In Silico Perturbation of Single Cells

One of the main goals of computational biology is to develop realistic models of cells, such that their behaviour can be studied in silico (i.e. in a computer simulation) and conclusions can be drawn on the actual biological phenomena we are considering. A perturbation, in this work, is defined as a change in the external environment or in the inner mechanisms of the cell. In order to produce actionable simulations it's imperative that the response of a model to a perturbation is as close as possible to what happens in reality. The aim of this work is to establish a metric of evaluation of different models, able to discern which among them behaves most similarly to experimental results. As we will see there is no agreed upon method in the literature, and the commonly employed strategies have some disadvantages that will be highlighted and improved upon. The main contribution is the development of a method taking full advantage of the characteristics of single cell data, mainly the joint probability distribution of the gene expression levels, that can now be estimated and could not have been with traditional bulk transcriptomics. With bulk transcriptomics in fact we can only determine the average expression levels of a given gene in the sample. Instead with single cell data we can appreciate the complex intertwining of the various genes' activity, since we can see for any given cell whether a certain gene tends to be co-expressed with others, and so on. Current methods are lacking on this point since they perform evaluations by aggregating data, in what's called pseudo-bulk, i.e. averaging the expression levels for any gene in a sample sequenced with scRNA-seq. This is almost equivalent to using bulk data, therefore I argue that there's room for improvement on this front and I propose one such technique in this manuscript.