Can deep learning predict inducibility of stress-responsive gene expression from DNA sequence in Arabidopsis thaliana?

Abstract

Noncoding regions mediate transcriptional adaptation to stress in plants, yet the genomic determinants driving gene expression responses remain poorly understood. Here, we evaluate the ability of two deep learning approaches, a convolutional neural network (CNN) and a transformer-based genomic language model pre-trained on plant genomes to predict changes in gene expression under various abiotic and biotic stress conditions. Using RNA-seq time-series data from Arabidopsis thaliana, we explored different strategies for summarizing expression dynamics to capture treatment-specific transcriptional changes relative to control conditions. Both models achieved low to moderate predictive performance, with pattern-triggered-immunity related treatments showing the strongest sequence-based predictability. While extending promoter regions upstream had a limited impact, including coding sequences significantly improved performance. Model interpretation revealed that the CNN recovered sequence features comparable to those identified by simple 6-mer based linear models, suggesting limited gains in regulatory insight from increased model complexity. These findings underscore both the promise and limitations of sequence-based models in uncovering the regulatory logic of induced plant stress responses.