QSAR - Summary: Quantum Chemistry in Drug Design

Dear Colleagues:
Attached is the summary of responses for my query. The original query and the 5 responses follows:

Original query-------------
Dear Friends:
While several QSAR related techniques and methodologies are appearing in drug
design Journals, very few talk about the more accurate quantum chemical
methods in drug design arena.

* What are the bottlenecks for the quantum chemical methods to get into the
area of drug design.

* For small molecules QC methods plays greater role, but for handling drug-
like molecules and handling several thousands of compounds, QC methods do not
see the limelight (correct me if i am wrong). Is the CPU-intensiveness alone
is the reason. Or is there any some conceptual gap in this. [ I hear someone
saying CPU-intensive is the reason and one has to wait for months to get
results ]

* Based on your experience/insight, can you think of some timeframe, say 5
years, 10 years down the line, Quantum chemical methods would play a major
role in the area of lead identification/optimization, or would you
say "prediction of future is difficult!".

I look forward to reading your views. Thanks.

S. Parthiban
Jubilant Biosys Ltd.
http://www.jubilantbiosys.com

Summary of responses:---------------------------

From: "Dr. N. SUKUMAR" <nagams_._rpi.edu>

The CPU bottleneck has certainly been a major factor, but I believe a
second factor is the descriptors commonly employed in drug design. If all
one uses in modeling are molecular-geometry-derived descriptors, atom
counts, topological descriptors and electrostatic potentials, then it
hardly seems worthwhile performing accurate quantum chemical computations,
especially in view of the enormous computational overhead. Ab initio
cmputations, however, can generate a lot more information at a fundamental
level, derived from the molecular wavefunction or electron density
distribution. There are a few research groups (ours among them) that have
investigated the use of electron-density-derived descriptors in drug
design. In the Transferable Atom Equivalents (TAE) method, first introduced
by Curt Breneman, we employ besides electrostatic potentials, electronic
kinetic energy densities, the Laplacian distribution introduced by Bader,
Fukui's function and Politzer's local average ionization potential. The
distributions of these electronic properties on the molecular van der Waals
surface (binned as histograms or encoded as wavelets) are used as
descriptors. These electron-density-derived descriptors have found success
in a number of applications, especially when used in combination with other
traditional descriptors. For small datasets of small molecules,
electron-density-derived descriptors can be readily determined from ab
initio computions, but for large pharmaceutical datasets and for
macromolecules, these descriptors can still be computed from an
atomic-fragment-based approach using the theory of Atoms In Molecules. This
is done in our RECON program, which employs atomic descriptors computed at
HF/6-31+G* level and is available for download from our website. Typical
CPU timings for RECON on a 1.7GHz Intel Pentium under linux are about 90
sec.for a set of 25 proteins and 7.5 min.for a 42,689 molecule dataset from
NCI -- comparable to times for computing topological descriptors. So I
would have to say that for such applications, CPU is no longer a limiting
factor.

Our protein chromatography studies are published in Mazza, et al,
Anal.Chem. 73, 5457-5461 (2001) and Song, et al, J.Chem.Inf.Comput.Sci. 42,
1347-1357 (2002), while the drug design applications are in various stages
of going to press and in press.

Dr. N. Sukumar
http://www.drugmining.com/
Rensselaer Department of Chemistry

From: "Patrick Bultinck" <Patrick.Bultinck-$-rug.ac.be>

Dear,

This is a very interesting question indeed. And as a matter of fact, there
is now in press a volume "Computational Medicinal Chemistry and Drug
Design" in press edited by two scientists working in pharmaceutical
industry and one professor of Utrecht University and myself. In this volume
a number of quantum chemical techniques (semi-empirical theory, wave
function theory, DFT, QM/MM, accuracy and applicability of methods, ...)
are described in quite some detail by eminent contributors.

Having some experience with computational medicinal chemistry focussing on
quantum chemistry, I would say that there are a number of different reasons
for quantum chemistry not always breaking through in drug design at the
speed one might expect. Here is my 0.02 eurocent worth :

- CPU demands are only a part of the problem but an important (and
expensive) one. This does not necessarily mean that drug molecules are too
big. With current codes and cpu power, one could handle several hundred
atoms with say Hartree-Fock. Moreover, linear scaling techniques allow us
to go even further with DFT, Hartree-Fock, local MP2, ... One must also
realize that many medicinal molecules are not so very big. After all, we
would like to get them in the blood and tissues, and so usually drugs are
not so big. Also remember that quantum chemistry is most often used to
design ligands and candidate-drugs, rather than actual drugs. QC plays a
role in the development steps. The drug is the thing you buy at your local
pharmacy, but this undergoes still quite a number of steps when leaving the
QC desktop (think of optimization for synthesis, clinical testing,
solubility, ...). The CPU problem will often arise when more accurate
techniques are needed, say some correlation level calculation is needed.
But a different problem is that sometimes the QC method does not scale too
dramatically, but the amount of data is too big. Think of some virtual
screening guy walking in and asking to calculate 10^6 molecules...

- There is sometimes a problem with what the qc results would mean. Suppose
you work out a conformational analysis on some nice level of theory and a
good basis set. What does this mean from a medicinal point of view ? Pretty
often, one observes that a critical structure known as the bio-active
conformation is not known. This then yields (sometimes) the conceptual
problem of "what do these results mean ?". In very simple student terms, I
would compare this to a simple equation A=B, and you need to solve for A
but do not know what B is.

- There are some problems related to the fact that the human body is not
composed of solvent free molecules under a wealth of approximations
(harmonic oscillator, rigid rotor, ideal gas, ... to name a few). This
means one needs to have the quantum chemical methods simulate in a better
way actual "drug circumstances". One needs to model solvents and
solute-solvent interactions for example. Many methods have been developped,
and conceptually the best and simplest one is explicit solvent modelling.
One then is confronted again with cpu time making this model often useless
in practice. Other models may also give rise to extra CPU demand, or may be
too crude in their approximations.

- A fairly social reason is sometimes (!!) the reluctant attitude of
(organic) chemists with respect to quantum chemistry. Although there
clearly are chemists which are bright in organic synthesis and quantum
chemistry, often quantum chemistry is still not considered on an equal
footing as the synthesis department. If then you have had to struggle to
get two quantum chemists in your industry, you can not expect these people
to make such a progress that they can easily convince the 200 other R&D
chemists that they will open the way to a new era in drug design.

- There are still some conceptual problems to be solved, or rather : there
still are a lot of conceptual problems to be solved. As a quantum chemist
we want to retrace things to wave functions/electron density as much as we
can. These are the all determining properties, and so we would like to
derive even e.g. QSAR models from these properties. Still much work is
needed in this, but there are advances all the time. The field of QSAR is
in fact a good example. QSAR is often done from known or tabulated (and
rule or fragment based combinations of) properties of a molecule, quite
often experimental properties. Quantum chemistry has entered this field by
allowing the calculation of properties which may not be accessible to the
experimentalist. Mati Karelson (also contributor to our volume) has written
a chemical review on this topic, describing many observable and
non-observable quantities of use in QSAR. Still deeper in the application
of QC in QSAR is the field known as quantum-QSAR. This field, largely
developed by Carbo-Dorca (http://iqc.udg.es) derives QSAR models from
purely quantum chemical ideas. Such an approach is quite interesting, but
often one needs to clear some "simple-looking" things like how to express a
thing like molecular similarity (the basis of quantum QSAR). This is a very
exciting area which will probably make it to the qc desktop in relatively
near future.

- Experimental techniques are being developed that make intensive use of
QC. As an example one can consider some spectroscopic techniques. Some
techniques give spectra which can hardly be interpreted "on sight", and
require some QC predicted spectra to decide what of the possible products
is actually present in the solution.

Now here is my 0.01 eurocent worth opinion on what way QC may break through
in drug design.

- With even increasing cpu power, and current developments in algorithms,
concepts etc., we will no doubt see an increase in the use of QC. My first
computer (when I was 11 years old and very chemistry ignorant) was a
Commodore 64. I am pretty sure you would have a hard time implementing even
simple models on these machines to work at a speed that would convince drug
design chemists. Now look at where we are with PC's. A good PC can do quite
some QC work nowadays. This evolution is likely to continue still, so the
problem of CPU requirements will reduce (that is: if we keep considering
similar molecules and similar QC models).

- There are exciting new areas that may contribute a lot to the use of QC
in drug design. Personally I am quite fascinated by conceptual DFT. These
may yield reactivity descriptors that can help you interpret/predict
reactivity orders and many more. Such areas may even be implemented
sometimes in highly efficient ways. We have recently published a number of
contribution on Electronegativity Equalization aimed at use in drug design.
This method may yield many reactivity descriptors for about a million
molecules/hour on a Pentium III machine. Such advances are bringing QC to a
speed that e.g. the virtual screening guys can live with. Another example I
work in with Carbo-Dorca is the field of molecular similarity and quantum
QSAR. Quantum QSAR models are working (not always) as good as classical (2D
or 3D) QSAR, but are based on purely electron densities.

So I think we are facing a nice future for QC in drug (ligand) design, and
it is already there. But on the other hand, it is not yet on an equal
footing as the many other aspects of drug design, and it would be naive to
think it will be next week. I am, however, sure that it will continue to
develop at an ever increasing speed.

Best regards,

Patrick Bultinck
Quantum Chemistry Group
Ghent University
Belgium
 

From: "Dr. Andreas Klamt" <andreas.klamt,,cosmologic.de>

My view of this topic is as follows: I am afraid, that realistic modeling of drug receptor interaction is still too demanding
for quantum chemical methods, because usually the enzyme is too large, and in addition a lot of conformations and many
compounds would have to be sampled.

The situation is different if you consider the ADME part: Here we can do good calculations of solubility and many kinds of
physiological partitioning properties quite efficiently using DFT methods. The advantage compared with force field models is
that you can get much better insight into the real physics using quantum chemistry. For examples see:

http://www.cosmologic.de/water_solubility.html

or
"COSMO-RS: a novel view to physiological solvation and partition questions", Andreas Klamt, Frank Eckert and Martin Hornig,
 Journal of Computer-Aided Molecular Design 15, 355-365 (2001)
and
"Prediction of aqueous solubility of drugs and pesticides with COSMO-RS", Andreas Klamt, Frank Eckert, Martin Hornig, Michael E.
Beck and Thorsten Bürger, J. Comp. Chem. 23, 275-281 (2002)

From: "Peter Gannett" <pgannett|-|hsc.wvu.edu>

I would say the problem is more complicated than just the amount of time required for a QC calc/compound. For example, you can optimized a compound geometry all you want but it is not necessarily the geometry adopted when the molecule is in the active site of an enzyme so there is not much use. Second, there are simple (minded) methods that clearly ignore a large number of interactions and still come out with useful information. CoMFA is an example of this. A rather simple set of compounds (training set of say 20 compounds) provides you with a reasonable ability to predict how to modify/improve on a drug's activity. It has proven to be a fairly powerful method though computationally, it is very inexpensive. So, it think the bottom line here that QC will not play an important role until it can be demonstrated that there are compelling reasons to implement it.

Pete Gannett

From: "Leif Norskov" <lnl a novo.dk>

Dear S. Parthiban.

In my opinion one simply cannot calculate properties that are
directly relevant for drug design by quantum chemical methods.
Waiting 5 or 10 years wouldn't change much.

However, there may be specific cases where one can find (empirically)
that there is a correlation between biological activity and
some electronic property such as a HOMO-LUMO energy difference.
In such cases QM can already today be useful.

Disclaimer: My scepticism towards quantum chemistry has prevented me
from ever trying - the opinion expressed above is pure fiction.

Best regards,

/Leif Norskov
 Novo Nordisk A/S
 LNL++novo.dk

From: "Jeremy R. Greenwood" <jeremy++compchem.dfh.dk>

Hi Parthiban,

Interesting questions.

> While several QSAR related techniques and methodologies are appearing in drug
> design Journals, very few talk about the more accurate quantum chemical
> methods in drug design arena.
 
Few, but some. I do it because I'm interested in fine-tuning hydrogen
bonding to substituted heteroaromatics, for which QM on model systems
helps develop SAR. And for philosophical reasons -- I look towards
the future and can see potential in combining the pure Platonic rationalism
of QM with the brute force of consumer-driven electronics. :)

> * What are the bottlenecks for the quantum chemical methods to get into the
> area of drug design.
 
In a way they're already involved since they're used to help
paramaterise forcefields.

Mostly I'd say it's the time and effort involved (especially if it's
industry) and there's some tradition/conservatism within academia as well
which slows the process. It takes time to introduce QM into undergraduate
pharmacology courses, it takes time for a critical mass of younger
quantum-enabled chemists to leak from theoretical chemistry departments
into the drug design arena and replace those for whom QM was never
an option in the past.

It seems that a lot of the best computational chemists are heading more
in the direction of e.g. inorganic chem / materials science in search
of harder problems, rather than scaling up the theories which work
well for small systems composed of 1st and 2nd row elements
in order to apply them to biological systems. Scale-up is less
glamourous; more engineering than pure science.

Then there's the fact that a lot of synthetic chemists are more
comfortable with classical concepts, and a lot of drug design is
ultimately done on paper by synthetic chemists. Not much gets
made without having them on board, and that takes a generation and
culture shift.

> * For small molecules QC methods plays greater role, but for handling drug-
> like molecules and handling several thousands of compounds, QC methods do not
> see the limelight (correct me if i am wrong). Is the CPU-intensiveness alone
> is the reason. Or is there any some conceptual gap in this. [ I hear someone
> saying CPU-intensive is the reason and one has to wait for months to get
> results ]
 
Currently I'd say it's mostly the CPU-intensiveness,
then maybe the fact that few interfaces are designed for it and
few codes are robust enough to handle e.g. input from 000's
of structures from MM from Corina.

> * Based on your experience/insight, can you think of some timeframe, say 5
> years, 10 years down the line, Quantum chemical methods would play a major
> role in the area of lead identification/optimization, or would you
> say "prediction of future is difficult!".

If Moore's Law holds for two more decades as it is predicted to do,
barring global economic meltdown, we can expect around 100 times
the current computing power per square inch in 2012, 10,000 times
the computing power in 2022. Still not good enough for full ab initio
treatments of whole enzymes with current methods, let alone full QM-MD
for binding.

But for e.g. a QM ligand in a MM(-MD) cavity, with linear scaling DFT,
more robust codes, new functionals -- quite do-able for a series,
or for developing excellent tailored 'scoring functions'.

And for ligand-only studies, for the ligand that takes a few days
now, you will be able to do your 00's or 000's of ligands accurately
in the same timeframe.

More likely to be in lead optimisation than in lead identification
or scaffold hopping -- when it comes to the millions/billions
of real compounds or enormous combinatorics of virtual compounds,
there's plenty of room for improvement and more to be gained by
improving MM & scoring functions than by burning cpu time on QM.

As with any new technology, we can say what kind of power will
be available with some confidence, but not how people will
choose to use it, so we can expect new kinds of applications
of the theory and the hardware that we haven't thought of yet.
I'll perhaps go out on a limb and say I think we can expect to
see a lot more high-end QM starting to creep into drug design
in the next 10-20 years, one way or another. For some kinds of
problems, it won't prove useful, and for others it will come to
dominate.

> I look forward to reading your views. Thanks.

(I look forward to hearing what the rest of the community thinks).

Jeremy
----------------------------------------------------------------------
Jeremy Greenwood jeremy_+_greenwood.net
Department of Medicinal Chemistry bh +45 35306117
Royal Danish School of Pharmacy fx +45 35306040
Universitetsparken 2, DK-2100 Copenhagen, Denmark ah +45 32598030
----------------------------------------------------------------------

 
Received on 2002-12-14 - 05:26 GMT