Predicting transcriptional burst kinetics from histone modifications with deep learning

Transcription is inherently bursty, and burst frequency and burst size provide mechanistically informative phenotypes that are not uniquely determined by mean expression alone. Yet it remains difficult to quantify, at genome scale, how histone-modification landscapes encode these burst kinetics. Here we show that DeepBurst, a deep learning model, predicts gene-specific burst-frequency and burst-size regimes from histone-modification profiles around transcription start sites. DeepBurst uses self-attention to learn combinatorial, position-dependent predictive patterns, with training labels inferred from single-cell transcriptome data using mechanistic bursting models. Across three human cell lines and two label-inference procedures, DeepBurst achieves consistent performance under chromosome-based splits and shows strong cross-cell type generalization. Model interpretation shows that burst frequency benefits from broader chromatin context, whereas burst size depends more strongly on promoter-proximal signals. Model-constrained in silico perturbations further identify histone-profile changes predicted to shift bursting regimes. DeepBurst therefore enables genome-wide mapping from readily measurable histone-modification landscapes to hard-to-assay transcriptional bursting dynamics, while generating testable hypotheses for future perturbation studies.