Closed UnixJunkie closed 4 years ago
Their atom type is almost: (partial-charge, VdW-radius, LJ-epsilon)
but the simplest way for us is to change bond distance in each atom pair to Euclidian distance. Then, in the Tanimoto, we need to use a triangle kernel function whose bandwidth has been optimized on a given dataset in some way.
test on a pre-docking experiment: does this improve the regressor performance compaired to working only in 2D
Since we need to vectorize the output, the triangular kernel is not acceptable. We can use linear binning or an histogram with overlaping bins.
giving up; this is just annoying
very interesting encoding in there "Zhu, F., Zhang, X., Allen, J. E., Jones, D., & Lightstone, F. C. (2020). Binding Affinity Prediction by Pairwise Function Based on Neural Network. Journal of Chemical Information and Modeling." https://pubs.acs.org/doi/abs/10.1021/acs.jcim.0c00026