I posted a short computational essay on Wolfram Community a little while ago (https://community.wolfram.com/groups/-/m/t/3727989) regarding what I call "molecular heatmaps" for visualizing atomic descriptors. This is based on previous work by #RDKit's Landrum and Riniker (cf. https://doi.org/10.1186/1758-2946-5-43).

Something I did not talk about in that essay is the virtue of reading journal articles that are outside of your usual purview (or, you might even say "comfort zone"). If I had completely restricted myself to the cheminformatics literature, I would not have found out about "blob functions" (cf. https://doi.org/10.1080/10867651.2001.10487549), which are often used in computer graphics. This isn't the first time I was able to profitably use knowledge from one subject to use in another one. It takes conscious effort, but I recommend having a wide reading appetite very warmly.

#chemistry #cheminformatics #visualization

Here's an #RDKit #cheminformatics quiz for you all. What do you think this code will output?

from rdkit import Chem
mol = Chem.MolFromSmiles("C" + "C(C)(C)" * 50 + "C")
pat = Chem.MolFromSmarts("[$([CD4H0X4](-*)(-*)(-*)-*)]")
print(len(mol.GetSubstructMatches(pat)))

No cheating by actually running the code! :) Feel free to explain your reasoning in the comments.

I just submitted a #cheminformatics preprint to ChemRxiv, based on the #RDKit count fingerprints, #chemfp, and some one-off R&D code I wrote over the last few months.

"Superimposed Coding of Count Fingerprints to Binary Fingerprints"

In short, my superimposed coding method gives k-recall@k nearest neighbor scores ~0.9 relative to using full count fingerprints and the multiset Tanimoto (aka MinMax, aka Ruzicka similarity). Recall can be over 0.95 w/ 8192 bits!

https://chemfp.com/SuperimposedCounts.pdf

#OpenBabel is dead, long live #RDKit!

https://github.com/RMeli/spyrmsd/issues/149

On a more serius note, it would be cool to have a cheminformatics library that actually works. Don't get me wrong, RDKit is very cool - but you can feel all the underlying problems it has when using it.

#Cheminformatics

Hey, @egonw - I'm working on a preprint.

How do I cite a source code file in the #rdkit and a commit message? FWIW, I use #Zotero.

"The RDKit implementation [of the multiset Tanimoto] was added in 2009, using fuzzy set operations already available for multiset Dice similarity."

"added in 2009" is commit 104efc5b607baa54ce0804c6a76d484bf9f78b57 at https://github.com/rdkit/rdkit/commit/423433a3e47df64af4a31888e835144e8b3a6c07#diff-d7a0f684fa993bfd84319df4d23b199973d13599b94ad6a4b3a6c79ed7d46719

"fuzzy set operations" is a reference to the two operations starting at https://github.com/rdkit/rdkit/blob/af4e6c05eca09efa8e8f61603937e0d997fc1499/Code/DataStructs/SparseIntVect.h#L132

Or am I overthinking?

#RDKit Atom Pair count fingerprints are wild! If you sum the per-record counts you'll see patterns like:

sum num_records
399 489
400 6
401 0
402 0
403 36
404 0
405 3
406 133423
407 9
408 6
409 42
410 6
411 0
412 36
...
493 83
494 1
495 3
496 116222
497 10
498 3
499 47
500 0

Not easy to plot! No doubt due to rings and chains causing highly repetitive path lengths. (406=2x7x29, 496=8x31)