BETA: a comprehensive benchmark for computational drug-target prediction.

Area of Science:

• Biomedical Informatics
• Computational Drug Discovery
• Network Biology

Background:

• Internal validation is standard for drug-target predictive models but can be biased with large, diverse datasets.
• Current evaluation methods may not capture model performance across various use-cases, including different data types and network structures.
• This limits the comprehensive assessment of predictive models for drug repurposing and target discovery.

Purpose of the Study:

• To introduce BETA, a novel benchmark for evaluating drug-target predictive models.
• To provide a large-scale, multipartite network and diverse evaluation strategies reflecting real-world scenarios.
• To enable a more robust assessment of computational methods in drug discovery.

Main Methods:

• Constructed BETA, a benchmark featuring a large multipartite network (0.97M concepts, 8.5M associations) and extensive similarities (62M drug-drug, protein-protein).
• Developed seven distinct evaluation strategies (344 Tasks) covering general, screening, category-based, and repurposing scenarios.
• Tested six state-of-the-art methods using chemical structure, gene sequence, and network-based inputs across all tasks.

Main Results:

• Analyzed best- and worst-performing cases to identify method limitations within the benchmark tasks.
• Demonstrated BETA's capability to differentiate the performance of various computational strategies.
• Highlighted specific methods' strengths and weaknesses across diverse drug discovery and repurposing tasks.

Conclusions:

• BETA serves as a crucial benchmark for selecting effective computational strategies in drug repurposing and target discovery.
• The benchmark provides a more realistic and comprehensive evaluation framework compared to traditional methods.
• Results from BETA can guide the development and optimization of predictive models for biomedical applications.