3 ms·
it's not hard to find other molecules with similar scaffolds. Simply train an unsupervised network (autoencoder) over SMILES strings taken from ZINC. Convert
by dekhn 5y ago
it's not hard to find other molecules with similar scaffolds. Simply train an unsupervised network (autoencoder) over SMILES strings taken from ZINC. Convert your molecule of interest to a vector using the encoder, then find N nearest neighbors and decode. Those molecules will likely have similar scaffolds.
There are other methods, graph substructure similarity is common.
More to your actual question, it was derived from a previous attempt (https://en.wikipedia.org/wiki/Lufotrelvir https://en.wikipedia.org/wiki/Lufotrelvir) which has a very similar structure, and I wasn't able to find prior information by doing a little literature searching to see if there are other similar compounds.
- whymauri 5y agoRight, it's not technically difficult to match a scaffold -- you can probably parallelize a big RDKit job, but it's nontrivial to scale to a gigascale set. And, like you mentioned, there's rule-based substructure matching (sometimes easier for those willing to venture a bit outside SMILES). Anyways, it was more a curiosity question -- thank you for the details on Lufotrelvir!
- dekhn 5y agoyou wouldn't run a big rdkit job, you'd just featurize all your molecules and then do vector comparisons. lots less CPU but a big risk of false negatives.
- whymauri 5y agowhat does 'featurize' mean here? do you mean fingerprint feature vectors? IIRC, the computation time to get a fingerprint vector vs. extracting a murcko scaffold are pretty similar, although it depends on the fingerprint.