on valid biological text based on preset thinking logic to obtain the initial reasoning result corresponding to the unknown biomolecule. The preset thinking logic is used to guide the reasoning process of the preset reasoning model.

[0225] The verification processing module 705 is used to verify the initial inference result and obtain the target inference result corresponding to the unknown biomolecule. The target inference result is used to indicate the potential drug target attributes of the unknown biomolecule.

[0226] In one possible implementation, the parsing processing module 702 is specifically used for:

[0227] Acquire agent templates, which provide a framework for task analysis, intent analysis, query rethinking, and combined queries to identify targets for unknown biomolecules;

[0228] Based on the intelligent agent template, context parsing is performed on biomolecular information to obtain multiple valid text segments;

[0229] Among them, task analysis is used to extract key elements from biomolecular information and form a task chain by logically sorting them according to basic features, functional inference, and target potential assessment; intent analysis is used to determine the task intent based on the task chain, focusing on exploring the potential of biomolecules as therapeutic targets for specific diseases, and obtaining intent analysis results; query rethinking is used to generate and optimize query sets based on intent analysis results; combined query is used to combine query sets according to logical association and priority, and initiate retrieval to biomedical databases to obtain multi-source related information.

[0230] In one possible implementation, the splicing processing module 703 is specifically used for:

[0231] The reordering module includes rule-driven or algorithm-driven modules; it concatenates multiple valid text segments to obtain valid biological text corresponding to the unknown biomolecule, including:

[0232] The sorting sequence number corresponding to each valid text segment is determined through rule-driven or algorithm-driven modules.

[0233] Based on the sorting number corresponding to each valid text segment, multiple valid text segments are concatenated to obtain valid biological text.

[0234] The rule-driven module determines the key inference dimensions corresponding to each valid text segment based on the dimension importance ranking rules, and determines the corresponding ranking number based on the key inference dimensions corresponding to each valid text segment; the algorithm-driven module determines the weight corresponding to each valid text segment, and determines the corresponding ranking number based on the weight corresponding to each valid text segment.

[0235] In one possible implementation, the inference processing module 704 is specifically used for:

[0236] The reasoning logic model is obtained by training the initial logic model with the domain knowledge of the target set. The domain knowledge of the target set includes known molecular information of multiple known biomolecules and their corresponding potential drug target attributes.

[0237] By using a reasoning logic model, the pre-set thinking logic and effective biological text are processed to obtain the initial reasoning result.

[0238] In one possible implementation, the inference processing module 704 is specifically used for;

[0239] The target analysis logic for biomolecules and multiple biological knowledge literatures are obtained. The target analysis logic is used to standardize and organize gene and protein information related to biopharmaceutical targets in the biological knowledge literatures.

[0240] By using target analysis logic, multiple biological knowledge documents are analyzed and processed to obtain multiple known biological texts corresponding to multiple known biomolecules, as well as the known drug target attributes corresponding to each known biomolecule.

[0241] By training the initial logic model using the known biological text corresponding to each known biomolecule and the known drug target attributes, a reasoning logic model is obtained.

[0242] In one possible implementation, the inference processing module 704 is specifically used for:

[0243] Using a biomedical pre-trained model, multiple known text segments and known target types are extracted from biological knowledge literature, and these multiple known text segments are concatenated into known biological text, which includes known biological sequences, known functions, and known expression data of known biomolecules.

[0244] Based on standardized reconstructed information, multiple identified known text segments are mapped into a structured data format;

[0245] Based on confidence-adjusted cases, obtain the known preliminary confidence levels corresponding to the molecular target association relationships between known biological texts and known target types;

[0246] Using the reasoning logic provided by mind chain question answering and targeted literature analysis, the initial confidence level corresponding to the known biological text is corrected to obtain the known target confidence level;

[0247] Known target types, known target confidence levels, and sources of biological knowledge literature are identified as the attributes of known drug targets.

[0248] In one possible implementation, the inference processing module 704 is specifically used for:

[0249] Obtain a preset synthesis template, which includes standardized reconstruction rules, confidence assessment rules, and medical question-and-answer generation rules for biomolecular analysis;

[0250] Using a large language model, based on biological knowledge literature and pre-set synthesis templates, the initial analysis logic is generated;

[0251] The initial analysis logic is validated through cross-evaluation using a large language model and expert calibration to obtain the validation results.

[0252] If the verification result is successful, the initial analysis logic will be determined as the target analysis logic.

[0253] The target determination device provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0254] Figure 8 A schematic diagram of the structure of an electronic device provided in this application. Please refer to [link / reference]. Figure 8 The electronic device 800 may include a processor 801 and a memory 802. Exemplarily, the processor 801 and the memory 802 are interconnected via a bus 803.

[0255] The 802 memory stores instructions executed by the computer;

[0256] The processor 801 executes computer execution instructions stored in the memory 802, causing the processor 601 to perform the target determination method as shown in the above method embodiment.

[0257] Accordingly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the target determination method of the above-described method embodiments.

[0258] Accordingly, embodiments of this application may also provide a computer program product, including a computer program, which, when executed by a processor, can implement the target determination method shown in the above method embodiments.

[0259] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0260] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0261] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0262] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0263] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0264] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0265] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0266] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0267] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for determining target points, characterized in that, include: Receive biomolecular information from unknown biomolecules; The biomolecule information is subjected to context parsing to obtain multiple valid text segments, which are key relevant information supporting the determination of the target of the unknown biomolecule. The reordering module concatenates the multiple valid text segments to obtain the valid biological text corresponding to the unknown biomolecule. Based on a preset thinking logic, the effective biological text is subjected to reasoning processing to obtain the initial reasoning result corresponding to the unknown biomolecule. The preset thinking logic is used to guide the reasoning process of the preset reasoning model. The initial inference result is verified to obtain the target inference result corresponding to the unknown biomolecule. The target inference result is used to indicate the potential drug target properties of the unknown biomolecule.

2. The method according to claim 1, characterized in that, The biomolecular information is subjected to context parsing to obtain multiple valid text segments, including: A smart agent template is obtained, which is used to provide a framework for task analysis, intent analysis, query rethinking, and combined query for the determination of the target of the unknown biomolecule; Based on the intelligent agent template, the biomolecular information is subjected to context parsing to obtain the multiple valid text segments; The task analysis is used to extract key elements from biomolecular information and form a task chain by logically sorting them according to basic features, functional inference, and target potential assessment; the intent analysis is used to determine the task intent based on the task chain, focusing on exploring the potential of biomolecules as disease treatment targets, and obtaining intent analysis results; the query rethinking is used to generate and optimize a query set based on the intent analysis results; and the combined query is used to combine the query set according to logical association and priority, and initiate a search in the biomedical database to obtain multi-source related information.

3. The method according to claim 1, characterized in that, The reordering module includes a rule-driven module or an algorithm-driven module; The multiple valid text segments are concatenated to obtain the valid biological text corresponding to the unknown biomolecule, including: The sorting sequence number corresponding to each valid text segment is determined through rule-driven or algorithm-driven modules. Based on the sorting number corresponding to each valid text segment, the multiple valid text segments are concatenated to obtain the valid biological text; The rule-driven module determines the key inference dimensions corresponding to each valid text segment based on the dimension importance ranking rule, and determines the corresponding ranking number based on the key inference dimensions corresponding to each valid text segment; the algorithm-driven module determines the weight corresponding to each valid text segment, and determines the corresponding ranking number based on the weight corresponding to each valid text segment.

4. The method according to claim 1, characterized in that, Based on a pre-defined logic, the valid biological text is subjected to reasoning processing to obtain the initial reasoning result corresponding to the unknown biomolecule, including: A reasoning logic model is obtained, which is a pre-trained model obtained by training an initial logic model with domain knowledge of the target set. The domain knowledge of the target set includes known molecular information of multiple known biomolecules and their corresponding potential drug target attributes. The reasoning logic model is used to process the preset thinking logic and the effective biological text to obtain the initial reasoning result.

5. The method according to claim 4, characterized in that, Obtain the reasoning logic model, including; The target analysis logic for biomolecules and multiple biological knowledge documents are obtained. The target analysis logic is used to standardize and organize gene and protein information related to biopharmaceutical targets in the biological knowledge documents. Through the target analysis logic, the multiple biological knowledge documents are analyzed and processed to obtain multiple known biological texts corresponding to multiple known biomolecules, as well as the known drug target attributes corresponding to each known biomolecule. The initial logic model is trained using the known biological text corresponding to each known biomolecule and the known drug target attributes to obtain the inference logic model.

6. The method according to claim 5, characterized in that, The target analysis logic includes standardized reconstructed information, confidence-adjusted cases, and public medical question-and-answer and targeted literature question-and-answer. For any given biological knowledge document, the target analysis logic analyzes and processes the biological knowledge document to obtain the known biological text and drug target attributes of the known biomolecules corresponding to the biological knowledge document, including: A biomedical pre-trained model is used to extract multiple known text segments and known target types from the biological knowledge literature, and then the multiple known text segments are concatenated into the known biological text, which includes known biological sequences, known functions, and known expression data of known biomolecules. Based on the standardized reconstruction information, the identified multiple known text segments are mapped into a structured data format; Based on the confidence adjustment case, obtain the known preliminary confidence level corresponding to the molecular target association relationship between the known biological text and the known target type; Using the reasoning logic provided by public medical question-and-answer with thought chain and targeted literature analysis, the initial confidence level corresponding to the known biological text is corrected to obtain the known target confidence level; The known target type, known target confidence level, and the literature source of the biological knowledge literature are determined as the attributes of the known drug target.

7. The method according to claim 6, characterized in that, The target analysis logic for obtaining biomolecules includes: Obtain a preset synthesis template, which includes standardized reconstruction rules, confidence assessment rules, and medical question-and-answer generation rules for biomolecular analysis; Using a large language model, initial analysis logic is generated based on the biological knowledge literature and the preset synthesis template; The initial analysis logic is validated through cross-evaluation using a large language model and expert calibration to obtain the validation results. If the verification result is successful, the initial analysis logic is determined as the target analysis logic.

8. A target determination device, characterized in that, It includes a receiving module, a parsing processing module, a splicing processing module, an inference processing module, and a verification processing module, among which, The receiving module is used to receive biomolecular information of unknown biomolecules; The parsing and processing module is used to perform context parsing processing on the biomolecule information to obtain multiple text segments, wherein the text segments are key relevance information supporting the determination of the target of the unknown biomolecule; The splicing processing module is used to splice the multiple text segments to obtain the valid biological text corresponding to the unknown biomolecule. The reasoning processing module is used to perform reasoning processing on the effective biological text based on a preset thinking logic to obtain the initial reasoning result corresponding to the unknown biomolecule. The verification processing module is used to verify the initial inference result to obtain the target inference result corresponding to the unknown biomolecule. The target inference result is used to indicate the potential drug target attributes of the unknown biomolecule.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 7.

Images