I am now back in the US after a week in Europe and late at night, with a disrupted sleep pattern and looking for something to soothe me to sleep I have been wandering through the rest of the NPC Browser data set [1,2,3] looking for more patterns in the data to see if I can come up with some general advice and cautions about how to build datasets of chemistry. I already have my “conclusion” in my head as to the best advice I can give to any government organization, and others, who are trying to build chemical databases but I will save this for later in the week. For now I want to highlight some of the issues to be careful of. Tonight’s focus is “structural depictions”.
Achieving high quality algorithmic 2D structural layout is difficult across large databases is difficult. All of the cheminformatics vendors have layout tools whereby a structure can be “cleaned” so that the layout on the page is visually appealing. If you have ever used any of the cleaning tools you might have discovered that they fail dismally with certain structures and you have to perform a layout manually. You might have discovered that when you clean the structure that stereocenters flip (something that should NEVER happen but, believe me, does!) However, there are some good tools that are available. OpenEye provide layout algorithms as part of their cheminformatics toolkit and it is certainly one I have experience of.
In any case, when creating a database of chemicals it makes good sense to use algorithms for layout rather than accept what is submitted. PubChem and ChemSpider, with tens of millions of structures, have to use algorithmic cleaning but if you have a small database it won’t take long to visually inspect and spot errors very quickly. There are many examples that SHOULD have been caught with the recent NGC HTS screening set.
The worst one of these that I found while simply browsing was that associated with “Silidianin” as shown below.
Here we see the structure of a bare hydroxyl group, but without a negative charge. However, the CAS number and various synonyms certainly don’t support the compound being a hydroxyl group. Indeed, this is a “0D structure” with all xy coordinates set to 0,0 and if the structure is cleaned in a drawing package then you see the structure shown on the left below. While it is the connection table for Sildianin it does not have the appropriate stereochemistry for Sildianin encoded into the structure with wedge and dashed-wedge bonds. The structure on the right has that shown.
I think you would agree it is more aesthetically pleasing and does communicate the bridged nature of the compound and carries the stereochemical information. However, there is something wrong with even this picture. Can anyone say what’s wrong?
Accurate structural representation in any database takes time, effort and, often, a skilled and careful eye to get right. Clearly the 0D structure is simply wrong and should have been caught. There are other offenders in the database that should have been caught also as shown below. There are LOTS more.
Building high quality chemical databases certainly is tricky…and can be very time-consuming to solve all of these issues.