Aid to discovery of new protein foldings

Large-scale sequencing projects produce an exploding number of known protein sequences. The current number is about 36,000 [Bairoch & Boeckmann, 92] sequences, but before the end of the century many more than 100,000 will have to be dealt with. This is in contrast to the far slower increase in the number of known protein structures, currently about 2,000 [Berstein et al., 77]. The rate of increasing is roughly 100 sequences/day and 1.5 3D structure/day. Thus, it is increasingly important to develop computational approaches to determine automatically (predict) the structure of proteins whose sequences are known. Because the general problem of prediction of protein fold is so difficult, researchers have tried to predict regular substructures forming the imaginary level of protein structural description known as secondary structure. Knowledge of the secondary structure can contribute significantly towards the goal of tertiary fold prediction. This knowledge can constrain the possible conformations of the protein [Cohen & Knutz, 89], provide a good starting point and reduces the search space in simulation of protein folding by molecular dynamics [Levitt, 83] or lattice models [Skolnick & Kolinski, 90], or can be used in predicting higher order structures (e.g. super secondary structures [Taylor & Thornton, 83], domains [Lathrop, 87]). The established methods for protein secondary structure prediction include hand-crafted expert rules [Lim, 74], biological predictive patterns [Cohen et al., 83, 86] [Presnell et al., 92], statistical Chou-Fasman theory [Chou & Fasman, 74], information theory-based GOR method [Garnier et al., 78]. More recent methods often make use of inductive learning techniques, whereby a system is trained with a set of sample proteins of known conformation and then uses what it has learned to predict the structure of previously unseen proteins. Both neural networks [Qian & Sejnowski, 88] [Kneller, 90] [Zhang et al., 92] and symbolic induction have been applied [King & Sternberg, 90] [Muggletonet al., 92] in the secondary structure prediction context. Despite the apparent practical importance of the secondary structure concept, the quarter of century long research efforts have shown the existence of a secondary structure prediction limit. Even if this limit has been recently ameliorated up to 70% [Cost & Salzberg, 93] [Leng & Bachanan, 93] [Rost & Sander, 93], this rate of accuracy is too low to be of practical use in constraining the conformation space for tertiary structure prediction. The main reason of the failure of the secondary structure prediction methods is that the formation of structure (including the secondary one) is only to a certain degree due to sequentially local interactions of amino acids. However, most methods known to date do rely on local information. It becomes widely recognized that to deal properly with protein structure prediction problem one should tackle differently the representation issues. Several attempts have been made to change the representations involved into the secondary structure prediction problem. The protein sequence usually presented by amino acids has only been examined in terms of physico-chemical properties of amino acids [Hunter, 91] [Cherkauer & Shavlik, 93]. The secondary structure elements have been replaced by alternative classes of local structure, which account for recognized helical and strand regions, as well as for novel categories such as Nand C-caps of helices and strands [Zhang et al., 93]. However, these first works addressing representation issues do not improve significantly the state-of-the-art. Obviously, much more representation work is needed. Biological knowledge representation is hard for several reasons: the tricky nature and ill-formalized character of the available biological knowledge the lack of general theoretical understanding in the field the fact that this knowledge comes in a raw form