Kipoi Seminar - Katherine Pollard, Gladstone Institute of Data Science & Biotechnology, UCSF, Biohub

Title: Human variant interpretation with sequence-to-activity models Abstract: The role of computational science in biomedical research has typically been downstream of experiments, where it plays important roles in signal processing, data integration, pattern detection, and hypothesis testing. But this is changing, and predictive models are now being used to generate and test hypotheses in silico. This change has been fueled by experimental technologies that drive rapid growth of biomedical databases couple with breakthroughs in computer hardware and software that have enabled the development of highly accurate, predictive modeling frameworks. My lab leverages these opportunities to decode how the human genome works, evolves, and breaks in disease. In this talk, I will share examples from human genomics, where we have built deep learning models to predict molecular phenotypes, such as epigenetic states, genome folding, and gene expression from DNA sequence alone. Then, we use the models to computationally predict how changing the DNA sequence alters genome function. This strategy leads to causal hypotheses and enables us to prioritize disease variants with predicted functional effects. Experiments designed using model outputs are accelerating the rate of discoveries, shedding light on genetic mechanisms in human evolution and disease. This prediction-first strategy exemplifies my vision for a more proactive, rather than reactive, role for computational science in biomedical research. ---------------------------------------------------------------------------------------------------- Add a comment regarding models sharing at NIH: https://osp.od.nih.gov/comment-form-r...