We describe our implementation of the Zhao, Morrison, and Parr method [Phys. Rev. A 50, 2138 (1994)] for the calculation of molecular exchange-correlation potentials from high-level ab initio densities. The use of conventional Gaussian basis sets demands careful consideration of the value of the Lagrange multiplier associated with the constraint that reproduces the input density. Although formally infinite, we demonstrate that a finite value should be used in finite basis set calculations. The potential has been determined for Ne, HF, N-2, H2O, and N-2(1.5r(e)), and compared with popular analytic potentials. We have then examined bow well the Zhao, Morrison, Parr potential can be represented using a computational neural network. Assuming v(xc)=v(xc)(rho), we incorporate the neural network into a regular Kohn-Sham procedure [Phys. Rev. A 140, 1133 (1965)] with encouraging results. The extension of this method to include density derivatives is briefly outlined.