43 KiB
Executable File
A literature-grounded executable multi-agent system for RNA-seq-guided plant gene-family analysis
中文题目建议：面向 RNA-seq 候选发现与可复现植物基因家族分析的文献知识驱动可执行多智能体系统
Alternative titles
- From papers to pipelines: a literature-grounded multi-agent system for reproducible plant gene-family analysis
- An evidence-aware executable multi-agent framework for RNA-seq-guided crop gene-family analysis
- Integrating literature RAG, RNA-seq discovery and executable workflows for plant gene-family analysis
中文说明：推荐使用第一个英文题目作为主标题。它同时包含 literature-grounded、executable、multi-agent、RNA-seq-guided 和 plant gene-family analysis，能准确覆盖本文的 RAG、bbav5 执行系统、RNA-seq 候选家族发现和文献方法复现四条主线。
Abstract
Plant gene-family analysis is a central task in agricultural genomics, but practical studies often require a long chain of heterogeneous steps, including gene identification, phylogenetic reconstruction, motif and domain analysis, promoter and synteny analysis, expression profiling, and functional candidate prioritization. These workflows become harder to reproduce when methods are described only in natural language in published papers, when RNA-seq analysis is disconnected from downstream gene-family interpretation, and when AI systems provide recommendations without executing auditable bioinformatics procedures. Here we present a literature-grounded executable multi-agent system that connects retrieval-augmented literature understanding with reproducible plant gene-family and RNA-seq workflows. The system consists of two coupled layers. The knowledge layer extracts source-relation-target triples, methodology observations, tools, parameters, inputs, outputs and evidence sentences from gene-family papers using a V15.3.4 extraction pipeline, semantic review, and graph/vector retrieval over V15 and PlantConnectome-derived knowledge bases. The execution layer, implemented in bbav5, maps literature-derived observations and user requests to executable stages, performs preflight validation, runs command-line bioinformatics tools, diagnoses failures, and records results through manifest-first artifact validation and reproducibility bundles. The built-in workflows cover both core gene-family stages and RNA-seq stages, providing the infrastructure for raw RNA-seq data to support candidate gene-family recommendation before downstream gene-family analysis.
We evaluate the framework along three axes: methodology extraction and alignment, executable workflow validation, and RNA-seq-guided candidate discovery design. In a curated methodology benchmark, the filtered evaluation achieved a stage-level F1 score of 0.951, tool-level F1 of 0.792, parameter accuracy of 0.825, and bbav5 alignment accuracy of 0.960, while evidence-sentence accuracy remained lower at 0.638 and is treated as a current limitation. In a real bbav5 full-chain SPL case, the system produced 770 final files and 104 manifest files across multi-species gene-family stages, supporting manifest-based audit of scoped outputs. A V15.3.4 staged extraction run over five papers produced 795 final-audit relations distributed across main knowledge-graph, background, method and dataset-metadata layers. We further describe a planned and partially implemented validation route in which RNA-seq raw-data processing recommends candidate gene families, one recommended family is selected for bbav5 gene-family analysis, and three published papers are used for methodology extraction and executable reproduction. Together, these results indicate that literature-grounded RAG is most useful when connected to executable, scope-aware and reproducible analysis systems rather than used as a standalone question-answering layer.
中文摘要说明：摘要采用“问题-系统-证据-边界”的结构。这里没有把 RAG 放成独立主角，而是把它定义为知识层；bbav5 是执行层；RNA-seq 是候选家族发现入口；三篇文章方法学抽取和复现是主要验证路线。证据句准确率较低、RNA-seq 复现未完全完成、PlantGeneBench 只作为可选补充，这些都没有夸大。
1. Introduction
1.1 Plant gene-family analysis in agricultural genomics
Gene families provide a natural unit for studying plant genome evolution, molecular function and agronomic traits. Many crop studies begin by identifying all members of a transcription factor, transporter, enzyme or resistance-gene family, and then combine phylogenetic, structural, regulatory and expression evidence to infer functional divergence and candidate genes for downstream experiments [REF: plant gene-family analysis]. A typical plant gene-family study includes homology-based or domain-based gene identification, multiple sequence alignment, phylogenetic tree construction, conserved motif discovery, Pfam or InterPro domain validation, exon-intron structure analysis, chromosome localization, intra- and interspecies synteny analysis, promoter cis-element analysis, expression profiling, protein-protein interaction analysis and evolutionary rate estimation. These steps are scientifically routine, but operationally complex: they rely on different tools, file formats, reference genomes, species identifiers and downstream visualization scripts.
中文说明：这一段先讲农业基因家族分析本身，不先讲 RAG。审稿人首先关心的是系统解决什么生物信息学问题，而不是用了什么 AI 技术。
1.2 RNA-seq turns gene-family analysis from annotation into candidate discovery
Classical gene-family analysis often starts from a preselected family. In functional genomics and breeding contexts, however, researchers frequently begin with a biological condition, tissue, stress treatment or phenotype, and use RNA-seq data to identify genes and pathways that respond to that condition. Connecting RNA-seq analysis to gene-family analysis therefore changes the task from "analyze this known family" to "discover candidate gene families that warrant detailed evolutionary and functional analysis." This requires a workflow that can process raw sequencing reads, perform quality control, alignment, quantification, sample-level quality assessment, differential expression, enrichment, gene set analysis or co-expression analysis, and then summarize candidate families for downstream gene-family workflows. Without this connection, expression evidence remains isolated from the comparative genomic analyses that are often needed to interpret gene-family expansion, diversification and regulation.
中文说明：这一段把 RNA-seq 放成生物学入口。它不是附属模块，而是从原始转录组数据推荐候选基因家族的入口。
1.3 Published methods are rich but difficult to execute reproducibly
Published plant gene-family papers contain valuable procedural knowledge: tool choices, parameter settings, databases, input data, outputs and manual validation steps. Yet these details are distributed across methods text, figure captions, supplementary files and implicit domain conventions. Manual reproduction requires reading the paper, extracting the computational steps, deciding which steps are executable, replacing obsolete tools with supported alternatives, preparing inputs and tracking outputs. This process is especially fragile when a paper combines computational analysis with RNA-seq, qRT-PCR, transient expression, yeast assays or other wet-lab validation. A useful AI system for agricultural bioinformatics should therefore do more than answer questions about a paper; it should convert literature-described methodology into structured observations, align those observations with an executable workflow, and clearly separate executable computational stages from manual or wet-lab stages.
中文说明：这一段是 RAG 进入文章的合理入口：文献方法学复现困难。因此 RAG 的定位是“把文献方法转成可执行流程的知识层”，不是泛问答。
1.4 Current AI and RAG systems remain weakly connected to execution
Recent large language model and retrieval-augmented generation systems have improved scientific question answering, literature summarization and code assistance [REF: scientific RAG; REF: LLM agents for bioinformatics]. However, many systems remain disconnected from audited execution. They may retrieve evidence or generate a recommended command, but they do not necessarily check whether required inputs exist, whether a dependency stage has completed, whether output artifacts match the expected contract, or whether a multi-species analysis is using the correct target-reference scope. Multi-agent systems can coordinate planning, execution and diagnosis, but in scientific workflows they require strong contracts, provenance and status definitions; otherwise an agent message such as "completed" may not correspond to real files, valid outputs or reproducible commands.
中文说明：这一段批判现有 AI/RAG 的边界：缺少真实执行、契约、manifest 和 scope identity。这样自然引出 bbav5。
1.5 Contributions
We present a literature-grounded executable multi-agent system for plant gene-family analysis. The system combines a RAG-based literature knowledge layer from gene-family-mas-v2 with the bbav5 executable analysis layer. The contribution is not a standalone RAG question-answering tool, nor merely a fixed gene-family pipeline. Instead, it is a closed loop that links literature-derived methodology, RNA-seq-guided candidate discovery, executable gene-family analysis, diagnosis and reproducibility.
The main contributions are:
- A literature knowledge layer that extracts method observations, biological SRT relations and evidence from gene-family papers, and aligns extracted methodology with bbav5 executable stages.
- An executable multi-agent analysis layer that plans, validates, runs and diagnoses gene-family and RNA-seq stages using real command-line tools.
- A scope-aware workflow identity model that propagates
scope_id,analysis_route,member_set_idandstage_contextacross planning, execution, manifests, status checks and result display. - A manifest-first reproducibility design that treats stage manifests and artifact paths as the source of truth for completed analysis.
- A validation strategy built around RNA-seq-guided candidate gene-family discovery and three-paper methodology-to-execution reproduction, with optional planning benchmarks and extension-stage proof-of-concept experiments.
中文说明：贡献排序经过降权处理：RAG 重要，但不是唯一主角；bbav5 执行和三篇文章复现是证据重心；extension stage 是扩展性证明，不作为主验证。
2. System overview
2.1 Dual-layer architecture
The system has two coupled layers. The first layer is a literature and knowledge layer implemented in gene-family-mas-v2. It parses gene-family papers, extracts biological and methodological knowledge, stores or retrieves evidence through graph and vector indices, and produces structured methodology observations. The second layer is an executable analysis layer implemented in bbav5. It receives user requests, structured methodology or selected candidate families, generates stage plans, validates inputs, executes analysis scripts, diagnoses failures and records reproducible outputs.
The coupling between the two layers is the key design choice. The RAG layer does not simply generate free-text answers. It produces structured records that can be aligned with the bbav5 stage catalog: normalized stage names, tools, parameters, inputs, outputs, evidence sentences and support status. The execution layer does not simply trust natural language recommendations. It runs preflight checks, chooses stages according to dependencies, executes CLI scripts and validates manifests and artifacts.
中文说明：系统不是“RAG 在前、bbav5 在后”的松散串联，而是结构化知识到可执行 stage 的耦合。
2.2 Workflow from literature or data to executable analysis
The system supports two primary entry points. In the literature-driven entry point, a paper is parsed and its methods are extracted into MethodObservation records. These records are aligned to bbav5 stages and classified as supported by default tools, supported by alternative tools, partially supported, external/manual, or unsupported. The executable subset can then be used to generate a reproduction plan. In the RNA-seq-driven entry point, raw sequencing data can be processed by the RNA-seq subsystem to generate expression, differential expression, enrichment, gene set or co-expression evidence. Candidate gene families are then recommended from these results and passed to the gene-family workflow for detailed analysis after manifest and artifact validation.
Both entry points converge on the same execution infrastructure: stage planning, preflight validation, command execution, manifest generation, artifact validation, diagnosis and reproducibility recording.
中文说明：这里明确两条入口：文献复现入口和 RNA-seq 候选发现入口，最后都进入 bbav5 执行层。
3. RAG-based literature knowledge layer
3.1 Document parsing and relation-rich chunk construction
The latest literature extraction line is represented by the V15.3.4 extractor. It uses MinerU-based PDF parsing and a raw-MinerU-markdown chunking strategy to preserve section boundaries before local cleanup. This design responds to a common problem in PDF-based scientific extraction: global text cleaning can flatten headings, captions and tables before the extraction model has a chance to identify major sections. By applying cleanup inside chunk paths rather than before section detection, the extractor preserves more of the document structure needed for method and relation extraction.
中文说明：这一段可以对应 src/rag/extractors/v15_3_4 的设计。注意不声称 MinerU 解析完美，只说它是当前主线解析器。
3.2 Source-relation-target extraction
The extractor represents biological and methodological knowledge as source-relation-target triples. The LLM-led extractor is designed for high recall: it extracts explicit molecular, genomic, expression, localization, synteny, motif, domain, promoter, dataset and method relations from text windows. Rule-based and PC-pattern-based components provide fallback, typed relation patterns and additional filtering. This hybrid design is useful for plant gene-family papers because the same fact may appear in prose, supplementary tables, figure captions or dataset-accession contexts.
The extracted relations are not immediately treated as final knowledge. They pass through canonicalization and semantic review steps that repair low-quality endpoints, normalize relation names, separate candidate relations from main knowledge-graph relations, and classify output layers such as main_kg, background_relations, method_relations and dataset_metadata.
中文说明：RAG 内容要写具体，不只是“我们用了 RAG”。这里强调 SRT、规则 fallback、PC pattern 和多输出层。
3.3 Evidence fidelity and semantic review
A central design principle of the V15.3.4 line is evidence fidelity. Earlier designs explored additional LLM-based evidence-local verification and minimization. In the current V15.3.4 configuration, the LLM-2 evidence-local verifier/minimizer is intentionally disabled, and the system preserves the evidence selected by the first extraction and by Phase 4.5/4.6 review. This is a conservative choice: it reduces the risk that a second model rewrites or over-compresses evidence in a way that weakens traceability. Audit files record relation repair, semantic review, evidence minimization decisions, supplementary table parsing and dataset accession context.
中文说明：这是一个很好的审稿点。我们不是盲目多加 LLM，而是为了证据可追溯关闭某个容易改写证据的步骤。
3.4 Dual knowledge base retrieval
The RAG layer can retrieve from two knowledge bases. The V15 knowledge base contains extracted SRTs from gene-family literature. The PC knowledge base contributes PlantConnectome-derived relation patterns and a broader biological relation substrate. The retrieval service combines Neo4j graph search and Qdrant vector search, with agentic query expansion, multi-hop graph exploration and reranking. The combined QA ablation design includes variants that remove agentic query expansion, multi-hop retrieval or reranking, allowing the contribution of these retrieval mechanisms to be tested.
中文说明：这一节写双知识库检索，但不要把全文写成 QA 论文。QA 是支撑 evidence retrieval 和方法学知识层的能力。
3.5 Methodology extraction and bbav5 alignment
For method reproduction, the system extracts normalized stages, tools, versions, parameters, inputs, outputs and evidence sentences from method sections. The methodology schema includes MethodObservation, ToolUsage, MethodEvidenceRef and MethodAlignment. The aligner maps normalized literature stages such as gene-family identification, phylogenetic analysis, motif analysis, synteny, RNA-seq quantification, differential expression, enrichment analysis, protein properties and PPI to bbav5 executable stages. It also separates wet-lab or manual validation stages, such as qRT-PCR, yeast assays, luciferase assays or transformation experiments, from computational stages that bbav5 can execute.
This alignment produces a support status for each extracted observation: supported_default, supported_alternative, partially_supported, external_or_manual or unsupported. This support status is important because reproducibility should not imply tool-identical reproduction when an original paper used obsolete or unsupported software. Instead, the system records whether bbav5 can execute the same analytical goal using its default or alternative tools.
中文说明：这一节是 RAG 和 bbav5 的接口。它解释为什么 RAG 结果能进入执行系统，而不是只变成回答文本。
4. bbav5 executable multi-agent analysis layer
4.1 Multi-agent execution workflow
bbav5 is an execution-oriented multi-agent system. Its LangGraph workflow includes routing, intent parsing, planning, preflight validation, execution, result parsing, validation, diagnosis and discovery/knowledge steps. The PlanningAgentV2 generates stage plans while considering completed stages, dependencies, stage capabilities, learning-based parameter recommendations and preflight input checks. The ExecutionAgent invokes stage tools through a registry that ultimately calls bbav5 CLI scripts or skill wrappers. The DiagnosisAgent classifies failures and proposes repair actions when a stage fails or when outputs are incomplete.
中文说明：bbav5 是“执行型 Agent”，这点要反复强调。不是聊天机器人调几个脚本，而是有 planning、preflight、execution、diagnosis、manifest status 的执行系统。
4.2 Built-in gene-family stages
The built-in gene-family workflow covers the main computational stages used in plant gene-family studies. These include gene-family identification, phylogenetic reconstruction, conserved motif analysis, Pfam/domain analysis, gene-structure analysis, promoter analysis, intra- and interspecies synteny analysis, karyotype visualization, combined tree-motif-domain-structure visualization, protein-protein interaction analysis, expression visualization, physicochemical protein analysis, ortholog detection, Ka/Ks analysis and summary generation. Each stage is implemented as a command-line entry point or skill wrapper, and successful execution produces a stage manifest that records artifacts and metrics.
中文说明：这里要把常规 gene-family stages 讲清楚，因为这是 bbav5 的生物信息学核心。
4.3 RNA-seq subsystem
bbav5 also includes an RNA-seq subsystem, which is essential for the proposed RNA-seq-guided candidate discovery workflow. The RNA-seq stages include quality control, alignment, quantification, sample-level quality assessment, differential expression analysis, GO/KEGG enrichment, fgsea, WGCNA and candidate-family recommendation. This subsystem is designed to process raw FASTQ files into expression matrices, differential gene sets, enriched pathways, gene sets or co-expression modules. After validation, these results can be summarized into candidate genes or candidate gene families for downstream gene-family analysis.
In the current manuscript plan, the RNA-seq subsystem is treated as a core capability rather than an auxiliary feature. However, RNA-seq-based reproduction of published papers should only be claimed where raw data, sample metadata, reference genome files, stage manifests and output artifacts have been verified. If some RNA-seq reproduction experiments remain untested, they should be reported as planned or pending validation rather than completed results.
中文说明：这一节回应你补充的重点。RNA-seq 是核心，但文章里必须区分“系统支持”和“已经完成真实复现”。
4.4 Scope-aware workflow identity
Plant gene-family workflows are vulnerable to identity errors. A target species, a reference species, a comparative species and a selected member set may share a family name but represent different biological scopes. bbav5 addresses this by propagating explicit workflow identity through scope_id, scope_mode, analysis_route, member_set_id and stage_context. These fields are used during planning, execution, manifest writing, status checks and result display. For example, target-only stages such as promoter, synteny, karyotype and protein-property analysis should not silently consume reference-combined outputs, whereas tree-structure stages such as phylogeny, motif, domain, structure and combo may operate under target-plus-reference scopes.
中文说明：这是 bbav5 很强的系统创新点。它解决多物种、多参考、亚家族确认时结果错配的问题。
4.5 Manifest-first status and reproducibility
The system treats manifests and artifact validation as the final source of truth for stage completion. Runtime messages from Celery or LangGraph are scheduling states; a stage is considered completed only when its manifest reports a success or partial-success status and the referenced artifacts resolve. The reproducibility manager records commands, environment information, file checksums and provenance. A typical reproducibility bundle contains commands.sh, environment.yml, checksums.sha256, progress.json and provenance.json.
中文说明：这一节要写得硬一些。多智能体系统容易“嘴上完成”，bbav5 用 manifest-first 避免这个问题。
4.6 Diagnosis and repair
The DiagnosisAgent contains stage-specific and issue-specific rules for common plant gene-family and RNA-seq failures. Examples include missing HMM/Pfam profiles, remote Pfam fetch failures, protein FASTA stop markers, genome-prepare content-type errors, low GFF-gene ID overlap, missing manifests, ortholog intermediate outputs not materialized, RNA-seq foreground/background ID mismatch, enrichment background absence and scope-identity contract mismatch. Repair policies are separated into automatic repair, user confirmation, report-only diagnosis and developer-fix categories.
中文说明：诊断模块不是泛泛的 error summary，而是针对生信流程常见失败设计的。
4.7 Extension stages as an extensibility proof of concept
In addition to built-in gene-family and RNA-seq stages, bbav5 contains an extension-stage mechanism. It can take already runnable user scripts and associated test data, convert them into controlled command templates through an API converter, bind inputs and parameters, execute validation runs, scan outputs and apply publication gates before storing the stage in the extension-stage store. This mechanism demonstrates that bbav5 is not limited to predefined workflows. However, unless an extension stage is included in the main RNA-seq or three-paper reproduction experiments, it should be treated as an engineering extensibility proof of concept rather than a primary biological validation result.
中文说明：这是实事求是的写法。extension stage 有真实代码支撑，但如果没参与主实验，不应拔高成主贡献第一梯队。
5. Experimental design
5.1 RNA-seq-to-candidate-family discovery
The first validation route starts from raw RNA-seq data. The intended workflow is: quality control of raw FASTQ files, read alignment to the reference genome, gene-level quantification, sample-level quality assessment, differential expression analysis, enrichment or gene set analysis, co-expression analysis where appropriate, and candidate-family recommendation. The output should include candidate genes, candidate gene families, supporting expression or enrichment evidence, and the corresponding manifests and reproducibility records.
中文说明：这一节是实验设计。正式投稿前最好补完整 FASTQ 到候选家族的真实 manifest。
5.2 Downstream gene-family analysis of a selected candidate family
After candidate-family recommendation, one recommended family is selected for downstream bbav5 gene-family analysis. The selected family should be processed through the executable gene-family workflow, including identification, phylogeny, motif, domain, structure, promoter, synteny or interspecies synteny, karyotype, expression integration, PPI, Ka/Ks where applicable and summary generation. The key evidence is not only the biological output, but also the stage manifests, artifact checks and reproducibility bundle.
中文说明：这会把 RNA-seq 的发现入口和基因家族分析主流程连接起来。
5.3 Three-paper methodology extraction and executable reproduction
The second validation route uses three published plant gene-family or RNA-seq-related papers. For each paper, the RAG layer extracts methods, tools, parameters, inputs, outputs and evidence sentences. The aligner maps extracted stages to bbav5 executable stages and assigns support status. The executable subset is then reproduced using bbav5 when required inputs are available. Wet-lab validation and unsupported manual procedures are retained as literature evidence and excluded from automatic execution.
中文说明：这是比 PlantGeneBench 更贴近 AIIA 审稿人的主验证。三篇文章要尽量补齐人工核对表。
5.4 Optional planning benchmark
PlantGeneBench can be used as a supplementary planning benchmark. It evaluates whether an agent maps plant gene-family requests to valid stages, dependencies, parameters and safety boundaries. Because the planned manuscript already includes three-paper methodology-to-execution reproduction, PlantGeneBench should not replace the real reproduction experiments. It is best used as supplementary evidence for planning robustness if time permits.
中文说明：PlantGeneBench 不应压过三篇文章复现。如果三篇文章复现扎实，可以不把 PlantGeneBench 放主文。
6. Results
6.1 Methodology extraction identifies executable stages from literature
In a curated methodology evaluation, the extraction and alignment pipeline showed strong stage-level and alignment performance. After filtering excluded stages, the benchmark contained 83 gold stage observations and 79 system stage observations. Stage extraction achieved 77 true positives, 2 false positives and 6 false negatives, corresponding to a precision of 0.975, recall of 0.928 and F1 score of 0.951. Tool extraction was more difficult, with a precision of 0.829, recall of 0.759 and F1 score of 0.792. Parameter extraction achieved an accuracy of 0.825. Alignment to bbav5 executable stages achieved an accuracy of 0.960.
Evidence extraction remained the weakest component. Evidence accuracy was 0.638, with 36 correct, 25 partial, 14 wrong and 1 missing evidence entries. This result indicates that the system can reliably identify stages and map them to executable workflows, but evidence-sentence localization still requires improvement and human review for high-stakes reproduction claims.
中文说明：这里使用已有 gold3 filtered 数字。证据准确率低不能隐藏，应该作为局限和后续改进点。
6.2 The V15.3.4 knowledge layer produces auditable relation layers
A V15.3.4 staged extraction run over five papers completed without recorded extraction errors and produced 795 final-audit relations. These relations were distributed across output layers: 442 main_kg relations, 178 background_relations, 91 method_relations and 84 dataset_metadata relations. This layered representation is useful for downstream use because experimental findings, prior knowledge, method statements and dataset metadata should not be mixed into a single undifferentiated relation set.
中文说明：这个结果是 RAG 知识层的量化支撑，但样本是 5 篇 staged run，不要写成大规模最终 benchmark。
6.3 bbav5 supports real multi-species gene-family execution with manifest audit
A real full-chain SPL case, job_1779966938704_aa384a, was recorded as a regression and validation baseline. The workspace contained 770 files in the final output directory and 104 manifest files in the baseline snapshot. The completed outputs covered genome preparation, configuration autofill, per-species identification and refinement, target-plus-reference phylogeny, motif, structure, domain, combo visualization, reference classification and subfamily evidence outputs. A later manifest/artifact audit found that inspected artifact paths resolved correctly when full_path was used first and relative paths were resolved against the job workspace.
The case also revealed realistic engineering constraints. Some compatibility manifests contained legacy naming patterns, and some stages required additional scope identity enrichment. These findings support the need for manifest-first status checks and compatibility-aware alias resolution rather than filename-only result discovery.
中文说明：这里既写成绩，也写暴露的问题。这样比只说“全流程跑通”更可信。
6.4 RNA-seq workflow provides the route to candidate gene-family recommendation
The RNA-seq subsystem is designed to process raw sequencing data into candidate gene-family recommendations through quality control, alignment, quantification, sample quality assessment, differential expression, enrichment, fgsea, WGCNA and candidate-family summarization. This workflow is the biological entry point for moving from condition-specific transcriptomic evidence to gene-family analysis.
At the current manuscript stage, this section should be completed with a concrete RNA-seq validation case. The final version should report the raw-data source, number of samples, reference genome and annotation, completed RNA-seq stages, differential expression or enrichment outputs, candidate families recommended by the system, and the selected family used for downstream gene-family analysis. Until those stage manifests and artifacts are verified, claims should be limited to implementation and planned validation rather than completed RNA-seq reproduction.
中文说明：这里先写成待补实验证据。如果你已经跑出结果，后续把这一节替换成真实结果表即可。
6.5 A selected candidate family can be analyzed by the executable gene-family workflow
Once a candidate family is selected from RNA-seq evidence, bbav5 can run the standard gene-family workflow to generate family members, phylogenetic trees, conserved motifs, Pfam/domain profiles, gene structures, promoter sequences and cis-elements, syntenic relationships, chromosome localization, expression visualizations and other downstream outputs. The important validation criterion is that each stage produces a manifest and that its artifacts resolve without cross-scope mismatch.
For the final manuscript, this section should use the selected candidate family from the RNA-seq validation case and report the actual stages completed, output files, major biological observations and manifest/provenance records. If the selected family is still under analysis, this section should remain a planned result rather than a completed claim.
中文说明：这节现在是结果模板。正式稿需要填入你选择的候选家族。
6.6 Three-paper reproduction links literature methods to executable workflows
The three-paper reproduction experiment is the main validation route for the literature-to-execution claim. For each paper, the system should report the number of extracted methodology observations, the number mapped to bbav5 stages, support-status distribution, executable stages, unsupported or manual stages, and the actual reproduction status. Computational stages such as identification, phylogeny, motif, domain, structure, synteny, enrichment or RNA-seq analysis can be reproduced if inputs are available. Wet-lab steps such as qRT-PCR, transient expression, yeast assays or luciferase reporter assays should be retained as literature evidence but excluded from automatic execution.
This design is stronger than a synthetic planning benchmark because it tests the full chain: paper parsing, method extraction, stage alignment, workflow execution and manifest audit. If RNA-seq reproduction from the three papers remains untested, the manuscript should explicitly distinguish gene-family stages that have been reproduced from RNA-seq stages that have only been aligned or planned.
中文说明：这一节是主验证，应优先补真实三篇文章表格。PlantGeneBench 只能补 planning，不替代复现。
6.7 Optional PlantGeneBench planning evaluation
PlantGeneBench-100 can be reported as supplementary planning evidence. In the existing pilot summary, the live API provider achieved a mean machine score of 0.9112 over 100 tasks, with layer-level mean scores of 0.9000 for biological interpretation, 0.9000 for boundary control, 0.9000 for evidence constraint, 0.9419 for execution planning and 0.8932 for task understanding. However, the same run recorded 45 provider failure flags. Therefore, these numbers should be treated as preliminary planning robustness results, not as the primary evidence for the manuscript.
中文说明：实事求是：可以写，但放补充，不压主线。
6.8 Extension-stage mechanism demonstrates engineering extensibility
The extension-stage mechanism provides a path for adding project-specific analyses beyond the built-in gene-family and RNA-seq stages. Existing extension-stage directories include examples such as protein sequence extraction, BLASTP protein search and HMM-search tests. The mechanism supports uploaded code, API-based conversion, input binding, validation runs, output scanning and publish checks. Unless a validated extension stage is used in the main RNA-seq or three-paper reproduction experiments, this result should be presented as proof-of-concept engineering extensibility rather than as a primary biological validation.
中文说明：拓展阶段有必要写，但降权。它是系统未来扩展性，不是当前主结果核心。
7. Discussion
7.1 RAG is most useful when connected to execution
The results support a practical view of RAG in agricultural bioinformatics. A standalone literature QA system can help researchers read papers, but it does not by itself produce reproducible analysis. In this framework, RAG is used as a structured knowledge layer: it extracts evidence, methods, tools and parameters, and passes them to an executable stage system. This design allows literature evidence to influence analysis planning while still requiring execution, manifests and artifact validation before a workflow is considered completed.
中文说明：这是对 RAG 定位的最终回答：RAG 是核心模块，但不是独立主角；它的价值在于连接执行。
7.2 RNA-seq-guided discovery expands the biological scope of gene-family analysis
Integrating RNA-seq analysis changes the system's biological role. Rather than only analyzing a user-specified family, the system can begin from transcriptomic evidence and recommend families for downstream analysis. This is important for agricultural studies where the starting point is often a stress treatment, tissue, developmental stage or phenotype. Candidate-family recommendation can prioritize families that are differentially expressed, enriched in relevant pathways, associated with co-expression modules or supported by gene-set evidence.
中文说明：这一段让文章从“流程平台”变成“发现候选基因家族的农业 AI 系统”。
7.3 Manifest-first reproducibility is essential for trustworthy multi-agent workflows
Multi-agent systems introduce flexibility but also increase the risk of opaque or inconsistent state. bbav5 addresses this by separating scheduling state from scientific completion state. A stage is not considered complete merely because an agent reported success; it must produce a manifest and resolvable artifacts. Scope-aware identity fields further prevent results from one species combination or member set from being reused in another. This design is especially important in multi-species gene-family analysis, where target-only and target-plus-reference outputs may share similar filenames but represent different biological scopes.
中文说明：这是系统可信性的工程核心。
7.4 Limitations
Several limitations remain. First, evidence-sentence localization in the methodology benchmark was weaker than stage or alignment extraction, indicating that human review remains necessary when exact evidence support is required. Second, the RNA-seq reproduction component should be fully validated with raw-data cases before strong claims are made about end-to-end transcriptome-driven discovery. Third, three-paper reproduction provides realistic evidence but remains a small sample; broader evaluation across gene-family types, crop species and publication styles is needed. Fourth, parts of the RAG layer depend on external LLM or API services, which may affect reproducibility unless model versions and prompts are recorded. Fifth, wet-lab validation steps cannot be automatically reproduced by bbav5 and should remain explicitly marked as manual or external. Finally, extension stages demonstrate useful engineering flexibility, but require additional systematic validation before being treated as a mature plugin ecosystem.
中文说明：局限要诚实。AIIA 审稿人通常会接受边界清楚的系统论文，但会反感过度声称。
7.5 Future work
Future work should complete the RNA-seq raw-data validation route, expand the three-paper reproduction experiment to a larger curated corpus, improve evidence-sentence fidelity, and add stronger human-reviewed benchmarks for methodology extraction and workflow alignment. The extension-stage mechanism should be evaluated with real user-contributed analyses, including cases where an original paper uses a method not covered by the built-in stage catalog. Containerized execution environments and public reproducibility bundles would further improve portability beyond the local deployment.
中文说明：未来工作围绕真实验证、证据质量、扩展阶段和环境可移植性。
8. Conclusion
We present a literature-grounded executable multi-agent system that connects RAG-based literature understanding with reproducible plant gene-family and RNA-seq workflows. The system uses RAG to extract methodology and evidence from papers, aligns extracted observations to bbav5 executable stages, provides RNA-seq infrastructure for candidate gene-family discovery, and executes gene-family workflows with scope-aware manifests and reproducibility bundles. The current evidence indicates strong stage extraction and workflow alignment, real multi-species gene-family execution with manifest audit, and a viable design for RNA-seq-guided candidate discovery and three-paper methodology reproduction. The broader implication is that RAG becomes more valuable for agricultural AI when it is embedded in executable, auditable and biologically grounded workflow systems.
中文结论说明：结论没有说“完全自动复现所有文章”或“完全解决基因家族分析”。它强调当前证据能支持的结论：RAG + 可执行 MAS + RNA-seq + manifest reproducibility 的闭环。
Proposed figures
Figure 1. Overall architecture of the literature-grounded executable MAS
Show input sources, the RAG knowledge layer, the bbav5 execution layer, and output artifacts. The left side should include papers, supplementary files and RNA-seq raw data. The middle should include V15.3.4 extraction, methodology observations, Neo4j/Qdrant retrieval and stage alignment. The right side should include planning, preflight, execution, diagnosis, manifest and reproducibility.
Figure 2. RAG-to-bbav5 methodology extraction and alignment workflow
Show PDF parsing, chunking, SRT extraction, methodology extraction, evidence review, MethodObservation, MethodAlignment and support status categories.
Figure 3. bbav5 executable multi-agent workflow
Show Router, Intent, PlanningAgentV2, PreFlight, ExecutionAgent, DiagnosisAgent, ResultParser and Discovery/Knowledge components.
Figure 4. RNA-seq-to-candidate-family discovery workflow
Show FASTQ, QC, alignment, quantification, sample QC, differential expression, enrichment/fgsea/WGCNA, candidate genes, candidate gene families and downstream gene-family analysis.
Figure 5. Scope-aware manifest and reproducibility model
Show how scope_id, analysis_route, member_set_id, stage_context, manifest and artifact paths connect across stages.
Figure 6. Three-paper methodology reproduction design
Show three papers, extracted methods, bbav5 alignment, executable stages, manual stages and reproduction status.
Supplementary Figure. Extension-stage generation workflow
Show uploaded runnable code, API conversion, command template, input binding, validation run, output scan, publish gate and registered extension stage.
Proposed tables
Table 1. bbav5 built-in and extension stage categories
Columns: stage category, stage IDs, biological purpose, input, output, environment, manifest support, validation status.
Table 2. Three-paper methodology extraction and bbav5 alignment
Columns: paper, extracted observations, mapped bbav5 stages, supported_default, supported_alternative, external/manual, unsupported, executable stages reproduced, pending stages.
Table 3. RNA-seq candidate-family recommendation outputs
Columns: RNA-seq stage, input, output, metric, candidate-family evidence, manifest path, validation status.
Table 4. Reproduction status by paper and stage
Columns: paper, literature stage, bbav5 stage, original tool, bbav5 tool, support status, execution status, manifest, notes.
Table 5. RAG/V15.3.4 extraction statistics
Columns: run, paper count, relation count, main_kg, background_relations, method_relations, dataset_metadata, audit files, limitations.
Supplementary Table. PlantGeneBench planning benchmark
Columns: provider, n, mean score, layer scores, provider failures, top error categories, interpretation.
Evidence checklist before submission
- Complete or document one RNA-seq raw-data workflow through candidate-family recommendation.
- Complete one selected candidate-family bbav5 gene-family analysis with manifests and final artifacts.
- Prepare three-paper methodology extraction outputs, manual review results and bbav5 alignment summaries.
- Record which stages from each paper were actually reproduced and which remained manual, unsupported or pending.
- Decide whether PlantGeneBench remains supplementary or is omitted.
- Add a citation pass to replace
[REF: ...]placeholders with verified references.