Deep Learning the cis-Regulatory Code to Predict and Design Plant Gene Expression

Abstract

Elucidating how non-coding cis-regulatory elements (CREs) orchestrate spatiotemporal gene expression remains a fundamental bottleneck in plant biology. This knowledge gap limits our ability to predict the functional consequences of genetic variation and hinders the rational design of beneficial agronomic traits. To bridge the genotype-to-phenotype gap, we leverage explainable artificial intelligence to decode the underlying DNA language of plant regulatory networks. In this presentation, I will introduce a unified deep learning framework encompassing two of our lab's flagship models, deepCIS and deepCRE. deepCIS is a multi-label convolutional neural network trained on extensive epigenomic and transcription factor (TF) binding data. It successfully deciphers the syntactic rules of DNA-protein interactions to enable highly accurate predictions of TF binding sites. Building upon this mechanistic foundation, deepCRE predicts steady-state gene expression profiles directly from cis-regulatory sequences. Validated across diverse model and crop species, including Arabidopsis, tomato, sorghum, and maize, deepCRE accurately captures the effects of natural genetic variants on promoter activity. Beyond in silico predictions, I will highlight the practical applications of these models in translational agriculture. By employing feature attribution methods and virtual mutagenesis, we demonstrate how deepCIS and deepCRE identify causal regulatory variants underlying quantitative traits. Furthermore, this AI-driven approach facilitates the selection of prime regulatory targets for CRISPR/Cas-mediated promoter editing. By virtually screening thousands of promoter modifications, our pipeline enables the rational design of synthetic regulatory elements to fine-tune gene expression, ultimately driving the development of climate-resilient crops.