Deep learning models to predict primitive streak formation in human development

Project Summary Congenital birth defects affect an estimated 3% of live births. To develop effective treatment strategies, a thorough understanding of early human development is necessary.

Our lab recently developed and in vitro model of human gastrulation, the process by which the three germ layers (endoderm, ectoderm, mesoderm) are formed around week three of gestation.

This so-called ?gastruloid? model is formed by treating human embryonic stem cells with purified differentiation factors that cause them to self-organize into a pattern resembling a gastrulating embryo.

One of the key events during this process is formation of the primitive streak?a migration of specialized mesenchymal stem cells along the embryonic midline that will form all mesodermal tissues including the heart, lungs, blood vessels, and cells of the circulatory system.

At the same time, cells on the periphery of the embryo begin forming extraembryonic mesoderm, which will ultimately become placental tissue.

Despite the critical importance of these cell fate changes, it is currently unclear which population of embryonic stem cells will differentiate to form primitive streak or extraembryonic mesoderm and how these cell fate decisions are determined.

The research objective of this fellowship proposal is to understand when and how human stem cells differentiate into primitive streak and extraembryonic mesoderm during gastrulation.

My overall approach is to use time-lapse fluorescence imaging to monitor differentiation decisions in real time and at single-cell solution.

I will then employ a specialized type of machine learning known as deep learning to accurately track the movement and signaling behavior of individual cells.

Next, I will develop a computational model that uses a cell?s image patterns to accurately predict how each cell ?chooses? between differentiation fates.

The two specific research aims are: 1) to identify the subpopulation of human embryonic stem cells that will commit to primitive streak; and 2) to determine the combination of intracellular and extracellular signaling events that govern differentiation to extraembryonic mesoderm.

The proposed work includes novel experimental procedures (specifically, real-time imaging of gastruloids formation in Aim 1) as well as unique neural network architectures that accurately predict binary cell fate outcomes of individual stem cells based on their signaling history. These methods will be generalizable to other biological systems.

The proposed training plan focuses on generating and applying cutting-edge statistical methods tasked with full single-cell feature data incorporation in order to make robust, theoretically and biologically sound predictions about human stem cell fate decisions.

A better understanding of early human development will inform future cellular therapies to prevent and treat congenital birth defects.

To support my training, I have assembled a strong mentorship team with expertise in stem cell biology, live-cell imaging, machine learning methodologies, and causal inference.