3
$\begingroup$

I was reading about molecular graph based generative models, which attempt to generate new molecules by training on a subset of a molecular dataset, such as QM9. In these works, there are 2 types of constraints that define the validity of the output:

  1. The sum of the order of the bonds that are incident to an atom may not exceed its valency.
  2. The output should be a single connected graph.

While we can go further and do DFT calculations to see if these are also stable structures, I was wondering if there are any other simple constraints that we can impose to check the validity of the output molecular graph, without the need for DFT-type calculations.

$\endgroup$
  • $\begingroup$ The idea of running DFT calculations implies that you have you graph together with some tentative geometry, not just connectivity. Is that so? $\endgroup$ – Ivan Neretin May 6 at 15:38
  • $\begingroup$ @IvanNeretin I can theoretically have, but I'm not. Essentially, I am training my model with graphs that only have the atom type information. As coordinates are also available in the training dataset, I can add them as additional features for each atom (that is something that I'm thinking of, but haven't found a potential benefit for that, yet). $\endgroup$ – Blade May 6 at 15:47
  • 1
    $\begingroup$ @Blade coordinates could be useful, along with atom type, to determine reasonable distances between atoms. You could have graph of two hydrogen atoms that satisfies your constraints, but if the atoms are 1000 angstroms apart, you can quickly determine that the graph doesn't represent a stable molecule. $\endgroup$ – Tyberius May 6 at 15:51
  • 1
    $\begingroup$ In terms of additional constraints, you will also want to have a minimum valency for each atom (C-H doesn't make sense as a whole molecule) and account for multiple bonds (quadruple bonded carbon doesn't makes sense, but quadruple bonded tungsten would). Beyond that, I can't think of any other obvious graphical constraints. You might be able to make some pseudogeometric considerations of things like a max/min ring size. $\endgroup$ – Tyberius May 6 at 23:32
  • 1
    $\begingroup$ As an aside, based on your previous questions, you might be interested in the new Materials Modeling SE. It might better address some of your questions that are more on the simulation/computation side. $\endgroup$ – Tyberius May 15 at 21:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.