Programmatic design and editing of cis-regulatory elements

The development of modern genome editing and DNA synthesis has enabled researchers to edit DNA sequences with high precision but has left unsolved the problem of designing these edits. We introduce Ledidi, a computational method that rephrases the discrete design task of choosing which edits to make as an easily solvable continuous optimization problem. Ledidi can use any pre-trained deep learning model to guide the optimization, yielding an edited sequence that exhibits the desired outcome while explicitly minimizing the number of edits. When applied in dozens of settings, we find that Ledidi's designs can precisely control transcription factor binding, chromatin accessibility, transcription, and enhancer activity in silico. By using several deep learning models simultaneously, we design cell type-specific enhancers and experimentally validate them in cellulo. Finally, we introduce the concept of an "affinity catalog'', where the design task is repeated multiple times across continuous variants of the design target. We demonstrate how these catalogs can be used to interpret deep learning models and the impact of starting template sequences, and also to design regulatory elements that control transcriptional dosage in a cell type-specific fashion. Overall design: The activity of developmental enhancer in S2 and ovarian somatic cells were measured using the STARR-seq method.

现代基因组编辑（genome editing）与DNA合成（DNA synthesis）技术的发展，使研究人员得以高精度编辑DNA序列，但却遗留了编辑方案设计这一尚未解决的核心难题。本文介绍计算方法Ledidi：其可将“选择编辑位点与类型”这一离散设计任务，重构为易于求解的连续优化问题。Ledidi可依托任意预训练深度学习模型指导优化流程，最终生成既符合预期功能、又能显性最小化编辑次数的编辑后序列。在数十种应用场景中验证后，我们发现Ledidi的设计可在计算机模拟（in silico）环境下，精准调控转录因子结合（transcription factor binding）、染色质可及性（chromatin accessibility）、转录（transcription）过程与增强子活性（enhancer activity）。通过同时调用多个深度学习模型，我们设计了细胞类型特异性增强子，并在细胞内（in cellulo）完成了实验验证。最后，我们提出“亲和力目录（affinity catalog）”的概念：针对设计目标的连续变体多次重复执行设计任务。我们展示了此类目录可用于解读深度学习模型以及起始模板序列的影响，同时还可用于设计以细胞类型特异性方式调控转录剂量的调控元件。总体设计：采用STARR-seq方法，检测S2细胞与卵巢体细胞中发育增强子的活性。