How the experiment may impact the data

Here’s one to file under “I am still trying to get my head around it”.

Back in April at the CDD community meeting Christopher (Chris) Lipinski presented some slides looking at kinase selectivity and the relationship with ligand efficiency. There seemed to be a general trend that more selective compounds had a better ligand efficiency. My colleagues at CDD have been digging deeper into this and will present a webinar Wed, Oct 22 from 2-3pm ET at which Chris and Matt Soellner will debate Entropic and Enthalpic Propensities Inherent in SBDD and HTS.

Now my involvement has been pretty limited to thinking of some interesting datasets to compare. Obviously my personal bias is towards the neglected diseases and anything that is in the public databases. For one I was interested to see how the >1000 whole cell Mycobacterium tuberculosis (Mtb) hits coming out of high throughput screens compared with ligands from structure based drug design studies (SBDD) for which there are examples in the PDB. A measure of enthalpy of the SBDD hits suggested it was higher than for the HTS hits. Because several datasets have been released on antimalarial HTS hits we can do the same comparison with SBDD hits.

Without trying to give too much away I would say some of the slides I have previewed were very interesting. Of course most will want to hear about the kinase data but lets think about what other questions could we ask. SBDD by its nature is trying to optimize the fit of compounds into a target, simplistically it is trying to get good interactions. Phenotypic HTS is not bothered about that, key determinants of activity are getting the molecule into the cell and then shutting down some target/s. So hydrophobicity is predominantly driving whole cell activity for Mtb, as we see many of the hits have a higher calculated logP (using whatever method you decide), although other properties may also be key.

So fundamentally depending on what kind of experimental approach we use to get Mtb active compounds we are biasing towards compounds with different physicochemical properties. We have come full circle. Target based approaches to antibacterial drug discovery have been a failure because one they found few hits and two the few hits did not have whole cell activity. It seems obvious now but target based drug discovery is really finding a needle in a haystack, trying to get very specific interactions while whole cell approaches ‘just’ need to get the compound in and perhaps have OKish affinity for one or more targets. Maybe the latter represents more of a complete system effect (more targets to interact with vs a single target).

So what does this say about our efforts using computational approaches to find compounds active against Mtb? Will they also have some of the same issues inherent in HTS and SBDD? For example docking molecules in a crystal structure as part of SBDD is going to drive towards very specific interactions, and if the method and scoring functions are poor then the hit rate will be very low. Machine learning methods are going to learn from just the mass of data you give them. So if you feed in whole cell data all you are going to do is basically replicate the physicochemical properties that allow you to get compounds into Mtb and hit a whole array of potential targets. Is there some middle ground here, a hybrid approach?

Perhaps running compounds through whole cell assays and just feeding those hits into SBDD as starting points? Then followed by feeding the resulting SBDD designs/hits into whole cell assays to ensure that there is a balance between specificity and ability to get into the cell. Perhaps this iterative approach would be more efficient computationally as a pipeline where the known whole cell hits are fed into docking against as many Mtb structures in the PDB as possible and those that have good scores would serve as a starting point for design.

Another question you could perhaps ask is are the compounds that we want to avoid in HTS (like PAINS) different in some way? Would they stand out from real HTS hits and real SBDD hits. Is a PAIN found by docking more useful than a PAIN found by HTS? Do they have different enthalpy scores?

Well I am sure the webinar will have others asking questions too. Its certainly got me thinking.



Leave a Reply

Your email address will not be published.

You may use these HTML tags and attributes: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>