*EASTBIO NPIF* Development of a machine learning-based approach for the analysis of DNA replication in primary mammalian cells

DNA replication is the fundamental process that allows transmission of the genetic information from mother to daughter cell. The correct execution of DNA replication is key to the maintenance of genome integrity and mistakes in this process are incompatible with successful embryonic development and underlie both aging and cell transformation. There are therefore multiple levels at which the process is controlled. One of the least understood aspects of the control of DNA replication is the temporal and spatial organisation of the activation of the origins of replication, the genomic regions where DNA replication starts. Origins of replication are not activated all at once, but follow a temporal program known as DNA replication timing. This program is cell-type specific and re-established every time a cell divides (reviewed in Rhind and Gilbert, 2013, CSH Persp Biol. 5, 1). The temporal control of origin’s activation, also corresponds to their spatial organisation, as the first origins that are activated are located centrally within the nucleus, while the origins that are activated later are peripheral. Whilst the tools to study the temporal aspect of DNA replication are already well developed, high throughput and standardised, the spatial organisation of DNA replication still relies on the slow and highly subjective manual classification of microscopy images. The Buonomo lab (http://sarabuonomolab.com/ ) has accumulated a large database of microscopy images of the different patterns of distribution of origins of replication that have been manually classified. These can therefore be used to train deep learning models in a supervised manner. The student will be encouraged to design and train novel deep neural network architectures and visualisation tools that can learn to classify images accurately and explain the predictions of the network by highlighting the responsible image regions. Most importantly, the Buonomo group has discovered the key protein that controls genome-wide replication timing and generated an allele that can be conditionally inactivated (Cornacchia et al., 2012, EMBO J. 31, 3678). The microscopy images of the mutant cells, that display a heavily altered replication timing, represent an invaluable tool to control the accuracy of the algorithm’s predictions and afford the chance to integrate biological perturbations in the learning process. While a large amount of imaging data is already available, the student will also be encouraged to acquire his/her own images, with the aim of developing sufficient laboratory skills and first-hand knowledge of the biological system being analysed.