how to download smi.anon file for placement in pipeline --sw_database

[JonasLi-19]

Hi, I plan to use fragmenstein to do a hits improvement by:
python -m fragmenstein pipeline
--receptor receptor.pdb
--hits hit1.sdf hit2.sdf hit3.sdf
--analogue_database enamine_real.db
--max_tasks 100
--cutoff 5
--sw_databases 'REAL-Database-22Q1.smi.anon'
--sw_length 100
--sw_dist 0.2
--use_pyrosetta
--output results
I have already use smallworld to search the available database and there is a REAL-Database-22Q1.smi.anon.
BUT I don't know how to get it?

[matteoferla]

That setting is for using the server with NextMove SmallWorld hosted by John Irwin at sw.docking.org.
The datasets that one can use are preset.
For a custom dataset you will have to rely on a local catalogue analogue search.
SmallWorld is used here because it is by edit distance, while Tanimoto is bad for smaller compounds.
But normal Tanimoto search algorithms are fine otherwise.

[JonasLi-19]

So if I want to use smallworld like you do in the example "pipeline", to get molecules similar to my Hits of VS, I need to download the python package: pip install -q smallworld-api ?
Otherwise I don't know how to prepare a local database as .smiles.anon format to use in "pipeline".

[SanFran-Me]

Hi, I have similar questions, because I don't know where is the Enamine database, as many Enamine database needs to purchase...
--analogue_database enamine_real.db How can we get the database to do analogue searching for "placement" without purchasing the Enamine?

[SanFran-Me]

If I have downloaded an large database, and two hits, can I just use

"from rdkit import Chem
from fragmenstein import Monster
hit1: Chem.Mol = ...
hit2: Chem.Mol = ...

monster = Monster([hit1, hit2])
monster.combine()"

And then use Tanimoto to search analogues?

[matteoferla]

@JonasLi-19 — Yes. The default pipeline uses the package smallworld-api, which is an SDK that queries the API of sw.docking.org. The *.smiles.anon datasets are present in sw.docking.org and you don't have to download anything. If you want to use a custom dataset then you'd need a different tool or a local version of SmallWorld (sold by NextMove). To see what databases there are without copying off the browser, you can do in Python this.

[matteoferla]

@SanFran-Me

I don't know where is the Enamine database, as many Enamine database needs to purchase

Vendor catalogues are mostly free, but require registration, but will be in different formats.
It is service providers that do sell the catalogues as a service.

For example:

• https://enamine.net/compound-collections/real-compounds/real-database
• https://lifechemicals.com/fragment-libraries/
• https://mcule.com/database/

can be downloaded with a browser happily.

The odd one out would be Enamine REAL SPACE, which either requires BioSolveIt to generate the compounds or can be bought as a stack of hard disks as it is so insanely huge. However, if you sign some paperwork you get the list of reaction and can do it yourself.

Some are secret, like Edelris Keymical fragment collection. In those cases, you generally cannot buy a given compound anyway.

How can we get the database to do analogue searching for "placement" without purchasing the Enamine?

The pipeline demo in the paper uses sw.docking.org as mentioned. The datasets are in the server.

And then use Tanimoto to search analogues?

Yes. And then the placement route can be done.
The placement will calculate the mapping of which atom is inherited from where.
In some cases a custom_map needs to be provided if something alien is needed —the "it made sense in my head" designs
The catch is that commands like Victor.get_plip_interactions and the pandas.dataframe wrappers (eg. or Laboratory.score) need to be called individually.

The only caveat is that Tanimoto similarity is sensitive to compound size: IIRC the scores for two 150Da that are one edit distance apart is 0.6. When dealing with fragment mergers this may be the case.
SmallWorld uses graph edit distance: however the wizardry that Roger Sayle & crew have done is impossible to replicate. The open methods are really slow and optimisation gets only so far —there is some mathemagical trick that goes on that I don't know, way more than just an octree hack. The neural network methods are not great.

A third approach would be pharmacophores.
I have had good results with judicious OpenEye ROCS use.
Open would be Pharmit server by David Koes, but I don't know if there's an API.
The caveat of some pharmacophoric distance metrics (e.g. USRCAT) is that aligning compounds by the principal moments of inertia is very sensitive to substitutions for fragments.