Pith. sign in

REVIEW 4 major objections 4 minor 98 references

MetaboLLM: a metabolomics-specialized large language model for biochemical knowledge integration and predictive metabolite graph construction

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that domain-specialized language models can organize heterogeneous biochemical knowledge into a metabolite graph that predicts patient outcomes better than curated pathway graphs, with AUC 0.8616 and 0.8123 in two…

desk verdict A serious and novel framework paper whose central graph comparison is weaker than claimed, but the work is well-engineered and deserves peer review. read the letter →

arxiv 2608.06253 v1 pith:VHFMKTGQ submitted 2026-08-06 cs.LG

Dohyun Ku , Min Gu Kwak , Francisco J. Pasquel , Jing Li This is my paper
classification cs.LG
keywords metabolomicslargelanguagemodelmetabolitegraphconstructionisomorphismnetworkretrieval-augmentedgenerationdomainadaptationpatient-levelpredictionbiochemicalknowledgeintegration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

MetaboLLM is a metabolomics-specialized language model built by harmonizing four public metabolomics databases into one corpus, then adapting general-purpose backbones to that corpus through continual pretraining and supervised fine-tuning, with structured retrieval of reaction and pathway context. The paper claims that this adaptation gives the model biochemical knowledge that general medical models lack, and that the model's generated descriptions can organize that knowledge into a graph. The decisive step is converting descriptions into topology: description embeddings are compared by cosine similarity, and the most similar metabolite pairs become edges in a shared graph whose nodes carry patient-specific metabolite abundances. A graph isomorphism network then uses this graph for prediction, reaching AUC 0.8616 for stress hyperglycemia after coronary artery bypass grafting and 0.8123 for postmenopausal hormone-regimen classification, outperforming conventional predictors and alternative graph constructions. If the claim holds, a specialized language model becomes not just a text generator but a way to turn dispersed, incomplete biochemical annotations into predictive and interpretable patient-level representations.

What carries the argument

The load-bearing mechanism is a description-to-graph conversion. For each measured metabolite, MetaboLLM produces a retrieval-grounded description; a biomedical text embedder turns each description into a vector $\mathbf{e}_i$, and the cosine similarity $S_{ij}=(\mathbf{e}_i\cdot\mathbf{e}_j)/(\|\mathbf{e}_i\|\|\mathbf{e}_j\|)$ ranks metabolite pairs, with the top-$k$ pairs selected as undirected graph edges. Sample-specific metabolite abundances form the node feature matrix $X_s$, and a graph isomorphism network updates node representations via $h_v^{(\ell)}=\mathrm{MLP}^{(\ell)}\big((1+\epsilon^{(\ell)})h_v^{(\ell-1)}+\sum_{u\in\mathcal{N}(v)}h_u^{(\ell-1)}\big)$, concatenating readouts from each layer before classification. The graph topology is shared across samples, so only the measured abundances vary; the model adaptation (continual pretraining on templated database records, then supervised fine-tuning on recall, classification, relation, and description tasks) is what makes the descriptions informative enough for this similarity topology to beat curated graphs.

What would settle it

A decisive check would be to rerun the stress-hyperglycemia pipeline after randomly permuting which of the 1,100 metabolite identities (and their descriptions and graph edges) are assigned to the measured abundance features, keeping everything else fixed; if the AUC stays near 0.8616 the biochemical content of the graph is not carrying the signal, whereas a clear drop would confirm that correct metabolite identity and semantic edges matter.

Watch

Extended reading notes

Core claim

The paper's central claim is that semantic similarity among MetaboLLM-generated biochemical descriptions is a better source of metabolite connectivity than curated pathway or reaction annotations for patient-level prediction. For each measured metabolite, the model writes a retrieval-grounded description that combines biological roles, reaction context, and pathway context; the descriptions are embedded, pairwise cosine similarities are computed, and the top-ranked pairs define the edges of a shared metabolite graph. Sample-specific measured abundances are placed on the nodes, and a graph isomorphism network combines the topology with node features for classification. This construction achieved the highest AUC in both applications, 0.8616 for stress hyperglycemia after coronary artery bypass grafting and 0.8123 for postmenopausal hormone-regimen classification, beating standard machine-learning baselines, curated pathway and reaction graphs, correlation graphs, random graphs, and graphs from an unadapted LLM with or without retrieval. The authors interpret the enrichment of high-importance edges for shared reactions and pathways, and the pathway enrichment of high-importance metabolites, as evidence that the learned topology is biochemically coherent rather than arbitrary.

Load-bearing premise

The measured metabolite features are assumed to be correctly identified and matched to the unified knowledge database, so the model descriptions, graph edges, and pathway interpretations are all built about the right molecules.

Editorial extensions

If this is right

  • Across four backbone families, the metabolomics-adapted model outperforms its unadapted and medically adapted counterparts on metabolomics knowledge and description tasks, so the benefit is not limited to a single language-model architecture.
  • Graphs from retrieval-grounded MetaboLLM descriptions consistently beat graphs from the unadapted backbone and from MetaboLLM without retrieval, showing that domain adaptation and retrieved evidence both contribute to the predictive topology.
  • Sparse graphs perform best, with peak AUC at 200 edges for stress hyperglycemia and 100 edges for hormone-regimen classification, implying that adding weak semantic links dilutes the signal.
  • High-importance edges are significantly enriched for shared pathways and reactions relative to randomized graphs, and top-ranked metabolites enrich for coherent pathways such as bile-acid, retinoid, steroid, and xenobiotic metabolism.
  • The metabolite-reaction-enzyme structure provides a natural route toward metabolite-protein-gene graphs and multi-omics prediction, as the paper itself notes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the same description-to-graph recipe could build predictive topologies for proteins, genes, or other molecular entities, provided a specialized language model can generate grounded textual descriptions for them.
  • Beyond the paper: the current descriptions are context-free; conditioning them on disease, tissue, or treatment context before edge selection is a natural follow-up that might sharpen outcome-specific predictions.
  • Beyond the paper: annotation uncertainty could be folded into the graph by making edge weights or inclusion depend on confidence in metabolite identity, rather than treating every matched feature as equally certain.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This paper presents MetaboLLM, a metabolomics-specialized large language model built by continual pretraining and supervised fine-tuning on a harmonized corpus of KEGG, HMDB, PubChem, and SMPDB, with retrieval-augmented description generation, and MetaboLLM-GIN, which converts the generated descriptions into a metabolite similarity graph and uses a graph isomorphism network for patient-level prediction. The authors report improved accuracy over base and medical LLMs on internal metabolomics tasks and on the external MetaBench benchmark, and report AUCs of 0.8616 for stress hyperglycemia after CABG and 0.8123 for postmenopausal hormone-regimen classification, outperforming conventional ML baselines, random/correlation/pathway/reaction graph baselines, and LLM-derived graph variants. They also provide pathway enrichment and edge-level biological validation. The paper is generally well written, transparent about limitations, and accompanied by public code, corpus, and model weights.

Significance. If the comparative claims hold, this is a useful demonstration that a domain-adapted LLM can serve as a knowledge representation layer for metabolomics, with public corpus, code, and model weights. Strengths include the multi-backbone adaptation study, the within-compound retrieval ablation, the external MetaBench transfer test, and the explicit discussion of limitations such as training/evaluation resource overlap, k-selection optimism, and annotation uncertainty. The main comparative claim, however, is currently weakened by the asymmetric graph baselines and a partly circular biological validation, so the paper needs additional experiments before the central claim can be accepted as stated.

major comments (4)
  1. [Methods: Alternative and ablation graph construction strategies] The comparison between the LLM-derived graph and the curated pathway/reaction baselines is asymmetric. When the candidate set exceeds the target size k, the pathway- and reaction-based graphs are obtained by sampling k edges uniformly at random for each seed, whereas the MetaboLLM graph always keeps the top-k pairs by description cosine similarity. This confounds the value of the language-derived ranking with the value of any selective sparsification of a larger candidate set; a baseline that picked the top-k curated edges by, for example, number of shared pathways or by chemical-structure similarity could in principle match or exceed the reported AUCs. I ask for a density-matched but biochemically ranked baseline, and ideally the same edge-count sweep for the MTBLS136 cohort that is shown for CABG in Fig. 3b. The authors' acknowledgment that k was selected on the same cohorts does not address this asymmetry.
  2. [Model interpretation and biological validation] The biological validation is partly circular. The RAG pipeline retrieves reaction- and pathway-level context and inserts it into the descriptions whose embeddings define the graph, and the validation then checks the same edges against shared KEGG/SMPDB reactions and pathways. The randomized-graph null distribution only controls for choosing edges uniformly at random; it does not control for the fact that the descriptions were constructed from the annotations being used as ground truth. The authors should validate edges generated without retrieval, or hold out a subset of pathways/reactions and test whether the graph predicts those withheld annotations, or use an independent source of functional relatedness such as HMDB disease annotations or literature co-occurrence as the validation target.
  3. [MetaboLLM evaluation protocol / Discussion] The internal benchmark is constructed from the same unified corpus used for CPT and SFT, so the reported gains on the 12 closed-form tasks and the five description tasks could partly reflect memorization of training data rather than transferable domain knowledge. The authors state this limitation in the Discussion, but they do not report an analysis of train/test overlap at the sample level or a split in which entire source databases are held out. Because the external MetaBench transfer uses PathBank, a source not in the corpus, that result is reassuring, but the Knowledge-task gains over unadapted backbones are small (e.g., 52.18 versus 51.42 for Qwen3-4B in Table 5), so a formal leakage analysis or held-out-source evaluation would materially strengthen the central knowledge-integration claim.
  4. [Stress hyperglycemia prediction] The clinical applications rest entirely on putative metabolite annotations. The authors explicitly flag that 'uncertainty in compound matching can propagate through description generation, graph construction, and pathway analysis,' but no sensitivity analysis is provided. For the small CABG cohort (n=61), a modest fraction of misannotated features could change the graph topology and the enrichment results. I request an analysis restricting to high-confidence annotations (for example, by Mummichog p-value or by requiring support from multiple annotation sources) or a bootstrap over the annotation set, to show that the main AUC and enrichment conclusions are stable.
minor comments (4)
  1. [Results: Stress hyperglycemia prediction] The statistical significance tests are paired t-tests across 20 seeds, but the five-fold cross-validation in the CABG cohort reuses the same patients across folds, so the effective number of independent observations is smaller than 20; a permutation test respecting the fold structure or a reported confidence interval would be more appropriate.
  2. [Results: Biochemical description generation] The description-generation metrics are computed against GPT-5-mini-generated references; the Discussion notes this, but the Results text should state more prominently that BERTScore-F1, ROUGE-L-F1, and BLEU-2 measure agreement with a teacher model rather than biochemical correctness.
  3. [Results: External benchmark evaluation] The sentence 'MetaboLLM-Qwen3-4B exceeded that range' in the MetaBench section should be softened, since the comparison values were obtained under different prompting and decoding settings, as the authors themselves note; the current wording overstates the direct comparability.
  4. [Abstract / Results] Minor editorial issues: 'Each MetaboLLM variants outperformed its corresponding...' and 'outperformed every evaluated medical models' contain subject-verb agreement errors and should be corrected.

Circularity Check

2 steps flagged · score 5.0 of 10

Partial circularity: graph-edge validation re-uses the same KEGG/SMPDB reaction and pathway annotations that were retrieved into the descriptions from which the edges are built, and internal description metrics partly grade agreement with GPT-5-mini teacher references; the central AUC comparisons are empirical and not circular.

  1. self definitional [Results: 'Biological validation of important graph edges'; Methods: 'Retrieval-augmented description generation' and 'Model interpretation and biological validation']
    "Retrieved reaction- and pathway-level contexts were incorporated into structured prompts for structure-rich metabolite description generation. ... An edge was considered a reaction match when both metabolites participated in at least one common KEGG reaction and a pathway match when they shared at least one KEGG or SMPDB pathway."

    The top-k edge set is computed from cosine similarity of descriptions produced with these RAG prompts, so reaction/pathway co-membership information is an input to the representation from which edges are selected. The edge validation then scores the same edges against the same KEGG reaction and KEGG/SMPDB pathway co-membership annotations. For the retrieval-grounded MetaboLLM graph, a high match ratio therefore largely confirms that the retrieved context was encoded in the text and embeddings, rather than providing independent evidence that the learned topology is biochemically coherent. The randomized-graph null only controls for edge density, not for the provenance of the descriptions.

  2. fitted input called prediction [Discussion, Limitations; Methods: 'Supervised fine-tuning for task alignment']
    "As supervised training and internal evaluation used GPT-5-mini-generated descriptions, the internal metrics partly measure agreement with teacher references rather than biochemical correctness."

    The SFT stage trains the model to emit GPT-5-mini-generated descriptions ('Responses for both regimes were generated using GPT-5-mini and appended to the conversational SFT samples'), and the internal description-generation tasks are scored against references from the same benchmark construction. Reporting these as description-generation performance therefore partly measures reproduction of the training/teacher distribution. The paper explicitly acknowledges this limitation, and the external PathBank-derived MetaBench task provides a mitigating but not fully matched check.

full rationale

The central AUC claims are not circular: the comparisons are empirical, the same GIN architecture is used across graph constructions, and the LLM-derived ablations (base vs adapted, RAG vs no-RAG) share the same top-k edge selection. The main circularity is in the biological validation: the RAG prompt injects the exact reaction and pathway context later used as ground truth for edge-precision and match-ratio tests, so those numbers are partly self-confirming. Internal knowledge and description benchmarks also overlap with the CPT/SFT resources, and the paper acknowledges that internal description metrics partly measure agreement with GPT-5-mini teacher references; the external MetaBench results provide independent support for the language-model claims. I found no load-bearing self-citation: the one co-authored citation (ref. 61) is secondary and accompanied by an independent reference (ref. 60). The curated-graph baselines sample random k-edge subsets when candidate sets exceed k, which is a fairness confound rather than a circularity, and it does not make the graph AUC equal to an input by construction. Overall, the derivation chain is partially circular in its validation and evaluation scaffolding, but the primary predictive results retain independent empirical content.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on public database quality, identifier harmonization, Mummichog annotation reliability, and the assumption that semantic similarity of LLM descriptions approximates functional biochemical relatedness. The edge count k and GIN and retrieval hyperparameters are set by performance or by hand, and the biological validation is not independent of the knowledge used to build the descriptions.

free parameters (3)
  • Edge count k for graph topology = 200 for CABG, 100 for hormone-regimen dataset
    Selected based on predictive performance within the same cohorts, as acknowledged in the Discussion: 'absolute performance at the selected graph size is likely optimistic.'
  • GIN hyperparameters per dataset = CABG: hidden 96, layers 2, dropout 0.5, LR 0.0006, weight decay 0.01; hormone: hidden 80, layers 3, dropout 0.75, LR…
    Chosen per dataset without a clearly described nested selection procedure, so outcome information may have influenced architecture choices.
  • Retrieval ranking parameters = bio weight 0.5, chem weight 0.5, max reaction metabolites 3, max pathways 3, metabolites per pathway 1
    Hand-set parameters that control which reaction and pathway context enters the descriptions, thereby influencing graph construction.
assumptions (4)
  • domain assumption KEGG, HMDB, PubChem, and SMPDB records are accurate, and union-find identifier matching correctly merges duplicate metabolite records.
    The entire corpus, training data, descriptions, and graph construction depend on the quality of cross-database harmonization described in Methods: Knowledge source integration.
  • domain assumption Mummichog putative annotations correctly identify the metabolites measured in the clinical cohorts.
    Only metabolites matched to the unified database are retained for prediction, and the authors note that annotation uncertainty can propagate through the pipeline.
  • ad hoc to paper Cosine similarity between LLM-generated description embeddings is a valid proxy for functional biochemical relatedness relevant to patient outcomes.
    This is the core graph-construction assumption introduced in Methods: LLM-based graph construction, and it is not independently validated against an external definition of metabolic relatedness.
  • domain assumption GNNExplainer importance scores identify edges and metabolites that are biologically meaningful rather than artifacts of model training.
    Used for biological interpretation; the authors concede that these scores indicate importance to model predictions, not causal biological effects.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MetaboLLM: a metabolomics-specialized large language model for biochemical knowledge integration and predictive metabolite graph construction." pith.science (2026). https://pith.science/paper/VHFMKTGQ

@misc{pith2026260806253,
  author       = {Pith},
  title        = {Pith review of: MetaboLLM: a metabolomics-specialized large language model for biochemical knowledge integration and predictive metabolite graph construction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VHFMKTGQ}},
  note         = {Machine review of arXiv:2608.06253}
}
read the original abstract

Metabolomics knowledge is distributed across heterogeneous resources and remains difficult to translate into predictive representations. We developed MetaboLLM, a metabolomics-specialized large language model adapted through continual pretraining, supervised fine-tuning, and structured retrieval, together with MetaboLLM-GIN, which converts generated biochemical descriptions into metabolite graphs for patient-level prediction using a graph isomorphism network. Across four backbone families, MetaboLLM outperformed corresponding base and medically adapted models on metabolomics knowledge, relational, and description tasks, and transferred to an external public benchmark. MetaboLLM-GIN achieved the highest AUC for stress hyperglycemia prediction after coronary artery bypass grafting (0.8616) and postmenopausal hormone-regimen classification (0.8123), outperforming conventional models, alternative graph constructions, and graphs generated from unadapted or non-retrieval LLM configurations. Model interpretation further produced biologically meaningful findings in both applications. These results show that domain-specialized language models can organize heterogeneous biochemical knowledge into predictive and interpretable metabolite graph representations.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

98 extracted references · 69 canonical work pages

  1. [1]

    Holmes, E., Wilson, I. D. & Nicholson, J. K. Metabolic phenotyping in health and disease. Cell 134, 714–717 (2008)

    2008

  2. [2]

    J., Yanes, O

    Patti, G. J., Yanes, O. & Siuzdak, G. Metabolomics: the apogee of the omics trilogy. Nature reviews Molecular cell biology 13, 263–269 (2012)

    2012

  3. [3]

    Dettmer, K., Aronov, P. A. & Hammock, B. D. Mass spectrometry‐based metabolomics. Mass spectrometry reviews 26, 51–78 (2007)

    2007

  4. [4]

    Alseekh, S. et al. Mass spectrometry-based metabolomics: a guide for annotation, quantification and best reporting practices. Nature methods 18, 747–756 (2021)

    2021

  5. [5]

    Wang, T. J. et al. Metabolite profiles and the risk of developing diabetes. Nature medicine 17, 448– 453 (2011)

    2011

  6. [6]

    H., Ivanisevic, J

    Johnson, C. H., Ivanisevic, J. & Siuzdak, G. Metabolomics: beyond biomarkers and towards mechanisms. Nature reviews Molecular cell biology 17, 451–459 (2016)

    2016

  7. [7]

    Wishart, D. S. Emerging applications of metabolomics in drug discovery and precision medicine. Nature reviews Drug discovery 15, 473–484 (2016)

    2016

  8. [8]

    Newgard, C. B. Metabolomics and metabolic diseases: where do we stand? Cell metabolism 25, 43– 56 (2017)

    2017

Show all 98 references
  1. [9]

    R., Warth, B., Spilker, M

    Guijas, C., Montenegro-Burke, J. R., Warth, B., Spilker, M. E. & Siuzdak, G. Metabolomics activity screening for identifying metabolites that modulate phenotype. Nature biotechnology 36, 316–320 (2018)

    2018

  2. [10]

    M., Ivanisevic, J., Giera, M

    Rinschen, M. M., Ivanisevic, J., Giera, M. & Siuzdak, G. Identification of bioactive metabolites using activity metabolomics. Nature reviews Molecular cell biology 20, 353–367 (2019)

    2019

  3. [11]

    & Tanabe, M

    Kanehisa, M., Furumichi, M., Sato, Y ., Ishiguro-Watanabe, M. & Tanabe, M. KEGG: integrating viruses and cellular organisms. Nucleic Acids Research 49, D545–D551 (2021). 31

    2021

  4. [12]

    Wishart, D. S. et al. HMDB 5.0: the Human Metabolome Database for 2022. Nucleic Acids Research 50, D622–D631 (2022)

    2022

  5. [13]

    Kim, S. et al. PubChem in 2021: new data content and improved web interfaces. Nucleic Acids Research 49, D1388–D1395 (2021)

    2021

  6. [14]

    Jewison, T. et al. SMPDB 2.0: Big Improvements to the Small Molecule Pathway Database. Nucl. Acids Res. 42, D478–D484 (2014)

    2014

  7. [15]

    Redestig, H. et al. Consolidating metabolite identifiers to enable contextual and multi-platform metabolomics data analysis. BMC bioinformatics 11, 214 (2010)

    2010

  8. [16]

    Kumar, A., Suthers, P. F. & Maranas, C. D. MetRxn: a knowledgebase of metabolites and reactions spanning metabolic models and databases. BMC bioinformatics 13, 6 (2012)

    2012

  9. [17]

    Braisted, J. et al. RaMP-DB 2.0: a renovated knowledgebase for deriving biological and chemical insight from metabolites, proteins, and genes. Bioinformatics 39, btac726 (2023)

    2023

  10. [18]

    & Mehl, F

    Moretti, S., Niknejad, A., Pagni, M. & Mehl, F. MetaNetX: a bridge between metabolic resources for enhanced curation and multi-omics data harmonization. Nucleic Acids Research 54, D617–D622 (2026)

    2026

  11. [19]

    Singhal, K. et al. Large language models encode clinical knowledge. Nature 620, 172–180 (2023)

    2023

  12. [20]

    Liu, F. et al. Application of large language models in medicine. Nature Reviews Bioengineering 3, 445–464 (2025)

    2025

  13. [21]

    Singhal, K. et al. Toward expert-level medical question answering with large language models. Nature medicine 31, 943–950 (2025)

    2025

  14. [22]

    Wu, C. et al. PMC-LLaMA: toward building open-source language models for medicine. Journal of the American Medical Informatics Association 31, 1833–1843 (2024)

    2024

  15. [23]

    Chen, Z. et al. Meditron-70b: Scaling medical pretraining for large language models. arXiv preprint arXiv:2311.16079 (2023)

    2023 arXiv

  16. [24]

    Xie, Q. et al. Medical foundation large language models for comprehensive text analysis and beyond. npj Digit. Med. 8, 141 (2025). 32

    2025

  17. [25]

    & Zhang, A

    Xiong, G., Jin, Q., Lu, Z. & Zhang, A. Benchmarking retrieval-augmented generation for medicine. in Findings of the Association for Computational Linguistics: ACL 2024 6233–6251 (2024)

    2024

  18. [26]

    & Zhang, R

    Li, M., Kilicoglu, H., Xu, H. & Zhang, R. Biomedrag: A retrieval augmented large language model for biomedicine. Journal of Biomedical Informatics 162, 104769 (2025)

    2025

  19. [27]

    Qiang, H. et al. Language model-guided anticipation and discovery of mammalian metabolites. Nature 1–10 (2026)

    2026

  20. [28]

    & Xie, P

    Guo, Z., Duan, D., Liang, Y ., Patil, A. & Xie, P. MetaboliteChat: A Unified Multimodal Large Language Model for Interactive Metabolite Analysis and Functional Insights. bioRxiv 2025.11. 07.687008 (2025)

    2025

  21. [29]

    Bekbergenova, M. et al. MetaboT: AI-based agent for natural language-based interaction with metabolomics knowledge graphs. arXiv preprint arXiv:2510.01724 (2025)

    2025 arXiv

  22. [30]

    Lu, Y . et al. MetaBench: a multi-task benchmark for assessing LLMs in metabolomics. in Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) 32647–32668 (2026)

    2026

  23. [31]

    Barupal, D. K. et al. MetaMapp: mapping and visualizing metabolomic data by integrating information from biochemical pathways and chemical and mass spectral similarity. BMC bioinformatics 13, 99 (2012)

    2012

  24. [32]

    & Perera-Lluna, A

    Picart-Armada, S., Fernández-Albert, F., Vinaixa, M., Yanes, O. & Perera-Lluna, A. FELLA: an R package to enrich metabolomics data. BMC bioinformatics 19, 538 (2018)

    2018

  25. [33]

    Knowledge graph and large language model for metabolomics

    Lu, Y . Knowledge graph and large language model for metabolomics. in Proceedings of the AAAI Conference on Artificial Intelligence vol. 39 29281–29282 (2025)

    2025

  26. [34]

    & Nothias, L.-F

    Féraud, M., Boukhajou, D., Gandon, F. & Nothias, L.-F. MetaboKG: An Analysis-centric Knowledge Graph Framework for Untargeted Metabolomics. arXiv preprint arXiv:2605.24706 (2026)

    2026 arXiv

  27. [35]

    & Celik, M

    Cakmak, A. & Celik, M. H. Personalized metabolic analysis of diseases. IEEE/ACM transactions on computational biology and bioinformatics 18, 1014–1025 (2020). 33

    2020

  28. [36]

    Vaida, M. et al. M-GNN: A graph neural network framework for lung cancer detection using metabolomics and heterogeneous graph modeling. International Journal of Molecular Sciences 26, 4655 (2025)

    2025

  29. [37]

    & Graham, C

    Yang, Z. & Graham, C. Hypercholesterolemia Risk Prediction from Serum Metabolomics Using a Metabolic Pathway-Integrated Graph Neural Network. bioRxiv 2025.10. 14.682480 (2025)

    2025

  30. [38]

    Monteiro, L. M. et al. PathwaySeeker: Evidence-Grounded AI Reasoning over Organism-Specific Metabolic Networks. bioRxiv 2026.04. 14.718256 (2026)

    2026

  31. [39]

    & Theis, F

    Krumsiek, J., Suhre, K., Illig, T., Adamski, J. & Theis, F. J. Gaussian graphical modeling reconstructs pathway reactions from high-throughput metabolomics data. BMC systems biology 5, 21 (2011)

    2011

  32. [40]

    Basu, S. et al. Sparse network modeling and metscape-based visualization methods for the analysis of large-scale metabolomics data. Bioinformatics 33, 1545–1553 (2017)

    2017

  33. [41]

    Benedetti, E. et al. A strategy to incorporate prior knowledge into correlation network cutoff selection. Nature communications 11, 5153 (2020)

    2020

  34. [42]

    Garcia-Gasulla, D. et al. The aloe family recipe for open and specialized healthcare llms. npj Digital Medicine (2026)

    2026

  35. [43]

    II-Medical-8B: Medical Reasoning Model

    Intelligent Internet. II-Medical-8B: Medical Reasoning Model. Hugging Face https://huggingface.co/Intelligent-Internet/II-Medical-8B (2025)

    2025

  36. [44]

    Qwen2.5 Meditron-3[7B]

    OpenMeditron Initiative. Qwen2.5 Meditron-3[7B]. Hugging Face https://huggingface.co/EPFLiGHT/Meditron3-Qwen2.5-7B (2025)

    2025

  37. [45]

    Sellergren, A. et al. Medgemma 1.5 technical report. arXiv preprint arXiv:2604.05081 (2026)

    2026 arXiv

  38. [46]

    Yu, H. et al. FinemedLM-o1: Enhancing medical knowledge reasoning ability of LLM from supervised fine-tuning to test-time training. arXiv preprint arXiv:2501.09213 (2025)

    2025 arXiv

  39. [47]

    Yang, A. et al. Qwen3 Technical Report. Preprint at https://doi.org/10.48550/arXiv.2505.09388 (2025)

    2025 doi arXiv

  40. [48]

    Team, G. et al. Gemma 3 Technical Report. Preprint at https://doi.org/10.48550/arXiv.2503.19786 (2025). 34

    2025 doi arXiv

  41. [49]

    Llama 3.2 3B Instruct

    Meta. Llama 3.2 3B Instruct. Hugging Face https://huggingface.co/meta-llama/Llama-3.2-3B- Instruct (2024)

    2024

  42. [50]

    Zhang, T., Kishore, V ., Wu, F., Weinberger, K. Q. & Artzi, Y . Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675 (2019)

    2019 arXiv

  43. [51]

    Rouge: A package for automatic evaluation of summaries

    Lin, C.-Y . Rouge: A package for automatic evaluation of summaries. in Text summarization branches out 74–81 (2004)

    2004

  44. [52]

    & Zhu, W.-J

    Papineni, K., Roukos, S., Ward, T. & Zhu, W.-J. Bleu: a method for automatic evaluation of machine translation. in Proceedings of the 40th annual meeting of the Association for Computational Linguistics 311–318 (2002)

    2002

  45. [53]

    Li, S. et al. Predicting network activity from high throughput metabolomics. PLoS computational biology 9, e1003123 (2013)

    2013

  46. [54]

    & Leskovec, J

    Ying, Z., Bourgeois, D., You, J., Zitnik, M. & Leskovec, J. Gnnexplainer: Generating explanations for graph neural networks. Advances in neural information processing systems 32, (2019)

    2019

  47. [55]

    L., Wang, Y ., Carter, B

    Stevens, V . L., Wang, Y ., Carter, B. D., Gaudet, M. M. & Gapstur, S. M. Serum metabolomic profiles associated with postmenopausal hormone use. Metabolomics 14, 97 (2018)

    2018

  48. [56]

    Wishart, D. S. et al. PathBank 2.0—the pathway database for model organism metabolomics. Nucleic acids research 52, D654–D662 (2024)

    2024

  49. [57]

    WERB, M. R. et al. Hormonal and metabolic responses during coronary artery bypass surgery: role of infused glucose. The Journal of Clinical Endocrinology & Metabolism 69, 1010–1018 (1989)

    1989

  50. [58]

    Roth-Isigkeit, A. et al. Persistent endocrine stress response in patients undergoing cardiac surgery. Journal of endocrinological investigation 21, 12–19 (1998)

    1998

  51. [59]

    E., Brismar, K., Barr, G

    Anderson, R. E., Brismar, K., Barr, G. & Ivert, T. Effects of cardiopulmonary bypass on glucose homeostasis after coronary artery bypass surgery. European journal of cardio-thoracic surgery 28, 425–430 (2005). 35

    2005

  52. [60]

    Neu, F. et al. Targeted metabolomic profiling reveals inflammation–associated longitudinal changes in plasma metabolites following on-pump coronary bypass surgery. Frontiers in medicine 12, 1673132 (2025)

    2025

  53. [61]

    Ku, D. et al. Metabolic Trajectories During Surgical Stress in Patients Undergoing Cardiac Surgery. bioRxiv (2025)

    2025

  54. [62]

    Schindler, R. et al. Plasma vitamin E and A changes during cardiopulmonary bypass and in the postoperative course. Langenbeck’ s Archives of Surgery 387, 372–378 (2003)

    2003

  55. [63]

    Graham, T. E. et al. Retinol-binding protein 4 and insulin resistance in lean, obese, and diabetic subjects. New England Journal of Medicine 354, 2552–2563 (2006)

    2006

  56. [64]

    Espeland, M. A. et al. Effect of postmenopausal hormone therapy on glucose and insulin concentrations. Diabetes care 21, 1589–1595 (1998)

    1998

  57. [65]

    Lindheim, S. R. et al. A possible bimodal effect of estrogen on insulin sensitivity in postmenopausal women and the attenuating effect of added progestin. Fertility and sterility 60, 664–667 (1993)

    1993

  58. [66]

    Jung, J. H. et al. Serum uric acid levels and hormone therapy type: a retrospective cohort study of postmenopausal women. Menopause 25, 77–81 (2018)

    2018

  59. [67]

    Hak, A. E. & Choi, H. K. Menopause, postmenopausal hormone use and serum uric acid levels in US women–the Third National Health and Nutrition Examination Survey. Arthritis research & therapy 10, R116 (2008)

    2008

  60. [68]

    Hu, E. J. et al. Lora: Low-rank adaptation of large language models. ICLR 1, 3 (2022)

    2022

  61. [69]

    & Ramsundar, B

    Ahmad, W., Simon, E., Chithrananda, S., Grand, G. & Ramsundar, B. ChemBERTa-2: Towards Chemical Foundation Models. Preprint at https://doi.org/10.48550/arXiv.2209.01712 (2022)

    2022 doi arXiv

  62. [70]

    SMILES, a chemical language and information system

    Weininger, D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J. Chem. Inf. Comput. Sci. 28, 31–36 (1988)

    1988

  63. [71]

    NeuML/pubmedbert-base-embeddings

    Mezzetti, D. NeuML/pubmedbert-base-embeddings. (2023)

    2023

  64. [72]

    {name_part} has the molecular formula {formula}. {description}

    Singh, A. et al. Openai gpt-5 system card. arXiv preprint arXiv:2601.03267 (2025). 36 Supplementary Information A. CPT Data Construction Details This appendix provides detailed information on the continual pretraining (CPT) data construction process, including database integra...

    2025 arXiv

  65. [73]

    Biochemical compounds: metabolite records with text representations, chemical and biological embeddings, and cross-database identifiers, including KEGG ID, HMDB ID, and PubChem ID. 50

  66. [74]

    Each record included reaction text, participating metabolite identifiers, and enzyme annotations

    Reactions: biochemical reaction records linking metabolites through enzymatic transformations. Each record included reaction text, participating metabolite identifiers, and enzyme annotations

  67. [75]

    This database structure enabled metadata-filtered retrieval of biochemically related metabolites before embedding-based ranking

    Pathways: metabolic pathway records with pathway descriptions and participating metabolite identifiers. This database structure enabled metadata-filtered retrieval of biochemically related metabolites before embedding-based ranking. B.3.2 Metadata-Filtered Candidate Retrieval ...

  68. [76]

    Query the reactions collection for all reactions containing the input metabolite

  69. [77]

    Extract all other metabolite IDs from these reactions as candidates

  70. [78]

    For each candidate, compute weighted similarity to the input: 𝑠𝑠𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑐𝑖𝑖𝑛𝑛𝑒𝑒𝑑𝑑 = 𝑤𝑤𝑐𝑐𝑖𝑖𝑐𝑐 ⋅ 𝑠𝑠𝑐𝑐𝑖𝑖𝑐𝑐 + 𝑤𝑤𝑐𝑐ℎ𝑒𝑒𝑐𝑐 ⋅ 𝑠𝑠𝑐𝑐ℎ𝑒𝑒𝑐𝑐

  71. [79]

    Filter candidates exceeding the duplicate similarity threshold (0.98)

  72. [80]

    Apply name-based deduplication to remove identical metabolites with different IDs

  73. [81]

    Select top-𝑘𝑘 candidates ranked by similarity

  74. [82]

    Step 3: Pathway-Based Retrieval

    Each selected metabolite is annotated with the reaction texts linking it to the input, providing explicit biochemical context. Step 3: Pathway-Based Retrieval. We identify metabolites that co-participate in pathways with the input:

  75. [83]

    Query the pathways collection for all pathways containing the input metabolite

  76. [84]

    If more pathways exist than the limit, rank by biological similarity to the input and select top pathways

  77. [85]

    You are a biomedical expert in metabolomics and biochemical pathways. Base your reasoning on the provided contexts and integrate general biochemical knowledge

    For each selected pathway: a. Extract metabolite IDs, excluding the input and previously selected metabolites b. Compute weighted similarity for each candidate c. Apply duplicate and name-based filtering d. Select top metabolites per pathway 51 This hierarchical approach ensur...

  78. [93]

    Its connections to other metabolites through the provided reactions and pathways

  79. [94]

    Its overall significance in cellular metabolism” Structure-poor Task Variants Base “Write a 300–350 word description covering:

  80. [97]

    Its overall significance in cellular metabolism” v1 “Provide a comprehensive description (300–350 words) addressing:

  81. [99]

    Its biochemical properties and known functions 54

  82. [100]

    Its overall significance in cellular metabolism” v2 “Compose a description of approximately 300–350 words that covers:

  83. [101]

    The biological role and functional significance of the input metabolite

  84. [102]

    Its biochemical properties and known functions

  85. [103]

    Its overall significance in cellular metabolism” B.3.4.3 Output Format Instructions RAG Output Format

  86. [106]

    Integrate across contexts rather than summarizing each separately

  87. [107]

    Non-RAG Output Format

    Explain why relationships exist, not just that they exist. Non-RAG Output Format

  88. [108]

    Begin directly with content, with no preamble

  89. [109]

    Write as flowing paragraphs, with no bullet points or headers

  90. [110]

    Explain the underlying biochemical mechanisms. B.3.5 Structure-Based Eligibility and Stratification A compound is eligible for retrieval-grounded (structure-rich) description if it has retrievable structure, defined as the union of three conditions: (i) ≥ 2 pathways and ≥ 1 re...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.