ESM4SL: Protein Language Models for Cancer Cell Line-specific Synthetic Lethality Prediction

Synthetic lethality (SL) refers to the relationship between two genes, where cell death occurs only when both genes are simultaneously perturbed. Identifying SL gene pairs has significant implications for targeted cancer therapy, but the number of identified SLs remains limited in most cell lines. Existing machine learning approaches struggle with challenges including data scarcity and weak generalization challenges when predicting SLs in specific cell lines. In this work, we present ESM4SL, a novel approach leveraging the pre-trained protein language model ESM-2 which incorporates rich protein sequence and evolutionary information, to alleviate the lack of generalization ability in predicting cell line-specific SL. Our method passes protein sequences through ESM-2 to get their representations, and integrates these representations using self-attention and cross-attention mechanisms for final prediction. Experimental results in multiple scenarios and various cell lines demonstrate that our method achieves state-of-the-art performance, making it an effective tool for SL prediction. To our best knowledge, this is the first time a protein language model is used as the core of a model for SL prediction. Our code is available at https://github.com/JieZheng-ShanghaiTech/ESM4SL.