Genome-scale prediction of context-specific synthetic lethality beyond protein interaction networks

Abstract

Identifying synthetic lethal (SL) interactions offers a principled framework for discovering disease-specific therapeutic targets. However, current machine learning approaches heavily rely on curated protein-protein interaction networks. Because these networks cover only [~]7,500 proteins, they severely restrict the search space of human gene pairs and introduce systematic biases toward well-characterized genes. To circumvent these limitations, we developed SLxGO, a network-independent machine learning framework that predicts SL interactions directly from semantic representations of Gene Ontology annotations encoded via BioBERT-derived embeddings. Benchmarked across multiple cross-validation schemes against eight state-of-the-art methods, SLxGO consistently achieved superior predictive ranking performance, maintaining robustness under cold-start conditions for previously unseen genes. Integrating cell line-specific transcriptional profiles extended this framework to context-dependent SL prediction across six distinct cell lines. Notably, we experimentally confirmed a context-specific EFNA1-SLC29A1 SL interaction in HeLa cells, alongside synergistic pharmacological validation of an ACVR1-SLC29A1 vulnerability. All predictions are hosted on SLiGO, an open-access database encompassing 30 million human gene pairs, establishing a comprehensive, genome-scale resource for context-specific vulnerability mapping across the human interactome.