菜单
Generate an interactive UMAP scatter plot of CultureMech CHEBI ingredients in KG-Microbe embedding space — each point is one unique ingredient, sized by frequency and colored by occurrence tier
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use this skill to quality-check CultureMech growth-media and solution recipes — verify schema/structure, MediaIngredientMech ingredient linkages and CHEBI mappings, solution references, data quality (placeholders, units, concentrations), and categorization. Use after creating/editing a recipe, before a KG export, or for periodic maintenance. Issues are graded P1 (blocking) → P4 (optional); P3/P4 can be auto-fixed.
Use this skill to manage CultureMech record identifiers — find the highest existing CultureMech id, mint the next id, and insert new media/solution records with correct id placement. Use when adding or importing recipe records, or reconciling id collisions.
Use this skill to integrate the MediaIngredientMech ingredient variant hierarchy (parent/child relationships — purity levels, hydrates, stereoisomers) into CultureMech recipe ingredient references. Use when linking recipe ingredients to their canonical MIM records or resolving variant forms.
Validate that every ontology ID carries its correct ontology LABEL across CultureMech recipe YAMLs and SSSOM mapping products. Uses LinkML schema bindings + linkml-term-validator for term.{id,label} (canonical label) and a shared OAK validator for products. Run report-then-enforce; triage drift as wrong-label vs wrong-id.
Scrape growth-medium recipes directly from the RIKEN JCM GRMD database and emit CultureMech-normalized YAML records. Use to close the gap between CultureMech's JCM coverage (originally ingested second-hand via MediaDive/TOGO) and the live JCM catalogue — e.g. when newly added high-GRMD media are missing, or to audit JCM completeness.
Run Edison Scientific deep research for a CultureMech medium record, then for each candidate organism returned by phase 1 run a per-organism follow-up that extracts the recipe, culture conditions, identifiers, and citation metadata from the primary publication. Produces a curation-focused report with proposed YAML changes.
| name | generate-ingredient-umap |
| description | Generate an interactive UMAP scatter plot of CultureMech CHEBI ingredients in KG-Microbe embedding space — each point is one unique ingredient, sized by frequency and colored by occurrence tier |
| category | visualization |
| requires_database | false |
| requires_internet | false |
| version | 1.0.0 |
| tags | ["umap","visualization","ingredients","chebi","embeddings","kg-microbe","d3js","html"] |
Produces an interactive HTML visualization (app/ingredient_umap.html) where each point
is a unique CHEBI ingredient observed across CultureMech media, positioned in 2D using
UMAP reduction of 512-dim KG-Microbe DeepWalk embeddings.
Run from CultureMech/ directory.
# Dry-run: count CHEBI ingredients, skip embedding/UMAP (fast)
just gen-ingredient-umap-dry
# Generate full visualization (takes ~2-5 min for embedding load + UMAP)
just gen-ingredient-umap
# Force cache reload (if embeddings changed)
just gen-ingredient-umap-force-reload
# Open in browser
open app/ingredient_umap.html
# Default (uses local embeddings, auto-detects unified mapping)
python scripts/generate_ingredient_umap.py
# Custom output path
python scripts/generate_ingredient_umap.py --output docs/ingredient_umap.html
# Filter to ingredients seen ≥5 times
python scripts/generate_ingredient_umap.py --min-count 5
# Custom UMAP parameters
python scripts/generate_ingredient_umap.py --n-neighbors 30 --min-dist 0.05
# Explicit embeddings path
python scripts/generate_ingredient_umap.py \
--embeddings-path /path/to/DeepWalk...tsv.gz
| File | Description |
|---|---|
scripts/generate_ingredient_umap.py | CLI entry point |
src/culturemech/visualization/ingredient_umap_generator.py | IngredientUMAPGenerator class |
src/culturemech/templates/ingredient_umap.html | Jinja2 + D3.js template |
app/ingredient_umap.html | Generated output |
.umap_cache/CHEBI_embeddings.pkl | Pickle cache (auto-created, ~800MB) |
Auto-detected in this order:
data/embeddings/DeepWalkSkipGramEnsmallen_degreenorm_embedding_512_v2_2026-05-26_00_56_15.tsv.gz (local, preferred)../CommunityMech/CommunityMech/data/embeddings/...tsv.gz (sibling repo fallback)First load parses the full 1.4M-node TSV and caches CHEBI-only nodes as pickle. Subsequent runs use the cache (fast).
CAS-RN and KG-Microbe node IDs shown in tooltips come from:
../culturebotai-claw/workspace/unified_ingredient_mapping.tsv
Auto-resolved from sibling repo. If the file is absent, tooltips omit CAS-RN.
| Metric | Count |
|---|---|
| CHEBI ingredients found in YAML files | 660 |
| Embedded (found in KG graph) | 650 (98.5%) |
| Missing embeddings | 10 (rare/new CHEBI IDs) |
| Top 100 tier | 100 |
| Top 500 tier | 400 |
| Other tier | 150 |
The generator extracts CHEBI IDs from each ingredient using this priority:
term.id starting with CHEBI: (native CultureMech field)chebi_term.id starting with CHEBI: (enriched by MIM sync)data/chemical_name_to_chebi_mapping_enhanced.json (fallback)| Event | Action |
|---|---|
| MIM sync added new CHEBI IDs to CultureMech | just gen-ingredient-umap |
| New media records imported | just gen-ingredient-umap |
| KG-Microbe embeddings updated (new release) | just gen-ingredient-umap-force-reload |
match-kg-microbe — matches CultureMech media (not ingredients) to KG-Microbe nodescross-repo-sync (culturebotai-claw) — sync MIM CHEBIs before regenerating (increases coverage)manage-ingredient-hierarchy — hierarchy enrichment also populates ingredient CHEBI fields