Stats
Actions
Tags
From chemaudit
Audits cheminformatics datasets with Fourches-style 5-component health scores, detects contradictory bioactivity labels in duplicates, compares versions, and generates curation reports from CSV/SDF files.
How this skill is triggered — by the user, by Claude, or both
Slash command
/chemaudit:chemaudit-dataset-intelligenceThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Whole-dataset auditing — operates on the file, not individual molecules. Four outputs:
Whole-dataset auditing — operates on the file, not individual molecules. Four outputs:
| Sub-score | Default weight | What it measures |
|---|---|---|
parsability | 0.25 | Fraction of rows with a valid, parseable SMILES |
stereo | 0.15 | Fraction with defined stereocenters where chirality is expected |
uniqueness | 0.20 | 1 - (duplicates / total) by canonical InChIKey |
alerts | 0.20 | Fraction of molecules alert-free against the default catalogs |
std_consistency | 0.20 | Fraction agreeing across three standardization pipelines (RDKit vs ChEMBL vs minimal) on a sampled subset |
Weights are customizable per audit; source of literature-backed defaults is DEFAULT_WEIGHTS in frontend/src/types/dataset_intelligence.ts.
Overall score is the weighted sum × 100.
For every InChIKey appearing ≥ 2 times with a numeric activity column, reports the max/min fold-difference. Threshold options: 3, 5, 10, 50, 100 (default: 10). Entries above the threshold are flagged as contradictory.
curl -sS -X POST http://localhost:8000/api/v1/dataset/upload \
-F "[email protected]" \
-F "smiles_column=canonical_smiles" \
-F "activity_column=pIC50"
Both column names are optional — auto-detection runs if omitted. Accepts .csv and .sdf. Response:
{
"job_id": "3f7a9b...",
"filename": "bioactivity.csv",
"file_type": "csv",
"status": "pending",
"message": "Dataset audit job submitted"
}
WebSocket:
const ws = new WebSocket(`ws://localhost:8000/ws/dataset/${job_id}`);
Poll:
curl -sS http://localhost:8000/api/v1/dataset/<job_id>/status
Response: {job_id, status, progress, current_stage, eta_seconds}. Stages: parse → health_audit → contradictions → standardization_sample → diff_ready.
curl -sS http://localhost:8000/api/v1/dataset/<job_id>/results
While processing, returns HTTP 202 with {job_id, status, message}. When complete, returns:
{
"job_id": "...",
"status": "complete",
"health_audit": {
"overall_score": 78.3,
"sub_scores": [
{"name": "parsability", "score": 0.985, "weight": 0.25, "count": 15, "total": 1000},
{"name": "stereo", "score": 0.62, "weight": 0.15, "count": 380, "total": 1000},
{"name": "uniqueness", "score": 0.93, "weight": 0.20, "count": 70, "total": 1000},
{"name": "alerts", "score": 0.81, "weight": 0.20, "count": 190, "total": 1000},
{"name": "std_consistency", "score": 0.88, "weight": 0.20, "count": 12, "total": 100}
],
"weights": {"parsability": 0.25, "stereo": 0.15, "uniqueness": 0.20, "alerts": 0.20, "std_consistency": 0.20},
"molecule_count": 1000,
"issues": [
{"row_index": 42, "smiles": "C(C)(C)(C)(C)C", "issue_type": "parse_error", "severity": "error", "description": "..."}
],
"property_distributions": {
"mw": {"bins": [...], "counts": [...]},
"logp": {"bins": [...], "counts": [...]},
"tpsa": {"bins": [...], "counts": [...]}
},
"std_pipeline_comparison": {...},
"std_sample_size": 100,
"dedup_groups": [{"inchikey": "...", "rows": [12, 88]}]
},
"contradictions": [
{
"inchikey": "BSYNRYMUTXBXSQ-UHFFFAOYSA-N",
"entries": [
{"row_index": 5, "smiles": "...", "activity": 7.2},
{"row_index": 203, "smiles": "...", "activity": 4.1}
],
"fold_difference": 1259.0,
"entry_count": 2,
"smiles": "CC(=O)Oc1ccccc1C(=O)O"
}
],
"numeric_columns": [
{"name": "pIC50", "priority": 1},
{"name": "MW", "priority": 2}
],
"curation_report": {...},
"curated_csv_available": true
}
Upload the comparison file against a completed primary job:
curl -sS -X POST http://localhost:8000/api/v1/dataset/<job_id>/diff \
-F "file=@bioactivity_v2.csv"
Response:
{
"added": [{"inchikey": "...", "smiles": "...", "row_index": 12, "properties": {...}, "changes": []}],
"removed": [...],
"modified": [{"inchikey": "...", "changes": [{"column": "pIC50", "old_value": 6.2, "new_value": 7.1}]}],
"added_count": 45,
"removed_count": 12,
"modified_count": 89,
"unchanged_count": 854,
"unique_columns_primary": 2,
"unique_columns_comparison": 3
}
Diff uses InChIKey as the primary key, so it's robust to SMILES notation differences.
# Full curation report JSON
curl -sS http://localhost:8000/api/v1/dataset/<job_id>/download/report \
-o curation_report.json
# Curated CSV — original rows + appended diagnostic columns
curl -sS http://localhost:8000/api/v1/dataset/<job_id>/download/csv \
-o curated.csv
Curated CSV appends: _health_issues (joined issue types), _is_duplicate (bool), _standardized_smiles, _alert_flags (joined catalog names).
POST /dataset/upload with smiles_column and activity_column pointing at your CSV./status or WS until complete.GET /results → read health_audit.overall_score (aim ≥ 70) and health_audit.sub_scores[] for the weakest component.contradictions[] (keeps only groups above the default 10-fold threshold — check entry_count and fold_difference per group).issues have severity="error", those rows need manual review.complete → capture job_id_v1.POST /dataset/<job_id_v1>/diff with the v2 file.added_count, removed_count, modified_count for an overview.modified[] contains changes[] arrays — look for activity-column changes.GET /download/report → JSON you can paste into an appendix or render as markdown.GET /download/csv → drop-in replacement dataset with diagnostic columns./dataset/upload: 3/min./dataset/<job_id>/status: 60/min./dataset/<job_id>/results: 30/min./dataset/<job_id>/diff: 10/min./dataset/<job_id>/download/*: 10/min.Only .csv and .sdf extensions accepted — .tsv, .txt, etc. are rejected by dataset-intelligence (unlike batch validation which accepts all four).
Check /api/v1/config for the deployment limit. Split the file or redeploy with a larger profile.
/resultsJob is still running. Keep polling /status (lighter endpoint) until status flips to complete or error.
Diff endpoint needs the primary job to finish first. Wait until the primary shows status=complete in /status.
activity_column auto-detected wrong columnSpecify it explicitly. The auto-detector ranks by name heuristics and numeric-column detection; numeric columns are returned in numeric_columns[] sorted by priority.
contradictions empty when you expect conflictsThe default fold threshold is 10. Lower-fold conflicts aren't reported unless you customize the threshold (currently set server-side; contact admin to change). Check entries[] length — duplicates below threshold are listed but won't appear as contradictions.
std_consistency score very lowAll three standardization pipelines (RDKit vs ChEMBL vs minimal) disagreed for many molecules. Usually means the dataset has unusual edge cases (metal complexes, large fragments, unusual charges). Look at std_pipeline_comparison for per-pipeline statistics.
curated_csv_available in resultsThe Celery worker failed mid-curation-report step. Re-upload. If it persists, check worker logs.
chemaudit-qsar-ready — once the audit reveals issues, curate per molecule with provenance.chemaudit-standardization — per-molecule investigation of why two SMILES share an InChIKey.chemaudit-diagnostics — pre-flight file integrity checks before upload.npx claudepluginhub kohulan/chemauditBatch processes CSV, TSV, TXT, or SDF molecule files via ChemAudit with Redis progress tracking, on-demand analytics (scaffold, clustering, MMP, taxonomy, R-group), and nine export formats.
References data quality dimensions with qsv checks and provides remediation decision tree for tabular CSV assessment and fixes.
Curates FiftyOne datasets: inspect schema, audit annotations, analyze class distributions, find duplicates, create subsets, and build train/val/test splits. Works with any CV dataset type.