Predicting Genetic Interactions Using Functional Interaction Networks

Iulia Veronica Gheorghe, MS (Advisor: Traver Hart, PhD)

         Mapping genetic interactions is central to understanding cellular systems and identifying therapeutic vulnerabilities, particularly in the context of cancer. Among these interactions, synthetic lethality, where simultaneous loss of two genes is lethal but loss of either alone is tolerated, offers a powerful framework for selectively targeting tumor-specific dependencies. In model organisms like S. cerevisiae, comprehensive double-knockout screens have revealed detailed genetic interaction maps, enabling systems-level insights into pathway structure, gene function, and cellular organization. Replicating this achievement in human cells, however, is complicated by the scale and complexity of the human genome. Recent advances in genome-wide CRISPR knockout screening have made it possible to systematically measure gene essentiality across hundreds of cancer cell lines, laying the foundation for coessentiality networks which are functional maps derived from correlated fitness profiles. Yet, existing coessentiality networks are often limited by static representations, lack of contextual resolution, and difficulty capturing context-dependent or multifunctional gene roles. This dissertation presents a framework for constructing and leveraging functional interaction networks to guide genetic interaction discovery.

          In the first phase of this dissertation, I systematically benchmarked data preprocessing strategies for generating coessentiality networks from genome-scale CRISPR knockout data. By evaluating combinations of essentiality scoring, normalization, and similarity measures, I identified an optimal pipeline that improves the biological accuracy of functional networks. Covariance normalization via whitening proved critical, and I showed that Pearson correlation on whitened data performs equivalently to ordinary least squares. This substantially improved functional enrichment and interpretability of network edges.

          Building on this optimized foundation, I developed a hierarchical clustering method based on partial correlation that captures both direct and conditional dependencies, allowing genes to belong to multiple modules and uncovering pleiotropic or context-specific roles. This approach captures both within- and between-module relationships, offering a scalable and interpretable representation of cellular organization. Benchmarking against alternative clustering methods demonstrated superior functional coherence, particularly at module sizes relevant for experimental screening.

          Finally, I applied insights from a high-resolution yeast genetic interaction network to guide the prediction of synthetic lethality-enriched modules in the human network. Features such as strong functional enrichment and module size proved most predictive of synthetic lethality. The resulting predictions focus on clinically relevant processes such as receptor tyrosine kinase signaling and DNA damage response pathways and were prioritized for screening using a multiplex CRISPR platform.

          Collectively, this work brings together strategies in data processing, network modeling, and synthetic lethality prediction to address challenges in mapping genetic interactions. By combining insights from model organisms with functional network analysis, it contributes both conceptual and practical advances to the field of functional genomics. This work provides a foundation for exploring gene function at scale and for informing the next generation of experiments aimed at understanding and uncovering therapeutic vulnerabilities.

Advisory Committee:

• Traver Hart, PhD, Chair
• Swathi Arur, PhD
• Jeffrey Chang, PhD
• Paul Scheet, PhD
• John Paul Shen, MD