Evolving Perl code for protein secondary structure prediction

Progress in the area of secondary structure prediction has been frustratingly slow[6]. The most accurate predictors at the moment are trained to predict one of three secondary structural states (helix, strand or coil) for each residue at position i using sequence information from a “window” of residues i ± 7. Information from more distant sequence positions should improve predictions further, since it is assumed that non-local interactions, like those that occur in sheet formation, can modulate the innate secondary structure preferences of a residue and its near neighbours. However, simply using a larger window does not help. First, the information content decreases rapidly as one moves away from i; because the likelihood that these residues are close in 3D space is also diminishing. Second, there is the problem of a using a fixed window to capture information from variable length secondary structures. Recently, attempts have been made to incorporate non-local information in secondary structure predictions. Baldi and coworkers[5] have used recurrent multi-pass neural networks and have shown that information from residues i±15 influences their predictions. Bystroff and coworkers[2] have taken another approach, which is to combine local predictors for secondary and supersecondary structures into a single large hidden Markov model which simultaneously takes into account context effects throughout the sequence. The more successful ab initio 3D predictors, such as Rosetta[1], may also have the side effect of producing more accurate secondary structure predictions. Unfortunately, these three approaches have not yet been shown to be superior to the established predictors in terms of percent predicted correctly into helix, strand or coil (Q3). The best Q3 currently stands at around 76%[4], and any future improvement on this will be a strong indicator of the successful incorporation of long-range and/or folding information into the predictors. This paper describes another attempt to increase secondary structure prediction accuracy using long-range information. The main assumption made in this work is that some form of computer program exists, at least in theory, which can do this. Such a program might mimic the folding dynamics in some way, perhaps in one, two or three dimensions using a reduced complexity model. For example, a predictor could have a simple rule: “predict weak-strand-region as strand if number-of-already-assigned-strands ≥ 2”. This rule could, for example, be applied after assigning “strong-strand-regions”, but before assigning regions with helical sequence patterns. The rationale here is that strands might be more likely to form in the context of an already forming sheet.