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

Synthetic lethality (SL) is a promising strategy for targeted cancer therapy, as inhibiting SL partners of genes harboring cancer-specific mutations can selectively eliminate cancer cells while sparing normal cells. Many SL prediction methods are based on supervised learning and have demonstrated strong predictive capability. However, these methods tend to underperform when trained on limited labeled data. Furthermore, existing SL prediction models are typically trained and tested under idealistically balanced class distributions. However, when these well-trained models are applied to real-world scenarios, where severe class imbalance is common and label distribution shifts occur, they often struggle to make accurate predictions. To address these issues, we propose MVCL4SL, a novel synthetic lethality prediction approach which integrates a multi-view neural network with a multi-view contrastive learning strategy. The multi-view features can describe genes comprehensively, and we propose a multi-view contrastive learning strategy to fuse the features, thereby strengthening the model's robustness and generalization capabilities. Experimental results indicate that MVCL4SL significantly outperforms six baselines across various cell lines, particularly in low-data training scenarios. MVCL4SL also showcases its robustness to a significant label distribution shift from training data. Our code is available at https://github.com/JieZheng-ShanghaiTech/MVCL4SL.