Presentation Abstract, exactly as submitted to the Conference site:
Though NMR spectra contain a lot of information about molecular structure, the link between an NMR spectrum and all compatible molecular structures is neither simple nor unique. It is still often mandatory to make use of a chemist's prior knowledge to reduce the number of 'admissible' molecules. With increasing sizes of the molecules of interest, we now encounter ever more complex spectra characterized by (i) incomplete spectral information, (ii) strong coupling, (iii) multiple matching spin systems and (iv) multiple molecules matching each spin system. The result is that interpretation of NMR spectra is again becoming an issue. nD and 13C spectra can help to resolve some of these problems but, due to longer acquisition times, they are less fit for high-throughput operations.
The standard approach calls for an expert equipped with tools such as spectral prediction, simulation and fitting programs. The art of predicting spectral parameters from a molecular structure has finally reached practical usefulness, but its natural partner, the approximation-free spectral simulation, though it has reached maturity many decades ago, is still limited to a dismally small number of coupled spins (~12). Consequently, simulation nowadays calls for approximate algorithms involving novel spin-system fragmentation techniques. Parameters fitting methods are also in evolution since, due to the enormous number of transitions involved in present-day problems, the old Laocoon-like approaches are ruled out in favor of algorithms based on interval functions.
The time-honored style of proceeding from a molecule to the spectrum is also becoming inadequate because it relies on expert guessing. We must face the fact that the starting point is not the molecule but the spectrum - plus a set of a-priori ideas about what molecules might be acceptable. An artificial intelligence should therefore proceed from the spectrum to a molecule, taking into account any prior knowledge, and come up with a set of probable solutions. Such a perspective will profoundly affect future algorithms for computer-aided evaluation and interpretation of NMR spectra. It calls for two novel categories of data-evaluation techniques which will complement prediction, simulation and fitting. One is a multiplet spin-coupling analysis aiming at the enumeration of spin-systems which are compatible with the given spectrum. The other task is finding out all molecules compatible with each of the spin systems. Together, these two items represent the deductive path of spectrum evaluation, while prediction-simulation-fitting embodies the more traditional verification path.
We will discuss the inroads accomplished by our group along both of these paths.
Addenda which did not fit the prescribed form:
Collaborators at Mestrelab Research who participate in this project:
- Santiago Dominguez, coordinator
- Nikolay Larin, mathematician
and others (please, see the last slide)