Benchmarking machine learning methods for synthetic lethality prediction in cancer.

Area of Science:

• Computational biology
• Genomics
• Drug discovery

Background:

• Synthetic lethality (SL) offers promising anticancer drug targets by exploiting cancer-specific dependencies.
• Machine learning (ML) methods are increasingly used for SL prediction to complement experimental screening.
• A comprehensive performance evaluation of existing ML-based SL prediction methods is lacking.

Purpose of the Study:

• To systematically benchmark 12 recent ML methods for SL prediction.
• To assess method performance across various data splitting, negative sampling, and task scenarios (classification and ranking).
• To identify limitations and provide guidance for selecting and developing SL prediction techniques.

Main Methods:

• Systematic benchmarking of 12 ML algorithms for SL prediction.
• Evaluation across diverse data splitting strategies and negative sampling techniques.
• Performance assessment on both classification and ranking tasks.

Main Results:

• All evaluated ML methods showed improved performance with enhanced data quality, such as excluding computational SLs and using gene expression for negative sampling.
• The SLMGAE method demonstrated superior performance among the tested algorithms.
• Significant limitations were identified in realistic scenarios, including cold-start independent tests and context-specific SL prediction.

Conclusions:

• Improving data quality is crucial for enhancing the performance of ML-based SL prediction.
• SLMGAE is a promising method, but current ML approaches face challenges in real-world, context-specific applications.
• The study provides valuable insights and resources for selecting and advancing ML techniques in SL virtual screening.