We manually identified and selected relevant literature through iterative keyword searches in scholarly databases, including Google Scholar, Semantic Scholar, and PubMed. Keywords and search terms were adapted and refined throughout the process to reflect emerging patterns and conceptual intersections across biology, KO, and AI. To broaden the research, alternative terms such as knowledge graph and knowledge representation were also used in place of KO. In addition to database searches, snowball sampling techniques (e.g., forward and backward citation tracking) were employed to capture high-impact studies. ArXiv and BioRXiv were further consulted to include preprints in AI and biological research.

The literature reviewed in this study is organized into two tiers. The first tier includes papers that describe and discuss conceptual frameworks for constructing and applying KO with AI technologies. We selected literature based on its theoretical significance and empirical contributions to understanding the evolving relationship between KO and AI-assisted biological research. This foundational set of work sets up the theoretical groundwork for examining science paradigms, classification theories, and the epistemological foundations of biological systematics, drawing from LIS, philosophy of science, and biological classification theory. Priority was given to influential publications that have shaped past and contemporary understanding of KOSs, the formation and transformation of scientific paradigms, and the evolution of biological classifications. Studies at the intersection of KR and AI research were also included to bridge traditional classification approaches with emerging computational methodologies.

The second tier focuses on empirical studies that demonstrate the active implementation of KOSs in contemporary biological research, particularly those incorporating AI technologies. Literature was excluded if it focuses solely on technical implementation details without addressing broader implications for KO theory or biological research paradigms. These studies include, for instance, efforts to construct KGs as advanced mapping approaches for organizing biological knowledge. Studies addressing KO in non-biological domains were excluded unless they offered theoretical insights directly applicable to biological classification systems. Papers presenting only preliminary results, purely speculative frameworks lacking empirical validation, or insufficient methodological detail for dimensional assessment were also excluded from the target paper pool, although some were kept as background literature when they contributed valuable theoretical insights.

Literature in the second tier, referred to as target papers, was evaluated using the proposed Knowledge Organization Analysis Framework (KOAF). This framework is built on three dimensions: functional sophistication (ranging from simple tasks such as terminology disambiguation to more advanced operations such as AI-driven knowledge discovery), automation degree in system construction (from manual curation only to fully automated systems), and reasoning and inference capability (from basic lookup functions to hypothesis generation). These three dimensions were designed to better structure the review of selected papers as well as serve as potential evaluation criteria for AI-powered bio-knowledge systems. The framework also helped us decide the target paper selection criteria: papers for the second tier should present concrete applications of KOSs in biological research contexts, demonstrate measurable impacts on research workflows or discovery processes, and provide sufficient technical detail to enable dimensional analysis within the KO analysis framework.

The final target paper pool includes 27 scholarly publications, spanning from 2003 to 2025 (see Appendix Table 4 for the full list). The analysis result on the target papers is visualized using the plotly R package (Sievert, 2020). We acknowledge that these 27 papers in the list may not be sufficiently inclusive of all existing literature that would qualify under the inclusion criteria. Given the rapidly expanding intersections of KO, AI, and biological research, a comprehensive coverage of all relevant studies is neither feasible nor necessary for the thematic analysis approach employed in this review. Instead, the papers included in this review were chosen to give representative examples across the three dimensions of the KO analysis framework, ensuring adequate coverage of different functional sophistication levels, degrees of automation in system construction, and reasoning capabilities. The temporal span reflects both the historical development of computational approaches to biological knowledge organization and the recent acceleration while keeping a focus on contemporary applications and emerging paradigms. The iterative selection process allowed refinement of inclusion parameters as emerging themes and conceptual intersections became clearer, ensuring that the final literature corpus represented both the theoretical foundations and empirical manifestations of knowledge organization systems in AI-augmented biological research.

In LIS, KO refers to activities such as describing, representing, and organizing documents, works, and concepts. In broader sense, KO can be understood as an institutional practice involved in the production and dissemination of knowledge. The narrower definition applies primarily within the LIS domain and encompasses document description, indexing and classification activities carried out in libraries, bibliographical databases, archives, and other forms of “memory institutions” (Hjørland, 2008).

Conventionally, librarians, archivists, and information specialists performed these practices. Today, technological advancements enable the embedding of well-structured KOSs into various tools and systems to perform the tasks that once required extensive human intervention. These developments have, in turn, driven advances in KOS design. With varying degrees of complexity and purpose, embedded KOSs help reduce ambiguities, control synonyms, and establish explicit semantic relationships between concepts based the types of information being handled (Zeng, 2008).

KO practices are not static. Their evolution reflects perceptual changes of similarity and difference among the objects we see. These categorization and classification abilities, fueled by cognition, in turn reinforce how we see things and the relations between them (Atran and Medin, 2008). In the field of biology, the science of classification formed its separate paradigm—systematics. Once known as “the Queen of Sciences”, systematics contributes to taxonomic theories that manage and interpret vast amounts of potentially conflicting information about the origins of life and organismal features, guided by diverse philosophical ontologies such as foundationalism and constructive empiricism (Brower and Schuh, 2021).

Linnaeus’s Systema Naturae endured for centuries because of its ability to convey, through language, the observed hierarchical structure of the living and inanimate world (Padial and De la Riva, 2021). Over time, several schools of philosophical perspectives emerged, from early taxonomy theory to Darwinism phylogenetic inference. Darwin paved the path towards a paradigmatic shift that gradually replaced both traditional classifying methods of species (e.g., by similarities or differences, degrees of divergence, or diagnosed populations) and the abstract notion of hierarchical levels attributed to natural or divine order during Linnaeus’s era. In contrast, phylogenetic trees offered a newer and more complete synthesis, deciding species classification by their phylogenetic lineage. Relationships cannot be directly observed but must be inferred through empirical evidence and patterns of variation (De Queiroz, 2007; Sites Jr and Marshall, 2004; Wiley, 1978). This framework shifted the focus of evolutionary biology from intrinsic reproductive isolation as the sole driver of speciation towards a more integrated genetic understanding of the origin of species (Padial and De la Riva, 2021).

Since systematics studies the observable morphological similarity and also fosters ontological discussions of nature and origins of biological diversity, biological classifications have significant scientific value as indices of accumulated knowledge (Brower and Schuh, 2021). As Foucault (1970) described in The Order of Things, the naturalists “distinguishes the parts of the natural bodies with his eyes, describes them appropriately according to their number, form, position, and proportion, and he names them”, emphasizing that the classification depends on what can be perceived and conceptualized within a given historical framework. Under this view, every classificatory system reflects not a timeless natural order but the epistemic conditions of its era. Brower and Schuh (2021) similarly argue that biological classifications embody the scientific perspectives and methodological constraints within which they are produced. Yet, past errors and the abandonment of outdated beliefs retain value, as they represent the ongoing confrontation with complex, indirect, fragmented evidence analyzed through assumptions that are themselves subject to refutation (Padial and De la Riva, 2021). Collectively, these perspectives reflect the underlying values of historically evolving systems of KO. Each system indexes how biological species were understood at a particular moment in scientific history.

When Kuhn first introduced the term paradigm, he argued that scientific paradigms emerge through the continual identification of anomalies and repeated experimentation that generate new narratives within existing domain. Observations evolve into rules, rules develop into theories, and theories show the invisible boundaries of scientific consensus that constitute paradigms. The discovery of knowledge, therefore, is not a linear process but one driven by recurring cycles of anomaly detection and conceptual reconstruction (Kuhn, 1962). Scientists owe their success to their ability to identify and solve problems using conceptual and instrumental techniques that already exist within the prevailing paradigm.

The contemporary scientific process, however, is different from the one Kuhn described. Unlike in Kuhn’s time, most researchers today are trained and socialized within paradigm-centered scientific communities, where their work is shaped by daily engagement with experiments and collaborative infrastructures. Modern science advances increasingly through networked accumulation and tools-mediated integration. The exponential growth of scientific knowledge gives rise to competition between scientists who pursue similar goals with comparable training, ambition, and resources (Wang and Barabási, 2021). To reach the frontiers of discovery more efficiently, scientists have narrowed their specialization and collaborate extensively to integrate diverse knowledge backgrounds essential for modern scientific research (Wang and Barabási, 2021).

Jones described this phenomenon as the “burden of knowledge”, arguing that innovators must undertake extensive education and training to reach the frontier of discovery (Jones, 2009). As scientific knowledge continues to expand exponentially, the structure of research itself has evolved. The growing complexity of contemporary science challenges the solitary model of inquiry that characterized earlier paradigms. One major response to this challenge is the rise of multidisciplinary collaboration. This increasing collaboration reflects the difficulty of conducting complex studies individually. This is not because scientists today are less productive (in fact, individual productivity is stably rising from 1900 to 2000), but because the amount of knowledge has been accumulating from the exponential growth of science and fewer new paradigm shifts happen within the existing fields (Gingras and Wallace, 2010).

Therefore, Kuhn’s view that emerging scientific paradigms must derive from new theories may not be entirely suitable for describing contemporary scientific discovery. Ankeny and Leonelli (2016) argued that paradigms are no longer useful as a framing concept for three reasons. First, paradigms tend to be highly static and inflexible in which changes only occur in dramatic fashion. Second, Kuhn considers conflicting paradigms to be incommensurable, therefore unable to coexist except in extreme moments of crisis. This does not align well with current interdisciplinary research, where different epistemic approaches often appear in parallel or at the intersection between paradigms to uncover questions. Third, Kuhn prioritizes theoretical knowledge as the primary output of science, while marginalizing applied science and technological innovations. Kuhn also limited scientific communication to textual documents, excluding other forms such as datasets, graphs, and charts from his definition of legitimate scientific outputs (Ankeny and Leonelli, 2016). A fourth point can be added to this argument, that is, the melding and interconnections of multidisciplinary scientific research do not follow the dramatic rupture with which Kuhn associated the emergence of new paradigms.

Instead, the new paradigms reflect a gradual convergence of existing domains, blurring disciplinary boundaries through the alignment of key concepts. These concepts, while semantically and distinctively defined, imply overlapping conceptual meanings. Much of today’s scientific progress, particularly in fields such as bioinformatics and computational biology, is driven less by theoretical contributions but more by revealing hidden patterns and relationships among large and complex data entities (Leonelli, 2016). In biology, Noble (2002) characterized this integrative approach through a “middle-out” framework for biological modeling, in which researchers work simultaneously upward from molecular mechanisms to physiological systems and downward from system functions to cellular components, with continuous integration occurring at intermediate levels. This approach transcends the traditional dichotomy between reductionist “bottom up” and holistic “top down” strategies. Data as empirical evidence have theoretical value as it embodies knowledge through classification (Leonelli, 2016). This coincides to Hjørland’s point that knowledge is embedded in larger frameworks of human activities (Hjørland, 2015). In data-intensive biology, research often starts at whichever level with sufficient data (e.g., pathways, cells, tissues, organs) and extends both vertically and horizontally to connect across multiple biological scales.

The application of NLP in biomedical research enables information extraction from electronic health records for hidden clinical narrative discovery (Houssein et al., 2021). Similarly, the development of GenBank, a public repository of annotated nucleotide sequences maintained by the NCBI, shows how scientific progress can be driven by the creation of shared data ecosystems that support reproducibility, retrieval, and the synthesis of biological knowledge (Benson et al., 2014). These examples of interdisciplinary fusion represent forms of convergence, not “paradigm shifts” in Kuhn’s conventional sense, but rather, gradual adaptations across domains facilitated by interoperable terminology, method, and data infrastructures.

This emerging paradigm of science combines individual efforts into a form of collective intelligence, defined as the general ability of a group to perform a wide variety of tasks and achieve deeper integration of scientific knowledge (Woolley et al., 2010). This shift raises questions about how organization of scientific knowledge within interdisciplinary paradigms are conceptualized evolve in a consistent pace with frontier discoveries. Effective classifications of knowledge enhance researchers’ capacity to identify connections across diverse literature or databases, thereby generating novel combinations of previously distinct insights (Szostak et al., 2016). As interdisciplinary research increasingly involves larger teams and more heterogeneous datasets, classification systems and KOSs at large are being adopted as foundational knowledge in science processes, even though such a role of KOSs is not always fully recognized by the researchers who adopted them.

Unlike paradigms that guide research through agreed identification and interpretation of concepts and epistemology (Kuhn, 1962), KOSs can support research by offering semantic alignment, enabling diverse teams to describe, use, and interpret data in consistent ways. Researchers who use KOSs implicitly accept, even if they might not be aware of it, the definition of entities and processes contained within KOSs at the moment they consulted it (Leonelli, 2016). For example, GO is one of the most successful bio-ontologies that offers comprehensive, structured, computer-accessible representation of genes and genetic functions (Ashburner et al., 2000; Pesquita et al., 2009). It provides a framework and a set of concepts for describing the functions of gene products by assigning GO annotations to the association between a specific gene product and a GO concept. Together, they make a statement pertinent to the function of that gene. There are three types of GO concepts in GO to represent the distinct aspects of how gene functions can be described: molecular function, cellular component, and biological process (Dessimoz and Škunca, 2017). The logic is that a gene encodes a gene product, and that gene product carries out a molecular-level process or activity (molecular function) in a specific location relative to the cell (cellular component), and this molecular process contributes to a larger biological goal (biological process).

A biologist who uses GO data for their research underlyingly means they accept the descriptive predicates that are in the GO framework, including definitions and conceptual boundaries. Descriptive metadata captured in KOSs thus have the same epistemic value as testable hypotheses: they are theoretical statements whose validity and meaning can only be interpreted and assessed with empirical evidence and the conditions under which that evidence has been produced. This crucial role in expressing disciplinary embodied knowledge qualifies KOS as a form of theory (Hjørland, 2015; Leonelli, 2016).

The convergence of scientific paradigms and the increasing reliance on structured relational KBs in interdisciplinary research reflect a growing need for shared understanding, often referred to as common knowledge (Neumann, 2010). This includes a collective foundation of terms, concepts, assumptions, and frameworks that enable coordination and collaboration across diverse domains. Although conceptual meanings may vary across disciplines, such ambiguity can be minimized using precise and formal structures of language. Conceptual clarity depends not only on empirical accuracy but also on the formal structure of the language used to express it (Reichenbach, 2012). In contemporary science, this shared structure is frequently established through human-mediated KR (Qin, 2020), where ontologies and classificatory systems mediate raw data and conceptual reasoning.

A related concept in linguistics and philosophy of AI is common sense, which serves as a measurement of whether an AI system is sufficiently intelligent to process and understand human language and its underlying meaning, thereby performing referring and reasoning (Browning and LeCun, 2023). For humans, common sense runs subconsciously in almost every decision in daily life. For instance, in the Winogard Schema Challenge, the following pair of sentences is used to test a machine’s capacity for common-sense reasoning:

The trophy doesn’t fit inside the suitcase because it is too large.

The trophy doesn’t fit inside the suitcase because it is too small.

Common sense allows us to effortlessly determine what “it” refers to in each sentence, based on the implied size difference. Although LLMs can pass such challenges, it does not necessarily imply that LLMs can truly understand the meaning of semantic knowledge and common sense (Browning and LeCun, 2023). Studies also show that LLMs can exhibit hallucinations by fabricating evidence that contradicts real-world facts (Huang et al., 2025b; Lu et al., 2024). While deploying RAG into LLMs can reduce hallucination, the performance can vary depending on how well models architecture uses external KBs, where in these situations they function as a form of semantic modeling.

KOSs map between model-theoretic representations of scientific domains and the natural language expressions researchers use to describe them. By codifying knowledge into structured forms, it enables conceptual regularity to support interpretation and semantic interoperability. Establishing connections between structured representations of knowledge and natural language is essential to enabling AI to not only process data, but to reason with it, thus truly achieving trustworthy AI-driven science inquiry.

In the context of KO, classification theories and systems are socially and culturally dependent artifacts that directly reflect the contextual realities in which they are developed (Beghtol, 2009). Classification theories serve as intellectual foundations for the practice of KO. Regardless of their specific names or forms, all KOSs are based on a set of principles that guide KO practices. The conceptual basis for a KOS defines its purpose and warrant, shaping both its content and structure. In this sense, classification theories rationalize the need for and functions of KOSs while defining the semantic relations of concepts they represent.

Some scholars have further argued that KOS itself can be regarded as a classification theory. The case of KBs well illustrates this perspective. As modern instantiations of KOSs, KBs have been viewed by logicist AI as theories because they comprise collections of terms with associated axioms designed to constrain unintended interpretations and to enable the derivation of new information from ground facts. KBs are often crafted in ways that both reflect common-sense knowledge in a declarative form while taking advantage of the reasoning capabilities used by a given automated system (Smith and Welty, 2001).

In systematics, the groundbreaking Linnaean taxonomy functioned as a conceptual theory in eighteenth-century botany, as stated in Data-Centric Biology: A Philosophical Study:

“[Bio-ontologies] constitute a form of scientific theorizing that has the potential to affect the direction and practice of experimental biology in the long term, and yet they do not aim to introduce new language into biological discourse but rather attempt to gather and express consensus on what constitutes established knowledge” (Leonelli, 2016, pp. 121–122).

This interpretation of bio-ontologies derives from pragmatic epistemology and a structuralist ontological stance, which view theories as tools that scientists employ to define, refine, search for, and evaluate answers to empirical and theoretical questions (Balzer and Moulines, 1996). Treating ontologies as classificatory theories emphasizes their role in enabling the dissemination and retrieval of research materials while providing systematic scrutiny and interpretation of empirical evidence. The theoretical value of descriptive predicates lies not in their capacity to explain specific phenomena, but rather in their ability to classify entities that sustain both stability and dynamism in the evolving understandings of life science.

Recent studies have proposed a paradigm shift from KOSs to knowledge organization ecosystems (KOEs), highlighting the collaboration of interdependent components such as individuals, knowledge organization models and organizations, within platform-based socio-technical systems (STSs) (Bagchi, 2021). The KOE framework encompasses three interrelated dimensions: individuals (designers, domain experts, transdisciplinary ethics experts); technology (representational appropriateness of knowledge models, inter and intra domain knowledge artifacts’ interoperability); and organizations (technology consortiums) (Bagchi, 2021). Informed by foundational works in socio-technical ethics and the philosophical ramifications of KO, this framework establishes a conceptual bridge between STS and KO theories, thereby illuminating their influences on classical AI problems (e.g., knowledge-based AI systems).

Biology is among the most classification-oriented natural sciences, so devoted to classifications such that the special division, biological systematics, emerged to exclusively develop classifications under the framework of “theory + philosophy + conceptual history” (Pavlinov, 2021). Since the 18th century, systematics and its effort to create natural systems of living beings was considered a branch of natural philosophy and the product of human cognitive activity. There are two main components in the conceptual history of systematics—anagenetic and cladogenetic. Anagenetic means a sequential progression of taxonomic groups from less to more, while cladogenetic means fragmentation and multiplication (Pavlinov, 2021). Two components correspond to the unification and diversification trends separately in systematics, thus developing into an open information system that allows the third component—network component, to play its role in exchange and combination.

Influenced by important scholars and their philosophical contributions, e.g., Charles Darwin and Ernst Haeckel, systematics embarks on different periods of theories. Classification systems were proposed suitable for direct interpretation of systematics. The structuralist approach to classification systems ($CS$) entails a union of taxonomic system ($TS$) and partonomic system ($PS$), each defined as an ordered pair of units and relations, optionally involving the background theory ($BT$) parameter for meaningful interpretations (Eqns. 1,2,3). Thus, the general definition of the classification system is:

(1)$$CS\supset \{TS\}\cup \{PS\}$$

(2)$$TS\supset BT\{T,C_T,R_T,R_{TC}\}$$

(3)$$PS\supset BT\{P,P_T,R_P,R_{TP}\}$$

where $T$ represents taxa (the taxonomic units being classified), $C_T$ represents taxonomic characters (the features used to characterize taxa), $R_T$ represents relations between taxa (such as similarity or kinship), and $R_{TC}$ represents taxon-character correspondence (how taxa relate to their defining characteristics). In the partonomic system, $P$ represents partons (the structural parts or components), $P_T$ represents the taxa to which partons are assigned, $R_P$ represents relations between partons (such as homology), and $R_{Tp}$ represents taxon-parton correspondence (how structural parts relate to their assigned taxa) (Pavlinov, 2021).

The new systematics era encountered fallacies by confusing the mission of taxonomy with population genetics. Population genetics is experimental and conducted in real time focusing on genetic phenomena within and between populations of living species, while taxonomy is comparative, historical, and non-experimental, and emphasizes at and above the species level. Some scholars predict the decay of systematics because of advancements in technology. Systematics seems vulnerable against trendy, modern, and easily fundable sciences like genetics or molecular science. However, narrowly focusing on one group or genus through molecular measures and an over-obsession in “DNA barcoding” of new species’ identification overlook the importance of conceptual understanding of species and the intrinsic relations leading to knowledge of natural life. Studies argue for new knowledge organization and management tools designed for taxonomists to return their focus to species, their characters and relationships. It is imperative that technology should be secondary to the pursuit of knowledge in science (Wheeler, 2023).

Apart from technology, informatics has also dramatically altered taxonomic working practices and workflows. One notable change is the discoverability and reusability of taxonomic data, both through publication venues, e.g., European Journal of Taxonomy, Pensoft Journals, databases, repositories, and KBs, e.g., DNA Data Bank of Japan, the European Molecular Biology Laboratory, GenBank, the International Barcode of Life, MorphoBank, TraitBank (Secretariat of the Convention on Biological Diversity, 2021). The DELTA (DEscription Language for TAxonomy) was developed in 2000 as an information language system specifically for biological taxonomy, providing standardization for encoding taxonomic information on biological species (Coleman et al., 2010). Domain-specific NLP algorithms are also designed for biological information extraction including cellular processes, taxonomic names, and morphological characters (Thessen et al., 2012). Advanced AI techniques such as LLMs offer even more diverse and powerful support to biological classification and KO.

Within biomedical discovery, the language interpretation capabilities of LLMs can provide a framework for integrating and synthesizing the complex evidence space and tools, systematizing and lowering the barriers to access and reason over multiple structured databases, repositories and KOSs, enriching the background knowledge through specialized ontologies and serving as interfaces to external analytical tools (Wysocki et al., 2024).

Even before the emergence of LLMs, earlier ML models employed KOSs for computational analysis in biomedicine. Bio-ontologies were used as reliable structured KBs with semantically rich relationships between properties for gene-disease association discovery and identification. Advanced genome sequencing technologies accelerated the process of exploring genomic variations (Piñero et al., 2020) and genetic markers detection (Chang et al., 2024). ML models accelerate the discovery in searching for primary causes of genetic diseases by predicting functional similarities of genes with their semantic similarity within KOSs. The backbone of this approach is the schema-guided, computer-accessible bio-ontologies such as GO and Disease Ontology (DO). By calculating the semantic distance between entities in GO, models such as Support-Vector Machines (SVM) can predict the functional similarity between gene products, therefore predicting genes that potentially share the same association with disease-triggering proteins (Kulmanov et al., 2021; Liu and Qin, 2025).

Such design has been applied in COVID-19 research and proved to be powerful. One of the key challenges during the pandemic was the diversity of concepts and the breadth of background knowledge needed to understand the disease and develop diagnostic and therapeutic solutions. Even a single research article could span multiple biomedical subfields as well as physics, chemistry, engineering, computer science, and social science. To address this, Hope et al. (2020) introduced SciSight, an exploratory search system designed to help researchers navigate the rapidly evolving COVID-19 literature. SciSight combined faceted navigation, entity extraction, and network visualizations to surface relationships between research topics, biomedical entities, and scientific groups. While not as formalized as a logic-based ontology, SciSight functioned as a lightweight, dynamic KOS that supported sense-making and collaborative discovery at scale. These approaches prove how KOSs can act as integrative infrastructures in highly dynamic, interdisciplinary problem spaces, setting the stage for their role in AI-assisted scientific reasoning.

Hope et al. (2021) also developed a method for extracting mechanistic knowledge from the COVID-19 Open Research Dataset (CORD-19) by combining automated machine reading with targeted expert curation. Their framework aggregated heterogeneous mechanistic assertions into a unified KB. It covers entities such as viral proteins, host factors, pathways, and disease processes. This enabled queries to trace causal chains and uncover cross-disciplinary connections, exemplifying that, once formalized into a structured, computable representation, diverse evidence can facilitate systematic exploration of emerging literature and support rapid hypothesis generation in crisis-driven research. Adding another layer to this ecosystem, Wang et al. (2021) demonstrated the value of domain-specific pretraining for vertical search in biomedical literature. Using large-scale corpora to adapt transformer models to biomedical language, they achieved substantial gains in retrieving relevant documents from specialized repositories. While their work focused on retrieval rather than explicit ontology construction, the improved access to precise, contextually relevant literature directly supported downstream KOS applications particularly in crisis contexts such as COVID-19, where prompt identification of relevant evidence was essential.

Supporting knowledge discovery in AI-assisted science inquiry and reasoning is another key function of Bio-KOSs. While LLMs have already shown exceptional capabilities in medical knowledge competency (Nori et al., 2023), baseline models such as OpenAI’s GPT-4 (Achiam et al., 2023) tend to hallucinate or respond with overly general answers due to the lack of domain-specific external KBs. For LLMs to achieve scientific ideation and fully automated reasoning, the framework for deep learning or adaptive inference must provide high-quality scientific artifacts such as scholarly publications, metadata as signals of quality and interest, and structured data as support (Hope et al., 2023; Zhou et al., 2024).

Complementing LLMs with RAG mechanisms enables contextual control over relevant background knowledge and facts, thereby enhancing analytical precision (Wysocki et al., 2024). The theoretical foundation of such an approach mirrors human scientific reasoning: new ideas are generated by building upon existing ones. Scholars synthesize accumulated knowledge and external evidence to form insights, answers and research directions (Hope et al., 2023). Similarly, KBs supply AI models with continuously updated knowledge derived from clinical, experimental, and population-level genetic research. Integrating multiple KOSs significantly improves the competence of AI-driven scientific ideation. Consequently, KR in biology must support not only classification but also reasoning capabilities.

Neumann (2010) introduced the modal logic as a means for computational systems to encode varying degrees of epistemic commitment, such as “believes”, “knows”, and “possibly”. In the context of biological discovery, such logic-based representations enable reasoning across incomplete data and evolving hypotheses, which is an essential capacity in a dynamic field of science characterized by ambiguity and exception. Biologists frequently work with layered hypotheses and partially overlapping models that cannot be verified fully in isolation but can gain epistemic strength when integrated through structured representation and logical rules. Bio-ontologies, with their formal semantics and intrinsic ability to manage unverified concepts and relations, have the potential to support both deductive reasoning and exploratory inference across relational structures. These systems do more than define what is known; they establish the infrastructure for discovering what is knowable. As Neumann (2010) emphasized, recognizing where knowledge comes from is just as important as knowing it in biomedical research.

In AI-driven scientific research, KOSs do more than serve as simple reflections of scientific understanding. Their value in providing quality data and knowledge infrastructures in AI for science is increasingly recognized by AI model and application developers. Meanwhile, rapid advancements in AI are introducing new methods and perspectives into the design and use of KOSs. The classification and entification of concepts in KOSs, as in the case of GO, are essential not only for data management and knowledge representation, but also for critically shaping the directions of scientific inquiry they enable (or constrain). When embedded into LLM pipelines, structured networks of entities and relations can affect the performance of AI models in executing complex research tasks. Since the release of GPT-4, numerous NLP studies have incorporated KOSs as external KBs within RAG frameworks to enhance LLM performance on advanced scientific tasks, including project workflow construction (Ramirez-Medina et al., 2025) and hypothesis or idea generation (Abdel-Rehim et al., 2025; Guo et al., 2024).

The first dimension is the level of functional sophistication of KOSs. We propose five distinct levels incrementally ranked from the most foundational to advanced (Table 1).

The second dimension, degrees of automation in KOS construction, describes the extent to which computational techniques such as ML, NLP, or rule-based extraction are involved during the building, integration, and maintenance of KOSs. It captures the shift from human expert-driven curation toward data-fueled automation pipelines in constructing KOSs. To compare the extent to which human ability and computational methods contribute to these processes, we assigned a score from 1 to 4 for each paper as an indicator for the degree of automation. This dimension captures both the technical mechanisms employed and the degree of human intervention needed throughout the processes of knowledge acquisition, integration, and maintenance (Table 2).

While the second dimension assesses automation during the construction phase, the third dimension evaluates the epistemic capability of the completed KOS on how effectively it can perform reasoning, inference, and discovery beyond its explicitly encoded content. It moves from systems that merely retrieve or represent information toward those that can generate new association, infer relationships, or support hypothesis formulation. Conceptually, this dimension parallels long-standing challenges in AI research. We use five tiers to rank the extent to which a KOS works in a spectrum of knowledge organization systems’ tasks, ranging from static information storage and retrieval to structured query answering, logical reasoning, and scientific discovery (Table 3). Each tier bears all the properties from its lower tier.