Thursday, November 19, 2026
Science
No Result
View All Result
  • Login
  • HOME
  • SCIENCE NEWS
  • CONTACT US
  • HOME
  • SCIENCE NEWS
  • CONTACT US
No Result
View All Result
Scienmag
No Result
View All Result
Home Science News

Language Models Map Distinct Vaccine Adverse Event Profiles from Package Inserts

November 19, 2026
in Science News
Kristina Jarvis
By Kristina Jarvis Scienmag Editorial Profile - Infectious Disease Medicine
Reading Time: 6 mins read
0
Language Models Map Distinct Vaccine Adverse Event Profiles from Package Inserts

Language Models Map Distinct Vaccine Adverse Event Profiles from Package Inserts

65
SHARES
587
VIEWS
Share on FacebookShare on Twitter
ADVERTISEMENT

A computational evaluation of 111 vaccine package inserts approved by the United States Food and Drug Administration has demonstrated that large language model text embeddings can identify nuanced adverse event patterns and distinguish live vaccines from non-live formulations. The investigation, led by Zhigang Wang, Xingxian Li, Jie Zheng, and Yongqun He, assessed how artificial intelligence tools process complex clinical safety summaries compared to established ontology-based frameworks. Standard vaccine safety surveillance has historically depended on labor-intensive manual curation to identify adverse events and connect them to formal biomedical terminologies. By deploying open-weight language models and dense vector representations alongside established biomedical vocabularies, the research team established an automated analytical pipeline capable of processing multi-page regulatory documents and extracting structured safety profiles across diverse biological platforms.

To assemble the core dataset, the investigators downloaded all accessible FDA vaccine package insert PDF files, focusing specifically on approved commercial formulations with complete documentation. Vaccines that lacked package insert documentation or were represented solely by regulatory approval letters were excluded from the analysis. The resulting cohort comprised 111 licensed vaccines, including a distinct subgroup of 15 influenza vaccines possessing separate brand names and detailed clinical summaries. Each vaccine was cataloged and mapped into the community-developed Vaccine Ontology, which provides standardized semantic classifications of vaccine platforms, specific adjuvants, and targeted pathogens. Of the 111 vaccines examined, 69 were already cataloged within the Vaccine Ontology, while the remaining 42 vaccines were systematically integrated using automated template functions governed by predefined design patterns. This ontological curation captured key metadata such as FDA indication, valency, recombinant design, adjuvant presence, and whether the vaccine was based on live attenuated or non-live mechanisms.

The first automated stage of the workflow utilized Meta’s open-weight Llama-3 8-billion parameter instruction-tuned language model to extract clinical safety text from the downloaded documents. The research team constructed a Python automation pipeline using the PyPDFLoader document reader to target the first two pages of each package insert, where the formal adverse reactions section typically resides. A zero-shot prompt instructed the model to isolate solely the text appearing directly under the adverse reactions heading, explicitly prohibiting conversational introductions, prefaces, and extrinsic background knowledge. Across the evaluation cohort, Llama-3 achieved over 80 percent extraction accuracy on its initial execution, successfully parsing 87 of the 111 package inserts without error. When initial runs produced minor extraction flaws due to the inherent stochasticity of autoregressive models, repeat parsing coupled with manual review swiftly corrected the discrepancies, substantially reducing the manual effort typically required to compile safety texts across expansive regulatory filings.

Following text extraction, the descriptive adverse reaction passages were converted into numerical representations using two distinct open-source embedding models: nomic-embed-text, developed by Nomic AI, and mxbai-embed-large, developed by mixedbread.ai. The nomic-embed-text model encoded the extracted clinical descriptions into 768-dimensional vector embeddings, while the mxbai-embed-large architecture mapped the identical text into 1,024-dimensional dense vectors. These high-dimensional numerical coordinates encapsulated the contextual and semantic relationships among clinical terms without requiring manual mapping to predefined lexical trees. Pairwise Pearson correlation coefficients were calculated between the embedding vectors of all 111 vaccines, establishing a metric of textual similarity across their reported adverse event profiles. The research team prioritized these specific local embedding architectures to avoid transmitting sensitive biomedical and regulatory text to third-party commercial cloud application programming interfaces, ensuring reproducible and secure data handling.

To evaluate the biological relevance of the resulting vector representations, the researchers performed hierarchical clustering on the pairwise correlation matrices using Euclidean distance and complete linkage algorithms. To identify the most coherent grouping structure across the full dataset of 111 vaccines, the authors applied the silhouette method, calculating average silhouette widths across candidate solutions ranging from 2 to 15 clusters. The mathematical optimization identified a seven-cluster configuration as optimal, yielding an average silhouette width of 0.14 and a Dunn Index of 0.219. These statistical metrics reflected a moderately well-defined grouping architecture characterized by consistent intra-cluster compactness and adequate separation between disparate vaccine classes. Furthermore, an examination of approval dates across the cohort revealed no disruptive temporal clustering artifacts, demonstrating that the learned vector relationships were driven by substantive adverse event terminology rather than regulatory writing trends from specific decades.

Parallel to the vector-based modeling, the investigators implemented a classical semantic similarity workflow to serve as a benchmark for the machine learning results. Adverse event descriptions were systematically mapped to standardized terms in both the Human Phenotype Ontology and the Ontology of Adverse Events using the BioPortal Annotator alongside an in-house tool named OntoRetriever. OntoRetriever was designed to map raw textual descriptions to Open Biomedical Ontologies formats while computing semantic distance confidence metrics to help curators inspect uncertain associations. To ensure complete accuracy in this reference dataset, clinical experts reviewed all term assignments. Semantic similarity between vaccine term sets was subsequently computed using Lin’s information content metric, implemented through the ontologyX software suite. In this formulation, the information content of any ontological concept is inversely proportional to its frequency and number of child concepts, allowing the similarity between two distinct medical terms to be quantified by calculating the information content of their most informative common ancestor.

A focused comparative analysis conducted on the 15 influenza vaccines revealed substantial structural differences between the two computational methods. Hierarchical clustering based on Lin’s semantic similarity produced high global similarity scores across the influenza cohort, yielding heatmaps dominated by high correlation values. In contrast, clustering driven by nomic-embed-text representations captured a more nuanced, differential adverse event profile, separating the vaccines into three distinct clusters when evaluated at an equivalent dendrogram threshold. In both analytical frameworks, the intranasally administered live attenuated formulation FluMist consistently segregated as an independent outlier possessing a unique symptom signature distinct from injectable preparations. The researchers noted that while ontology-based metrics are constrained by static hierarchical linkages and fixed parent-child definitions, language model embeddings capture flexible contextual nuances, polysemous medical terminology, and descriptive language variations embedded directly within package insert texts.

To test whether latent biological characteristics were embedded within the adverse event text representations, the authors constructed predictive machine learning models to classify each vaccine as either live or non-live. In this classification task, non-live denoted all formulations lacking replication-competent organisms, including inactivated pathogens and messenger RNA platforms. Using standardized training and testing splits comprising 70 percent and 30 percent of the data, the authors applied LASSO logistic regression with 10-fold cross-validation to select the most informative embedding coordinates while mitigating overfitting. On the complete dataset, the optimized LASSO model utilizing nomic-embed-text features achieved an area under the receiver operating characteristic curve of 0.93 on the training split, 0.85 on the test split, and 0.91 across the combined cohort. In an overall 10-fold cross-validation assessment, the model demonstrated an accuracy of 81.89 percent, a sensitivity of 80.00 percent, and a specificity of 83.06 percent.

When identical classification procedures were implemented using the 1,024-dimensional vectors from mxbai-embed-large, predictive performance dropped, yielding area under the curve metrics of 0.83 on the training set, 0.78 on the test set, and 0.82 across the complete data. Similarly, alternative linear support vector machines trained on the selected features underperformed relative to the LASSO regression, an outcome attributed to excessive model complexity on a constrained sample size. To interrogate which embedding dimensions governed the LASSO model’s decisions, the authors performed Shapley Additive Explanations analysis using the shapr package in R. While this interpretability framework successfully ranked the mathematical contribution of specific vector components, the authors observed that individual embedding dimensions in 768-dimensional latent space cannot be mapped directly to discrete medical terms, representing an interpretability constraint inherent to deep learning text representations.

Detailed examination of the adverse event profiles that differentiated the two vaccine classes explained the underlying biological signal detected by the regression model. Formulations categorized as live vaccines frequently elicited systemic, inflammatory, and upper respiratory responses, prominently characterized by fever and generalized fatigue. Within the influenza group, the live attenuated FluMist Quadrivalent formulation was uniquely linked with reported adverse reactions of asthma and recurrent wheezing, matching patterns seen in broader pharmacovigilance databases. Conversely, non-live formulations predominantly correlated with localized mechanical and muscular reactions, including injection site pain, swelling, erythema, and localized neuromuscular discomfort. An analysis of misclassified formulations confirmed this clinical alignment: live vaccines such as the PRIORIX measles, mumps, and rubella vaccine and the IXCHIQ chikungunya vaccine were misclassified as non-live due to prominent local injection site pain, whereas non-live vaccines like SPIKEVAX and Gardasil were predicted as live due to widespread systemic complaints of fever and fatigue.

The authors outlined several methodological and structural limitations that bound the immediate generalizability of their findings. The primary constraint stems from the source material, which was restricted exclusively to 111 FDA-approved vaccine package inserts written under United States regulatory standards and formatting conventions. Safety documentation compiled under international regulatory authorities or published in other languages may employ divergent reporting structures and vocabularies that could alter embedding geometries and affect clustering stability. Furthermore, inactivated vaccines outnumber live attenuated formulations within the licensed FDA inventory, creating an inherent class imbalance that the researchers had to mitigate through class weighting during model training. The study also excluded supplementary manufacturing parameters, excipient inventories, and storage criteria from the feature set, leaving open the question of how non-clinical formulation variables might interact with adverse event vector embeddings in future pharmacovigilance research.

Subject of Research: Automated extraction and semantic profiling of vaccine adverse events using large language models and biomedical ontologies

Article Title: Unveiling differential adverse event profiles in vaccines via LLM text embeddings and ontology semantic analysis

Article References: Wang, Z., Li, X., Zheng, J., & He, Y. (2025). Unveiling differential adverse event profiles in vaccines via LLM text embeddings and ontology semantic analysis. Journal of Biomedical Semantics, 16(1), Article 10. https://doi.org/10.1186/s13326-025-00331-8

Image Credits: AI Generated

DOI: 10.1186/s13326-025-00331-8

Keywords: vaccine safety, large language models, text embeddings, adverse events, biomedical ontologies, Llama-3, nomic-embed-text, mxbai-embed-large, Human Phenotype Ontology, Ontology of Adverse Events, Vaccine Ontology, hierarchical clustering, LASSO regression, package inserts, machine learning, pharmacovigilance, semantic similarity

Cite Scienmag News

Kristina Jarvis. (November 19, 2026). Language Models Map Distinct Vaccine Adverse Event Profiles from Package Inserts. Scienmag. https://scienmag.com/language-models-map-distinct-vaccine-adverse-event-profiles-from-package-inserts/

Kristina Jarvis. "Language Models Map Distinct Vaccine Adverse Event Profiles from Package Inserts." Scienmag, 19 November 2026, https://scienmag.com/language-models-map-distinct-vaccine-adverse-event-profiles-from-package-inserts/. Accessed 19 November 2026.

Kristina Jarvis. "Language Models Map Distinct Vaccine Adverse Event Profiles from Package Inserts." Scienmag. November 19, 2026. https://scienmag.com/language-models-map-distinct-vaccine-adverse-event-profiles-from-package-inserts/

Tags: adverse eventsbiomedical ontologieshierarchical clusteringHuman Phenotype Ontologylarge language modelsLASSO regressionLLaMA-3Machine learningmxbai-embed-largenomic-embed-textOntology of Adverse Eventspackage insertspharmacovigilancesemantic similaritytext embeddingsVaccine Ontologyvaccine safety
Share26Tweet16
Previous Post

AI Integration in Text Linguistics Boosts Undergraduate Research Competence in Kazakhstan Study

Next Post

Micropulse 520-nm Laser Shows Short-Term Retinal Improvements in Acute Chorioretinopathy Series

Related Posts

Surgery Fails to Outperform Conservative Care for Pain in Fibrous Dysplasia
Medicine

Surgery Fails to Outperform Conservative Care for Pain in Fibrous Dysplasia

November 19, 2026
Food Package Label Awareness Shows No Link to Body Weight Blame
Medicine

Food Package Label Awareness Shows No Link to Body Weight Blame

November 19, 2026
Autogenous Dentin Matrix Matches Xenograft Dimensional Stability in Ridge Preservation Trial
Science News

Autogenous Dentin Matrix Matches Xenograft Dimensional Stability in Ridge Preservation Trial

November 19, 2026
Sepsis Reprograms Blood Stem Cells and Impairs Immune Metabolism in Survivors
Science News

Sepsis Reprograms Blood Stem Cells and Impairs Immune Metabolism in Survivors

November 19, 2026
Study Unveils Surgical Distribution and Sex Disparities Across Israeli Ophthalmology Hospitals
Medicine

Study Unveils Surgical Distribution and Sex Disparities Across Israeli Ophthalmology Hospitals

November 19, 2026
Nearly Half of Surveyed Dhaka Children Had Prior Professional Autism Suspicion Reported
Pediatry

Nearly Half of Surveyed Dhaka Children Had Prior Professional Autism Suspicion Reported

November 19, 2026
Next Post
Micropulse 520-nm Laser Shows Short-Term Retinal Improvements in Acute Chorioretinopathy Series

Micropulse 520-nm Laser Shows Short-Term Retinal Improvements in Acute Chorioretinopathy Series

  • Mothers who receive childcare support from maternal grandparents show more optimized

    Mothers who receive childcare support from maternal grandparents show more parental warmth, finds NTU Singapore study

    27656 shares
    Share 11059 Tweet 6912
  • University of Seville Breaks 120-Year-Old Mystery, Revises a Key Einstein Concept

    1061 shares
    Share 424 Tweet 265
  • Bee body mass, pathogens and local climate influence heat tolerance

    682 shares
    Share 273 Tweet 171
  • Researchers record first-ever images and data of a shark experiencing a boat strike

    546 shares
    Share 218 Tweet 137
  • Groundbreaking Clinical Trial Reveals Lubiprostone Enhances Kidney Function

    531 shares
    Share 212 Tweet 133
Science

Embark on a thrilling journey of discovery with Scienmag.com—your ultimate source for cutting-edge breakthroughs. Immerse yourself in a world where curiosity knows no limits and tomorrow’s possibilities become today’s reality!

RECENT NEWS

  • Surgery Fails to Outperform Conservative Care for Pain in Fibrous Dysplasia
  • Food Package Label Awareness Shows No Link to Body Weight Blame
  • Autogenous Dentin Matrix Matches Xenograft Dimensional Stability in Ridge Preservation Trial
  • Sepsis Reprograms Blood Stem Cells and Impairs Immune Metabolism in Survivors

Categories

  • Agriculture
  • Anthropology
  • Archaeology
  • Athmospheric
  • Biology
  • Biotechnology
  • Blog
  • Bussines
  • Cancer
  • Chemistry
  • Climate
  • Earth Science
  • Editorial Policy
  • Marine
  • Mathematics
  • Medicine
  • Pediatry
  • Policy
  • Psychology & Psychiatry
  • Science Education
  • Science News
  • Social Science
  • Space
  • Technology and Engineering

Subscribe to Blog via Email

Enter your email address to subscribe to this blog and receive notifications of new posts by email.

Join 5,151 other subscribers

© 2026 Scienmag - Science Magazine

Welcome Back!

Login to your account below

Forgotten Password?

Retrieve your password

Please enter your username or email address to reset your password.

Log In
No Result
View All Result
  • HOME
  • SCIENCE NEWS
  • CONTACT US

© 2026 Scienmag - Science Magazine

Discover more from Science

Subscribe now to keep reading and get access to the full archive.

Continue reading