Scalability of Mass Spectrometry Based Metabolomics

source

Scalability of mass spectrometry-based metabolomics for natural extracts libraries exploration - current status, challenges, and opportunities

Terms

• Natural products (NPs): Bioactive or structurally diverse compounds produced by living organisms (plants, microbes, marine organisms) that are often mined for drugs.
• Natural extracts libraries (NELs): Collections of solvent-derived extracts from many organisms, usually stored in well plates for high-throughput screening and chemical profiling.
• Metabolomics: Large-scale measurement of small molecules (metabolites) in samples; here focused on mass spectrometry for complex natural extracts.
• Mass spectrometry (MS): Technique measuring ions by mass-to-charge ratio (m/z) to detect and quantify molecules.
• LC-ESI-MS: Liquid chromatography + electrospray ionization + mass spectrometry; a common workflow to separate and detect metabolites.
• Dereplication: Identifying known compounds early to avoid rediscovering them.
• Feature: A detected signal interpreted as a putative compound-related ion (defined by m/z, retention time, intensity; may include MS/MS info).
• Feature detection: Converting raw MS data into a list/table of features (peaks) per sample.
• Feature alignment: Matching the “same” feature across many runs/samples (accounting for retention time drift, noise, missingness).
• Gap filling: Recovering missing intensities for aligned features by revisiting raw data in expected m/z/RT windows.
• Ion mobility spectrometry (IMS): Extra separation step distinguishing ions by shape/size/charge; adds another data dimension (and complexity).
• DDA (data-dependent acquisition): MS/MS collected for selected (often most intense) precursor ions; can miss low-abundance compounds.
• DIA (data-independent acquisition): MS/MS collected broadly across all ions; increases coverage but yields chimeric spectra.
• Chimeric spectra: MS/MS spectra containing fragments from multiple precursors due to co-isolation/co-elution.
• Spectral library: Reference MS/MS spectra for known compounds used for matching/annotation.
• Structural library: Database of candidate molecular structures (e.g., PubChem) used for in silico annotation.
• Cosine similarity: Standard score for comparing MS/MS spectra by overlap of fragment peaks/intensities.
• Modified cosine: Cosine variant that also matches peaks shifted by precursor mass difference (useful for analogues).
• Analogue search: Searching for “related” compounds (shared scaffold/family) rather than exact matches.
• Molecular networking (MN): Graph where spectra/features are linked by spectral similarity to reveal families of related molecules.
• Tanimoto similarity: Fingerprint-based structural similarity (Jaccard index) used for comparing molecules efficiently.
• Fingerprint (e.g., ECFP, MAP4): Bit-vector representation of a molecule’s substructures for fast similarity computations.
• Maximum common subgraph (MCS/MCES): Graph-based structural similarity measures; more interpretable but computationally expensive.
• Knowledge graph: Graph database linking samples, metadata, spectra, annotations, bioactivity, taxonomy for scalable querying/reuse.
• MassQL: Query language designed for mass spectrometry patterns in raw/processed data (m/z, RT, intensity logic).
• SPARQL: Query language for RDF/knowledge graphs; enables rich metadata + relationship queries.
• Non-structural annotation: Assigning useful properties (bioactivity likelihood, biosource, class) without full structure ID.
• Substructure annotation: Inferring recurring fragments/motifs corresponding to parts of molecules (even when full ID is unknown).
• MS2LDA / Mass2Motifs: Topic-modeling approach that discovers recurring fragment/loss patterns interpreted as substructures.
• De novo annotation: Predicting structures for unknowns without database matches (often via generative/ML models).

1) Core thesis: scalability is the bottleneck (and opportunity)

• NELs are powerful because they can contain huge biochemical diversity, but current NP workflows still behave “molecule-first”: they often end up deeply investigating only a small number of molecules/features.
• The review argues for shifting from molecule-first exploration to computationally enabled, scalable exploration of entire extract libraries—so we can prioritize better, reduce rediscovery, and make decisions earlier from MS data.
• “Scalability” for NELs isn’t just speed/throughput; it includes:
  • Time: fast enouh processing to avoid analysis bottlenecks.
  • Quality: broad, reliable metabolome coverage and usable annotation.
  • Data retrieval: being able to query, visualize, and act on results (decision support), not just generate tables.

2) Why NELs are different from biomedical/environmental metabolomics

• Biomedical metabolomics often scales by number of samples (large cohorts) but within a relatively narrower chemical space.
• Environmental/NP metabolomics scales by both chemical diversity and matrix complexity; NELs intensify this because they combine:
  • many extracts,
  • high structural novelty,
  • sparse distribution of individual compounds across the library,
  • and a high cost of “manual interpretation per molecule.”
• Therefore, NEL scalability must emphasize: coverage + prioritization + novelty assessment from complex, diverse mixtures.

3) Scalability across the pipeline (MS acquisition → processing → annotation → decisions)

3.1 Data acquisition: start clean or suffer forever

• No downstream method can fully rescue low-quality acquisition; scalable workflows require futureproofed acquisition choices.
• Key acquisition constraints:
  • Throughput vs resolution/coverage trade-offs are unavoidable and must be managed intentionally.
• Example approach: MultiplexMS (pool + computational deconvolution) to accelerate acquisition by exploiting the fact that many natural products are sparsely distributed across extracts.

Resolution & separation

• Chromatography (LC) remains essential for complex natural extracts: reduces co-elution, ion suppression, helps detect isomers, improves sensitivity for low-abundance metabolites.
• Skipping chromatography increases throughput but causes “data congestion” and worsens interpretability.
• Alternatives (CE, SFC, nano/microflow LC) are promising but adoption is constrained by infrastructure and throughput.
• Future direction: faster, automated, mixed-mode separations that better cover neglected chemical space (e.g., polar metabolites).

Ion mobility (IMS)

• Adds a valuable third dimension for distinguishing similar metabolites.
• But it increases data complexity and lacks standard, seamless integration in many workflows.
• Needs faster IMS platforms and better standardized software pipelines to be truly scalable for NELs.

Ionization (ESI; polarity)

• Positive mode is common but can add noise and in-source fragmentation.
• Negative mode provides complementary coverage.
• Polarity switching can improve coverage but may reduce scan rate and complicate alignment.
• Big need: smarter/adaptive strategies to optimize polarity per sample (or computational integration across modes).

Fragmentation strategies (CID, EAD, UVPD; DDA vs DIA)

• Lack of standardized collision energies hurts cross-lab/library matching.
• CID often yields incomplete structural info → lots of unannotated features.
• Multimodal fragmentation (EAD/UVPD) could enrich structural signal.
• DDA biases toward abundant ions; DIA improves breadth but produces chimeric spectra and increases computational burden.
• Notably: DIA hasn’t yet been widely used in published NEL exploration workflows (as described here), likely due to complexity.

Detection hardware: ToF vs Orbitrap

• ToF: fast acquisition, strong dynamic range, good for throughput-heavy contexts.
• Orbitrap: ultra-high resolution and mass accuracy but slower at top resolution (scan speed trade-off).
• Instrument choice shapes downstream computational needs and what questions are feasible.

3.2 Data processing: turning thousands of files into usable feature tables

Data formats & interoperability

• Scalable metabolomics depends on open, lossless, interoperable formats (e.g., mzML and related standards) and tooling ecosystems (pyOpenMS, matchms, spectrum_utils, MSnbase, etc.).
• A forward-looking requirement: formats and metadata conventions that remain compatible with AI/ML workflows.

Feature detection

• Feature detection extracts signals across RT (and possibly IMS) into features (m/z, RT, intensity).
• Scalability constraints are driven by sample count and data density (high scan rate, high resolution, IMS).
• Modern tools mitigate RAM constraints via memory mapping (offloading to disk while keeping random-ish access).
• Example: mzmine’s memory mapping and internal data model improvements dramatically reduced memory usage and enabled parallel processing at larger scales.

Feature alignment: project-centric vs sample-centric paradigms

• Project-centric alignment
  • Align all samples into one consensus feature table, then gap fill.
  • Benefits: consistent reference frame, easier batch-effect handling, simplifies missingness management, includes poorly fragmented features for stats/prioritization.
  • Requires careful algorithm design for concurrency; alignment/gap fill can be bottlenecks if poorly implemented.
• Sample-centric + knowledge graphs
  • Treat each sample as a core entity; integrate across studies via knowledge graphs.
  • Benefits: avoids “siloed projects,” supports reuse, aggregation, and continuous growth of a shared knowledge base.
  • Especially suited to long-term, multi-project NEL efforts and ML/AI-driven pattern discovery.

Feature grouping

• Groups adducts, in-source fragments, multimers (reducing redundancy).
• More sophisticated correlation/shape-based grouping improves accuracy but can become computationally expensive at large scale.
• Filtering low-abundance features improves dashboard usability—but risks removing rare/novel chemistry (a key tension for NEL discovery).

3.3 Annotation: scaling from “few IDs” to “useful knowledge at library scale”

Two similarity pillars: structural vs spectral

• Structural similarity
  • Fingerprints + Tanimoto: fast and scalable but can be counterintuitive due to lossy representations.
  • Graph approaches (MCS/MCES): more interpretable but computationally hard; heuristics (e.g., myopic-MCES) and hybrid methods try to bridge speed and richness.
• Spectral similarity
  • Many tasks rely on scoring large numbers of MS/MS comparisons:
    • exact library matching,
    • analogue search,
    • molecular networking,
    • repository-scale linking.
  • Cosine similarity remains the baseline; correct peak matching matters (avoid bin-edge artifacts; avoid double-counting matches).
  • Greedy peak matching is typically “good enough” and far cheaper than optimal assignment (e.g., Hungarian algorithm).
  • GPU implementations can massively accelerate similarity computation.

Exact search vs analogue search

• Exact search: match within small precursor tolerance → smaller search space.
• Analogue search: by nature tends toward many-vs-many comparisons → quadratic explosion if naive.
• Practical GNPS scaling trick: cap allowable mass shift (e.g., ~200 Da) to cover common biotransformations, trading recall for runtime.

Indexing + repository-scale tricks

• Fragment ion indexing: pre-index fragments to filter candidate comparisons quickly; used to accelerate cosine and entropy-based scores (e.g., “flash entropy”).
• Nearest neighbor suspect libraries
  • Use molecular networking connections from annotated spectra to propose hypotheses for unannotated “suspect” spectra at massive scale.
  • Key idea: propagate knowledge via spectral neighborhoods + plausible mass differences/modifications.
  • Outcome: expanding “annotation reach” beyond what classical reference libraries can do.

ML/deep-learning similarities and retrieval

• Learned embeddings (Spec2Vec, MS2DeepScore) aim to be more tolerant of multiple modifications and low direct peak overlap.
• Once embeddings are computed once per spectrum, similarity computation becomes cheap (cosine in embedding space).
• Model quality depends on curated training data → community datasets and ML-ready benchmarks matter (e.g., initiatives like MassSpecGym).

Tool spotlight: MS2Query

• Combines multiple scores/features with an ML model to rank candidates efficiently.
• Promises faster analogue search without needing strict mass-shift caps, enabling broader library consideration.

Spectral + structural annotation tools

• Spectral libraries (MassBank, GNPS sublibraries, specialized NP libraries) remain essential but coverage is still limited vs NP chemical space.
• Hybrid/in silico approaches (MetFrag, CFM-ID, FIORA; large in silico DBs like ISDB-LOTUS) help bridge gaps.

Substructure annotation (when full ID fails)

• MS2LDA discovers Mass2Motifs; useful but interpretation can become a human bottleneck at scale.
• MotifDB: curated motifs enable more efficient screening (fewer “free” motifs).
• MESSAR: supervised substructure recommendation via association-rule mining.
• Scaling limits: memory use and human validation time → pushes toward more modular, community-driven, scalable ecosystems (e.g., MS2LDA 2.0 + improved visualization/annotation guidance).

De novo annotation for unknowns

• Since unknowns dominate, generative approaches (e.g., MSNovelist) are positioned as important for expanding beyond known databases, though they bring heavier computational/validation demands.

Non-structural annotation: making progress without full structures

• Predict bioactivity/toxicity or biosource associations from spectral/metadata patterns.
• Taxonomy-informed annotation (e.g., TIMA) improves confidence by using biological plausibility.
• MASST-style repository searches (foodMASST/microbeMASST/plantMASST) contextualize unknowns by linking spectra to known biosources across public datasets.

4) Querying + decision-making: the “action” layer

Query languages as a scalability lever

• MassQL: MS-native querying over m/z/RT/intensity patterns; designed to be accessible to non-programmers.
• SPARQL: knowledge-graph querying enabling rich, relational questions that combine spectral + biological + metadata context (example: IDSM with similarity extensions).
• The review frames a future where MassQL (signal-level) and SPARQL (semantic/metadata-level) are combined into hybrid query workflows.

Prioritization tools and strategies

• Inventa: scores extracts for structural novelty using multiple components (feature specificity/annotation, literature signals, class detection, ML-based dissimilarity).
• Rational library minimization: reduce redundancy across organisms by iteratively selecting extracts that maximize new scaffold coverage; huge cost savings without losing hit rate (in the cited example).
• FERMO: integrates heterogeneous data types to filter/prioritize features using novelty/uniqueness/specificity/phenotype association; separates scalable computation from optional dashboard visualization. Annotation time can balloon with huge libraries → targeted libraries recommended until faster search methods are integrated.
• MS2DECIDE: “white-box” decision-theoretic ranking that combines GNPS + SIRIUS + ISDB-LOTUS outputs and agreement between suggested structures to estimate “knownness”/novelty.
• msFeaST: integrates stats + feature grouping + interactive chemical map; warns about human/UI scalability limits beyond ~5–10k features → encourages thresholding and careful selection before dashboard-driven exploration.

5) Outlook: what “scalable NEL metabolomics” should become

• Future progress requires coordinated advances in:
  • acquisition quality and throughput (better separations, smarter fragmentation, scalable modes like multiplexing),
  • computational efficiency (indexing, embeddings, GPU acceleration),
  • annotation breadth (repository-scale propagation, hybrid spectral+structural, substructure and non-structural annotation),
  • data infrastructure (open formats, interoperability, reusable pipelines),
  • governance/community (open science, shared resources like LOTUS/ENPKG-like infrastructures).
• The end-goal is an iterative “virtuous cycle”:
  • use prior knowledge (spectra/taxonomy/literature) → prioritize extracts → generate new data → refine priorities → rank features → elevate to semantic knowledge → enable complex queries → repeat.
• A recurring theme: scalability is not just compute—it’s also human interpretability, reproducibility, modular software design, and community-maintained ecosystems.