Item

Abstract

We develop a methodology for automated extraction of network information from a large dataset containing images of Drosophila terminal cells. The dataset contains images of larvae grown with different mutations prohibiting the expression of one of four genes: Rab8, Myospheroid, Crumbs and Rhea. Larvae are also distinguished based on their genetic background and growing temperature. The dataset is composed of over 500 images which is a novelty for this field of research. This enables us to find statistically highly significant results. We apply a supervised learning approach to quantify the effect on discernability of each of the three growing conditions. Using an unsupervised learning approach we find hidden phenotypes spanning several of the already known phenotypes induced by the larva’s genotype. We find that most of the information contained in network growth patterns is strongly tied to network size. By analyzing deviations from the size dependence of network realization we establish four main growth characteristics we call phenotypic trends. We are also able to find very simple models describing cell branching behaviour and distributions of tube lengths and tube radii.