École Nationale Supérieure de
Techniques Avancées

Résumé

Using a Deep Learning model to predict chromatin accessibility gives better results than with more standard approaches. When trained with hundred of thousands of DNA sequences, such a model can learn the complex grammar of DNA. However, a deep learning model works like a "black box" because it is difficult to determine which features of the input are used to make the prediction. This report describes the training and optimization of the Basset model with Keras, a python Deep Learning library, and the implementation of DeepLIFT, a tool recently developed to facilitate the interpretation of deep learning models. DeepLIFT gives importance scores to the nucleobases of the input DNA sequences of our model. Studying the scores makes it possible to find relevant patterns or motifs for gene expression. We did manage to generate the importance scores but we could not extract the motifs in studied DNA sequences. We then tried to find interactions between patterns within a sequence using the DFIM method. We did not find any clear interaction using DFIM but the work is still in progress