Knowledge graph embeddings (KGEs) have emerged as a powerful paradigm for learning
latent representations of biomedical entities and their relationships, enabling tasks such as
drug target prediction and drug-target interaction (DTI) discovery. However, systematic
comparisons across model architectures and embedding dimensionalities on real-world
biomedical graphs remain limited, particularly under rigorous and fair evaluation
protocols. In this study, we present a comprehensive comparative analysis of three
prominent KGE models — TransE, ComplEx, and TriModel — applied to a
heterogeneous biomedical knowledge graph derived from DrugBank, encompassing drugs,
proteins, and their relational associations. All models are trained and evaluated across
three embedding dimensions (100, 200, and 300) under two complementary evaluation
tasks: drug target prediction, assessed with Mean Reciprocal Rank (MRR) and Hits@K,
and DTI binary classification, assessed with AUC-ROC, AUC-PR, and F1-score. To
ensure a fair DTI evaluation, we introduce a degree-matched negative sampling strategy
that generates hard negatives by pairing each drug with proteins of similar training-set
connectivity, mitigating the trivial-rejection bias inherent in uniform random sampling.
Our results show that TriModel consistently achieves the highest drug target prediction
performance (MRR = 0.609, Hits@10 = 0.719 at dimension 300), while ComplEx leads in
DTI classification (AUC-ROC = 0.799 at dimension 300), suggesting that complex-valued
representations capture interaction-relevant structure more effectively. TransE lags behind
both models on all metrics and dimensions. Across all models, performance improves
monotonically with embedding dimension, indicating that higher-dimensional spaces
provide meaningful representational benefits within the tested range. These findings offer
actionable guidance for selecting KGE architectures in biomedical applications and
highlight the importance of task-specific evaluation beyond standard link prediction
benchmarks.