We are interested in understanding how cell fate—the future identity of a cell—is specified during development. Many molecular processes, such as transcriptional regulation, epigenetics, intracellular signaling, and RNA regulation, are known to be involved in cell-fate specification. Furthermore, cell-fate specification is governed by complex interconnected gene networks. How such complex networks lead to the apparently simple and reliable phenomenon of cell-fate choice is an important problem in biology. Our long-term goal is to determine the rules governing cell-fate specification at the level of the mechanistic details of the regulation of individual genes and, more broadly, at the level of the organization of complex gene networks.

Our long-term goal is to determine the rules governing cell-fate specification at the level of the mechanistic details of the regulation of individual genes and, more broadly, at the level of the organization of complex gene networks

Blood-Cell Development as Model System

Our current work is focussed on cell-fate specification in mouse hematopoiesis. During hematopoiesis, the hematopoietic stem cell, through a series of intermediates with progressively restricted fate potential, gives rise to all the major types of cells found in blood. Hematopoiesis is one of the premier model systems for the study of both gene regulation and epigenetics as well as complex gene regulatory networks (GRNs).


We are a highly collaborative and multidisciplinary team developing novel combinations of empirical and computational approaches for understanding gene regulation and GRNs. We utilize functional genomics approaches such as RNA-Seq, ATAC-Seq, Hi-C, including single-cell versions, as well as genome editing to characterize chromatin states and gene expression during differentiation. We define and test the rules of cell-fate specification rigorously by coupling these empirical techniques with predictive computational modeling and machine learning approaches.

We couple functional genomics approaches with predictive computational modeling to rigorously define and test the rules of cell-fate specification

Research Directions

Decoding the gene regulatory logic of individual enhancers

The regulatory logic—the transcription factors (TFs), their roles, and their binding sites—of most genes remains to be discovered. The decoding of regulatory logic is challenging because of its complexity—genes may be regulated by multiple enhancers and each enhancer may, in turn, be jointly regulated by several TFs exerting positive or negative influence over the target gene.

We developed a data-driven computational approach to solve the problem of decoding regulatory logic in a scalable manner by training mechanistic models of transcription on functional genomics datasets. Our approach does not require a priori knowledge of the identities or the regulatory roles, activation or repression, of the TFs regulating a CRM. The TFs regulating a CRM and their regulatory roles are inferred in silico by testing many alternative models, each realizing a potential regulatory scheme, against quantitative reporter data.

We applied our approach to a gene essential for neutrophil development, Cebpa, whose regulatory logic was poorly understood at the time. We identified 4 enhancers, of which 3 were not known previously. Our in silico approach inferred a comprehensive map of the regulation of the Cebpa locus, with activation by PU.1, C/EBP family TFs, Gfi1, and Egr1, which was verified experimentally.

A map of Cebpa constructed by a combination of mechanistic modeling and functional genomics

Future Direction: Having demonstrated the utility of our approach by learning the regulation of an important developmental gene, we are exploring ways of scaling regulatory decoding to a genomic scale by utilizing high-throughput reporter assays.

The architecture of hematopoietic gene regulatory networks

One of the main ideas about how multipotential progenitors choose between the red-blood cell and myeloid lineages is that two so-called master regulators, Gata1 and PU.1, function as a bistable switch by repressing each other’s expression. Although simple, this model is based largely on genetic and biochemical analyses conducted at steady state, which may not capture causality.

We took an alternative, machine-learning based, approach of inferring genetic architecture by training mechanistic dynamical models on gene expression time series data. We constructed a model of a 12-gene GRN including key lineage-specifying TFs, cytokine signaling effectors, and cytokine receptors using a high-temporal resolution dataset of erythrocyte-neutrophil differentiation. The model can quantitatively predict the consequences of several genetic perturbations and the inferred architecture is consistent with prior evidence.

Our analysis of the transient dynamics of gene regulation produced a surprise. In contrast to the bistable-switch model, which posits that PU.1 initiates the differentiation of neutrophils, our model predicts that PU.1 is activated much later in the differentiation process by C/EBPα and Gfi1, TFs known to be necessary for myeloid differentiation. This prediction was also validated in an independent single-cell RNA-Seq dataset. Besides providing insights into the causality of events and the network architecture underlying lineage choice, this work establishes the our modeling approach as a powerful tool for understanding cell-fate choice.

Future Direction: The extensive amount of lineage antagonism and transdifferentiation documented in the hematopoietic system implies that the specification of a lineage not only depends on the factors expressed in it but also on factors expressed in other lineages. Accurate modeling of lineage decisions therefore must integrate multiple lineages. We are expanding the scope of our work to the choices between multiple hematopoietic lineages and to datasets generated by newer single-cell techniques.

Technology development: Fast Inference of Gene Regulation
Empirical spatiotemporal gap gene expression patterns (A) and FIGR simulations (B)

Hematopoietic cell-fate specification relies on large densely interconnected gene networks. A major bottleneck in predictively modeling large networks using differential equations is that parameter inference is computationally expensive. For example, inferring the parameters of a 12-gene network takes about 30 hours on 10 CPUs. Larger networks are nearly impossible to infer since the computational demands grow very rapidly with the number of genes. We have developed a new method for parameter inference, Fast Inference of Gene Regulation (FIGR), that utilizes supervised classification techniques and is vastly more efficient than previous methods. For example, FIGR achieves a 600x speedup on the gap gene network of Drosophila.

Future Direction: FIGR enables differential equation modeling of much larger networks than was previously possible and we are utilizing it our efforts to infer hematopoietic gene networks from multiple lineages.

Non-additive control of gene expression by multiple enhancers

Classical enhancers have long been regarded as acting additively in a distance independent manner, but recent experiments do not support the classical assumptions and reveal non-additive behavior. Non-additive responses could be a result of interference between enhancers at the level of promoter-enhancer looping or chromatin structure. We are combining genome editing, epigenetic profiling, and computational modeling to investigate the mechanisms underlying non-additive co-regulation of genes by multiple enhancers.