A Method For The Inverse Qsar/Qspr Based On Artificial Neural Networks And Mixed Integer Linear Programming

In this study,we propose a novel method for the inverse QSAR/QSPR. Given a set of chemical compounds G and their values a(G) of a chemical property, we define a feature vector f(G) of each chemical compound G. By using a set of feature vectors as training data, the first phase of our method constructs a prediction function psi with an artificial neural network (ANN) so that psi(f(G)) takes a value nearly equal to a(G) for many chemical compounds G in the set. Given a target value a* of the chemical property, the second phase infers a chemical structure G* such that a(G*) = a* in the following way. We compute a vector f* such that psi(f*) = a*, where finding such a vector f* is formulated as a mixed integer linear programming problem (MILP). Finally we generate a chemical structure G* such that f(G*) = f*. For acyclic chemical compounds and some chemical properties such as heat of formation, boiling point, and retention time, we conducted some computational experiments with our method.