Accurate interaction detection in combinatorial screens with complex phenotypes

Project Summary Combinatorial screens characterize the molecular pathways of the cell by identifying interactions between perturbations. Quantifying interactions requires a notion of non-interaction (i.e., a null interaction model), which is well-established for combinatorial screens measuring cell survival and other simple phenotypes. However,

recent advances in single-cell omics and automated microscopy have enabled combinatorial screens with high- dimensional phenotypic profiles. These phenotypic profiles consist of features with unknown, non-linear behavior, such as gene expression or image features. As a result, there is no well-defined null interaction model

for combinatorial screens with high-dimensional phenotypes. The goal of my proposal is to establish a null interaction model for high-dimensional phenotypes. My approach, termed DENIM (Data-driven Estimation of Null Interaction Models), aims to use robust estimation models to learn an additive representation for null perturbations and improve quantification of interactions across combinatorial

screen datasets. In addition, this framework has a key application in drug target identification by enabling accurate prioritization of chemical-genetic interactions from microscopy-based phenotypic profiles. My central hypothesis is that null interaction models which accurately model the diverse behavior of high-

dimensional features can substantially improve interaction detection. In Aim 1, I will develop and critically assess DENIM using both simulated and real-world combinatorial datasets with gene expression or high-content imaging readouts. I will evaluate DENIM against existing baselines on the ability to accurately interaction strength in

simulation, as well as recapitulate known interactions between gene pairs or drug pairs. Preliminary data from simulation demonstrates that DENIM can accurately identify and learn non-linear perturbation trajectories, which leads to improvement in performance relative to the standard baseline (i.e., vector addition). In Aim 2, I will

leverage the improved prioritization of interactions provided by DENIM to perform drug target identification. This target identification platform requires large-scale genetic perturbations and the study of chemical-genetic interactions. To this end, preliminary results show that large-scale genetic perturbations can be accurately

conducted using pooled genetic screening followed by in situ sequencing for guide demultiplexing. Moreover, I analyzed a pilot screen on a well-established MEK1/2 inhibitor, where I accurately recovered target pathway components, with similar performance using drug-relevant or drug-agnostic biomarkers. Together, completion of

this proposal will result in a rigorous framework for high-dimensional interactions, benchmarking of null interaction models, and a novel platform for chemical-genetic target identification using microscopy-based phenotypic profiles.