"Multi-scale Deep Tensor Factorization Learns a Latent Representation of the Human Epigenome"

Jacob Schreiber, University of Washington

Monday June 11th
12:00 - 1:00PM
Clark S361
Lunch provided

ABSTRACT: The human epigenome has been experimentally characterized by dozens of assays across hundreds of cell types resulting in thousands of measurements describing each genomic position. While these measurements have provided insight into important biological phenomena, the massive amount of data can complicate computational analyses. We introduce Avocado, a multi-scale deep tensor factorization approach that learns a low dimensional latent representation of the human epigenome from the data contained in the Roadmap compendium. This representation is trained to impute epigenomics experiments, and we first show that Avocado outperforms previous approaches at this imputation task in a variety of evaluation schemes. Next, we show that this representation is useful past the task that it was trained on in the settings of predicting gene expression, promoter-enhancer interactions, the boundaries between topologically associating domains (TADs), and at predicting regions that interact frequently in the three-dimensional structure of the genome (FIREs). We then use feature attribution methods to better understand how Avocado works, and find that it encodes a rich functional landscape.

BIO: Jacob Schreiber is a Ph.D. student and NSF IGERT Big Data Fellow at the University of Washington, where he studies the application of scalable machine learning to genome science problems. In his free time, he is also the author of software packages pomegranate and rambutan, and a core developer for scikit-learn.

ABOUT THE SCGPM: The Stanford Center for Genomics and Personalized Medicine (SCGPM) seeks to advance genomic technology so that someday both genetic and molecular profiling will become powerful and routine tools for predicting disease risk and monitoring and treating a wide range of pathologies. Towards this mission, the SCGPM serves to centralize and develop collaborative intellectual and technological resources that promote genomic research and analysis, predict drug response, educate physicians, and examine the ethics of personalized medicine. These efforts include large basic science projects such as ENCODE that decipher the human genome as well as clinical research projects such as the sequencing of cancer genomes and individuals with inherited diseases. Through these efforts, the Center aims to bring genomics to the clinic.

For more information about the SCGPM, go to http://scgpm.stanford.edu.