Geometric deep learning: Application to chemoinformatics
Keywords: graph theory, deep learning, machine learning, Python.
Duration : 8 to 10 months
Conversely to machine learning on data encoded as vectors, learning a prediction function on graphs arises different scientific bottlenecks. First of all, by their non euclidean representation, use of classic machine learning methods is non trivial. Some methods propose to embed the graphs onto an euclidean space. However, such projections induce a loss of structural information which may be difficult to control.
In the two last decades, some methods aimed to design graphs kernels or graph edit distance approximation methods to avoid an arbitrary representation of graphs as vectors. GREYC and LITIS laboratories from Normandie University collaborate on the definition of these methods. Since the emerging of deep learning methods, most of proposed approaches were defined using vectors as inputs. As a consequence, graphs were mostly apart of the scope of application of these methods, and graph based machine learning can not profit of impressive deep learning advances.
However, since the pioneer work of Gori and Scarselli [1, 2], some propositions were made to bridge the gap between graphs and deep learning. One particular application of graph based machine learning is chemoinformatics. Molecular compounds are naturally encoded as graphs and graph based methods are of thus methods of choice when predicting properties of molecules. Within AGAC regional project, the two laboratories GREYC and LITIS put in common their expertise to design and develop new machine learning methods based on graphs to be used in chemoinformatics. The project includes aspects of graph edit distance and kernels. A logic continuation is to study how geometric deep learning may help to improve results on chemoinformatics. To achieve this goal, we are recruiting a post doctoral researcher for 8 to 10 months.
This project will be supervised in close collaboration by LITIS (Rouen, France) and GREYC (Caen, France) laboratories which have a strong exper-
tise on graph based machine learning methods. The chemical expertise will be brought by COBRA laboratory (Rouen, France).
Details on the position:
Location: LITIS laboratory in Rouen (Normandy).
Date of desired start January 2019
Duration: 8 to 10 months
Salary: about 2200 euros/month net
• Updated CV
• cover letter explaining the candidate’s qualifications for the position,
• Letter of support (if applicable)
 Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph domains. In Neural Networks, 2005. IJCNN’05.
Proceedings. 2005 IEEE International Joint Conference on, volume 2, pages 729–734. IEEE, 2005.
 Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model. Trans. Neur.
Netw., 20(1):61–80, January 2009.
(c) GdR 720 ISIS - CNRS - 2011-2018.