Representation learning of the epigenome

Abstract:
Assays of the human epigenome capture the regulatory state of cells in health and disease. With tens of thousands of experiments completed, it is now feasible to train large-scale representation models. In this talk, I will share how we approach this challenge by first abstracting all epigenome data into genomic intervals. I will describe several types of neural embedding models we use to investigate what biological questions can be addressed with epigenome region embeddings. I will also describe our efforts to curate, standardize, annotate, and share genomic interval data to make it more broadly useful for machine learning and beyond.

Contact

  Donna McMinn
  dlp18@psu.edu