Dr. Anshul Kundaje
"Deciphering regulatory syntax and genetic variation with deep learning models"
Abstract: I will present efficient, compact deep learning models of regulatory DNA based on a consistent architectural backbone that exhibit high predictive performance, stability and interpretability when trained on regulatory experiments spanning a diversity of target readouts ranging from transcription factor binding, chromatin accessibility, nascent transcription and massive parallel reporter activity. I will show how we use these models to (1) debug experiments by detecting cryptic experimental biases, (2) debias the experiments by learning and correcting these biases, (3) reveal underlying biological causal sequence syntax, (4) extract biophysical parameters ab-initio (5) prioritize, predict and interpret the effects of regulatory common and rare genetic variants. Our models are competitive with or outperform much larger multi-task supervised models and fine-tuned DNA language models.
Biography: I will present efficient, compact deep learning models of regulatory DNA based on a consistent architectural backbone that exhibit high predictive performance, stability and interpretability when trained on regulatory experiments spanning a diversity of target readouts ranging from transcription factor binding, chromatin accessibility, nascent transcription and massive parallel reporter activity. I will show how we use these models to (1) debug experiments by detecting cryptic experimental biases, (2) debias the experiments by learning and correcting these biases, (3) reveal underlying biological causal sequence syntax, (4) extract biophysical parameters ab-initio (5) prioritize, predict and interpret the effects of regulatory common and rare genetic variants. Our models are competitive with or outperform much larger multi-task supervised models and fine-tuned DNA language models.
Dr. Allon Wagner
