SpectraCheck is the spectroscopy intelligence engine inside MolTrace. NMR, LC-MS, HRMS, and MS/MS arrive as one evidence stack — typed, traceable, multi-modal by default. Every numerical claim links back to the spectrum it came from.
The pipeline, end-to-end
Every stage emits a typed Pydantic model with stable JSON keys. Every emission is pinned to the immutable raw archive by SHA-256 + recipe hash — designed so you can replay a report from any prior date and reproduce bit-identical output.
Raw FID archive (Bruker / Agilent-Varian) lands in the immutable vault. SHA-256 hashed. Vendor metadata extracted: SFO1/BF1 (Bruker) or sfrq/reffrq (Varian).
Outputs
Archive · vendor · field_mhz · acquisition params
Fourier transform + apodization + phasing + baseline correction. Every parameter recipe-hash-linked to the unchanged raw archive — designed for byte-for-byte reproducible replay.
Outputs
Processed spectrum (ppm + intensity) · processing_metadata
Legacy detector (default) or GSD-Prompt-3 (experimental opt-in). Both surface the same envelope: peaks, environments, category_counts, environment_counts.
Outputs
Peaks · environments · category mix · per-peak fit metrics
Each peak auto-classified: compound / solvent / impurity / artifact / ¹³C satellite. Curated impurity-shift tables match labels (residual CHCl₃, methanol, acetic acid, BHT, …).
Outputs
Category · category_reason · impurity_match / solvent_hit
39-layer evidence stack scores candidate SMILES across NMR, HRMS, MS/MS, predicted shifts, fragmentation trees, and reaction history. Cross-modal contradictions surface as warnings.
Outputs
Ranked candidates · DP4 confidence · multi-layer evidence
Structure-elucidation report composer designed for regulatory submission. Every numerical claim is hyperlinked to its source. Human signoff required before release.
Outputs
Audit-grade report · ALCOA+ ledger entry
At a glance
Raw FID ──► Process FID ──► Detect peaks ──► Categorise
│ │ │ │
│ │ │ ▼
│ │ │ Cross-modal score
│ │ │ (HRMS · MS/MS · 2D NMR)
│ │ │ │
▼ ▼ ▼ ▼
SHA-256 recipe_hash per-peak QC candidate ranking + DP4
immutable reproducible (χ²ᵣ, RMSE, with audit-grade trail
vault processing FWHM, S/N) and human signoff gateModalities supported
A pharmaceutical R&D group operates across NMR + LC-MS + HRMS + MS/MS simultaneously. SpectraCheck fuses these as one evidence stack — not separate apps — and uses cross-modal contradictions (HRMS exact mass disagreeing with NMR-implied formula) as first-class warnings.
Layers 22 · 24
Accepts
Raw FID · JCAMP-DX · CSV · vendor exports
Layers 23 · 25
Accepts
Processed 2D NMR · vendor archives
Layers 29 · 30 · 31 · 32
Accepts
mzML · mzXML · processed peak lists
Layers 35 · 36 · 37
Accepts
mzML · mzXML · processed peak tables
Auto-classification
Peaks aren't just numbers. SpectraCheck applies a 5-category taxonomy — backed by curated impurity-shift tables and solvent residual-peak windows — so reviewers can scan a peak list and immediately see what's signal, what's noise, and what needs follow-up.
Categories are exposed as structured fields ( peak.category, peak.solvent_hit, peak.impurity_match) — not buried in a confidence number. Each impurity match cites the reference shift, the observed shift, and the delta.
Target analyte signals — passed to candidate scoring.
Residual CHCl₃ at 7.26, DMSO-d₆ at 2.50, CD₃OD at 3.31. Auto-masked from candidate scoring.
Curated shift-table match — methanol, ethyl acetate, acetic acid, DMF, dichloromethane, BHT, …
Phasing distortions, sidebands, t₁ noise. Excluded from candidate scoring; flagged for reviewer.
Spinning sidebands of strong ¹H peaks at natural-abundance ¹³C positions. Down-weighted.
Two detectors, one envelope
We ship a stable, evidence-pipeline legacy detector as the default. The experimental GSD-Prompt-3 backend (industry-standard peak detection with auto-classification) ships as opt-in until it clears our internal validation gate. Both return the same { peaks, environments, category_counts } envelope so consumer code is detector-agnostic.
| Capability | Legacy (default) | GSD (experimental) |
|---|---|---|
| Peak detection on NMRShiftDB2 corpus | Median Δ = 14 | Median Δ = 19 (Δ = 5 compound-only) |
| Solvent auto-detect (structured field) | Inferred via impurity-match table | 94.4% on our 18-fixture internal benchmark |
| Per-peak fit QC (χ²ᵣ, RMSE, FWHM, S/N, baseline σ) | Now exposed (Phase 24, normalised pending) | Native — normalised to baseline σ |
| Multiplicity + J-values + integration | Native — multiplet notation per peak | Not exposed |
| Candidate structure matching · DP4 ranking | Full evidence pipeline | Peak detection only — no candidate matching |
| Category classification (5-set) | Open string (detection + chemical-region) | Closed enum, native |
| Promotion status | Default backend | Experimental · opt-in · gated |
Full A/B methodology + per-fixture numbers documented in Field notes and the technical white paper §3.1.
Internal validation benchmarks
Every figure below comes from our internal benchmark corpora. The regression gate runs in CI; any drift larger than 50% on any single fixture fails the build with the fixture_id called out by name.
94.4%
Solvent auto-detect (internal benchmark)
On our 18-fixture internal NMRShiftDB2-derived benchmark (17 of 18 fixtures with a reference). Strict gate target: 95%.
Δ ≤ 2
Compound-count median
Median absolute peak-count delta vs expert-curated references on the HMDB-style multiplet-line corpus. Strict gate cleared.
20
Fixture regression gate
Every detector change runs against a curated corpus before merge. CI fails by fixture_id when any single fixture drifts > 50%.
39
Evidence layers
Built additively Weeks 22 → 39. Every layer's output is a typed Pydantic model with stable JSON keys.
The closing loop
Most analytical platforms stop at "the spectrum has been processed." Ours doesn't. A real worked example with acetic acid impurity:
Impurity peak at 2.10 ppm matched to acetic acid CH₃ within 0.001 ppm. Confidence 93%.
Acetic acid is ICH Q3C Class 3 — no need for action below 5000 ppm. Action item raised only if observed concentration crosses threshold.
Next experiment's reaction recipe receives the impurity limit as a Bayesian prior. Solvent + workup updated automatically.
Re-acquired spectrum confirms impurity below threshold. Audit ledger records every step from FID hash to recipe update.
Audit & compliance
SpectraCheck's controls are designed to support compliance — by design, not by accident. The audit ledger, the immutable raw vault, the recipe-hash provenance, and the human-signoff release gate were designed to support ICH Q2(R2) ALCOA+, the FDA's January 2025 AI framework, and the EMA reflection paper from day one.
Immutable raw vault
Every FID is SHA-256 hashed, vault path policy enforced, and never overwritten.
Recipe-hash provenance
Every processing run links a recipe hash to the unchanged raw archive. Designed for bit-identical replay.
Human signoff queue
No regulatory document is released without an explicit qualified-human attribution.
ALCOA+ audit ledger
Attributable · Legible · Contemporaneous · Original · Accurate · Complete · Consistent · Enduring · Available.
Cross-modal contradiction warnings
HRMS exact mass disagreeing with NMR-implied formula raises a first-class warning before signoff.
Tenant isolation by default
Controls designed to support SOC 2 Type II, data-residency controls designed to support GDPR, role-scoped audit-event ledger.
Open SpectraCheck on the platform or schedule a 30-minute walkthrough on a real analyte from your workflow.
FAQ
SpectraCheck ingests raw FID archives from Bruker and Agilent-Varian (parsed via the open-source nmrglue library), plus JCAMP-DX, CSV, and vendor exports for NMR, and mzML, mzXML, and processed peak lists for LC-MS, HRMS, and MS/MS.
It ranks candidate SMILES across a multi-layer evidence stack (1H/13C, 2D NMR, HRMS, MS/MS, predicted shifts, fragmentation trees) and runs an automated structure-verification layer, but the platform frames every output as decision support, never a proof of identity.
It aggregates cross-modal evidence into a ranked candidate list and exposes a DP4/DP5 panel, but MolTrace states the result is decision support — not proof of identity, and not a calibrated DP4/DP5 probability. A human reviewer weighs it before signoff.
Every numerical claim links back to its source spectrum; raw FIDs are SHA-256 hashed in an immutable vault, processing is recipe-hash-linked for reproducible replay, and no report releases without human signoff recorded in an ALCOA+-aligned ledger — controls designed to support 21 CFR Part 11 and ICH Q2(R2), not a compliance claim.
Each peak is auto-classified as compound, solvent, impurity, artifact, or 13C satellite against curated Fulmer/Gottlieb residual-solvent and impurity-shift tables, so non-analyte signals are masked from candidate scoring and flagged for the reviewer.
NMR (1H/13C, with 2D COSY/HSQC/HMQC/HMBC behind a feature flag), HRMS, MS/MS, and LC-MS features fuse into one evidence stack, and cross-modal contradictions — such as an HRMS exact mass disagreeing with the NMR-implied formula — surface as first-class warnings.