The main principle behind this tool is to predict whether a mutation will have an impact on a protein's interaction with another. It will mainly use deep learning approaches.
To do so, we encode each protein in a supervised manner to increase the relative weight of the mutations.
The './draft' directory contains proof-of-concept code supplied by Quentin Ferré and meant to be mainly edited and reviewed by him.
We may work with conjoined triad (representaition of amino acid composition) or with FASTA sequences.
Use a supervised encoder to create a new representation of the proteins ,alone or in relation with their mutants.
The supervised aspect allow us to emphasize the mutations and their impact on interactivity.
It is then possible to extract outputs from the intermediary layer to, well, extract this representation.
Trying to maximize output for the neurons in the representation layers allows us to see where each neuron pays attention.
Such a representation can be used in an Adaboost model.
It is also possible to create a simple decision tree. While less precise, it is graphically visualizable.
Keras Documentation : https://keras.io/
Zacharie Menetrier zacharie.menetrier@gmail.com
Quentin Ferré quentin.ferre@inserm.fr