{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,7,27]],"date-time":"2026-07-27T06:01:41Z","timestamp":1785132101726,"version":"3.55.0"},"reference-count":39,"publisher":"Frontiers Media SA","license":[{"start":{"date-parts":[[2026,7,27]],"date-time":"2026-07-27T00:00:00Z","timestamp":1785110400000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":["frontiersin.org"],"crossmark-restriction":true},"short-container-title":["Front. Bioinform."],"abstract":"<jats:p>Large language model (LLM) text embeddings have recently been used for supervised learning on tabular data by serializing each observation into text and then converting the text into a dense fixed-length vector. This strategy is appealing for biomedical tabular data, which are often mixed-type and contain missing values, because it produces a complete numeric representation even when some original entries are missing. However, its suitability for unsupervised tasks remains unclear. Here we evaluate the use of LLM-derived text embeddings for dimension reduction of tabular data, focusing on biological and clinical datasets. We compare an LLM embedding-based approach with a direct tabular approach that computes dissimilarities directly from the original variables. Because unsupervised dimension reduction has no ground-truth low-dimensional target, we assess performance through stability under perturbation. Across multiple datasets and analysis settings, the LLM embedding-based approach is consistently less stable than the direct tabular approach. In particular, small amounts of additional missingness and random permutation of feature order can substantially alter the resulting low-dimensional representation. These results suggest that the straightforward use of LLM text embeddings is not reliable for unsupervised dimension reduction of tabular data.<\/jats:p>","DOI":"10.3389\/fbinf.2026.1851023","type":"journal-article","created":{"date-parts":[[2026,7,27]],"date-time":"2026-07-27T05:40:23Z","timestamp":1785130823000},"update-policy":"https:\/\/doi.org\/10.3389\/crossmark-policy","source":"Crossref","is-referenced-by-count":0,"title":["Instability of LLM text embeddings for unsupervised dimension reduction of tabular data"],"prefix":"10.3389","volume":"6","author":[{"given":"Jun","family":"Li","sequence":"first","affiliation":[{"name":"Department of Applied and Computational Mathematics and Statistics, University of Notre Dame","place":["Notre Dame, IN, United States"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Yixuan","family":"Gou","sequence":"additional","affiliation":[{"name":"The Awty International School","place":["Houston, TX, United States"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Shawn","family":"Su","sequence":"additional","affiliation":[{"name":"The John Cooper School","place":["The Woodlands, TX, United States"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Mujun","family":"Xu","sequence":"additional","affiliation":[{"name":"Round Rock High School","place":["Round Rock, TX, United States"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Brendan","family":"Chen","sequence":"additional","affiliation":[{"name":"Shanghai American School Puxi","place":["Shanghai, China"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"1965","published-online":{"date-parts":[[2026,7,27]]},"reference":[{"key":"B1","doi-asserted-by":"publisher","first-page":"2095","DOI":"10.1007\/s00405-023-08299-w","article-title":"Machine learning for risk stratification of thyroid cancer patients: a 15-year cohort study","volume":"281","author":"Borzooei","year":"2024","journal-title":"Eur. Archives Oto-Rhino-Laryngology"},{"key":"B2","volume-title":"National Health and Nutrition Examination survey","year":"2020"},{"key":"B3","first-page":"483","article-title":"Simple and effective embedding model for single-cell biology built from ChatGPT","volume-title":"Nat. Biomed. Eng.","author":"Chen","year":"2025"},{"key":"B4","doi-asserted-by":"publisher","first-page":"1","DOI":"10.1038\/s41592-025-02707-1","article-title":"A visual\u2013omics foundation model to bridge histopathology with spatial transcriptomics","volume":"22","author":"Chen","year":"","journal-title":"Nat. Methods"},{"key":"B5","doi-asserted-by":"publisher","first-page":"24","DOI":"10.1038\/s44386-025-00027-0","article-title":"Genetic evidence informs the direction of therapeutic modulation in drug development","volume":"2","author":"Chen","year":"","journal-title":"Npj Drug Discov."},{"key":"B6","doi-asserted-by":"crossref","first-page":"315","DOI":"10.1007\/978-3-540-33037-0_14","article-title":"Multidimensional scaling","volume-title":"Handbook of Data Visualization","author":"Cox","year":"2008"},{"key":"B7","doi-asserted-by":"publisher","first-page":"1470","DOI":"10.1038\/s41592-024-02201-0","article-title":"scgpt: toward building a foundation model for single-cell multi-omics using generative ai","volume":"21","author":"Cui","year":"2024","journal-title":"Nat. Methods"},{"key":"B8","doi-asserted-by":"crossref","first-page":"11763","DOI":"10.52202\/068431-0855","article-title":"Lift: language-Interfaced fine-tuning for non-language machine learning tasks","volume":"35","author":"Dinh","year":"2022","journal-title":"Adv. Neural Inf. Process. Syst."},{"key":"B9","doi-asserted-by":"publisher","first-page":"617","DOI":"10.1109\/tsmc.1979.4310090","article-title":"Pattern recognition with partly missing data","volume":"9","author":"Dixon","year":"2007","journal-title":"IEEE Trans. Syst. Man, Cybern."},{"key":"B10","doi-asserted-by":"publisher","first-page":"e215","DOI":"10.1161\/01.cir.101.23.e215","article-title":"Physiobank, physiotoolkit, and physionet: components of a new research resource for complex physiologic signals","volume":"101","author":"Goldberger","year":"2000","journal-title":"circulation"},{"key":"B11","doi-asserted-by":"publisher","first-page":"857","DOI":"10.2307\/2528823","article-title":"A general coefficient of similarity and some of its properties","volume":"27","author":"Gower","year":"1971","journal-title":"Biometrics"},{"key":"B12","doi-asserted-by":"publisher","first-page":"33","DOI":"10.1007\/bf02291478","article-title":"Generalized procrustes analysis","volume":"40","author":"Gower","year":"1975","journal-title":"Psychometrika"},{"key":"B13","first-page":"5549","article-title":"Tabllm: few-shot classification of tabular data with large language models","volume-title":"Proceedings of the 26th International Conference on Artificial Intelligence and Statistics","author":"Hegselmann","year":"2023"},{"key":"B14","volume-title":"UCI Machine Learning Repository","year":"1983"},{"key":"B15","doi-asserted-by":"publisher","first-page":"58","DOI":"10.21037\/jlpm.2018.06.01","article-title":"Using machine learning techniques to generate laboratory diagnostic pathways\u2014a case study","volume":"3","author":"Hoffmann","year":"2018","journal-title":"J. Laboratory Precis. Med."},{"key":"B16","article-title":"TabPFN: a transformer that solves small tabular classification problems in a second","volume-title":"International Conference on Learning Representations (ICLR)","author":"Hollmann","year":"2023"},{"key":"B17","doi-asserted-by":"publisher","first-page":"1462","DOI":"10.1038\/s41592-024-02235-4","article-title":"Assessing gpt-4 for cell type annotation in single-cell rna-seq analysis","volume":"21","author":"Hou","year":"2024","journal-title":"Nat. methods"},{"key":"B18","first-page":"6488","article-title":"Representation learning for tabular data: a comprehensive survey","volume-title":"IEEE Transactions on Pattern Analysis and Machine Intelligence","author":"Jiang","year":"2026"},{"key":"B19","volume-title":"Eicu Collaborative Research Database Demo","author":"Johnson","year":"2021"},{"key":"B20","volume-title":"MIMIC-IV Clinical Database Demo","author":"Johnson","year":"2023"},{"key":"B21","volume-title":"Finding Groups in Data: An Introduction to Cluster Analysis","author":"Kaufman","year":"2009"},{"key":"B22","doi-asserted-by":"publisher","DOI":"10.1038\/s41551-025-01602-6","article-title":"Leveraging multi-modal foundation models for analysing spatial multi-omic and histopathology data","author":"Liu","year":"2026","journal-title":"Nat. Biomed. Eng"},{"key":"B23","doi-asserted-by":"publisher","first-page":"9455","DOI":"10.1038\/s41467-025-64506-8","article-title":"UNICORN: Towards universal cellular expression prediction with a multi-task learning framework","volume":"16","author":"Liu","year":"2025","journal-title":"Nat. Commun."},{"key":"B24","article-title":"UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction","volume-title":"arXiv","author":"McInnes","year":"2018"},{"key":"B25","first-page":"861","article-title":"Nuclear feature extraction for breast tumor diagnosis","author":"Nick Street","year":"1993","journal-title":"Biomed. image"},{"key":"B26","volume-title":"Embeddings Faq","year":"2024"},{"key":"B28","doi-asserted-by":"publisher","first-page":"465","DOI":"10.1038\/s41746-025-01869-8","article-title":"Enhancing ehr-based pancreatic cancer prediction with llm-derived embeddings","volume":"8","author":"Park","year":"2025","journal-title":"Npj Digit. Med."},{"key":"B29","volume-title":"NHANES: Data from the US National Health and Nutrition Examination Study","author":"Randall","year":"2014"},{"key":"B30","doi-asserted-by":"publisher","first-page":"1","DOI":"10.1613\/jair.346","article-title":"Improved heterogeneous distance functions","volume":"6","author":"Randall Wilson","year":"1997","journal-title":"J. artificial intelligence research"},{"key":"B31","first-page":"3982","article-title":"Sentence-bert: sentence embeddings using siamese bert-networks","volume-title":"Proceedings of the 2019 Conference on Empirical Methods in Natural language Processing and the 9th International Joint Conference on Natural language Processing","author":"Reimers","year":"2019"},{"key":"B32","volume-title":"scipy.spatial.procrustes","year":"2026"},{"key":"B33","first-page":"16857","article-title":"MPNet: Masked and Permuted Pre-training for Language Understanding","volume":"33","author":"Song","year":"2020","journal-title":"Advances in Neural Information Processing Systems"},{"key":"B34","article-title":"LLMs are also effective embedding models: an in-depth overview","volume-title":"arXiv","author":"Tao","year":"2025"},{"key":"B35","doi-asserted-by":"publisher","first-page":"1930","DOI":"10.1038\/s41591-023-02448-8","article-title":"Large language models in medicine","volume":"29","author":"Thirunavukarasu","year":"2023","journal-title":"Nat. Med."},{"key":"B36","doi-asserted-by":"publisher","first-page":"401","DOI":"10.1007\/bf02288916","article-title":"Multidimensional scaling: I. theory and method","volume":"17","author":"Torgerson","year":"1952","journal-title":"Psychometrika"},{"key":"B37","doi-asserted-by":"publisher","first-page":"884","DOI":"10.1109\/TBME.2009.2036000","article-title":"Accurate telemonitoring of parkinson\u2019s disease progression by noninvasive speech tests","volume":"57","author":"Tsanas","year":"2009","journal-title":"IEEE Trans. Biomed. Eng."},{"key":"B38","first-page":"5998","article-title":"Attention is all you need","volume":"30","author":"Vaswani","year":"2017","journal-title":"Adv. Neural information processing systems"},{"key":"B39","volume-title":"GENPHIRE: Enhancing Disease Risk Prediction Using Large language Model","author":"Yao","year":"2025"},{"key":"B27","doi-asserted-by":"publisher","first-page":"21","DOI":"10.1038\/s43588-025-00912-8","article-title":"Scouter predicts transcriptional responses to genetic perturbations with large language model embeddings","volume":"6","author":"Zhu","year":"2026","journal-title":"Nat. Comput. Sci."}],"container-title":["Frontiers in Bioinformatics"],"original-title":[],"link":[{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fbinf.2026.1851023\/full","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,7,27]],"date-time":"2026-07-27T05:40:27Z","timestamp":1785130827000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fbinf.2026.1851023\/full"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,7,27]]},"references-count":39,"alternative-id":["10.3389\/fbinf.2026.1851023"],"URL":"https:\/\/doi.org\/10.3389\/fbinf.2026.1851023","relation":{},"ISSN":["2673-7647"],"issn-type":[{"value":"2673-7647","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,7,27]]},"article-number":"1851023"}}