Research Projects

Synthetic lethality (SL) occurs when the simultaneous disruption of two genes leads to cell death, while inactivation of either gene alone is non-lethal (O'Neil et al. 2017). The principle of SL, wherein drug therapy can be used to inhibit the activity of a gene mutated in cancer cells, has been proposed as a means to develop novel precision cancer therapies for more than 25 years (Hartwell et al. 1997). However, till date, only PARP1-inhibitor therapies in BRCA1/2-mutated cancers have been demonstrated to be clinically successful (Lord and Ashworth 2017; Pujade-Lauraine et al. 2023; Färkkilä et al. 2020; Carreira et al. 2021), highlighting the need for exploration of SL-based precision medicine in other cancer types. Experimental determination of SL involves drug screening (Chan and Giaccia 2011; Tang et al. 2022) , RNAi screening (Luo et al. 2009) and CRISPR/Cas9 screenings (McDonald et al. 2017). However, given the large number of possible pairwise gene combinations (~200 million in human cells) among different genetic contexts, it is infeasible to screen all potential SL gene pairs experimentally. As a result, computational approaches have emerged as an alternative to predict SL gene pairs. Computational models capable of predicting SL can be divided into four categories: statistics-based methods, network-based methods, classical machine learning-based methods and deep learning-based methods (Wang et al. 2022). Statistical models and network based approaches rely on significance indicators from the analysis of diverse and heterogeneous data to predict SL. This forms a severe limitation in instances when experimental data is unavailable in specific cancer cell lines. Classical machine learning (ML) methods, such as the random forest based approach demonstrated by De Kegel et. al. (De Kegel et al. 2021) and deep learning methods such as the semi supervised neural network based EXP2SL (Wan et al. 2020), have also been used to make SL predictions. Recent advances in deep learning (DL) methods have improved our ability to decipher the complex relationships between inputs and outputs in DL models, encouraging the development of DL-based SL prediction algorithms. While these methods have improved upon traditional SL prediction tools, these limitations in addition to poor interpretability and inability to process genes with unseen features (Wang et al. 2022) leaves room for improvement in the SL prediction domain. Additionally, although many SL prediction algorithms exist, till date only MVGCN-iSL (Fan et al. 2022) and ELISL (Tepeli et al. 2024) have been specifically designed to account for cell line specificity. Despite this, both the models were outperformed by SLMGAE (in 293T and OVCAR8 cells) and NSF4SL (HeLa cells) in a recent benchmarking study (Feng et al. 2024). The same benchmarking study showed that all seven models evaluated in the study showed significantly poor performance in SL ‘Ranking’. In this project, we propose to develop SLxGO, a gradient boosting based algorithm to improve cell line specific SL prediction with particular focus on ranking ability. The project also aims at validating SL predictions using CRISPR/Cas9 screen data, knockouts, knockdown and pharmacological inactivation experiments to explore the precision medicine potential of the findings.