Practical Fragments has featured a number of posts comparing
various fragment-finding methods. In some cases there is good agreement, while
in others – not so much. Computational methods can in theory sample the
greatest swath of diversity space: a virtual library can be orders of magnitude
larger than any physical library. In a recent paper in J. Chem. Inf. Model. Gregg Siegal at ZoBio and Leiden University
and Jens Carlsson at Stockhom University and their colleagues compare the
performance of virtual screening with a biophysical method.
The target they chose, the A2A adenosine receptor
(A2AAR) is a GPCR implicated in a variety of diseases. It also has
the advantage of multiple published co-crystal structures with either agonists
or antagonists bound, making it a good candidate for computational screening.
The researchers began by conducting a computational screen
of 500 fragments using DOCK 3.6 against the crystal structure of an
antagonist-bound A2AAR and ranked these according to how well they
scored. Next, the researchers physically screened the same library of 500
fragments against A2AAR using an NMR-based screening method called
TINS (see also here). This resulted in a whopping 94 primary hits, which were
followed up in a radioligand displacement assay to yield 5 confirmed hits with
Ki values ranging from 14-600 micromolar. Happily, 4 of the 5 hits from
the TINS screen were within the top 5% scoring hits identified in silico.
This is satisfying at first glance, but what does it say
about the other top-scoring computational hits? Computational screening
virtually docks fragments in many possible positions, or poses, which are automatically
evaluated. Manual inspection of the top 50 in silico hits showed that, in some
cases, the best poses had desolvated polar groups, which would presumably be
energetically unfavorable. Indeed, identifying the “correct” pose seems to be a
general problem with docking fragments.
But some of the top-scoring fragments looked fine by visual
inspection, so 5 of these were tested in a radioligand displacement assay.
Surprisingly, 3 of these were active, with Ki values ranging from 18-128
micromolar. In other words, these were false negatives in the primary TINS
assay.
Having found hits that had been missed using a biophysical
screen, the researchers then docked 328,000 commercially available fragments
against the target – an exercise that took only seven hours on a computer
cluster. Of the top hits, 22 were purchased and tested in the radioligand
displacement assay, and a remarkable 14 of these were active, with Ki
values ranging from 2-240 micromolar. (I do wonder how much chemical intuition played a role in choosing hits to purchase.)
Interestingly, all of the 14 hits from docking had
respectable ligand efficiencies (LE > 0.3 kcal/mol/atom, with a single
exception). This is consistent with previous fragment docking studies that show that the best results are obtained with the most
ligand-efficient fragments. It’s also a nice feature; after all, these are
exactly the kind of hits you would hope to find, though of course you want to
first filter out any garbage from your virtual library.
This paper provides more evidence that computational
approaches can find fragment hits for GPCRs, at least relatively “druggable”
ones with good structural characterization. It is also a useful reminder of the
importance of using multiple methods, to avoid both false positives and false
negatives.
Finally, if you haven't already voted on your fragment-finding methods, please do so on the right side of the page!
Finally, if you haven't already voted on your fragment-finding methods, please do so on the right side of the page!
Posts
2 comments:
This is very interesting to me because it has always been my understanding that TINS has a very low false rate. I haven't read the paper (yet), but do they have any comment on that?
Below find our responses to some of the interesting questions posed by Dan and Teddy
This is satisfying at first glance, but what does it say about the other top-scoring computational hits? Computational screening virtually docks fragments in many possible positions, or poses, which are automatically evaluated. Manual inspection of the top 50 in silico hits showed that, in some cases, the best poses had desolvated polar groups, which would presumably be energetically unfavorable. Indeed, identifying the “correct” pose seems to be a general problem with docking fragments.
A: We agree that docking of fragments is more challenging than leads or drugs in some aspects. There are typically more ways to dock a fragment into a site and, for reasons outlined in the paper, it may also be more challenging to predict their affinities.
But some of the top-scoring fragments looked fine by visual inspection, so 5 of these were tested in a radioligand displacement assay. Surprisingly, 3 of these were active, with Ki values ranging from 18-128 micromolar. In other words, these were false negatives in the primary TINS assay.
A: Compound 7 was not observed in the TINS spectra, likely due to non-specific association with the detergent micelle which we have reported in the past (Früh et al, Chem. Biol., 17, 2010, p. 881). Non-specific associate of fragments (compounds in general) remains a problem with biophysical screening of membrane proteins, although this can be reduced by using nanodiscs to solubilize the protein. A benefit of TINS is that such compounds are not reported as false positives, although in this case it did result in a hit being missed. Compound 6 exhibited a T/R ratio (a measure of target specific binding) that was just over the cutoff used to define hits and therefore inclusion as a hit or not was somewhat arbitrary. Compound 8 consistently exhibited a T/R indicative of preferential binding to the A2aR, but was not within the range considered to be a hit. TINS, like any ligand observed NMR assay, is sensitive to koff. This is very nicely explained in the review of Bernd Meyer (Angewandte Chemie, 2003, 42, p.864). The typically large koff associated with weakly binding fragments lends itself well to ligand observed NMR assays like TINS. Moreover, the extreme sensitivity created by immobilizing the target in TINS enables the assay to pick up fragments with KD’s > 10 mM. With this sensitivity, few weakly binding compounds miss detection earning the assay a reputation for few false negatives. In the present situation, we can speculate that fragments 6 & 8 dissociate sufficiently slowly that the binding signal is becoming stoichiometric and that is why they are not defined as hits. Particularly for GPCRs where the orthosteric binding site may be a very deep pocket in the trans-membrane domain, it is possible that koff can be a limiting factor. Therefore we strongly advocate orthogonal screening methods.
Having found hits that had been missed using a biophysical screen, the researchers then docked 328,000 commercially available fragments against the target – an exercise that took only seven hours on a computer cluster. Of the top hits, 22 were purchased and tested in the radioligand displacement assay, and a remarkable 14 of these were active, with Ki values ranging from 2-240 micromolar. (I do wonder how much chemical intuition played a role in choosing hits to purchase.)
A: The docking screen was carried out in two steps. In the first, docking was used to exclude 99.8% of the library, leaving 500 (0.2%) candidates for experimental testing. In a second step, we always visually inspect the 500 top-ranked compounds. This is standard procedure in our and many other groups. This step involves several criteria, of which some could considered to be “chemical intuition”: Inspection of receptor-ligand complementarity, consideration of energy terms not included in the scoring function (entropy, ligand internal energy), diversity of selected compounds, and purchasability of the chemicals.
Gregg and Jens
Post a Comment