Dictionary Lookups and Optical Structure Recognition Versus Structure Drawing. Which is Less Error Prone?

Luqidcarbon has put up a recent blog posting about the speed by which he/she can draw structures in ChemDraw and asked for challengers. PRM has commented in Chemical SpeedDrawing. The challenge is outlined below…

A few people took the challenge and commented on the blog posting. Me too, but I cheated.

I’m of the opinion that for “KNOWN” structures the best thing to do would be to “look it up” in an appropriate reference collection. Now, I could go to ChemSpider for all of them (or PubChem) but I can predict that because the curation process is not as advanced as I would like it to be we will have issues with multiple structures with the same connectivities but with different stereocenter definitions (incomplete in general). Just to check I did so…

THC is there but a search on THC turns up three structures. Since we’re trying to reproduce the right one we can visually inspect and download. Takes me about 20 seconds to locate, inspect and download into my drawing package

Penicillin G…little more complex because there is more stereochemistry to inspect. ChemSpider gives 5 hits, 1 is now curated out since it’s a dimer. 1 is a potassium salt and the other three have the right connectivities but different stereo displays…easy to download one and change the stereo to match the structure of interest..about 30 seconds.

(+)-Discodermolide gives 6 hits..all differing in their stereochem. Nevertheless….I can download the structure and edit the stereo in about half the time. The correct structure on ChemSpider is here, with the InChIKey AADVCYNFEREWOS-USOWQXFZDY.

What’s the message?..it MAY be faster to download the “framework” and match the steroechemistry in the drawing you are trying to copy rather than draw from ground zero. Why? Notice the structure of discodermolide drawn by liquidcarbon is WRONG as far as I can tell….see below for the detail but there is a methyl instead of a hydroxy on his/her structure.

Rather than use ChemSpider I went to a HIGHLY CURATED data set of names and structures called the ACD/Dictionary (Don’t forget I used to work at ACD/labs so am biased to this product since I managed it. That Ssaid, it is VERY well curated…thousands of hours of work!). It’s included with a copy of the commercial ChemSketch. There are alternatives on the market, for example ChemDraw could likely reverse all the names above to the appropriate structures but I don’t have the software to check. But, bottom line, there are faster ways for “known structures” than redrawing.

Fifteen seconds TOTAL is what it took me to put the structures for all three on a page. Oh, and it had the OH in place as shown. What’s this error in structure worth to you in terms of time and proliferation? A chemist costs at least $150 per hour. If the incorrect structure proliferates from misdrawing whet will it cost a company??

A drawing package like this is a few hundred bucks for corporate and just over a hundred bucks for academia (see here). By the way…I am NOT on commission…but I’ll take COFFEE!!!! ChemDraw has nam-structure conversion too as does ACD/labs and they can all convert common names to structures..with differing degrees of accuracy.

So, forget the redrawing when there are other solutions for “knowns”.

What about for unknowns…when you are looking at a NEW structure in a paper, or on a website and it’s complex …should you redraw? NO, NO, NO.

I have blogged previously about software for converting structure images to real structures (connection tables). Today I had the pleasure of sharing space with Stephen Boyer from IBM who originally patented scan and convert for chemical structures. I am interested in the possibility of using BATCH processing of structure images to convert structures from Open Access articles and adding them to ChemSpider. We have access to 50,000 Open Access articles…and lots of structures therein. If we can batchwise convert then we can populate the database and index by structure.

With this in mind I have started working with Simbiosys to test and validate their CLiDE software. My first tests give pretty good results and I am only in “learning mode”. I have just taken delivery of a beta version of their latest software and am yet to install so I asked them to grab the image from the webpage and convert. the results are below.

The conversion to connection tables lost a couple of stereo bonds BUT the connection itself is there and it is EASY to put the stereo back in. Plus, it lost a proton on THC but that is easy to add back in. That’s pretty impressive!

Now imagine a work flow where all you do is open up a PDF file, activate CLiDE, select the structure with a rectangular drag and it recognizes and converts to a connection table and all you do is check it and ‘touch it up”. That 150 seconds of drawing time that liquidcarbon talks about could be reduced to about 30 seconds AND, hopefully, no hydroxyls will be lost.

I am definitely putting effort into CLiDE validation at present as it appears to be well-advanced but not yet perfect (hey, it’s software!). I WILL be supporting their efforts to make it better…I want to use it for my own selfish needs and I judge it is performing well. I’ll keep you informed.