Machine Learning and Friends Lunch: Jacob Schreiber, Modeling and Design of DNA Regulatory Elements
Content
Speaker
Dr. Jacob Schrieber (UMass Chan Medical School)
Abstract
Recent advances in predictive modeling of genomic modalities have resulted in state-of-the-art performance, but the size and complexity of these models have made them ill-suited for many important tasks, such as designing regulatory DNA with specific properties. In this talk, I will introduce a new predictive model, Cherimoya, which uses an order of magnitude fewer parameters than the closest competitor while still outperforming it. This is made possible through the Cheri Block, a new layer that is a middle ground between the transformer and a dilated convolution. Then, I will describe our DNA design algorithm, Ledidi, that minimally edits template DNA to achieve functional outcomes, and showcase recent design successes.
Speaker Bio
Dr. Jacob Schreiber is an Assistant Professor in the Department of Genomics and Computational Biology at UMass Chan Medical School, where he runs the Programmable Genomics Laboratory, and a Visiting Scientist at the Broad Institute.