Genomic language models improve cross-species gene expression prediction and accurately capture regulatory variant effects in Brachypodium mutant lines

Predicting gene expression from cis-regulatory DNA sequences at the promoter and terminator regions is a central challenge in plant genomics. This capability is also a prerequisite for assessing the effects of regulatory mutations on gene expression. Here, we developed deep learning sequence-to-expression (S2E) models that leverage context-aware sequence embeddings from the PlantCaduceus genomic language model instead of one-hot encoding of sequences, to predict gene expression across 17 plant species. To further improve predictions, we integrated chromatin accessibility data as auxiliary regulatory features. First, we evaluated our models to predict gene expression on unseen gene families via cross-validation, demonstrating our models prediction accuracy across all species outperforms PhytoExpr, the current state-of-the-art (SOTA) S2E model in plants (Pearson R=0.82 vs. R=0.74). We then validated variant effect predictions using an experimental dataset across 796 Brachypodium mutant lines, specifically designed to test predictions at single-base resolution. Our models outperformed SOTA S2E models in predicting between-gene expression differences (regression coefficient {beta}=0.78 vs. {beta}=0.57). Remarkably, they also accurately predicted the effects of single-nucleotide mutations on within-gene expression, while SOTA S2E models showed only weak associations (regression coefficient {beta}=0.38 vs. {beta}=0.08). Our results demonstrated the value of context-aware DNA sequence embeddings for predicting regulatory variant effects in plants. They also reveal a persistent accuracy gap in S2E models when moving from between-gene to allelic variation, a challenge that needs to be addressed in future S2E studies.