MS-Net: Multi-Similarity based network annotation for untargeted metabolomics

doi:10.21203/rs.3.rs-8174529/v1

Download PDF

Research Article

MS-Net: Multi-Similarity based network annotation for untargeted metabolomics

https://doi.org/10.21203/rs.3.rs-8174529/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Confident metabolite annotation remains a critical bottleneck in untargeted LC-MS metabolomics, as experimental spectral libraries cover only 5–20% of detected features. While in silico tools generate extensive candidate lists, top-ranked predictions often fail to reflect true identities, resulting in high false annotation rates. We present MS-Net (Multi-Similarity Network-based annotation), an accessible workflow integrating mass spectral similarity networks, molecular structure similarity (Tanimoto metrics), and taxonomic knowledge to prioritize annotations within vast candidate spaces. MS-Net employs a composite Link Score combining full-molecule and scaffold Tanimoto similarities with MS/MS cosine similarity and in silico confidence metrics. High-confidence annotations seed iterative propagation throughout the network. Applied to a Cannabis sativa dataset (2,595 initial features reduced to 1,297 after filtering, from 118,000 candidates), MS-Net resolved the annotation space to 1,275 confident assignments. notably, 53% of final annotations were rescued from lower in silico ranks (2–50), demonstrating the algorithm's ability to correct ranking errors. The workflow enables reproducible, offline annotation prioritization suitable for systems biology integration.

Structural Biology

Analytical Biochemistry

Computational Biology

metabolomics

mass spectrometry

annotation

molecular networking

Tanimoto similarity

natural products

KNIME

Cannabis

cannabinoids

The increasing application of metabolomics, both as a standalone discipline and integrated with other omics, underscores its maturity¹. Untargeted LC-MS fingerprints from biological matrix yield hundreds to thousands of features (mz x RT pairs). One of the main bottlenecks is properly annotating all detected signals with sufficient confidence to ensure robust biological contextualization (e.g. pathway enrichment analysis, multi-omics inference models…).

A dereplication strategy involves signal deconvolution, ions linkage deciphering using tools such as Ion Identity Network² or MS-CleanR³ followed by annotation relying on a “body of evidence” approach. The formalization of annotation confidence Levels was proposed by the Metabolomic Standard Initiative⁴ and refined to account for new online or computational approaches⁵. Level 1 relies on an authentic standard match, while Level 2 involves a spectral match from external libraries. The above annotation Levels include in silico approaches (Level 3) up to Level 4 (only partial match) and Level 5 for unknown spectroscopic signals. The main limiting step is the discrepancy between publicly available mass spectral libraries and the chemical diversity encountered in biological samples. For instance, the last version of FragHub mass spectral library integration⁶ (10.5281/zenodo.17235587, v12), encompasses two million spectra for 150,000 unique INCHIKEY, while chemical space produced by living organisms is estimated to be millions of molecules⁷. As a result, experimental spectral matches (Levels 1 and 2) cover only 5 to 20% of all detected signals in untargeted experiments. The remaining signal annotations are based on chemical catalog queries using in silico fragmentation approaches (Level 3). For each feature, putative matches are ranked according to a specific scoring system depending on the software used. For instance, MS-Finder⁸ leverages rule-based fragmentation to predict fragments from SMILES structures, and MetFrag⁹ employs combinatorial bond breaking. It ranks putative structures based on weighted sum fragment matches, intensity correlation, and neutral losses. Sirius-CSI¹⁰ combines fragmentation trees with machine learning to predict molecular fingerprints, ranking candidates via Bayesian probability and fingerprint similarity. The identification accuracy reached between 10 and 30% for the first hit (Top 1) to 90% for the top 20 according to the CASMI challenge¹¹ and other benchmark¹². As a result, despite the advancements in computational tools for metabolite annotation, these solutions may lead to numerous false annotations in the context of untargeted LC-MS metabolomics, dealing with hundreds to thousands of features. Several strategies have been proposed to address this challenge based on mass spectral similarity networks provided by the GNPS facility².

For instance, Network Annotation Propagation (NAP)¹³, leverages molecular network topology to re-rank structural candidates even without spectral library matches. Taxonomically Informed Metabolite Annotation (TIMA)¹⁴ improves confidence in annotations by considering the taxonomic position of biological sources. ConCISE¹⁵ integrates molecular networking, spectral library matching, and in silico class predictions to provide accurate classifications for subnetworks. Inventa¹⁶ calculates a priority score to highlight structural novelty within extracts. More recently, MS2DECIDE¹⁷ aggregates data from GNPS, Sirius, and ISDB-LOTUS¹⁸ using multi-criteria decision analysis to prioritize annotation and highlight feature novelty potential. Another avenue has been proposed using multilayer networks, which combines knowledge-based metabolic reaction, correlation networks, and mass spectral similarity to enable global metabolite annotation from knowns to unknowns¹⁹. These approaches collectively aim to streamline feature identification by leveraging computational insights and expert knowledge to enhance the efficiency and confidence of MS/MS-based annotations.

Although these network-based strategies have enhanced annotation confidence, their adoption in routine metabolomics workflows faces practical barriers, including requirements for bioinformatics expertise, dependencies on web servers or dedicated computational infrastructure, and outputs with limited chemical metadata that restrict seamless integration with pathway enrichment analysis and or multi-omics approaches²⁰.

To address this challenge, we propose MS-Net (Multi-Similarity Network-based annotation). This user-friendly workflow combines mass spectral and Tanimoto similarity networks with correlation analysis and taxonomic data to drive annotation prioritization. Additionally, MS-Net integrates positive and negative ionization chromatographic fingerprints with several filtering options. Finally, the workflow generates enriched metadata annotation, statistical tables, and map results on mass spectral and Tanimoto similarity networks. This approach is completely offline and compatible with output files from widely used software in the field, such as MS-Dial, MZmine, Sirius-CSI, and MS-Finder (Fig. 1).

Workflow Architecture and Implementation

MS-Net was developed using the Knime visual programming interface²¹ and needs three primary inputs: (1) feature intensity tables (peak height or area), (2) mass spectral similarity (MSS) network edge lists, and (3) multi-Level annotation hits encompassing experimental library matches (Levels 1-2) and in silico predictions (Levels 3-4). These last collectively define a putative chemical space that may comprise several thousand candidate structures per dataset (Figure 1A,B).

Annotation Confidence Scoring System

The workflow begins by normalizing confidence scores across annotation Levels to establish a unified ranking system. Level 1 receives the highest confidence score and is assigned to features showing perfect concordance with authentic standards based on MS/MS spectral similarity (similarity > 0.95) and retention time agreement (ΔRT < 0.2 min). Level 2 annotations derive from spectral library matches and are subdivided according to matching quality: Level 2a for high-confidence matches (similarity > 0.85) and Level 2b for moderate-confidence matches (similarity > 0.7). For in silico annotations, MS-Net enriches structural candidates with taxonomic information by querying Coconut 2.022. The workflow performs InChIKey-based matching to identify compounds reported from user-specified taxonomic sources (genus and family Levels). Candidates with confirmed biosource origins or matching the target chemical class are assigned Level 3a, while remaining in silico matches receive Level 3b. Spectral library matches exhibiting lower similarity (similarity < 0.7) or significant precursor mass discrepancies are classified as MS/MS analogs (Level 4). Features lacking any structural annotation remain at Level 5 (unknown) (Figure 1B). Confidence scores are normalized to a 0–100 scale to enable cross-Level comparison. The normalization scheme employed Level-specific transformations: Level 1 = 95 + (similarity score− 0.95) × 100; Level 2a = 85 + (similarity score− 0.85) × 100; Level 2b = 70 + (similarity score − 0.70) / 0.15 × 14; Level 3b (in silico and de novo) = 40 + confidence score × 40; Level 4 (MS/MS analogs) = 30 + (similarity score − 0.5) × 90; Level 5 = 0. These transformations ensure that experimental matches consistently receive higher scores than computational predictions while preserving score discrimination within each Level.

Feature Filtering and Data Reduction

Prior to network-based annotation propagation, users may optionally apply filtering strategies to reduce dataset complexity while retaining chemically informative features. The workflow supports retention time clustering using the MS-CleanR³ algorithm. Within each RT cluster, the most intense features and/or those with the highest network connectivity (degree) are preferentially retained, effectively removing redundant signals (Figure 1C). Optionally, a cosine filter threshold can be added to constrain the MSS network. Finally, putative annotations between two nodes may be filtered according to XlogP calculated for each candidate. For each feature pair, the XlogP is compared to the edge delta retention time. In C18 mode, only feature pairs exhibiting XlogP trends consistent with retention time order are conserved.

Network-Based Annotation Propagation

MS-Net constructs a seed subnetwork comprising only high-confidence annotations (Levels 1, 2a, and 3a), which serves as the foundation for propagating structural assignments throughout the entire MSS network. To rank competing structural candidates for each feature pair connected in the MSS network, we developed a composite scoring metric that integrates spectral, structural, and computational evidence into a unified Link Score. The structural similarity component employs two complementary Tanimoto measures calculated from molecular fingerprints: Tanimoto_Full (based on Morgan, PubChem or RDkit) captures overall molecular similarity, including substituents, while Tanimoto_Murcko (scaffold fingerprints) emphasizes core structural frameworks. These metrics are dynamically weighted according to their relative informativeness. When the absolute difference between these two measures exceeds 0.1 (|Tanimoto_Full - Tanimoto_Murcko| > 0.1), the higher value receives greater weight, reflecting either the dominance of substituent patterns (Tanimoto_Full) or core scaffold similarity (Tanimoto_Murcko). When this difference falls within ±0.1, equal weights are applied. The resulting Combined_Tanimoto is then scaled by the MS/MS cosine similarity to produce the Adjusted_Structural score, which accounts for both molecular structure and spectral concordance. In parallel, an InSilico_Combined score is calculated as the arithmetic mean of confidence scores described above. The final Link Score integrates these two components through a user-tunable weighting parameter:

Link Score = (1 - α) × Adjusted_Structural + α × InSilico_Combined

The parameter α allows users to control the relative contributions of spectral-structural evidence versus in silico prediction scores. Lower α values (e.g., 0.3) prioritize structural and spectral similarity, while higher values give more weight to computational prediction confidence. For each feature pair in the MSS network, the workflow evaluates all candidate structures and selects the annotation with the highest Link Score. This process propagates iteratively from high-confidence seeds to their direct neighbors and subsequently through the entire connected network. Features that remain outside the MSS network are ranked using a simplified metric combining their best in silico score with the maximum Tanimoto similarity to any annotated compound within the MSS network.

Metadata Enrichment and Output Generation

The final annotated feature list is optionally enriched with chemical metadata using ClassyFire and NPClassifier ontologies, providing hierarchical chemical classifications (kingdom, superclass, class, subclass). Database identifiers for each annotated feature are retrieved using the Chemical Translation Service²³, ensuring compatibility with downstream pathway enrichment tools and multi-omics integration platforms (Figure 1E).

MS-Net generates four primary outputs: (1) a comprehensive annotated feature table with confidence Levels, structural information, and chemical ontology; (2) a feature height/area table; (3) an MSS network edge table for visualizing spectral similarity relationships; and (4) a Tanimoto-based network edge table connecting structurally related compounds, including links between unknown features and their nearest annotated structural neighbors. This latter output enables users to infer structural motifs for unannotated features. Both networks are enriched with putative chemical reactions between two neighboring nodes using a delta mass match to a predefined list from Metanetter 2²⁴. Optionally, features acquired in positive and negative ionization modes can be merged based on user-defined retention time and m/z tolerances.

Application to Cannabis Metabolomics Dataset

Inflorescences of three medical-grade Cannabis sativa L. chemotypes were selected to evaluate MS-Net's annotation capabilities. Cannabis represents an ideal model system for several reasons. First, the species exhibits remarkable chemical diversity, encompassing a wide array of cannabinoids, terpenes, and phenolic compounds (primarily flavonoids and hydroxycinnamic acids)^25–27. The phytochemistry of cannabis is well-documented, with numerous metabolomic studies demonstrating robust discrimination among cultivars and chemotypes^28–30.

The traditional morphology-based classification (Indica vs. Sativa) has been superseded by a chemotype system based on the relative concentrations of Δ9-tetrahydrocannabinol (THC) and cannabidiol (CBD). This framework defines three main chemotypes: Type I (THC-predominant, <0.5% CBD), Type II (balanced THC:CBD ratio), and Type III (CBD-predominant, <1% THC). Although expanded classifications include Type IV (cannabigerol-predominant) and Type V (cannabinoid-free), Types I–III remain the most extensively characterized^31,32.

Crucially for algorithm evaluation, cannabinoids exhibit exceptional structural diversity—estimated at 120–150 distinct phytocannabinoids—while sharing highly similar core molecular scaffolds³³. For instance, the major cannabinoids THC, CBD, and cannabichromene (CBC) are constitutional isomers (C₂₁H₃₀O₂) differing only in cyclization patterns: THC features a pyran ring, CBD a cyclohexene ring, and CBC a benzopyran structure. This structural similarity, combined with well-elucidated biosynthetic pathways, provides an ideal benchmark for evaluating network-based annotation algorithms. The biosynthetic pathway initiates with cannabigerolic acid (CBGA)—formed by CBGA Synthase-catalyzed condensation of olivetolic acid and geranyl pyrophosphate—which serves as the universal precursor for nearly all other cannabinoids³⁴. Additionally, in planta, cannabinoids exist predominantly as carboxylic acids (THCA, CBDA, CBCA), with decarboxylation to neutral forms occurring upon heating.

To demonstrate MS-Net's capabilities, we applied the workflow to a comprehensive untargeted metabolomics study of Cannabis sativa L., analyzing three distinct chemotypes: Bedrocan® (THC-dominant, Type I), Bedrolite® (CBD-dominant, Type III), and Bediol® (THC/CBD-balanced, Type II). Initial data processing detected 2,595 features across positive and negative ionization modes.

Feature Filtering and Chemical Space Reduction

We applied a sequential filtering strategy to reduce dataset complexity while retaining chemically meaningful signals. First, Ion Identity Networking identified and collapsed redundant adducts and isotopes by selecting the most informative precursor ions ([M+H]⁺ or [M-H]⁻) from each ion cluster. Subsequently, MS-CleanR-based retention time clustering further consolidated co-eluting features (Figure 2B). This filtering reduced the dataset to 1,297 unique features while maintaining comprehensive chemical coverage.

In silico annotation using Sirius-CSI (top 50 candidates per feature) and MSNovelist (top 20 de novo structures per feature) generated a putative chemical space encompassing more than 118,000 candidate structures (Figure 2A). This expansive search space highlights the challenge of confident structural assignment in untargeted metabolomics: without prioritization strategies, the likelihood of selecting incorrect annotations from such large candidate pools is substantial.

Network-Based Annotation Prioritization

We seeded the MSS network using Level 1 (authentic standards), Level 2a (high-confidence spectral matches), and Level 3a annotations (taxonomically informed candidates from the Cannabis genus, Cannabaceae family, or cannabinoid chemical class). PubChem-based fingerprint was selected to calculate Tanimoto similarities. The Link Score algorithm was configured with α = 0.3 to prioritize structural and spectral similarity over raw in silico ranking scores.

To evaluate the algorithm's performance, we examined the agreement between MSS network topology (cosine similarity) and structural similarity (Tanimoto scores). Before annotation prioritization, the raw chemical space exhibited a mean absolute distance of 0.55 between these metrics, with the highest density occurring between 0.6 and 0.8 (Figure 2C). This discrepancy reflects that spectral similarity does not always correlate with structural similarity, particularly when in silico tools generate diverse candidates. Restricting to only the top-ranked in silico candidate (top 1) dramatically reduced the mean distance to 0.3, but at the cost of excluding potentially correct structures ranked lower. Expanding to the top 10 or top 20 candidates achieved better performance, with maximum density centered around 0.2, indicating strong agreement between spectral and structural similarities. Notably, incorporating the top 10 de novo structures from MSNovelist further improved concordance, suggesting that machine learning-generated candidates can complement database-constrained searches for features representing novel or underrepresented chemical scaffolds. Finally, top 50 in silico and Top 20 de novo candidates per feature were selected for annotation prioritization.

Global Dataset Annotation and Chemical Space Coverage

MS-Net reduced the initial chemical space from 118,000 candidates to 1,275 confidently annotated compounds across 1,297 features (Figure 2D). Analysis of annotation rank distribution revealed that 47% of features were assigned their top-ranked in silico candidate, indicating strong agreement between computational predictions and network-guided prioritization. An additional 30% of annotations fell within ranks 2–20, demonstrating the algorithm's ability to rescue correct structures initially ranked lower due to limitations in in silico fragmentation models. The remaining 23% of annotations were ranked above position 20 (Figure 2F).

The final annotation distribution by confidence Level showed: 9 Level 1 (authentic standards), 58 Level 2a (high-confidence spectral matches), 31 Level 2b (moderate-confidence spectral matches), 43 Level 3a (taxonomically informed in silico annotations), 1051 Level 3b in silico and 71 3b de novo matches, 4 Level 4 (MS/MS analogs), and 26 Level 5 (unknown) (Figure 2E). This distribution reflects the typical annotation coverage achievable in specialized metabolomics studies, where experimental spectral libraries cover only a fraction of detected features, necessitating extensive in silico inference.

Chemotype Discrimination

Principal component analysis (PCA) of the annotated feature matrix (n = 18 samples, p = 1,297 features) revealed clear separation of the three cannabis chemotypes, with the first two principal components explaining 90 % of total variance (Figure 2G). Sparse partial least squares discriminant analysis (sPLS-DA) identified 60 discriminant features that robustly distinguished the chemotypes (Figure 2H).

Chemical ontology classification using NPClassifier revealed distinct natural product pathway enrichments for each chemotype (Figure 2I). Bedrocan® (THC-dominant) exhibited enrichment in phenylpropanoids and terpenoid pathways, consistent with high Levels of THC and related cannabinoids. Bedrolite® (CBD-dominant) showed elevated Levels of amino acid derivatives and shikimate pathways. Bediol® (balanced THC/CBD) displayed an intermediate metabolite profile with enrichment in polyketide derivatives.

Case Study: Cannabinoid Subnetwork Annotation

The initial MSS network is seeded with annotation Level 1, 2a and 3a (green dots, figure 3A). The cannabinoid MSS subnetwork illustrates the algorithm's discriminative power (Figure 3B). MS-Net successfully prioritized structurally coherent annotations, including close derivatives differing primarily in hydroxylation patterns, methyl substitutions, or double bond positions—chemical variations consistent with known cannabinoid biosynthetic pathways. Two de novo structures from MSNovelist were also integrated, representing potential novel cannabinoid scaffolds warranting further investigation (Figure 3B). An illustration of MS-net prioritization algorithm is displayed between cannabichromenic acid (CBC-A, Level 1 authentic standard) and its close neighbor, accounting for annotation le 3b (Figure 3C), displaying a pseudomolecular ion at m/z 301.145 (molecular formula C₁₈H₂₂O₄, 0.9 ppm mass accuracy), which matched 70 putative in silico candidates. Reliance solely on the top-ranked in silico candidate from Sirius-CSI would have resulted in an incorrect structural assignment. However, the Link Score algorithm identified cannabiorcichromenic acid, originally ranked 50th, as the most probable annotation based on its high Tanimoto similarities to CBC-A, which is consistent with the MSS cosine similarity of 0.89.

Exploitation of the Tanimoto structural similarity network through the study of cannabinoid biosynthesis pathway

Figure 4 illustrates how the MSNet workflow efficiently clusters metabolites according to their structural similarities, providing a comprehensive view of the metabolic landscape. The upper panel (A) presents the complete Tanimoto structural similarity network. Within this network, each node represents a distinct metabolite, and the edges delineate the Tanimoto similarity score between them. The organization of the network reveals well-defined clusters of metabolites corresponding to distinct biosynthetic routes, reflecting the algorithm’s capacity to capture pathway-level organization within complex metabolomic data. Overall, the similarity network is organized into three main subnetworks, five large clusters, and several smaller satellite clusters. An expanded sub-network (B) was isolated by selecting the cannabinoids and their precursors, as well as their close neighbors. This network highlights a local region of the network, illustrating how cannabinoids and precursors with related chemical structures or biosynthetic origins are tightly interconnected. The lower panel (C) focuses on the three main biosynthetic pathways that specifically yield cannabinoids: the olivetolic acid, orsellinic acid, and divarinic acid pathways. Each pathway is represented as a sequence of enzymatic and chemical reactions converting early intermediates into key cannabinoids. The three precursors associated with the three main biosynthetic pathways were detected within the extracts, along with the majority of metabolites produced via the olivetolic acid pathway and a subset originating from the divarinic and orsellinic acid pathways.

MS-Net was designed to streamline the LC-MS-based untargeted metabolomic workflow from raw data acquisition to a complete annotated table enriched with metadata based on outputs from processing software commonly used in the field. Our results demonstrate that MS-Net not only provides confident structural annotations but also generates biologically interpretable outputs with enriched metadata suitable for downstream pathway enrichment analysis and multi-omics integration.

The core algorithm of MS-Net leverages Tanimoto similarity metrics starting from confidently annotated nodes (Level 1, 2a, and 3a), which serve as seeds to propagate annotations into the MSS networks. Consequently, the results will be impacted by the quality of annotation used as input. Annotation Level 1 is dependent on the internal library and generally limited to synthetic standards. Annotation Level 2a, encompassing high experimental MS/MS similarity scores, will be influenced by the quality and coverage of the mass spectral library used. Initiatives such as FragHub⁶, MSnLib³⁵, MassBank³⁶, or GNPS², dramatically improved matching results. However, metabolome coverage of mass spectral libraries is still limited compared to natural product catalogs³⁷. To extend putative seeds among the MSS network, we implement InChiKey-based matches against the Coconut 2.0²² chemical catalog or user-choice chemical ontology, which increases the number of seeds to more than 90 compounds in the case of Cannabis extract profiling (Supplementary data: ResultsNEGPOS.xlsx). This well-studied plant³⁸ is perfectly suitable for the MS-Net approach, while understudied organisms may suffer from less accurate annotation results.

The type of molecular fingerprint used highly influences the Tanimoto-based similarity result³⁹. Three contrasted type of fingerprints are available within MS-Net: Morgan fingerprints which capture local circular substructures with a radius of 2 by default to provide highly discriminative power for structural similarity, RDKit (Daylight-like) encodes molecular paths through graph providing a balanced general-purpose representation, while PubChem uses 881 predefined substructural features that tend to emphasize common functional groups and yield higher similarity scores between diverse molecules. This last have been used for the present study, and according to our experience, Morgan or Pubchem molecular fingerprints allowed the detection of subtle substructure changes compared to RDKit type.

The MS-Net approach is also highly dependent on the MSS network topology. A poorly clustered network, constructed with a low similarity threshold (e.g., 0.6), may result in spurious results. For the Cannabis dataset, the modified cosine score algorithm with a threshold of 0.7 was used, which provided an equilibrated clustered MSS network (Fig. 3A). Still, changing spectral clustering measures, such as Bonanza available in MSdial, impacts the final results (data not shown).

In the frame of Cannabis dataset, the homogeneous distribution of compounds across biosynthetic pathways within the similarity network demonstrates that MSNet is effective not only in establishing structural relationships among compounds but also in grouping them coherently (Fig. 4A). Among the annotated compounds, cannabinoids represent a chemically important class. Cannabinoids are secondary metabolites biosynthesized in the glandular trichomes of female cannabis inflorescences through three well-characterized biosynthetic routes: the olivetolic acid pathway, the divarinic acid pathway, and the orsellinic acid pathway. The MSNet annotation workflow successfully identified the three precursors of these metabolic routes (olivetolic, divarinic, and orsellinic acids), along with a series of downstream cannabinoids. The olivetolic acid pathway, for example, was almost fully reconstructed, as shown in Fig. 4C. Notably, compounds with nearly identical structures were accurately annotated, such as Δ⁹-THC and Δ⁸-THC—two isomers with identical molecular masses. Three cannabinoids associated with the other two biosynthetic routes were annotated. It is important to emphasize that these results reflect the metabolic fingerprint of the polar phases of cannabis inflorescence extracts at the time of extraction. Consequently, cannabinoids that were not annotated (non-circled) may have been absent from the extracts due to degradation (either in the extracts or in the inflorescences) or may not have been biosynthesized by the plant. (In planta, cannabinoids are predominantly produced in their acidic forms and are subsequently converted into their neutral analogues through decarboxylation induced by light or heat.) Overall, these findings demonstrate that MSNet is an effective tool to capture known biosynthetic pathways among the initial annotation chemical space.

Sample preparation process

Three cultivars of medicinal C. sativa L. dried female inflorescences were purchased from Bedrocan International (Veendam, Netherlands) including a THC-dominant type Bedrocan® variety (batch: 20C30EY20E13), a CBD-dominant type Bedrolite® variety (batch: 20I14FR20L02) and a THC/CBD-intermediate type Bediol® variety (batch: 19L16FB20K04) according to OMC (Office of Medicinal Cannabis, Netherlands) and ANSM (Agence Nationale de Sécurité du Médicament et des produits de santé, France) requirements and authorization. Samples were stored in airtight containers in the dark at 25 °C.
For each variety, 500 mg of material were weighed, flash-frozen in liquid nitrogen, and finely ground using a pre-chilled mortar and pestle. The resulting frozen powder was extracted at 25 °C using a biphasic solvent system. Briefly, 1.5 mL of a methyl tert-butyl ether/methanol (75:25 v/v) mixture was added to the sample, followed by sonication in an ultrasonic bath for 3 min. Subsequently, 1.5 mL of a water/methanol (75:25 v/v) mixture was added, and the sample was sonicated again under identical conditions. The extracts were then centrifuged at 240 g for 3 min at 5 °C and the resulting supernatants were collected. Both the polar and apolar fractions were evaporated using a SpeedVac SPD vacuum concentrator (Thermo Fisher Scientific) at 35 °C and 50 mbar. The dried residues were subsequently reconstituted in acetonitrile/water (50:50 v/v) to a final concentration of 10 mg/mL. Extractions were performed in triplicate. The extracts were stored at -20 °C in amber glass bottles between each use.

Standards preparation

A Level 1 library was constructed by injecting analytical grade standards of cannabinoids, which were prepared from commercially available solutions at 1 mg/mL of cannabigerolic acid (CBGA), cannabigerol (CBG), cannabidiolic acid (CBDA), cannabidiol (CBD), cannabichromenic acid (CBCA), cannabichromene (CBC), Δ-9-tetrahydrocannabinolic acid (d9THCA), Δ-9-tetrahydrocannabinol (d9THC), Δ-8-tetrahydrocannabinol (d8THC), tetrahydrocannabivarin (THCV), cannabidivarinic acid (CBDVA), cannabidivarin (CBDV), cannabicyclol (CBL) were purchased from Cerilliant (Austin, Texas, USA). Cannflavin A at 1 mg/mL was obtained from Sigma-Aldrich (Darmstadt, Germany). All standard solutions were diluted to a concentration of 12 μg·mL⁻¹ in acetonitrile.

UHPLC-HRMS/MS analysis

Chromatographic separation was carried out on a Vanquish UHPLC system (Thermo Fisher Scientific, Waltham, MA, USA) equipped with a Luna Omega Polar C18 analytical column (150 × 2.1 mm, 1.6 µm; Phenomenex, Torrance, CA, USA). The chromatographic system was coupled with a Vanquish Diode Array Detector (DAD), a Charged Aerosol Detector (CAD) and a Q-exactive Plus mass spectrometer (Thermo Fisher Scientific, Waltham, MA, USA). A dual-pump system was installed, where the first pump managed the separation at the column Level, and the second pump system generated a counter-gradient at the column outlet to maintain a constant mixture of 50% water and 50% acetonitrile. The separation was performed at a flow rate of 0.4 mL·min⁻¹ and at a column oven temperature of 40 °C, using a mobile phase composed of H₂O + 0.05% formic acid (A) and acetonitrile + 0.05% formic acid (B). A 1 µL injection of extracts at 10 mg·mL⁻¹ was introduced into the system. The applied gradient was as follows: 0 to 0.5 min, 2% B; 0.5 to 18 min, 98% B; 18 to 21 min, 98% B; 21 to 21.5 min, 2% B; and 21.4 to 24 min, 2% B.

Mass spectrometry detection was performed using a Q Exactive Plus instrument equipped with a heated electrospray ionization source (HESI-II) operating in positive and negative ionization modes with a resolution of 35,000 (MS1) and 17,500 (MS2). The collision energy was set to 10, 20, and 40 eV in stepped mode for data-dependent MS/MS acquisition. The capillary temperature was 300°C, and the mass scan range was 100–1,500 m/z for both MS1 and MS2. Data-dependent acquisition targeted the four most intense precursor ions per scan cycle.

Data Processing and Feature Detection

Raw LC-MS data files were processed using MZmine (version 4.7, mzio GmbH). Mass detection was performed separately for MS1 and MS2 scans using the "Factor of lowest signal" algorithm with noise factors of 5.0 and 2.5, respectively. Chromatographic peaks were built using the ADAP chromatogram builder with the following parameters: minimum consecutive scans = 4, minimum intensity for consecutive scans = 10,000, minimum absolute height = 50,000, and m/z tolerance of 0.002 Da or 10 ppm (scan-to-scan). Chromatograms were smoothed using Savitzky-Golay filtering (5-scan window in the retention time dimension). Feature deconvolution was performed using the local minimum resolver with a minimum absolute height of 2.E6 and 1.E6 in positive and negative mode respectively, peak top-to-edge ratio of 1.8, peak duration range of 0.04–3.0 min, and minimum of 5 data points per feature. MS/MS spectra were paired to features using a precursor m/z tolerance of 0.01 Da or 10 ppm, retention time tolerance matching feature edges, and minimum relative feature height of 0.25. Isotope grouping was performed using the ¹³C isotope filter with m/z tolerance of 0.005 Da or 10 ppm, retention time tolerance of 0.05 min, monotonic shape validation, and maximum charge state of 2. Ion Identity Networking (IIN) was performed to identify adducts, in-source fragments, and neutral losses. Network refinement was applied with minimum network size of 1, deletion of networks lacking monomer ions, and removal of networks with fewer than 2 links. Retention time correction was performed using aligned features as reference points (m/z tolerance: 0.005 Da or 10 ppm; RT tolerance: 0.1 min; minimum standard intensity: 500,000). Feature lists were aligned using the join aligner with m/z tolerance of 0.0015 Da or 10 ppm (sample-to-sample), RT tolerance of 0.15 min, and m/z weight-to-RT weight ratio of 3:1. Alignment incorporated isotope pattern comparison (minimum score: 0.8) and MS/MS spectral similarity (weighted cosine ≥ 0.7) when available. Gap filling was performed using the multithreaded peak finder with intensity tolerance of 20%, m/z tolerance of 0.002 Da or 10 ppm, RT tolerance of 0.1 min, and minimum of 3 scans per feature. Duplicate features were filtered using an averaging mode with m/z tolerance of 0.005 Da or 10 ppm and RT tolerance of 0.1 min. Blank subtraction was applied requiring minimum detection in 2 blank samples, with features retained only if they showed ≥2-fold higher intensity in samples compared to blanks (based on maximum blank intensity). Final feature filtering retained only features meeting all of the following criteria: detected in at least 6 samples within any sample class (out of 6 replicates per chemotype), presence of at least 2 isotopic peaks with validated ¹³C pattern and presence of MS/MS spectrum. Spectral networking was performed using the modified cosine algorithm with the following parameters: m/z tolerance = 0.003 Da or 10 ppm, maximum precursor m/z delta = 600 Da, minimum matched signals = 4, minimum cosine similarity = 0.6. Our In-house library of cannabinoids standards was used with a minimum cosine similarity = 0.85, precursor m/z tolerance = 0.01 Da or 20 ppm, spectral m/z tolerance = 0.015 Da or 25 ppm, RT tolerance = 0.2 min, minimum matched signals = 4 using weighted cosine similarity (MassBank weighting), removed precursor ions, performed ¹³C deisotoping (m/z tolerance: 0.001 Da or 10 ppm, monotonic shape, max charge: 2), and cropped spectra to m/z overlap regions. The feature quantification table, mass spectral similarity network and annotation table were exported for further MS-Net processing. Processed MS/MS data were exported in .mgf format for further Sirius processing.

Corresponding MZbatch used for MZmine processing are available here: https://zenodo.org/records/17669288

Sirius-CSI annotation

Data in MGF format were imported into SIRIUS-CSI (v6.3, Lehrstuhl Bioinformatik Jena). Molecular formula prediction was performed using pre-configured Orbitrap parameters with allowed elements limited to C, H, N, O, P, and S, using a "database search" strategy. CSI:FingerID structural prediction queried PubChem as the primary database with bio-databases enabled for further taxonomically enriched matches. FragHub spectral databases were included for experimental spectral matching and analog detection using a precursor m/z tolerance of 20 ppm. Only features achieving minimum confidence scores <0.2 in the initial annotation round were selected for subsequent de novo structure prediction using MSNovelist. Results were exported in CSV format with quoted strings, retaining the top 50 ranked candidates per feature for both in silico and de novo matches.

MS-Net parameters

Output files from MZmine (feature quantification table, MSS network edge list, spectral library annotations) and SIRIUS-CSI (top 50 in silico candidates and top 20 de novo structures from MSNovelist per feature) were imported into MS-Net (developed in Knime Analytics Platform, v5.2). Features matching authentic standards (spectral cosine > 0.95, ΔRT < 0.2 min) were assigned Level 1. Spectral library matches were stratified as Level 2a (cosine > 0.85) or 2b (0.7 < cosine ≤ 0.85). In silico candidates were enriched with taxonomic information by querying Coconut 2.0 (linked to NCBI taxonomy) using InChIKey-based matching. Among the top 10 in silico candidates per feature, those matching taxonomic criteria (Cannabis genus, Cannabaceae family) or classified as cannabinoids by NPClassifier were elevated to Level 3a; remaining in silico matches received Level 3b. Spectral library matches with cosine < 0.7 or significant precursor mass errors were classified as Level 4 (MS/MS analogs). Features were filtered using Ion Identity Network (IIN) results to remove redundant adducts and isotopes, followed by MS-CleanR-based RT clustering (ΔRT ≤ 0.01 min). Within each cluster, the top 2 features by network degree and the top 2 by peak intensity were retained. The MSS network was constrained to edges with cosine similarity ≥ 0.7 and ΔRT ≤ 8 min between connected nodes. High-confidence annotations (Levels 1, 2a, 3a) seeded the MSS network for iterative annotation propagation. For each feature pair, candidate structures were ranked using the Link Score formula: Link Score= (1 - α) × Adjusted_Structural + α*× InSilico_Combined, where Adjusted_Structural integrates dynamically weighted Tanimoto similarities (full-molecule and Murcko scaffold fingerprints-PubChem) modulated by MS/MS cosine similarity, and InSilico_Combined represents the mean of available in silico confidence scores. The weighting parameter was set to α = 0.3, prioritizing structural-spectral evidence (70%) over in silico ranking (30%). The top 5 candidates by Link Score were retained per feature before looping through all MSS network. Redundant annotations with identical InChIKey identifiers and pearson correlation among samples >0.7 were consolidated by selecting the candidate with the highest mean peak height. Positive and negative mode feature lists were merged using ΔRT ≤ 0.05 min, Δm/z ≤ 0.002 Da, and a minimum Pearson correlation ≥ 0.6 across sample intensities. Final annotations were enriched with chemical ontology classifications from ClassyFire (kingdom, superclass, class, subclass) and NPClassifier (pathway, superclass, class). Database identifiers (PubChem CID, KEGG, HMDB, ChEBI) were retrieved using the Chemical Translation Service. Natural product-likeness scores were calculated using the NPlikeness calculator. Structural similarity networks were constructed using Tanimoto coefficient ≥ 0.8, retaining the top 2 nearest neighbors per feature.

Statistical analysis

Multivariate data analysis was performed using the mixOmics R package v6.3.2 through the MetaboStat_AgX Shiny interface (available at https://zenodo.org/records/17352817). Prior to analysis, missing values were imputed using the global minimum intensity value divided by 10. Feature intensities were normalized using total sum normalization (TSN), followed by unit variance scaling (autoscaling). Principal component analysis (PCA) was performed to assess overall variance structure and sample clustering. Partial least squares discriminant analysis (PLS-DA) models were validated using 10-fold cross-validation repeated 100 times to assess classification performance and prevent overfitting. Model quality was evaluated using the classification error rate. Sparse PLS-DA (sPLS-DA) was applied to identify discriminant features, selecting the top 30 features per component (60 total across two components) based on variable importance in projection (VIP) scores. Hierarchical clustering (Euclidean distance, Ward's linkage) of the top 60 discriminant features generated the clustered heatmap visualization. Chemical ontology enrichment analysis was performed on features characteristic of each chemotype according to their coefficient score using NPClassifier pathway classifications. The results were visualized as pie charts showing pathway distribution per chemotype.

Code availability

MS-Net workflow, tutorials and MZbatch files used for this study are available here: https://zenodo.org/records/17669288

Data availability

Raw LC-MS data acquired in positive and negative mode are available here: https://zenodo.org/records/17671960

Borges, R. M. et al. Quantum Chemistry Calculations for Metabolomics: Focus Review. Chem. Rev. 121, 5633–5670 (2021).
Schmid, R. et al. Ion identity molecular networking for mass spectrometry-based metabolomics in the GNPS environment. Nat. Commun. 12, 3832 (2021).
Fraisier-Vannier, O. et al. MS-CleanR: A Feature-Filtering Workflow for Untargeted LC–MS Based Metabolomics. Anal. Chem. 92, 9971–9981 (2020).
Sumner, L. W. et al. Proposed minimum reporting standards for chemical analysis: Chemical Analysis Working Group (CAWG) Metabolomics Standards Initiative (MSI). Metabolomics 3, 211–221 (2007).
Charbonnet, J. A. et al. Communicating Confidence of Per- and Polyfluoroalkyl Substance Identification via High-Resolution Mass Spectrometry. Environ. Sci. Technol. Lett. 9, 473–481 (2022).
Dablanc, A. et al. FragHub: A Mass Spectral Library Data Integration Workflow. Anal. Chem. acs.analchem.4c02219 (2024) doi:10.1021/acs.analchem.4c02219.
Medema, M. H., de Rond, T. & Moore, B. S. Mining genomes to illuminate the specialized chemistry of life. Nat. Rev. Genet. 22, 553–571 (2021).
Tsugawa, H. et al. Hydrogen Rearrangement Rules: Computational MS/MS Fragmentation and Structure Elucidation Using MS-FINDER Software. Anal. Chem. 88, 7946–7958 (2016).
Ruttkies, C., Schymanski, E. L., Wolf, S., Hollender, J. & Neumann, S. MetFrag relaunched: incorporating strategies beyond in silico fragmentation. J. Cheminformatics 8, 3 (2016).
Hoffmann, M. A. et al. High-confidence structural annotation of metabolites absent from spectral libraries. Nat. Biotechnol. 40, 411–421 (2022).
CASMI 2022. Revisiting CASMI https://fiehnlab.ucdavis.edu/casmi (2022).
Zhu, B. et al. Knowledge-based in silico fragmentation and annotation of mass spectra for natural products with MassKG. Comput. Struct. Biotechnol. J. 23, 3327–3341 (2024).
da Silva, R. R. et al. Propagating annotations of molecular networks using in silico fragmentation. PLOS Comput. Biol. 14, e1006089 (2018).
Rutz, A. et al. Taxonomically Informed Scoring Enhances Confidence in Natural Products Annotation. Front. Plant Sci. 10, 1329 (2019).
Quinlan, Z. A. et al. ConCISE: Consensus Annotation Propagation of Ion Features in Untargeted Tandem Mass Spectrometry Combining Molecular Networking and In Silico Metabolite Structure Prediction. Metabolites 12, 1275 (2022).
Quiros-Guerrero, L.-M. et al. Inventa: A computational tool to discover structural novelty in natural extracts libraries. Front. Mol. Biosci. 9, 1028334 (2022).
Mejri, Y. et al. MS2DECIDE: Aggregating Multiannotated Tandem Mass Spectrometry Data with Decision Theory Enhances Natural Products Prioritization. Chemistry–Methods 202400088 (2025) doi:10.1002/cmtd.202400088.
Allard, P.-M. et al. Integration of Molecular Networking and In-Silico MS/MS Fragmentation for Natural Products Dereplication. Anal. Chem. 88, 3317–3323 (2016).
Zhou, Z. et al. Metabolite annotation from knowns to unknowns through knowledge-guided multi-layer metabolic networking. Nat. Commun. 13, 6656 (2022).
Lê Cao, K.-A. & Welham, Z. M. Multivariate Data Integration Using R: Methods and Applications with the mixOmics Package. (Chapman and Hall/CRC, Boca Raton, 2021). doi:10.1201/9781003026860.
Berthold, M. R. et al. KNIME-the Konstanz information miner: version 2.0 and beyond. AcM SIGKDD Explor. Newsl. 11, 26–31 (2009).
Chandrasekhar, V. et al. COCONUT 2.0: a comprehensive overhaul and curation of the collection of open natural products database. Nucleic Acids Res. 53, D634–D643 (2025).
Wohlgemuth, G., Haldiya, P. K., Willighagen, E., Kind, T. & Fiehn, O. The Chemical Translation Service—a web-based tool to improve standardization of metabolomic reports. Bioinformatics 26, 2647–2648 (2010).
Burgess, K. E. V., Borutzki, Y., Rankin, N., Daly, R. & Jourdan, F. MetaNetter 2: A Cytoscape plugin for ab initio network analysis and metabolite feature classification. J. Chromatogr. B https://doi.org/10.1016/j.jchromb.2017.08.015 (2017) doi:10.1016/j.jchromb.2017.08.015.
Radwan, M. M., Chandra, S., Gul, S. & ElSohly, M. A. Cannabinoids, phenolics, terpenes and alkaloids of cannabis. Molecules 26, 2774 (2021).
Andre, C. M., Hausman, J.-F. & Guerriero, G. Cannabis sativa: the plant of the thousand and one molecules. Front. Plant Sci. 7, 19 (2016).
Pereira Francisco, V. et al. Development of GC–MS coupled to GC–FID method for the quantification of cannabis terpenes and terpenoids: Application to the analysis of five commercial varieties of medicinal cannabis. (2024).
Vásquez-Ocmín, P. G. et al. Cannabinoids vs. whole metabolome: Relevance of cannabinomics in analyzing Cannabis varieties. Anal. Chim. Acta 1184, 339020 (2021).
Aliferis, K. A. & Bernard-Perron, D. Cannabinomics: Application of Metabolomics in Cannabis (Cannabis sativa L.) Research and Development. Front. Plant Sci. 11, (2020).
Zandkarimi, F. et al. Comparison of the cannabinoid and terpene profiles in commercial cannabis from natural and artificial cultivation. Molecules 28, 833 (2023).
De Meijer, E. P. M. & Hammond, K. M. The inheritance of chemical phenotype in Cannabis sativa L. Euphytica 145, 189–198 (2005).
Hazekamp, A. & Fischedick, J. T. Cannabis‐from cultivar to chemovar. Drug Test. Anal. 4, 660–667 (2012).
Hanuš, L. O., Meyer, S. M., Muñoz, E., Taglialatela-Scafati, O. & Appendino, G. Phytocannabinoids: a unified critical inventory. Nat. Prod. Rep. 33, 1357–1392 (2016).
De Ronne, M. & Torkamaneh, D. Discovery of major QTL and a massive haplotype associated with cannabinoid biosynthesis in drug‐type Cannabis. Plant Genome 18, e70031 (2025).
Brungs, C. et al. MSnLib: efficient generation of open multi-stage fragmentation mass spectral libraries. Nat. Methods 22, 2028–2031 (2025).
Neumann, S. et al. MassBank: an open and FAIR mass spectral data resource. Nucleic Acids Res. gkaf1193 (2025) doi:10.1093/nar/gkaf1193.
Marti, G. Lessons from Mass Spectral Library Integration: Addressing Metadata Gaps and Expanding Chemodiversity. (2025).
Wishart, D. S. et al. Chemical Composition of Commercial Cannabis. J. Agric. Food Chem. acs.jafc.3c06616 (2024) doi:10.1021/acs.jafc.3c06616.
Julian Pollmann Florian Huber. Count your bits: more subtle similarity measures using larger radius count vectors. https://doi.org/10.1101/2025.06.16.659994.

The authors declare no competing interests.

Download PDF

Version 1

posted

You are reading this latest preprint version

MS-Net: Multi-Similarity based network annotation for untargeted metabolomics

Status:

Version 1

Abstract

Figures

Introduction

Results

Workflow Architecture and Implementation

Annotation Confidence Scoring System

Feature Filtering and Data Reduction

Network-Based Annotation Propagation

Metadata Enrichment and Output Generation

Feature Filtering and Chemical Space Reduction

Network-Based Annotation Prioritization

Global Dataset Annotation and Chemical Space Coverage

Chemotype Discrimination

Case Study: Cannabinoid Subnetwork Annotation

Discussion

Material and methods

UHPLC-HRMS/MS analysis

Data Processing and Feature Detection

Sirius-CSI annotation

MS-Net parameters

Statistical analysis

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1