Increased appreciation of 5-hydroxymethylcytosine (5hmC) as a stable epigenetic mark, which defines cell identity and disease progress, has engendered a need for cost-effective, but high-resolution, 5hmC mapping technology. Current enrichment-based technologies provide cheap but low-resolution and relative enrichment of 5hmC levels, while single-base resolution methods can be prohibitively expensive to scale up to large experiments. To address this problem, we developed a deep learning–based method, “DeepH&M,” which integrates enrichment and restriction enzyme sequencing methods to simultaneously estimate absolute hydroxymethylation and methylation levels at single-CpG resolution. Using 7-week-old mouse cerebellum data for training the DeepH&M model, we demonstrated that the 5hmC and 5mC levels predicted by DeepH&M were in high concordance with whole-genome bisulfite–based approaches. The DeepH&M model can be applied to 7-week-old frontal cortex and 79-week-old cerebellum, revealing the robust generalizability of this method to other tissues from various biological time points.

5-羟甲基胞嘧啶 (5hmC) 作为一种稳定的表观遗传标记，它定义了细胞身份和疾病进展，越来越受到重视，从而产生了对具有成本效益但高分辨率的 5hmC 映射技术的需求。当前基于富集的技术提供了廉价但低分辨率和相对富集的 5hmC 水平，而单碱基分辨率方法在扩大到大型实验时可能过于昂贵。为了解决这个问题，我们开发了一种基于深度学习的方法“DeepH&M”，它集成了富集和限制性内切酶测序方法，以同时估计单 CpG 分辨率下的绝对羟甲基化和甲基化水平。使用 7 周龄小鼠小脑数据训练 DeepH&M 模型，我们证明了 DeepH& 预测的 5hmC 和 5mC 水平 M 与基于全基因组亚硫酸氢盐的方法高度一致。DeepH&M 模型可应用于 7 周龄的额叶皮层和 79 周龄的小脑，揭示了该方法对来自不同生物时间点的其他组织的强大普遍性。