Harnessing computational intelligence for synthetic lethality: A roadmap from network biology to interpretable deep...

Area of Science:

• Oncology
• Computational Biology
• Genetics

Background:

• Synthetic lethality (SL) offers targeted cancer therapy by exploiting genetic vulnerabilities.
• Experimental screening for SL is challenging due to the vast number of gene interactions.
• Computational approaches are crucial for efficient SL discovery.

Purpose of the Study:

• To provide a comparative synthesis of computational SL discovery architectures.
• To establish evidence-based best practices for evaluating SL prediction models.
• To discuss synthetic rescue and its role in drug resistance for clinical translation.

Main Methods:

• Review and synthesis of computational SL discovery methods, including network-based models, graph transformers, and knowledge graph reasoning.
• Analysis of methodological evaluation strategies: data splitting, metric selection, and negative sampling.
• Integration of multi-omics data and interpretable deep learning for predicting SL candidates.

Main Results:

• A comparative overview of diverse computational SL discovery architectures is presented.
• Best practices for rigorous methodological evaluation are provided, addressing class imbalance and negative sampling.
• A framework for prioritizing SL candidates robust to resistance mechanisms is proposed.

Conclusions:

• Computational methods are essential for navigating the complexity of synthetic lethality discovery in precision oncology.
• Standardized evaluation and consideration of resistance mechanisms are key for clinical translation of SL therapies.
• This review offers a roadmap to bridge computational predictions with actionable clinical insights.