Thermodynamic instability of siRNA duplex is a prerequisite for dependable prediction of siRNA activities

Abstract

We developed a simple algorithm, i-Score (inhibitory-Score), to predict active siRNAs by applying a linear regression model to 2431 siRNAs. Our algorithm is exclusively comprised of nucleotide (nt) preferences at each position, and no other parameters are taken into account. Using a validation dataset comprised of 419 siRNAs, we found that the prediction accuracy of i-Score is as good as those of s-Biopredsi, ThermoComposition21 and DSIR, which employ a neural network model or more parameters in a linear regression model. Reynolds and Katoh also predict active siRNAs efficiently, but the numbers of siRNAs predicted to be active are less than one-eighth of that of i-Score. We additionally found that exclusion of thermostable siRNAs, whose whole stacking energy (ΔG) is less than −34.6 kcal/mol, improves the prediction accuracy in i-Score, s-Biopredsi, ThermoComposition21 and DSIR. We also developed a universal target vector, pSELL, with which we can assay an siRNA activity of any sequence in either the sense or antisense direction. We assayed 86 siRNAs in HEK293 cells using pSELL, and validated applicability of i-Score and the whole ΔG value in designing siRNAs.

In-silico Approaches for RNAi Post-Transcriptional Gene Regulation: Optimizing siRNA Design and Selection

“…Also in this work they pointed out a very important threshold as the exclusion of Thermostable siRNA (with stacking energy (whole ∆G) < -34.6 k.cal) improved the score accuracy of not only i-Score but also DSIR, Biopredsi and ThermoComposition21 (Ichihara et al 2007). …”

“…These scoring techniques predict siRNA efficiency more accurately than the older tools. Although they use completely different algorithms to evaluate siRNA efficiency, they have very close accuracy compared to the rest of the second generation algorithms, as described by (Ichihara et al 2007). As in the comparative study done in the Ichihara's work all the second generation (except for Scales) and only Reynold and Katoh from the first generation achieved 90% successful prediction.…”

“…There are several methods for scoring and predicting designed siRNA activity, some of them are more accurate than the others; however, they are classified into two groups (Ichihara et al 2007…”

“…This class has been developed mainly through experimental data observation, as the existence of dataset with fully annotated experimentally siRNAs with their different efficiency enabling sophisticated data mining handling of this data, was not available until the dataset of Novartis that was introduced by Huesken (Huesken et al 2006) and used for training of several scoring tool as: Biopredsi (Huesken et al 2006), DSIR (Vert et al 2006), ThermoComposition21 (S.A. Shabalina et al 2006), i-Score (Ichihara et al 2007) and Scales (Matveeva et al 2007). These scoring techniques predict siRNA efficiency more accurately than the older tools.…”

“…Although ANN used in this work provided ambiguity to the module and prevented further development (due to the complexity of the model), it was considered, at that time period, the best way to handle all these different parameters. The server based Biopred model was later simulated and released as Biopredsi excel-based tool together with i-Score, which is going to be illustrated later on (Ichihara et al 2007). …”