From known chemical space to unannotated metabolites: a cluster-guided retention-time driven framework for biologically informed annotation Authors: Dipendra Bhandari, Henry A Paz, Keith Henderson, Kiran Kumar Adepu, Ahmad Mani-Varnosfaderani, Hailemariam Abrha Assress, Brian D Piccolo, Renny S Lan, Elisabet Børsheim, Colin D Kay, Sree V Chintapalli Published: 2026 ABSTRACT AbstractIntroductionUntargeted metabolomics often results in a significant portion of unannotated metabolites, or “metabolic dark matter,” which hinders biological interpretation.ObjectivesA two-step analytical approach was developed to systematically prioritize and interpret unannotated metabolites using plasma LC–MS/MS data from pregnant women with obesity as a biologically relevant test dataset.MethodsThe first step involved clustering 1,021 known metabolites into ten structurally coherent groups based on the Tanimoto similarity, thus defining the biologically relevant chemical space of the dataset. These metabolites were further characterized by Absorption, Distribution, Metabolism, and Excretion (ADME) profiling, protein target prediction, molecular docking and Kyoto Encyclopedia of Genes and Genomes pathway mapping analysis, to establish biological plausibility and functional perspective. Candidate structures for 1,836 unannotated features were retrieved from PubChem using molecular formula and molecular weight matching within a ±0.5 Da tolerance.ResultsThis search yielded 569,115 candidate structures, of which 368,197 unique structures were retained after curation. Tanimoto coefficient filtering reduced the candidate pool to 19,868 structurally plausible candidates, and retention time-based prioritization further refined this set to 418 high confidence candidate annotations, including 83 database-supported candidates identified through HMDB and LIPID MAPS structure database cross-referencing. RT-based prioritization effectively distinguished positional isomers sharing the same molecular formula by incorporating agreement between predicted and experimentally observed retention times.ConclusionThis improved discrimination among structurally similar candidates, expanded metabolite annotation confidence, and provided a scalable framework for prioritizing dark matter metabolites in untargeted metabolomics.Graphical abstractClustered-based workflow integrating chemical similarity and retention time to prioritize and annotate unknown metabolitesSupplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s11306-026-02520-7.Keywords: Metabolomics, In silico, Target prediction, Gene, Protein, Retention time, Dark matter Introduction Metabolomics is the global analysis of metabolites and small molecules in a biochemical system (Eshawu & Ghalsasi, 2024; Johnson & Gonzalez, 2012; Karakas et al., 2025; Liu & Locasale, 2017; Qiu et al., 2023). Metabolites are not only substrates or end-products of biochemical reactions, but are also commonly classified as active regulators, and in some cases metabolite levels represent sensitive indicators of physiological changes (Dai et al., 2022; Milanesi et al., 2020; Pang & Hu, 2023; Wang et al., 2020). However, a significant part of the metabolome has yet to be characterized; thus, these metabolites are referred to as metabolomic dark matter; DM (Monge et al., 2019). DM includes unidentified metabolites that may be of pivotal importance in biological functions but are eliminated through existing systems of identification due to the constraints of current databases and analytical approaches (Monge et al., 2019). To identify and screen metabolites, various complementary technologies have been developed to improve metabolite prediction. These methods are effective when a reference spectrum (authentic standard) or a spectrum in a curated database exists. However, unannotated metabolite features can arise from at least two different situations. First, there are “known unknowns” that are already present in chemical or metabolite databases but remain unassigned when using standard LC–MS/MS. Second, other features may correspond to “unknown unknowns”, such as the absence of metabolites in existing databases, metabolites resulting from microbial modification or exposure to xenobiotics, and metabolites with novel structures. It is crucial to differentiate between these two categories since the former can be retrieved using improved database search and orthogonal prioritization, while the latter may need further structural validation. To overcome this limitation, in silico fragmentation tools mimicking MS/MS fragmentation patterns were employed to predict and rank candidate molecular structures in accordance with their probability (Blaženović et al., 2018). Methods such as Compound Structure Identification: Fingerprint Identification (CSI:FingerID), Sum formula Identification by Ranking Isotope patterns Using mass Spectrometry (SIRIUS), and Competitive Fragmentation Modeling for Metabolite Identification (CFM-ID) tools use tandem mass spectrometry (MS/MS) fragmentation patterns to predict molecular structures through screening of large chemical libraries (Allen et al., 2014; Böcker et al., 2009; Deng et al., 2024; Dührkop et al., 2015; Zhou et al., 2022). Although these models enhance automation process and improve discovery of features, they demand large and high-quality training data. Nevertheless, these approaches are still limited due to spectral similarity, inability to differentiate structural isomers, limitation of the reaction coverage, biotransformation specificity, and unclear extrapolation to non-model organisms and environmental systems (Zhang et al., 2025; Zhou et al., 2022). Despite the progress of analytical technologies, important technical and conceptual hurdles continue, especially when attempting to identify unknown spectral features. A significant drawback is the absence of reference spectra for metabolites that are not yet abundant or common in the environment, or in biological organisms. This is further complicated by the fact that there are also both structural (positional) isomers and stereoisomers, with almost identical masses and similar fragmentation patterns (Idle & Gonzalez, 2007; Monge et al., 2019). Consequently, the high structural complexity of metabolites poses a major obstacle to confident metabolite identification, especially where no reference standards are commercially available (Zhong et al., 2022). The current spectral databases frequently lack completeness, and cannot capture rare or hitherto uncharacterized compounds, and therefore cannot be used to compare experimental data to known metabolites. Compounding the problem, instrumental variability, including different ionization behavior and fragmentation patterns, brings another level of complexity requiring harmonized protocols and large-scale data curation (Abiead et al., 2025; Lin et al., 2020). Furthermore, numerous unknown metabolites could be produced via non-canonical biosynthetic routes or post-translational modifications that are not reflected in existing pathway databases (Cakmak & Ozsoyoglu, 2007; Cobbold et al., 2021). The interpretation of these features is further complicated by analytical artifacts, including in-source fragmentation and ion rearrangements, which can produce signals that mimic genuine metabolites, making biological interpretation challenging (Eicher et al., 2020; Novoa-del-Toro & Witting, 2024; Verpoorte et al., 2022; Xu et al., 2015). While existing approaches for metabolomics annotation can identify structurally plausible candidates, most methods depend on spectral or structural similarity search methods which fail to offer interpretation of biologically plausible or relevant unknown metabolites. Structurally related metabolites produced in endogenous biotransformation reactions or from pathway-level metabolic remodeling are often not easily prioritized and biologically contextualized. Therefore, the incorporation of orthogonal information such as chromatographic behavior, physicochemical properties and biological relevance could create a more biologically meaningful framework for identifying and interpreting metabolic DM, in addition to the traditional spectral matching methods. To achieve this goal, the current study proposes an integrated approach for prioritizing DM which combines formula and molecular weight-based candidate retrieval, structural projection of candidates into known-metabolite chemical neighborhoods, retention time (RT) prediction using machine learning models, and cluster-level biological contextualization (Fig. 1). The workflow is designed to supplement rather than compete with traditional MS/MS-based annotation and is intended to include an additional prioritization strategy for features that do not have spectral evidence or are ambiguous or are lacking in spectral evidence. Metabolites were grouped into structurally coherent clusters using plasma LC–MS/MS data, followed by the filtering of candidate structures identified from PubChem using Tanimoto similarity and ranking based on agreement of the predicted and observed retention time. A key outcome of this study the development of a transparent and biologically informed computational pipeline for prioritizing DM candidates for future validation. Workflow for the dual-phase analytical framework for known and DM metabolite interpretation. Counts represent different workflow levels: metabolite features, retrieved PubChem candidate structures, unique candidate structures, TC-filtered candidate annotations, and final RT-prioritized candidates. The number of candidate-cluster annotations can be higher than the number of unique molecules when multi-cluster mapping is used. A stepwise summary of the workflow counts, and corresponding counting units is provided in Supplementary Table S2 Materials and methods Plasma metabolite collection Metabolite samples were collected from a study examining the effects of exercise training during pregnancy on plasma metabolite responses to acute exercise (Allman-Tucker et al., 2025; Taylor et al., 2025). As part of a randomized controlled trial conducted at Arkansas Children’s Nutrition Center; ACNC (NCT02125149), 119 sedentary pregnant women with obesity (28 ± 5 years) were assigned to either an exercise program (EX; n = 66; combined aerobic and resistance training three times weekly from ~13 weeks to delivery) or a standard-of-care control group (CON; n = 53). Submaximal fitness assessments were performed at ~12 and 24 weeks of gestation, with workload increasing to a rate of perceived exertion (RPE) of 15 on the Borg scale for all participants (Borg, 1982). Immediately post-exercise, fingerstick blood samples were collected and stored at −80 °C. Plasma samples were thawed at 4 °C, and 50 µL aliquots were placed on a QuanRecovery 96-well plate. Pooled quality control (QC) samples, solvent blanks, and a process blank were included. Extraction was performed with 350 µL of ice-cold acetonitrile:methanol:water (160:160:30 µL) containing 500 ng mL−1 of internal standards D6-trans-cinnamic acid and D5-Tryptophan to control variability introduced during sample preparation. After vortexing (10 min, 4 °C) and centrifugation (4,347 × g, 15 min, 4 °C), 100 µL supernatant was collected for each polarity mode, dried under vacuum, and reconstituted in 50 µL water:methanol (95:5) with 1000 ng mL−1 sulfadimethoxine and (Allen et al., 2014) C10- Kynurenine sulfate to monitor instrument measurement variability. A total of 235 samples were analyzed across three batches with randomized injections. QC samples were injected every tenth run to monitor performance. The electrospray source was cleaned between batches, and calibration was performed using Pierce FlexMix standards per manufacturer instructions. In the present manuscript, these samples were used as a biologically relevant LC–MS/MS dataset for developing and evaluating the computational DM prioritization workflow; differential analysis between exercise and control groups was not the primary objective of this study. LC–MS/MS metabolomics Instrumental analyses were performed on an Orbitrap Exploris 480 mass spectrometer (Thermo Fisher Scientific, Waltham, MA) equipped with high- and low-flow HESI probes, coupled to a Vanquish Ultra-High-Performance Liquid Chromatography (UHPLC) system. Chromatographic separation was achieved using an ACQUITY Premier Charged Surface Hybrid (CSH) C18 column (1.7 µm, 2.1 × 100 mm; Waters, USA) at 40 °C with a flow rate of 0.3 mL min–1. The mobile phase consisted of water (A) and acetonitrile (B) with 0.1% formic acid, using a 15-min gradient and an injection volume of 5 µL. The LC gradient was programmed as follows: 0–2 min, 0–40% B; 2–8 min, 40–98% B; 8–10 min, held at 98% B; 10–10.5 min, returned to 0% B; and 10.5–15 min; equilibrated at 0% B. Data were acquired in both positive (3.6 kV) and negative (3.5 kV) ESI modes (sheath/aux/sweep gas = 35/10/1 Arb; vaporizer and transfer tube = 350 °C). Full MS1 scans were collected over m/z 60–900 (positive)/58–870 (negative) at 180,000 resolution, AGC 5e5, MIT 100 ms MS/MS spectra were acquired on pooled QCs using the AcquireX iterative DDA workflow (top 5 precursors, 2 m/z window, stepped HCD 20/40/60 V, resolution 30,000, AGC 1e6, MIT 50 ms). Instrument control and data acquisition were managed using Xcalibur v4.4.16.14. Data collection, processing, and compound annotation Raw LC–MS/MS data were processed using FreeStyle v1.8 and Compound Discoverer v3.3 (Thermo Fisher Scientific, Waltham, MA). Peak detection for internal standards was performed with the Integrated Spectral Component and Isotope Search (ISCIS) algorithm. Data was processed in Compound Discoverer (CD) v3.3 following a node-based workflow for spectrum selection, RT alignment, compound detection and grouping, gap filling, QC correction, background subtraction, formula prediction, and spectral annotation. Metabolite annotation was performed using the default Compound Discoverer workflow, with spectral matching against the mzVault and mzCloud libraries. The spectral matching scores and the number of matched fragment ions for all Level 1 and Level 2 metabolite identifications are provided in Supplementary Table S1. Features were detected within m/z 50–1200 Da (S/N > 1.5), aligned to pooled QC samples, and filtered using 5 ppm mass tolerance and 0.2 min RT windows. Chemical formulas for the unknown compounds were predicted by setting the mass tolerance to 5 ppm and intensity tolerance of 30% for isotope pattern matching. Spectral annotation employed mzCloud, mzVault, and NIST databases (10 ppm mass tolerances), and ChemSpider searches across Kyoto Encyclopedia of Genes and Genomes (KEGG), Human Metabolome Database (HMDB), and BioCyc within 5 ppm. RT matching was supported by the search Mass List node, which matched detected features within 5 ppm mass error and 0.6 min RT to our in-house library of 339 standards. Annotation confidence was assigned as Level 1 (MS/MS spectral matches to the in-house spectral library and RT), Level 2 (putative annotations based on either MS/MS matches to spectral libraries (mzVault, mzCloud, or NIST) or RT match with isomeric compounds matching both MS/MS spectra and RT), or Level 3 (no chemical identification with annotations solely based on molecular formula or accurate-mass match in ChemSpider without MS/MS or RT confirmation). Features with mass error > ±0.5 ppm or a QC-corrected relative standard deviation of 20%, or features detected in blanks, were eliminated. To ensure data robustness, only features with a peak quality score of at least 5 in a minimum of 50% of the samples, and a peak intensity of at least 500,000 in at least 50% of the samples, were retained for downstream analysis. Clustering and similarity exploration A total of 2857 metabolites were obtained from high-resolution untargeted metabolomics (Supplementary Figure S1). Based on the established metabolite annotation confidence tiers, metabolites classified to levels (1, 2 & 3 metabolites; n = 1021) were collectively retained as “known metabolite features” for downstream structural clustering analyses. Although Level 3 annotations do not represent definitive compound-level identifications which include multiple candidate structures within the same chemical class, they still provide valuable structural and biochemical information, including molecular formula, spectral characteristics, and metabolite class relationships. Instead of viewing Level 3 annotations as metabolite names, we considered them to provide a weak definition of structure-based neighborhoods that share similar structural scaffolds and physicochemical properties. The remaining 1836 unannotated features which are devoid of chemical identity or structural assignment, were classified as DM or “unknown metabolites” (Supplementary Table S1). The dataset was assembled at the level of the LC–MS features, so duplicate entries of putative compound names or candidate structures were not automatically recognized. In certain cases, the candidate annotations had the same putative name or formula but were associated with different experimental features with different RTs and/or ion forms and/or adduct states and/or modes of polarity and/or potential isomeric assignments. For this reason, these entries did not undergo collapse to a compound-level row but rather were kept as individual LC–MS feature-level annotations. This distinction is significant as it allows the present workflow to focus on candidate structures for individual experimental features based on feature-specific m/z and experimentally observed RT. To organize the known metabolites and capture their underlying chemical relationships, we applied the Silhouette Optimized Molecular Clustering (SOMoC) method (Prada Gori et al., 2022). These structurally coherent known-metabolite clusters served as a chemical reference framework to guide the exploration of unannotated DM metabolites for interpreting unknown features. Since the unknown metabolites lack definitive chemical names or structures and are characterized only by formula, delta mass, calculated molecular weight (MW), Mass/Charge (m/z), MS/MS peaks, and experimentally calculated RT, their direct structural interpretation remains challenging. To define this reference chemical space, known metabolites (Levels 1–3) were first represented using their SMILES structures and grouped into structurally coherent clusters based on molecular similarity. Metabolite structures for which no compounds were known from PubChem were then processed using this structural reference landscape to find structurally related and biologically relevant candidates. (Kim et al., 2025). To improve clarity across the multi-step workflow, we distinguished the major counting units used throughout the manuscript, including LC–MS metabolite features, known metabolite features, PubChem-retrieved candidate structures, unique candidate structures retained after curation, TC-filtered candidate structures, RT-prioritized candidate annotations, and database-supported candidate annotations. These categories are summarized in (Supplementary Table S2) because the number of candidate-cluster annotations can exceed the number of unique structures when the same candidate maps to more than one known-metabolite reference cluster. Tanimoto coefficient: A structure-based search strategy was used to identify plausible molecular identities for unannotated DM features. As the initial features were not annotated, the candidate structures were retrieved from PubChem based on the assigned chemical formula and associated molecular weight, allowing a mass tolerance of ±0.5 Da. This initial PubChem search generated numerous structural candidates for each unannotated feature. Before fingerprint-based similarity analysis, candidate entries with no available structural information such as missing or incorrect SMILES, duplicate entries, or chemically ambiguous entries were eliminated. The structural similarity between the PubChem-generated candidate structures and known metabolites was then calculated using the TC approach. Molecular fingerprints were created from SMILES representations using RDKit (v2024.03.5) Morgan circular fingerprints with a radius of 2 and a fingerprint size of 2048 bits (Landrum et al., 2024). These extended connectivity fingerprint (ECFP)-like fingerprints represent local atomic environments and molecular substructures with binary vectors. All known metabolites and candidate structures generated with PubChem were processed with the same fingerprints to enable direct comparison of TC values. Chirality was not explicitly represented; thus, TC values represent similarity in the two-dimensional fingerprint features and do not necessarily indicate stereochemical identity. For any two compounds, A and B, the Tanimoto coefficient was calculated as: where a and b represent the number of fingerprints feature present in compounds A and B, respectively, and c represents the number of shared fingerprint features. The values of the TC are between 0 and 1, with closer values to 1 denoting a higher similarity to the molecular fingerprint representation (Ma et al., 2011). The pairwise TCs were computed among each PubChem candidate structure and all known metabolites in the preexisting reference clusters for the cluster-based structural projection. The reported TC value for each candidate-cluster assignment is the maximum pairwise TC, which is the similarity between the candidate and the nearest known metabolite in the cluster. Therefore, the TC values reported in Supplementary Table S3 are not cluster-wide average TC values. If a PubChem candidate and at least one known metabolite in the assigned cluster has the same morgan fingerprint representation, a TC value of 1.0 will be obtained. This is not an indication that the candidate is the same molecule as all the other metabolites in the cluster, nor does it necessarily prove the identity of the molecule in all stereochemical, tautomeric, or ionization states. To ensure that a spectrum of similarity was captured, from the most similar metabolites to moderately similar compounds that could be biotransformation-related analogs or endogenous metabolites, candidate structures were retained in the range of 0.50–1.00 for TC. Higher TC values were seen as greater structural similarity to known metabolites, with moderate TC values increasing the searchable chemical space to include chemically related but potentially modified metabolites. In cases where a PubChem-derived candidate structure exceeded the selected TC threshold when compared with known metabolites from more than one reference cluster, the candidate was retained as a multi-cluster annotation rather than forced into a single cluster. These multi-cluster annotations reflect shared structural motifs across known-metabolite clusters. As there was no universally accepted threshold value for TC to be applied when prioritizing candidates for DM, TC ≥ 0.5 was set as an exploratory lower threshold for retaining structurally similar candidates without unrestricted expansion of PubChem candidates (Bajusz et al., 2015). The TC is widely used for molecular fingerprint-based similarity calculations in cheminformatics, and prior work has shown that it performs reliably among commonly used similarity metrics for fingerprint comparison. Thus, TC ≥ 0.5 was not considered to be structural identity, but rather to be a permissive threshold for discovery-oriented candidate recovery in this study. In addition, more stringent thresholds, TC ≥ 0.6 and ≥ 0.7, were assessed and tested in parallel to increasingly conservative levels of structural similarity. The structural plausibility step was primarily performed with the TC-based filtering. These candidate structures were then further prioritized based on orthogonal criteria like system specific RT prediction and assessment of biological relevance. This combined approach enabled the evaluation of the candidate DMs both as structural analogues of known metabolites in terms of their experimental chromatographic properties, and in the context of the biological process involvement. Because different steps of the workflow operate on different counting units, we distinguished detected metabolite features, PubChem-retrieved candidate structures, unique candidate structures, candidate-cluster annotations, unique chemical formulas (UCFs), and database-supported annotations. Candidate-cluster annotations may exceed the number of unique molecules because the same candidate structure can map to more than one reference cluster when it exceeds the TC threshold in multiple clusters. Thus, the number of raw candidates should not be seen as the number of unique metabolite identities. Absorption, distribution, metabolism, and excretion prediction (ADME) Estimation of the ADME properties were performed only for known metabolites using SWISS-ADME platform (Daina et al., 2017) that includes gastrointestinal absorption (GI), blood–brain barrier permeability (BBB), skin permeability, and overall bioavailability scores. GI and BBB predictions were obtained using the Brain or IntestinaL EstimateD (BOILED-Egg) model, while skin permeability was estimated using the Potts and Guy method (Potts & Guy, 1992). The bioavailability score was calculated based on structural and physicochemical rules, where a standard score of ≥ 0.55 suggests favorable bioavailability, while a lower score (< 0.55) implies limited systemic uptake. This analysis was used to biologically characterize the structurally coherent clusters identified from the known metabolite data, and to define a physiologically relevant reference space for interpreting the metabolome data. Importantly, only known metabolite clusters were used to calculate ADME properties and describe the major biological and physicochemical properties of each chemical neighborhood. Target prediction and molecular docking To add functional and biological context into the metabolite clustering framework, known metabolites from each structural cluster were subjected to in silico protein target prediction to identify potential protein-target interactions associated with each chemical neighborhood. Specific target information for metabolites was retrieved from BindingDB (Liu et al., 2006), Swiss Target Prediction (Daina et al., 2019), and GlassDB (Yue et al., 2023) which identify probable protein partners based on experimental ligand-target interactions data (along with specific binding site residue information) and structural similarity metrics. If no direct target predictions were available from SwissTargetPrediction, BindingDB, or GlassDB, docking studies were carried out using protein targets identified from structurally similar metabolites in the same structural cluster, with the assumption that structurally similar metabolites may have similar binding propensities. The predicted targets of all three resources have been compiled and consolidated, with higher confidence given to targets that were predicted across multiple resources. When predictions differed between databases, targets were prioritized using the amount of overlap between platforms, evidence of the targets being metabolic, and reports of experiments demonstrating the metabolic targets. It was hypothesized that metabolites with only indirect targets might have biological interactions with other members of the same cluster because they share structural scaffolds and physicochemical properties which is in line with the structure–activity relationship (SAR) principle (Harren et al., 2022). To further evaluate these predicted interactions, molecular docking analyses (Trott & Olson, 2010) were performed to characterize potential binding orientations, binding affinities, and key molecular interactions at active or allosteric protein binding sites. For docking analyses, candidate ligands were prepared from SMILES-derived structures and converted into three-dimensional conformations using in-house script using RDKit library. Protein structures were retrieved from the protein data bank (PDB), and the structures were processed by removing nonessential waters or ligands, adding hydrogens, assigning charges and converting the file to the required docking format. Docking was carried out with AutoDock Vina (v1.1.2) (Trott & Olson, 2010) with grid boxes centered on binding site reported from the literature, predicted active site or a representative region of the ligand binding region (Daina et al., 2019; Liu et al., 2006; Yue et al., 2023). Binding energies measured in kcal/mol were only considered as a relative estimation in terms of plausibility of interaction to be computed. Protein structures, including the corresponding PDB codes and chain identifiers, were explicitly listed for each cluster in (Supplementary Tables S4). Docking scores were only used as relative estimates of the plausibility of binding in each protein-metabolite system and were not considered as experimental binding affinities. Thus, docking data was not used to validate direct protein-metabolite interactions but rather to provide a basis for biological context. Metabolic pathway mapping and network integration To gain biological context for the clusters of metabolites defined by structure, curated biological databases were used to map known metabolites to metabolic and signaling pathways. Importantly, pathway mapping, ADME profiling and molecular docking were not used directly to filter or annotate DM metabolites. Instead, these analyses were carried out at the level of the known metabolite clusters to help delineate the wider biological and functional context of structurally similar metabolites. The KEGG (Kanehisa & Goto, 2000) was used to map known metabolites within each structural cluster to pathways, with the aim of identifying metabolic functions, signaling pathways, and enzyme-associated processes that were enriched within each cluster. Furthermore, the Search Tool for Interacting Chemicals (STITCH) database (Kuhn et al., 2008) was used to predict metabolite-protein interaction networks and to gain systems-level biological context of metabolites that have structural similarities to the protein of interest. Metabolites having better predicted biological relevance (based on favorable protein target predictions and molecular docking interactions) were selectively analyzed by KEGG pathway mapping analysis. An in-house Python workflow was developed to automate these analyses, and process them in batches, integrate pathway annotation and visualize the metabolite-protein interaction network across clusters. Retention time prediction model development The RT was used to provide complementary physicochemical information to help narrow down structurally plausible candidates, since chromatographic retention behavior is strongly related to the molecular properties of the compounds, such as polarity, hydrophobicity, hydrogen bonding capacity, and molecular size. Hence, RT prediction was included as an extra filtering step to minimize structural assignments flagged by similarity search. A library of 339 authenticated metabolite standards generated by the ACNC Metabolomics Core facility was used to build an in-house RT prediction model. Of the broader in-house RT standard library, these 339 compounds contained valid SMILES structures and experimentally measured RT values under the same LC–MS/MS conditions as the unknown metabolite features and were therefore retained for model development. To assess the model’s generalizability and reduce overfitting, the dataset was randomly split into two subsets: 80% for training and 20% (80:20) for testing with random state value of 42, resulting in 271 compounds for model training and 68 compounds for independent test-set evaluation A total of 217 RDKit molecular descriptors were generated from SMILES representations using RDKit.Chem.Descriptors. These descriptors comprised both molecular, structural, topological, physicochemical, electronic, and charge-related or fragment-based descriptors. Missing descriptor values were filled with the mean imputation, and descriptors were standardized prior to model building. Model performance was evaluated on the test set and reported using mean absolute error (MAE) and R2. To avoid ambiguity, the term “metabolite feature” refers to an LC–MS-detected feature defined by m/z, RT, and associated annotation information, whereas “molecular descriptor” refers to the RDKit-derived variables used for machine learning-based RT prediction. To enhance the prediction robustness and minimize the reliance on single modeling method, two complementary machine learning models, namely CatBoost Regressor (Prokhorenkova et al., 2018) and Stacking Regressor (Breiman, 1996), were implemented (Fig. 2). The CatBoost model was chosen for its ability to be applied to nonlinear relationships and heterogeneous descriptor spaces, while the Stacking Regressor used Random Forest, Gradient Boosting base learners, and RidgeCV as a meta-model to enhance generalization and reduce model-specific bias. Hyperparameter optimization for the CatBoost model was performed using Optuna with five-fold cross-validation to minimize MAE, and model performance was evaluated using MAE and R2 on the independent test dataset. Simultaneously, SHAP (SHapley Additive exPlanations) analysis was used to reveal the most important molecular descriptors that influence the variation in RT (Bernal et al., 2025). Following model optimization and validation, RT values were predicted for candidate structures that passed the prior Tanimoto similarity filtering step (TC ≥ 0.5). Estimated RTs from both models were then combined and compared with the experimentally measured RTs of the unknown metabolite features. The CatBoost and Stacking Regressor predictions were averaged using an unweighted consensus approach because both models showed comparable independent test-set performance, and no external validation dataset was available to justify model-specific weighting without introducing additional bias. Determination and validation of optimal cluster numbers based on SMILES derived structural similarity. Silhouette scores were calculated across varying cluster numbers to evaluate clustering quality using Tanimoto similarity of molecular fingerprints. The optimal cluster number (K = 10; silhouette score = 0.7591, indicated by the red arrow) was selected, as it produced the highest silhouette score during SOMoC optimization, indicating the best balance of intra-cluster similarity and inter-cluster separation among the tested cluster solutions while minimizing over-partitioning Priority score calculation based on retention time deviation To prioritize template structures within DM metabolite annotation following TC filtering, an RT-based confidence-scoring scheme was applied using a custom script in R. The two predicted RT values (CatBoost and Stacking ensemble models) were averaged to get a consensus RT estimate of each unknown metabolite feature. This consensus was compared with the experimentally observed RT (Expt_RT) to compute the RT deviation (diff_rt). A weighting factor (wt1) was then calculated as the inverse squared RT deviation, i.e., wt1 = (diff_rt)⁻2, assigning higher weight to candidates whose predicted RT values more closely match the experimental value. To account for multiple candidate structures per metabolite (indexed by feature ID; Supplementary Table S2), candidate weights were scaled by the within-feature total to yield a priority score (priority = wt1/Σwt1) within feature ID that represents each candidate’s relative support within a feature ID based on RT agreement. Results and discussion Known metabolite dataset Representative SMILES of each known metabolite (n = 1021) identified with annotation confidence Levels 1, 2, and 3 were retrieved from the PubChem database. Sixteen metabolites were excluded because of ambiguous or unresolved annotations (including metabolites with multiple inconsistent compound assignments, missing stereochemical or structural information, naming differences between databases and missing corresponding structures in PubChem), leaving 1005 metabolites in the curated set. Clustering was subsequently performed using the SOMoC framework, which gave ten clusters (assigned with numbers from Clus0 to Clus9) based on the maximum silhouette score during SOMoC optimization (silhouette score of 0.7591), closely agreed on intra-cluster similarity and inter-cluster separation (Fig. 2). Rather than removing the potential isomeric forms (which often display different RT values), they were intentionally retained to preserve biologically relevant structural diversity and better represent the underlying chemical complexity of the metabolomics dataset. This approach was especially relevant to the capture of metabolites resulting from biotransformation reactions, as slight structural changes can give rise to essentially identical but structurally different molecules with different functional properties. The resulting ten clusters varied in size, ranging from smaller, chemically specific groups such as Clus0 (31 metabolites) to broader clusters like Clus2 (187 metabolites). The clusters collectively gave a structured reference framework to classify known metabolites based on common structural features and biochemical relationships. Although metabolite names were first extracted from Compound Discoverer annotations, PubChem was specifically used to search and standardize SMILES and structural metadata for the metabolites, in subsequent similarity analysis (Supplementary Table S4). Structural coherence of these clusters and high silhouette scores allowed for interpretation of cluster level biochemical characteristics and exploration of shared biological functions, protein target relationships, and pathway associations among structurally similar metabolites. Importantly, clustering of the compounds was conducted solely based on structural similarity as determined from SMILES, while the clusters were biologically characterized and interpreted by subsequent docking analyses, pathway mapping, and protein target prediction studies downstream of the clustering process. Cluster-level biological contextualization of known metabolites To further assess the physiological plausibility of the known metabolites comprising each cluster, we computed their ADME-related properties. Since the present dataset was generated from plasma rather than brain or cerebrospinal fluid, BBB permeability was not used to infer central nervous system relevance. Instead, BBB predictions were considered only as a general physicochemical indicator related to lipophilicity, polarity, and potential tissue-distribution behavior. Rather than calculating an arithmetic mean of all metabolites, the cluster-level trends were interpreted based on the major distribution and frequency of the metabolite properties in each cluster. Overall, cluster-level ADME profiling showed that polar peptide-like and phosphoethanolamine-containing clusters generally exhibited lower predicted absorption or permeability, whereas clusters enriched in lipophilic aromatic compounds, steroidal metabolites, fatty acid derivatives, acylcarnitines, indoles, and purines displayed more variable but often favorable predicted bioavailability and permeability profiles (Supplementary Table S4). Using SwissTargetPrediction, BindingDB, and GlassDB databases, several metabolites from the known clusters were associated with predicted protein targets together with corresponding binding site residue information. Many structurally related metabolites in the same cluster, however, did not provide direct target predictions, which probably indicates that the existing cheminformatic databases and/or structural representation were not fully covered. Though lacking direct target predictions, a few metabolites showed favorable docking scores with protein targets of structurally related cluster members. For example, in Clus0, metabolites lacking direct SwissTargetPrediction hits were docked against representative targets, including cytochrome P450 enzymes, protein kinases, protein-tyrosine phosphatases, and vitamin D receptor-associated proteins identified from cluster representatives. Molecular docking binding energies (kcal/mol) were calculated to determine the strength of the interactions between the metabolites and the protein molecules using AutoDock and were interpreted as relative indicators of predicted binding strength rather than definitive experimental binding affinities. Scores below − 7.0 kcal/mol were interpreted as having strong binding affinity, scores between − 5.0 and − 7.0 kcal/mol were considered as moderate binding affinity, and scores above − 4.0 kcal/mol were interpreted as weak or nonspecific interactions (Wong et al., 2022). As these docking scores are computational estimates, they were only used for prioritization of metabolite-protein pairs for downstream biological interpretation and future experimental validation rather than as definitive evidence of physiological binding. Clus0 docking results showed several positive metabolite-protein interactions, which further justified the potential of down streaming the biological interpretation to structurally related metabolites, which do not have direct target annotations. The docking results were therefore interpreted only as relative indicators of the potential metabolite-protein interaction plausibility (Supplementary Tables S4). However, DM metabolites were not directly subjected to ADME prediction, target prediction, molecular docking, and pathway mapping to assign final candidate structures. These analyses were instead performed only on known metabolite clusters only to serve as biological and physicochemical context for the chemical neighborhoods into which dark matter candidates were projected. Thus, these analyses should be used as a second layer of contextualization and not as a validation of the candidate annotations. Pathway-level prioritization and visualization of bioactive metabolites (cluster-specific vs. combined-clusters) Known metabolites were prioritized and evaluated for biological relevance by integrating ADME-related properties, molecular docking scores, and pathway-level associations. Preliminary filtering within each structural cluster retained metabolites with favorable ADME-related properties, including high gastrointestinal absorption, BBB permeability where biologically relevant, acceptable bioavailability scores (≥ 0.55), and reasonable skin permeability values (log Kp between − 20 and − 1). The metabolites with predicted protein targets, and a binding energy < − 4.0 kcal/mol, were kept for further pathway analysis. Importantly, this prioritization workflow was used separately for every structural cluster instead of the whole set of metabolites. This cluster-based strategy enabled structurally related metabolites with related chemical scaffolds to be tested in coherent chemical neighborhoods and ensured that structurally unrelated compounds did not interfere with the evaluation of structurally related metabolites. To extend biological interpretation beyond metabolites with direct target annotations, structurally related metabolites within each cluster were docked against protein targets identified from representative cluster members, based on the premise that metabolites with similar structures may exhibit comparable binding tendencies. Biological influence was estimated quantitatively by calculating a dual-weighted prioritization score, which combined (i) number of unique protein targets associated with the metabolite, and (ii) number of enriched pathways associated with the unique protein targets. As a result, metabolites with larger target interaction profiles and/or more associated pathways were prioritized first. Pathway influence bar plots and metabolite-pathway heatmaps were generated for each cluster (Supplementary Figure S2), revealing biologically coherent and often cluster-specific pathway enrichment patterns. All metabolite clusters were also analyzed together in a combined-cluster framework to compare patterns of metabolite organization at a localized versus global level. In this analysis, metabolite-target-pathway relationships from all clusters were pooled and reanalyzed with the top 10% highest-ranking pathway associations. The analysis of the combined sets gave a global overview of the metabolic themes and pathways that occurred repeatedly across all the clusters but also led to a higher degree of overlap in pathways and a lower level of functional specificity. Pathway overlap across metabolites was summarized using a global similarity heatmap (Fig. 3), while hierarchical clustering of the top enriched pathways revealed highly interconnected functional groups with limited separation (Fig. 4A). The relative contribution of the top enriched pathways is shown in Fig. 4B. Likewise, the combined network graph (Supplementary Figure S3) created using a force-directed layout algorithm showed significant overlaps in nodes and convergence of pathways around the highly connected metabolites, making it more difficult to separate different biological modules. Cluster-specific analysis was the primary strategy for biological interpretation and metabolite prioritization because it preserved structural coherence and improved pathway specificity, whereas complementary combined-cluster analysis provided system-level insights into global pathways and central metabolites shared across multiple biochemical processes. Global multiset Jaccard similarity across combined metabolite clusters. All ten optimized metabolite clusters were integrated and analyzed at a fixed top 10% threshold to assess global pathway overlap Integrated analysis of the top 10% pathways across all known metabolites. A Hierarchical clustering heatmap of enriched pathways across metabolites and B pie chart showing the relative distribution of enriched pathways Unknown metabolite dataset Structural space was expanded to include unannotated metabolites that had chemical formulas and molecular weights assigned to their features, and were subsequently searched against the PubChem database with a mass tolerance of ± 0.5 Da. Because the molecular structures corresponding to these uncharacterized features had not yet been assigned, PubChem searching was performed at the formula/mass level rather than by structure, allowing plausible candidate structures to be retrieved based on resolved mass and predicted chemical formula. This search returned 569,115 candidate structures with associated compound names and formulas, representing possible structural matches for 1,021 unknown features. However, 815 formulas did not return valid PubChem matches under the applied criteria. Following the filtration of invalid, duplicated, and chemically ambiguous database entries (those with incomplete structural information or non-single-component records, which could cause a fingerprint-based similarity calculation), 368,197 unique candidate structures were left for further analysis. To avoid de novo clustering of unknown metabolites, candidate structures were mapped on to the ten pre-existing structural clusters defined from known metabolites (Supplementary Figure S1). Molecular fingerprints were generated for each of the PubChem candidate structures using the same fingerprinting method used for the known metabolite set. Each candidate structure was then compared to the known metabolites in each reference cluster using pairwise TC values. If the candidate had a TC similarity to at least one known molecule in the same known-metabolite cluster with the threshold chosen, the candidate was annotated for that cluster. If a candidate displayed TC similarity with known metabolites in multiple clusters, the metabolite was left as a shared candidate annotation for multiple relevant clusters to avoid being assigned to just one of the multiple clusters. These workflow levels are summarized in (Supplementary Table S2) to distinguish feature-level counts from structure-level and candidate-annotation-level counts. Structural projection was assessed using three thresholds, from exploratory to stringent (TC ≥ 0.5, ≥ 0.6, and ≥ 0.7). The TC ≥ 0.5 threshold was used as a permissive discovery-oriented cutoff to maximize recovery of structurally related candidate annotations, whereas TC ≥ 0.6 and TC ≥ 0.7 were used to evaluate increasingly conservative candidate sets. As expected, lower TC thresholds increased structural coverage, while higher threshold reduced candidate numbers but increased structural stringency. Therefore, TC ≥ 0.5 was used for broad candidate recovery, but candidate confidence was not assigned based on TC alone. Final prioritization also incorporated RT agreement and database cross-referencing where available. Structural redundancy was also assessed based on unique chemical formulas (UCFs) since PubChem retrieval returned more than one structure for the same chemical formula, while some of the retrieved candidates were similar to known metabolites within more than one reference cluster. 195 features were identified as unknown features and mapped to known-metabolite clusters, with the highest UCF numbers being observed for clusters Clus5 (31) and Clus9 (27) at the highest value of TC ≥ 0.7. The number of unknowns increased to 271 at TC ≥ 0.6, where structurally related analogs were more widely recovered. Importantly, several clusters had a significant number of candidate structure annotations with relatively few UCF annotations, suggesting that there was significant redundancy in the retrieved structures. Despite the use of multiple TC thresholds, 604 unknown features could not be confidently mapped to any known-metabolite cluster at TC ≥ 0.5. Additional relaxation below the TC < 0.5 level may yield more structurally dissimilar candidates but would be at greater risk for false-positive annotation and decreased chemical interpretability. Thus, a minimal exploration threshold for structural projection was established as TC ≥ 0.5, and increasingly conservative candidate assignments as TC ≥ 0.6 and TC ≥ 0.7 were considered (Fig. 5). However, the final confidence in candidate annotations was not determined by TC alone, as it was also validated by using additional orthogonal approaches such as RT prediction and downstream prioritization. Distribution of total candidate structures and unique chemical formulas (UCFs, unique chemical formulas) across clusters at varying Tanimoto similarity. Bar plots illustrate the number of total retrieved structures (white bars) and unique chemical formulas (UCFs; black bars) assigned to clusters (Clus0-Clus9) under three Tanimoto Coefficient (TC) thresholds: 0.5, 0.6, and 0.7. Lower TC thresholds (0.5) resulted in substantially larger candidate pools across clusters, reflecting broader structural inclusion. Increasing the TC threshold progressively reduced the number of retrieved structures, indicating stricter structural similarity constraints Tanimoto coefficient scoring followed by RT priority scoring for DM metabolite prediction At TC ≥ 0.5, 19,868 candidate structures were retained, with most candidates falling within the moderate TC range, indicating structural relatedness to known metabolites but not definitive metabolite identity. Hence, the prediction of RT was added as another orthogonal validation step after filtering by TC. The RT that describes the chromatographic behavior and physicochemical properties, such as polarity, hydrophobicity, and molecular interactions with the stationary phase, was used as a complementary parameter for ranking structurally similar candidates. Global or publicly available RT prediction models were not applied because RT is highly dependent on the LC–MS method conditions, such as the composition of the mobile phase, gradient, column chemistry, column flow rate, etc. A CatBoost regression model achieved a MAE of 0.34 and an R2 of 0.945 on the independent test set, indicating strong agreement between predicted and experimentally observed RT values (Fig. 6A). A Stacking Regressor combining Random Forest, Gradient Boosting, and RidgeCV models produced an MAE of 0.41 and an R2 of 0.927 (Fig. 6B), indicating comparable predictive performance between the two models. Q2 (cross-validated coefficient of determination) was not calculated in the current workflow, therefore, model performance was reported using independent held-out test-set MAE and R2 rather than cross-validated Q2. Prediction of retention time (RT) using machine learning models. A CatBoost Regressor model showing strong agreement between predicted and experimental RT values (mean absolute error [MAE] = 0.34 min, R2 = 0.945). B Stacking Regressor model demonstrating comparable predictive performance (MAE = 0.41 min, R2 = 0.927). Each point represents an individual metabolite, with dashed lines indicating the line of identity The trained RT models were then applied to 108,512 candidate structures that passed TC similarity filtering. To prioritize candidates within each feature group, a consensus predicted RT was calculated by averaging the CatBoost and Stacking Regressor predictions. This RT deviation was used as the weighting factor, with smaller prediction errors being given more weight. To derive a final priority score based on the RT, these weights were normalized within the candidate group, comprising all the candidate structures that share a specific feature or molecular formula. This combined TC-RT scoring scheme enabled the selection of top candidates on both structural and chromatographic consistency. Predictions are likely to be more accurate for candidate structures that are within the chemical and descriptor space spanned by the training compounds, as the prediction models were trained on real standards found in the in-house library. Candidate molecules that had a significantly greater size, unusual functional groups, or descriptor profiles not covered in the training set were thus interpreted with greater caution. However, the RT model was not employed for structural identification, but rather for candidate ranking, and prediction uncertainty was not formally estimated in the current workflow. Future implementations could involve the addition of prediction intervals, applicability domain filtering, external validation sets or authentic-standard confirmation to extend our ability to quantify confidence in RT-based prioritization. Structural annotation and biological relevance of prioritized DM metabolites We prioritized 418 unique molecules from the PubChem-derived candidate pool that matched 1,021 unannotated DM features after performing dual filtering based on the TC of structure similarity and system-specific RT prediction. For clarity, the TC values reported in Supplementary Table S2 are the maximum pairwise TC computed between each PubChem candidate structure and the most similar metabolite in an existing reference cluster. These values do not represent the average TC between a candidate molecule and all molecules within a cluster. Hence, if a candidate molecule has the same Morgan fingerprint representation as at least one known metabolite in the assigned cluster, using the selected fingerprinting method, the TC value will be 1.0. The TC value of 1.0 should thus be interpreted as a complete fingerprint match with the closest reference metabolite, and not as a complete match with all molecules of the cluster or as definitive chemical identity across all stereochemical, tautomeric, or ionization states. To further test the biological plausibility for these prioritized structures, an automated database cross-referencing pipeline was developed, based on the Human Metabolome Database (HMDB) and the LIPID MAPS Structure Database (LMSD). Out of 418 unique candidate molecules, 83 were annotated in at least one of these curated metabolomics or lipidomics databases, highlighting their relevance as known, or previously reported biochemical entities. The candidates to be annotated, based on databases, were not interpreted as definitive candidates but as higher confidence annotations, which passed the structural and chromatographic filtering. Their recovery further demonstrates that the TC-RT workflow preserves biologically plausible metabolites that may not have been recovered by standard compound matching due to incomplete mass spectral coverage, poor spectral evidence, and/or poor database representation. The final 83 candidates supported by HMDB/LIPID MAPS should be considered as putative candidates for the database supported annotations, not the experimentally confirmed metabolites. These candidates were selected as they agreed on formula/molecular weight matching, structural relatedness (based on TC), RT consistency and metabolite-database cross-referencing. But authentic standard confirmation or targeted MS/MS validation or additional orthogonal structural data would be necessary for definitive metabolite identification. The candidates supported by the database were linked to biologically coherent chemical classes, based on the previously constructed known-metabolite structural network and cluster organization. The latter consisted of acylcarnitines, fatty acid derivatives, amino acid conjugates, hydroxylated lipids, long-chain amines, amides, sterol-related molecules, and complex lipids. Some candidates were annotated to more than one reference cluster because they exceeded the TC threshold against known metabolites in multiple clusters. This multi-cluster mapping is indicative of common structural features between related reference neighborhoods and does not mean that the candidate molecule was averaged over all these molecules or is equally similar to each of them. The structural similarity criteria of TC combined with the RT prioritization criteria were especially notable for the formula C31H61NO4. Although 186 PubChem structures were retrieved for this formula, only 20 structures corresponding to 3-(methyltricosanoyloxy)-4-(trimethylazaniumyl)butanoate positional isomers passed TC-based filtering. The main difference between the two types of candidates was the location of the methyl substitution in a 23-carbon fatty acid chain, with each one corresponding to a different HMDB identifier, ranging from HMDB0240967 to HMDB0240986. Due to their identical molecular formula, it is impossible to distinguish the various compounds based on their formula alone. But the chromatographic properties can be modified by small structural changes, like the position of the methyl group, the branching or the unsaturation state of the chain. Thus, the ability to predict RT for the system could serve as an orthogonal criterion for ranking these near-isomeric candidates. Predicted RT values that most closely matched the experimentally determined RT were assigned the highest normalized priority scores, so that the workflow would be more likely to prioritize the most chromatographically similar structures in each isomeric group. High-scoring candidates, especially those with scores close to 1.0, were supplemented with metabolite classes that have known biological significance: acylcarnitines, amino acid conjugates, long-chain fatty acid derivatives, hydroxylated lipids, long-chain amines, and amides. This pattern supports the value of integrating molecular fingerprint similarity with predicted chromatographic behavior. Candidates were not only prioritized if they agreed with the known metabolite’s RT value by TC, but also required to agree with the experimentally observed RT. Consequently, the TC-RT approach helped to enhance confidence in candidate assignments, with structures being chosen that were chemically sensible and chromatographically coherent. One interesting finding from the cross-referencing between HMDB and LMSD was that there were many matches in the moderate ranges of the similarity index, and not just at the highest ranges. Particularly, there were 103 annotation records in the range of TC 0.50–0.59 and 173 in the range of TC 0.60–0.69. These intermediate TC values are NOT to be interpreted as negligible similarity. Rather, they may be related to structural changes induced by biotransformation, such as hydroxylation, methylation, oxidation, elongation of the chain, shortening of the chain, and β-oxidation. Molecular fingerprints encode structural features, and when the molecule undergoes biotransformation, the altered chemical structure could also result in a different fingerprint representation, as well as a decreased TC similarity to a known reference compound, even if the molecule structurally or functionally belongs to the same class. Many of the candidates associated with these formulas were acylcarnitines, fatty acid derivatives, hydroxylated lipids, or lipid-like metabolites. Under these circumstances, compounds can have the same molecular formula (or metabolite class) but differ in the position of the substitution, length of chain, branching, or unsaturation. This structural variation may affect RT behavior and could be associated with different biological functions. Thus, medium TC values with high RT agreement can suggest possible metabolite candidates which might otherwise be rejected by strict TC values (e.g., TC ≥ 0.7). Long-chain amines and amides were also found among the prioritized candidates which are consistent with lipid-associated signaling, metabolism associated with membranes and bioactive amide chemistry. This multi-cluster mapping, however, is to be understood as an overlap in structure between reference clusters, rather than as indicative of a unique cluster to which the compound belongs. Candidates not recorded in LMSD but in HMDB were interpreted with caution. The presence or absence in LMSD and HMDB does not imply a different biological category since not all the resources in each database are covered, and lipid entries may be better represented in lipid-specific databases. However, LMSD-only matches were helpful, and some of the prioritized candidates matched lipid-like molecules in dedicated lipidomics databases. These are phospholipids, sterol derivatives, and extended acyl-chain compounds, which highlighted the importance of using complementary metabolite databases for DM annotation. Cross-referencing the databases helped to increase the confidence of a subset of prioritized candidates, but not all the PubChem-derived candidates were identified in HMDB or LMSD. These unmatched candidates could be metabolite candidates not included in existing metabolomics databases, previously unreported metabolites in the host organism, or because of microbial modification, newly detected metabolite candidates, exposure or xenobiotic metabolites, or lower confidence structural assignments that need additional confirmation. Therefore, absence from HMDB or LMSD was not treated as evidence of biological irrelevance. Instead, these candidates were retained as lower-confidence annotations that require additional validation. These assignments are considered as selected possible assignments, but not as actual structural identifications. TC similarity and RT agreement increase confidence in candidate ranking, but definitive metabolite identification still requires MS/MS confirmation, authentic standard comparison, or additional orthogonal structural evidence. Future work will involve expanding candidate evaluation with additional curated resources, such as FooDB, ChEBI, and other specialized biochemical resources. A lower TC threshold of less than 0.5 might lead to more recovery of structurally divergent or missed metabolites, but might also suggest a high level of similarity, possibly overinterpreting the data. Thus, the lower bound of 0.5 was kept for the present workflow as an exploratory limit, whereas the consistency of RT and the support of the database were applied to increase the confidence of the priorities for DM annotations. One limitation of the present approach is that the formal false positive and false negative rates of the DM candidates could not be computed as there were no definitive identities of the unannotated features. Similarly, no negative or null-control candidate set was used in the present analysis. Future applications might use decoy structures, a random approach in assigning formula to structure, authentic standards or a targeted MS/MS validation to determine false-positive rates to further quantify candidate confidence. The current workflow involved in the assignment of potential false positives involved several orthogonal criteria, such as formula/molecular weight matching, TC-based structural projection, RT comparison, and database cross referencing where available. The other limitation includes the application of the biological contextualization layers in this workflow. Database- and model-dependent analyses such as ADME prediction, protein-target prediction, molecular docking, and pathway mapping can be impacted by limited or incomplete pathway coverage, lack of ligand-target interaction information and uncertainty around predicted binding poses or scores. Structural similarity also does not guarantee identical protein binding or biological activity. These analyses were therefore only employed for the purpose of providing biological context to known metabolite clusters, not definitive evidence for the identity of the metabolites. Experimental validation will be required to confirm whether prioritized candidates participate in the predicted pathways or protein-interaction networks. Computational implementation and scalability The workflow was implemented using open-source Python and R software, with fingerprint, descriptor generation with RDKit, RT prediction models from machine learning, and custom Python and R scripts for TC filtering and priority scoring. While the approach is high computational complexity as large candidate spaces are derived from PubChem and then searched, fingerprinted, scored and ranked, it is scalable by batch processing and modular analysis. This approach allows for flexibility in the workflow when applied to other metabolomics data sets, thresholds for TCs and different validation criteria, whilst maintaining transparency and reproducibility. Conclusion The analysis introduced here is a two-phase analytical approach that offers a detailed interpretation of the plasma metabolome. This study presents a cluster-guided and RT-informed workflow for prioritizing dark matter metabolite candidates in untargeted plasma metabolomics. Using known metabolites, structurally coherent chemical neighborhoods were defined and candidates for unannotated features were removed by maximum pairwise Tanimoto similarity, resulting in further ranking by agreement between predicted and experimentally observed retention time. This TC-RT approach resulted in less redundancy of candidates and better discrimination of structurally similar molecules, including positional isomers. The framework is designed to be used alongside the traditional MS/MS based annotation to provide extra prioritization for features with little, ambiguous, or nonexistent spectral data. Overall, this approach provides a transparent and biologically informed strategy for prioritizing candidate dark matter metabolites for future experimental validation.. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file1 (JPG 3445 KB) Supplementary file2 (PDF 4770 KB) Supplementary file3 (JPG 9583 KB) Supplementary file4 (XLSX 13107 KB) Supplementary file5 (XLSX 11 KB) Supplementary file6 (XLSX 3707 KB) Supplementary file7 (XLSX 299 KB) Acknowledgements The authors gratefully acknowledge the ACNC MEND and Physical Activity Cores for sample collection, and the ACNC Metabolomics Core for their expertise in sample preparation and metabolomic analyses. The authors also acknowledge the support of the ACNC Biostatistics and Data Innovation Core (BaDI). Author contributions DB and HAP: Writing: review & editing, visualization, validation, data curation, formal analysis, KH, KKA, AMV, HAA, BPD, RLS: Data curation, formal analysis and editing EB: Editing, resources CKD: Methodology, investigation, review & editing, funding acquisition SVC: Writing original draft, review & editing, visualization, validation, software, methodology, investigation, formal analysis, data curation, conceptualization. Funding This work was supported by the USDA-ARS under Project Number 6026-10700-001-000D. Data availability Data supporting the findings of this study are available within the paper and Supplementary Information files along with deposition in the Metabolomics Workbench. This study is available at the NIH Common Fund's National Metabolomics Data Repository (NMDR) website, the Metabolomics Workbench, https://www.metabolomicsworkbench.org where it has been assigned Study ID ST005101. The data can be accessed directly via its Project DOI: https://doi.org/10.21228/M8XZ9B. Declarations Conflict of interest The authors declare no competing interests. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Dipendra Bhandari and Henry A. Paz contributed equally to this work and shared first authorship. References Associated Data Data Citations Supplementary Materials Supplementary file1 (JPG 3445 KB) Supplementary file2 (PDF 4770 KB) Supplementary file3 (JPG 9583 KB) Supplementary file4 (XLSX 13107 KB) Supplementary file5 (XLSX 11 KB) Supplementary file6 (XLSX 3707 KB) Supplementary file7 (XLSX 299 KB) Data Availability Statement Data supporting the findings of this study are available within the paper and Supplementary Information files along with deposition in the Metabolomics Workbench. This study is available at the NIH Common Fund's National Metabolomics Data Repository (NMDR) website, the Metabolomics Workbench, https://www.metabolomicsworkbench.org where it has been assigned Study ID ST005101. The data can be accessed directly via its Project DOI: https://doi.org/10.21228/M8XZ9B.