 So it turns out we can do that knowledge-based statistical scoring on a grid 2 in a very easy way. Just pick out all those structures from the protein data bank, put them on a grid, anything that has a small compound bound, check what is the probability of having a carbon in that grid point, what is the probability of having an oxygen, the probability of having a nitrogen or a hydrogen, maybe sulfur, the few most common atoms. That creates an effective, statistics-based potential for each type of atom on each grid, and then I can take a random new molecule and quickly assess is it likely to be good or bad to have this molecule? Not perfect, but super fast, in particular when we use fast Fourier transforms to handle all the translations on the grid.