From Platts (+ at +) cardiff.ac.uk Thu May 1 08:39:58 1997 Received: from crane.cf.ac.uk for Platts %! at !% cardiff.ac.uk by www.ccl.net (8.8.3/950822.1) id HAA17099; Thu, 1 May 1997 07:58:32 -0400 (EDT) Received: from thor.cf.ac.uk by crane.cf.ac.uk with SMTP (PP); Thu, 1 May 1997 12:53:27 +0100 Received: from localhost (sacjap #*at*# localhost) by thor.cf.ac.uk (8.8.5/8.6.12) with SMTP id MAA12637 for ; Thu, 1 May 1997 12:57:54 +0100 (BST) Date: Thu, 1 May 1997 12:57:54 +0100 (BST) From: James Platts To: chemistry.,at,.www.ccl.net Subject: calculation of chemical shifts (summary) Message-ID: MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Hi All, Many thanks to all who responded to my original post... > Sorry if this is a FAQ. A colleague wants me to predict the chemical >shift of a nucleus in a large-ish molecule (50+ atoms). All the literature >I have seen on such calculations indicate that one needs to use a large >basis set and preferably some estimate of correlation, which is out for a >molecule of this size. > Does anyone have any ideas how I could get a reasonable estimate of this >number? Would a semi-empirical treatment be worthwhile? The general consensus is that although "standard" correlation methods are clearly out of the question, DFT could well give us the accuracy we need (Salahub particularly has worked on this it seems). This is the approach I intend to follow, though my supercomputing time is already running low :-( . Also, the quality of the geometry used for the NMR calculation is crucial - maybe even to the point where one should use a better basis set for this than for the final run. "Locally dense" basis sets, which aim to saturate the region of interest and treat the rest of the molecule less accurately, sound like a good idea to me. Finally, empirical schemes based on additivity got a couple of plugs. Free demo's from http://www.acdlabs.com or http://www.acdlabs.co.uk. Here are some of the replies... --------------------------------------------------------------------------- From: Rainer Koch It depends on the molecule. To my knowledge, there isn't a recommended method for all kind of molecules. Usually, you are right, but in a submitted paper, we found that some medium basis sets (6-31+G**) using either B3LYP or HF on a "good" geometry gives better results than extensive basis sets. The systems studied were heterocyclic rings, the errors were about 3 ppm, the program used G94. If you are interested in more details, let me know. I would recommend an MP2 geometry but if you only have a HF/6-31G*-optimized structure (50+ atoms...) you also get reasonable results... I tried some semiempirical packages as well, but their data are not consistent and reliable. --------------------------------------------------------------------------- From: "Ananikov V.P." In most cases with the semiempirical methods you can not calculate correct values of the chemical shift, especially if the molecule contains charged atoms, i.e. N+, O-, etc. The ab-initio calculations are quit successful, but it is necessary to use high quality bases sets( 6-311G, DZV, TZV) and often polarization functions. Concerning your question, I believe, there are two possible solutions in the case of the large molecules: 1) Chemical shift prediction using additivity relationships and empirical correlations. These methods are very fast, the accuracy is about 0.5-5ppm. I am using ACD CNMR/HNMR programs for these purpose. 2) Ab-initio chemical shift prediction based on RPA algorithm. LORG and SOLO methods are implemented in RPAC program. I can send you more information about these packages if you are interesting in. --------------------------------------------------------------------------- From: Steven Creve You could try a DFT calculation, see for publications of Salahub. --------------------------------------------------------------------------- From: Alexander Backes If I understand you correctly, the main problem you have is the largeness of the molecule and the question of finding a quantum chemical method for it which yields reasonable results. In fact, electron correlation methods are NOT out for a molecule of this size. DFT methods are fast and allow you to calculate molecules conatining 50 atoms and more in reasonable time. You have to perform a geometry optimization on your molecule first; the quality of this optimization is very important for further calculation of chemical shifts. So, if you use ab initio methods, and you have this large molecule, I recommend DFT methods. B3LYP or BLYP are common DFT-functionals, yielding very good results; they are implemented in Gaussian 94, for example. The use of basis set depends on the atoms of your system. If there is only hydrogen, carbon or oxygen, for example, a 6-31G(d) or 6-311G(d) basis set is suitable. If there are transition metals or higher main group atoms, you must choose special basis sets. Second, for the calculation of chemical shifts, implementation of electron correlation is indispensable. Again, DFT methods are suitable (GIAO-B3LYP or GIAO-BLYP, for example, also implemented in Gaussian 94). The basis sets you use are dependent on your atoms, as mentioned above. For carbon, hydrogen and silicon we obtained very good results with the 6-311G(d,p) basis set, using GIAO-B3LYP. A semi-empirical treatment of such a large molecule is surely faster than an ab initio calculation. If your system contains the usual main group elements (C, H, O, N etc.), an AM1 or PM3 geometry optimization could yield a reasonable structure, which you can use for the calculation of chemical shifts. But there may be bad results easily. For myself, if I want to get a sufficient chemical shift prediction, I would only use an ab initio method, for both optimization and shift calculation (for the latter ab initio is the only sufficient method, because all other methods predict the shift by using databases). But it's like always in the calculation business: When there is no experience, trial and error... =;-) --------------------------------------------------------------------------- From: Georg Schreckenbach yes, the chemical shift pops up every now and then in the CCL. Anyways, you are right, you will need a large basis set, and you will need, depending on your moecule, correlation effects. For the latter, may I suggest to use density functional theory. It does account for correlation in an approximate way, yet "large" molecules are still feasible. As far as the basis sets are concerned, people have experimented with "locally dense" basis sets, i.e., putting extended basis sets onto the atom of interest plus some neighboring nuclei, and somewhat smaller basis sets elsewhere. I am not up to date though with the pro's and con's of this approach. Besides, make sure to use a distributed-gauge-origin method like GIAO and IGLO, basis set requirements are completely out of hand otherwise. --------------------------------------------------------------------------- From: penk &$at$& cc.vchgroup.de You have been asking for SW support in predicting NMR shifts. Our SW system SpecInfo can predict C-NMR shifts for molecules consisting of up to 100 non-hydrogen atoms (up to 255 in the next version). The prediction is based on a large database with 100,000 experimental C-NMR spectra. You are very welcome to test SpecInfo by sending me the structure either as Molfile or as a drawing by Fax, and I will send you the prediction for all C-atoms (free of charge of course). I garantuee that I will keep the structure confidential and will keep no copies of it. --------------------------------------------------------------------------- From: Doug Fox While the properties are basis set dependent we have done some studies which suggest that there are some reasonable compromises. Notably we have done HF/6-31G* NMR properties on Taxol, 112 atoms, with G94 and get pretty reasonable agreement. We do recommend getting a better structure so perhaps B3LYP/6-31+G(d,p) for the structure optimization followed by a single point NMR at HF/6-31G*. See Cheeseman, et. al. JCP vol 104, pp 5497 (1996) or contact the Gaussian, Inc. office for a reprint. --------------------------------------------------------------------------- From: Victoria Barclay You don't mention if this is proton, C-13, or other. If it's H or C, you might want to take a look at our web site, http://www.acdlabs.com, and even try out our interactive predictor for a portion of the molecule. (Our software goes to 255 nuclei, but not the demos, alas.) --------------------------------------------------------------------------- From: Ernest Chamot You asked about predicting the chemical shift of a nucleus in a molecule too large to treat with a large basis set and correlation. You are correct, in that to get decent chemical shifts the experience has been that correlation is required (even in the geometry optimization). So I agree it doesn't look like you're going to be able to do this by first principles. There are a couple alternatives. Two programs, ACD/CNMR by Advanced Chemistry Development Inc. and ChemIntosh/ChemWindow by SoftShell both offer prediction of NMR chemical shifts based, I believe, on a group contribution method. They actually don't do too badly on a few pathological cases I've checked them with, but there is an occasional miss. The program, VAMP, (a derivative of AMPAC) by Oxford Molecular has a feature to calculate NMR chemical shifts based on a semiempirical calculation. It uses a built-in neural net model. --------------------------------------------------------------------------- From: Matt Dimmic I've been doing NMR calcs for a little while now, and there are a few approximations which you can use to make your calculations less expensive. 1) Use locally dense basis sets. For your center of interest and the atoms in its vicinity, use a large basis set, while for the other 40 or so use a compact basis set. It's my opinion that, unless your system is highly conjugated or delocalized, this should get you reasonable estimates. 2) Try DFT methods instead of, say, MP2. I'm not too strong on the theory, but I believe DFT includes correlation implicitly, and sometimes yields a faster calculation. 3) Examine the shifts relative to each other, rather than absolute or relative to a standard. From the literature I've noticed that absolute shifts are difficult to calculate properly, but (because the environments are similar) the relative shifts should be correlative. While this is probably shaky science on its own, when correlated with other evidence it can be helpful. --------------------------------------------------------------------------- From: Jeremy R Greenwood Have you considered an empirical approach? I was fairly impressed with the new software from ACDlabs: http://www.acdlabs.com