The implications of higher (or lower) success in secondary structure prediction of chain fragments

Abstract

The accompanying paper by Kihara (2005) provides clear data illustrating what has long been believed: Local secondary structure formation is affected by long-range interactions. To predict secondary structures more accurately, long-range tertiary information should be taken into account. This is an important contribution, as it substantiates what was commonly expected. And the investigator performs this job in a systematic, comprehensive way, providing an excellent reference for future papers on the limitations in all current schemes for secondary structure prediction. As the paper notes, secondary structure prediction is one of the oldest problems addressed through what is now called bioinformatics research (Chuo and Fasman 1978; Garnier et al. 1978). Today it is generally accepted that formation and subsequent association of secondary structures do not necessarily constitute the first steps in protein folding. Nevertheless, accurate prediction of the secondary structure elements is a tremendous asset in correct tertiary prediction. Current work on secondary structure prediction uses methods such as neural networks (Rost and Sander 1994), support vector machines (Hua and Sun 2001; Ward et al. 2003), machine learning algorithms, and hidden Markov models. Older methods are also enhanced (Kloczkowski et al. 2002), and new servers are constructed. Yet, despite considerable effort, current prediction success does not exceed the upper 70%–80%. This value is calculated with respect to the secondary structure assignment, which also suffers from inconsistencies between the different methods. This barrier has been attributed to long-range effects. Kihara (2005) demonstrates these effects by calculating the residue contact order (RCO) (Plaxco et al. 1998) on a large data set of nonhomologous proteins. This allows him to probe a potential relationship between the residue separation in terms of the RCO and the accuracy of the secondary structure prediction. The results presented in this paper reinforce several interesting conclusions relating to prediction accuracy. As is broadly accepted, folding is not a random search; rather, the landscape is funnel-shaped, which implies that building-block fragments of the protein have preferred conformations (Tsai et al. 2000). The conformations of the building blocks are stabilized by mutual interactions, short- or long-range, and their population times are related to the success of the prediction of the secondary structures. Thus, (1) if the residue has long-range interactions, then according to Figure 5 in the Kihara paper (2005), its population time in a given secondary structure is variable, as implied by low accuracy. In contrast, if the residue has short-range interactions, its population time is high; (2) if the prediction success of the secondary structure is high, the folding rate is fast, since it has higher population time (if the RCO is unchanged). The probability of success or confidence in the prediction can be used in scoring. The results provided by Kihara (2005) convincingly illustrate that to achieve higher accuracy in the prediction, the (~20) residue windows frequently used in the prediction of the secondary structures are not good enough. They are too short to account for long-range interactions. And consideration of long-range interactions is essential for a higher rate of success in the prediction. Yet, longer windows are difficult to implement in the calculation since they require a much larger data set. Nevertheless, the Kihara paper provides a real impetus to proceed in this direction.