I develop rigorous mathematical and machine learning methodologies, including generative probabilistic models, deep learning architectures, and information theoretical approaches, to decode the spatial organization and temporal evolution of genetic and epigenetic features underlying various tissue types and diseases.
Please describe one or two of your most interesting projects.
We pioneered the mathematical model of spatial gradient modeling in the spatial gene expression of tissues. Many tissues form certain geometries, such as layered structure in cortex, concentric circles in olfactory bulb. The spatial gradient should reflect the tissue geometries. Assuming that a 2D tissue has a one-dimensional axis that describes the intrinsic geometry, we formulated the intrinsic axis learning problem using conformal map when two boundaries of the tissue geometry are available, and additionally modeled the gene expression as a piecewise function along the intrinsic axis. We subsequently developed neural network models to learn one or multiple intrinsic axes in an unsupervised way relaxing the requirement of the tissue boundaries.
We develop a nested hidden Markov random field and hidden Markov model to simultaneously learn the spatial organization of tumor clones and allele-specific copy number aberrations, one typical cancer drivers, of each tumor clone. The spatial organization of genetically distinct tumor clones further allows inferring the temporal tumor evolution and movement in space.
What makes you excited about your data science and AI research?
I believe data science and AI research are not only a tool for prediction but also a way for understanding the complex world. By consolidating mathematical models and automated learning scheme in AI, we can uncover hidden relationships and rules embedded in the large data that is otherwise invisible.