Machine Learning for classifying proteins by shape.
ABSTRACT (less than 100 words):
New neural networks architectures are being designed for handling 3D objects directly rather than by considering multiple 2D projections, most notable the family of PointNet models from Stanford and of GCNN (Geodesic CNN). This thesis shall exploit such architectures in the context of macromolecular shapes and, in particular, proteins. Despite the success of such architectures on rigid 3D objects, namely furniture and CAD models, molecules seem significantly harder given the intricacy of their structure but also the variety of representations possible. This thesis shall juxtapose different representations mostly based on the backbone sequence of atoms, and may focus on specific protein families and/or specific categories of structure. Our models shall be trained on the extensive structural databases such as CATH and SCOP. This work is inscribed in the LAMBDA and GRAPES European networks.
NAME & POSITION OF THE SUPERVISOR:
Professor, Department Informatics & Telecoms, NKUA
Adjunct researcher, ATHENA Research
LAB/GROUP, DEPARTMENT, INSTITUTION where the thesis will be executed:
Lab of Geometric and Algebraic Algorithms
Department Informatics & Telecoms, NKUA