Niccolo Cirone — University of Turin #
The intrinsic dimension of cell differentiation #
Despite the striking simplicity of the idea behind the Waddington Landscape, the identification of a quantity that approximates the differentiation potential is still an open area of research. Our goal is to identify a proxy for this differentiation potential avoiding biological assumptions, strong pre-processing and sophisticated approaches, focusing instead on the statistical and geometrical properties of single cell RNA-sequencing datasets.   The mRNA count matrix contains the coordinates of N cells in the high-dimensional expression space, where the dimensions are the sequenced genes. Because of biological constraints, data points are not uniformly spread along each direction, but are rather embedded in a subspace of dimension lower than that of the original expression space.  To prove it, we focus on the intrinsic dimension of this kind of data, which can be interpreted as the dimension of the manifold from which they are supposed to be drawn. Besides that, our main claim is that the intrinsic dimension reflects the level of potency of a cell population. We relied on different approaches to estimate intrinsic dimension, from projective (based on the decomposition of covariance matrix) to fractal ones: the main estimator we used is 2NN, a local quantity that leverages the statistics of distances between first and second nearest neighbors.  We studied several datasets referring to different biological processes and animal models, and we observed that the intrinsic dimensionality actually decreases during specialization, justified by a progressive structuring of data due to gene regulation mechanisms.