Deep learning has the potential to make a significant impact in basic biology and cancer, but a major challenge is understanding the reasons behind their predictions. My research develops methods to interpret this powerful class of black box models, with a goal of elucidating data-driven insights into the underlying mechanisms of sequence-function relationships.
Deep learning is being applied rapidly in many areas of genomics, demonstrating improved performance over previous methods on benchmark datasets. Despite the promise of deep learning, it remains unclear whether improved predictions will translate to new biological discoveries because of their low interpretability, which has earned them a reputation as a black box. Understanding the reasons underlying a deep learning model’s prediction may reveal new biological insights not captured by previous methods. Our group develops methods to interpret high-performing deep learning models to distill knowledge that they learn from big, noisy, biological sequence data. Our goal is to elucidate biological mechanisms that underlie sequence-function relationships for gene regulation and protein (dys)function. Recently, we have teamed up with other members of the CSHL Cancer Center to investigate the sequence basis of epigenomic differences across healthy and cancer cells.
Making AI algorithms show their work
May 13, 2021
AI machines are often better than humans at discerning patterns. CSHL researchers developed a way to find out why.
AI researchers ask: What’s going on inside the black box?
February 8, 2021
Although researchers have figured out how to train computers to recognize things, they have yet to understand how machines make those predictions.
Peter Koo wants to understand how machines learn biology
September 20, 2019
Dr. Peter Koo joins the CSHL faculty as an assistant professor. His focus is on exploring how artificial intelligence integrates with biology and genomics.