Project Summary Cis-regulatory elements (CREs) are crucial components of transcriptional regulation and the rapid growth of genomic data has enabled researchers to annotate CREs across many biological contexts. However, despite the comprehensiveness of these collections, understanding the rules dictating how CREs regulate genes remains a major unresolved problem in genomics. Therefore, to better understand gene regulation, we are proposing to develop a new framework where CRE-gene interactions are modeled as graphs. This will enable researchers to accomplish a wide range of computational tasks such as comparisons between cell types, predictions of new interactions, and predictions of gene expression. Specifically, this pilot project aims to evaluate the feasibility and generalizability of a CRE-interaction graph approach for predicting gene expression. We will build CRE-interaction graphs in three biological contexts using public datasets, including those generated by Common Fund projects, by integrating genomic interaction data, such as CRISPR perturbations and Hi-C loops, with annotated CREs. Then to demonstrate the utility of these graph models, we will use graph neural networks to predict gene expression, testing different algorithms and gene expression qualifications to maximize model performance. Finally, we will use feature attribution methods and prediction explainer algorithms to interpret our models to gain a better understanding of the mechanisms regulating transcription. The project will not only lead to a better model for predicting gene expression, but also establish a flexible framework for future research on gene regulation. The project will also produce a resource for the computational and machine learning community and improve the utility of existing resources.

Project Narrative Cis-regulatory elements (CREs) play a key role in transcriptional regulation. While the growing availability of genomic data has facilitated the annotation of CREs, the rules and mechanisms which govern how CREs regulate genes is still not fully understood. Here we propose a new graph-based framework for modeling CRE-gene interactions, which will allow for improved understanding of gene regulation and the creation of new resources for the computational and machine learning community.