Abstract 897: Synthetic lethality screening with Recursive Feature Machines

Abstract

Abstract Synthetic lethality refers to the concept that simultaneous perturbation of gene pairs leads to cell death but individual perturbation does not. Synthetically lethal gene pairs (SL pairs) provide a potential avenue for selective targeting of cancer cells based on underlying genetic vulnerabilities. The rise of large scale gene perturbation screens such as the Cancer Dependency Map (DepMap) offers the opportunity to identify SL pairs automatically using machine learning. Yet, a key difficulty in using machine learning models for this task is the lack of labeled data in the form of known SL pairs. Thus, prior approaches framed SL pair identification as a feature learning problem where the goal is to identify the genomic features most influential for predicting low cellular viability under a gene knockout. These prior approaches have primarily utilized random forests since these machine learning models are one of the only nonlinear models for which feature importances are provided explicitly. On the other hand, if we could identify features learned by state-of-the-art models on screening tasks such as kernel machines, we would be better powered in finding SL pairs. In this work, we present a computationally efficient and effective pipeline for SL pair screening built on a recently developed class of feature learning kernel machines known as Recursive Feature Machines (RFMs). We show that our pipeline more accurately recovers experimentally verified SL pairs than prior work including the best model from the DepMap portal and previous random forest based approaches (see the table below). Moreover, our pipeline identifies several new candidate SL pairs for further analysis, opening new avenues for targeting genetic vulnerabilities in cancer. Rank of verified SL pairs across unsupervised methods (lower is better with a minimum value of 1). Experimentally Verified SL Pairs Pearson Correlation PARIS DepMap RFMs (Ours) SMARCA2/SMARCA4 1 1 1 1 ARID1A/ARID1B 1 1 1 1 STAG1/STAG2 1 1 1 1 CREBBP/EP300 1 1 1 1 VPS4A/VPS4B 7 > 10 5 1 DDX17/DDX5 4 1 1 1 ENO1/ENO2 > 10 1 1 1 SMARCC1/SMARCC2 > 10 1 1 1 UBB/UBC > 10 1 > 10 1 MAGOH/MAGOHB 1 1 1 1 FAM50A/FAM50B 1 1 1 1 Citation Format: Adityanarayanan Radhakrishnan, Cathy Cai, Barbara A. Weir, Christopher Moy, Caroline Uhler. Synthetic lethality screening with Recursive Feature Machines [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 897.