Functional features defining the efficacy of cholesterol-conjugated, self-deliverable, chemically modified siRNAs

Abstract

Progress in oligonucleotide chemistry has produced a shift in the nature of siRNA used, from formulated, minimally modified siRNAs, to unformulated, heavily modified siRNA conjugates. The introduction of extensive chemical modifications is essential for conjugate-mediated delivery. Modifications have a significant impact on siRNA efficacy through interference with recognition and processing by RNAi enzymatic machinery, severely restricting the sequence space available for siRNA design. Many algorithms available publicly can successfully predict the activity of non-modified siRNAs, but the efficiency of the algorithms for designing heavily modified siRNAs has never been systematically evaluated experimentally. Here we screened 356 cholesterol-conjugated siRNAs with extensive modifications and developed a linear regression-based algorithm that effectively predicts siRNA activity using two independent datasets. We further demonstrate that predictive determinants for modified and non-modified siRNAs differ substantially. The algorithm developed from the non-modified siRNAs dataset has no predictive power for modified siRNAs and vice versa. In the context of heavily modified siRNAs, the introduction of chemical asymmetry fully eliminates the requirement for thermodynamic bias, the major determinant for non-modified siRNA efficacy. Finally, we demonstrate that in addition to the sequence of the target site, the accessibility of the neighboring 3' region significantly contributes to siRNA efficacy.

Figure 1.

Efficacy distribution of the panel of chemically modified, asymmetric, self-delivering siRNAs (sdRNAs). (A) sdRNAs are asymmetric siRNAs, consisting of a 20-nucleotide antisense strand and a 15-nucleotide sense strand, in which all pyrimidines are 2′-fluoro (antisense) and 2′-O-methyl (sense) modified. The 3′ terminal backbone is phosphorothioated (six linkages in antisense and two in sense). The 3′ end of the sense strand is conjugated to cholesterol. (B) The efficacy of 356 sdRNAs targeting 17 genes was evaluated using dual luciferase reporter in HeLa cells at 1 μM (passive uptake) at 48 h (n = 3, mean ± SD).

Figure 2.

Development of an algorithm for the prediction of sdRNA efficacy. (A) The positional base preference matrix was generated using three functionality cutoffs (17%, 24% and 35% functional versus >44% non-functional compounds) for the 50-base regions comprising the siRNA-targeting site. Matrix weight values are color-coded by value as indicated by color bar below matrices. Analyzed mRNA positions corresponding to siRNA-targeting region (shaded) are indicated at the top. The location of cleavage site between positions 10 and 11 is indicated with a black arrow. (B) Using linear regression analysis (R 3.4.1), the scoring algorithm was generated for shown positional preference matrices (P < 0.001, see Materials and Methods). Algorithm performance is visualized as positive predictive power (PPP) versus sensitivity curves. PPP is calculated as the percent of correctly predicted (functional) sequences versus total predicted sequences for each algorithm value. Sensitivity is calculated as the percent of functional sequences selected vs total functional sequences present in the dataset for each algorithm value. sdRNA compounds with >44% gene expression remaining were defined as non-functional. The 17%/NF-preference matrix-based algorithm demonstrates the best performance with 96% PPP at 25% sensitivity. Black line shows performance of the control algorithm (see Methods). (C) The efficacy of individual sdRNA compounds selected by the 17%/NF scoring algorithm at 25% sensitivity (n = 3, mean ± SD). sdRNA IDs are indicated along the x-axis.

Figure 3.

sdRNA algorithm accurately predicts efficacy using two independent datasets. The sdRNA (24%/NF) algorithm was applied to predict the efficacy of (A) 50 sdRNAs targeting five genes (qPCR, sdRNAs inducing ∼ <25% target gene expression are defined as functional). sdRNA algorithm predicts efficacy with ∼60% accuracy at ∼30% sensitivity. (B) 94 sdRNAs targeting the huntingtin gene (QuantiGene (34), sdRNAs inducing ∼ <25% target gene expression are defined as functional). sdRNA algorithm predicts efficacy with ∼80% accuracy at ∼25% sensitivity. Black line shows performance of the control algorithm (see Methods).

Figure 4.

Algorithms derived from naked siRNA do not have predictive power for modified (sdRNAs) and vice versa. (A) The positional base preference matrix was generated from non-modified (orange) (5) and chemically modified (green) sdRNA. Sequences were aligned based on the 5′ end of the antisense strand. Matrix weight values are color-coded by value as indicated by shaded bar below the matrices. Analyzed mRNA positions corresponding to siRNA-targeting region (shaded) are indicated at the top. Black arrow indicates the location of cleavage site between positions 10 and 11. (B) The ability of algorithms derived from non-modified and modified sdRNA datasets to predict the efficacy of non-modified and modified siRNAs was calculated using PPP vs sensitivity plots. (C) The thermodynamic flexibility of the non-modified and chemically modified siRNAs was estimated by averaging GC content over a sliding window of four bases. Thermodynamic bias is indicated as the difference between the relative thermodynamic flexibility at 5′ and 3′ ends of the siRNA duplex. Chemically modified siRNAs do not display conventional thermodynamic bias.

Figure 5.

mRNA local thermodynamic flexibility in the 3′ region outside the siRNA-targeting site contributes to sdRNA efficacy. (A) The frequency of AU at each position (black bars) in the siRNA-targeting region and surrounding 5′ and 3′ regions was computed by subtracting the frequency of AU in non-functional siRNAs from that in functional (< 24% mRNA expression remaining) siRNAs. The background (grey area) was simulated using AU frequency in the randomly distributed training dataset of 356 siRNAs. The 80% confidence interval of the simulated background is shown. Analyzed mRNA positions are indicated at the top along with corresponding siRNA-targeting region (shaded area; positions 1-20). The location of cleavage site between positions 10 and 11 is indicated with a black arrow. (B) The frequency of AU at each position was averaged over a four-base region (black line). The average AU frequency was computed over each region (grey dotted line). The background (grey solid line) was averaged over a four-base region.

Figure 6.

2′-O-methyl modification at position 14 of the antisense strand negatively modulates sdRNA efficacy. (A) The frequency of 2′-O-methyl modification per position of the antisense strand in functional (defined as < 24% gene expression remaining) versus non-functional (defined as >44% gene expression remaining) sdRNAs. (B) The efficacy of sdRNA targeting MAP4K4 with and without 2′-O-methyl modification in position 14 of the antisense strand. MAP4K4 expression was analyzed by qPCR in HeLa cells treated with sdRNAs for 72 h (n = 3, mean ± SD; one-way ANOVA P < 0.001).