École Nationale Supérieure de
Techniques Avancées

Abstract

The project I undertook relates to models of protein evolution, which are used to determine the evolutionary links between living beings. I had to find a method enabling to find the optimal parameters for a new cutting-edge model. First, I used a simulated dataset for which the optimal parameters were known. Step by step, I tried to find a method that determines the best class frequencies and the optimal number of classes. I compared different methods, combined them, and selected the best ones. In the end, the results are almost perfect; they are really close to the known optimal parameters. Then I tested my two best methods on real data (with three different datasets). They seem to lead to really good mixture models that are fairly simple to run and yet give better results than the current mixture models.