DeepChrom: A Diffusion-Based Framework for Long-Tailed Chromatin State Prediction

Chromatin state reflects distinct biological roles of the genome that can systematically characterize regulatory elements and their functional interaction. Despite extensive computational studies, accurate prediction of chromatin state remains a challenge because of the long-tailed class imbalance. Here, we propose a deep-learning framework, DeepChrom, to predict long-tailed chromatin state directly from DNA sequence. The framework includes a diffusion-based model that balances the samples of different classes by generating pseudo-samples and a novel dilated CNN-based model for chromatin state prediction. On top of that, we further develop a novel equalization loss to increase the penalty on generated samples, which alleviates the impact of the bias between ground truth and generated samples. DeepChrom achieves outstanding performance on nine human cell types with our designed paradigm. Specifically, our proposed long-tailed learning strategy surpasses the traditional training method by 0.056 in Acc. To our knowledge, DeepChrom is pioneering in predicting long-tailed chromatin states by the diffusion-based model to achieve sample balance.