We introduce two interoperable resources that facilitate semantic data integration and machine learning in biomedical research: the Med2RDF knowledge graph and the Tabulae dataset preparation system. Med2RDF transforms heterogeneous biomedical databases into RDF using a unified ontology and publishes them via the RDF Portal with SPARQL endpoints and FAIR-compliant metadata. It addresses key integration issues such as identiefir heterogeneity and vocabulary inconsistency. Tabulae builds on this semantic infrastructure by enabling users to ecfiiently generate machine learning-ready tabular datasets from complex, RDF-based sources. It abstracts the complexities of querying and integrating diverse information, including compound and protein features, into a unified and readily usable format for machine learning algorithms. We demonstrate the utility of these resources through a case study on compound-protein interaction (CPI) prediction using Random Forest regression. This illustrates how semantic integration combined with machine learning can support ecfiient drug discovery, and highlights Tabulae's unique capability in bridging the gap between rich knowledge graphs and the data formats required by modern machine learning workoflws. Together, Med2RDF and Tabulae provide a scalable and reusable framework for semantic data-driven research in biomedicine.
The landscape of biomedical research is characterized by an ever-growing volume of diverse and heterogeneous datasets, ranging from chemical compounds and protein sequences to
0000-0003-2391-0384 (T. Katayama); 0000-0001-7883-3756 (S. Kawashima); 0000-0001-8195-5893 (Y. Moriya);
0000-0003-1095-8864 (R. Kojima); 0000-0002-9569-8370 (T. Koyama); 0000-0002-2555-7345 (M. Kamada)
© 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
clinical observations and scienticfi literature. While these resources contain highly
complementary information, their fragmentation and semantic inconsistencies pose signicfiant
challenges to comprehensive data analysis, hindering advanced applications like drug discovery
and clinical decision support. Traditional approaches to data integration often fall short due to
disparities in identiefir schemes, property vocabularies, and conceptual modeling. To address
these fundamental challenges, the Semantic Web technologies, particularly the Resource
Description Framework (RDF), ofer a powerful paradigm for expressing and linking such
disparate data into a unified, machine-readable format, thereby enabling sophisticated
crossdatabase queries and inference [
The Med2RDF project is a signicfiant initiative dedicated to constructing a semantically
integrated biomedical knowledge graph by converting diverse life science databases and
documents into RDF. Initial work on Med2RDF has been previously presented, detailing its
conceptual framework and early applications [
Building upon these foundational eforts in semantic data integration, we have developed Tabulae (https://github.com/dbcls/tabulae), a novel system designed to facilitate the generation of integrated datasets from various underlying databases for machine learning applications (Figure 1). Tabulae aims to bridge the gap between complex, semantically integrated biomedical data and the tabular formats oeftn required by machine learning algorithms. While the underlying mechanisms may involve sophisticated data acquisition techniques, including querying multiple databases via SPARQL endpoints or leveraging pre-processed knowledge graphs, Tabulae provides a streamlined approach for users to obtain readily usable datasets. This system aggregates crucial information, such as compound properties (e.g., SMILES, Ro5, AlogP, MW, HBA, HBD) and protein details (e.g., UniProt IDs, sequences) from sources like chembl_protein table. By abstracting the complexities of data integration, Tabulae enables researchers to quickly access and prepare high-quality, pre-integrated data suitable for immediate use in computational experiments, signicfiantly accelerating the initial data preparation phase in various biomedical applications, including drug discovery.
To demonstrate the utility of Tabulae in practical drug discovery workoflws, we present a case study focused on Compound-Protein Interaction (CPI) prediction using a machine learningbased approach. For this analysis, a subset of approximately 10,000 data points with speciefid activity values ("=") was randomly sampled from the Tabulae-acquired chembl_protein table. Given the computational demands of protein language models, the analysis was executed within a T4 GPU runtime environment. The methodology involved a comprehensive feature engineering pipeline to transform raw chemical and biological data into numerical representations suitable for machine learning. Compound features were generated from SMILES strings using Morgan Fingerprints, resulting in a feature dimension of 512. SMILES (Simpliefid Molecular Input Line Entry System) is a textual representation of chemical structures, while Morgan Fingerprints are widely used chemical descriptors that encode structural information of a molecule into a xfied-length binary vector. For protein features, amino acid sequences were converted into embeddings using the ESM2 protein language model (facebook/esm2_t6_8M_UR50D), yielding a feature dimension of 320. ESM2 is a state-of-the-art protein language model developed by Meta AI, capable of generating rich contextual embeddings from protein sequences, capturing complex functional and structural information. These distinct compound and protein feature vectors were then concatenated to form a unified input feature set, resulting in a total feature dimension of 832. This combined feature set was then utilized to construct a regression model aimed at predicting selected concentration response activity values.
A Random Forest Regressor with 100 estimators was employed for model construction. The dataset was split into 80% for training and 20% for testing to ensure robust evaluation of the model's generalization capabilities. The constructed model was evaluated on the unseen 20% test set, with its predictive performance assessed using widely accepted metrics: Mean Squared Error (MSE) and R-squared (R2) score. Furthermore, the predictive accuracy was visually conrfimed through a scatter plot comparing true and predicted activity values, efectively illustrating the model's ability to capture and predict complex CPI relationships. This case study not only highlights the practical utility of Tabulae as a valuable resource for drug discovery research by providing easily accessible, integrated data, but also exempliefis a robust machine learning workflow for predicting CPIs. The successful demonstration of this approach underscores its potential to accelerate the identification of novel drug candidates and to streamline the early stages of the drug discovery process.
In conclusion, the Med2RDF project and the RDF Portal have established a robust framework for the semantic integration and dissemination of heterogeneous life science data, laying the groundwork for a comprehensive biomedical knowledge graph. Complementing these eforts, Tabulae provides a crucial link by enabling the ecfiient extraction and preparation of integrated, machine learning-ready datasets from these rich data resources. The successful application of Tabulae-derived data in the Compound-Protein Interaction prediction task validates its utility in real-world scenarios, showcasing its power in generating insights for drug discovery. We anticipate that this integrated approach, from knowledge graph construction to data preparation for machine learning, will be broadly applicable to numerous life science and medical applications, including target identicfiation, drug repurposing, personalized medicine, and systems biology research, ultimately advancing data-driven decision-making across the biomedical domain. The authors thank Dr. Yoji Shidara of Enishi Tech Inc. for his contributions to the implementation of Tabulae.
During the preparation of this work, the authors used ChatGPT and NotebookLM in order to: Text translation, Grammar and spelling check. Aeftr using these services, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.