henhaobio
Sign in

Analytical ChemistryScientific Communication & EducationPeer-reviewed journal article

A Large Language Model-Powered Map of Metabolomics Research

Olatomiwa O. BifariniD, Varun S. Yelluru, Aditya Simhadri, Facundo M. FernándeziD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

We present a comprehensive map of the metabolomics research landscape, synthesizing insights from over 80,000 publications. Using PubMedBERT, we transformed abstracts into 768-dimensional embeddings that capture the nuanced thematic structure of the field. Dimensionality reduction with t-SNE revealed distinct clusters corresponding to key domains such as analytical chemistry, plant biology, pharmacology, and clinical diagnostics. In addition, a neural topic modeling pipeline refined with GPT-4o mini reclassified the corpus into 20 distinct topics--ranging from "Plant Stress Response Mechanisms" and "NMR Spectroscopy Innovations" to "COVID-19 Metabolomic and Immune Responses." Temporal analyses further highlight trends including the rise of deep learning methods post-2015 and a continued focus on biomarker discovery. Integration of metadata such as publication statistics and sample sizes provide additional context to these evolving research dynamics. An interactive web application (https://metascape.streamlit.app/) enables dynamic exploration of these insights. Overall, this study offers a robust framework for literature synthesis that empowers researchers, clinicians, and policymakers to identify emerging research trajectories and address critical challenges in metabolomics, while also sharing our perspectives on key trends shaping the field.

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 1 (19 Mar 2025) · CC BY 4.0

PubMedBERT embeddings of 80,656 metabolomics abstracts, projected with t-SNE and grouped into 20 GPT-4o mini-labelled topics, map the field's domains and trends (1998 to early 2024), browsable online.

The metabolomics literature has grown fast and spread across many subjects, and the authors argue that conventional reviews and keyword database searches show only narrow slices of it. They adapt the embedding-map approach that González-Márquez and colleagues applied to biomedical literature, and add BERTopic clustering with GPT-4o mini labels to classify the papers that journal-title categories leave unassigned.

Key results

  1. ShownBERTopic clustering of the PubMedBERT embeddings, with GPT-4o mini writing the cluster labels, assigns many more papers to a theme than journal-title categories do. It yields 20 topics, and plant stress responses is the largest.

    65,676 publications assigned to a topic (the rest are outliers) vs ~40,000 labelled by the 18 journal-keyword categories, which left 41,721 unclassified. Largest topic: 'Plant Stress Response Mechanisms', 19,258 papers. Largest journal-derived categories: analytical chemistry 8,981, plant biology 4,648, pharmacology 4,229. · Global View of Metabolomics Corpus; Publication Classification; Topic Modelling with a Large Language Model; Tables

  2. Shownt-SNE preserves the structure of the 768-dimensional embeddings better than UMAP when reducing them to two dimensions, so the authors use t-SNE for the main map.

    k-NN accuracy 0.56 for t-SNE vs 0.54 for UMAP. k-NN recall (overlap of the 10 nearest neighbours in the original and 2D spaces) 0.33 vs 0.14. The classifier used 10 neighbours and was tested on a 1% hold-out; both projections used a single seed (random_state = 42). · Global View of Metabolomics Corpus; Embeddings Dimensionality Reduction; Tables

  3. ShownThe number of papers in the corpus rose steadily from 1998, with a sustained high-growth phase from 2011 and the largest single-year jump from 2020 to 2021, which the authors tentatively link to COVID-19.

    About 12,000 corpus papers in 2023. Mean year-over-year increase: 447.58 publications. 2011 to 2021 classed as consistently high growth and 1998 to 2002 as low growth, with thresholds set at the mean rate of change plus or minus one standard deviation. Early 2024 papers excluded. · Publication Analytics; Analysis of Publication Statistics; Synthesis of Findings and Perspectives; Tables

  4. SuggestedSample sizes extracted from the abstracts are mostly small and cluster in clinical and animal studies. The authors connect this to the shortage of clinically validated metabolite biomarkers.

    A regex for 'n = <number>' finds most reported sizes below 50 (n = 1849 in Figure 5e); only 51 exceed 1000, and these tend to be systematic meta-reviews. In support, the authors cite a pancreatic cancer meta-analysis in which 87% of 655 candidate biomarkers came from single studies. · Research Trend Discovery; Synthesis of Findings and Perspectives; Figure legends

  5. SuggestedKeyword overlays show COVID-19 metabolomics as one dense cluster from 2020 onward. Deep learning papers are sparse, sit in data-science regions and mostly date from after 2015, while chemometrics appears across several decades.

    Keyword and regex filters ('COVID-19|SARS-CoV-2', 'deep learning OR neural network', 'chemometrics OR multivariate analysis', 'biomarker discovery') drawn on t-SNE maps coloured by publication year. The patterns are described visually; no counts or statistical tests are given. · Research Trend Discovery; Figure legends

Methods, data and reproducibility

PubMed records matching 'metabolomics' or 'metabonomics' (1998 to early 2024) were embedded with PubMedBERT (mean of the last hidden states, 768 dimensions) and reduced to 2D with t-SNE and UMAP. Domains came from journal-title keywords. Topics came from BERTopic with HDBSCAN clusters, whose c-TF-IDF keywords GPT-4o mini turned into labels. Regex queries pulled out p-values and sample sizes.

  • PubMed abstracts and metadata matching 'metabolomics' or 'metabonomics', 1998 to early 2024, retrieved with Entrez Direct; deposited on Zenodo with PubMedBERT, t-SNE and UMAP embeddings · n = 80,656 publications · 10.5281/zenodo.15020144
Methods & tools
PubMedBERT (microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract-fulltext), Hugging Face Transformers 4.38.1, PyTorch 2.1.0, scikit-learn 1.2.2 (t-SNE, KNeighborsClassifier, CountVectorizer), umap-learn 0.5.5, BERTopic 0.16.3, HDBSCAN 0.8.29, GPT-4o mini (OpenAI API client 1.51.1), Entrez Direct (NCBI E-utilities), WordCloud 1.9.3
Software & environment
Python with Transformers 4.38.1, PyTorch 2.1.0+cu121, scikit-learn 1.2.2, umap-learn 0.5.5, BERTopic 0.16.3, HDBSCAN 0.8.29, OpenAI API client 1.51.1 and WordCloud 1.9.3; t-SNE and UMAP run with random_state = 42.
Assay conditions & compute
Embeddings computed on an NVIDIA Tesla T4 (CUDA 12.2) in Google Colab; topic modelling run in a high-memory GPU environment that is not specified further; run times not reported.
Reproducibility gaps
Topic labels come from calls to OpenAI's GPT-4o mini API, so a re-run may not reproduce them exactly. The labelling prompt (Appendix B) and the journal keyword mappings (Appendix A) are referenced but not included in this text.
Limitations (4)
  • Assigning research domains from journal-title keywords leaves papers in multidisciplinary journals unlabelled (41,721 publications); the authors added LLM-labelled topics to address this. · stated by the authors
  • The analysis uses only abstracts and basic PubMed metadata; the authors note that full text, author affiliations or funding sources could sharpen the classification. · stated by the authors
  • The 20 GPT-4o mini topic labels and the HDBSCAN clusters (minimum cluster size 500) are not checked against human annotation, topic-coherence metrics or other clustering settings. · noted in this summary
  • The trend claims (deep learning after 2015, more recent mechanistic studies, the COVID-19 surge) rest on visual reading of keyword-filtered maps, without counts or statistical tests. · noted in this summary

Read this if you want a ready-made embedding map and topic labels for the metabolomics literature, or a template for building a PubMedBERT, t-SNE and BERTopic/LLM map of another field.

Citations
6
FWCI
1.80
Field percentile
top 15%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISPhongwattan '23Kidder '26Dang '26Lu '26Mannochio-R '26Banerjee '26
65% semantic matchBioinformatics
Thiptanawat Phongwattana, Jonathan H. Chan·bioRxiv·Jun 27, 2023· 3 citations