Prediction of Molecular Properties by Recursive Neural Networks. Application to the Glass Transition Temperature of Acrylic Polymers

In the past few years, a novel approach in cheminformatics for the Quantitative Structure-Property Relationship (QSPR) analysis of Physical, chemical and biological properties of chemical compounds was developed at the University of Pisa. This methodology is based on the direct treatement of molecular structure, without using numerical descriptors, and employs recursive neural networks. In subsequent studies it was successfully used to predict varoius properties of different classes of compounds. It is a promising tool in the evaluation of existing substances, as well as in the design of new materials. This master thesis fucuses on the prediction of the properties of polymers, a problem not easily treatable with traditional methods based on molecular descriptors. The study explores different representational issues and show the accuracy and flexibility of the structure-based QSPR approach