Bio-KGvec2go: Biomedical KG Embeddings

• Bio-KGvec2go is a web-based platform that creates and serves pre-trained knowledge graph embeddings for biomedical ontologies with dynamic updates.
• It extends the original KGvec2go by incorporating the Gene Ontology and Human Phenotype Ontology along with multiple embedding models to reduce retraining effort.
• The service features version-aware storage and easy API/GUI access, enabling reproducible analyses and supporting diverse biomedical applications.

Bio-KGvec2go is a web-based service for generating, hosting, and distributing pre-trained knowledge graph embeddings for biomedical ontologies, specifically the Gene Ontology and the Human Phenotype Ontology. It extends the original KGvec2go framework from general-purpose RDF knowledge graphs to dynamically updated biomedical resources, with support for multiple embedding model families, version-aware storage, and API- and GUI-based access to vectors, similarity scores, and nearest-neighbor queries. Its central objective is to provide up-to-date embeddings with minimal computational effort required from users, thereby reducing repeated retraining, improving reproducibility across ontology versions, and supporting sustainable computing in biomedical AI workflows (Ahmad et al., 9 Sep 2025).

1. Conceptual basis and motivation

Biomedical knowledge is inherently relational: diseases have symptoms, genes encode proteins, proteins participate in pathways, and drugs target proteins. Knowledge graphs represent such structure as entities and relations, while ontologies formalize the domain vocabulary and its semantics. In biomedicine, the Gene Ontology and the Human Phenotype Ontology function as core infrastructure for annotation, integration, and reasoning, and both are subject to continuous curation and extension (Ahmad et al., 9 Sep 2025).

Knowledge graph embedding methods address the mismatch between graph-structured knowledge and the fixed-length real-valued vectors expected by many machine learning systems. In the setting described for Bio-KGvec2go, embeddings are used to preserve graph structure and, to varying degrees, semantics in low-dimensional vector spaces. The motivating use cases explicitly include protein function prediction, protein–protein interaction prediction, gene–disease association discovery, patient similarity, disease diagnosis support, and link prediction or classification over biomedical graphs (Ahmad et al., 9 Sep 2025).

The service is framed as a response to several limitations in existing public resources. General-purpose services such as the original KGvec2go do not provide embeddings for GO or HP. Many published embedding releases correspond to a single static snapshot, whereas biomedical ontologies change monthly or every few months. Without a hosted service, users must monitor releases, download updated files, implement and tune embedding pipelines, and manage versioning themselves. Bio-KGvec2go is intended to remove that burden by centralizing training and serving embeddings as a reusable resource (Ahmad et al., 9 Sep 2025).

A plausible implication is that Bio-KGvec2go occupies an infrastructural position rather than a task-specific one: it is designed primarily to supply standardized representations and provenance rather than to prescribe a single downstream analytical workflow.

2. Relationship to KGvec2go

Bio-KGvec2go is explicitly described as an extension of KGvec2go. The original KGvec2go is a web service exposing pre-trained knowledge graph embeddings through a simple HTTP API so that downstream applications do not need to train models themselves or download and maintain large embedding files locally. Its architecture is RESTful, lightweight for clients, and model-agnostic at the API level. It was implemented in Python, used Flask as the web framework, and could be deployed behind Apache HTTP Server (Portisch et al., 2020).

The original service served embeddings for four non-biomedical knowledge graphs: ALOD, DBpedia, WordNet, and CaLiGraph according to the Bio-KGvec2go paper, and it used RDF2Vec as its embedding method. Its API supported retrieval of entity vectors, computation of cosine similarity, and nearest-neighbor lookup for related entities. It was designed to run efficiently even on devices with limited CPU and RAM (Ahmad et al., 9 Sep 2025).

Bio-KGvec2go preserves the central architectural idea of “embeddings as a service” while extending the system along three axes. First, it adds biomedical ontologies, namely GO and HP. Second, it broadens model support beyond RDF2Vec to include TransE, TransR, DistMult, HolE, and BoxE, using PyKEEN for those models and pyRDF2Vec for RDF2Vec. Third, it introduces dynamic updating and version management: ontology releases are periodically fetched from official sources, checksums are compared, all models are retrained when a change is detected, and multiple ontology versions are hosted concurrently (Ahmad et al., 9 Sep 2025).

The result is not merely a biomedical rebranding of KGvec2go. It alters the service model from static serving of pre-trained embeddings to a version-aware, update-sensitive system aligned with the release practices of biomedical ontologies.

3. Supported ontologies, graph construction, and versioning

Bio-KGvec2go currently supports two biomedical ontologies: the Gene Ontology and the Human Phenotype Ontology. These were selected because they are described as among the most widely used and impactful biomedical ontologies, which maximizes the utility of a reusable embedding service (Ahmad et al., 9 Sep 2025).

Ontology	Scope	Structural characteristics
Gene Ontology (GO)	Protein/gene function	More than 40,000 classes; three domains; relations include is_a, part_of, regulates
Human Phenotype Ontology (HP / HPO)	Phenotypic abnormalities in human hereditary diseases and related clinical concepts	More than 18,000 classes; directed acyclic graph; all edges are is_a

For GO, the ontology is organized into Biological Process, Molecular Function, and Cellular Component. The graph view used by the service treats GO classes as nodes and semantic relations such as is_a, part_of, and regulates as edges, with is_a described as dominating. Monthly official releases are available through https://release.geneontology.org/, and the ontology undergoes additions, deprecations, and restructuring over time. The applications cited for GO include protein function prediction, protein–protein interaction prediction, and gene–disease association prediction (Ahmad et al., 9 Sep 2025).

For HP, the ontology covers phenotypic abnormalities, past medical history, mode of inheritance, clinical course, clinical modifiers, and frequency. It is represented as a directed acyclic graph in which all edges are is_a. Releases occur roughly every 1–2 months on GitHub at https://github.com/obophenotype/human-phenotype-ontology/releases. The cited application areas include patient similarity computation, genotype–phenotype association prediction, and gene–disease association prediction (Ahmad et al., 9 Sep 2025).

As of the paper’s timeframe, the system hosts embeddings for six versions of each ontology, with the first version dating back to 2023 and subsequent versions appearing approximately every six months. For each ontology version and each embedding model, the vectors are 200-dimensional. The service preserves historical versions while using the most up-to-date version by default for similarity and nearest-neighbor queries (Ahmad et al., 9 Sep 2025).

The ontology-to-graph transformation is deliberately limited. The embedding input consists of ontology classes and class–class relations. There is no explicit handling of datatype literals, numeric or string attributes, complex OWL axioms, or reasoning-derived edges. This suggests that the current representation emphasizes graph topology and direct ontological relations rather than the full expressivity of OWL semantics.

4. Embedding models and training configuration

Bio-KGvec2go includes six knowledge graph embedding models spanning several modeling paradigms: TransE, TransR, DistMult, HolE, RDF2Vec, and BoxE. The service therefore combines translational distance models, semantic matching models, a random-walk-based method, and a geometric box-based model within a single deployment framework (Ahmad et al., 9 Sep 2025).

The translational models are presented through their standard formulations. TransE models a triple $(h, r, t)$ using entity vectors $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$ and a relation vector $\mathbf{r} \in \mathbb{R}^d$, with scoring function

$$f_{\text{TransE}}(h,r,t) = - \left\| \mathbf{h} + \mathbf{r} - \mathbf{t} \right\|_{p}.$$

TransR introduces a relation-specific projection matrix $M_r \in \mathbb{R}^{d_r \times d_e}$, yielding projected representations $\mathbf{h}_r = M_r \mathbf{h}$ and $\mathbf{t}_r = M_r \mathbf{t}$, and uses

$$f_{\text{TransR}}(h,r,t) = - \left\| \mathbf{h}_r + \mathbf{r} - \mathbf{t}_r \right\|_{p}.$$

The stated conceptual role of TransR is to capture different relation semantics in their own embedding subspaces (Ahmad et al., 9 Sep 2025).

Among semantic matching models, DistMult is defined by

$$f_{\text{DistMult}}(h,r,t) = \sum_{i=1}^{d} h_i \, r_i \, t_i,$$

equivalently $$\mathbf{h}^{\top} \operatorname{diag}(\mathbf{r}) \mathbf{t}$$, and is noted as being best suited to symmetric relations because the score is symmetric in $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$0 and $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$1. HolE uses circular correlation

$\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$2

with score

$\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$3

The service description characterizes HolE as providing rich interactions while remaining efficient (Ahmad et al., 9 Sep 2025).

RDF2Vec differs from triple-scoring models. It transforms an RDF graph $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$4 into sequences produced by random walks or Weisfeiler-Lehman subtree kernels, then trains a neural LLM such as Word2Vec. The semantics are captured through co-occurrence in walks rather than an explicit triple scoring function. BoxE represents entities as points in $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$5 and relations as sets of axis-aligned hyperrectangles, allowing representation of logical patterns such as hierarchy and intersections, though the exact scoring formulas are not given in the paper (Ahmad et al., 9 Sep 2025).

Training is standardized across models. TransE, TransR, DistMult, HolE, and BoxE are trained with PyKEEN; RDF2Vec is trained with pyRDF2Vec. All models use the default hyperparameters of the respective libraries except for two explicit settings: the number of epochs is set to 100 and the embedding dimension is set to 200. The paper does not specify exact loss functions or negative sampling schemes, stating instead that these follow the default settings of the underlying libraries (Ahmad et al., 9 Sep 2025).

The contrast with the original KGvec2go is substantial. KGvec2go used only RDF2Vec, trained via random, duplicate-free walks of depth 8 and a skip-gram Word2Vec model with window size $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$6, $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$7 iterations, negative sampling, and $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$8 negative samples, again producing 200-dimensional vectors (Portisch et al., 2020). Bio-KGvec2go generalizes from a single embedding family to a small model suite, enabling direct comparison of distinct representational biases over the same ontology snapshots.

5. System architecture, update mechanism, and access patterns

Bio-KGvec2go combines an automated data pipeline, an embedding training pipeline, a storage and versioning layer, and a web API and UI layer derived from KGvec2go. The data pipeline begins with source registration: GO is obtained from the GO releases site and HP from the HPO GitHub releases. A scheduler periodically invokes a script that downloads the latest ontology file, computes a checksum, and compares it with previously stored checksums. If the checksum differs, the system recognizes a new ontology version, stores the ontology file, derives a version identifier from the release, and registers associated metadata (Ahmad et al., 9 Sep 2025).

When a new version is detected, the embedding training pipeline constructs a graph in which ontology classes are entities and ontological relations are edges. For each model in $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$9TransE, TransR, DistMult, HolE, RDF2Vec, BoxE$\mathbf{r} \in \mathbb{R}^d$0, a training job is executed using PyKEEN or pyRDF2Vec with dimension $\mathbf{r} \in \mathbb{R}^d$1 and $\mathbf{r} \in \mathbb{R}^d$2 epochs. The resulting class embeddings are serialized as JSON files whose keys are ontology class identifiers such as GO IDs or HP IDs and whose values are 200-dimensional floating-point arrays. Trained models are archived in addition to the raw vectors (Ahmad et al., 9 Sep 2025).

The storage layer supports multiple ontology versions simultaneously. The service maintains six versions per ontology in the deployment described in the paper, preserves older versions for reproducibility and ontology evolution studies, and uses the most recent version by default for interactive similarity and nearest-neighbor functions. Embeddings stored on Zenodo are accompanied by metadata following the PROV standard, including the input ontology and version, the KGE model, and hyperparameters such as epochs and dimension (Ahmad et al., 9 Sep 2025).

The web tier is built on the Python/Flask architecture inherited from KGvec2go and can be deployed behind Apache HTTP Server. It exposes a web interface and REST endpoints for three principal functions: downloading embeddings, computing semantic similarity between classes, and retrieving the top closest concepts. The service endpoint is given as https://www.bio.kgvec2go.org, while the source code is available at https://github.com/ritatsousa/biokgvec2go, and trained models and embeddings are archived at https://zenodo.org/records/15865665 (Ahmad et al., 9 Sep 2025).

For similarity queries, two classes are provided by identifier or text label, and the service retrieves vectors from the most recent ontology version and computes cosine similarity: $\mathbf{r} \in \mathbb{R}^d$3 The output is a scalar in $\mathbf{r} \in \mathbb{R}^d$4, where $\mathbf{r} \in \mathbb{R}^d$5 indicates identical direction, $\mathbf{r} \in \mathbb{R}^d$6 orthogonality, and $\mathbf{r} \in \mathbb{R}^d$7 opposite direction. For nearest-neighbor queries, the target vector is compared against all other class vectors, ranked by cosine similarity, and the top 10 are returned together with identifier, label, similarity score, and a direct URL for further exploration (Ahmad et al., 9 Sep 2025).

The dynamic behavior of the system should be interpreted carefully. The update mechanism is periodic and checksum-based, but the paper does not fix a global update frequency. It also does not implement incremental or local-update embedding methods. Every detected ontology change triggers full retraining of all models for that ontology version (Ahmad et al., 9 Sep 2025).

6. Applications, evaluation status, limitations, and future directions

Bio-KGvec2go is positioned as a reusable infrastructure for ontology-aware biomedical machine learning. The paper identifies several downstream uses for GO embeddings: predicting uncharacterized protein functions, predicting protein–protein interactions, and supporting drug target discovery by relating functional annotations. For HP embeddings, the paper cites gene–disease association discovery and disease diagnosis workflows in which a patient’s phenotypic profile is represented in HP space and compared with disease profiles or phenotype sets via similarity (Ahmad et al., 9 Sep 2025).

The service also has direct relevance for ontology curation and semantic annotation. Its similarity and “Top Closest Concepts” functions support annotation support, helping experts identify semantically similar GO or HP terms during labeling. The same facilities can aid ontology development and quality assurance by revealing clusters of highly similar terms that may indicate redundancies, near-duplicates, or missing or misclassified relations. Because multiple ontology versions are preserved, the resource can also be used for comparative analysis of embeddings over time, investigation of changes in embedding-space geometry, and explicit study of concept drift in biomedical knowledge (Ahmad et al., 9 Sep 2025).

Despite these practical use cases, the paper presents Bio-KGvec2go primarily as an infrastructure and resource contribution rather than as a benchmark study. It does not report quantitative evaluation results such as Mean Reciprocal Rank or Hits@$\mathbf{r} \in \mathbb{R}^d$8, does not compare the six supported models on GO or HP in task-specific settings, and does not provide ablation studies. The stated support for embedding quality is indirect, resting on the established use of these model families in the literature and on prior applications of GO and HP embeddings to biomedical tasks (Ahmad et al., 9 Sep 2025).

This lack of direct benchmarking distinguishes Bio-KGvec2go from the original KGvec2go, which evaluated embeddings against semantic similarity datasets using Spearman’s rank correlation and reported that a combination of multiple models outperformed the best individual model on WS-353 and MEN, while WordNet alone slightly outperformed the combined setting on SimLex-999 (Portisch et al., 2020). Bio-KGvec2go does not replicate such comparative evaluation for biomedical ontologies.

The current limitations are explicitly delimited. Support is restricted to GO and HP. The update mechanism relies on full retraining rather than incremental dynamic embedding techniques. Only class-level graph structure is modeled; data properties, complex OWL axioms, and reasoning-derived edges are omitted. Search is limited to identifier- or label-based lookup with case and whitespace normalization, without typo tolerance or advanced query assistance. The paper also notes the absence of task-specific quantitative evaluation (Ahmad et al., 9 Sep 2025).

The authors outline several future directions: extending the service to additional biomedical ontologies and knowledge graphs such as disease, drug, pathway, and multi-species phenotype ontologies; incorporating more embedding models, including newer geometric or GNN-based approaches; improving search with autocomplete and tolerance to minor misspellings or typos; and enhancing dynamic updating mechanisms. A plausible implication is that future versions could move from periodic full regeneration toward more scalable update-aware embedding pipelines, but this is described as a natural next step rather than an implemented feature (Ahmad et al., 9 Sep 2025).