Artificial Intelligence-Driven Methods and Systems for Constructing Disease Gene Pharmacology Association Databases

Abstract

This invention discloses an artificial intelligence-driven method and system for constructing a disease gene pharmacology association database, belonging to the interdisciplinary fields of computer software, bioinformatics, and precision medicine. This invention collects multi-source heterogeneous data through a deep reinforcement learning intelligent crawler, and achieves data standardization and quality control through a GNN-DTI model and a three-level labeling system; it employs a MySQL hierarchical storage engine to support high concurrency and large-scale expansion; it combines a T5 model and D3.js to provide natural language retrieval and visual interaction, and uses a multi-language SDK and containerized deployment to achieve pluggable expansion. This invention realizes a closed loop of data acquisition, processing, storage, and application, with a data confidence level exceeding 0.85, millisecond-level query response, and supports six major scenarios including precision medicine, providing efficient decision support for medical, pharmaceutical, and scientific research fields.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of computer software, bioinformatics and precision medicine, and in particular to an artificial intelligence-driven method and system for constructing a disease gene pharmacology association database. Background Technology

[0002] With the rapid development of high-throughput sequencing technologies (such as RNA-seq and single-cell sequencing), electronic medical record systems, and pharmacovigilance platforms, disease-related gene expression data, drug molecular structure information, and evidence of drug-disease-gene associations are experiencing exponential growth. This data is typically stored across multiple heterogeneous platforms. For example, TCGA (The Cancer Genome Atlas) provides cancer genomics and transcriptome data; GEO (Gene Expression Omnibus) stores high-throughput gene expression datasets; PubMed serves as a literature repository containing a vast amount of biomedical articles; DrugBank aggregates medicinal chemistry and target information; DisGeNET integrates disease-gene association evidence; and the FDA pharmacovigilance database records adverse reaction reports. These well-known technologies form the foundation of bioinformatics and pharmacology research.

[0003] In the prior art, the technical solutions most similar to the proposal in this application mainly include heterogeneous data integration systems based on manual or semi-automated tools. For example, some existing solutions use script crawlers (such as the Python-based Scrapy framework) to manually collect data from the aforementioned platforms: for literature sources such as PubMed, abstracts or full texts are downloaded through keyword searches (such as "disease name AND drug name"), and then rule-based named entity recognition tools (such as regular expression matching) are used to extract disease, drug, and gene entities, and simple relational tables are manually constructed; for structured databases such as DrugBank, data fields are downloaded in batches using API calls (such as RESTful interfaces), and field mapping is performed through Excel or simple scripts; for datasets such as GEO, raw expression files are downloaded via FTP, and format conversion is handled manually. These solutions typically store the collected data in a single relational database (such as the standard table structure of MySQL), support basic SQL queries, and provide retrieval functions through a web interface (such as a simple Flask-based front-end). Another close alternative is some open-source knowledge base projects (such as Bio2RDF or PharmGKB), which link multi-source data through semantic web technologies (such as the RDF format): first, the ontology mapping is manually defined (such as unifying gene symbols from different sources to the HGNC standard), and then the association is retrieved using the SPARQL query language, but this relies on periodic batch updates rather than real-time collection, and the visualization is limited to static charts.

[0004] These existing technical solutions primarily rely on manual intervention and fixed rules in their implementation. They differ from the proposal in that: existing solutions have static data collection processes that are susceptible to changes in data sources (e.g., lacking dynamic adjustment mechanisms to combat anti-scraping measures); standardization relies on manual rules rather than AI models (e.g., lacking semantic relationship prediction via graph neural networks); their storage architectures are singular and do not support hierarchical optimization (e.g., lacking master-slave replication or graph database collaboration); their retrieval interfaces are limited to keyword matching rather than natural language processing; and their scalability is limited (e.g., lacking a containerized SDK to support custom applications). These differences make it difficult for existing solutions to efficiently handle data heterogeneity and scale growth.

[0005] These methods have significant drawbacks when processing disease gene expression and pharmacological association data, as follows: Compared to the proposal in this application, heterogeneous data integration systems based on manual or semi-automated tools have the following disadvantages: The data acquisition process relies on static script crawlers (such as Python's Scrapy framework) and manual keyword searching, lacking dynamic adjustment mechanisms (such as intelligent strategies based on deep reinforcement learning), making it susceptible to being blocked by anti-crawling mechanisms of data sources, leading to acquisition interruptions or low efficiency; data standardization relies on manual rules and simple tools (such as regular expression matching or Excel field mapping), unable to handle complex semantic relationships and cross-source differences (such as batch effects in gene expression profiles or inconsistent formats of drug structures), resulting in significant differences in data format, units, and naming rules, making standardization difficult and prone to errors; data quality control relies on manual intervention, unable to automatically detect conflicts, missing data, or... Redundancy (such as the lack of confidence prediction and adversarial training in graph neural networks) leads to a high proportion of data noise (potentially exceeding 30%), making it difficult to guarantee quality; the storage architecture uses a single relational database (such as the standard MySQL table structure), which does not support master-slave replication, hierarchical caching (such as Redis), and graph database collaboration (such as Neo4j), making it unable to support high-concurrency scientific research queries, resulting in long query response times and poor scalability; retrieval and application extensions are limited to basic SQL queries and simple web interfaces (such as the Flask front end), lacking natural language interfaces, intelligent visualizations (such as D3.js interactive graphs), and pluggable SDKs (such as Python / Java / R supporting containerized deployment), making it difficult to directly serve downstream scenarios such as precision medicine and drug repositioning, resulting in insufficient application scalability.

[0006] Compared to the proposal in this application, the disadvantages of this scheme for open-source knowledge base projects (such as Bio2RDF or PharmGKB) include: data collection relies on periodic batch updates and manually defined ontology mappings, lacking real-time intelligent collection (such as customized extraction templates and integrity verification), making it susceptible to changes in data sources, resulting in poor data timeliness and incomplete collection; data standardization relies on semantic web technologies (such as RDF format and manual mapping), making it impossible to integrate AI models (such as semantic relationship prediction and three-level quality labels in GNN-DTI), leading to inefficiency in handling data conflicts and redundancy, and quality control relying on manual review; the storage architecture is limited to linked data formats, does not support hierarchical optimization (such as partitioned indexes and sharding), and is difficult to handle TB-level data growth and high-concurrency access; the retrieval interface relies on the SPARQL query language, lacks natural language processing and interactive visualization (such as path highlighting and Jupyter plugins), resulting in a poor user experience and lack of support for multi-language queries; application extension is limited to static charts and federated queries, lacks webhook mechanisms and multi-language SDKs, and cannot seamlessly integrate third-party tools or support custom model deployment, making it difficult to extend to scenarios such as clinical trial optimization or adverse reaction early warning.

[0007] These shortcomings result in existing solutions being inefficient and inaccurate when faced with exponentially growing heterogeneous data, failing to meet the needs of precision medicine decision-making. Therefore, the technical problems that this application proposes to address include: achieving automated and intelligent heterogeneous data collection to avoid blocking and improve efficiency; unifying data formats and filtering noise through AI-driven standardization and quality control (such as GNN models and ComBat algorithms); building a hierarchical storage engine (such as master-slave replication and Neo4j integration) to support high concurrency and large-scale expansion; providing intelligent retrieval and visual interaction (such as T5 models and D3.js interfaces) to enhance user experience; and developing pluggable application interfaces (such as SDKs and Webhooks) to support multi-scenario derivative applications, thereby building an efficient and reliable integrated database system. Summary of the Invention

[0008] To overcome the above shortcomings, the present invention aims to provide an artificial intelligence-driven method and system for constructing a disease gene pharmacology association database, addressing the problems of scattered disease gene data sources, high standardization difficulty, low retrieval efficiency, and insufficient application scalability in existing technologies. This system utilizes AI technologies such as deep learning and graph neural networks to achieve intelligent collection, standardized processing, hierarchical storage, intelligent retrieval, and scalable applications of heterogeneous data, constructing a unified knowledge base of disease gene expression and pharmacology associations to support precision medicine decision-making, drug development, and clinical trial optimization.

[0009] To achieve the above objectives, the technical solution of the present invention is as follows: In a first aspect, the present invention provides an artificial intelligence-driven method for constructing a disease gene-pharmacology association database, the method comprising the following steps: S1, based on a dynamic crawling strategy using deep reinforcement learning and a multi-source customized extraction template, combined with MD5 hash integrity verification, yields a multimodal dataset containing text, numerical representation spectra, and structured association tables. This dataset serves as the direct processing object for S2.

[0010] By designing an intelligent crawler strategy based on deep reinforcement learning, the number of concurrent threads (e.g., initially set to 5-10 threads, automatically halved when the latency exceeds 500ms) and the crawling interval (minimum 1 second, maximum 5 seconds) are dynamically adjusted according to the robots protocol, access frequency limits, and real-time response latency of the target data source, in order to avoid being blocked and improve the collection efficiency. For different types of data sources, customized extraction templates are used: For literature data (such as PubMed articles), a pre-trained BERT-BiLSTM-CRF model is used for Named Entity Recognition (NER) to extract disease entities (such as "Alzheimer's disease"), drug entities (such as "aspirin"), gene entities (such as "APOE"), and relational terms (such as "treatment", "inhibit expression", "upregulation"), and triple relations (such as <drug, target, gene>) are extracted through dependency parse trees; For structured databases (such as DrugBank, DisGeNET), a RESTful API parser is automatically generated based on Swagger documentation to achieve field-level mapping and batch downloading; For semi-structured APIs (such as the GEO dataset), parameter queries and response parsing are automatically adapted in accordance with the OpenAPI specification, supporting pagination and incremental updates. During the collection process, data integrity verification (such as MD5 hash verification) is introduced to ensure that the collected raw data is not lost or tampered with, and finally a multimodal dataset is generated, including text, numerical expression profiles, and structured association tables.

[0011] S2 takes the multimodal dataset output by S1 as the processing object, and uses the GNN-DTI graph neural network model, adversarial training and ComBat algorithm, format unification and semantic alignment standardization methods, combined with a three-level quality labeling system to obtain gene-drug-disease association data with 20%-30% noise data filtered out. This data serves as the core storage object of S3.

[0012] A proprietary GNN-DTI (Graph Neural Network for Drug-Target Interaction) model was designed, where nodes represent drug, disease, and gene entities, and edges represent semantic relationships such as "targeting," "treatment," and "expression correlation." The training objective is to predict the confidence of edges (using the cross-entropy loss function and incorporating L2 regularization to prevent overfitting). To improve model robustness, adversarial examples were introduced (e.g., adding noise to the input graph via an FGSM attack) to train the model to distinguish between real relationships and spurious noise. Standardization processes included: applying the ComBat algorithm to batch effect correction for numerical gene expression profiles (e.g., RNA-seq data) to eliminate platform differences; uniformly converting drug molecular structures (e.g., SMILES strings) to InChIKey format to ensure cross-source consistency; and using Word2Vec embedding for semantic alignment of text entities. Quality control employs a three-tiered labeling system: Gold level (confidence ≥ 0.9, cross-validated by multi-source evidence); Silver level (0.5-0.9, predicted by the model but requiring manual review); and Bronze level (< 0.5, marked as low reliability and requiring manual verification). This process filters out approximately 20%-30% of noisy data, ensuring a high-quality foundation for the database.

[0013] S3 uses the gene-drug-disease association data output by S2 as the storage object. Based on master-slave replication architecture, Redis caching, data source + year and month partitioning strategy and multi-type index optimization, combined with Neo4j graph database auxiliary storage, it obtains a high-performance storage architecture that supports read-write separation, millisecond-level query response and TB-level data horizontal expansion.

[0014] A master-slave replication architecture is adopted. The master database stores all Gold and Silver level data, using the InnoDB engine to support row-level locking and transactional consistency (ACID properties) to handle high-concurrency writes. The slave database is set to read-only mode and uses the MyRocks engine for compressed storage, reducing disk usage by more than 50%. Redis is introduced as a caching layer to store frequently accessed query results (such as gene-drug association subgraphs), with a TTL (Time To Live) of 2 hours and support for LRU eviction policies. The partitioning strategy is based on "data source + year and month" (e.g., partition key is source_year_month) to facilitate fast data rollback and historical queries. Index optimizations include: full-text search using inverted indexes (based on the Lucene plugin); range queries (such as expression value thresholds) using B-tree indexes; and graph structure data auxiliary storage in the Neo4j database, supporting the Cypher query language. Through a layered design, read and write separation is achieved, reducing query response time to milliseconds and supporting horizontal scaling (Sharding) to cope with terabyte-level data growth.

[0015] S4 uses the structured and graph-structured data stored in S3 as the retrieval data source. Based on the T5-small model, a finely tuned natural language interface, the D3.js interactive visualization tool, a RESTful API, and Jupyter plugin integration, it achieves intelligent retrieval capabilities that support multilingual queries, mixed sentence conversion, dynamic interaction, and personalized recommendations.

[0016] The natural language interface is fine-tuned based on the T5-small model, converting user queries (such as "find genes and drugs related to diabetes") into hybrid query statements (SQL for structured data + graph query for relational paths), and improving accuracy through Beam Search decoding. The graphical interface uses the D3.js library, supporting interactive operations: node drag-and-drop (force-directed layout algorithm), subgraph filtering (edge ​​filtering by confidence threshold), and path highlighting (shortest path algorithm such as Dijkstra). It also provides a RESTful API, conforming to the OpenAPI 3.0 standard, with selectable JSON, CSV, or RDS return formats, and supports pagination and authentication (OAuth 2.0). An integrated Jupyter plugin, using ipywidgets, enables interactive visualization, allowing users to adjust query parameters in real time and view dynamic graphs. This module supports multilingual queries (English / Chinese) and records user behavior logs to optimize the recommendation algorithm.

[0017] S5 is based on the intelligent retrieval service of S4 and the stored data of S3. It is supported by multi-language SDKs, Docker+Kubernetes containerized deployment, gRPC protocol calls and Webhook event subscription mechanism, and has a scalable interface architecture that supports seamless integration of third-party tools and custom application development.

[0018] A multi-language SDK is provided: Python SDK (installable via pip install aiddb-sdk, supporting asynchronous queries); Java SDK (Maven dependency, integrated with Spring Boot); R SDK (released via CRAN, compatible with Bioconductor). Algorithm extensions are deployed using containerization (based on Docker and Kubernetes), calling custom models (such as user-uploaded custom GNN variants) via the gRPC protocol. An open webhook mechanism allows subscription to events (such as "new drug-disease evidence" or "quality label change"), with event pushes using JSON payloads and supporting retries and dead-letter queues. This interface ensures system scalability, allowing users to seamlessly integrate third-party tools for custom application development.

[0019] Secondly, the present invention also provides an artificial intelligence-driven system for constructing a disease gene-pharmacology association database, the system comprising: The heterogeneous data source intelligent acquisition module is used to obtain a multimodal dataset containing text, numerical expression spectrum and structured association table by combining a dynamic crawling strategy based on deep reinforcement learning and a multi-source customized extraction template with MD5 hash integrity verification. This dataset is the direct processing object of S2, as described in S1 above. The AI ​​data standardization and quality control module is used to process the multimodal dataset output by S1. Based on the GNN-DTI graph neural network model, adversarial training and ComBat algorithm, format unification, semantic alignment standardization methods, and a three-level quality labeling system, it obtains gene-drug-disease association data that filters out 20%-30% of noise data. This data serves as the core storage object of S3, as described in S2 above. The MySQL hierarchical storage engine module is used to store the gene-drug-disease association data output by S2. Based on master-slave replication architecture, Redis caching, data source + year and month partitioning strategy and multi-type index optimization, combined with Neo4j graph database auxiliary storage, a high-performance storage architecture that supports read-write separation, millisecond-level query response and TB-level data horizontal expansion is obtained, as described in S3 above. The intelligent retrieval and visualization interaction module is used to retrieve structured and graph-structured data stored in S3 as the data source. It integrates a natural language interface finely tuned based on the T5-small model, the D3.js interactive visualization tool, the RESTful API, and the Jupyter plugin to achieve intelligent retrieval capabilities that support multilingual queries, mixed sentence conversion, dynamic interaction, and personalized recommendations, as described in S4 above. The pluggable application extension interface module is used to support the intelligent retrieval service of S4 and the storage data of S3. Based on multi-language SDK, Docker+Kubernetes containerized deployment, gRPC protocol call and Webhook event subscription mechanism, it provides an extensible interface architecture that supports seamless integration of third-party tools and custom application development, as described in S5 above.

[0020] Based on the above methods and systems, this database can be used to develop the following six core applications, further expanding its practical value in the field of precision medicine: 1. Precision Medication Decision Support System. Inputs include the patient's genomic VCF file, clinical phenotype description (e.g., "hypertension with diabetes"), and medication history. The processing flow is as follows: first, gene-drug-disease triples are matched (using GNN embedding similarity calculation); then, drug candidates are ranked using a random forest model (considering efficacy, side effects, and interactions); finally, an XGBoost model is applied to predict personalized dosages (trained based on historical data). The output is a medication regimen report, including a list of recommended drugs, evidence levels (Gold / Silver), and potential risk warnings. This application can reduce medication error rates by 15%-20% and is suitable for clinical decision support.

[0021] 2. Drug Relocation Discovery Engine. Utilizing the link prediction capabilities of GNNs (trained on a known drug-disease map), this engine discovers candidates for "drug repurposing" (e.g., aspirin for cancer prevention); it also integrates with the ClinicalTrials.gov registry to screen indications already in Phase II trials; and it automatically generates reports including relocation path visualization, confidence scores, and links to relevant literature evidence. This engine accelerates drug development cycles and could potentially save hundreds of millions of dollars in costs.

[0022] 3. Disease Subtyping and Prognostic Prediction. Integrating gene expression profiles, drug response data, and survival curves, a multi-task deep learning model (based on the ResNet architecture) is trained: Task 1 is subtype classification (K-means initialization), Task 2 is risk scoring (Cox proportional hazards model), and Task 3 is treatment recommendation (based on reinforcement learning). Outputs include molecular subtype labels (e.g., "hormone receptor-positive breast cancer"), 5-year survival prediction, and personalized treatment plans. This application improves prognostic accuracy to over 85% and supports precise tumor subtyping.

[0023] 4. Optimization of Clinical Trial Patient Recruitment. The system parses trial inclusion and exclusion criteria (such as the XML file corresponding to "NCT number") and converts them into structured Boolean expressions (e.g., "Age > 18 AND BRCA1 mutation"); it automatically matches patients' electronic medical records (EHRs) using TF-IDF vector similarity to generate a recruitment priority list; and it provides a visual funnel plot (rendered using D3.js) to show the conversion rate from screening to enrollment. This optimization can reduce recruitment time by 30% and improve trial efficiency.

[0024] 5. Early warning of adverse drug reactions. By integrating FAERS real-time data streams and combining gene mutation signals (e.g., CYP2D6 polymorphism) to construct a Bayesian risk model, when new adverse event signals exceed a threshold (e.g., PRR>2), an early warning notification (including the chain of evidence and statistical p-value) is sent to pharmaceutical companies or regulatory agencies. This application helps prevent drug recalls and protect patient safety.

[0025] 6. Automated construction of scientific knowledge graphs. New evidence is written to Neo4j in real time (nodes / edges are updated using Cypher statements); it supports SPARQL queries (federated queries of multi-source graphs), GNN inference (predicting potential associations), and visualization exploration (force-directed layout + heatmap); it exports OWL / RDF format for reuse by downstream systems (such as genome browsers). This construction tool simplifies the research process and supports interdisciplinary collaboration.

[0026] The technical solution of this invention achieves a closed loop from data collection to application through AI-driven modular design, improving the accuracy (average confidence level >0.85) and efficiency (query response <1s) of the database, and supports extended applications, making it suitable for medical institutions, pharmaceutical companies and research units.

[0027] By adopting the above technical solution, the beneficial technical effects obtained by the present invention are as follows: This invention, through an AI-driven modular design, constructs an integrated database system linking disease gene expression and pharmacology, which significantly outperforms existing technologies in the following aspects: (i) Efficient and intelligent data acquisition: Through intelligent crawling strategies based on deep reinforcement learning, the anti-crawling mechanisms and access restrictions of the data source are dynamically adapted (such as dynamically adjusting the number of threads and crawling interval). Compared with traditional static script crawlers, the acquisition efficiency is improved by about 30%-50%, and the blocking is effectively avoided, ensuring the stability and integrity of data acquisition.

[0028] (ii) High-quality data standardization and quality control: The GNN-DTI graph neural network model and adversarial training technology are adopted, combined with the ComBat algorithm and semantic alignment, to automatically handle format differences, batch effects and semantic ambiguity of heterogeneous data sources. The data noise filtering rate reaches 20%-30% and the confidence level averages above 0.85. Compared with traditional manual rule standardization, it significantly improves data consistency and reliability.

[0029] (III) High-performance hierarchical storage and query: Through master-slave replication architecture, Redis caching, Neo4j graph database and partitioned index optimization, read-write separation and millisecond-level query response are achieved, supporting TB-level data growth and high-concurrency access. Compared with a single relational database solution, query latency is reduced to 1 / 10 of the original, and storage space usage is reduced by more than 50%.

[0030] (iv) Intelligent search and interactive experience: Based on the T5-small model, natural language query and D3.js interactive visualization support multilingual input, dynamic graph display and evidence backtracking. Compared with traditional keyword search, the user query efficiency is improved by about 40%, and the interactive experience is more intuitive, which is suitable for scientific research and clinical scenarios.

[0031] (v) Powerful application scalability: Through multi-language SDKs (Python / Java / R), containerized deployment (Docker / Kubernetes) and Webhook mechanism, the system can seamlessly integrate third-party tools and support custom model extensions. Compared with existing static knowledge base solutions, the development cycle is shortened by 30%-50%, and it can flexibly support six core applications such as precision medicine, drug repositioning, and clinical trial optimization.

[0032] (vi) Practical value in multiple scenarios: By integrating gene-drug-disease triplet and multi-task deep learning, this system significantly improves the application effect of precision medicine. For example, it reduces the medication error rate by 15%-20%, shortens the drug repositioning cycle (saving hundreds of millions of dollars in costs), improves the prognostic accuracy of disease subtypes to over 85%, optimizes the recruitment efficiency of clinical trials by 30%, and provides early warning of adverse drug reactions and automated knowledge graph construction, providing medical institutions, pharmaceutical companies and research units with efficient and reliable decision support tools. Attached Figure Description

[0033] Figure 1 The present invention provides an overall architecture flowchart of an artificial intelligence-driven integrated database system for disease gene expression and pharmacology. Detailed Implementation

[0034] The present invention is illustrated by the following examples.

[0035] like Figure 1 As shown, Figure 1 The data stream, as demonstrated, begins from multiple heterogeneous data sources (such as TCGA, GEO, PubMed, DrugBank, DisGeNET, and FDA), and is processed through AI data standardization (including intelligent collection, quality control, and standardization of heterogeneous data sources) to generate high-quality gene-drug-disease association data. The data then enters a MySQL hierarchical storage engine (including a master database, slave databases, Redis cache, and Neo4j graph database), ultimately supporting intelligent retrieval, interactive visualization, and pluggable application extension modules.

[0036] like Figure 1 The image shows an embodiment of the present invention: an artificial intelligence-driven integrated database system and method for disease gene expression and pharmacology association, comprising: Step S1: Based on the dynamic crawling strategy of deep reinforcement learning and the multi-source customized extraction template, combined with MD5 hash integrity verification, a multimodal dataset containing text, numerical representation spectrum and structured association table is obtained. This dataset is used as the direct processing object in S2.

[0037] Data was collected from heterogeneous data sources, including TCGA (Cancer Genome Atlas, providing cancer genome and transcriptome data), GEO (Gene Expression Database, storing high-throughput gene expression datasets), PubMed (Biomedical Literature Repository, containing a vast number of articles), DrugBank (Medicinal Chemistry and Target Information Repository), DisGeNET (Disease-Gene Association Evidence Integration Repository), and FDA (Pharmacovigilance Database, recording adverse reaction reports). The system employs a dynamic scheduler based on deep reinforcement learning (RL) to manage the data collection process. State variables include target site response latency, failure rate, rate limit, and the estimated probability of being blocked in the nearest window. The action space includes adjustments to the number of concurrent threads (initially set to 5-10 threads, automatically halved when the response latency exceeds 500ms, down to a minimum of 1 thread), adjustments to the crawling interval (an adaptive range of 1 second to 5 seconds, dynamically changing based on latency), UA (User Agent) pool rotation (pool size k=20), and agent switching (rotation window w=100 requests). The reward function integrates "number of valid documents per unit time," "failure penalty" (e.g., maximum retry limit r=3, exponential backoff base b=2), and "blocking penalty" (probability estimation based on historical access logs). Compared to traditional static crawlers, this mechanism improves data collection efficiency by approximately 30%-50% and effectively avoids blocking.

[0038] For different data sources, customized extraction templates are used: For literature data (such as PubMed articles), a pre-trained BERT-BiLSTM-CRF model is used for named entity recognition (NER) to extract disease entities (such as "Alzheimer's disease"), drug entities (such as "aspirin"), gene entities (such as "APOE"), and relational terms (such as "treatment", "inhibit expression", "upregulation"), and triple relations (such as <drug, target, gene>) are extracted through dependency parse tree; entity standard names are aligned to authoritative ontologies (such as MeSH terms, HGNC gene symbols, DrugBank preferred names), and a reversible mapping table is established to reduce semantic ambiguity; the extracted triple records confidence information, including the source literature PMID, paragraph position, intra-sentence offset, and context window. Structured API data (such as DrugBank and DisGeNET) is automatically generated into a RESTful API parser based on Swagger documentation, enabling field-level mapping, batch downloading, pagination, and breakpoint resumption. Semi-structured API data (such as the GEO dataset) automatically adapts parameter queries and response parsing according to the OpenAPI specification, supporting incremental updates (collecting only changes since the last version). During the collection process, data integrity verification (such as calculating MD5 / SHA256 fingerprints) and lineage records (source_id, version, ingest_time) are introduced to ensure that the original data is not lost or tampered with, ultimately generating a multimodal dataset, including text, numerical expression spectra, and structured association tables. After the raw data enters the lake (Raw Zone), it undergoes light normalization (field renaming, unit unification such as log2(TPM+1), timestamp normalization to UTC) and missing value labeling (NA / NULL identifiers and traceable reason codes, such as "source missing" or "parsing failed"), and then enters the "candidate evidence queue" to await subsequent standardization processes.

[0039] Step S2 uses the multimodal dataset output from S1 as the processing object. Based on the GNN-DTI graph neural network model, adversarial training and ComBat algorithm, format unification, semantic alignment standardization, and a three-level quality labeling system, gene-drug-disease association data with 20%-30% noise data is obtained. This data serves as the core storage object in S3.

[0040] A proprietary GNN-DTI (Graph Neural Network for Drug-Target Interaction) model was used to perform consistency tests and confidence scores on the relationship between "drug-target-disease (and symptoms / phenotype)". During graph construction, node types included Drug, Gene, Disease, Phenotype, and Trial; edge types included targets, treatments, associates_with, upregulates / downregulates, contraindicated_for, and adverse_event_of. Each node contained structured attributes (such as the InChIKey / ATC code for drugs, Entrez ID for genes, and MeSH / DOID identifiers for diseases), and edges included features such as source evidence count, textual evidence strength, and statistical significance (such as p-value).

[0041] Model Structure Example: A heterogeneous graph convolutional neural network (Hetero-GNN) and relational attention (R-GAT) mechanism are stacked in 2-3 layers, with a hidden layer dimension d=128. Residual connections and LayerNorm are combined to improve stability. The loss function is cross-entropy with a temperature parameter (temperature scaling τ=2.0), and threshold calibration uses Temperature Scaling. The training set consists of known high-confidence relations (multi-source cross-validation) and negative sampling pairs (negative sampling ratio 1:3). Adversarial training is introduced (e.g., FGSM attack injects perturbations into node / edge features, Dropout=0.2, L2 regularization=1e-5) to improve model robustness and transferability. The output confidence R∈[0,1] is used to generate an evidence attribution (attention heatmap) for interpretability demonstration.

[0042] Standardization processes include: applying the ComBat algorithm to batch effect correction for numerical gene expression profiles (such as RNA-seq data) to eliminate cross-platform differences (formula: Y_corrected = Y - B * β, where B is the batch design matrix and β is the estimated parameter); uniformly converting drug molecular structures (such as SMILES strings) to InChIKey format to ensure cross-source consistency and deduplication; and using Word2Vec embedding (or sub-word embedding) for semantic alignment of text entities, supporting cross-language (such as mixed Chinese-English vocabularies). Quality control employs a three-tiered labeling system: Gold level (R≥0.9, cross-validated by multi-source evidence, directly added to the database); Silver level (0.5≤R<0.9, predicted by the model but requiring manual review, entering a semi-automatic review queue); and Bronze level (R<0.5, marked as low reliability, retained in the archive but not included in the main index). Conflict resolution strategy priority: multi-source consistency > clinical trial evidence > high-quality reviews > single-source original studies; when conflicts are unresolved, the relationship is locked and marked "further evidence required". Quality metrics are evaluated on a batch-by-batch basis, assessing precision / recall, cross-batch stability (Kappa coefficient), duplicate entries, and evidence drift. When batch quality falls below a threshold (e.g., precision < 0.85), the system rolls back to the previous stable snapshot. All quality control decisions and parameter snapshots (model version, thresholds, training data windows) are archived and made traceable. This step filters out approximately 20%-30% of noisy data, ensuring a high-quality foundation for the database.

[0043] Step S3 uses the gene-drug-disease association data output from S2 as the storage object. Based on master-slave replication architecture, Redis caching, data source + year and month partitioning strategy and multi-type index optimization, combined with Neo4j graph database auxiliary storage, a high-performance storage architecture supporting read-write separation, millisecond-level query response and TB-level data horizontal expansion is obtained.

[0044] The architecture employs a four-layer structure: master-slave-cacher-graph. The master database stores all Gold and Silver level data, using the InnoDB engine to support row-level locking and transactional consistency (ACID properties) to handle high-concurrency writes. The slave database is set to read-only mode and uses the MyRocks engine for compressed storage, reducing disk usage by more than 50%. Redis is introduced as a caching layer to store hot query results (such as gene-drug association subgraphs), with a TTL (Time To Live) of 2 hours, supporting LRU eviction policies and cache breakdown protection (mutex locks and short-term null value caching). Graph structure data is supplementarily stored in the Neo4j database, supporting the Cypher query language for complex path retrieval.

[0045] The database is divided into sub-databases and tables by topic: DB_Disease (disease database), DB_Drug (drug database), DB_Gene (gene database), and DB_Clinical (clinical database). Key tables include disease (disease_id, mesh_id, name, synonyms, …), drug (drug_id, inchikey, atc, name, …), gene (entrez_id, symbol, alias, …), relation (src_id, dst_id, rel_type, evidence_score R, evidence_ref, lvl), expression (dataset_id, gene_id, sample_id, value_norm, batch), and ae_signal (drug_id, meddra_term, prr, ror, pval, window), etc. The partitioning strategy is based on "data source + year and month" (e.g., the partition key is source_year_month), which facilitates fast data rollback and historical queries; index optimization includes: using inverted indexes for full-text search (based on the Lucene plugin); using B-tree indexes for range queries (e.g., expression value thresholds); using hash indexes for equality queries; and creating composite indexes for high-frequency dimensions (e.g., idx_rel_type_src_dst, idx_gene_symbol, idx_drug_inchikey).

[0046] Query acceleration: Cross-topic joins prioritize materialized views or result snapshots to reduce latency; complex relationship retrieval achieves sub-second response times through a hybrid strategy of "SQL pre-screening + graph database path expansion" (such as Dijkstra's shortest path algorithm). Versioning and rollback: Incremental data is stored as "monthly partition + version number (ver)"; a full database snapshot ID (snapshot_id) is generated before ingestion, and abnormal batches can be rolled back based on snapshots. Backup strategy is "daily incremental + weekly full + off-site disaster recovery," with a recovery point objective (RPO) ≤ 24 hours. Through layered design, read / write separation is achieved, reducing query response time to milliseconds, and horizontal scaling (Sharding) is supported to cope with TB-level data growth.

[0047] Step S4 uses the structured and graph-structured data stored in S3 as the retrieval data source. Based on the T5-small model-fine-tuned natural language interface, D3.js interactive visualization tool, RESTful API and Jupyter plugin integration, intelligent retrieval capabilities supporting multilingual queries, mixed sentence conversion, dynamic interaction and personalized recommendations are obtained.

[0048] The natural language interface is fine-tuned based on the T5-small model, performing intent recognition and slot filling. It converts user queries (such as "find genes and drugs related to diabetes") into hybrid query statements (SQL for structured data filtering + graph query for relational path calculation), and improves accuracy through Beam Search decoding (width 5). Execution is performed after secondary constraints from the rule base and permission system. The graphical interface uses the D3.js library and supports interactive operations: node dragging (force-directed layout algorithm), subgraph filtering (edges are filtered by confidence threshold, such as adjusting the threshold slider to R≥0.7), path highlighting (shortest path algorithm such as Dijkstra); it supports linked charts and tables, and evidence backtracking (clicking on an edge pops up source literature, PMID, and paragraph fragments). It also provides a RESTful API, conforming to the OpenAPI 3.0 standard, with selectable return formats of JSON, CSV, or RDS, supporting pagination (50 items per page by default), authentication (OAuth 2.0), and rate limiting (100 times / minute). This module integrates a Jupyter plugin and uses ipywidgets for interactive visualization, allowing users to adjust query parameters (such as expression thresholds or relation types) in real time and view dynamic graphs. It supports multilingual queries (English / Chinese) and logs user behavior to optimize recommendation algorithms (e.g., pre-caching of popular queries based on behavior logs).

[0049] Step S5, based on the intelligent retrieval service of S4 and the stored data of S3, and supported by multi-language SDK, Docker+Kubernetes containerized deployment, gRPC protocol calls and Webhook event subscription mechanism, obtains an extensible interface architecture that supports seamless integration of third-party tools and custom application development.

[0050] A multi-language SDK is provided: Python SDK (installable via pip install aiddb-sdk, supporting asynchronous queries and batch graph retrieval); Java SDK (Maven dependency, integrated with Spring Boot, providing a quick start template); R SDK (released via CRAN, compatible with Bioconductor workflows, supporting bioinformatics pipeline integration). Algorithm extensions are deployed using containerization (based on Docker images and Kubernetes orchestration), calling custom models (such as user-uploaded custom GNN variants, supporting seamless hot updates and switching) via the gRPC protocol. An open Webhook mechanism allows subscription to events (such as "new drug - disease evidence" or "quality label change"), with event pushes using JSON payloads, supporting retries (up to 3 times, exponential backoff) and a dead-letter queue (storing failure events for auditing). This interface ensures system scalability, allowing users to seamlessly integrate third-party tools (such as interface with hospital HIS / PACS, following FHIR standards) to achieve custom application development. Security and Compliance: Data transmission uses TLS 1.2+, and encryption at rest uses Transparent Data Encryption (TDE); access control uses RBAC and fine-grained row and column level permissions; audit logs and tamper detection are provided; patient-level data is de-identified (spoofed, noise-injected, and access-isolated).

[0051] The typical business process in this embodiment includes precision medication decision support: inputting the patient's VCF file, clinical phenotype (e.g., "hypertension with diabetes"), and medication history; the processing flow is as follows: ① gene variation analysis and phenotype mapping; ② calculation of candidate drug similarity based on GNN-DTI; ③ random forest ranking (considering efficacy and side effects); ④ XGBoost prediction of personalized dosage; outputting a recommendation list, evidence level, and risk warning. Similarly, drug retargeting uses link prediction to screen for "drug repurposing," combined with evidence stratification from ClinicalTrials.gov; disease subtyping uses multi-task deep learning (ResNet architecture, K-means initialization, Cox risk model), outputting molecular subtype labels, 5-year survival rate, and recommended treatment.

[0052] Example of operating parameters in this embodiment: acquisition layer N_max=16, r=3; NLP sampling learning rate 3e-5, sequence length 512; GNN layer number L=3, τ=2.0; storage latency threshold 200ms; API example such as GET / api / v1 / relations?src=EGFR&rel=targets&conf>=0.9 returns a list of drugs.

[0053] Experimental results of this embodiment: Example 1: 327 articles were retrieved from PubMed, 1482 triplets were extracted, 1093 of which were Gold level, and the query time was reduced to 180ms; Example 2: Osimertinib was recommended for NSCLC patients, with a predicted PFS of 19.2 months; Example 3: Imatinib predicted a probability of 0.82 for "systemic sclerosis".

[0054] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by program instructions and related hardware, and the program can be stored in a computer-readable storage medium, such as ROM / RAM, disk, optical disk, etc.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention (such as replacing GNN with a lightweight knowledge distillation model, using NewSQL / column engine for the storage layer, using WebGL graph rendering for the front end, and extending the API to GraphQL endpoints) should be included within the scope of protection of the present invention.

[0056] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by program instructions and related hardware. The program can be stored in a computer-readable storage medium, such as ROM / RAM, disk, optical disk, etc. The program stored in the above storage medium can efficiently execute the technical solutions of the present invention, ensuring high-performance operation and scalability of the system.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention, such as replacing the graph neural network model with a lightweight model, replacing the storage engine with NewSQL or a columnar database, replacing the front-end rendering technology with WebGL, or extending the API endpoint to GraphQL, should be included within the scope of protection of the present invention.

Claims

1. A method for constructing an artificial intelligence-driven disease gene-pharmacology association database, characterized in that, include: S1, based on a dynamic crawling strategy using deep reinforcement learning and a multi-source customized extraction template, combined with MD5 hash integrity verification, yields a multimodal dataset containing text, numerical representation spectrum, and structured association table. This dataset serves as the direct processing object for S2. S2 takes the multimodal dataset output by S1 as the processing object, and uses the GNN-DTI graph neural network model, adversarial training and ComBat algorithm, format unification and semantic alignment standardization methods, combined with a three-level quality labeling system to obtain gene-drug-disease association data that filters out 20%-30% of noise data. This data serves as the core storage object of S3. S3 uses the gene-drug-disease association data output by S2 as the storage object. Based on master-slave replication architecture, Redis caching, data source + year and month partitioning strategy and multi-type index optimization, combined with Neo4j graph database auxiliary storage, it obtains a high-performance storage architecture that supports read-write separation, millisecond-level query response and TB-level data horizontal expansion. S4 uses the structured and graph-structured data stored in S3 as the retrieval data source. Based on the T5-small model, the natural language interface is finely tuned, the D3.js interactive visualization tool is integrated with RESTful API and Jupyter plugin, resulting in intelligent retrieval capabilities that support multilingual queries, mixed sentence conversion, dynamic interaction and personalized recommendations. S5 is based on the intelligent retrieval service of S4 and the stored data of S3. It is supported by multi-language SDKs, Docker+Kubernetes containerized deployment, gRPC protocol calls and Webhook event subscription mechanism, and has a scalable interface architecture that supports seamless integration of third-party tools and custom application development.

2. An artificial intelligence-driven system for constructing a disease gene-pharmacology association database, characterized in that, include: The heterogeneous data source intelligent acquisition module is used for dynamic crawling strategies based on deep reinforcement learning and multi-source customized extraction templates. Combined with MD5 hash integrity verification, it obtains a multimodal dataset containing text, numerical expression spectrum and structured association table. This dataset is used as the direct processing object of S2. The AI ​​data standardization and quality control module is used to process the multimodal dataset output by S1. Based on the GNN-DTI graph neural network model, adversarial training and ComBat algorithm, format unification, semantic alignment standardization methods, and a three-level quality labeling system, it obtains gene-drug-disease association data that filters out 20%-30% of noise data. This data serves as the core storage object of S3. The MySQL tiered storage engine module is used to store gene-drug-disease association data output by S2. Based on master-slave replication architecture, Redis caching, data source + year and month partitioning strategy and multi-type index optimization, combined with Neo4j graph database auxiliary storage, it achieves a high-performance storage architecture that supports read-write separation, millisecond-level query response and TB-level data horizontal expansion. The intelligent retrieval and visualization interaction module uses structured and graph-structured data stored in S3 as the retrieval data source. Based on the T5-small model, the natural language interface is finely tuned, the D3.js interactive visualization tool is integrated with RESTful API and Jupyter plugin to achieve intelligent retrieval capabilities that support multilingual queries, mixed sentence conversion, dynamic interaction and personalized recommendations. The pluggable application extension interface module is used to support the intelligent retrieval service of S4 and the storage data of S3. Based on multi-language SDK, Docker+Kubernetes containerized deployment, gRPC protocol call and Webhook event subscription mechanism, it provides an extensible interface architecture that supports seamless integration of third-party tools and custom application development.

Images