Not too much time this week to dwell in detail on papers, but here are some that I found interesting:
Richard Friesner and his colleagues at Columbia investigate the entropic contribution of arginine side chains by measuring local conformational dynamics by NMR spin relaxation. They compare these results to detailed MD simulations to infer that the aliphatic side chain maintains enough entropy for the charged amino group to fruitfully engage in rigid salt-bridge formation, thus minimizing the entropic penalty.
On the other side of the Atlantic, Angelo Vedani and his colleagues in Basel have derived a QSAR model for psychotropic drugs binding to the glucocorticoid receptor using their programs Yeti and QUASAR. The GR is a nuclear hormone receptor that can modulate a variety of key processes by binding small molecules. The model was validated on 110 compounds representing 4 chemical classes.
An interesting paper on the conformation of a common and important consensus tripeptide motif in a glycoprotein compares measured IR intensities and frequencies using a technique called ion-dip IR spectroscopy to frequencies calculated using high-level ab initio theory (MP2/6-311++G**). The Oxford authors do the comparison for the wild type motif as well as for two mutants and ask a "Why nature selected this specific motif" kind of question by looking at the viability of the resulting conformations. It looks fine, but I am a little skeptical about all those gamma turns that they see, a common artifact of inadequate treatment of solvation in both force field and ab initio approaches.
And again on the medicinal front, a group in Europe does some fragment-based computational design and find a PPAR agonist. Detailed knowledge of protein structure aids in explaining why one bioisostere works and another one does not.
That GR paper... So you give a program a known binding site, list of essential residues and 88 structures with given affinities mostly within narrow range (lets set aside the dubious practical value of converting IC50s to Ki from a set of disparate experiments). Then you give the program 22 very similar structures and find that it predicts the affinity in 50-70% cases. And when you do this with two programs, you find that they agree with each other on average roughly to an order of magnitude.
ReplyDeleteAs a crystallographer who is skeptical of predictive power of "paper biochemistry", I find it inexcusable that a crucial experiment has not been done and not even hinted at in the paper: Take two intelligent and educated humans in place of computer programs. If human output will have the comparable predictive power as that of computers (and at least in this particular case I belive it will be), it would mean that all the high-tech hokery-pokery with MD, flexible docking and what have you is simply worthless excercize in using CPU cycles. Has this been done? If not, why?
I largely agree with you. I am not a fan of QSAR, and a lot of the conclusions seem could be validated by experienced organic chemists too. I also second the suspicion about using IC50s derived from Ki values from disparate experiments. Here is a nice article about whether QSAR is really of use in Drug Discovery:
ReplyDeleteIs QSAR relevant to drug discovery?
Arthur M Doweyko
IDrugs 2008 11(12):894-899
Address
Bristol-Myers Squibb Co, Research and Development,
Computer-Aided Drug Design, PO Box 4000, Princeton, NJ 08543, USA
Email: arthur.doweyko@bms.com
The concept of quantitative SAR (QSAR) is inherently imbued with an expectation of predictivity, novel insights and the generation
of useful hypotheses, particularly as applied to the drug discovery process. However, even recently developed QSAR models often appear
to be flawed, characterized by mediocre predictive power and undecipherable descriptors. As a result, users may be able to derive only
a vague notion of which molecular features are correlated to activity. Consideration of several precautions is necessary to attempt to
circumvent the misuse and misunderstanding of the QSAR technique. Issues related with QSAR include an erroneous association of
correlation with causation, the close relationship between large numbers of descriptors and the effect of chance factor, the misuse of the
'leave-one-out' paradigm, and finally, the QSAR enigma, wherein the predictivity of a model is not necessarily a measure of a model's utility.