Research

(Khyati embryo image) Developing embryos need to regulate genes at the right time and place to form the intricate structures of the body plan. These gene cis-regulatory instructions are encoded in DNA sequences called enhancers. They can be measured by a variety of assays, yet we cannot identify their function from sequence alone. Decades of research have focused on understanding the mechanisms by which transcription factors read out cis-regulatory sequences and regulate gene transcription, but this knowledge is insufficient to make accurate predictions about when and where cis-regulatory sequences instruct gene regulation in the embryo. If we could decipher the cis-regulatory code, this would unlock an enormous amount of information encoded in the human genome, much of which influences our disease disposition and response to treatments.

(BPNet overview figure). We have shown that neural networks trained on high-resolution transcription factor binding data make highly accurate predictions and learn cis-regulatory rules consistent with mechanistic studies. At the same time, they also reveal novel unexpected patterns that can be experimentally validated and inform models of gene regulation. We are expanding this approach to many cell types and genomic assays, including chromatin accessibility, nucleosome occupancy, histone modifications, 3D chromatin organization, RNA polymerase II initiation and pausing, and nascent transcripts. The goal is to learn the sequence rules that drive each gene regulatory process such that we can eventually build models that read the entire cis-regulatory code for many cell types.

(Charles MD image) Our models suggest that transcription factors have two modes of cooperativity, one through strictly spaced motifs and one through motifs with soft syntax within nucleosome distance. To further understand these mechanisms, we combine the results from our neural network model with MD simulations and in vivo CRISPR experiments.

(Kalean embryo image) Our models suggest transcription factors pioneer chromatin proportional to motif affinity and in many cases act on nucleosomes cooperatively. We are currently investigating thermodynamic models for this type of cooperativity, how it affects transcription factor binding and how it might change the properties of enhancers.

(Charles image) We have created novel neural network models that predict the genome-wide occupancy of nucleosomes without the experimental bias imparted by MNase. Leveraging these models, we study the motifs that position or deplete nucleosomes and how they affect gene regulation. We also investigate how sequences not corresponding to transcription factor motifs might influence gene regulatory processes.

We take a variety of approaches to fill the remaining gaps in how the right combination of transcription factors form active enhancers, which induce nearby promoters to transcribe a gene. This includes interpreting models of histone modifications and enhancer RNA found at active enhancers, and analyzing high-resolution data of RNA polymerase II initiation and pausing.

We have previously developed ChIP-nexus technology to map transcription factors at the highest possible resolution in vivo. We are currently expanding our capabilities for single-cell and single-molecule approaches.