Research Recipes

Research Recipes

Complete research workflows that ScienceClaw executes autonomously. Each Recipe defines a multi-step pipeline from query to report. When a user's prompt matches a Recipe trigger, execute the full pipeline without asking for confirmation.

When to Use

• User asks to "分析 X 在 Y 中的作用" → gene-landscape
• User asks to "评估 X 的成药性" → target-validation
• User asks to "综述 X" or "survey X" → literature-review
• User provides expression data for analysis → diff-expression
• User asks about treatment options → clinical-query
• User asks to profile a researcher → person-research
• User asks "X 的新适应症" or "drug repurposing" → drug-repurposing
• User asks "跑个 MD 模拟" or "molecular dynamics" → molecular-dynamics

Recipe Execution Rules

1. Detect the matching Recipe from trigger patterns
2. Create a project directory: ~/.scienceclaw/workspace/projects/<slug>-<YYYY-MM-DD>/
3. Execute each step, reporting substantive progress (numbers, not filler)
4. Generate METHODS.md at the end
5. List all output files and offer export options
6. Suggest 2-3 scientifically-motivated follow-ups

Recipe 1: gene-landscape

Trigger: "分析 X 在 Y 中的作用", "investigate X in Y", "X 在 Y 中的角色/功能"

Steps

Step 1 — Literature Search (PubMed + OpenAlex + Semantic Scholar)

bash: echo "=== PubMed ===" && \
curl -s "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=pubmed&retmode=json&retmax=30&sort=relevance&term=GENE+AND+DISEASE" && \
echo -e "\n=== OpenAlex ===" && \
curl -s "https://api.openalex.org/works?search=GENE+DISEASE&per_page=15&sort=relevance_score:desc&select=id,title,authorships,publication_year,cited_by_count,doi" && \
echo -e "\n=== Semantic Scholar ===" && \
curl -s "https://api.semanticscholar.org/graph/v1/paper/search?query=GENE+DISEASE&limit=15&fields=title,authors,year,abstract,citationCount,externalIds"

Fetch abstracts for top 10 PMIDs. Report: "PubMed 检索到 N 篇，OpenAlex N 篇，S2 N 篇，获取前 10 篇摘要..."

Step 2 — Pan-cancer Expression (TCGA via cBioPortal)

Query TCGA expression data across cancer types. Use cBioPortal REST API or direct TCGA data. Generate a boxplot of expression across cancer types.

Report: "TCGA 数据获取完成，GENE 在 N/33 癌种中显著上调..."

Step 3 — Survival Analysis

For cancer types with significant expression changes, run Kaplan-Meier survival analysis:

library(survival); library(survminer)
# Split by median/optimal cutoff
# KM plot + log-rank p-value + Cox HR with 95% CI

Report: "STAD 生存分析完成：HR=X.XX (95%CI: X.XX-X.XX, p=X.XXX)..."

Step 4 — Immune Infiltration Correlation

Query TIMER2.0 or compute from TCGA data. Correlate gene expression with immune cell infiltration scores (M1/M2 macrophages, CD8+ T cells, Tregs, etc.).

Report: "免疫浸润分析完成，与 M2 巨噬细胞最强相关 (r=X.XX, p<X.XXX)..."

Step 5 — Pathway Enrichment

Use co-expressed genes (top 200 by Pearson correlation) for GO/KEGG enrichment via Enrichr or local R (clusterProfiler).

Report: "通路富集分析完成，N 条 GO terms，N 条 KEGG pathways 显著..."

Step 6 — Report Assembly

Compile all findings into reports/GENE_DISEASE_report.md:

• Executive summary (key findings in 3-5 bullets)
• Expression profiling results (with boxplot figure)
• Survival analysis results (with KM figure)
• Immune microenvironment analysis
• Pathway enrichment results (with dot plot)
• Limitations and caveats
• References (GB/T 7714 format from actual search results)

Generate METHODS.md. List all files. Offer export options. Suggest follow-ups.

Output

reports/GENE_DISEASE_report.md + METHODS.md + figures/ (boxplot, KM curve, enrichment dot plot)

Recipe 2: target-validation

Trigger: "评估 X 的成药性", "X 能不能做靶点", "druggability of X", "target validation for X"

Steps

1. Literature search — PubMed + OpenAlex for target + druggability/therapeutic
2. Protein info — UniProt (function, domains, subcellular location) + AlphaFold (structure availability)
3. Protein interactions — STRING network (top 10 interactors, confidence > 0.7)
4. Existing compounds — ChEMBL (bioactivity data, IC50/EC50) + PubChem
5. Drug info — DrugBank (approved/investigational drugs targeting this protein)
6. Clinical trials — ClinicalTrials.gov (interventional trials mentioning the target)
7. Patent landscape — Quick search via web_search for patents
8. Report — Druggability assessment with evidence grading (Strong/Moderate/Weak/None)

Output

reports/GENE_target_validation_report.md with sections for each step, evidence summary table, and overall druggability verdict.

Recipe 3: literature-review (Parallel Chapter Architecture)

Trigger: "综述 X", "survey X", "X 的研究进展", "review the literature on X"

Overview

Uses an outline-first, chapter-by-chapter deep research approach inspired by DeepResearchAgent. Instead of a linear pipeline, the review is structured as: broad scan → outline → per-chapter deep dive → cross-chapter synthesis.

Steps

Step 1 — Broad Scan & Outline Generation

Search all sources (PubMed 50 + OpenAlex 30 + Semantic Scholar 30 + Asta 15 + bioRxiv/arXiv if relevant). Deduplicate by DOI > PMID > title. Read all abstracts. Cluster papers into 4–6 themes. Generate a structured outline:

## Outline: [TOPIC] Literature Review

1. Introduction and Historical Context
2. [Theme A]: [Description] (N papers)
3. [Theme B]: [Description] (N papers)
4. [Theme C]: [Description] (N papers)
5. Current Challenges and Controversies
6. Future Directions and Emerging Opportunities

Report: "广搜完成：PubMed N 篇，OpenAlex N 篇，S2 N 篇，Asta N 篇。去重后 N 篇。聚类为 N 个主题，生成了 outline。"

Step 2 — Per-Chapter Deep Dive

For each chapter in the outline, independently:

• Run chapter-specific keyword searches (narrower, more targeted than Step 1)
• Trace citation chains from the most relevant papers in that theme
• Read full text for the top 2–3 papers per chapter (via Jina Reader or Europe PMC)
• Extract key findings, methodological approaches, contradictions, and gaps
• Write the chapter section progressively to the report file

Report per chapter: "第 N 章「[Theme]」深度搜索完成：新增 N 篇文献，阅读 N 篇全文。"

Step 3 — Cross-Chapter Synthesis

After all chapters are written:

• Check for contradictions or redundancies between chapters
• Ensure consistent citation numbering across the entire review
• Write Executive Summary (key findings in 5–8 bullets with numbers)
• Write "Research Gaps and Open Questions" section
• Write "Future Directions" section grounded in specific data gaps identified

Step 4 — Publication Trend & Figures

• Year-by-year publication count chart (matplotlib bar chart)
• Optional: citation network visualization for top papers
• Optional: keyword co-occurrence heatmap

Step 5 — References & Finalization

• GB/T 7714 format, numbered by order of appearance
• Verify every citation against tool results (zero fabrication)
• Generate METHODS.md

Output

reports/literature_review_TOPIC.md (structured review, 4000–8000 words) + figures/publication_trend_TOPIC.png + METHODS.md

Recipe 4: diff-expression

Trigger: "分析这个表达矩阵", "差异表达分析", "DEG analysis", "differential expression"

Steps

1. Read data — Load the user's expression matrix (CSV/TSV/Excel)
2. QC — Sample distribution, missing values, normalization check
3. Differential expression — DESeq2 (count data) or limma (normalized), with fold change and adjusted p-value
4. Volcano plot — Highlight significantly up/down-regulated genes (NPG palette)
5. Heatmap — Top 50 DEGs, clustered by sample and gene
6. GO/KEGG enrichment — Up-regulated and down-regulated gene sets separately
7. Report — Summary of DEG counts, top genes, enriched pathways

Output

figures/volcano_plot.png, figures/heatmap_top50_deg.png, figures/go_enrichment.png, data/deg_results.csv, reports/deg_analysis_report.md

Recipe 5: clinical-query

Trigger: "X 的最新治疗方案", "X 怎么治", "treatment for X", "X 的临床指南"

Steps

1. Clinical trials — ClinicalTrials.gov API (interventional, Phase 2-4, recruiting/completed)
2. Treatment guidelines — PubMed search for "X treatment guidelines" OR "X clinical practice guideline", recent 3 years
3. Drug options — DrugBank search for approved drugs for the disease
4. Summary table — Drug name, mechanism, approval status, trial phase, key efficacy data

Output

Direct response in chat (this is a quick-to-medium task). If results are extensive, create a project and report.

Recipe 6: person-research

Trigger: "调研 X 教授", "X 的学术背景", "profile of Dr. X", "X 发了哪些文章"

Steps

1. Author search — OpenAlex author API (name match, affiliation, h-index, works count, cited_by_count)
2. Publication list — PubMed author search, top 10 by citation
3. Citation metrics — Total citations, h-index, i10-index (from OpenAlex)
4. Research themes — Extract top keywords/concepts from their works
5. Report — Academic profile with metrics, top papers, research focus areas

Output

reports/researcher_profile_NAME.md

Recipe 7: drug-repurposing

Trigger: "帮我找 X 的新适应症", "X 能不能重定位到 Y", "drug repurposing for X", "X 的老药新用", "repurpose X"

Steps

1. Drug profile — DrugBank (approved indications, mechanism, targets) + ChEMBL (bioactivity data, off-target hits) + PubChem (chemical properties, similar compounds)
2. Target network analysis — STRING/OpenTargets: map drug targets to PPI network. Identify "unexpected targets" → map to new diseases via OpenTargets disease associations
3. Clinical evidence mining — ClinicalTrials.gov: ALL trials for this drug (including failed/terminated ones for off-label indications) + PubMed: case reports, retrospective studies, off-label use reports
4. Patent landscape — web_search: "[drug] patent expiry", "[drug] generic availability". Patent expiry = higher repurposing viability
5. Safety profile — OpenFDA adverse event reports. Note: some "adverse effects" may indicate therapeutic potential (e.g., metformin weight loss → obesity, metformin anti-cancer signals)
6. Repurposing candidate ranking — Score each candidate indication by: target evidence (0-30) × clinical evidence (0-30) × safety (0-20) × patent status (0-20). Output ranked table with evidence summaries

For detailed pipeline implementation, refer to the drug-repurposing skill.

Output

reports/drug_repurposing_DRUGNAME.md (ranked candidates with evidence) + figures/target_network_DRUGNAME.png

Recipe 8: molecular-dynamics

Trigger: "跑个分子动力学模拟", "molecular dynamics for X", "binding free energy of X", "蛋白配体模拟", "MD simulation"

Steps

1. Structure retrieval — Fetch from PDB or AlphaFold by protein ID/gene name
2. System preparation — Force field assignment, solvation, ionization via OpenMM/AmberTools
3. Energy minimization — Minimize with self-correction (verify potential energy convergence)
4. Equilibration — NVT then NPT ensemble equilibration
5. Production run — Record trajectory (length depends on system size, default 10 ns)
6. Analysis — RMSD, RMSF, hydrogen bonds, radius of gyration, binding free energy (MM/PBSA if ligand present)
7. Report — Figures + structural insights + stability assessment

Note: Requires OpenMM or GROMACS to be installed. Check via scienceclaw doctor. If not available, report and suggest installation.

For detailed pipeline implementation, refer to the molecular-dynamics skill.

Output

reports/md_simulation_PROTEIN.md + figures/rmsd_*.png + figures/rmsf_*.png + data/trajectory_*.dcd