Systematic prioritisation of context-specific paralog pair vulnerabilities in cancer

Abstract Background Genome-wide CRISPR screening has enabled the development of dependency maps in hundreds of cancer cell lines, facilitating the identification of genetic vulnerabilities associated with specific biomarkers and supporting precision oncology efforts. However, current dependency maps largely capture single-gene effects and systematically miss vulnerabilities where functional compensation between paralogs masks the underlying dependency. Combinatorial screens have revealed that paralog pairs are often synthetic lethal but that these effects are highly context-dependent across tumour types and genetic backgrounds. To enable the clinical translation of paralog synthetic lethality, it is therefore necessary to identify which paralog pairs constitute actionable vulnerabilities in specific cancer contexts. Methods We developed a machine learning classifier to predict cell-line-specific synthetic lethality between paralog pairs using features derived from transcriptomics, genomics, gene essentiality, and network context. We designed an evaluation framework spanning three biologically relevant scenarios: predicting synthetic lethality for seen paralog pairs in unseen cell lines, for unseen pairs in seen cell lines, and for unseen pairs in unseen cell lines. We applied the model to 33,419 pairs across 1,005 cancer cell lines and validated predictions using independent combinatorial CRISPR screens. Results We found that the cell-line-specific expression and essentiality of paralogs and their interaction partners are informative for predicting synthetic lethal interactions. Our model generalised to unseen paralog pairs and to unseen cell lines, though the combination of both was the most challenging scenario. The agreement between predicted and experimentally observed interactions was comparable to that between independent experimental studies, suggesting that predictive accuracy may be constrained as much by experimental reproducibility as by the model itself. When applied to HER2-amplified breast cancer, the model recovered known synergies and predicted novel biomarker-associated vulnerabilities. Conclusions We present a comprehensive, context-resolved resource of paralog synthetic lethal vulnerabilities across 1,005 cancer cell lines. This resource enables systematic, disease-stratified prioritisation of paralog-pair dependencies. Genome-scale predictions are made available through an interactive web portal ( https://cancergenetics.github.io/paralogmap/ ) to support hypothesis generation, guide targeted combinatorial screening, and facilitate the identification of clinically actionable paralog targets.