Multi-view Contrastive Learning for Cell Line-specific Synthetic Lethality Prediction.

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Synthetic lethality (SL) offers targeted cancer therapy by exploiting gene dependencies.
• Supervised learning models for SL prediction face challenges with limited data and class imbalance.
• Real-world applications of SL prediction models are hindered by data limitations and distribution shifts.

Purpose of the Study:

• To develop a robust and generalizable synthetic lethality prediction method.
• To address the limitations of existing SL prediction models in low-data and imbalanced scenarios.
• To improve the accuracy and reliability of identifying synthetic lethal gene pairs.

Main Methods:

• Proposed MVCL4SL, a novel approach integrating multi-view neural networks and multi-view contrastive learning.
• Utilized multi-view features for comprehensive gene description and feature fusion.
• Employed contrastive learning to enhance model robustness and generalization.

Main Results:

• MVCL4SL significantly outperformed six baseline methods across various cell lines.
• Demonstrated superior performance in low-data training scenarios.
• Showcased robustness against significant label distribution shifts from training data.

Conclusions:

• MVCL4SL provides a powerful and reliable method for synthetic lethality prediction.
• The multi-view approach enhances model performance under challenging data conditions.
• This method holds promise for advancing targeted cancer therapy development.