Interpretable Metadata-Based Microbial Risk Prediction Using Large Language Models

Hyunwoo Ted Yoo; Gail L Rosen

doi:10.1145/3765612.3767197

Back

Interpretable Metadata-Based Microbial Risk Prediction Using Large Language Models

Conference proceeding

Open access

Interpretable Metadata-Based Microbial Risk Prediction Using Large Language Models

Hyunwoo Ted Yoo and Gail L Rosen

BCB '25: Proceedings of the 16th ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics, pp 1-10

10 Dec 2025

DOI: https://doi.org/10.1145/3765612.3767197

Featured in Collection : Research Supported by Drexel Libraries' OA Programs

Files and links (1)

url

https://doi.org/10.1145/3765612.3767197View

Published, Version of Record (VoR)Open Access via Drexel Libraries Read and Publish Program 2025CC BY-NC V4.0, Open

Abstract

Large language models

microbial risk prediction

metadata analysis

ontology classification

environmental biosurveillance

Environmental metadata provides a widely accessible signal for microbiome analysis, especially in settings where genomic sequencing is unavailable. However, traditional machine learning models often struggle to make reliable predictions using metadata alone. The potential of large language models (LLMs) to effectively utilize such metadata remains underexplored. In this work, we investigate whether LLMs can perform accurate microbial ontology classification and pathogen risk prediction using metadata alone. We evaluate models such as ChatGPT-4o, Claude 3.7, Grok-3, and LLaMA 4 on tasks including ontology classification based on the Earth Microbiome Project Ontology (EMPO 3) and binary prediction of E. coli contamination risk, across both zero-shot and few-shot settings. To further understand how these models make predictions, we conduct a comprehensive ablation study to identify which environmental features such as rainfall, turbidity, or temperature most influence performance. Our results show that LLMs not only achieve competitive predictive performance but also exhibit distinct patterns of reliance on specific metadata features. These patterns enable insights into biological decision factors without requiring access to genomic sequences. Our findings highlight the potential of LLMs as robust and interpretable tools for metadata-based microbiome analysis and environmental biosurveillance.

Metrics

8 Record Views

Details

Title: Interpretable Metadata-Based Microbial Risk Prediction Using Large Language Models
Creators: Hyunwoo Ted Yoo - Drexel University, Electrical and Computer Engineering
Gail L Rosen (Corresponding Author) - Drexel University, Electrical and Computer Engineering
Publication Details: BCB '25: Proceedings of the 16th ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics, pp 1-10
Conference: 16th ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics (Philadelphia, Pennsylvania, United States, 12 Oct 2025–15 Oct 2025)
Publisher: Association for Computing Machinery
Number of pages: 10
Grant note: National Science Foundation: 2107108 Center for Produce Safety
This work was supported in part by a National Science Foundation grant #2107108 and by the Center for Produce Safety under a grant entitled "Development of a risk ranking tool for evaluating hazards and risks related to agricultural water subpart E".
Resource Type: Conference proceeding
Language: English
Academic Unit: Electrical and Computer Engineering
Web of Science ID: WOS:001658892200003
Scopus ID: 2-s2.0-105025531912
Other Identifier: 991022136866304721