Sunday, September 28, 2014

PM6, DFTB, GAMESS, enzyme design, protein-ligand binding, and protein structure validation

The publication of this paper got me thinking and this blog post is an attempt to organize my thoughts.

A while back my group developed a method for high-throughput calculation of the effect of mutations on enzyme barrier heights.  The newly developed PM6 method was key to this: it was fast enough and predicted the right reaction mechanism, which AM1 and PM3 did not.

Around that time Mike Gilson and Stefan Grimme published studies were they predicted good absolute binding energies of host-guest complexes using hydrogen bond- and dispersion-corrected PM6. Again PM6 was fast enough to compute the vibrational frequencies needed for the free energy correction.

Parallel to this we have been working in protein structure determination using chemical shifts.  I am starting to suspect that the best way to validate the predicted position of key side chains is to compute their chemical shifts ab initio.  The first step this process might be a complete or partial geometry optimization using something like PM6-D3H+, followed by an QM/MM-like ab initio chemical shift calculation.  Or perhaps a similar check of all side chains.

Using PM6 in MOPAC gave all sorts of problems. At that time the only other option for PM6 was Gaussian, without dispersion and hydrogen bond corrections. So the only option was to put PM6 in GAMESS, or more precisely:
1. implement PM6
2. implement dispersion and hydrogen bond correction
3. interface to PCM
4. linear-scaling FMO implementation.
5. interface to QM/MM?

+Jimmy Charnley Kromann has done 1-2 and +Casper Steinmann has done 3, both for elements up to F. Heavier elements require d-functions.  We have the d-integral code.

Now Nishimoto, Fedorov, and Irle have implemented dispersion-corrected DFTB and FMO2-DFTB in GAMESS, leaving the implementation of a hydrogen bond correction and interface to PCM.

Where does that leave us?
Is DFTB as good as PM6?
For binding and frequencies the answer appears to be "yes" (see for example Table 5 and Table I). For TSs I'm not really sure, but perhaps a little better (Table 4).  We should definitely test this on our little database.
One worry is that DFTB is available for fewer elements than PM6.

What's more work: implementing PM6 for the remaining elements, and d-integral PCM implementation, and FMO2 or implementing hydrogen bond corrections and the PCM interface for DFTB? 
I think the latter.  However I should definitely ask the authors if they have similar plans and when we can get our hands on their code.  Should we decide to go ahead with FMO-PM6 this paper will be very helpful.

What about QM/MM?
AMBER now has dispersion and hydrogen bond corrected PM6 and DFTB, complete with interface to a continuum solvent method and, of course, the AMBER FF. Do we really need to treat all the atoms with SE-QM?  If not, we could go straight to applications!

One problem is that we have no real experience with AMBER. How easy is it to use? Would we get the help we need if we get stuck?  How easy is it to modify the code?

There is really only one way to find out.

This work is licensed under a Creative Commons Attribution 4.0