4
$\begingroup$

I'm trying to use a method for evaluating similarity of molecules. I have seen many publications in the machine learning community have used Tanimoto coefficient. However, when I read about it, it states that this is a measure of similarity for molecular fingerprint representation.

  1. I was wondering if this means that the measure of distance is not for the molecule itself but for its representation?
  2. If so, is there a measure that compares two molecules based off of their inherent structure rather than the particular representation of choice?

Also, any other suggestions on the choice of measure would be appreciated.

$\endgroup$
  • 3
    $\begingroup$ Related: jcheminf.biomedcentral.com/articles/10.1186/s13321-015-0069-3 $\endgroup$ – Tyberius Apr 2 '20 at 3:37
  • 1
    $\begingroup$ I'm not sure I understand what you mean by "inherent structure rather than the particular representation." If you're storing on the computer, you necessarily are making choices about the representation. $\endgroup$ – Geoff Hutchison Apr 2 '20 at 13:10
  • $\begingroup$ Of course the similarity measure depends on the representation. If I picked some 3D distance-based representation, I'd want some other similarity calculation (e.g. Euclidian) $\endgroup$ – Geoff Hutchison Apr 2 '20 at 13:13
  • $\begingroup$ @GeoffHutchison By inherent I mean the structural similarity. Taking two images as an example, I can find the Euclidian distance between the two images by subtracting corresponding pixel values. But does it give any measure of how similar these two images are? Do these images belong to the same category? Are both apples or one is apple the other one orange? $\endgroup$ – Blade Apr 2 '20 at 17:16
  • $\begingroup$ @GeoffHutchison In the molecular space, let's take alanine dipeptide as an example. Each configuration can be categorized by the characteristic conformation. So I find dihedral angles and I can find a measure of similarity in the space of alanine dipeptide molecules with different configurations. $\endgroup$ – Blade Apr 2 '20 at 17:21
6
$\begingroup$

I was wondering if this means that the measure of distance is not for the molecule itself but for its representation?

Inherently, you must decide on a representation of a molecule in cheminformatics. In most cases, the representation is a "2D" valence-bond connection table (e.g., SMILES, SD file, ChemDraw, etc.) familiar to anyone who has taken an organic chemistry class.

So yes, a similarity measure depends on the representation. Tanimoto coefficients are usually based on comparing binary fingerprints.

If so, is there a measure that compares two molecules based off of their inherent structure rather than the particular representation of choice?

I'm not sure how to answer this question since a representation forces choices about an "inherent structure" of a molecule. If I pick SMILES, I have no coordinates, implied hydrogen atoms, and choices about implied bond orders aromaticity, etc.

I think perhaps your question is about fingerprints - that "Is is there a similarity measure between a molecular graph - without generating fingerprints as an intermediate step?"

In that case, I think graph similarity measures are what you want, e.g. from Peter Willett, "RASCAL: Calculation of Graph Similarity using Maximum Common Edge Subgraphs" The Computer Journal (2002) pp 631–644

The reason people use fingerprint similarity measures is that they're easy to implement, and have worked fairly well for a wide variety of tasks.

$\endgroup$

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.