Imputing connections of random gene networks from time series data using ANNs
Sammanfattning: This thesis presents the architecture of a convolutional neural network which is trained to impute the connections of randomly generated gene regulatory networks under varying amounts of regularisation. The generated gene networks are simulated from 10 different starting conditions for each set of connections in order to obtain multiple time series. The generated time series are fed into the neural network for classification of the connections of each node. Ternary classification labels connections as inhibiting, absent, or promoting, whereas binary classification labels connections as present or absent. The performance of the neural network is evaluated by testing its accuracy on data it had not previously seen, known as test data. Ternary classification is unable to obtain a test accuracy above ~54% and binary classification cannot increase test accuracy beyond ~64%. Despite a relatively low test accuracy, the time dynamics which are obtained from the predicted networks perform much better than randomly generated networks. Furthermore, some of the higher test accuracy scores were associated with distributions of guesses which were heavily biased towards only guessing zeroes, whereas somewhat lower test accuracy scores had very realistic guess distributions. While the results did not produce a particularly high numerical accuracy, the time dynamics obtained from the predicted connections of the neural network resemble the real dynamics, which indicates that there is potential in the method. Making some improvements, such as changing the way gene networks are generated and creating super learner ensembles, has potential to substantially increase performance.
HÄR KAN DU HÄMTA UPPSATSEN I FULLTEXT. (följ länken till nästa sida)