Individual mutation information-based intelligent decision-making system for precise targeted medication of tumors

Abstract

The invention relates to the technical field of tumor treatment, in particular to an intelligent decision-making system for tumor precise targeted medication based on individual mutation information, which comprises a data processing layer, a variation annotation and function prediction layer, a knowledge base integration layer and a scheme decision-making engine and report visualization module. According to the intelligent decision-making system for tumor precise targeted medication based on individual mutation information, a rule engine and a prediction model are combined, dynamic priority ranking is output, multi-model fusion decision making is achieved, and clinical scene deep adaptation is achieved by predicting primary and secondary drug resistance, calculating liver and kidney function adjusting dosage and generating a combined medication time sequence scheme; through an individualized drug delivery scheme, combination drug use optimization is achieved, a visual clinical report is generated, clinical executable operation is further strengthened, and through algorithm quantification, a dynamic knowledge graph, AI auxiliary decision making and a clinical operation closed loop, the next-generation technical research direction of a tumor precise drug use system can be represented.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tumor treatment, in particular to an intelligent decision-making system for precise tumor targeted drug use based on individual mutation information. BACKGROUND

[0002] The intelligent decision-making system for precise tumor targeted drug use integrates genomic sequencing data, clinical information and multi-dimensional knowledge base, and provides personalized targeted drug use scheme decision support tools for tumor patients through artificial intelligence algorithm. The core is to analyze tumor specific mutations, match targeted drugs and predict efficacy. It has wide application prospects in the fields of cancer early screening and prevention, drug resistance mechanism cracking and cancer vaccine research and development.

[0003] According to the search, the patent for invention with the Chinese patent publication number CN109686456A discloses a tumor precise drug use interpretation system and method. The tumor precise drug use interpretation system and method can collect clinical level and literature level tumor drug use guidance evidence through the literature review subsystem, ensure the full use of data, and can update the drug use guidance evidence in the knowledge management subsystem in real time, ensure the accuracy and timeliness of the data, and can interpret the tumor precise drug use according to the variation data, can provide drug use guidance from the individual multi-omics gene detection result level, maximize the availability of drug use guidance, meet the needs of clinical guidance, and assist clinicians in making clinical drug use and treatment decisions.

[0004] However, the scheme decision recommendation of the tumor precise drug use interpretation system and method stays in static evidence matching, cannot realize dynamic optimization, and cannot provide dose calculation, drug administration timing planning, toxicity prediction and other key clinical operation support problems, so the intelligent decision-making system for precise tumor targeted drug use based on individual mutation information is proposed. SUMMARY

[0005] (I) Technical problems solved

[0006] In view of the deficiencies in the prior art, the present application provides an intelligent decision-making system for precise tumor targeted drug use based on individual mutation information, which has the advantages of realizing clinical scene depth adaptation through multi-model fusion decision-making and further strengthening the clinical executable operation, and solves the problems of low intelligent decision-making efficiency and high difficulty of clinical landing execution of the existing tumor precise drug use interpretation system and method in the above background art.

[0007] (II) Technical solutions

[0008] To achieve the above-mentioned through multi-model fusion decision realizes the clinical scene depth adaptation, further strengthens the clinical executable operation purpose, the present application provides the following technical scheme: the intelligent decision system of tumor precision targeted drug use based on individual mutation information, including data processing layer, unified multi-source heterogeneous data format, executes quality control and standardization processing, provides high quality, structured patient molecular and clinical data basis for subsequent analysis;

[0009] Variation annotation and function prediction layer, analyze the biological function of gene variation, combine population frequency and database evidence for clinical classification, and screen out key driving variation;

[0010] Knowledge base integration layer, dynamically aggregate drug-target-variation association, clinical guidelines and evidence-based research, and other authoritative knowledge, and construct a structured knowledge graph that can be queried in real time and evidence graded;

[0011] Scheme decision engine, integrate patient variation spectrum, clinical characteristics and knowledge base, generate priority-ordered drug recommendations through rule engine and model, predict drug resistance risk and toxicity, and recommend specific drug administration scheme;

[0012] Report visualization module, convert complex analysis into intuitive clinical decision report through visualization tools, support doctor-patient communication and execution.

[0013] Preferably, the data processing layer includes an interface management module, a data quality control module, a standardization conversion module, a variation detection module, a variation basic annotation module and a clinical information structured module, the interface management module accesses data input source, and the input source includes:

[0014] 1) raw file output by sequencer; 2) variation calling file provided by external laboratory; 3) clinical data derived from hospital information system and electronic medical record; 4) other omics data file;

[0015] The data quality control module calculates and evaluates the data from the interface management module, including:

[0016] 1) calculate the average Q value of each base position by FastQC, expressed as Where Q i is the Phred score of the i-th base, N is the number of bases, and the Phred quality score is Q = -10·log 10 (P error ), P error is the probability of base recognition error, to measure the probability of base recognition correctness;

[0017] 2) check the distribution of A, T, C and G in each position, and calculate the percentage of nucleotides in each position;

[0018] 3) Calculate the proportion of identical sequence reads, expressed as Duplication Rate = (Number of duplicate reads) / (Total number of reads), where high duplication rate can be caused by PCR amplification bias or low complexity regions;

[0019] 4) Scan the reads by FastQC to match the known adapter sequences, detect the contamination of sequencing adapter sequences appearing in the reads;

[0020] 5) Calculate the GC base percentage of the entire sample reads, expressed as GC% = (Count(G) + Count(C)) / (Count(A) + Count(T) + Count(G) + Count(C)) * 100%;

[0021] 6) Calculate the ratio of total reads to data volume to ensure that the preset coverage depth requirement is met;

[0022] The standardized conversion module uniformly converts data of different formats into the internal standard format of the system, specifically including:

[0023] 1) For FASTQ data, trim the sequencing adapter sequences and low-quality base fragments;

[0024] 2) For VCF data, verify the VCF file format, detect the reference genome version of the input VCF, and ensure uniform standards for genome coordinate conversion;

[0025] 3) For clinical data, standardize the data through structure conversion, field mapping, and value standardization;

[0026] The variant detection module uses alignment marking and variant calling technology to identify genomic variants from the quality-controlled and standardized FASTQ data; the variant basic annotation module performs basic annotation on the variants in the VCF file to ensure that key information uses the system uniform standard; the clinical information structured module further structures, entity associates, and quality improves the clinical data after standardization mapping.

[0027] Preferably, the variant annotation and functional prediction layer includes a functional impact prediction module, a population frequency filtering module, a driver assessment module, a clinical classification module, and an annotation integration module. The functional impact prediction module predicts the potential impact of the variant on gene function according to the output of the variant basic annotation module. The specific prediction methods include:

[0028] 1) Use SIFT method to compare homologous protein sequences, calculate the degree of tolerance to changes at specific amino acid positions, and output the actual score SIFT score and the predicted value SIFT predWhen SIFT score <0.05, the variant is determined to be harmful to the gene function;

[0029] GERP++ method is used to measure the selection constraint intensity of the genomic site in the evolution process, and the output is the rejection substitution score GERP++ RS , and the higher the positive value is, the more harmful the variant is to the gene function;

[0030] 2) PolyPhen-2 method is used to predict the impact of missense variants by combining sequence conservation, differences in physical and chemical properties of amino acids, and protein structure information, and the output is the actual score PolyPhen-2 score , and the higher the positive value is, the more harmful the variant is to the gene function; pred ;

[0031] 3) CADD method is used to integrate various annotation information, and SVM is used to train to distinguish simulated neutral variants and simulated pathogenic variants, and the output is the scaled score CADD phred , and the higher the score is, the more harmful the variant is to the gene function;

[0032] REVEL method is used to integrate the output results of the prediction tool, and a random forest model is used for training to optimize the prediction of pathogenicity of rare missense variants, and the output is the actual score REVEL phred , and the higher the score is, the stronger the predicted pathogenicity is;

[0033] 4) Deep neural network based on Inception architecture is used to predict the impact of variants on splice donor, acceptor, branch point and polypyrimidine region, and ΔScore is calculated to represent the impact intensity;

[0034] The population frequency filtering module queries the allele frequency AF of each variant according to the database file, and the expression is AF=(Number of Allele Counts) / (Total Allele Counts in Population)*100%, and a preset AF threshold rule is used for filtering and marking, and a variant list is output to add a population AF field and mark that the preset common threshold is exceeded, wherein the threshold filtering step includes:

[0035] 1) Based on the global AF rule filtering threshold, the variants with AF>0.1% or AF>1% are excluded;

[0036] 2) Based on the specific population AF rule filtering threshold, the patient's ethnic background is filtered;

[0037] 3) For genes related to recessive genetic diseases, the frequency of homozygotes or hemizygotes in healthy populations is calculated.

[0038] Preferably, the driver assessment module queries the database API or local mirror for evidence of oncogenicity, frequency and drug association information, and performs driver gene analysis on the sample using the MutSigCV tool, whose driver gene prediction algorithm includes:

[0039] 1) Construct a background mutation model according to the factors of gene length, sequence context, replication time and expression level, identify genes with significantly higher mutation frequency than background mutation rate in a specific cancer type based on mutation frequency pattern, and calculate the number of mutations to test whether it is higher than expected;

[0040] 2) Identify mutation-enriched genes that are clustered in specific functional domains or have specific functional characteristics based on functional impact pathways;

[0041] 3) Evaluate the oncogenic potential of individual variations according to database evidence and algorithm prediction results;

[0042] The clinical grading module evaluates each variation according to the pre-set rule engine, and outputs a clinical significance grade for each variation, and the grading rules include:

[0043] 1) Tier I level: Variants-drug matching has the highest level of evidence, showing significant clinical benefit in the patient's tumor type;

[0044] 2) Tier II level: Variants-drug matching has strong evidence, but may not be in the same cancer type or the evidence strength is slightly lower than I level;

[0045] 3) Tier III level: There is preclinical or preliminary clinical evidence to support potential benefit, but the clinical evidence is not sufficient;

[0046] 4) Tier IV level: There is only evidence that the variant is a cancer driver, but there is no direct matching effective targeted therapy;

[0047] 5) Tier V level: There is evidence that the variant is not supported as a basis for clinical decision-making;

[0048] The annotation integration module integrates annotation results from different tools and databases to generate a structured variant interpretation report, and processes annotation information that has redundancy or conflict, including:

[0049] 1) For basic annotation conflicts, including different database gene symbols and transcript priorities, the pre-set standard is preferred;

[0050] 2) For functional prediction conflicts, including SIFT prediction of harmful and PolyPhen-2 prediction of benign, prefer to use CADD to set a voting mechanism, combined with specific location information to handle conservatively, marked as Conflicting or VUS;

[0051] 3) For database evidence conflict, according to evidence level priority, timeliness priority and annotation conflict for processing.

[0052] Preferably, the knowledge base integration layer includes a data collection and analysis module, an entity recognition module, a knowledge fusion module, an evidence level quantification module, a knowledge graph construction module and a query interface module. The data collection and analysis module automatically obtains raw data from various heterogeneous knowledge sources. The input sources include structured databases, semi-structured guidelines and unstructured literature. By inputting raw heterogeneous data, the structured information units are parsed;

[0053] The entity recognition module accurately identifies and links biological and medical entities to unique identifiers in authoritative databases in the preliminary structured data based on the input preliminary structured entries, including:

[0054] 1) Named entity recognition: a. Based on dictionary rules, the pre-defined biomedical dictionary is accurately matched, and the regular expression is used to identify and match ambiguous variants;

[0055] b. According to the pre-trained biomedical NER model, the input sentence or text segment is outputted by the model to output the entity label sequence of each word;

[0056] 2) Entity linking: a. According to the mention string, the dictionary / index is queried to generate a list of possible Concept ID candidates; b. The similarity between the mention and each candidate concept is calculated using context information to select the most matching ID;

[0057] The knowledge fusion module inputs the structured knowledge entries, and judges whether different entries describe the same core fact through key field matching and similarity calculation, and fuses different knowledge entries. The evidence level quantification module inputs the fused knowledge records, and according to the comprehensive quantification formula, the quantification rules are designed to give the knowledge records the evidence level score, and the formula is represented as:

[0058]

[0059] Among them, Reg Score represents regulatory approval, Guide Score represents guideline recommendation, Study Score represents research type, Sample Size represents patient sample size, represents statistical significance, Consensus Score represents independent verification, Recency Score represents timeliness, w veg , w gude , wstudy , w sample , w stat , w cons , w time are weights;

[0060] The knowledge graph construction module organizes the fused and graded knowledge into a knowledge graph network and stores it efficiently according to the knowledge records quantified by evidence levels. The query interface module provides interface types including RESTful AP and Graph Query Endpoint interfaces.

[0061] Preferably, the scheme decision engine includes a driver variation screening module, a knowledge base query matching module, a drug reaction prediction module, and a medication recommendation generation module. The driver variation screening module obtains a driver variation list by inputting a structured variation annotation report and a structured clinical archive, specifically including: 1) associating variation data with clinical data, and screening graded variations as candidate driver events based on clinical significance grading; 2) extracting key information of each driver variation, and outputting a driver variation list and associated clinical context;

[0062] The knowledge base query matching module inputs the driver variation list and the clinical context, queries the knowledge graph to seek related targeted drugs and evidence, including:

[0063] 1) querying the relationship edges that meet the conditions in the knowledge graph for each variation in the driver variation list, and querying related evidence based on the relationship edges;

[0064] 2) calling the query interface module of the knowledge base integration layer to input parameters, receiving and parsing the returned JSON data, which contains a drug list, drug effects, and evidence details;

[0065] The drug reaction prediction module predicts the sensitivity or drug resistance probability of a candidate drug according to a pre-trained model, combined with the patient's comprehensive molecular features and clinical features, specifically including:

[0066] 1) inputting a candidate drug list, patient molecular feature vectors, and patient clinical feature vectors;

[0067] 2) training an independent model for each candidate drug, including an input layer, a feature selection layer, a prediction layer, and an output layer, to predict efficacy indicators;

[0068] The drug recommendation generation module recommends a combination of targeted drugs based on pathway analysis, model prediction and knowledge graph reasoning, sets a drug combination generation strategy including overcoming drug-resistant combinations, vertical and horizontal blocking and synthetic lethal combinations, predicts the synergy index and joint efficacy score of any two or more candidate drug combinations according to the combination efficacy prediction model, and the efficacy prediction model input is the patient molecular characteristics and drug pair characteristics, predicts the additive toxicity risk of the combination according to the combination toxicity prediction model, and sets screening evaluation criteria including drug synergy potential, toxicity acceptability, mechanism of action rationality and pharmacodynamic feasibility.

[0069] Preferably, the report visualization module includes a report content assembly module, a key driver variation visualization module, a drug recommendation and evidence chain display module, a clinical regimen detailed description module and an interactive export module. The report content assembly module organizes and arranges all content blocks according to a pre-defined medical report template, including patient information summary, detection overview, key molecular discovery, targeted therapy recommendation, detailed treatment regimen, molecular mechanism illustration and evidence appendix, and the workflow is:

[0070] 1) Create a report object according to the template definition; 2) Extract the corresponding fields from the input data and fill in each part of the template; 3) Call the visualization elements generated by other sub-modules and insert them into the corresponding positions; 4) Call the natural language summary sub-module to generate summary text and perform final format checking;

[0071] The key driver variation visualization module displays the mutation panorama of the patient's tumor through a chart type, including: 1) According to the mutation landscape graph, the variation state of multiple genes in multiple samples is displayed in the form of a heat map;

[0072] 2) Use the variant allele frequency distribution graph to display the proportion of driver variations in different VAF ranges to assist in judging subclone structure;

[0073] 3) According to the key variation function impact score graph, the harmfulness score of the key driver variation in different prediction tools is displayed;

[0074] The drug recommendation and evidence chain display module clearly presents the priority-ordered treatment regimens and the evidence and reasoning chain behind them, and the specific display forms are:

[0075] 1) Highlight the preferred solution through a priority-ordered table;

[0076] 2) The evidence chain Sankey diagram visualizes the connection between the detected driver variation, the affected biological function and pathway, the targeted drug, the expected efficacy and the evidence level flow;

[0077] 3) Evidence Detail Card in the report or interactive interface, by clicking on each recommended solution to expand, which contains the strongest evidence summary, prediction model results, risk details and availability details;

[0078] The clinical regimen specification module converts the optimized treatment regimen into detailed specifications, and its output forms include structured dosing schedules, medication guides and electronic prescription interfaces; the interactive export module provides a user interface for reports, supports interactive data exploration and multiple format exports.

[0079] (III) beneficial effects

[0080] Compared with the prior art, the present application provides an intelligent decision-making system for precise targeted drug use of tumors based on individual mutation information, which has the following beneficial effects:

[0081] 1. The intelligent decision-making system for precise targeted drug use of tumors based on individual mutation information combines rule engines and prediction models to output dynamic priority rankings, realizes multi-model fusion decision-making, predicts primary and secondary drug resistance, calculates liver and kidney function adjusted doses, and generates combination drug timing schemes, thereby realizing deep adaptation to clinical scenarios.

[0082] 2. The intelligent decision-making system for precise targeted drug use of tumors based on individual mutation information realizes combination drug optimization and generates visual clinical reports through individualized dosing regimens, further enhances clinical executability, and realizes end-to-end automated closed-loop operation from data processing to clinical execution. BRIEF DESCRIPTION OF DRAWINGS

[0083] Figure 1 The figure is a schematic diagram of the system modules of the present application. DETAILED DESCRIPTION

[0084] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the embodiments of the present application and the accompanying drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0085] Embodiment one

[0086] In this embodiment, the interface management module receives input data from different sources and in different formats, and performs preliminary classification and routing, and the workflow is as follows:

[0087] 1) Interface definition: provide standardized API interface, file upload interface and database connector;

[0088] 2) Data routing: automatically identify data types according to file types and contents, and distribute to corresponding processing pipelines;

[0089] 3) Metadata extraction: extract key metadata such as sample ID, sequencing platform, sequencing date, batch information, etc. for subsequent quality control and batch correction;

[0090] 4) Data integrity check: check if necessary files are complete, necessary clinical fields exist, output, and sorted raw data files and their associated metadata, ready for corresponding pipelines.

[0091] Evaluate the quality of raw sequencing data through the data quality control module to ensure that the data is suitable for subsequent variant detection. The workflow is as follows:

[0092] 1) Use the tool FastQC to generate QC reports for each FASTQ file, and use the system to parse the QC reports and calculate key indicator values;

[0093] 2) According to the preset QC threshold rules, including:

[0094] a. Pass: all key indicators meet the standards;

[0095] b. Warning: some indicators are close to the threshold or have potential problems;

[0096] c. Fail: key indicators are seriously out of standard, trigger alarm, data requires resequencing;

[0097] Output QC report, QC status, only in Pass or Warning and manual review passed data into the next step.

[0098] Convert data of different sources and formats into system internal standard format through standardized conversion module, the workflow is as follows:

[0099] 1) ASTQ data: use tools to remove sequencing adapter sequences and read end low-quality bases, ensure that all FASTQ files follow the same standard;

[0100] 2) VCF data: use tools to verify if VCF files meet the specifications, ensure that key fields exist and format is correct, detect or require input VCF reference genome version, if inconsistent with system standard, use tools for genome coordinate conversion, remove variants that do not pass quality filter;

[0101] 3) Clinical data: convert CSV / Excel / JSON / HL7, etc. into internal unified structured database table or data model, map source data field name to system defined standardized field name, mark missing values, according to the rules for simple interpolation or keep missing.

[0102] Genomic variations are identified from the quality controlled and normalized FASTQ data by the variant detection module, whose workflow is as follows:

[0103] 1) Alignment: using alignment tools to map the sequencing reads to the reference genome, generating BAM or SAM files;

[0104] 2) Post-alignment processing: sorting BAM files, marking duplicate reads, local realignment, or base quality recalibration;

[0105] 3) Variant calling: a. SNV / Indel: using tools to detect point mutations and small indels; b. CNV: using tools to detect copy number variations based on BAF; c. SV / Fusion: using tools to detect structural variations and gene fusions based on read pairs, split reads, coverage depth, etc.

[0106] The standardized VCF files output by the pre-processing module are input into the variant basic annotation module to add basic biological background information to each variant in the VCF, whose workflow is as follows:

[0107] 1) Gene information annotation: gene symbol, gene ID, gene type;

[0108] 2) Transcript information annotation: transcript ID, transcript biotype, whether it is a canonical transcript;

[0109] 3) Region information annotation: the functional region of the genome where the variant is located, and for splice site variants, the distance to the nearest exon boundary is calculated;

[0110] 4) Nucleotide change annotation: HGVScDNA nomenclature;

[0111] 5) Protein change annotation: HGVS protein nomenclature, predicted amino acid change type;

[0112] 6) dbSNPID annotation: rs number of known SNPs.

[0113] The workflow of the clinical information structuring module is as follows:

[0114] 1) Entity association: different sources of clinical records are integrated together through unique patient id and sample id associations to form a panoramic view of the patient;

[0115] 2) Timeline construction: key clinical events are organized according to the time axis;

[0116] 3) Key feature extraction: a. Calculate ECOG or PS score, extract staging information according to staging criteria; b. Extract important molecular marker historical results; c. Extract comorbidity list and its severity; d. Extract liver and kidney function indicators.

[0117] Example two

[0118] In this embodiment, in the functional impact prediction module, the system configures and runs multiple prediction tools, the tools read the basic annotation results, extract the sequence around the variant site or rely on the pre-calculated genome-wide score file, and then calculate the prediction score and classification of each tool.

[0119] The population frequency filtering module inputs the variant list containing basic annotation and functional prediction results, and filters out common variants using large-scale population databases, including: 1) gnomAD contains tens of thousands to hundreds of thousands of healthy individuals or unselected population genomic data; 2) 1000Genomes Project population variant database; 3) ESP large-scale exome database; 4) dbSNP database.

[0120] The driver assessment module assesses the role of the variant in cancer development according to the inputted population frequency filtered variant list, and the database query method includes:

[0121] 1) Cancer somatic mutation database: COSMIC somatic mutation database contains hot spot mutations, functional annotation, frequency information;

[0122] cBioPortal: Integrates TCGA, ICGC and other large cancer genome projects visualization analysis platform, provides frequency and clinical association of variants in specific cancer;

[0123] 2) Cancer knowledge base: OncoKB provides mutation-drug matching and evidence level supported by FDA / NCCN guidelines; CIViC collects literature and expert interpretation of clinical evidence; CGI provides driver gene, biomarker, targeted drug information.

[0124] Example three

[0125] In this embodiment, the specific workflow of the data collection and analysis module includes:

[0126] 1) API call: For sources that provide API, write scripts to call API regularly, pass parameters as needed to get JSON / XML format response;

[0127] 2) Web crawler: For websites without API or API restricted, use frameworks to simulate browser access to pages, parse HTML / DOM structure, locate and extract target data, handle pagination, login, anti-crawl mechanism;

[0128] 3) File download and parsing: download PDF guidelines, Excel / CSV data files, and send them to the subsequent parsing module;

[0129] 4) For JSON / XM format data, use the corresponding parsing library to extract fields; for HTML format data, use the HTML parser to extract target text and attributes; for PDF format data, use OCR + text extraction library to extract text and identify tables; for CSV / Excel format data, use data processing library to read table data; for natural language text format data, perform basic cleaning;

[0130] 5) Map genes to standard symbols, map drugs to standard generic names, standardize description strings to system internal variant representation, map tumor types to standard codes, extract key information from evidence descriptions, including drug effects, evidence types, efficacy indicators, and statistical significance, extract evidence publication dates, guideline version dates, and drug approval dates, and standardize them to date format.

[0131] The knowledge graph construction module stores the knowledge graph in a graph database, and the specific process includes:

[0132] 1) According to the data model, convert the fused and quantified knowledge records into nodes and edges; 2) Import the nodes and edges into the selected graph database, and establish indexes to speed up queries.

[0133] Example Four

[0134] In this embodiment, the knowledge base query matching module generates a preliminary candidate drug list, where each entry contains the standard name of the drug, the main target point, the associated driver variation, the predicted effect, and the evidence list. The knowledge base query matching module also includes a rule engine-based preliminary screening and framework construction module and an evidence-based priority sorting module.

[0135] The rule engine-based preliminary screening and framework construction module inputs the preliminary candidate drug list and constructs a rule library, which includes rules such as excluding drug contraindications, avoiding repeated and ineffective treatments, mandatory priority, handling black box warnings, and coexistence of drug resistance markers. The rule engine traverses the rule library and applies each rule to the preliminary candidate drug list.

[0136] The evidence-based priority sorting module sorts the drugs in the candidate drug list that have passed the rule screening and enhancement based on the evidence strength returned by the knowledge base and the matching degree in the patient's clinical context. Specifically, it includes:

[0137] 1) Calculate the score of each candidate drug with all associated evidence quality and matching degree, the formula is:

[0138]

[0139] where e is each piece of evidence record traversing the drug association, Evidence Weight Evidence quality score, EPS j The maximum value of the product of the evidence matching degree score and the evidence quality score of all associated evidence of the drug is taken to emphasize the strongest evidence; the weighted average of the evidence matching degree score and the evidence quality score is taken to consider the overall evidence strength; and the sum of the evidence matching degree score and the evidence quality score is taken to consider the number of evidences.

[0140] The drug reaction prediction module further comprises a drug resistance prediction module and a drug action toxicity prediction module. The drug resistance prediction module analyzes the primary drug resistance signal in the current mutation profile of the patient, predicts the possible secondary drug resistance mechanism after using the candidate drug, and proposes avoidance strategies in the recommendation. The prediction method comprises:

[0141] 1) Check whether there is a known mutation that is primarily resistant to the candidate drug to detect coexisting drug resistance mutations;

[0142] 2) Use a pre-trained model to predict the primary drug resistance probability of the patient to the candidate drug;

[0143] 3) Query the known drug resistance mechanism in the knowledge base, use the training set containing the pre-treatment molecular characteristics and post-treatment drug resistance mechanism data to train the model to predict the most likely drug resistance mechanism, and predict the secondary drug resistance;

[0144] 4) For the detected primary drug resistance, strongly warn and suggest avoiding using in the recommendation of the corresponding candidate drug, and for the predicted secondary drug resistance, suggest monitoring in the recommendation of the candidate drug;

[0145] In the drug action toxicity prediction module, the drug resistance information, liver and kidney function, combined drug use, and drug metabolism enzyme genotype are input to evaluate the potential interaction and toxicity risk between candidate drugs and the patient's own condition, specifically including:

[0146] 1) Drug-drug interaction check: query the DDI knowledge base to check whether there is an interaction between any two drugs in the list and obtain the interaction effect;

[0147] 2) Genotype-based toxicity and metabolism prediction: based on pharmacogenomics, if the patient has a slow metabolism type PM, query the database and predict the change in drug metabolism rate, and then predict the increased toxicity risk or reduced efficacy;

[0148] 3) Liver and kidney function adjustment: based on the pharmacokinetic model or the dose adjustment rule, according to the creatinine clearance rate index, calculate the dosage adjustment of the drug under the current state of the patient;

[0149] 4) Integration of DDI severity, PGx risk, liver and kidney function adjustment requirement, drug intrinsic toxicity profile, patient general condition to overall toxicity risk score, which is formulated as:

[0150]

[0151] where Severity DDIj represents the most severe DDI grade involved in the drug, PGx Riskj represents the risk grade predicted based on genotype, represents the dose adjustment degree, Baseline Toxj represents the drug intrinsic toxicity grade, w ddi , w pgx , w organ , w drug are all weight coefficients.

[0152] The drug recommendation generation module outputs a drug combination list, each of which contains a drug list in the combination, combination principle, predicted synergy degree, predicted combination toxicity risk, evidence level, dosing sequence, dose adjustment suggestion and monitoring requirement.

[0153] Example Five

[0154] In this embodiment, the template content of the report content assembly module includes:

[0155] 1) Patient information summary: patient ID, age, gender, sample ID, tumor type, current line, key clinical status-PS score and liver and kidney function;

[0156] 2) Test overview: test platform, test gene panel, quality control index and overall TMB / MSI status;

[0157] 3) Key molecular findings: driver variant summary table including gene, variant type, clinical significance classification-Tier and COSMIC hotspot and allele frequency, other potential related variants and drug resistance related variants;

[0158] 4) Targeted therapy recommendations: prioritized recommended regimen list, detailed reasons for each regimen, expected efficacy and risk overview;

[0159] 5) Detailed treatment regimen: specific dosing regimen of the first choice regimen including drug, dose, route, cycle and schedule, alternative regimen and supportive treatment suggestion;

[0160] 6) Molecular mechanism illustration: signal pathway diagram affected by key driver variants;

[0161] 7) Evidence appendix: key literature supporting recommendations, guideline abstract, clinical trial information.

[0162] The interaction export module includes the following export functions:

[0163] 1) PDF report: static, printable, format compliant with medical document archiving requirements, using libraries such as wkhtmltopdf, WeasyPrint, PDFKit to convert HTML reports to PDF;

[0164] 2) Structured data export: JSON and XML format reports containing all raw data and analysis results;

[0165] 3) Image export: PNG and SVG format reports for individual saving of Oncoprint, pathway maps, etc.

[0166] In summary, the tumor precision targeted drug intelligent decision system based on individual mutation information combines rule engines and prediction models to output dynamic priority rankings, implement multi-model fusion decision-making, predict primary and secondary drug resistance, calculate liver and kidney function to adjust dosages, and generate combination drug timing plans, achieve deep adaptation to clinical scenarios, optimize combination drug use through individualized dosing plans, and generate visual clinical reports to further enhance clinical executability, and achieve end-to-end automated closed-loop operations from data processing to clinical execution.

[0167] The related modules involved in the system are hardware system modules or functional modules combining existing computer software programs or protocols with hardware. The computer software programs or protocols involved in the functional modules are known to those skilled in the art and are not improvements of the system. The improvement of the system is the interaction or connection relationship between the modules, i.e., the improvement of the overall structure of the system to solve the corresponding technical problems of the system.

[0168] Although embodiments of the present application have been shown and described, it is to be understood that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent decision system for precise tumor targeting drug use based on individual mutation information, characterized in that, The data processing layer includes unified multi-source heterogeneous data format, quality control and standardization processing, and provides high-quality and structured patient molecular and clinical data for subsequent analysis; The variant annotation and function prediction layer analyzes the biological function of gene variants, combines population frequency and database evidence for clinical classification, and screens out key driver variants; The knowledge base integration layer dynamically aggregates drug-target-variant associations, clinical guidelines, and evidence-based research, and constructs a structured knowledge graph that can be queried in real time and evidence-based; The scheme decision engine integrates patient variant spectrum, clinical characteristics, and knowledge base, generates priority-ranked drug recommendations through rule engine and model, predicts drug resistance risk and toxicity, and recommends specific dosing regimens; The report visualization module converts complex analysis into intuitive clinical decision reports through visualization tools, supporting physician-patient communication and implementation. 2.The intelligent decision-making system for precise tumor targeting drug use based on individual mutation information according to claim 1, characterized in that, The data processing layer includes an interface management module, a data quality control module, a standardization conversion module, a variant detection module, a variant basic annotation module, and a clinical information structured module. The interface management module accesses data input sources, including: 1) raw files output by sequencers; 2) variant calling files provided by external laboratories; 3) clinical data derived from hospital information systems and electronic medical records; 4) other omics data files; The data quality control module calculates and evaluates data from the interface management module, including: 1) The average Q value of each base position is calculated by FastQC, expressed as where Q i is the Phred score of the i-th base, N is the number of bases, and the Phred quality score is expressed as Q = -10 · log 10 (P error ), P error is the probability of base recognition error, to measure the probability of base recognition correctness; 2) Check the distribution of A, T, C, and G in each position and calculate the percentage of each nucleotide in each position; 3) Calculate the proportion of completely identical sequence reads, represented as Duplication Rate = (Number of duplicate reads) / (Total number of reads), where high duplication rates may be due to PCR amplification bias or low complexity regions; 4) Scan the read with FastQC to match known adapter sequences and detect contamination where sequencing adapter sequences appear in the read; 5) Calculate the GC base percentage of the entire sample read, represented as GC% = (Count(G) + Count(C)) / (Count(A) + Count(T) + Count(G) + Count(C))*100%; 6) Calculate the ratio of total read count to data volume to ensure that the pre-set coverage depth requirement is met; The standardization conversion module converts different formats of data into internal standard format, including: 1) For FASTQ data, trim sequencing adapter sequences and low-quality base fragments; 2) For VCF data, verify VCF file format, detect input VCF reference genome version, and ensure uniform standard for genome coordinate conversion; 3) For clinical data, standardize data through structure conversion, field mapping, and value standardization; The variation detection module adopts alignment marking method and variation calling technology to identify genomic variations according to the quality control and standardized FASTQ data; the variation basic annotation module performs basic annotation on the variations in the VCF file to ensure that the key information uses a unified standard; and the clinical information structured module further structures, entity associates and quality improves the standardized mapped clinical data. 3.The intelligent decision-making system for precise tumor targeting medication based on individual mutation information according to claim 2, characterized in that, The variation annotation and function prediction layer includes a function impact prediction module, a population frequency filtering module, a driving evaluation module, a clinical classification module and an annotation integration module. The function impact prediction module predicts the potential impact of the variation on the gene function according to the output of the variation basic annotation module. The specific prediction method includes: 1) SIFT method is used to compare homologous protein sequences, calculate the degree of tolerance of specific amino acid positions, and output the actual score SIFT score SIFT pred , when SIFT score <0.05, it is determined that the variation is harmful to the gene function; GERP++ was used to measure the strength of selection constraint on genomic sites during evolution, and the output is the rejection substitution score GERP++ RS The higher the positive value, the more harmful the variation to gene function; 2) PolyPhen-2 method combined with sequence conservation, amino acid physicochemical property differences and protein structure information to predict missense variant effects, output actual score PolyPhen-2 score vs. predicted value PolyPhen-2 pred ; 3) CADD method is used to integrate multiple annotation information, and SVM is used to train to distinguish simulated neutral variation and simulated pathogenic variation, and output scaling score CADD phred The higher the score is, the more harmful the variation is to the gene function; The output results of the prediction tools are integrated by the REVEL method, and a random forest model is trained to optimize the pathogenicity prediction of rare missense variants, and an actual score REVEL is output phred The higher the score is, the stronger the predicted pathogenicity is; 4) Based on the Inception architecture, a deep neural network is used to predict the impact of the variation on the splice donor, acceptor, branch point and polypyrimidine region, and a ΔScore is calculated to represent the impact strength; The population frequency filtering module queries the allele frequency AF of each variation according to the database file, the expression of which is AF=(Number of Allele Counts) / (Total Allele Counts in Population)*100%, and adopts a preset AF threshold rule for filtering marking. The output variation list adds a population AF field and marks the variations that exceed the preset common threshold. The threshold filtering step includes: 1) Based on the global AF rule filtering threshold, the variations with AF>0.1% or AF>1% are excluded; 2) Based on the specific population AF rule filtering threshold, the patient's ethnic background is filtered; 3) For genes related to recessive genetic diseases, the frequency of homozygotes or hemizygotes in healthy populations is calculated. 4.The intelligent decision-making system for precise tumor targeting medication based on individual mutation information according to claim 3, characterized in that, The driving evaluation module queries the database API or local mirror using the variation identifier to obtain oncogenic evidence, frequency and drug association information, and uses the MutSigCV tool to analyze the driving genes of the sample. The driving gene prediction algorithm includes: 1) According to the factors of gene length, sequence background, replication time and expression level, a background mutation model is constructed to identify genes with mutation frequency significantly higher than the background mutation rate in a specific cancer type based on mutation frequency patterns, and to calculate the mutation number test whether it is higher than expected; 2) Based on the function impact pathway, identify mutation enrichment genes that are aggregated in specific functional domains or have specific functional characteristics; 3) According to the database evidence and algorithm prediction results, the oncogenic potential of individual variations is evaluated; The clinical classification module evaluates each variation according to a preset rule engine, and outputs a clinical significance classification for each variation. The classification rules include: 1) Tier I level: variation-drug matching has the highest level of evidence, showing significant clinical benefit in the patient's tumor type; 2) Tier II level: variation-drug matching has strong evidence, but may not be in the same cancer type or the evidence strength is slightly lower than I level; 3) Tier III level: there is preclinical or preliminary clinical evidence to support potential benefit, but the clinical evidence is not sufficient. 4) Tier IV: only evidence that the variant is a driver of cancer exists, but no direct match to an effective targeted therapy; 5) Tier V: evidence exists that the variant is not a driver for clinical decision making; The annotation integration module integrates the annotation results from different tools and databases, generates a structured variant interpretation report, and processes the annotation information with redundancy or conflict, including: 1) For basic annotation conflicts, including different database gene symbols and transcript priorities, the preset standard is used for processing; 2) For functional prediction conflicts, including SIFT prediction of harmful and PolyPhen-2 prediction of benign, a CADD setting voting mechanism is used, combined with specific location information conservation processing, marked as Conflicting or VUS; 3) For database evidence conflicts, according to the priority of evidence level, timeliness and conflict marking are processed.

5. The intelligent decision system for precision tumor targeted drug use based on individual mutation information according to claim 1, characterized in that, The knowledge base integration layer includes a data collection and analysis module, an entity recognition module, a knowledge fusion module, an evidence level quantification module, a knowledge graph construction module, and a query interface module. The data collection and analysis module automatically obtains raw data from various heterogeneous knowledge sources, and its input sources include structured databases, semi-structured guidelines, and unstructured literature. By inputting raw heterogeneous data, the structured information unit is parsed; The entity recognition module accurately identifies and links biological and medical entities to unique identifiers in authoritative databases based on the input preliminary structured entries in the preliminary parsed structured data, including: 1) Named entity recognition: a. Based on dictionary rules, the pre-defined biomedical dictionary is accurately matched, and the regular expression is used to identify and match ambiguous variants; b. According to the pre-trained biomedical NER model, the entity label sequence of each word is outputted by inputting the sentence or text fragment; 2) Entity linking: a. According to the mention string, the dictionary / index is queried to generate a list of possible Concept ID candidates; b. The similarity between the mention and each candidate concept is calculated using context information to select the most matching ID; The knowledge fusion module inputs structured knowledge entries, and determines whether different entries describe the same core fact through key field matching and similarity calculation, and fuses different knowledge entries; The evidence level quantification module inputs the fused knowledge records, and according to the comprehensive quantification formula, the quantification rules are designed to assign evidence level scores to the knowledge records, and the formula is represented as: The knowledge graph construction module organizes the fused and graded knowledge into a knowledge graph network and efficiently stores it according to the knowledge records quantified by the evidence level, and the query interface module provides interface types including RESTful AP and GraphQuery Endpoint interfaces. where Reg Score represents regulatory approval, Guide Score represents guideline recommendation, Study Score represents study type, Sample Size represents patient sample size, represents statistical significance, Consensus Score represents independent validation, Recency Score represents recency, w veg , w gude , w study , w sample , w stat , w cons , w time are all weights; ​ 6.The intelligent decision-making system for precise tumor targeting medication based on individual mutation information according to claim 5, characterized in that, The scheme decision engine includes a driving variation screening module, a knowledge base query matching module, a drug reaction prediction module and a drug recommendation generation module, the driving variation screening module obtains a driving variation list by inputting a structured variation annotation report and a structured clinical archive, specifically including: 1) associate variation data with clinical data, and screen out hierarchical variations as candidate driving events based on clinical significance classification; 2) extract the key information of each driving variation, and output the driving variation list and the associated clinical context; The knowledge base query matching module inputs the driving variation list and the clinical context, queries the knowledge graph to seek related targeted drugs and evidence, including: 1) query the relationship edge that meets the condition in the knowledge graph for each variation in the driving variation list, and query the related evidence based on the relationship edge; 2) call the query interface module of the knowledge base integration layer to input parameters, receive and parse the returned JSON data, which includes a drug list, drug effects and evidence details; The drug reaction prediction module predicts the sensitivity or drug resistance probability of the candidate drug according to the pre-trained model combined with the comprehensive molecular characteristics and clinical characteristics of the patient, specifically including: 1) input the candidate drug list, patient molecular feature vector and patient clinical feature vector; 2) train an independent model for each candidate drug, including an input layer, a feature selection layer, a prediction layer and an output layer, to predict the efficacy index; The drug recommendation generation module recommends a combination of targeted drugs based on pathway analysis, model prediction and knowledge graph reasoning, sets the drug combination generation strategy including overcoming drug resistance combination, vertical and horizontal blocking and synthetic lethality combination, predicts the synergy index and joint efficacy score of any two or more candidate drug combinations according to the combination efficacy prediction model, and the input of the combination efficacy prediction model is patient molecular characteristics and drug pair characteristics, predicts the additive toxicity risk of the combination according to the combination toxicity prediction model, and sets the screening evaluation criteria including drug synergy potential, toxicity acceptability, mechanism of action rationality and pharmacodynamic feasibility. 7.The intelligent decision-making system for precise tumor targeting medication based on individual mutation information according to claim 1, characterized in that, The report visualization module includes a report content assembly module, a key driving variation visualization module, a drug recommendation and evidence chain display module, a clinical scheme detailed description module and an interactive export module, the report content assembly module organizes and arranges all content blocks according to the pre-defined medical report template, including patient information summary, detection overview, key molecular discovery, targeted therapy recommendation, detailed treatment scheme, molecular mechanism illustration and evidence appendix, and the workflow is: 1) create a report object according to the template definition; 2) extract the corresponding fields from the input data and fill in the template parts; 3) call the visualization elements generated by other submodules and insert them into the corresponding positions; 4) call the natural language summary submodule to generate summary text and perform final format checking; The key driving variation visualization module displays the mutation panorama of the patient's tumor through chart types, including: 1) display the variation state of multiple genes in multiple samples in the form of a heat map according to the mutation landscape graph; 2) use the variation allele frequency distribution graph to display the proportion of driving variations in different VAF ranges to assist in judging subclonal structure; 3) For key driver variants, show their deleteriousness scores in different prediction tools according to the key variant functional impact score map; The drug recommendation and evidence chain display module clearly presents the prioritized treatment options and the evidence and reasoning chain behind them, in the form of: 1) Highlight the first choice by a prioritized table; 2) Visualize the flow from the detected driver variants to the impacted biological functions and pathways to the targeted drugs to the expected efficacy and evidence level by an evidence chain Sankey diagram; 3) Evidence details cards are unfolded by clicking on each recommended option in the report or interactive interface, which contains the strongest evidence summary, prediction model results, risk details, and availability details; The clinical regimen specification module converts the optimized treatment options into detailed specifications, which are output in the form of structured dosing schedules, medication guides, and electronic prescription interfaces; the interactive export module provides a user interface for reports, supports interactive data exploration, and supports multiple formats for export.

Images