RNA pesticide research and development evaluation method and system based on multi-expert agent cooperation

Abstract

The application discloses a kind of RNA pesticide research and development evaluation method and system based on multi-expert agent cooperation, and relates to RNA pesticide intelligent research and development technical field.The method includes the following steps: transcriptome mutual comparison is carried out to target pest and co-target pest, and shared similar genes are extracted according to preset screening conditions;Expert agent is used to independently evaluate the value of candidate genes according to structured evidence planning task;According to the ranking and level of candidate genes, the ranking and level of candidate genes are obtained according to the comprehensive recommendation score of candidate genes;According to the ranking and level of candidate genes, candidate double-stranded RNA fragments are generated and the function pre-score is calculated, the candidate double-stranded RNA fragments are screened according to the function pre-score, and the target hit and off-target risk evaluation of the candidate fragments reserved by preliminary screening are carried out;Structural research report is generated by comprehensively evaluating process information.The method of the application can efficiently output the RNA pesticide research and development scheme which can be explained, traced back and experimentally verified.

Description

Technical Field

[0001] This invention relates to the field of intelligent R&D technology for RNA pesticides, and in particular to an RNA pesticide R&D evaluation method and system based on multi-expert agent collaboration. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] RNA pesticide development is a systematic project encompassing "target identification, mechanism determination, sequence design, safety assessment, and product commercialization." For target pests, the number of candidate genes is large, and evidence from cross-species RNAi, literature, functional annotation, paralogous relationships, and non-target risks are scattered across different data sources. Relying solely on manual screening or a single dsRNA design tool makes it difficult to simultaneously consider target reliability, silencing effect, safety, and commercialization feasibility.

[0004] In the current RNA pesticide R&D process, the various stages from target discovery to product transformation are severely fragmented. There is a lack of an intelligent collaborative system that can integrate multi-source heterogeneous data and run through target evaluation, sequence design and transformation constraints. As a result, R&D relies heavily on human experience and fragmented tools, making it difficult to systematically balance lethality, functional redundancy avoidance, sequence safety and formulation feasibility. Specifically, existing dsRNA design tools typically use known single-gene sequences as input, failing to automatically construct candidate gene pools from multi-species transcriptomes, cross-species RNAi evidence, functional annotations, and paralogous relationships. Furthermore, functional annotations and RNAi experimental evidence are scattered and conflicting, lacking a unified weighting mechanism, resulting in low efficiency for manual screening. Simultaneously, the functional compensation risks arising from paralogous and alternative pathways are rarely considered as independent evaluation dimensions, and sequence design lacks quantitative trade-offs for multiple objectives such as fragment length, GC content, siRNA activity, thermodynamic asymmetry, key site preference, on-target coverage, and off-target risk, often relying on single scores or empirical decisions. More critically, transformation constraints such as formulation delivery and ecological exposure are often considered only after candidate fragments are determined, leading to frequent obstacles in later implementation. Moreover, even with the introduction of general intelligent agents, the lack of clear expert division of labor, tool usage constraints, and evidence auditing mechanisms for RNA pesticides makes the evaluation process difficult to trace, results in poor consistency of conclusions, and hinders the formation of a closed-loop, auditable decision-making chain from target discovery to product transformation. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for evaluating RNA pesticide development based on multi-expert agent collaboration. This system enables the automatic completion of candidate gene construction, target value assessment, dsRNA fragment design, safety and transformation feasibility assessment, and output of interpretable, traceable, and experimentally verifiable RNA pesticide development solutions, under the constraints of multi-source transcriptome resources, functional annotation, literature evidence, homology comparison, and non-target transcriptome.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: The first aspect of this invention provides a method for evaluating the research and development of RNA pesticides based on multi-expert agent collaboration, comprising the following steps: Transcriptome sequence resources and associated annotation resources of target pests, co-target pests and non-target species are obtained, and the transcriptome resources are standardized to form species-level transcriptome sequence files and annotation tables in a unified format. Transcriptome cross-comparison is performed between target pests and co-target pests, and shared similar genes are extracted according to preset screening conditions to obtain candidate genes. Design multiple expert agents, use the expert agents to independently evaluate the value of candidate genes based on structured evidence planning tasks, and perform consistency checks on the scores obtained by the expert agents. The candidate genes are ranked and ranked based on the scores of those that pass the consistency check. Candidate double-stranded RNA fragments are generated based on the sequencing and ranking of candidate genes, and functional pre-scores are calculated. Candidate double-stranded RNA fragments are screened based on functional pre-scores to obtain candidate fragments that are initially retained. The target hit and off-target risks of the candidate fragments that are initially retained are assessed to obtain dual-mode fragment recommendation results. A structured R&D report is generated by comprehensively evaluating information from the process.

[0007] Furthermore, the specific steps for transcriptome comparison between the target pest and co-target pests, and extraction of shared similar genes according to preset screening conditions, are as follows: Construct a common target alignment index; Based on the co-target alignment index, a forward alignment is performed from the target pest to the co-target pest to obtain a candidate hit set; Perform a validity determination of the candidate hits based on the candidate hit set; Calculate the support score for a single species; Calculate cross-species co-target support; The reasons for entering the candidate pool are generated, thus forming a candidate gene pool.

[0008] Furthermore, the expert agents include target discovery expert agents, target quality expert agents, and functional redundancy expert agents. The target discovery expert agent is used for lethal potential assessment, the target quality expert agent is used for function and pathway coreness assessment, and the functional redundancy expert agent is used for single target effectiveness risk assessment.

[0009] Furthermore, the specific steps for generating candidate double-stranded RNA fragments and calculating functional pre-scores based on the sequencing and ranking of candidate genes are as follows: Extract the target gene coding sequence based on the annotation and generate candidate double-stranded RNA fragments; For each candidate double-stranded RNA fragment, a virtual enzyme cleavage simulation was performed to obtain a set of small interfering RNAs; A multi-criteria hierarchical scoring algorithm for small interfering RNAs was used to pre-score the function of small interfering RNA sets.

[0010] Furthermore, the small interfering RNA multi-criteria stratified scoring algorithm uses the positive chain window as the scoring object, first performs hard constraint elimination, and then performs functional pre-scoring calculation according to the stratification of basic score addition, dynamic score addition, and site deduction.

[0011] Furthermore, the specific steps for assessing the target hit and off-target risks of the candidate fragments retained in the initial screening are as follows: Candidate double-stranded RNA fragments selected and retained after functional pre-scoring and their derived small interfering RNA sequences generated by virtual Dicer processing simulation were compared with the transcriptomes of target pests, co-target pests, and non-target species to obtain control-related coverage evidence and safety-related risk evidence for each candidate fragment. For each candidate fragment, the prevention and control-related coverage evidence and the safety-related risk evidence are used to generate recommendation results based on a dual-mode recommendation objective function.

[0012] A second aspect of this invention provides an RNA pesticide research and evaluation system based on multi-expert agent collaboration, comprising: The data acquisition module is used to acquire transcriptome sequence resources and associated annotation resources of target pests, co-target pests and non-target species, and to standardize the transcriptome resources to form species-level transcriptome sequence files and annotation tables in a unified format. The module performs transcriptome cross-comparison between target pests and co-target pests, and extracts shared similar genes according to preset screening conditions to obtain candidate genes. The agent evaluation module is used to design multiple expert agents, utilize the expert agents to independently evaluate the value of candidate genes based on structured evidence planning tasks, and perform consistency checks on the scores obtained by the expert agents. The candidate gene screening module is used to calculate the comprehensive recommendation score of candidate genes based on the scores of passing the consistency check, and to obtain the ranking and grade of candidate genes based on the comprehensive recommendation score of candidate genes. The fragment recommendation module is used to generate candidate double-stranded RNA fragments based on the order and grade of candidate genes and calculate the functional pre-score. The candidate double-stranded RNA fragments are screened based on the functional pre-score to obtain the candidate fragments retained in the initial screening. The target hit and off-target risk of the candidate fragments retained in the initial screening are assessed to obtain the dual-mode fragment recommendation results. The report generation module is used to comprehensively evaluate process information and generate structured R&D reports.

[0013] A third aspect of the present invention provides a computer-readable storage medium storing a computer program adapted for loading by a processor and executing the steps of the RNA pesticide research and evaluation method based on multi-expert agent collaboration as described in the first aspect of the present invention.

[0014] A fourth aspect of the present invention provides a computer device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the RNA pesticide research and evaluation method based on multi-expert agent collaboration as described in the first aspect of the present invention.

[0015] A fifth aspect of the present invention provides a computer program product or computer program comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps of the RNA pesticide research and evaluation method based on multi-expert agent collaboration as described in the first aspect of the present invention.

[0016] The above one or more technical solutions have the following beneficial effects: This invention proposes a method and system for evaluating RNA pesticide development based on multi-expert intelligent agent collaboration. It enables the automatic completion of candidate gene construction, target value assessment, dsRNA fragment design, and safety and transformation feasibility evaluation under the constraints of multi-source transcriptome resources, functional annotation, literature evidence, homology alignment, and non-target transcriptome. It outputs interpretable, traceable, and experimentally verifiable RNA pesticide development plans. This invention organizes information previously scattered across literature, transcriptome, functional annotation, BLAST alignment, dsRNA design rules, and application validation experience into a unified intelligent decision-making process. Compared to manual screening or single-point dsRNA design tools, this invention improves candidate gene screening efficiency, enhances target scoring interpretability, reduces functional redundancy and off-target risks, identifies transformation barriers early, and provides traceable candidate solutions for subsequent experimental validation.

[0017] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the RNA pesticide research and evaluation method based on multi-expert agent collaboration in Embodiment 1 of the present invention. Detailed Implementation

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0021] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] Example 1: This invention provides a method for evaluating RNA pesticide development based on multi-expert agent collaboration. It uses a large language model or other agent model capable of task planning as the task center, and insect transcriptome, functional annotation, literature databases, Basic Local Alignment Search Tool (BLAST), Gene Ontology (GO), siRNA rule model, and transcriptome alignment tool as external capability modules. A closed loop is formed through a Query Planner, multiple expert agents, and Critic evidence auditing. This embodiment does not simply generate text suggestions, but rather organizes the RNA pesticide development process into a plannable, callable, auditable, and iterative intelligent design process.

[0023] like Figure 1 As shown, it includes the following steps: S1: Obtain transcriptome sequence resources and associated annotation resources of target pests, co-target pests and non-target species, and standardize the transcriptome resources to form species-level transcriptome sequence files (FASTA format, FAST-All, rapid sequence format) and annotation tables in a unified format. Compare the transcriptomes of target pests and co-target pests, and extract shared similar genes according to preset screening conditions to obtain candidate genes.

[0024] In one specific implementation, this embodiment starts from multi-source insect transcriptome resources to uniformly obtain species-level transcriptome FASTA, CDS coordinates, and transcript annotations. For insect species with RefSeq or reference genome annotations, transcriptome resources can be obtained from NCBI Datasets; for species with insufficient NCBI transcriptome resources, genome, GFF3, and protein annotations can be obtained from insect-specific databases such as InsectBase, and mRNA transcriptomes can be extracted. By filtering low-information annotations and comparing transcriptomes, a candidate gene pool that can enter the intelligent evaluation stage is constructed.

[0025] S1.1: Obtain transcriptome sequence resources and associated annotation resources for target pests, co-target pests and non-target species.

[0026] Specifically, the system receives user input regarding target pests, co-target pests, and non-target species. Target pests are the primary targets for RNA pesticide control; co-target pests are other pests for which the same dsRNA fragment is desired to have control potential; and non-target species are those used for ecological safety assessment and may include beneficial insects, pollinating insects, natural enemy insects, aquatic organisms, or other species of concern under regulatory requirements.

[0027] The data acquired in this embodiment includes transcriptome sequence resources and their associated annotation resources: transcriptome sequence files, genome sequences, Universal Feature Format (UPF) version 3 annotation files, protein annotations, coding sequence coordinates, species names, assembly versions, database numbers, and download records. The transcriptome sequence files can be in FASTA format; the UFF version 3 annotation files can be in GFF3 format.

[0028] The data sources for this embodiment are existing publicly available databases. For species with reference sequence databases or reference genome annotations, transcript sequences can be obtained from the National Center for Biotechnology Information (NCBI) dataset service. For species with insufficient public transcripts but whose annotations are included in insect-specific databases, genome, GFF3 annotations, and protein annotations can be obtained from publicly available specialized databases such as the Insect Genome Database, and messenger RNA transcriptomes and CDS coordinates can be extracted accordingly.

[0029] The purpose of this step is to establish a unified data foundation for subsequent target screening, fragment design, and ecological safety assessment. It enables the complementarity of legitimate data from multiple sources, increases species coverage, and provides symmetrical and traceable input for candidate gene evaluation and off-target assessment.

[0030] S1.2: Standardize transcriptome resources to form species-level transcriptome sequence files and annotation tables in a unified format.

[0031] Specifically, transcriptome, GFF3, protein annotation, and CDS information from different sources are standardized to form species-level transcriptome sequence files and annotation tables in a unified format.

[0032] More specifically, it includes the following steps: (1) Uniform sequence key: Use the stable transcript number, accession number or custom number in the FASTA header line as the subsequent matching key to establish the mapping relationship between transcript number, gene number and protein number.

[0033] (2) Unify sequence characters: unify sequence characters to uppercase, process soft mask characters, gap characters and uncertain bases, and delete empty sequences, completely repeated sequences and obviously low-quality sequences.

[0034] (3) Unified annotation fields: An annotation table with consistent fields is generated from GFF3 annotation, protein annotation or database annotation, including at least species name, transcript number, gene number, functional description, CDS start and end positions and annotation source.

[0035] (4) Generate CDS coordinate table: For transcripts derived from genome and annotation files, generate CDS coordinate table simultaneously; for data sources that already contain CDS, verify whether the CDS coordinates match the transcript length.

[0036] (5) Generate processing logs: record data source, version, download time, filtering rules and standardized parameters.

[0037] This step outputs species-level transcriptome sequence files, transcript annotation tables, CDS coordinate tables, and processing logs. Its purpose is to eliminate naming and format differences between different databases, ensuring that subsequent BLAST alignment, GO parsing, RAG retrieval, dsRNA fragment searching, and non-target alignment are all performed based on the same set of standardized sequences and annotation fields.

[0038] S1.3: Compare the transcriptomes of the target pest and the co-target pest, and extract shared similar genes according to the preset screening conditions to obtain candidate genes.

[0039] Specifically, transcriptomes of the target pest and co-target pests are compared, and shared similar genes are extracted according to preset screening criteria. Based on the obtained standardized transcriptomes of the target pest and co-target pests, and annotation tables, a cross-species co-target candidate pool construction algorithm is executed. This embodiment incorporates alignment hit rate, coverage, cross-species support count, hit region length, and subsequent dsRNA design requirements into the candidate gene construction rules, thereby forming a traceable, scoreable, and subsequent fragment design-ready candidate gene pool. For example, a hit length greater than 500 bp, an identity greater than 80%, and an E-value less than or equal to 1 × 10⁻⁶ can be used. -5 The rules are used as the initial threshold to form a candidate gene pool.

[0040] Specifically, the following steps are included: S1.3.1: Construct a common target alignment index.

[0041] A comparison database was constructed using standardized transcriptomes from one or more co-target pests, and the species name, database version, transcript number, and annotation source were recorded. If multiple co-target pests exist, species-level comparison indexes were established for each pest, and a cross-species hit summary table was created to provide underlying data support for subsequent cross-species data fusion.

[0042] S1.3.2: Perform a forward comparison from the target pest to the co-target pest based on the co-target comparison index to obtain a candidate hit set.

[0043] Using the transcriptome sequence of the target pest as the query sequence, nucleic acid sequence alignment is performed on the transcriptomes of each co-target pest to obtain a candidate hit set. For each hit... The system extracts the expected value. Sequence consistency Hit length Query coverage Target coverage Multidimensional feature vectors, including alignment score, hit species, and hit transcript number.

[0044] S1.3.3: Perform a validity determination of the candidate hits based on the candidate hit set.

[0045] A hit is determined by a combination of significance, consistency, length, and coverage. A valid hit can be determined using equation (1): (1).

[0046] in, Indicates the first The comparison hit; Indicates the expected value; Indicates sequence consistency; Indicates the length of the hit; Indicates the coverage of the query sequence; Indicates the coverage of the hit sequence; This is an indicator function; it takes the value 1 if the condition is met, and 0 otherwise. , , , , This is a preset threshold.

[0047] The aforementioned thresholds are not arbitrarily set, but are determined by the subsequent dsRNA fragment design length, cross-species co-targeting requirements, and sequence alignment reliability. Specifically, Can be set to This is used to exclude random matches; It can be set to 80% or higher to ensure that the hit regions across species have sufficient similarity to trigger an effective RNAi cascade reaction; It can be set to 300bp or 500bp to ensure that the length of the hit region can support the design of subsequent dsRNA fragments of 200bp to 500bp, and to leave redundant space to avoid complex secondary structures. and The introduction of this method is used to prevent false positives caused by high uniformity in local short fragments, ensuring that the alignment region dominates in both the target gene and the co-target gene. For closely related co-target pests, this can improve... Alternatively, a coverage threshold can be set; for transcripts with low quality or in initial screening scenarios, the threshold can be appropriately relaxed, and low-confidence-risk markers can be added to the candidate pool.

[0048] S1.3.4: Calculate the support score for a single species.

[0049] For candidate genes In co-target species The effective hit set in The system selects the highest valid support score as the supporting evidence for the candidate gene in that species. The support score can be calculated according to formula (2): (2).

[0050] in, Indicates co-target species For candidate genes Support points; , , , Representing sets The best overall performance is achieved by considering the expected value, consistency, length, and query coverage. Indicates the reference length; , , , As the weight. Preferably, can be The result is obtained after normalization, and the sum of all weights is 1. This formula integrates statistical significance, sequence similarity, relative length, and coverage into a single quantitative indicator.

[0051] S1.3.5: Calculate cross-species co-target support.

[0052] Statistical candidate genes The cross-species co-target support is calculated by determining the number of co-target species that provide effective support and combining the support scores from each species. .For example: (3).

[0053] in, This represents a collection of pest species that share a common target. The higher the value, the more likely the candidate gene is to support a multi-pest co-control design.

[0054] S1.3.6: Generate candidate pool entry reasons, forming candidate gene pool.

[0055] Based on the number of effective hits, the number of species hit, the highest consistency, the longest hit region, and whether it covers the designable length of subsequent dsRNAs, an entry reason field is generated for candidate genes. Entry reasons include, but are not limited to: "single co-target hit", "multiple co-target hits", "highly consistent short fragment hits", "moderately consistent long fragment hits", "hit covering the designable dsRNA region", and "target pest-specific region candidate". This process makes the screening logic interpretable and provides clear branching criteria for downstream design.

[0056] Target pest genes that meet the valid hit criteria and have designable regions are added to the candidate gene pool. The wide table of the candidate pool should include at least: target pest gene ID, transcript ID, functional annotation, co-target species name, co-target hit ID, expected value, sequence consistency, hit length, query coverage, target coverage, and single species support score. Cross-species co-target support Reasons for entering the candidate pool and risk marking.

[0057] This embodiment improves upon the traditional method of "manually selecting similar genes after sequence alignment" by introducing a candidate gene pool construction algorithm for RNA pesticide design. This algorithm not only determines sequence similarity but also further assesses whether the similar regions are sufficient to support subsequent dsRNA fragment design, whether they are supported by multiple co-target species, whether target pest-specific regions exist, and the specific reasons for the candidate gene's inclusion in the candidate pool. Its superiority lies in improving the reproducibility, interpretability, and seamless integration with subsequent fragment design steps of candidate gene screening. The S3 output is no longer just a list of similar sequences but a co-target candidate pool containing cross-species evidence, scoring, and design adaptation information. Specifically, this invention realizes a paradigm shift from "static sequence alignment" to "dynamic design and adaptation." By introducing coverage and absolute length constraints, it ensures that each gene in the candidate pool is not only sequence-similar, but that the similar regions are sufficient to support the subsequent engineered extraction of dsRNA fragments. At the same time, a multidimensional quantitative evaluation model for RNA pesticides is constructed, deeply coupling the biological characteristics of RNA interference technology's sensitivity to mismatches with the design requirements of dsRNA to avoid complex secondary structures. In addition, by introducing an "entry reason" mechanism, complex algorithmic decisions are transformed into readable standardized tags, enabling downstream dsRNA design to be automatically routed to the corresponding design branch directly based on the tags, significantly improving the level of intelligence and collaborative efficiency.

[0058] S2: Design multiple expert agents, use the expert agents to independently evaluate the value of candidate genes based on structured evidence planning tasks, and perform consistency checks on the scores obtained by the expert agents.

[0059] In one specific implementation, a search plan and evidence grading strategy are generated based on candidate genes and their annotation information. For example, for genes with direct RNAi evidence, priority is given to searching experimental literature from the same species; for genes lacking evidence from the same species, the search is expanded to include cross-species homologous gene evidence, key life process evidence, and functional redundancy evidence. Target discovery expert agents, target quality expert agents, and functional redundancy expert agents independently analyze candidate genes and output lethal potential, target quality, and redundancy risk-related scores and evidence summaries, respectively. Consistency checks are performed on the conclusions of multiple expert agents. If contradictions are found between literature evidence and GO annotations, BLAST results indicate high redundancy but the target score is too high, or a particular expert's evidence is insufficient, audit opinions and revision instructions are generated, and the corresponding expert agents are selectively rerun to form a traceable evidence loop.

[0060] S2.1: Design multiple expert intelligent agents.

[0061] In this embodiment, the expert agents include a target discovery expert agent, a target quality expert agent, and a functional redundancy expert agent. The target discovery expert agent is used for lethal potential assessment. This agent searches for RNAi-related evidence using the RAG tool in the literature, mines direct evidence within the same species and clues about homologous targets across species, and assesses the potential for silencing to lead to lethality, developmental inhibition, or decreased fitness. The target quality expert agent is used for function and pathway coreness assessment. This agent analyzes gene function using GO tools and protein function annotation to determine whether candidate genes participate in key life processes or core pathways, and assesses their functional necessity, pathway cascade effects, and target reliability. The functional redundancy expert agent is used for single-target effectiveness risk assessment. This agent identifies highly similar paralogous genes, isoforms, or alternative pathways using BLAST tools and literature evidence, and analyzes whether functional redundancy may weaken the RNAi silencing phenotype.

[0062] S2.2: Utilize expert agents to independently evaluate the value of candidate genes based on structured evidence planning tasks.

[0063] S2.2.1: Planning task for generating structured evidence for candidate genes.

[0064] After obtaining the candidate gene pool, if expert agents directly assemble search terms and score them synchronously, evidence acquisition and conclusion generation become coupled, making it difficult to distinguish between direct RNA interference evidence from the same species and indirect evidence from across species. Therefore, this invention adds an evidence planning step before expert evaluation, for each candidate gene... Planning task for generating structured evidence : (4).

[0065] in, Candidate genes Functional annotations, sequence similarity annotations, and a set of synonyms; For optional gene database name, alias context, and species name context, This indicates that the generation rules of the large language model are invoked. Its function is not to directly output target scores, but to specify "what evidence to collect, at what level to collect evidence, and which expert agent to process the evidence".

[0066] Furthermore, Represented as a set of three types of tasks: (5).

[0067] in, A task set for retrieving evidence of RNAi lethal potential; For a set of evidence retrieval tasks with functional redundancy; Parameters are called for the functional annotation tool. For the two categories of retrieval tasks—lethal and redundant—they are further subdivided according to the level of evidence: (6).

[0068] in, Level 1 evidence requires linking the target species to the gene or protein entity, and is used for direct experiments within the same species or as corroborating evidence within the species. Level 1 evidence allows for expanded recalls under insect homology or despecies constraints, but only as a supplement when direct evidence is lacking, and must not be used in conjunction with... Interpretation of mixed evidence. Please provide at least the species field, protein name field, gene ontology tool query field, and homologous protein location field.

[0069] During generation, the system first... Perform cleaning and name entity parsing, using [methods] when necessary. Merge official aliases; then generate the aforementioned tiered tasks separately and remove duplicates. If automatic generation fails, proceed according to the rule template. The rollback yields the minimum completeness. To ensure the process is not interrupted. Once completed, With candidate genes Binding: The target discovery expert agent is primarily responsible for execution. The target quality expert agent mainly performs The functionally redundant expert agent mainly performs .

[0070] Therefore, evidence planning and expert scoring are decoupled, allowing subsequent assessments and audits to be conducted in accordance with the established procedures. The check examines whether evidence was collected according to plan, whether evidence was used beyond its proper scope, and whether indirect evidence was misinterpreted as direct evidence. The advantage of this step lies in the fact that, before the expert agent generates a conclusion, it uses a structured task to constrain the source of evidence, its level, and tool usage parameters, ensuring the traceability and verifiability of candidate gene evaluation.

[0071] S2.2.2: Multi-expert agents independently evaluate the value of candidate genes.

[0072] In completing the structured evidence planning task Subsequently, this embodiment does not rely on a single model to provide a conclusion on the superiority or inferiority of a target at once. Instead, it sets up multiple expert agents to evaluate the same candidate gene within their respective evidence boundaries. Independent assessments are conducted, and traceable sub-value representations are output separately. The purpose is to separate the three different types of questions—"whether it is lethal or severely suppressive," "whether the function is critical enough," and "whether it is easily compensated for after silence"—to avoid confusion of evidence types and the indivisibility of scoring rationale.

[0073] Specifically, for each candidate gene The system activates the following three expert agents and executes them respectively. The corresponding task in the middle.

[0074] (1) A target discovery expert agent is used for lethal potential assessment. This expert mainly performs... In the lethality-related search task, combined with evidence such as RNA interference, double-stranded RNA silencing, knockdown, and lethality or developmental inhibition obtained through literature search enhancement, the likelihood of silencing leading to death, developmental arrest, or a significant decrease in fitness is assessed, and a lethality potential score is output. And a summary of evidence. Preferably, the level of evidence in the literature is determined first, and then a score is awarded within the corresponding upper limit. Direct experimental evidence of the same gene in the same species can support a higher score, while when there is only prior evidence of homology, same pathway or function across species, the upper limit of the score is lowered accordingly to suppress inflated scores without strong evidence.

[0075] (2) Target quality expert agent, used for functional and pathway core assessment. This expert mainly performs... The system performs functional annotation parsing tasks and calls a gene ontology query tool (based on the online query results of UniProt) to obtain gene annotation information. Combining molecular function, biological process, cellular component, and network location information, it determines whether candidate genes are located in critical life processes, core pathway nodes, or physiological bottlenecks related to pest control, and outputs target quality scores. And functional description. The expert provided annotation-driven target quality priors, which do not replace lethal experimental evidence.

[0076] (3) A functionally redundant expert agent is used for single-target effectiveness risk assessment. This expert mainly performs... The system performs redundant related retrieval tasks and calls a local paralogous sequence alignment tool to execute sequence alignment algorithms. It also enhances the generated results by combining literature searches on redundant topics, identifying compensatory factors such as highly similar paralogous genes, isozymes, alternative pathways, and co-expression of gene family subtypes, and outputs a functional redundancy risk score. .in, The larger the value, the higher the risk of compensation and phenotypic attenuation after single-target silencing.

[0077] The assessments of the three experts are independent of each other and can be formalized as follows: (7).

[0078] in, , , The evaluation process involves three categories of experts: lethal potential, target quality, and functional redundancy. Each expert is prohibited from rewriting the sub-results of other experts. All outputs are saved as structured records, including at least gene identifiers, sub-scores, evidence summaries, and citations of the tools or literature used, for subsequent evidence auditing and overall ranking.

[0079] The scoring uses a uniform unit, all mapped to a score of 1 to 10.

[0080] Through this step, the present invention obtains independent value representations of candidate genes in three orthogonal dimensions: lethal potential, target quality, and functional redundancy. This provides decomposable and comparable inputs for subsequent evidence consistency audits and comprehensive recommendation scoring, rather than an uninterpretable single black-box score.

[0081] S2.3: Perform a consistency check on the scores obtained by the expert agent.

[0082] The three expert agents each output a lethal potential score. Target quality score Functional redundancy risk score Subsequently, this embodiment does not directly proceed to the comprehensive ranking. Instead, the Critic agent checks the consistency between the expert narrative and the scoring rationale against the tool's evidence. If necessary, selective reruns are initiated only for experts with problematic narratives. The aim is to form a closed loop of "assessment-audit-revision," reducing the risk of missing evidence, conclusions exceeding acceptable levels, or obvious contradictions between experts going undetected.

[0083] It is important to emphasize that this step does not recalculate the overall recommendation score, nor does it replace the individual scores given by the three experts; its output is the audit opinion and revision instructions. The evidentiary benchmarks used by the audit should at least include the structured evidence planning task. The execution results, as well as objective materials generated by the aforementioned steps, such as gene ontology, protein annotation, literature search enhancement to generate hits, and paralogous sequence alignment.

[0084] For candidate genes The consistency check should preferably cover the following aspects: 1. Whether the target quality description is consistent with the gene ontology and protein annotation; 2. Whether the lethal reasoning is consistent with the level of evidence and abstract content of the adopted literature, and whether there are any cases of "giving an excessively high lethal judgment without direct evidence" or other situations that exceed the appropriate level; 3. Whether the redundancy assessment is consistent with the collateral comparison and the background of the redundant literature; 4. Are there any obvious narrative conflicts among the three experts? For example, the comparison of related species strongly suggests high redundancy, but the redundancy risk score is abnormally low; or the target discovery expert agent claims to have direct lethal evidence of the same species, but the literature does not support it.

[0085] Audit results can be formalized as follows: (8).

[0086] in, Indicate whether revision is needed, when This indicates that revisions are needed; This indicates the group of experts who need to rerun the process. These are actionable revision instructions that specify which instrumental evidence should be used to revise the narrative and scores. A wide range of scores is not considered abnormal in itself; it is only considered abnormal when the instrumental evidence is significantly inconsistent with expert conclusions, or when there is substantial conflict among experts. .

[0087] when At that time, the system only... The experts in the process returned to the corresponding evidence collection and evaluation process, and will Inject its hint constraints, and overwrite the original sub-item outputs with the revised results; expert results not included in the subset remain unchanged. It is preferable to set the upper limit of audit-rerun rounds to 1 to avoid repeated disturbances that could lead to unstable results. If, after a rerun, no new necessary revisions are triggered, then proceed to the subsequent comprehensive recommendation score and level determination.

[0088] Through this step, this embodiment ensures that the results of multi-expert evaluations undergo a consistent and verifiable evidence check before being ranked, and controls the scope of revisions and computational costs through selective reruns, thereby improving the auditability and robustness of target decisions.

[0089] S3: Calculate the comprehensive recommendation score of candidate genes based on the scores of passing the consistency check, and obtain the ranking and grade of candidate genes based on the comprehensive recommendation score of candidate genes.

[0090] In one specific implementation, after completing consistency checks and necessary selective reruns, lethal potential is assessed based on the audited itemized scoring. Target quality Risk of functional redundancy Calculate the comprehensive recommendation score for candidate genes. Based on this, a ranking and grade are output to determine which genes will be included in subsequent double-stranded RNA fragment design. A comprehensive recommendation score is calculated and a target ranking is output, which can mark candidate genes as Recommended, Pending, or Not Recommended.

[0091] because This indicates redundancy risk, and a larger value indicates a higher risk. The complementary amount is used as a low-redundancy advantage in the overall score. When each item uses a 1 to 10-point scale, the overall recommendation score can be calculated using formula (9): (9).

[0092] And satisfy: (10).

[0093] in, Candidate genes Overall recommendation rating; Rate the lethal potential; Target quality scoring; Assess the risk of functional redundancy. , , These are the weights for lethal potential, target quality, and low redundancy advantage, respectively. The preferred default weights are... , , The ranking is influenced by lethal potential, target quality, and low redundancy advantage; the weights can also be adjusted according to pest type, crop scenario, or experimental strategy.

[0094] In equation (9) This indicates that the lower the redundancy risk, the greater the advantage of low redundancy.

[0095] The overall score is calculated deterministically by the program according to fixed rules, rather than being generated by calling a large model again, thus ensuring that tied scores are ranked in the same order and that results are reproducible. After calculation, according to... Candidate genes were sorted in descending order and divided into three levels: 1. Recommendation: Prioritize designing subsequent double-stranded RNA fragments.

[0096] 2. Pending: Can be considered as an alternative depending on experimental resources and research needs.

[0097] 3. Not recommended: In principle, do not proceed to the design stage of high-cost segments.

[0098] Preferably, a threshold is used on a scale of 1 to 10. , :when When it is determined to be recommended; when When it is determined to be pending; when It is judged as not recommended at this time.

[0099] This step combines the interpretable component judgments of multi-expert agents with deterministic weighting on the program side, making the final sorting stable, reproducible, and auditable, and providing clear target access boundaries for subsequent segment design.

[0100] S4: Generate candidate double-stranded RNA fragments based on the order and grade of candidate genes and calculate the pre-score of function. Screen the candidate double-stranded RNA fragments based on the pre-score of function to obtain the candidate fragments retained in the initial screening. Then, assess the target hit and off-target risks of the candidate fragments retained in the initial screening to obtain the dual-mode fragment recommendation results.

[0101] In one specific implementation, the priority target is transformed into dsRNA candidate fragments that can be experimentally validated. First, the target CDS is extracted, and a full CDS variable-length window search is performed to obtain the coordinates, length, and sequence of the candidate dsRNA fragments. Then, virtual Dicer cutting and siRNA functional rule scoring are performed. For each candidate fragment, virtual Dicer cutting is performed to generate potential siRNA windows, and siRNA activity characteristics are calculated based on rules such as guanine-cytosine content (GC content, referring to the proportion of guanine G and cytosine C in the sequence), thermodynamic asymmetry, seed region and cutting-related site preference, low complexity, and consecutive base anomalies. On-target / off-target comparisons are then performed in the transcriptomes of the target pest, co-target pests, and non-target species. Two categories of candidate fragments are output: control-priority and safety-priority. The candidate fragments and their siRNA windows are compared with the transcriptomes of the target pest, co-target pests, and non-target species to obtain on-target coverage and off-target risk. Based on different application objectives, two types of recommendation results are output: the prevention-first mode improves the targeted hit coverage among common target pests; the safety-first mode reduces the potential off-target risk of non-target species.

[0102] S4.1: Generate candidate double-stranded RNA fragments based on the sorting and ranking of candidate genes and calculate pre-functional scores.

[0103] In one specific implementation, for target genes identified as recommended or pending, this embodiment further converts them into experimentally verifiable double-stranded RNA (dsRNA) candidate fragments. Before entering the cross-transcriptome targeting and off-target evaluation, fragment enumeration, virtual cleavage enzyme processing simulation, and functional pre-scoring based on the small interfering RNA multi-criteria hierarchical scoring algorithm proposed in this embodiment are completed. A candidate window is systematically generated within the coding sequence design space, and the potential silencing efficacy is estimated using interpretable and reproducible rules, compressing the candidate size for subsequent expensive alignments.

[0104] Specifically, the following steps are included: S4.1.1: Extract the target gene coding sequence based on the annotation and generate candidate double-stranded RNA fragments.

[0105] First, the target gene coding sequence (CDS) is extracted based on the annotation, preferably requiring the designed interval to fall entirely within the coding region. Then, a variable-length window search is performed across the entire coding sequence: the fragment length is preferably between 200 and 500 base pairs, and can be stratified by length, for example, 200 to 299, 300 to 399, or 400 to 500 base pairs; the system moves to the set step size and adjusts the length accordingly. For each window... ,in Starting point To obtain a candidate fragment set, record the coordinates, length, and nucleotide sequence. .

[0106] Compared to sliding over a fixed length or manually selecting a single interval, this step makes the design space coverage more complete and retains comparable candidates for different length layers.

[0107] S4.1.2: Perform virtual cleavage enzyme processing simulation on each candidate double-stranded RNA fragment to obtain a set of small interfering RNAs.

[0108] For each candidate fragment The system performs virtual enzyme cleavage, i.e., Dicer processing simulation, to generate potential small interfering RNA (siRNA) windows according to a preset product length; preferably, 21 nucleotides are used as the main product, and auxiliary windows of 16, 26, or more nucleotides can be set. This transforms fragment-level double-stranded RNA into a scoreable set of siRNAs. This serves as the basic unit for subsequent functional pre-scoring.

[0109] S4.1.3: Utilize a multi-criteria hierarchical scoring algorithm for small interfering RNAs to pre-scorrate the small interfering RNA set calculation function.

[0110] This embodiment proposes a multi-criteria hierarchical scoring algorithm for small interfering RNA (SRNA) designed for double-stranded RNA. The algorithm uses the positive strand window as the scoring object, first performing hard-constraint elimination, and then calculating functional pre-scoring based on stratification using basic scoring, dynamic scoring, and site deduction.

[0111] S4.1.3.1: Hard elimination.

[0112] For any window sequence: If the A / U ratio at positions 15 to 19 is zero, it is directly identified as a rejection window. Hard rejection is used to eliminate low-quality windows with insufficient terminal asymmetry and difficulty in forming an effective leader chain preference.

[0113] S4.1.3.2: Window-level hierarchical scoring.

[0114] For windows that are not eliminated, calculate the total score according to formula (11): (11). in: Basic bonus points Defined as: (12).

[0115] in, This indicates whether the guanine-cytosine content falls within the preferred range. For example, the preferred range is 30% to 52%. If it meets the requirement, the value is 1; otherwise, the value is 0. This indicates whether the internal hairpin determination is successful. If there is no stable hairpin or the hairpin melting temperature is lower than the preset threshold, the value is 1; otherwise, the value is 0. This represents the base preference score for key sites, with a preference count for the number of terms where the 19th base is A, the 3rd base is A, and the 10th base is U. .

[0116] Dynamic bonus points Defined as: (13).

[0117] in, Display window The number of A / U bits from the 15th to the 19th bit, with an upper limit of 5, is used to characterize the asymmetric advantage corresponding to the 3' end being rich in A / U.

[0118] Site deduction Defined as follows: if the 19th bit is G / C, add 1; if the 13th bit is G, add 1. .

[0119] like If the condition is met, the window will be added as an eliminated window; otherwise, it will be retained. .when Time is used as an efficient window; when This is recorded as a super window. For example, It can be set to 6. It can be set to 9.

[0120] Equation (10) incorporates compositional constraints, secondary structure constraints, key site preferences and terminal asymmetry into the same additive framework, making the quality of a single small interfering RNA decomposable and reproducible; hard elimination and deduction suppression are obviously unfavorable configurations, while dynamic bonus rewards are more likely to have active asymmetric characteristics.

[0121] It should be noted that in this embodiment, the 19th position being A, the 3rd position being A, and the 10th position being U are preferred as key sites. This is based on the sequence preference characteristics in the interaction process between small interfering RNA and the RNA-induced silencing complex (RISC) in the RNA interference mechanism.

[0122] Specifically, the functional activity of small interfering RNA (sRNA) is influenced not only by its overall base composition but also by guide strand selection, 5' end thermodynamic asymmetry, and the base composition at specific sites. Position 19, located near the end of the guide strand, is associated with guide strand preference. Positions 3 and 10 are in important regions related to target recognition and cleavage of sRNA, and their base preferences have been used in existing RNA interference activity evaluation studies to predict silencing efficiency. Therefore, this embodiment uses these sites as local sequence-level features to construct an interpretable sRNA activity pre-scoring index.

[0123] S4.2: Screen candidate double-stranded RNA fragments based on functional pre-scores to obtain candidate fragments retained in the initial screening.

[0124] Define fragments above window-level scores. The aggregation feature. For eliminated or unscored windows, the score is treated as 0, recorded as... .definition: (14), (15).

[0125] in, , , These are the number of high-efficiency, super, and low-quality windows, respectively; low-quality windows include discarded windows or... The window. The segment function pre-score, i.e. the first stage comprehensive ranking score, can be calculated according to formula (16): (16).

[0126] Equation (16) rewards high efficiency and super window density while preserving average activity, and doubles the penalty for poor window density, in order to reduce the risk of generating a large number of inefficient small interfering RNAs after double-stranded RNA is processed, triggering competitive inhibition of the silencing complex.

[0127] Each length layer according to Candidate fragments are sorted in descending order, and the top few are selected for subsequent targeted hit and off-target risk assessment, thus forming a two-stage design of "rule-based pre-screening + transcriptome assessment".

[0128] This embodiment achieves a computable transformation from recommended target sites to candidate double-stranded RNA fragments: variable-length search of the entire coding sequence ensures the integrity of the design space; virtual cutting enzymes simulate adaptor fragments and small interfering RNA products; and a multi-criteria hierarchical scoring and fragment-level pre-scoring constitute an interpretable rule engine. Compared to methods that only use length or GC content for screening, this step further considers terminal asymmetry, secondary structure, key site preference, and poor window density, thereby providing higher-quality input for subsequent dual-mode recommendations prioritizing prevention and control and prioritizing safety.

[0129] S4.3: Target hit and off-target risk assessment is performed on the candidate fragments retained in the initial screening to obtain dual-mode fragment recommendation results.

[0130] After generating and pre-scoring candidate double-stranded RNA fragments, this embodiment further compares the target hit and off-target risks at the transcriptome level for the candidate fragments retained in the initial screening. Based on the same feature set, it outputs two categories of recommendation results: prevention-first and safety-first, according to the application objectives. The purpose is to transform "accurate targeting" and "minimum damage" into calculable indicators, and to form differentiated solutions through objective function switching, rather than two unrelated black-box processes.

[0131] S4.3.1: The candidate double-stranded RNA fragments retained after functional pre-scoring and their derived small interfering RNA sequences generated by virtual Dicer processing simulation are compared with the transcriptomes of the target pest, co-target pests, and non-target species, respectively, to obtain the control-related coverage evidence and safety-related risk evidence for each candidate fragment. Specifically, the double-stranded RNA fragments are compared with the transcriptomes of the target pest, co-target pests, and non-target species, and the target coverage, co-target coverage, and potential off-target risk-related indicators are obtained based on the comparison results.

[0132] The following indicators are statistically analyzed: 1. Targeted hit metrics: such as the number of hit events, coverage, and hit quality, forming a target coverage score. Common target hit score ; 2. Off-target risk indicators: such as the number and quality of potential hit events in the non-target transcriptome or genes outside the target genome, forming an off-target risk score. .

[0133] In off-target evaluation, self-hit interference from the candidate fragment's own source sequence is eliminated, thus... This more accurately reflects the risk of accidentally harming other genes or species. This step outputs prevention-related coverage evidence and safety-related risk evidence for each candidate fragment, serving as input for dual-mode sorting.

[0134] S4.3.2: For each candidate fragment, the evidence related to prevention and control and the evidence related to safety risks are used to generate recommendation results based on a dual-mode recommendation objective function.

[0135] On the same set of candidate features, two application modes, namely prevention and control priority mode and safety priority mode, are achieved by switching weights.

[0136] The prevention-priority mode prioritizes maximizing the target and the hit rate among common targets, and can be calculated using equation (17): (17).

[0137] The safety-first mode prioritizes minimizing the potential off-target risk, and can be calculated using equation (18): (18).

[0138] in, Representing fragments Priority scoring for prevention and control; Representing fragments Safety priority score; Indicates the target coverage score; Indicates the off-target risk score; , , , For pattern weights.

[0139] In the preferred example, the prevention and control mode is taken as follows: , Safe mode , In the event of a draw, the functional pre-match score will be used. As a sequence key, it ensures that the intrinsic quality of the sequence still participates in the decision-making process.

[0140] More generally, it can also be written as: (19).

[0141] It should be noted that in this embodiment... Candidate double-stranded RNA fragment The comprehensive evaluation score can be used in both modes, and is a universal evaluation framework. Equation (19) is a universal comprehensive evaluation form, which is an extension of equations (17) and (18); by adjusting the weights of different indicators, a prevention-first mode and a safety-first mode can be formed respectively. This indicates the activity index of small interfering RNA, reflecting the potential of candidate fragments to produce effective small interfering RNA after processing. The target coverage score reflects the sequence coverage ability of the target pest and co-target pests. This indicates a multi-species coverage capability index, used to reflect the co-target control potential of candidate fragments; This represents a sequence quality index, used to reflect the sequence composition characteristics of candidate fragments; It represents the off-target risk score, which is used to reflect the potential risks in non-target species or non-target genes. , , , , These represent the weight coefficients corresponding to each evaluation indicator, and their values ​​are preset or adjusted according to the specific application objectives. The scope of prevention and control scenarios is expanding. and Increased security scenarios Both modes can be combined with an activity threshold and an upper limit for the proportion of low-quality small interfering RNA.

[0142] The principle of this step is as follows: the difference in the target is reflected as the difference in weight; the target and off-target features are calculated only once, and the objective function is executed twice during sorting, so that the two sets of results are comparable and interpretable, and are connected with the pre-scoring in step S8 to form a complete design chain of "rule coarse sorting + transcriptome fine sorting".

[0143] According to and Sort the results in descending order; or, in safety-first mode, output the results according to the risk minimization rule. The output includes: 1. Priority segments and alternative segments for epidemic prevention and control; 2. Safety-priority recommended clips and alternative clips; 3. Coordinates, length, sequence, and other details of each segment. , , , And key small interfering RNA quality indicators.

[0144] The above outputs are used for subsequent transformation constraint assessment and structured report generation.

[0145] Through the above steps, this embodiment advances double-stranded RNA candidates from "sequence-based availability" to "application-scenario-based decision-making": it can both quantify the potential for co-target control and explicitly constrain ecological off-target risks; the two modes share the same alignment features, only switching the objective function, avoiding redundant calculations and incomparable results. Compared to outputting a single optimal fragment, this step is more adaptable to different R&D and regulatory needs such as strong control and strict ecological constraints.

[0146] S5: Comprehensively evaluate process information and generate a structured R&D report.

[0147] In one specific implementation, after outputting the recommended results for two types of double-stranded RNA—prioritizing prevention and control and prioritizing safety—this invention enters the final solution integration stage: transformation constraints such as sequence-level quantitative evaluation, formulation delivery, and ecological validation are layered and merged, generating an experimentally verifiable and traceable structured R&D report. The aim is to further transform "calculated optimal fragments" into "practical and discussable R&D solutions," while avoiding directly rewriting sequence scores with unverifiable textual preferences.

[0148] This embodiment integrates fragment length, GC content, siRNA activity, high-efficiency / low-quality siRNA ratio, target hit coverage, off-target risk, delivery formulation, and ecological validation recommendations to form a principal dsRNA protocol, alternative protocols, indicator comparisons, and a structured R&D report. Specifically, it includes the following steps: S5.1: Merge information according to the hierarchical merging principle.

[0149] This step does not involve simply piecing together various types of information, but rather merging them according to the following three principles: First layer: Sequence layer quantitative master sorting.

[0150] Functional pre-scoring Prioritize scoring for prevention and control Safety Priority Score Based on this, the primary and alternative fragments are identified for each of the two modes. The sequence-level conclusions remain consistent with the aforementioned deterministic rules and alignment metrics.

[0151] The second layer: Conversion layer constraints and risk labeling.

[0152] After the primary and alternative segments are determined, the system introduces conversion-related evidence and suggestions, including at least: Delivery and formulation recommendations: Based on publicly available knowledge and search results, we provide directional recommendations on oral administration, spraying, carrier materials, stability, etc.

[0153] Ecological and validation pathway recommendations: Provide suggested items for non-target screening, indoor bioassay, and progressive validation.

[0154] This layer is attached to the scheme by default in the form of annotations and descriptions, and the evidence in this layer is not allowed to directly rewrite the sequence quantitative score. When a pre-set hard constraint conflict is triggered, such as when the search shows prior rights risks that are highly related to the candidate strategy, the corresponding fragment can be downgraded to an alternative or a risk mark can be added.

[0155] The third layer: structured aggregation of the report layer.

[0156] The target assessment evidence chain and fragment design results are organized and output according to fixed chapters to form a unified deliverable. The above merging relationship can be summarized as equation (20): (20).

[0157] in, Indicates targeting candidate genes A structured research and development solution; This indicates the selection of the best sequence for layer sorting; Indicates transformation constraint annotation; This indicates a structured merging of "sorting results + constraint descriptions", rather than arbitrary reweighting of scores.

[0158] S5.2: Generate structured R&D reports.

[0159] The report should include at least: 1. Recommended target list with recommended, pending, and not recommended levels, along with lethal potential ratings. Target quality score Functional redundancy risk score Overall Recommendation Score ; 2. Whether evidence planning, multi-expert evaluation, and evidence auditing trigger revised summaries to form a traceable chain of evidence; 3. Prioritize and select double-stranded RNA fragments for both prevention and safety, including coordinates, length, sequence, guanine-cytosine content, small interfering RNA activity, proportion of high-efficiency small interfering RNA and proportion of low-quality small interfering RNA; 4. Evidence of target hits in target pests and co-target pests, and off-target risks in non-target species or non-target genes; 5. Delivery formulation recommendations, ecological screening, and subsequent experimental validation pathways.

[0160] The dosage, median lethal concentration, and other information in the literature are for indoor exploration reference only and should not be used as recommendations for field application or product labeling.

[0161] Through this step, this embodiment completes a full closed loop from multi-source transcriptome and intelligent target evaluation to dual-mode double-stranded RNA design, and then to pre-integration of transformation constraints. Compared to simply outputting a list of sequences or untraceable natural language suggestions, this step forms a research and development deliverable aligned with "indicators-evidence-solutions," facilitating experimental verification and iterative write-back, thereby improving the systematic nature and engineering feasibility of early-stage RNA pesticide research and development decisions.

[0162] This embodiment realizes intelligentization of the entire RNA pesticide development chain. It does not only design dsRNA fragments for a given gene, but starts from multi-source transcriptome resources and integrates candidate gene mining, intelligent target evaluation, dsRNA fragment design, sequence and application evaluation, and structured report output.

[0163] The method in this embodiment improves the interpretability of target selection. RNAi lethal potential, target quality, and low redundancy advantages are independently evaluated by different expert agents, and consistency is checked through the Critic evidence audit module, ensuring that each recommended target has a corresponding chain of evidence.

[0164] This embodiment reduces the risk of failure due to functional redundancy. By incorporating paralogous genes, isoforms, and phenotypic compensation caused by alternative pathways as independent risk dimensions into the score, it is possible to screen out or reduce the priority of high-redundancy risk targets at an early stage.

[0165] This embodiment of the method balances control effectiveness with ecological safety. It simultaneously performs transcriptome comparisons of dsRNA candidate fragments with those of target pests, co-target pests, and non-target species, and outputs two recommendation categories: control priority and safety priority, adapting to different application scenarios.

[0166] This embodiment allows for the early application of product transformation constraints. By incorporating delivery formulation and ecological validation recommendations into the final scheme evaluation, it avoids discovering delivery limitations or ecological risks only after the candidate fragment has entered subsequent experiments.

[0167] This embodiment of the method forms an iterative R&D decision-making closed loop. Through Query Planner, expert agents, Critic auditing, and selective rerun mechanisms, the system can self-revise based on evidence conflicts or scoring anomalies, rather than outputting an untraceable conclusion all at once.

[0168] To better illustrate the superiority of the method in this embodiment, a dsRNA scheme for a certain lepidopteran pest was designed: Using a specific lepidopteran pest as the target pest, and selecting transcriptomes from several co-target pests and non-target species as input, the system first retrieves transcriptome, genome annotation, and protein annotation resources for the target pest and related species from public databases, generating a unified species-level transcriptome FASTA. Subsequently, the system performs transcriptome cross-comparison, screening for shared similar genes based on rules of hit length greater than 500 bp, identity greater than 80%, and E-value less than or equal to 1 × 10^-5, and constructing a candidate gene pool.

[0169] For candidate genes, the Query Planner generates a search plan. The target discovery expert agent retrieves RNAi experimental evidence and cross-species homologous target clues; the target quality expert agent analyzes GO functions and key life processes; and the functional redundancy expert agent identifies paralogous genes and alternative pathways. The Critic evidence auditing module checks the consistency of the three expert outputs. If a candidate gene has highly similar paralogous genes but the target discovery expert agent gives a high recommendation, the functional redundancy expert agent is required to supplement the analysis, and if necessary, lower the overall ranking of that gene.

[0170] CDS (Cellular Dataset) is extracted from the top-ranked target genes, and a full CDS variable-length window search is performed to generate multiple candidate dsRNA fragments. For each fragment, the system performs virtual Dicer cutting, calculates the GC (Gross Genetic Computation), thermodynamic asymmetry, key site preference, and low-quality filtering indicators of the potential siRNA, and compares it with the transcriptomes of target pests, co-target pests, and non-target species. Finally, the system outputs control-priority fragments and safety-priority fragments, respectively, and provides the coordinates, length, sequence, score, evidence of target hit, off-target risk warning, and suggestions for further validation for each fragment.

[0171] The output structured synthesis report may include at least: (1) A list of recommended target genes, including three categories of markers: Recommended, Pending, and Not Recommended; (2) Lethal potential, target quality, functional redundancy risk and comprehensive score of each recommended target; (3) The main dsRNA fragment and the alternative dsRNA fragment, including coordinates, length, sequence, GC content, siRNA activity, and the ratio of high-efficiency / low-quality siRNA; (4) On-target coverage of target pests and co-target pests; (5) Off-target risk in non-target species; (6) Recommendations for delivery formulation, recommendations for ecological risk screening and subsequent experimental verification pathways.

[0172] Through the above implementation methods, this embodiment can organize information that was originally scattered across literature, transcriptomics, functional annotation, BLAST alignment, dsRNA design rules, and application validation experience into a unified intelligent decision-making process. Compared with manual screening or single-point dsRNA design tools, this invention can improve the efficiency of candidate gene screening, enhance the interpretability of target scoring, reduce functional redundancy and off-target risks, identify transformation barriers in advance, and provide traceable candidate solutions for subsequent experimental validation.

[0173] Example 2: Embodiment 2 of the present invention provides an RNA pesticide research and evaluation system based on multi-expert agent collaboration, comprising: The data acquisition module is used to acquire transcriptome sequence resources and associated annotation resources of target pests, co-target pests and non-target species, and to standardize the transcriptome resources to form species-level transcriptome sequence files and annotation tables in a unified format. The module performs transcriptome cross-comparison between target pests and co-target pests, and extracts shared similar genes according to preset screening conditions to obtain candidate genes. The agent evaluation module is used to design multiple expert agents, utilize the expert agents to independently evaluate the value of candidate genes based on structured evidence planning tasks, and perform consistency checks on the scores obtained by the expert agents. The candidate gene screening module is used to calculate the comprehensive recommendation score of candidate genes based on the scores of passing the consistency check, and to obtain the ranking and grade of candidate genes based on the comprehensive recommendation score of candidate genes. The fragment recommendation module is used to generate candidate double-stranded RNA fragments based on the order and grade of candidate genes and calculate the functional pre-score. The candidate double-stranded RNA fragments are screened based on the functional pre-score to obtain the candidate fragments retained in the initial screening. The target hit and off-target risk of the candidate fragments retained in the initial screening are assessed to obtain the dual-mode fragment recommendation results. The report generation module is used to comprehensively evaluate process information and generate structured R&D reports.

[0174] Example 3: Embodiment 3 of the present invention provides a computer-readable storage medium storing a computer program adapted for loading by a processor and executing the steps in the RNA pesticide research and evaluation method based on multi-expert agent collaboration as described in Embodiment 1 of the present invention.

[0175] Example 4: Embodiment 4 of the present invention provides a computer device, the device comprising: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the steps in the RNA pesticide research and evaluation method based on multi-expert agent collaboration as described in Embodiment 1 of the present invention.

[0176] Example 5: Embodiment 5 of the present invention provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in the RNA pesticide research and evaluation method based on multi-expert agent collaboration as described in Embodiment 1 of the present invention.

[0177] The steps and methods involved in Examples 2, 3, 4 and 5 above correspond to those in Example 1. For specific implementation methods, please refer to the relevant description section of Example 1.

[0178] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0179] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The computer-readable storage medium can be any available medium that a computer can access or a data processing device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, an optical medium, or a semiconductor medium, etc.

[0180] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for evaluating the research and development of RNA pesticides based on multi-expert agent collaboration, characterized in that, Includes the following steps: Transcriptome sequence resources and associated annotation resources of target pests, co-target pests and non-target species are obtained, and the transcriptome resources are standardized to form species-level transcriptome sequence files and annotation tables in a unified format. Transcriptome cross-comparison is performed between target pests and co-target pests, and shared similar genes are extracted according to preset screening conditions to obtain candidate genes. Design multiple expert agents, use the expert agents to independently evaluate the value of candidate genes based on structured evidence planning tasks, and perform consistency checks on the scores obtained by the expert agents. The candidate genes are ranked and ranked based on the scores of those that pass the consistency check. Candidate double-stranded RNA fragments are generated based on the sequencing and ranking of candidate genes, and functional pre-scores are calculated. Candidate double-stranded RNA fragments are screened based on functional pre-scores to obtain candidate fragments that are initially retained. The target hit and off-target risks of the candidate fragments that are initially retained are assessed to obtain dual-mode fragment recommendation results. A structured R&D report is generated by comprehensively evaluating information from the process.

2. The RNA pesticide R&D evaluation method based on multi-expert agent collaboration as described in claim 1, characterized in that, The specific steps for comparing the transcriptomes of the target pest and co-target pests, and extracting shared similar genes according to preset screening conditions, are as follows: Construct a common target alignment index; Based on the co-target alignment index, a forward alignment is performed from the target pest to the co-target pest to obtain a candidate hit set; Perform a validity determination of the candidate hits based on the candidate hit set; Calculate the support score for a single species; Calculate cross-species co-target support; The reasons for entering the candidate pool are generated, thus forming a candidate gene pool.

3. The RNA pesticide R&D evaluation method based on multi-expert agent collaboration as described in claim 1, characterized in that, The expert agents include target discovery expert agents, target quality expert agents, and functional redundancy expert agents. The target discovery expert agent is used for lethal potential assessment, the target quality expert agent is used for function and pathway coreness assessment, and the functional redundancy expert agent is used for single target effectiveness risk assessment.

4. The RNA pesticide research and evaluation method based on multi-expert agent collaboration as described in claim 1, characterized in that, The specific steps for generating candidate double-stranded RNA fragments and calculating functional pre-scores based on the sequencing and ranking of candidate genes are as follows: Extract the target gene coding sequence based on the annotation and generate candidate double-stranded RNA fragments; For each candidate double-stranded RNA fragment, a virtual enzyme cleavage simulation was performed to obtain a set of small interfering RNAs; A multi-criteria hierarchical scoring algorithm for small interfering RNAs was used to pre-score the function of small interfering RNA sets.

5. The RNA pesticide research and evaluation method based on multi-expert agent collaboration as described in claim 4, characterized in that, The small interfering RNA multi-criteria stratified scoring algorithm uses the positive chain window as the scoring object. It first performs hard constraint elimination, and then performs functional pre-scoring calculation according to the stratification of basic score addition, dynamic score addition, and site deduction.

6. The RNA pesticide R&D evaluation method based on multi-expert agent collaboration as described in claim 1, characterized in that, The specific steps for assessing the target hit and off-target risks of the candidate fragments retained in the initial screening are as follows: Candidate double-stranded RNA fragments selected and retained after functional pre-scoring and their derived small interfering RNA sequences generated by virtual Dicer processing simulation were compared with the transcriptomes of target pests, co-target pests, and non-target species to obtain control-related coverage evidence and safety-related risk evidence for each candidate fragment. For each candidate fragment, the prevention and control-related coverage evidence and the safety-related risk evidence are used to generate recommendation results based on a dual-mode recommendation objective function.

7. An RNA pesticide research and evaluation system based on multi-expert agent collaboration, characterized in that, include: The data acquisition module is used to acquire transcriptome sequence resources and associated annotation resources of target pests, co-target pests and non-target species, and to standardize the transcriptome resources to form species-level transcriptome sequence files and annotation tables in a unified format. The module performs transcriptome cross-comparison between target pests and co-target pests, and extracts shared similar genes according to preset screening conditions to obtain candidate genes. The agent evaluation module is used to design multiple expert agents, utilize the expert agents to independently evaluate the value of candidate genes based on structured evidence planning tasks, and perform consistency checks on the scores obtained by the expert agents. The candidate gene screening module is used to calculate the comprehensive recommendation score of candidate genes based on the scores of passing the consistency check, and to obtain the ranking and grade of candidate genes based on the comprehensive recommendation score of candidate genes. The fragment recommendation module is used to generate candidate double-stranded RNA fragments based on the order and grade of candidate genes and calculate the functional pre-score. The candidate double-stranded RNA fragments are screened based on the functional pre-score to obtain the candidate fragments retained in the initial screening. The target hit and off-target risk of the candidate fragments retained in the initial screening are assessed to obtain the dual-mode fragment recommendation results. The report generation module is used to comprehensively evaluate process information and generate structured R&D reports.

8. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the RNA pesticide research and evaluation method based on multi-expert agent collaboration as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1-6: the RNA pesticide research and evaluation method based on multi-expert agent collaboration.

10. A computer device, characterized in that, include: A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the RNA pesticide research and evaluation method based on multi-expert agent collaboration as described in any one of claims 1-6.

Images