ELISL: Early-Late Integrated Synthetic Lethality Prediction in Cancer

Abstract

Anti-cancer therapies based on synthetic lethality (SL) exploit tumor vulnerabilities for treatment with reduced side effects. Since simultaneous loss-of-function of SL genes causes cell death, tumors with known gene disruptions can be treated by targeting SL partners. Computational selection of promising SL candidates amongst all gene combinations is key to expedite experimental screening. However, current SL prediction models: (i) only use tissue type-specific molecular data, which can be scarce/noisy, limiting performance for some cancers; and (ii) often rely on shared SL patterns across genes, showing sensitivity to prevalent gene selection bias. We propose ELISL, Early-Late Integrated models for SL prediction using forest ensembles. ELISL models ignore shared SL patterns, and integrate context-specific data from cancer cell lines or tumor tissue with context-free functional associations derived from protein sequence. ELISL outperformed existing methods and was more robust to selection bias in 8 cancer types, with prominent contribution from sequence. We found better survival for patients whose tumors carried simultaneous mutations in a BRCA gene together with an ELISL-predicted SL gene from the HH, FGF, or WNT families. ELISL thus arises as a promising strategy to discover SL interactions with therapeutic potential.