Cell type directed design of synthetic enhancers

Abstract

Transcriptional enhancers act as docking stations for combinations of transcription factors and thereby regulate spatiotemporal activation of their target genes. A single enhancer, of a few hundred base pairs in length, can autonomously and independently of its location and orientation drive cell-type specific expression of a gene or transgene. It has been a long-standing goal in the field to decode the regulatory logic of an enhancer and to understand the details of how spatiotemporal gene expression is encoded in an enhancer sequence. Recently, deep learning models have yielded unprecedented insight into the enhancer code, and well-trained models are reaching a level of understanding that may be close to complete. As a consequence, we hypothesized that deep learning models can be used to guide the directed design of synthetic, cell type specific enhancers, and that this process would allow for a detailed tracing of all enhancer features at nucleotide-level resolution. Here we implemented and compared three different design strategies, each built on a deep learning model: (1) directed sequence evolution; (2) directed iterative motif implanting; and (3) generative design. We evaluated the function of fully synthetic enhancers to specifically target Kenyon cells in the fruit fly brain using transgenic animals. We then exploited this concept further by creating "dual-code" enhancers that target two cell types, and minimal enhancers smaller than 50 base pairs that are fully functional. By examining the trajectories followed during state space searches towards functional enhancers, we could accurately define the enhancer code as the optimal strength, combination, and relative distance of TF activator motifs, and the absence of TF repressor motifs. Finally, we applied the same three strategies to successfully design human enhancers. In conclusion, enhancer design guided by deep learning leads to better understanding of how enhancers work and shows that their code can be exploited to manipulate cell states.

Transfer learning identifies sequence determinants of regulatory element accessibility

“…We further acknowledge that further work on model interpretation is needed to arrive at a sequence code for regulatory activity. Recent developments to this end (Shrikumar et al 2019; Avsec et al 2021b; de Almeida et al 2022; Taskiran et al 2022) show great promise, and we expect that integration of such analyses with the transfer learning scheme of ChromTransfer will be important for future efforts to understand the regulatory code.…”

EUGENe: A Python toolkit for predictive analyses of regulatory sequences

“…We compared this to predicted scores generated with the same feature implantation approach using a dinucleotide shuffled version of the 16bp sequence containing the TATA box motif, a random 16bp one-hot encoded sequence, and a 16bp all zeros input. We performed the in silico evolution experiments on the same set of 310 promoter sequences 31,62 . In each round, we first used in silico saturation mutagenesis to identify the mutation that increased the model score by the largest positive value (delta score).…”

“…However, due to the inherent flexibility of PL, more advanced users can customize almost all aspects of their model training strategy. For instance, custom training loops can be defined for models that implement multipart loss functions or use multiple optimizers [60][61][62] (e.g. for variational autoencoders and generative adversarial networks respectively).…”

Multiplex profiling of developmental enhancers with quantitative, single-cell expression reporters

“…Reporter assays have been applied to mammalian differentiation models (e.g., neuronal (21), naive to epiblast (22)), but these remain essentially simple trajectories. Single-cell chromatin accessibility data from systems containing extensive cell type heterogeneity can be used to train models predicting differential accessibility from DNA sequence (23)(24)(25), with promise to also correlatively predict cell-typespecific expression (26). However, these models remain one step removed from the functional outcome and are inherently limited given that differentially accessible genomic regions commonly lack autonomous expression-enhancing activity (27).…”