(140b) Using Atom Modules and Modularity of Molecules for Property Prediction

This paper proposes the use of modularity as a molecular descriptor combined with the atom groups or "modules" which are found by applying community detection algorithms to molecules represented as graphs or networks of atoms. These descriptors are then used to develop neural networks to predict physical properties of oxygenated organic molecules derived from biomass. The potential for this approach will be shown and demonstrated for the modeling of viscosity, lubricity and cetane number showing improved performance in respect to alternative models in terms of means squared error (MSE). Modularity on its own can be useful to estimate viscosity for n-alkanes, esters, isoalkanes, aldehydes, aromatics, and cycloalkanes. This was due to the capacity of modularity to capture the structural features of the molecules employed in the data set. As such, modularity can be exploited in the screening, design, and engineering of green chemicals derived from biomass for specific applications requiring tailored properties.