Analytical instruments produce more spectra than experts can interpret. Deep models now read raw MS/MS, NMR, IR and chromatographic signals directly: they sequence peptides and antibodies without a database, rank candidate structures for unknown molecules in metabolomics or non-target screening, and integrate chromatography peaks consistently. The benefit is faster, more reproducible interpretation, with the chemist or analyst still deciding on identity and release.
Problem & value
- Today: peptide identification relies on database search (protein FASTA required), small-molecule identification on matching against spectral libraries that cover only a small fraction of known and possible chemistry, and NMR elucidation on expert manual work. Chromatography peaks in QC labs are often re-integrated by hand, a classic data-integrity finding.
- AI value: (1) de novo sequencing for antibodies, novel proteins, unsequenced species and immunopeptidomics where no database exists; (2) "dark matter" metabolites and unknown pollutants annotated beyond library matches; (3) consistent peak integration and system-suitability checks; (4) first-pass structure proposals from NMR/IR to shorten expert time.
- Evidence: the open models Casanovo and InstaNovo (Noble lab, InstaDeep/DTU) show that spectrum-to-sequence transformers work on real proteomics data [1][2]; CSI:FingerID/SIRIUS and MIST-style models predict molecular fingerprints and structures from MS/MS [3][4]. Published gains are benchmark-specific, so measure on your own instruments and matrices.
Variants
- Proteomics / biopharma: de novo peptide and antibody sequencing, sequence-variant and PTM discovery, immunopeptidomics, biosimilar characterisation.
- Metabolomics & natural products: compound annotation, class prediction, molecular networking.
- Non-target screening: water, food, feed and consumer-product contaminants; supports REACH-type work (see SCI-015).
- NMR/IR/Raman structure elucidation: synthetic chemistry, process chemistry, polymer and materials labs; pairs with SCI-004 and SCI-007.
- QC chromatography: automated peak detection and integration in regulated labs (GxP).
- Clinical/forensic toxicology: drug and metabolite screening. Keep to library/annotation level; results feed human-signed reports only.
How it works
- instrument raw data
- vendor-neutral conversion (ProteoWizard msConvert
- mzML; NMR
- JCAMP/nmrML)
- signal processing (centroiding, baseline/phase correction, deisotoping, alignment)
- model inference MS/MS peptides: spectrum
- sequence (transformer)
- FDR/rescoring MS/MS small molecules: spectrum
- formula
- fingerprint/structure candidates
- ranking NMR/IR: spectrum (+ formula)
- structure candidates
- predicted-vs-measured shift check chromatography: baseline + peak detection
- integration
- suitability flags
- confidence, FDR control and identification level (e.g. MSI levels 1–5)
- expert review (spectrum viewer, annotated evidence)
- report / LIMS / ELN
- Key models: encoder-decoder transformers on peaks (Casanovo, InstaNovo); fingerprint- or graph-based predictors (SIRIUS/CSI:FingerID/CANOPUS, MIST); contrastive spectrum-structure embeddings for library-free retrieval; CNN/transformer models for NMR and IR; classical signal processing remains the backbone for chromatography, with ML for peak classification.
- Human-in-the-loop: all novel identifications, structure proposals and any integration change in a regulated lab are reviewed; AI output is a ranked hypothesis, not a result.
- Orthogonal checks: formula consistency, isotope pattern, retention-time prediction, predicted vs. measured NMR shifts and, for final proof, reference standards.
Data preparation
Spectra live on instrument PCs, lab file shares and in chromatography data systems (CDS), LIMS or ELN, often in proprietary vendor formats (.raw, .d, .wiff), with sample metadata in Excel or on paper. Typical gaps: missing or inconsistent sample metadata, no common format, undocumented acquisition methods, instrument drift between batches, scattered reference libraries, and unclear access rights on shared instrument PCs.
Preparation steps: inventory instruments and data flows (PREP-001, PREP-005); convert to open formats with ProteoWizard msConvert (mzML) or JCAMP-DX/nmrML (PREP-007); digitise paper logbooks and sample sheets at source (PREP-008, PREP-006); structure sample metadata into tidy tables (PREP-011); harvest labels from past confirmed identifications (PREP-020); build instrument-matched test sets with the analysts (PREP-018). Tools: msConvert, matchms, Nextflow, a FAIR data catalogue; in regulated labs keep audit trails intact. See DATA-003 and GUIDE-003.
Expect 40-60 percent of effort in conversion, metadata curation and test-set creation.
Implementation paths
Which path for whom
- Enterprise: C - open workflows (OpenMS, SIRIUS, Casanovo) are reproducible, avoid vendor lock-in on raw data and run on HW2; buy (A) where validated GxP software is mandatory for release testing.
- Developer: C - open weights, containers and workflow managers are the natural base for building and benchmarking.
- Public: C - university and agency labs rely on open formats and open tools, and sovereign data handling is easy to demonstrate.
Path A — Buy (off-the-shelf)
- Instrument vendors: Thermo Fisher (Compound Discoverer, Proteome Discoverer, Chromeleon), Waters (UNIFI, waters_connect, Empower), Agilent (MassHunter, OpenLab), SCIEX (OS, Biologics Explorer), Bruker (MetaboScape, TopSpin, Compass).
- Specialist software: Bioinformatics Solutions (PEAKS, Canada), Protein Metrics Byos, Mnova by Mestrelab (Spain, EU: NMR, MS), ACD/Labs (Spectrus, incl. structure elucidation), Spectronaut and DIA-NN-based pipelines (Biognosys, Switzerland), MSAID (Germany, EU, spin-out of TUM; Proteomics software with ML rescoring), Genedata (Expressionist, Switzerland).
- Choose A when: the lab already runs one vendor's instruments, needs validated software (GxP, 21 CFR Part 11 / EU Annex 11) and has no data-science team.
Path B — Build on cloud APIs
- Cloud storage and batch compute in EU regions (AWS, Azure, Google; or EU providers such as OVHcloud, IONOS, STACKIT) run converters and open tools at scale; managed ELN/LIMS data lakes (e.g. Benchling, Genedata, TetraScience data platform) provide harmonised data.
- A frontier reasoning LLM (Claude, GPT, Gemini, Mistral) acts as orchestrator and explainer via tool calls: runs SIRIUS/Casanovo, drafts review summaries, writes analysis scripts. It must not be treated as a spectral interpreter on its own: general LLMs are unreliable at reading raw spectra.
- Choose B when: you need pipeline integration and elastic compute but no custom model training.
Path C — Self-host (open source)
- MS peptides: Casanovo, InstaNovo (open weights), plus Sage, MSFragger, MaxQuant, DIA-NN, Percolator/Mokapot for rescoring; OpenMS and pyOpenMS as workflow base; Nextflow/nf-core (e.g. nf-core/quantms) for reproducible runs.
- MS small molecules: SIRIUS (CSI:FingerID, CANOPUS, open-source; Friedrich Schiller University Jena, Germany, EU), MZmine, MS-DIAL, matchms, GNPS and MassBank (public libraries), MassBank of North America; NMR: nmrglue, NMRium (web, open-source), plus open prediction models.
- Chromatography: open-source libraries (pybaselines, scipy-based peak finders, OpenChrom).
- Hardware: HW2 (one 24 GB GPU) suffices for inference and fine-tuning of Casanovo-size models; raw-data storage and CPU search usually dominate (tens of TB at facility scale).
- Ops effort: needs bioinformatics/cheminformatics staff; reproducibility via containers and workflow managers.
Path D — Custom training
Worth it for lab-specific instruments, unusual chemistries (glycans, cross-links, lipids, polymers) or non-standard fragmentation. Needs thousands to millions of annotated spectra (public: MassIVE, PRIDE, GNPS, MassBank, NIST-licensed libraries), synthetic data from simulators/predictors and held-out, instrument-matched test sets. Fine-tune existing open models first.
State of the art (2026)
- De novo peptide sequencing moved from curiosity to routine for antibodies and immunopeptidomics; transformer models (Casanovo, InstaNovo) are open, and commercial tools (PEAKS) embed deep learning [1][2]. Peptide-level precision still drops for long peptides, PTMs, low-intensity spectra and different instruments than training data.
- Small-molecule annotation: structure retrieval/generation from MS/MS improves over library search, but correct top-1 structure rates remain modest on novel compounds; community challenges such as CASMI show this honestly [4]. Treat outputs as candidate lists with identification confidence levels (Schymanski scale).
- NMR/IR: ML shift prediction is accurate enough to rank candidates; complete de novo elucidation of complex natural products from 2D NMR is improving but still expert-led.
- Chromatography: automated integration works for well-behaved peaks; co-elution, drift and noisy baselines need review; regulators expect audit trails for reprocessing.
- Limits and failure modes: distribution shift between instruments, hallucinated but plausible sequences/structures, overconfident scores, isomers indistinguishable by MS, leaked test data in benchmarks, proprietary raw formats and library licensing restrictions.
Shared resources & combinations
- Hardware/stack: shares HW2 GPU and workflow infrastructure with SCI-008, SCI-009, SCI-017.
- Upstream/downstream: SCI-007 (self-driving labs generate and consume spectra), SCI-004 (retrosynthesis plus product confirmation), SCI-002 (metabolite ID in drug discovery), SCI-015 (safety assessment of detected substances), SCI-017 (proteomics biomarkers), SCI-022 (test-data analytics), PRED-014 (soft sensors use spectra, e.g. NIR, but do not elucidate structure).
- Data layer: open formats (mzML, nmrML, AnIML, Allotrope) and a FAIR data catalogue are the shared prerequisite for all of these.
Effort, cost & KPIs
- Path A: weeks to deploy; licences per seat/instrument, validation effort for GxP.
- Path B: 2–4 months; compute and storage egress dominate cost.
- Path C: 3–6 months with 1–2 specialists; GPU modest, storage and curation significant.
- Path D: 6–12+ months; data curation and annotation are the main cost.
- KPIs: identification rate at fixed FDR, top-k structure accuracy on a labelled test set, expert review time per sample, reintegration/manual-edit rate, reproducibility across batches, turnaround time, share of unknowns annotated.
People & change
Start with the lab head and the analysts who interpret spectra every day: they know the failure modes and must own acceptance rules. Add IT early for storage, instrument-PC access and containers, and IT security for networked instruments. In GxP labs involve quality assurance and legal/compliance for validation, audit trail and Annex 11 / Part 11 expectations. The data protection officer (DPO) is needed only when samples are clinical, forensic or biobank material; then a DPIA (Art. 35 GDPR) is likely and cloud vendors need an Art. 28 processing agreement.
Works council: depends. Chromatography and instrument software log user actions, and the Federal Labour Court holds that objective suitability to monitor behaviour suffices for co-determination (section 87(1) no. 6 BetrVG). Inform the council in good time (section 90(1) no. 3), agree that logs serve data integrity only, not performance review. Public labs involve the staff council.
AI literacy: the AI Act Art. 4 requires measures that support staff AI literacy; run a short training on confidence levels, FDR and out-of-distribution failures. Choose one instrument and one question as pilot. Typical fear: "peak integration will be automated and my job with it" - address by keeping every identification and any reintegration under analyst sign-off.
Guides: GUIDE-008, GUIDE-009, GUIDE-004.
Risks, regulation & compliance
- EU AI Act: typically minimal risk (research and analytical tooling); AI literacy duty (Art. 4). Higher classes apply only if the output becomes a safety component or part of a regulated medical device or IVD (Annex I; MDR/IVDR), or in clinical decision contexts. No Art. 5 prohibition is engaged by normal analytical use.
- GDPR: mostly no personal data; clinical, doping, forensic or biobank samples (proteomics, metabolomics) can be personal or health data (Art. 9): needs legal basis, pseudonymisation, DPA with cloud vendors and EU hosting.
- GxP / data integrity: EU GMP Annex 11, FDA 21 CFR Part 11, ALCOA+ principles; validated software, audit trail for every reprocessing, locked methods, and documented AI model versions.
- Sector rules: accreditation (ISO/IEC 17025) and method validation for routine reporting; identification must follow community reporting standards (MSI/Schymanski levels).
- Dual-use awareness: unknown-compound identification can touch controlled substances; apply institutional export-control and access policies.
- Mitigations: orthogonal confirmation, reference standards, locked model versions, blinded re-testing and drift monitoring.
Getting started
- Pick one question (e.g. antibody sequence confirmation, unknown peaks in a water screen) and assemble 50–200 well-annotated runs as a test set.
- Convert to open formats and run baselines: a database search or library match, then Casanovo/InstaNovo or SIRIUS for comparison.
- Compare against ground truth at fixed FDR; log failure modes by instrument and matrix.
- Add a review UI (spectrum plus evidence) and define acceptance rules and identification levels.
- Version models, data and workflow (Nextflow, containers); then scale or fine-tune.
Sources
- Casanovo project — https://github.com/Noble-Lab/casanovo
- InstaNovo — https://github.com/instadeepai/InstaNovo
- SIRIUS / CSI:FingerID / CANOPUS documentation — https://v6.docs.sirius-ms.io/
- CASMI contest (critical assessment of small molecule identification) — http://www.casmi-contest.org/
- OpenMS — https://openms.de/
- GNPS — https://gnps.ucsd.edu/
- MassBank — https://massbank.eu/
- EU GMP Annex 11 (EudraLex Vol. 4) — https://health.ec.europa.eu/medicinal-products/eudralex/eudralex-volume-4_en
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj