I have recently been reading a paper, in which given dihedral angles of alanine dipeptide molecule $\psi$ and $\phi$, the conformation is catagorized as either $\beta$-1, $\beta$-2, or $\alpha$. I was trying to see if it is possible to extend the same idea to QM9 molecules and cluster them into different groups based on their properties? The properties provided for the QM9 dataset include:

  1. Norm of the dipole moment ($\mu$)
  2. Norm of static polarizability ($\alpha$)
  3. Energy of the electrons in the highest occupied molecular orbital (HOMO energy), lowest unoccupied molecular orbital (LUMO energy), and their difference (Gap)
  4. Zero point vibration energy (zpve)
  5. Atomization energy (U)
  6. Electronic spatial extent (R2)

The answer that I'm looking for is for instance something in the lines of: "based on the $\delta$ values, molecules can be divided into to two groups of $\gamma$-1 and $\gamma$-2". In other words does any of these properties inherently divide the dataset into k groups?

  • 1
    $\begingroup$ This seems a little too broad at this point. For example, you could divide them into molecules without a dipole (or below some threshold) and molecules with a dipole. The same could be done for any of these properties. It would help to know what you hope to accomplish with the groups. $\endgroup$ – Tyberius Dec 25 '19 at 2:29
  • 1
    $\begingroup$ @Blade: It doesn't seem you'll get an answer here. Perhaps a Stack Exchange dedicated only to computational physical chemistry or materials science, where there's no questions like "what is an acid?" and the real experts in the field are concentrated and engaged, would be helpful. You might be interested in committing so that you can be part of the private beta of this: area51.stackexchange.com/proposals/122958/… $\endgroup$ – user1271772 Dec 26 '19 at 19:42
  • 1
    $\begingroup$ Thanks for the suggestions, I'll try both. Just for future readers, let me correct your sentence : I'm essentially just looking for something else I can predict and group. Previously: INPUT: molecular configuration of a single molecule, PREDICTION: cluster based on conformation type. Now: INPUT: different molecules in QM9 dataset, including atom types and bonds, PREDICTION: this is my question, what people would be interested to see as a prediction, what are common predictions that people address in their works. $\endgroup$ – Blade Dec 26 '19 at 20:01
  • 1
    $\begingroup$ @Tyberius Also, I liked the k-means clustering idea very much, the only additional improvement would be if the k was actually meaningful. Meaning that for instant I can divide HOMO energy into "k= any number that I like", but it would have been nicer if HOMO energy could have inherently been divided into k meaningful groups, not a random number that I picked. I guess I'll modify my question now as "if any of these properties inherently divide the dataset into k groups." $\endgroup$ – Blade Dec 26 '19 at 20:08
  • 1
    $\begingroup$ IMHO, the problem is two-fold: (1) that the molecules in QM9 are an enumeration of possible molecules up to 9 non-first-row atoms, not a set of known molecules and (2) it's not obvious what the purpose of the clusters would be. In the case of Ramachandran plots, they indicate the shape of the secondary structure. Since QM9 is something of an artificial group, I can't think of an obvious application for clusters from it. $\endgroup$ – Geoff Hutchison Dec 27 '19 at 3:39

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.