There are four main types of KOS. Based on the level of sophistication, the simplest type is term lists, such as glossaries and dictionaries. The second type is metadata-like models, including gazetteers, directories, and authority files. Classification and categorization go to the next level of sophistication because of the embedded relations between concepts or classes in classification/categorization schemes, taxonomies, and subject headings. The relationship models are the most sophisticated among the four KOS types. Ontologies and semantic networks are the two members of the relationship model, which possess all traits of a KOS that one can hope for: explicitly representing concepts/entities with unambiguous terms while embedding relations between concepts/entities (Zeng, 2008). In the sense of representing a conceptual system via a logical theory, an ontology consists of an annotated and indexed set of formal propositions or assertions about things, a collection of assertions that are called a theory in logic (Guarino & Giaretta, 1995). This special property of ontologies is as closely as it can be in fitting the knowledge representation ideology in AI, which emphasizes sufficiently precise notation, adequacy and expressiveness of representation schemes (Bench-Capon, 1990).

Ontologies sometimes are also called knowledgebases because they are essentially a collection of symbolic structures representing the world based on our cognition (Levesque & Lakemeyer, 2022). Such representation is semantically rich with not only unambiguous vocabularies for entities and individuals in the entities, but also explicit relations between the entities that go beyond hierarchical and associative relations that are the only available relation types in many traditional KOS. Not all ontologies, however, can be called knowledgebases. For example, Schema.org is an ontology developed in collaboration among Google, Microsoft, Yahoo, and Yandex to be used for representing web content as structured data (Google et al., 2024). The purpose of Schema.org is to provide a standard representation scheme for classes of entities and properties these entities possess. Since it is just a representation scheme, it does not contain data (i.e., individual members of classes). Therefore, it is not a knowledgebase and cannot be used for modelling or reasoning purposes.

While many traditional KOS are not designed for problem-solving nor as a data source for ML modelling purposes, many ontologies in the biomedical domain have disrupted the tradition and evolved into knowledgebases. For instance, gene ontology (GO) (Ashburner et al., 2000) and disease ontology (DO) (Bello et al., 2018) use axioms to model knowledge in order to define the semantic interpretation of the presented entities, rules, and class constraints and present multiple relations between concepts. Knowledge organization and AI communities, two research fields that were once separated and did not have much communication before, are now drawing closer and converging through advances in semantic web technologies and data science (Qin, 2020). This trend symbolizes a change to knowledge organization practices. More importantly, it is a signal to knowledge organization moving towards a more interoperable, practical, and heterogeneous identity that exceeds the simple purpose of storing knowledge.

One application of biomedical ontologies is in disease-gene 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), generating vast amount of data. This data is preserved and organized into KOS by biocurators who apply knowledge organization theories and practices, which creates semantically rich data sources for applying ML algorithms to analyse larger and complex data sets (Libbrecht & Noble, 2015). Complex diseases with a strong genetic influence often have multiple aetiologies with the involvement of possibly hundreds of different genes. Supervised ML methods can trace hidden relationships among disease-causing genes in existing datasets to discriminate unknown disease genes from non-disease genes (Asif et al., 2018). This advancement plays a crucial role in disease diagnosis and form the basis for clinical decision-making (Chang et al., 2024). This approach was once difficult to proceed due to the lack of: (a) structured KOS with semantically rich relationships between properties; and (b) a powerful computational hardware with feasible models to analyse large heterogeneous data sets. Problem (a) was greatly improved through making data FAIR (findable, accessible, interoperable, reusable) (Wilkinson et al., 2016) and encoding languages (e.g. XML and JSON) for representing machine-readable knowledge and reflecting reality. Problem (b) now is being addressed by numerous supervised and unsupervised ML models that are developed and applied to assist in areas such as genetics and molecular science.

One major issue in identifying disease-gene association is how to define an unknown gene can cause similar outcomes as a disease-associated gene already known. Biologists found that functional similar genes tend to contribute to similar phenotypes. For instance, etiologically relevant genes disrupted by genetic variants in ASD patients tend to aggregate in specific biological processes (Voineagu & Eapen, 2013). This means disease-causing genes and disease-candidate genes may belong to the same tree path in GO DAG structure (Asif et al., 2018). Gene products that share highly overlapping GO terms may have higher functional similarities. If a gene product is on the same tree branch with gene products that are associated with disease-related GO annotation function terms, it indicates this gene may have a higher probability to be a disease-causing gene. Thus, the comparison of gene products’ function similarities is transformed into the comparison of semantic similarities of GO terms. Adopting ontology annotation data can provide a means to compare entities on aspects that otherwise not be comparable.

Typically there are two ways to compare terms in graph-structured ontologies such as GO: edge-based or node-based (Pesquita et al., 2009). Edge-based approaches are based on counting the number of edges in the graph path between two terms (Rada et al., 1989). This can be problematic for biological data because the approach is based on two conditions: (i) nodes and edges in the biological ontology are uniformly distributed, and (ii) edges at the same level in the ontology correspond to the same semantic distance between terms. However, biological knowledge can rarely meet these two conditions where terms at the same tree structure level share the same scale and weight.

Node-based approaches are more commonly accepted in biological domain. There have been several statistical measurements and equations developed. Resnik proposed the Information Content (IC) to quantify the informativeness of a concept c as negative the log likelihood (Resnik, 1999):

However, a drawback of Resnik’s method is that it ignores the information contained in the structure of the ontology by only concentrating on the information content of a term derived from the corpus statistics. For biological ontologies, the specificity of a GO term is usually determined by its location in the GO graph. A GO term’s semantics (biological meanings) are inherited from all its ancestor terms (Wang et al., 2007). For instance, shown in figure 1, GO:0002839 positive regulation of immune response to tumour cell is a child term of GO:0002418 immune response to tumour cell, the latter is a child term of GO:0002347 response to tumour cell and GO:0006955 immune response, both of which are the child terms of GO:0008150 biological processes. Because GO term is the aggregation of all its parent terms, GO:0002839 positive regulation of immune response to tumour cell should possess the characteristics of both GO:0002418 response to tumour cell and GO:0006955 immune response. Not to mention biomedical ontologies usually have various edge length i.e. edges at the same level convey different semantic distances, various depth i.e. terms at the same level have different level of details, and various node density i.e. some areas of the ontology have a greater density of terms than others (Pesquita et al., 2009).

Wang proposed a metrics that is specifically for encoding biological terms’ semantics by aggregating the semantic contributions of all its ancestor terms including itself in the GO graph. Wang believes Resnik’s IC approach focus is more applicable for knowledge in natural language such as bird and crane, forest and graveyard but is not the ideal option for biological knowledge in ontologies (Wang et al., 2007). IC neglects the logic that if two GO terms share the same parent are near the root of the ontology i.e., terms that are more general, they should have larger semantic difference than two terms having the same parent and being far away from the root of the ontology because the latter are more specific terms. GO is constructed in such way that if the child GO term describes the gene product, then all its parent terms must also apply to that gene product (Wang et al., 2007). DO contains logical definitions (axioms) to describe relevant disease drivers, constructed with specific relational ontology (RO) terms, to create a restriction between a DO term and another open biological and biomedical ontology (OBO) Foundry ontology term. Thus, DO has the ability to ‘infer’ the child terms originated from a parent term based on a known parent-child relationship (Qin & Liu, 2024). Therefore, it is reasonable to aggregate the biological meanings of all its ancestor terms when determining its semantic similarity with another GO term.

Next, the issue is transforming semantic similarity of GO terms into the functional similarity of gene products. Since one gene product may have more than one GO annotation, we must consider the contributions from the semantically similar terms that annotate the genes separately (Wang et al., 2007). Wang first defines the maximum semantic similarity between one GO term and a set of GO = {go₁,go₂, … ,go_k}as:

Then, given two gene products G1 and G2 annotated by sets of GO terms: GO1 = {go₁₁, go₁₂, …, go_1m} and GO2 = {go₂₁, go₂₂, …, go_2n} respectively, the functional similarity of G1 and G2 can be defined as:

In this study, we applied Wang’s approach to calculate the functional similarity of gene products for predicting candidate genes that lead to ASD disease. GO annotation data were collected using the org.Hs.eg.db R package (Carlson et al., 2019). Only terms for biological processes are selected (N = 28,140). Next, we mapped all the GO terms and their associated gene products (N = 157,247). Since our aim is investigating the construction and quality of GO as a knowledge ontology in biomedical research, we only try to replicate part of the workflow by Asif et al., using three of the ML models – support-vector machines (SVM), random forest (RF), and gradient boosting (GB). We randomly selected 1,000 gene products from the mapping as our train set for the models. For our test set, we also randomly selected 20 candidate gene products obtained from the Simons foundation autism research initiative (SFARI, https://gene.sfari.org/) gene database (N = 1,176), in which 15 are categorized by SFARI as high confidence disease genes (HD) and 5 are categorized as low confidence (LD) disease genes. This categorization is used to compare with results from the ML classification models. Table 2 shows the 20 test gene products and their categories. Before we allow the model to conduct semantics similarity calculation, we removed SFARI’s categorization for the models to predict and then verify their overall performance by comparing its results with SFARI’s classification. The semantic similarity measures between GO terms were implemented using the GOSemSim R package (Yu et al., 2010).