Cross-Species Prediction of Histone Modifications in Plants via Deep Learning

The regulation of gene expression in plants is governed by complex interactions between cis-regulatory elements and epigenetic modifications such as histone marks. While deep learning models have achieved success in predicting regulatory features from DNA sequence, their cross-species generalizability in plants remains largely unexplored. In this study, we systematically evaluated the ability of deep learning models to predict histone modifications across plant species using a multi-stage framework based on the Sei architecture. We trained species-specific models for Arabidopsis (A. thaliana), rice (O. sativa), and maize (Z. mays), achieving high within-species predictive performance (AUROC > 0.94) and strong agreement between predictions and experimental ChIP-seq profiles. However, cross-species predictions showed reduced performance with increasing phylogenetic distance, highlighting limited model transferability between monocots and dicots. To improve generalization, we built family-level models by training on multiple species across the Poaceae and Brassicaceae families. These models demonstrated robust predictive power in completely unprofiled species--those entirely unused in training set--highlighting the models adaptability to novel plant genomes based solely on conserved regulatory syntax. In contrast, cross-family models produced inconsistent results, with reliable performance only in species sharing conserved regulatory features. Lastly, we developed an easy-to-use pipeline that predicts genome-wide chromatin signals directly from DNA sequences. Our findings demonstrate that phylogenetically informed model training significantly improves cross-species epigenomic prediction, offering a scalable computational strategy for functional annotation in non-model and agriculturally important plants.