Using DNNs to Predict Mouse Neuronal Connectivity based on Gene Expression

For my full report, please contact me at ashley.h.kim@columbia.edu

Abstract: Recent advances in barcoding and neuroimaging technologies have allowed for increasingly precise and detailed datasets that match gene expression to neural connectivity. The purpose of this project is to utilize deep learning methods on one such dataset in order to predict a mouse’s motor and audio neuron connectivity based on data about its gene expression. Experimentation with various neural network architectures indicated that a simple two-layer multilayer perceptron network was the most effective for predicting both audio and motor neural connectivity types, with audio neuron classification accuracy rates consistently ~50% and motor neuron rates ~60%. This limited accuracy is to be expected given the small number of input features compared to output features and the fairly limited scope of the connectivity-gene expression data. Given the number of classification categories (36 for motor, 12 for audio; 2.78% and 8.33% expected accuracy of random guess rates respectively), the final neural network performed exceedingly well compared to an untrained network.