A statistical analysis of RNA folding algorithms through thermodynamic parameter perturbation

Abstract

Computational RNA secondary structure prediction is rather well established. However, such prediction algorithms always depend on a large number of experimentally measured parameters. Here, we study how sensitive structure prediction algorithms are to changes in these parameters. We found already that for changes corresponding to the actual experimental error to which these parameters have been determined, 30% of the structure are falsely predicted whereas the ground state structure is preserved under parameter perturbation in only 5% of all the cases. We establish that base-pairing probabilities calculated in a thermal ensemble are viable although not a perfect measure for the reliability of the prediction of individual structure elements. Here, a new measure of stability using parameter perturbation is proposed, and its limitations are discussed.

Figure 1

Schematic representation of an RNA secondary structure. The solid line represents the backbone of the molecule while the dashed lines symbolize base pairs. Each such structure can be decomposed into stacking loops (s), bulges (b), interior loops (i), hairpin loops (h) and multi-loops (m) as indicated.

Figure 2

The average scaled tree distances of various natural sequences for different perturbations of the parameter ε. The error bars denote the statistical error after averaging over 100 different realizations of perturbed free-energy parameter sets. The solid line shows the average over all sequences. It becomes obvious that already at the experimental uncertainty of ε ≈ 0.3 kcal/mol ∼30% of the structure is predicted unreliably.

Figure 3

Correlation between the probability to find a base pair in an mfe structure with perturbed parameters and the probability of the base pair in the thermal ensemble for the sequence with accession number Y13474. Each point represents a base pair and its position represents its respective probabilities of forming. The different plots show how the two probabilities become more correlated as the strength ε of the parameter perturbations is increased.