Knowledge Organization in AI-Era Biology

Biological knowledge organization systems (KOSs), such as taxonomies, controlled vocabularies, ontologies, and knowledge graphs, were traditionally developed to classify and represent existing knowledge. Today, these systems are increasingly integrated with ML, NLP, and LLMs. They are no longer only resources that researchers search or retrieve information from. In many cases, they have become part of the computational process through which scientific knowledge is analyzed and discovered.

In 2025, while preparing for my candidacy, I conducted a thematic review to understand the transition into AI-assisted science and how the roles of biological KOSs have changed alongside scientific and technological development. Existing classifications of KOSs could not fully describe what is happening in AI-augmented biological research. A simple ontology may be structurally sophisticated but still rely heavily on manual expert curation. Another system may automatically integrate large amounts of biological data but have relatively limited reasoning ability. More recently, some systems combine ontologies, knowledge graphs, and AI models to generate predictions or even scientific hypotheses.

To capture these differences, I developed the Knowledge Organization Analysis Framework (KOAF), placing KOSs on a single scale from simple to complex across three dimensions:

(1) Functional sophistication: from basic terminology disambiguation and data standardization to AI-driven knowledge discovery.

(2) Automation in system construction: from completely manual expert curation to highly automated knowledge extraction and integration.

(3) Reasoning and inference capability: from simple information retrieval to logical reasoning, cross-domain inference, and eventually AI-supported hypothesis generation.

I then applied this framework to representative studies published between 2003 and 2025.

The patterns showed earlier biological KOSs primarily helped researchers establish common terminology and organize heterogeneous data (what traditional systematics and phylogenetics do). Formal ontologies such as Gene Ontology later provided richer semantic structures that machines could process and reason over. More recent systems however, increasingly combine ontologies and knowledge graphs with AI models to integrate knowledge across domains, identify previously hidden relationships and support scientific discovery.

To me, the most interesting part of this transition is that KOSs are beginning to move beyond their traditional role as passive knowledge repositories. When an ontology or knowledge graph becomes part of an AI system that generates a prediction, suggests a biological relationship, or contributes to a new hypothesis, the organization of knowledge itself begins to influence what can be discovered.

In the paper, we describe these advanced KOSs as increasingly functioning as epistemic agents in scientific research.

This also raises new questions. If AI models rely on organized domain knowledge to perform scientific reasoning, how should this knowledge be structured and validated? How much of scientific discovery comes from the AI model, and how much comes from the underlying knowledge organization system? And when these systems contribute to new scientific claims, how can we ensure their transparency, accountability, and trustworthiness?

The full paper, “Knowledge Organization in AI-Era Biological Sciences: A Thematic Review,” is published in Knowledge Organization. https://doi.org/10.31083/KO46546

An interactive visualization of the representative studies mapped to the KOAF framework is available at: https://qiaoyiliu.github.io/KOAF/targetpaper3D.html

The full list of literature mapped to KOAF:

Title	Author(s)	Year
eVOC: A Controlled Vocabulary for Unifying Gene Expression Data	Kelso et al.	2003
The Unified Medical Language System (UMLS): integrating biomedical terminology	Bodenreider	2004
The OBO Foundry: coordinated evolution of ontologies to support biomedical data integration	Smith et al.	2007
Integrating phenotype ontologies across multiple species	Mungall et al.	2010
Darwin Core: An Evolving Community-Developed Biodiversity Data Standard	Wieczorek et al.	2012
The Semanticscience Integrated Ontology (SIO) for biomedical research and knowledge discovery	Dumontier et al.	2014
Semantics in Support of Biodiversity Knowledge Discovery: An Introduction to the Biological Collections Ontology and Related Ontologies	Walls et al.	2014
Knowledge Discovery from Biomedical Ontologies in Cross Domains	Shen & Lee	2016
Neuro-symbolic representation learning on biological knowledge graphs	Alshahrani et al.	2017
The DisGeNET knowledge platform for disease genomics: 2019 update	Piñero et al.	2020
Combining lexical and context features for automatic ontology extension	Althubaiti et al.	2020
BiOnt: Deep Learning using Multiple Biomedical Ontologies for Relation Extraction	D. Sousa & Couto	2020
Evolving knowledge graph similarity for supervised learning in complex biomedical domains	R. T. Sousa et al.	2020
Towards Semantic Integration for Explainable Artificial Intelligence in the Biomedical Domain:	Pesquita	2021
Extracting a Knowledge Base of Mechanisms from COVID-19 Papers	Hope et al.	2021
The Human Phenotype Ontology in 2021	Köhler et al.	2021
Biomedical knowledge graph embeddings for personalized medicine: Predicting disease-gene associations	Vilela et al.	2023
The Gene Ontology knowledgebase in 2023	The Gene Ontology Consortium et al.	2023
Building a knowledge graph to enable precision medicine	Chandak et al.	2023
Gene-associated Disease Discovery Powered by Large Language Models	Chang et al.	2024
A framework for integrating biomedical knowledge in Wikidata with open biological and biomedical ontologies and MeSH keywords	Turki et al.	2024
An LLM-based Knowledge Synthesis and Scientific Reasoning Framework for Biomedical Discovery	Wysocki et al.	2024
UniEntrezDB: Large-scale Gene Ontology Annotation Dataset and Evaluation Benchmarks with Unified Entrez Gene Identifiers	Miao et al.	2024
Unifying Large Language Models and Knowledge Graphs: A Roadmap	Pan et al.	2024
TaxonGPT: Taxonomic Classification Using Generative Artificial Intelligence	H. Huang et al.	2024
Building a literature knowledge base towards transparent biomedical AI	Y. Huang et al.	2025
Network for knowledge Organization (NEKO): An AI knowledge mining workflow for synthetic biology research	Xiao et al.	2025

Data Literacy and Academic Library Pedagogy

Back in my first year of PhD, I became curious about how academic libraries support their patrons' evolving information needs. While information literacy has been studied extensively, data literacy remains an area where many people continue to struggle, especially as generative AI and data-intensive tools become more integrated into everyday academic work. To me, data competency felt like an important and timely area to explore.

I began applying for IRB approval for a study with Heather Charlotte Owen, the Data Librarian at the University of Rochester (UR) Libraries and a close friend of mine. We both completed our MLIS degrees at Syracuse University and share a strong passion for coding and data processing. Heather contributed significantly in organizing UR Libraries’ data events, trainings, and workshops that serve students, faculty, staff, and the broader Rochester community. Together, we designed a follow-up survey for participants who attended these workshops, aiming to measure their progress in data access, retrieval, interpretation, and reuse.

Over the course of a year, we were fortunate to receive more than 100 valid responses from students, faculty and staff, and members of the Rochester community. Their honest and anonymous reflections allowed us to examine how data literacy training affects people’s confidence and emotional experience with data. Our analysis showed a clear pattern: participants who attended these workshops not only improved their data literacy, but also experienced lower cognitive overload, the mental strain of processing too much information at once. In turn, this reduction in cognitive overload led to a meaningful decrease in information anxiety, the stress people feel when they cannot access, understand, or manage the information they need. The signs were clear: learning how to navigate data made library patrons feel more capable and less overwhelmed.

We also found that different groups benefited in different ways. Students and participants in health or social professions tended to show the strongest gains in confidence after training, while STEM professionals and some faculty/staff sometimes became more aware of their own limitations—an important reminder that data literacy support must be tailored to diverse needs. Still, across all groups, the evidence was clear: academic libraries play a vital role in helping people build data skills and reduce the anxieties that often accompany data-intensive environments.

As academic libraries continue to expand their data services, there is a growing need to explore how creative and user-centered pedagogy can make data literacy training even more effective. Not all learners engage with data in the same way, and not all disciplines face the same challenges. Understanding these needs will be essential for helping patrons navigate an increasingly data-driven and AI-augmented world with confidence rather than anxiety.

Some of our findings were accepted to the Association for Information Science and Technology (ASIS&T) 2025 annual conference in Washington D.C. Here is a glimpse to our poster and submission:

https://asistdl.onlinelibrary.wiley.com/doi/abs/10.1002/pra2.1467

Using Gene Ontology and ML Models for Gene-Disease Association Identification

Genomic data in large genomic knowledgebases (KB) such as Gene Ontology (GO) are being integrated into clinical diagnosis and disease prediction. This type of integration is particularly useful in predicting complex diseases with highly heterogeneous genotypes that make biological marker identifications difficult. Several types of machine learning (ML) models have been applied to identify relatively small number of disease-associated genetic sequence amongst the large number of common variants carried by an individual [1]. Biomedical knowledge organization systems (KOSs) containing relationships between variants, genes, and diseases promise higher precision and performance of ML models.

I use GO as a case study of biomedical KOS that provides rich annotation data in directed acyclic graph (DAG) structure as train dataset for ML algorithms to identify potential disease-triggering gene products. Different from approaches in bioinformatics, my research focus is not on designing models and packages. Rather, I discuss the data quality, ontology structure, crosslink with external resources e.g. Disease Ontology, to evaluate the current design of KOS for biological research, which applies theories in knowledge organization and library science. Past findings on this area were contributed by either bioinformatics or computer science scholars. My role is to reveal the importance of knowledge work in bridging these two communities, and discuss the usage of ontology data in LLM to achieve trustworthiness and precision.

Currently, I test collecting GO annotation data to identify potential gene products that may be associated with Autism disease using ML algorithms - Random Forest, Support-Vector Machine, and Gradient Boosting. A demo of this preliminary step will be presented at the DCMI 2024 NKOS workshop (https://www.dublincore.org/conferences/2024/sessions/nkos-workshop/).