Magic methyls and magic carpets

A few days ago, there was this post by Derek Lowe, reviewing a recent paper on magic methyls and their occurrence and impact in medicinal chemistry practice. They're called 'magic' because, although methyls are relatively insignificant in terms of size, polarity or lipophilicity, the addition of one in a compound can sometimes have a dramatic impact in its potency - much more that it would be attributed to any simple desolvation effects.

More generally, the 'magic methyl' phenomenon pops up in discussions about the validity of the molecular similarity principle, descriptors, QSAR - almost everything in the applied Chemoinformatics field - and belongs to the general class of 'activity cliffs'. 

Methylation is a chemical transformation, and transformations along with their impact on a property of choice can be easily mined and studied using the so-called Matched Molecular Pairs analysis (MMPA). We already have a comprehensive database of all the matched pairs and transformations in ChEMBL, so it was relatively straightforward to extract all the methylations (H>>CH3) recorded in ChEMBL_17 and analyse their impact in binding affinity. (b.t.w., MMPs are coming to the ChEMBL interface soon, so look out for this feature if you are interested in this area).

So, in more detail, I extracted all the H>>CH3 pairs and joined them with their pActivities (Ki, IC₅₀, EC₅₀) against human proteins as reported in the literature (our data validity flags were quite useful in this case). The trick here is to only consider molecule pairs tested against the same assay, so that their respective activities are directly comparable and one can safely subtract one from the other.

I ended up with 37,771 data points - much more than another recent publication that looked at this. Here's how the histogram of Delta pActivity (log units) looks like:

As you can see, the scale tilts slightly to the left of zero, meaning that methylation has overall neutral to negative effect on binding affinity. This is not the first time people see this. There are indeed, however, several examples (~2.3K out of 37.8K, to be exact) of magic methyls with more than 10-fold increase in activity. More about this later.

Some of you will ask: 'OK, but what about the context? - methylation of a carbon, nitrogen or oxygen is not the same'. You're right, it's not. So I trellised the above plot by a perception of context - i.e. whether the methylation happens next to an aromatic/aliphatic C or N or next to an oxygen:

The same trend, more or less, is observed with the exception of the aromatic carbon context, whereby methylation seems to have more favourable effect that expected by the overall distribution. Perhaps that could be explained by introducing torsional and planarity changes, etc. For a more thorough explanation of this, see here. 

Here are some examples of 'magic methyls' in the literature:

The take home message is: Magic methyls, unlike magic carpets, do exist but there are also equally as many, or even more, 'nasty' methyls. However, both of them are just a rather small minority compared to the 'boring' methyls - i.e. methyls with minimal or zero impact on potency.

It's just human nature to remember the few exceptions and outliers and forget the vast evidence to the contrary. However, isolating and understanding such edge cases and black swans is what could make the difference in drug discovery. 

George

Some Gamification of ChEMBL

Here's a small toy app built from the ChEMBL API - really just a technology exemplar - but also pointing towards some interesting potential crowdsourced things that could be done - for example community classification of the drug-likeness of compounds, or the identification of complex toxicophores - the sort of thing where capturing 'expert' tacit knowledge is needed - and the best way to learn is through analysis of examples.



So, have a play, and see how well you do.



There's a small interesting story behind this app - I originally thought we should do a dirty hack (a mapped gif - the shame of it) - but no, the development pixies, did a proper job, and in a very short time....



The first load of the app is slow. So settle down, clear your mind, and get ready to test your skills. The obvious next feature for this is one of those buttons to tweet a score, saying something like - "I got to level 7 on the ChEMBL-brain-O-thon"



Oh, the url for this won't be stable, and may well disappear completely, so be quick and waste some valuable work time now!



jpo

Antibacterial Targets - Evidence for exclusion of targets for which host orthologues exist

One of the classic mantras for the genomics-based discovery of novel anti-bacterials is to ignore targets for which orthologues exist in the host (human) genome, but my hunch was that the majority of antibacterial mechanisms have clear host orthologues (they do, as you'll see below). I haven't come across any really simple papers supporting this dogma in the past, so decided to have a quick look this morning. I am the unwilling host for an oral bacterial infection myself at the moment, but I must stress that the throat above is not mine, but an anonymous one from Teh Interweb!



So, using a book I've just picked up at the ACS in Indy, I went through and did some quick analysis - the prose in the book is great, informal, and very very readable - buy it!

%T Antibacterial Agents: Chemistry, Mode of Action, Mechanisms of Resistance and Clinical Applications
%A R.J. Anderson
%A P.W. Groundwater
%A A. Todd
%A A.J. Worsley
%I Wiley
%D 2012
%O ISBN 978-0-470-97245-8

I went through, and at a drug class level, assigned the distinct mechanisms into three target classes.

1. Orthologue of antibacterial target is present in humans.


2. Orthologue of antibacterial target is absent in humans.


3. Antibacterial acts through a non gene-derived target mechanism.

The counts for these three states are 8, 4, 3 or in a simple graphical form

So, no real great evidence to focus on bacteria specific genes - in fact it's 2:1 in favour of targets for which orthologues exist in the host. The key would seem to be more exploiting physicochemical differences between bacterial and human cells (e.g. acidity), or exploit differences in the binding sites. I guess the dogma arose for a couple of reasons - firstly everyone knowns about penicillins, and secondly, it is an easy filter to apply bioinformatically, and finally it just seems like a perfectly sensible thing to do with respect to elimination of mechanism-based toxicity - and so is often done.



On this latter point, that of mechanism-based toxicity, this can be really important, but remember, a drug dosed to a human is not magically attracted only to relatives of the bacterial target, it will sample and equilibrate across all accessible binding sites of all proteins, and drugs will have side-effects and toxicity related to 1) binding to orthologues, 2) binding to paralogues, and 3) binding to anything else. A nice example of this is the paper from Science earlier this year on the side effects of sulphonamide antibacterials via inhibition of host sepiapterin reductase.



To be clear what I did here. Firstly - I used the chapters in the Antibacterial Agents book to define a class, so the graph can be plotted in many other ways - however, I was interested in distinct mechanisms. Secondly, several antibacterials don't target proteins, but various parts of the ribosome - these are gene derived, so count above as gene products, but they are not proteins (well they are riboproteins). Even if you strip these RNA targets out though (there are 5) the numbers are still not compelling for the need to avoid host target orthologues - the ratio would be 3:4 instead of 8:4.



What are the implications of removing this bacterial specific filter? Well that is more than a quick job on a Sunday morning and two cups of Lapsong Suchong - but my feeling is it might be quite significant.



jpo