Why do molecular modelling?

For successful communication of research results, e.g. in a poster at a conference, good-quality pictures of calculated models are now becoming the norm
Modelling provokes you to do good, well-focussed experiments in the synthetic laboratory
Modelling is not likely to be accurate enough to replace synthetic experiments, but it can be very useful in selecting the best experiments to try in the time available

A model is useful if the guidance it has given you, in carrying out your experiments, turns out to be correct, with possibly a few explainable exceptions
If, statistically, the model does not predict the results you find, it is not useful and you should not publish it

To predict the shape of molecules to see whether they will fit the geometry of proteins, etc., and hence have possible medicinal value
Protein structures can be displayed as molecular models, e.g. by Chime, or by most modelling programs
Protein structures, e.g. from crystallography, are available from the Brookhaven Protein Databank

To select, on the basis of energetics, which of several possible intermediates, or reaction routes in general, is most likely to explain observed products in a synthesis

To model NMR chemical shifts, particularly ¹³C or ³¹P, to assign observed spectra to possible products
Can be used where multiplicities, or accumulated knowledge about shifts or couplings, is not sufficient to distinguish between the possibilities
This is most useful for distinguishing isomers, conformers, or rotamers stable on the NMR timescale, where observed differences in shifts can be compared with calculated differences in shieldings

Geometry can be compared with crystallography results for the same compound
It is not likely to be exactly the same in solution as in the crystal
If the calculated major conformer in solution has a geometry similar to that in the crystal, it is unlikely to be so by accident, so the model is probably right
If it is quite different, then it may be correct, but you need some different evidence to test the model, e.g. by observation of different reaction pathways which correspond to the different geometry

Non-bonded distances in the model can be compared with values from solution state NMR Nuclear Overhauser Effect experiments

Relative energies can be compared with observed product or conformer distributions, if equilibrium has been established
As above, if the agreement is good, it is unlikely to be accidental; if it is poor, you cannot draw useful conclusions