Data Analysis Project Presentation

  • Gates Hillman Centers
  • ASA Conference Room 6115
  • Machine Learning Masters Program
  • Machine Learning Department
  • Carnegie Mellon University
Project Presentations

Recovering Developmental Dynamics from Single-Cell Data via Penalized Principal

Modern single-cell technologies offer a detailed view of the conditions and states of thousands of cells at the individual (single-cell) level. For asynchronous cell processes, such as hematopoeisis, the full spectrum of developmental stages can be captured in a single snapshot, and a common objective is to align the cells along a trajectory representing the developmental process. Unlike most methods that require prior dimensionality reduction, our variational approach operates directly in the original high-dimensional space to output a regularized curve that efficiently approximates the data. We apply our method to a single-cell mass cytometry dataset from human bone marrow samples to recover trajectories representing B cell development. We obtain results that closely resemble those of a popular method for this task, and we find that computed trajectories from five human samples exhibit consistent expression patterns of canonical features. As our approach allows, we also analyze the role of features in the trajectory, and finally we discuss additional ways of direct curve comparison.

Faculty Committee: Ziv Bar-Joseph, Dejan SlepĨev

For More Information, Please Contact: