UC7 — FT-MS molecular-formula assignment & van Krevelen chemical space — drive debrief

Driven via /drive-scenario against local Galaxy (backend 8080, dev client 5173), 2026-06-20.

TL;DR

• Cleanest UC yet. All 9 RECETOX MFAssignR tools were already installed at the exact shipped pins (1.1.2+galaxy*) — no version migration, the shipped IWC mfassignr.ga imported and invoked unchanged. 10-step linear chain, every job green.
• MVP is a faithful reproduction + a genuinely good figure. From one 30,401-peak negative-mode mass list: noise estimate → isotope filter → first-pass CHO → recalibration → final assignment → van Krevelen diagram (the money shot). 3,036 unambiguous CHO formulas; the VK cloud is centered exactly where natural dissolved organic matter sits.
• The recalibration story is quantitative and real. Mean absolute mass error on assigned formulas tightens 1.39 ppm (first pass) → 0.49 ppm (after recalibration) — ~3× sharper. That is the QC narrative, backed by the MZplot.
• Stretch landed on-graph. An tp_awk pass bins each unambiguous formula by its van Krevelen region → a quantitative compound-class composition table: lignin/CRAM-like 57.2%, protein/amino-sugar 12.4%, tannin 12.2%, lipid 9.6%, condensed-aromatic 5.1%, carbohydrate 1.8%, unsaturated-HC 1.5%, other 0.2%. A textbook DOM fingerprint — the qualitative plot turned into a ranked number.
• Honest framing throughout: descriptive single-sample characterization (no condition contrast); assignment is partial (3,263 unassigned vs 3,036 unambiguous, ~52% of the candidate masses scored — far fewer relative to the 25,675 monoisotopic peaks, most of which never reach a candidate formula); shipped chemistry is CHO-only so heteroatom-class binning would be degenerate; region boundaries are literature conventions.
• Clean linear extraction. Walking the provenance backward from the VK plot recovered the whole assignment+recalibration workflow as a 10-step .ga that round-trips clean — the no-map-over extraction story the collection-heavy vignettes don’t show.

Artifacts

• History 4a56addbcc836c23 (30 datasets).
• Page b847e822bdc195d0, slug uc7-ftms-van-krevelen-molecular-formula (/published/page?id=b847e822bdc195d0), published. 7 sections, all genuine on-graph directives (history_dataset_as_image / _as_table).
• Extracted workflow UC7_mfassignr_van_krevelen_extracted.ga (10 steps; from-history provenance extraction; re-import round-trips clean, 9 tool steps + 1 data input, all connections intact).
• Figures (usecases dir): uc7_van_krevelen.png (money shot, 2100²), uc7_mzplot_recal.png (recalibration error-m/z QC), uc7_snplot.png (S/N QC), uc7_page_full.png (whole published page).

Pipeline as driven

1. Upload mfassignr_input.txt (Zenodo 13768009) via /api/tools/fetch — 30,401 peaks (mass + intensity), negative mode.
2. Import + invoke shipped mfassignr.ga unchanged (single “Feature table” input).
3. Noise — KMDNoise = 346.07 (carried forward; HistNoise = 317.35 concordant); SNplot QC figure.
4. Isotope filter — IsoFiltR → 25,675 monoisotopic / 4,726 isotopologue peaks.
5. First-pass CHO — MFAssignCHO → 3,097 formulas as recalibration anchors.
6. Recalibration — RecalList → FindRecalSeries → Recal; MZplot; mean |err_ppm| 1.39 → 0.49.
7. Final assignment — MFAssign (CHO-only) → 3,036 Unambig / 0 Ambig / 3,263 None + plots collection (VK, MSgroups, errorMZ, msassign). The money-shot VK is the final MFAssign element cd04f282f6a11f18 (hid28) — not the near-identical first-pass CHO VK (hid24); they sit adjacent in the history, so embed by the MFAssign-created id.
8. Stretch — tp_awk_tool bins Unambig by O/C (col 35) + H/C (col 36) into VK regions → compound-class composition table.
9. Extraction — from-history workflow extraction (job_ids backward from VK) → 10-step .ga; round-trip re-import verified.

Gotchas / notes

• From-history extraction API id formats are mixed (legacy controller): job_ids must be encoded ids (controller decodes them); dataset_ids must be raw HID integers (e.g. [1]), passed through and int()-ed in extract_steps. Passing encoded dataset ids → ValueError: invalid literal for int(); passing raw int job ids → “unable to decode”. An empty job_ids/dataset_ids extraction silently yields a 0-step workflow.
• Page directives reject title= on history_dataset_as_table (400 “Invalid argument to Galaxy directive [title]”) — drop the arg.
• Full-page screenshot can blank large images mid-capture (they showed as pink boxes); the images were actually complete (2100²) — re-shoot after confirming img.complete.
• Shipped MFAssign is CHO-only (Nx=0, Sx=0, Px=0, Ox=30) → 3,029 CHO + 7 CH; the meaningful composition axis is van Krevelen region, not heteroatom class. Enabling N/S/P would extend the stretch.
• Ambig=0 is a config artifact (ambig_bool=off), not absence of ambiguity — multi-candidate masses fall into None. Stated honestly on the page.

Key IDs

History 4a56addbcc836c23; input h5… 82cdcb0c41950210 (hid1). SNplot 8eb3358bbd1313c6 (hid6), monoisotopic hid7, first-pass CHO Unambig hid10. MZplot 0453c666c6590847 (hid18). Final Unambig 3e2dfd57e568189f (hid20), None 04f11d4cb8d28870 (hid22), VK cd04f282f6a11f18 (hid28). VK composition 174eabd470156671 (hid31). Extracted workflow UC7_mfassignr_van_krevelen_extracted.ga.