Computational methods for predicting synthetic lethality in cancer

Abstract

Cancer, the second leading cause of death worldwide, is primarily driven by genetic mutations. While targeted therapies offer the advantage of minimizing damage to normal cells by selectively targeting cancerous ones, they are limited by the challenge of directly targeting all oncogenic drivers, many of which are not easily druggable. To overcome these limitations, an alternative strategy has emerged: targeting non-essential genes that exhibit synthetic lethality (SL) with cancer-driving mutations. SL occurs when the simultaneous inactivation of two genes leads to cell death, while inhibition of either gene alone does not. Notably, the therapeutic effect of Olaparib was approved, as it inhibits the PARP family in BRCA1/2-mutated cancers, establishing SL as a viable therapeutic strategy. However, the impracticality of experimentally screening all possible gene pairs in human cancer spurred researchers to adopt statistical and deep learning methodologies to reduce the time and resources required for prediction. While these efforts yielded numerous models, they often suffered from two major limitations: first, they failed to account for the heterogeneous cancer environment (i.e., cell line specificity), compromising result reproducibility; and second, the interpretability offered via attention mechanisms lacked a logical, causal connection to the underlying SL pathogenesis. To overcome these fundamental limitations, this thesis presents an integrated approach. I first propose an SL prediction model utilizing a newly constructed knowledge graph, cancer-specific multiomics data, and a graph-based deep learning architecture to enhance prediction performance and enable cancer-type-specific SL identification. Furthermore, the thesis introduces a novel mechanism inference framework built on a foundation of the largest existing knowledge graph to date. This framework applies GraphRAG combined with causal reasoning Chain-of-Thought (CoT) prompt engineering within a Large Language Model (LLM) to logically infer and articulate SL mechanisms in human-understandable language. Through this comprehensive workflow, I successfully predict clinically applicable SL pairs and, crucially, infer their mechanisms based on causal relationships and biological context. This dual success—prediction and causal mechanism inference—is expected to significantly accelerate the target identification process, transforming the initial stage of drug development.