A pipeline for gene expression profiles analysis to predict physical connections through the brain regions

The aim of this thesis is to perform an integrative analysis of gene expression and connectivity data using machine learning techniques. The implemented algorithm is able to detect the physical connection’s degree between brain areas exploiting patterns of gene expression profiles. In this study, I used gene expression and connectivity data (axonal projections) available on the Allen Mouse Brain Atlas (AMBA) resources combined with the connection’s intensity reported on Brain Architecture Management System (BAMS) database. When the viral tracer is injected in a brain region, called source region, it produces axonal projections in several target regions. Expression and connectivity data are provided as grid data, a 3D matrix representing mouse brain volume. Each item of the matrix is a voxel, that stores gene expression or axonal projection’s intensity quantified as energy level. The first phase of the work was preliminary data downloading step. I have obtained grid data for 3318 genes and 2333 injection sites for coronal section from Allen Brain Atlas. Projection data were reconducted to an unique value representing the connectivity’s intensity between the source and the target region. This value was obtained calculating the median of all the projection energy values for each source-target brain regions. In order to create a model capable to recognize the level of connection of two areas, N source-target regions are selected in accordance with their connectivity intensity and the type of analysis to perform. These data undergo an ad hoc pipeline in order to obtain M vectors representing gene expression profiles for the selected source-target regions and their labels. These are used as dataset to feed a MultiLayer Perceptron. Initial experiments consisted in a binary classification between connected and unconnected regions. . To perform the binary classification, I assigned label “1” (codifying connected regions) to the source-target vectors obtained by pairs with maximum intensity on BAMS, otherwise I assigned label “0” (codifying unconnected condition). The trained model has shown accurate performances suggesting that a strong correlation between gene expression and axonal connectivity exists. More in detail, accuracy, recall and f1_scores reached 1.0 value on the test set after few epochs. Verified this first hypothesis, I designed a second set of experiments to test if gene expression profiles can contain enough information to predict the intensity of connections between regions. The dataset was created selecting only one source region and some of its multiple targets with different connection’s intensity. The regression produced real output values representing connection’s intensity that have been revealed to be accurate. In fact, encouraging results have been observed reporting a mean square error equal to 3.02∗10-4±0,00077∗10-4 on the test set. Since the model has proved to be promising, I’m working on further experiments. These focus to the generalization to a wider dataset composed of cortex and cerebellum’s data (approximately 58 regions). The purpose of this thesis is to predict connection’s intensity for the couples of brain regions for which BAMS matrix does not report complete connectivity data. The complete matrix is obtained analyzing gene expression and connectivity’s energy data provided by AMBA.