About MolTrace
MolTrace is a scientific intelligence company building the audit-ready evidence engine for pharmaceutical R&D. Every numerical claim we surface — peak, score, candidate, compliance verdict — is designed to be reachable, reproducible, and human-signed-off.
Why we exist
In our experience, routine structure elucidation can consume many hours per non-trivial small molecule, even with experienced analysts and modern NMR. Much of that is peak picking, integration adjustment, candidate ranking, and assembling the result into a reviewable narrative. The acquisition takes minutes; the rest is friction.
Published NMR-derived structures are not always reproducible, largely because the chain of custody from FID to regulatory submission is invisible by default — phase corrections done by eye, integration regions adjusted post-hoc, peak lists re-edited in spreadsheets.
Meanwhile the FDA's January 2025 AI framework and the EMA's reflection paper put new credibility burdens on every AI-derived claim in a regulatory submission. R&D groups are being asked to simultaneously: adopt more AI, prove it's reproducible, document every parameter that drove an assignment, and keep the human chain of decisions visible to inspectors.
The status-quo toolchain — spreadsheets, ad-hoc desktop processing, email-attached PDFs — doesn't address any of those four constraints. MolTrace is built to address all four.
The four commitments
Every architectural decision in the platform derives from one of these four. If a feature can't be justified against them, we don't ship it.
Every claim shown in the UI is reachable, by hyperlink, to its underlying data: the source spectrum file, the picked peaks, the SMILES candidate, the literature citation that justifies the chemical-shift window, and the human reviewer who released the final report. There is no confidence number with no audit trail anywhere in the system.
No regulatory document is released without an explicit human signoff. AI accelerates evidence assembly; humans make the call. This is designed to support both the FDA AI credibility framework (Stage 4 — human oversight gates) and the EMA reflection paper on AI in the medicinal-product lifecycle.
Where a community-maintained, peer-reviewed library exists, we use it: RDKit for cheminformatics, nmrglue for vendor FID parsing, mzML for MS interoperability, Pydantic for typed contracts, FastAPI for routing, Next.js for the UI. Proprietary code is confined to the evidence-orchestration and confidence-aggregation layers, where the additive value lives.
A pharmaceutical R&D group operates across NMR + LC-MS + HRMS + MS/MS + reaction history simultaneously. MolTrace fuses these as one evidence stack — not as separate apps — and uses cross-modal contradictions (e.g. HRMS exact mass disagreeing with NMR-implied formula) as first-class warnings.
The numbers we publish
Many marketing pages claim 95%+ accuracy without showing their work. We publish the corpus, the gate, and the regression test that catches us when we drift.
39
Evidence layers
Each layer is additive, typed, and never overwrites a prior layer's contract. Built incrementally Weeks 22 → 39.
94.4%
Solvent auto-detect (internal corpus)
Illustrative — internal NMRShiftDB2 20-fixture validation corpus, not an audited benchmark. The strict promotion gate target is 95%; the framework continues to validate the algorithm on HMDB-style references.
20
Fixture regression gate
FE-produced A/B JSON is wired into a backend CI test. Any detector drift > 50% on any fixture fails CI by fixture_id.
6–48 hrs
Today's bottleneck
Routine 1D NMR structure elucidation per non-trivial small molecule can take this long. In our experience, much of that is cognitive overhead — peak picking, integration, candidate ranking, report-writing.
What we won't ship
The product loop
Most analytical platforms stop at "the spectrum has been processed." Ours doesn't. Spectroscopy evidence flows directly into regulatory action items, which become constraints on the next round of reaction optimization.
Spectroscopy evidence engine. Raw FID → processed spectrum → peaks classified by category, with audit-grade fit metrics per peak.
Open moduleCloses the loop between spectroscopy evidence and regulatory action. Dossiers, traceability, ALCOA+ audit ledger, designed to support ICH Q2(R2).
Open moduleTurns regulatory action items into reaction-optimization constraints. Bayesian + ML-guided next-experiment recommendations under impurity limits.
Open moduleThe closing loop
Raw FID ─► Processed spectrum ─► Peaks + categories ─► Multi-modal evidence
│
▼
┌─► Next experiment (Repho) ◄─ Regulatory action items ◄─┘
│ (Regentry)
│
└─ with impurity / solvent / nitrosamine constraints fed back as priorsUnder the hood
Where a community-maintained library exists, we use it. Proprietary code is confined to the evidence-orchestration and confidence-aggregation layers — that's where the additive value lives.
This isn't a cost-savings choice. It's a regulatory one: open dependencies are inspectable, reproducible, and survive vendor turnover.
RDKit
Cheminformatics — SMILES canonicalisation, descriptors, substructure matching
nmrglue
Vendor-agnostic FID parsing (Bruker / Agilent-Varian)
mzML / mzXML
Open mass-spectrometry data interchange
lmfit
Voigt / Lorentzian fitting with per-peak QC residuals
Pydantic
Typed API contracts — the FE↔BE binding contract
FastAPI
Backend routing layer (Python 3.13)
Next.js 16 / React 19
Application UI (Vercel deployment)
Plotly
Spectrum rendering with static-plot anti-shake
Recent ships
Every existing endpoint and regression test stays green as new evidence layers land. These are the most recent capabilities to clear the gate.
May 28, 2026
Reduced χ², RMSE, FWHM, S/N, and baseline σ now exposed per peak on the same regulatory-tier surface GSD already provides.
May 27, 2026
Multiplet-line-granularity references that match how peak-pickers actually count. Gate criteria met on this specific internal reference corpus.
May 27, 2026
On internal benchmark hardware, processing of large 98,304-point FIDs runs meaningfully faster than the prior implementation.
May 27, 2026
Both detectors now expose { peaks, environments, environment_count, category_counts }. The FE renders both through one detector-agnostic panel.
May 27, 2026
Resolves the corpus-vs-detector granularity mismatch documented in §3.1 of the technical paper.
Where we work
Three regions we focus on, each chosen for its proximity to the pharma R&D + regulatory community.
Boston / Cambridge, MA
Americas
A region we focus on for its Kendall Square pharma corridor and proximity to the FDA Boston District office.
London, UK
EMEA
A region we focus on for its MHRA regulatory ecosystem and the broader EMA reflection-paper community.
Singapore
APAC
A region we focus on for its APAC pharma-manufacturing corridor and regional CRO base.
Compliance, in plain language
Designed to support SOC 2
Controls designed to support a SOC 2 Type II audit (security, availability, and confidentiality). SOC 2 is not yet held.
Designed to support GDPR
Tenant data residency + processing notices designed to support the EU framework.
Designed to support ICH Q2(R2)
Audit ledger + immutable raw vault + recipe-hash provenance built to support the ALCOA+ data-integrity principles.
Designed to support GxP validation
Releases gated by a human signoff queue with reviewer attribution per artefact.
Designed to support FDA AI framework (Jan 2025)
Risk-based credibility framework built to support explicit traceability, model documentation, and human oversight.
Designed to support EMA reflection paper
AI-derived evidence in submissions is designed to be reproducible, version-controlled, and subordinate to expert review.
Read the methodology, browse the modules, or talk to a human about how MolTrace would fit your evidence chain.