Nikolaus Fortelny ist Postdoctoral Fellow in der Christoph Bock Group am CeMM – Research Center for Molecular Medicine of the Austrian Academy of Sciences in Wien.
Abstract:
Recent advancements in single cell technologies have led to the generation of sequencing data on thousands to millions of individual cells, numbers that are likely to scale up exponentially in the future. While the resulting datasets enable the study of biological systems at unprecedented detail, they also pose computational challenges in the integration, normalization, and especially interpretation of the data. The increase in scale of biological data is accompanied by rapid progress in the field of machine learning. Algorithms that rely on large datasets have reached human level in image recognition, complex games such as GO, and other tasks. In this talk, I will demonstrate how machine learning helped us to understand complex biological processes in a study of trans- and de-differentiation of pancreatic islet cells and a study of patient heterogeneity of drug responses in chronic lymphocytic leukemia. Complex machine learning algorithms such as deep neural networks are powerful predictors. However, the rules learned to generate predictions are often difficult to interpret in such black box algorithms. Their applicability could be significantly increased if it was possible to extract rules in addition to the predictions made. In the second part of this talk, I will demonstrate how interpretability can be enabled by integrating prior biological network knowledge into black box algorithms. Applied to a dataset of T cell receptor stimulation as well as the entire Human Cell Atlas, this modified algorithm learned rules that correspond to signaling proteins and pathways regulating cell state.