Medicine research and development industry knowledge base system and method based on artificial intelligence

Abstract

The invention discloses a medicine research and development industry knowledge base system and method based on artificial intelligence, and belongs to the technical field of medicine information processing and artificial intelligence. The system comprises a data acquisition module, a data preprocessing module, a knowledge extraction module, a knowledge integration module, a knowledge reasoning module, a knowledge updating module and an intelligent center layer. The data acquisition module acquires multi-modal medicine research and development data, the preprocessing module performs cleaning, standardization and feature alignment, the knowledge extraction module adopts a deep learning model based on Transformer to extract entity relationships, the knowledge integration module checks conflicts through a chemical rule knowledge base and constructs a knowledge graph, and the knowledge graph is used for establishing a knowledge database. The knowledge reasoning module performs association prediction by using a graph neural network and a time sequence model, the knowledge updating module realizes parameter optimization and incremental updating based on human feedback reinforcement learning, and the intelligent central layer overall plans all the modules to work cooperatively based on a large language model arrangement framework. According to the method, multi-source heterogeneous data can be integrated, domain semantics can be deeply understood, a knowledge structure is dynamically evolved, intelligent decision support is provided, and the efficiency and accuracy of medicine research and development knowledge management and application are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of pharmaceutical information processing and artificial intelligence technology, and more specifically, to a knowledge base system and method for the pharmaceutical R&D industry based on artificial intelligence. Background Technology

[0002] Pharmaceutical R&D is a typical knowledge-intensive field. The amount of research literature, clinical trial reports, regulatory documents, and internal corporate data generated during the R&D process is growing exponentially. How to efficiently extract, integrate, and utilize professional knowledge from massive amounts of multi-source, heterogeneous data has become a key bottleneck restricting the efficiency of new drug development. Existing technologies have attempted to construct pharmaceutical knowledge graphs for knowledge management, such as using models like BERT for entity relation extraction, but they generally suffer from the following shortcomings: First, insufficient multi-source data integration capabilities; structured experimental data and unstructured text, images, and other modal information are difficult to model uniformly, leading to information silos. Second, a lack of depth in domain semantic understanding; traditional natural language processing models lack biochemical mechanism constraints, making it difficult to accurately analyze complex reaction formulas, target interactions, and clinical terminology. Third, static and fixed knowledge bases; unable to track the latest global pharmaceutical developments in real time and lacking self-evolution capabilities. Fourth, broken decision support chains; extracted knowledge is difficult to transform into interpretable and traceable R&D recommendations.

[0003] In recent years, large language models and deep learning technologies have shown potential in intelligent question answering and knowledge reasoning, but the adaptability of existing systems in the pharmaceutical R&D field remains weak. While some patents propose process decision-making methods based on knowledge graphs, they are mostly geared towards general production scenarios and lack customization for the specific needs of pharmaceutical R&D, such as chemical rule constraints, clinical trial timeline characteristics, and evidence level screening. Furthermore, existing knowledge update mechanisms largely rely on manual maintenance and lack a closed-loop design that leverages user feedback to achieve model self-evolution. Therefore, there is an urgent need for an intelligent pharmaceutical R&D knowledge base system that can deeply integrate domain expertise, possess multimodal processing capabilities, and dynamically evolve. Summary of the Invention

[0004] In view of the above-mentioned technical problems in related technologies, this invention proposes an artificial intelligence-based knowledge base system and method for the pharmaceutical R&D industry, which can overcome the above-mentioned shortcomings of the prior art.

[0005] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows: An AI-based knowledge base system for the pharmaceutical R&D industry; This AI-based pharmaceutical R&D industry knowledge base system includes: a data acquisition module, a data preprocessing module, a knowledge extraction module, a knowledge integration module, a knowledge reasoning module, a knowledge update module, and an intelligent central layer; The data acquisition module is used to acquire multimodal pharmaceutical R&D data from multiple data sources; The data preprocessing module is used to clean, standardize, disambiguate entities, and align features of the acquired data. The knowledge extraction module uses a deep learning model based on the Transformer architecture to extract entities and relationships between entities from preprocessed data. The knowledge integration module integrates a chemical rule knowledge base, which is used to perform rule-based verification and conflict correction on the extracted knowledge in order to construct a structured knowledge graph. The knowledge reasoning module, based on graph neural networks and time-series models, is used to perform association reasoning and prediction on the knowledge graph; The knowledge update module, based on a human feedback reinforcement learning mechanism, is used to optimize system parameters and drive knowledge updates based on user feedback. The intelligent central layer is built on a large language model orchestration framework and is used to schedule tasks and orchestrate workflows for the data acquisition module, data preprocessing module, knowledge extraction module, knowledge integration module, knowledge reasoning module, and knowledge update module, so as to collaboratively complete the knowledge construction, service, and evolution process.

[0006] Furthermore, the data acquisition module adopts a web crawler framework with an integrated reinforcement learning dynamic scheduler, which is used to dynamically adjust the crawling frequency and depth of different data sources based on the stage of the R&D project, the authority of the data source, and historical usage feedback. The data preprocessing module executes a timeliness-relevance joint weighted algorithm to calculate a comprehensive weight for each piece of acquired data. The comprehensive weight is determined based on the semantic similarity between the data and the input technical text, the time decay factor, the domain influence factor of the data source, and the level of clinical evidence. The data preprocessing module also includes a multimodal feature alignment unit, which is used to align the semantic features of image data and text data through a cross-modal attention network.

[0007] Furthermore, in the aforementioned timeliness-relevance joint weighted algorithm, the formula for calculating the comprehensive weight W is: ; Where S is the cosine similarity between the input technical text and the crawled data, R is the time decay factor, I is the domain influence factor, L is the clinical evidence level, and α, β, and γ are adjustable hyperparameters.

[0008] Furthermore, the knowledge extraction module uses a model pre-trained on biomedical text for domain-adaptive fine-tuning, and introduces syntactic dependency tree mask and biochemical mask matrix to jointly constrain the calculation of attention weights; The knowledge extraction module adopts a multi-task learning architecture, with a named entity recognition head, a relation extraction head, and an event extraction head connected in parallel after a shared Transformer encoder.

[0009] Furthermore, the chemical rule knowledge base in the knowledge integration module includes at least a reaction rule layer, a physicochemical property constraint layer, and a safety and compliance layer; The knowledge integration module performs multi-dimensional evaluation of the extracted knowledge through an iterative verification function. The dimensions evaluated by the verification function include the free energy of chemical reactions, safety score, synthesis feasibility score, and intellectual property risk score. When a knowledge conflict is detected, the intelligent central layer invokes the orchestrated ReAct workflow to perform retrieval and correction operations.

[0010] Furthermore, the knowledge reasoning module constructs a heterogeneous graph knowledge graph and uses a relational graph convolutional network for node representation learning; simultaneously, it uses a temporal model to encode the clinical trial time series data and updates the obtained representation vectors as dynamic attributes to the knowledge graph.

[0011] Furthermore, the reward model in the knowledge update module calculates the reward value based on multi-dimensional feedback labels, and adds a KL divergence penalty term during the strategy optimization process to constrain the strategy update; the knowledge update module is also used to identify knowledge blind spots based on system confidence and R&D attention, and trigger the data acquisition module to perform targeted data supplementation.

[0012] According to another aspect of the present invention, a method for implementing a knowledge base system for the pharmaceutical R&D industry based on artificial intelligence is provided; The implementation method of this AI-based pharmaceutical R&D industry knowledge base system includes: Knowledge construction phase: Multi-source heterogeneous data is acquired and processed through the data acquisition module and data preprocessing module; knowledge is extracted, verified and integrated through the knowledge extraction module and knowledge integration module to construct a pharmaceutical R&D knowledge graph; Knowledge service and reasoning stage: The intelligent central layer parses the user's natural language query and schedules the knowledge reasoning module and the underlying knowledge base to perform collaborative retrieval and reasoning, generating and outputting structured results with knowledge tracing and confidence assessment; Knowledge evolution stage: User feedback is collected through the knowledge update module, and the system parameters are optimized and the knowledge graph is incrementally updated based on the human feedback reinforcement learning mechanism.

[0013] Furthermore, in the knowledge construction stage, the knowledge integration module uses a chemical rule knowledge base to perform conflict detection and verification on the extracted "drug-action-target" triplet. The output of the verification function is a multi-dimensional evaluation vector that includes reaction Gibbs free energy, safety score, synthesis feasibility score and intellectual property risk score.

[0014] Furthermore, in the knowledge evolution stage, the knowledge update module updates the model parameters based on the near-end policy optimization algorithm. Its optimization objective function includes a KL divergence penalty term to constrain the difference between the new policy and the original security policy.

[0015] The beneficial effects of this invention are as follows: By constructing a hierarchical architecture of "data-knowledge-service" with an intelligent central layer as the core, and coordinating multimodal data preprocessing, domain-adaptive knowledge extraction, chemical rule and symbolic reasoning fusion, temporal graph joint reasoning, and reinforcement learning update mechanism based on human feedback, this invention effectively integrates multi-source heterogeneous pharmaceutical data, significantly improves the semantic understanding of complex medical terms and biochemical relationships, and then constructs a dynamic knowledge graph with self-evolution capabilities. This provides researchers with traceable, explainable, and high-confidence decision support, greatly improving the efficiency and intelligence level of knowledge acquisition in pharmaceutical research and development. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0017] The AI-based pharmaceutical R&D industry knowledge base system according to an embodiment of the present invention includes: a data acquisition module, a data preprocessing module, a knowledge extraction module, a knowledge integration module, a knowledge reasoning module, a knowledge update module, and an intelligent central layer; The data acquisition module is used to acquire multimodal pharmaceutical R&D data from multiple data sources; The data preprocessing module is used to clean, standardize, disambiguate entities, and align features of the acquired data. The knowledge extraction module uses a deep learning model based on the Transformer architecture to extract entities and relationships between entities from preprocessed data. The knowledge integration module integrates a chemical rule knowledge base, which is used to perform rule-based verification and conflict correction on the extracted knowledge in order to construct a structured knowledge graph. The knowledge reasoning module, based on graph neural networks and time-series models, is used to perform association reasoning and prediction on the knowledge graph; The knowledge update module, based on a human feedback reinforcement learning mechanism, is used to optimize system parameters and drive knowledge updates based on user feedback. The intelligent central layer is built on a large language model orchestration framework and is used to schedule tasks and orchestrate workflows for the data acquisition module, data preprocessing module, knowledge extraction module, knowledge integration module, knowledge reasoning module, and knowledge update module, so as to collaboratively complete the knowledge construction, service, and evolution process.

[0018] According to an embodiment of the present invention, in a specific implementation of the AI-based pharmaceutical R&D industry knowledge base system, the data acquisition module adopts a web crawler framework with an integrated reinforcement learning dynamic scheduler, which is used to dynamically adjust the crawling frequency and depth of different data sources based on the R&D project stage, the authority of the data source, and historical usage feedback. The data preprocessing module executes a timeliness-relevance joint weighted algorithm to calculate a comprehensive weight for each piece of acquired data. The comprehensive weight is determined based on the semantic similarity between the data and the input technical text, the time decay factor, the domain influence factor of the data source, and the level of clinical evidence. The data preprocessing module also includes a multimodal feature alignment unit, which is used to align the semantic features of image data and text data through a cross-modal attention network.

[0019] In a specific embodiment of the AI-based pharmaceutical R&D industry knowledge base system according to an embodiment of the present invention, the formula for calculating the comprehensive weight W in the timeliness-relevance joint weighting algorithm is as follows: ; Where S is the cosine similarity between the input technical text and the crawled data, R is the time decay factor, I is the domain influence factor, L is the clinical evidence level, and α, β, and γ are adjustable hyperparameters.

[0020] According to an embodiment of the present invention, in a specific implementation of the knowledge base system for the pharmaceutical R&D industry based on artificial intelligence, the knowledge extraction module uses a model pre-trained on biomedical text for domain-adaptive fine-tuning, and introduces syntactic dependency tree mask and biochemical mask matrix to jointly constrain the calculation of attention weights; The knowledge extraction module adopts a multi-task learning architecture, with a named entity recognition head, a relation extraction head, and an event extraction head connected in parallel after a shared Transformer encoder.

[0021] According to an embodiment of the present invention, the knowledge base system for the pharmaceutical R&D industry based on artificial intelligence, in a specific implementation, includes at least a reaction rule layer, a physicochemical property constraint layer, and a safety and compliance layer in the knowledge integration module. The knowledge integration module performs multi-dimensional evaluation of the extracted knowledge through an iterative verification function. The dimensions evaluated by the verification function include the free energy of chemical reactions, safety score, synthesis feasibility score, and intellectual property risk score. When a knowledge conflict is detected, the intelligent central layer invokes the orchestrated ReAct workflow to perform retrieval and correction operations.

[0022] According to an embodiment of the present invention, in a specific implementation of the AI-based pharmaceutical R&D industry knowledge base system, the knowledge reasoning module constructs a heterogeneous graph knowledge graph and uses a relational graph convolutional network for node representation learning; simultaneously, a time series model is used to encode clinical trial time series data, and the obtained representation vector is used as a dynamic attribute to update the knowledge graph.

[0023] According to an embodiment of the present invention, in a specific implementation of the AI-based pharmaceutical R&D industry knowledge base system, the reward model in the knowledge update module calculates the reward value based on multi-dimensional feedback labels, and adds a KL divergence penalty term to constrain the strategy update during the strategy optimization process; the knowledge update module is also used to identify knowledge blind spots based on system confidence and R&D attention, and trigger the data acquisition module to perform targeted data supplementation.

[0024] Secondly, the implementation method of the artificial intelligence-based pharmaceutical R&D industry knowledge base system according to the embodiments of the present invention, applied to the above-mentioned system, includes: Knowledge construction phase: Multi-source heterogeneous data is acquired and processed through the data acquisition module and data preprocessing module; knowledge is extracted, verified and integrated through the knowledge extraction module and knowledge integration module to construct a pharmaceutical R&D knowledge graph; Knowledge service and reasoning stage: The intelligent central layer parses the user's natural language query and schedules the knowledge reasoning module and the underlying knowledge base to perform collaborative retrieval and reasoning, generating and outputting structured results with knowledge tracing and confidence assessment; Knowledge evolution stage: User feedback is collected through the knowledge update module, and the system parameters are optimized and the knowledge graph is incrementally updated based on the human feedback reinforcement learning mechanism.

[0025] According to an embodiment of the present invention, in a specific implementation of the knowledge base system for the pharmaceutical R&D industry based on artificial intelligence, during the knowledge construction stage, the knowledge integration module uses a chemical rule knowledge base to perform conflict detection and verification on the extracted "drug-action-target" triplet. The output of the verification function is a multi-dimensional evaluation vector that includes reaction Gibbs free energy, safety score, synthesis feasibility score, and intellectual property risk score.

[0026] According to an embodiment of the present invention, in a specific implementation of the knowledge base system for the pharmaceutical R&D industry based on artificial intelligence, during the knowledge evolution stage, the knowledge update module updates the model parameters based on the proximal strategy optimization algorithm. The optimization objective function includes a KL divergence penalty term to constrain the difference between the new strategy and the original safety strategy.

[0027] To facilitate understanding of the above technical solutions of the present invention, the following describes the above technical solutions of the present invention in detail through the overall system architecture and specific embodiments.

[0028] System overall architecture: This invention provides an artificial intelligence-based knowledge base system and method for the pharmaceutical R&D industry. The system is an intelligent platform employing a layered "data-knowledge-service" architecture. Its core innovation lies in introducing a "large language model orchestration framework (Dify) as the intelligent hub," which, in conjunction with a lower-level domain-specific deep learning model and a top-level symbolic rule engine, constructs a dynamic knowledge entity that is perceptible, reasonable, verifiable, and evolvable. The system architecture, from top to bottom, includes: Interaction and Service Layer: Provides interfaces for natural language question answering, intelligent search, analysis report generation, and R&D suggestion push.

[0029] The intelligent central layer is responsible for task scheduling, workflow orchestration, multi-model collaborative decision-making, and closed-loop management of human-machine feedback (RLHF). This is the "brain" of the system, responsible for combining underlying capabilities into practical applications to solve complex R&D problems. In the knowledge integration module, when a conflict is detected between 'Drug A - Action - Target B' and the rule base, Dify not only invokes the ReAct workflow but also dynamically loads the latest published compliance guidelines, updates the rule base in real time, and feeds back to the knowledge extraction module to readjust entity relationship weights.

[0030] Core Knowledge Processing Layer: Contains the six core modules described in this invention (data acquisition, preprocessing, extraction, integration, reasoning, and updating), and is responsible for the full lifecycle management of knowledge.

[0031] Data and infrastructure layer: Provides access to multi-source heterogeneous data, and supports distributed storage and computing resources.

[0032] Enhanced technical details for each module: 1. Multimodal data preprocessing and feature alignment; Dynamic priority crawling strategy: The data acquisition module integrates Scrapy and Selenium, and has a built-in reinforcement learning-based dynamic scheduler. The scheduler adjusts the crawling frequency and depth of each data source in real time based on the development project stage, the authority of the data source, and historical usage feedback.

[0033] A timeliness-relevance weighted algorithm for the pharmaceutical field: Based on the existing time decay factor, it further introduces a domain influence factor I(d) and a clinical evidence level L(d). The comprehensive weighting formula is enhanced as follows: Here, α, β, and γ are adjustable hyperparameters. This algorithm ensures that the knowledge base is not only "new," but also "excellent" and "relevant."

[0034] Multimodal Feature Alignment: For image data such as chemical structural formulas, pathological slides, and protein electrophoresis images, a pre-trained Vision Transformer (ViT) is used to extract a visual feature vector Vimg. For the text description, a biomedical BERT is used to extract a text feature vector Vtext. A cross-modal attention network is used to learn the joint representation of the two, Vjoint = Attention(Vimg, Vtext), to achieve semantic alignment between the image and text, providing a unified feature input for subsequent knowledge extraction. During the ViT and BERT feature alignment stage, a chemical structure similarity loss term is added to ensure that the image and text are consistent at the chemical semantic level.

[0035] The data preprocessing module not only cleans the data but also incorporates a time-sensitive weighting algorithm specific to the pharmaceutical field. For the n captured data points, the overall weight is calculated as follows: ; Here, represents the cosine similarity between the input technical text and the crawled data, R is the time decay factor, and tpub is the publication time. This algorithm ensures that the knowledge base always prioritizes referencing the latest scientific research results and clinical trial data.

[0036] 2. Knowledge Extraction Module: A hybrid task architecture based on Transformer; Pre-trained model fine-tuning: Instead of directly using the general BERT, BioBERT or PubMedBERT is used as the base model, and further domain-adaptive fine-tuning is performed on a self-constructed pharmaceutical R&D entity-relationship annotation dataset.

[0037] A grammar- and structure-aware attention mechanism is introduced: in addition to the biochemical mask matrix Mbio, a syntactic dependency tree mask Msyn is also introduced. When calculating attention, the model is forced to focus more on words that are syntactically directly related, thereby improving the accuracy of extracting complex relationships in long and difficult sentences. The final attention weights are calculated as follows: Where o represents element-wise multiplication (Hadamard product), and λ is the balance coefficient.

[0038] Multi-task learning head design: Three task heads are connected in parallel after a shared Transformer encoder. Named Entity Recognition Header: Employs a CRF layer to identify entities such as drugs, targets, diseases, and side effects.

[0039] Relationship Extraction Header: Employs a pointer network to identify relationships between entities such as "inhibition", "activation", "cause", and "combined use".

[0040] Event Extraction Header: For clinical events, identify the complete event chain of "drug administration-response-outcome".

[0041] The knowledge extraction module adopts an improved Transformer architecture, which achieves synchronous extraction of entities and relationships through multi-task learning.

[0042] Semantic embedding generation: The BERT model is used to transform the token into a high-dimensional vector $H$.

[0043] Attention mechanism constraint: To capture the long-distance dependence between drug molecules and target proteins, a scaled dot product attention formula is introduced: ; Mbio is a biochemical mask matrix introduced in this invention, used to shield potential pseudo-links that do not conform to chemical bonding rules.

[0044] 3. Knowledge Integration Module: Chemical Rule Engine and Dynamic Validation; Digitalization and Hierarchical Structure of Chemical Rule Knowledge Base: Reaction rule layer: Based on SMILES strings and RDKit toolkit, it encodes organic reaction rules (such as functional group transformation, protection and deprotection).

[0045] Physicochemical property constraint layer: Set reasonable ranges for parameters such as solubility, LogP, and pKa.

[0046] Safety and Compliance Layer: Integrates known drug-drug interaction (DDI) pairs, high-risk genotoxicity warning structures (such as warning groups), and incompatibilities tables included in pharmacopoeias of various countries.

[0047] Detailed breakdown of the iterative verification and correction process: Conflict detection: The extracted "drug A-action-target B" triplet is matched with a rule base. If A is found to be structurally highly similar to a known drug C (judged by the Tanimoto coefficient of molecular fingerprint), and C has no known interaction with B, it is marked as a "potential new action" requiring high-level evidence; if the interaction of A and B is found to contradict known pathway knowledge, it is marked as a "conflict".

[0048] Intelligent correction suggestion generation: When the system detects that "reagent X is added under strong acid conditions" in the extraction experimental steps, but X is known to be easily hydrolyzed, it will not only issue an alarm, but its built-in Dify-arranged ReAct workflow will automatically execute: Think -> Search (pharmacopoeia and chemical databases) -> Act (generate suggestion: "It is recommended to replace reagent X with a more stable precursor Y, or add it slowly at low temperature"), along with the literature evidence found.

[0049] The validation function f_val is extended: based on the original key indicators, a synthetic feasibility score Synth(R) (based on inverse synthesis analysis algorithm) and an intellectual property risk score IP(R) (based on patent database similarity retrieval) are added. f_val outputs a multi-dimensional evaluation vector to comprehensively measure the feasibility of a research and development hypothesis.

[0050] The core of this module lies in combining "connectivity" with "symbolicism (rule base)". The knowledge integration module evaluates the extracted reaction formulas or recipes through an iterative verification function fval: ; Where G is the reaction Gibbs free energy (to determine spontaneity), Safety(R) is the safety score, and T is the preset threshold.

[0051] Rule conflict correction: If the extracted "drug A + drug B" has a compatibility incompatibility, the system will automatically call the ReAct operator arranged by Dify, search the pharmacopoeia database for correction, and generate a warning message with the cited source.

[0052] 4. Knowledge Reasoning Module: Based on GRU and Graph Neural Network (GNN) Construction of Dynamic Heterogeneous Graph Neural Networks: In a knowledge graph G=(V,E), nodes V and edges E have multiple types (heterogeneous graph). Relational Graph Convolutional Networks (R-GCN) are used to distinguish different types of relations, thereby learning node representations more accurately.

[0053] Time-Graph Fusion Reasoning: For clinical trial time-series data, a bidirectional GRU is used to capture past and future dependencies.

[0054] The efficacy indicator sequence of a patient group in a clinical trial that changes over time is encoded into a comprehensive representation vector htrial using GRU.

[0055] This vector `htrial` is used as a dynamic attribute of the "drug-disease" edge corresponding to this clinical trial and updated in the knowledge graph. This allows the graph to express not only static relationships but also "the strength / trend of a certain relationship in a specific trial".

[0056] In the knowledge reasoning module, in addition to R-GCN, a new causal graph attention layer is added. Through intervention-effect prediction task training, it identifies the causal chain of "drug → target → pathway → disease" and outputs the causal strength confidence.

[0057] Zero-sample / few-sample inference capability: For novel targets or rare diseases, when direct association evidence is extremely scarce, the system employs a meta-learning strategy. Utilizing a large number of known "drug-target-pathway-disease" association patterns in the atlas, it learns model initialization parameters that can quickly adapt to new tasks. When encountering new entities, only a small number of labeled samples are needed for relatively reliable association predictions.

[0058] To achieve drug interaction (DDI) prediction, this system constructs a pharmaceutical knowledge graph G=(V,E), where V represents the drug / target entity and E represents the interaction relationship.

[0059] Temporal Feature Mining: Time-series data from clinical trials are processed using Gated Cyclic Units (GRUs). The update gate zt formula for the GRU is: zt=σ(Wz.[ht-1,xt]); Inference Engine: Utilizes Graph Convolutional Neural Networks (GCNs) to perform evolutionary reasoning on drug nodes, predicting potential new drug indications. Its hierarchical transformation formula is as follows: H(l+1)=σ() This allows us to uncover potential "drug-protein-disease" connections hidden in the literature.

[0060] 5. Knowledge Update Module: RLHF Feedback Closed Loop; Fine-grained reward model design: The reward value *r* comes not only from the binary feedback of "useful / useless". The system defines a set of multi-dimensional feedback labels, such as: "Accurate information", "Clear logic", "Innovative suggestion but high risk", and "Outdated evidence". Each label corresponds to a reward sub-item, which together constitute the total reward. This allows the model to understand more specific human preferences.

[0061] Safety barriers and stability training: To prevent the model from "going astray" in pursuit of high rewards during the RLHF optimization process (such as fabricating seemingly reasonable but false information), a KL divergence penalty term is added to the objective function of the PPO algorithm to constrain the difference between the new policy and the original safety policy, ensuring the stability and security of knowledge updates.

[0062] Active learning-driven data acquisition: The knowledge update module and analysts' usage logs identify knowledge areas with low system confidence but high R&D attention. The data acquisition module is then proactively triggered to supplement data and retrain models for these "knowledge blind spots," achieving targeted self-improvement.

[0063] This invention introduces reinforcement learning based on human feedback (RLHF).

[0064] Reward Model: When developers rate the decision suggestions provided by the system as "useful / useless", the system converts them into a reward value $r$.

[0065] Policy optimization: Update the model parameters θ using the Proximal Policy Optimization (PPO) algorithm. The objective function is: J(θ) = ; This gives the knowledge base system the ability to "self-evolve," and as it is used more often, the medical advice it provides will be more aligned with the company's R&D preferences.

[0066] The system workflow and methods are as follows: The method described in this invention achieves the following process through the coordinated operation of the above-mentioned systems: 1. Knowledge Construction Stage: Starting from multi-source data, through preprocessing, extraction, and integration, a high-quality, structured initial pharmaceutical R&D knowledge graph that has been validated by chemical and biomedical rules is constructed.

[0067] 2. Knowledge Service and Reasoning Stage: Users pose questions or hypotheses using natural language. The system then dispatches resources through the intelligent hub (Dify). a. Semantic understanding and query decomposition: The problem is parsed using a large language model and broken down into sub-tasks such as "entity recognition", "relationship query", "evidence summarization", and "comparative analysis".

[0068] b. Multi-engine collaborative retrieval and reasoning: Path queries are performed in the graph database, full-text semantic retrieval is performed in the literature database, and the reasoning module is called to predict time-series data and potential associations.

[0069] c. Comprehensive answer generation and source tracing: Integrate the results of each subtask to generate a structured report, and automatically label the knowledge source and confidence score of each key conclusion.

[0070] 3. Knowledge Evolution Stage: Users provide feedback on the system output, or researchers input new experimental data. This new information enters the RLHF closed loop and active learning process, driving incremental updates of the knowledge graph and parameter optimization of the inference model, enabling the entire system to achieve a spiral-like intelligent evolution.

[0071] Example 1: This invention provides an AI-based knowledge base system and method for the pharmaceutical R&D industry, comprising a data acquisition module, a data preprocessing module, a knowledge extraction module, a knowledge integration module, a knowledge reasoning module, and a knowledge update module. The data acquisition module is responsible for acquiring data from multiple data sources, including scientific literature, clinical trial reports, regulatory documents, and internal company materials. The acquired data includes various multimodal data formats such as text, image, and audio. Specifically, the data acquisition module uses the Scrapy framework combined with Selenium to crawl open databases, such as the Chinese Academy of Sciences Open Data Platform, CNIPA, and the National Bureau of Statistics of China. The data preprocessing module preprocesses the acquired data, including data cleaning, standardization, synonym replacement, and entity disambiguation, ensuring data consistency and accuracy and providing a foundation for subsequent knowledge extraction. Specifically, the data preprocessing module uses the natural language processing tool TF-IDF to simultaneously extract keywords from the input technical text and the crawled data, assigning time weights to each crawled data entry and using time-weighted filtering to identify technical content associated with the keywords in the input technical text. The knowledge extraction module employs techniques such as reinforcement learning, machine learning, and Dify to perform deep analysis and mining on preprocessed data, including entity recognition, relation abstraction, and semantic parsing, to construct an initial knowledge structure. Specifically, the knowledge extraction module uses a pre-trained Transformer model, with the following model tasks: Task 1, identifying technical terms; Task 2, determining the technical field to which the input text belongs; and Task 3, generating semantic embeddings and optimizing term representations.

[0072] The knowledge integration module utilizes a chemical rules knowledge base to verify and correct the knowledge extraction results. Through an iterative verification process, it ensures the accuracy and rationality of the knowledge. Specifically, this includes: first, establishing a chemical rules knowledge base based on chemical principles and expert experience; then, matching the knowledge extraction results with the chemical rules knowledge base to identify parts that conflict with known chemical rules or may cause side reactions; and finally, correcting parameter combinations based on the verification results, including adjusting reagent ratios or reaction conditions, or generating warning messages.

[0073] The knowledge reasoning module is implemented based on a deep learning model. Through in-depth analysis and mining of knowledge structures, it provides intelligent decision support and knowledge services. Specifically, it includes: first, building a deep learning model, including a natural language processing model and a machine learning model; then, inputting the knowledge structure into the model for training, training the model to recognize and understand the relationships between knowledge; and finally, using the trained model for reasoning and prediction to generate derived results from the knowledge.

[0074] The knowledge update module is responsible for the system's continuous learning and self-improvement, including feedback and self-learning mechanisms. By analyzing and integrating user feedback information, combined with the self-learning capabilities of deep learning models, the system can continuously optimize and improve the accuracy and reliability of its knowledge.

[0075] Example 2: The data acquisition module of this invention obtains data from the following data sources: scientific literature from the Chinese Academy of Sciences Open Data Platform; clinical trial reports from ClinicalTrials.gov; regulatory documents from the National Medical Products Administration; and internal company documents from the company's internal server. The acquired data includes a variety of multimodal data such as literature texts, clinical trial report texts, regulatory document texts, internal company report texts, research paper images, clinical trial data, drug development flowcharts, and patent documents.

[0076] The data preprocessing module of this invention performs the following preprocessing on the acquired data: First, it cleans the text data, removing stop words, punctuation, and special characters such as numbers; then, it performs standardization, unifying data of different formats into a unified format; next, it performs synonym replacement, replacing synonyms with unified vocabulary expressions; finally, it performs entity disambiguation, unifying entities such as organizations / companies / drugs into standard entity representations. The knowledge extraction module of this invention uses the Transformer model for knowledge extraction, with the following specific steps: First, it segments the input text into words, using BERT's WordPiece segmenter to split the text into sub-word units and defining special tokens; then, it maps each token to an ID in the vocabulary and adds positional encoding to each token; next, it converts the TokenIDs in the input layer into embedding vectors, captures the contextual relationships between tokens through the Transformer encoder, and generates context-dependent hidden representations; finally, it performs term recognition through a conditional random field and outputs the recognition results.

[0077] The knowledge integration module of this invention establishes a chemical rule knowledge base, including the following rules: 1. Rules governing the hydrolysis of drugs; 2. Drug metabolism rules; 3. Rules governing the pharmacological mechanisms of drugs; 4. Rules governing adverse drug reactions; 5. Rules for the use of medications.

[0078] The extracted knowledge is matched with a chemical rules knowledge base to identify conflicting or potentially risky parameter combinations, generating warning messages or correction suggestions.

[0079] The knowledge reasoning module of this invention constructs a deep learning model, including: 1. Natural Language Processing Model: Using the BERT model for text feature extraction; 2. Machine learning model: Using a GRU network for knowledge relationship mining. The knowledge structure is input into the model for training, enabling the model to recognize and understand the relationships between knowledge points. The trained model is then used for reasoning and prediction to generate derived results, such as the interaction between drug A and drug B, or the metabolic process of drugs.

[0080] The knowledge update module of this invention includes a feedback mechanism, comprising: 1. User feedback mechanism to receive user evaluations and suggestions on system output; 2. The system employs a self-learning mechanism, optimizing the model's parameter combinations through iterative training on massive amounts of data. These mechanisms enable the system to continuously optimize and improve the accuracy and reliability of its knowledge.

[0081] In summary, by utilizing the technical solutions described above in this invention, a hierarchical "data-knowledge-service" architecture centered on an intelligent central layer is constructed. This architecture integrates multimodal data preprocessing, domain-adaptive knowledge extraction, chemical rule and symbolic reasoning fusion, temporal graph joint reasoning, and a reinforcement learning update mechanism based on human feedback. This effectively integrates multi-source heterogeneous pharmaceutical data, significantly improves the semantic understanding of complex medical terms and biochemical relationships, and constructs a dynamic knowledge graph with self-evolutionary capabilities. This provides researchers with traceable, explainable, and highly confident decision support, greatly improving the efficiency and intelligence level of knowledge acquisition in pharmaceutical research and development.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A knowledge base system for the pharmaceutical R&D industry based on artificial intelligence, characterized in that: include: The system includes a data acquisition module, a data preprocessing module, a knowledge extraction module, a knowledge integration module, a knowledge reasoning module, a knowledge update module, and an intelligent central layer. The data acquisition module is used to acquire multimodal pharmaceutical R&D data from multiple data sources; The data preprocessing module is used to clean, standardize, disambiguate entities, and align features of the acquired data. The knowledge extraction module uses a deep learning model based on the Transformer architecture to extract entities and relationships between entities from preprocessed data. The knowledge integration module integrates a chemical rule knowledge base, which is used to perform rule-based verification and conflict correction on the extracted knowledge in order to construct a structured knowledge graph. The knowledge reasoning module, based on graph neural networks and time-series models, is used to perform association reasoning and prediction on the knowledge graph; The knowledge update module, based on a human feedback reinforcement learning mechanism, is used to optimize system parameters and drive knowledge updates based on user feedback. The intelligent central layer is built on a large language model orchestration framework and is used to schedule tasks and orchestrate workflows for the data acquisition module, data preprocessing module, knowledge extraction module, knowledge integration module, knowledge reasoning module, and knowledge update module, so as to collaboratively complete the knowledge construction, service, and evolution process.

2. The AI-based pharmaceutical R&D industry knowledge base system according to claim 1, characterized in that, The data acquisition module adopts a web crawler framework with an integrated reinforcement learning dynamic scheduler, which is used to dynamically adjust the crawling frequency and depth of different data sources based on the stage of the R&D project, the authority of the data source, and historical usage feedback. The data preprocessing module executes a timeliness-relevance joint weighting algorithm to calculate a comprehensive weight for each piece of acquired data. The comprehensive weight is determined based on the semantic similarity between the data and the input technical text, the time decay factor, the domain influence factor of the data source, and the level of clinical evidence. The data preprocessing module also includes a multimodal feature alignment unit, which is used to align the semantic features of image data and text data through a cross-modal attention network.

3. The AI-based pharmaceutical R&D industry knowledge base system according to claim 2, characterized in that, In the aforementioned timeliness-relevance joint weighted algorithm, the formula for calculating the comprehensive weight W is: ； Where S is the cosine similarity between the input technical text and the crawled data, R is the time decay factor, I is the domain influence factor, L is the clinical evidence level, and α, β, and γ are adjustable hyperparameters.

4. The AI-based pharmaceutical R&D industry knowledge base system according to claim 1, characterized in that, The knowledge extraction module uses a model pre-trained on biomedical text for domain-adaptive fine-tuning, and introduces syntactic dependency tree mask and biochemical mask matrix to jointly constrain the calculation of attention weights; The knowledge extraction module adopts a multi-task learning architecture, with a named entity recognition head, a relation extraction head, and an event extraction head connected in parallel after a shared Transformer encoder.

5. The AI-based pharmaceutical R&D industry knowledge base system according to claim 1, characterized in that, The chemical rule knowledge base in the knowledge integration module includes at least a reaction rule layer, a physicochemical property constraint layer, and a safety and compliance layer. The knowledge integration module performs multi-dimensional evaluation of the extracted knowledge through an iterative verification function. The dimensions evaluated by the verification function include the free energy of chemical reactions, safety score, synthesis feasibility score, and intellectual property risk score. When a knowledge conflict is detected, the intelligent central layer invokes the orchestrated ReAct workflow to perform retrieval and correction operations.

6. The AI-based pharmaceutical R&D industry knowledge base system according to claim 1, characterized in that, The knowledge reasoning module constructs a heterogeneous graph knowledge graph and uses a relational graph convolutional network for node representation learning. At the same time, it uses a time series model to encode the time series data of clinical trials and updates the obtained representation vectors as dynamic attributes to the knowledge graph.

7. The AI-based pharmaceutical R&D industry knowledge base system according to claim 1, characterized in that, The reward model in the knowledge update module calculates reward values ​​based on multidimensional feedback labels and incorporates a KL divergence penalty term during strategy optimization to constrain strategy updates. The knowledge update module is also used to identify knowledge blind spots based on system confidence and R&D attention, and trigger the data acquisition module to supplement targeted data.

8. A method for implementing an artificial intelligence-based knowledge base system for the pharmaceutical R&D industry, characterized in that, Applied to the system as described in any one of claims 1-7, the method comprises: Knowledge construction phase: Multi-source heterogeneous data is acquired and processed through the data acquisition module and data preprocessing module; knowledge is extracted, verified and integrated through the knowledge extraction module and knowledge integration module to construct a pharmaceutical R&D knowledge graph; Knowledge service and reasoning stage: The intelligent central layer parses the user's natural language query and schedules the knowledge reasoning module and the underlying knowledge base to perform collaborative retrieval and reasoning, generating and outputting structured results with knowledge tracing and confidence assessment; Knowledge evolution stage: User feedback is collected through the knowledge update module, and the system parameters are optimized and the knowledge graph is incrementally updated based on the human feedback reinforcement learning mechanism.

9. The method for implementing an artificial intelligence-based pharmaceutical R&D industry knowledge base system according to claim 8, characterized in that, In the knowledge construction phase, the knowledge integration module uses a chemical rule knowledge base to perform conflict detection and verification on the extracted "drug-action-target" triplet. The output of the verification function is a multi-dimensional evaluation vector that includes reaction Gibbs free energy, safety score, synthesis feasibility score and intellectual property risk score.

10. The method for implementing an artificial intelligence-based pharmaceutical R&D industry knowledge base system according to claim 8, characterized in that, During the knowledge evolution stage, the knowledge update module updates the model parameters based on the near-end policy optimization algorithm. Its optimization objective function includes a KL divergence penalty term to constrain the difference between the new policy and the original security policy.

Images