Bioinformatic analyses of whole genome sequences highlight the problem of identifying the biochemical and cellular functions of the many gene products that are at present uncharacterised. Determination of their three-dimensional structures, either experimentally or by prediction, provides a powerful tool to address function, since it is at this level that biological activity is expressed. Here, we discuss the current approaches to protein structure prediction from sequence data, including the ab initio prediction of new folds, methods of fold recognition and comparative modelling based on homology. The value and limitations of such models are also explored. A major factor for the future will be the growth of the database of experimentally determined protein structures, through structural genomics projects. The prospects for this approach are also discussed, together with our experience in a pilot structural genomics project focused on proteins from Mycobacterium tuberculosis, the cause of tuberculosis (TB).