Managing the sequence-specificity of antisense oligonucleotides in drug discovery

Abstract

All drugs perturb the expression of many genes in the cells that are exposed to them. These gene expression changes can be divided into effects resulting from engaging the intended target and effects resulting from engaging unintended targets. For antisense oligonucleotides, developments in bioinformatics algorithms, and the quality of sequence databases, allow oligonucleotide sequences to be analyzed computationally, in terms of the predictability of their interactions with intended and unintended RNA targets. Applying these tools enables selection of sequence-specific oligonucleotides where no- or only few unintended RNA targets are expected. To evaluate oligonucleotide sequence-specificity experimentally, we recommend a transcriptomics protocol where two or more oligonucleotides targeting the same RNA molecule, but with entirely different sequences, are evaluated together. This helps to clarify which changes in cellular RNA levels result from downstream processes of engaging the intended target, and which are likely to be related to engaging unintended targets. As required for all classes of drugs, the toxic potential of oligonucleotides must be evaluated in cell- and animal models before clinical testing. Since potential adverse effects related to unintended targeting are sequence-dependent and therefore species-specific, in vitro toxicology assays in human cells are especially relevant in oligonucleotide drug discovery.

Figure 1.

All drugs affect the transcriptome. (A) Global transcriptome effects in vitro upon treatment with drugs. Microarray data for 25 antisense oligonucleotides (AONs) were retrieved from public repositories (Supplementary Table S2), and for 315 FDA-approved small molecule compounds (SMCs) from the connectivity map repository (4) (Supplementary Table S1). Analysis of the SMCs revealed that those with first level Anatomical Therapeutic Chemical codes L (antineoplastic and immunomodulating agents) and P (antiparasitic products) affected significantly more genes than the others (P < 0.001, Wilcoxon test). They are therefore shown in a separate column. Both AONs and SMCs were typically evaluated at several different doses and in multiple cell lines (Supplementary Tables S1 and S2). For this analysis, all microarray data were preprocessed in the same manner, using robust multiarray averaging (129). Data from each study were preprocessed individually, except for the connectivity map data, where each SMC with its designated vehicle controls was preprocessed individually. Microarray data derived from seven different Affymetrix platforms. To allow comparisons across microarray types, a set of 11 367 genes assayed on all platforms was identified and the number of genes that changed expression by >50% in this common set of genes was calculated for each AON or SMC. (B) Schematic connecting drugs with their primary intended and unintended targets and the net downstream secondary effects. Here, we only consider RNA molecules as intended targets of AONs, and proteins as intended targets of small molecule compounds. In principle, and as unintended targets, AON binding to proteins and SMC binding to RNAs are also possible.

Figure 2.

Summary of major factors affecting gapmer activity on intended and unintended RNA targets: Targetable sequence space, target dynamics, target site accessibility, hybridization between gapmer and target, and target degradation by RNase H. Refer to subsections in the section headed “Determinants for RNase H-activity on unintended target RNA" for details.

Figure 3.

The effect of mismatches on affinity depends on the identity of the mismatched base pairs and the sequence context. The vertical axis indicates standard free energy, ΔG°. On each side of the axis are indicated ΔG° values for three examples of trinucleotide bases, which are paired either fully matched or with one central mismatch. On the left-hand side, the trinucleotide is GGC, and on the right-hand side it is CGA. The two mismatch-examples are G/G and G/T on both the left- and right-hand sides. Thermodynamic parameters for DNA–DNA binding were used to calculate ΔG° values (81).

Figure 4.

Model solutions of reactions (1a) and (1b) leading up to the relationship between binding affinity and potency. (A) Time-resolved numerical solution of the relative concentrations of free target RNA, [T], free gapmer oligonucleotide, [O], and the duplex between gapmer and RNA, [OT]. At a fixed time point, denoted by the vertical grey line, the concentration of RNA target is recorded, and in (B) plotted as a function of the total concentration of gapmer. From this curve, the gapmer concentration at which a half-maximal effect is achieved can be identified, and in (C) plotted as a function of the dissociation constant, K_d, between free gapmer and target RNA, and duplex (solid line). The relationship between binding affinity and potency presented in Pedersen et al. (48), is sketched as a dashed line.

Figure 5.

Experimental design to identify effects of unintended RNA targets using transcriptomics. Each box represents the entire transcriptome. Black dots and filled circles indicate targets and secondary effects. (A) Upon treatment with AON A, the central dot and largest circle indicate intended target and secondary effects. In addition, three unintended targets with corresponding secondary effects are shown. (B) When treating with AON B, the intended target and secondary effects are similar to AON A (compare A and B), but the unintended targets are different. (C) When comparing transcriptome changes upon treatment with AON A versus AON B, pharmacologically-induced changes derived from silencing the intended target are similar, whereas unintended effects are unique to each oligonucleotide.

Figure 6.

Different search paradigms yield very different results when studying the effects of oligonucleotide length on specificity. (A) When searching for matches with at least a certain number, x, of matching characters, specificity appears to decrease with length, whereas (B) when searching for hits with no more than a certain number, y, of mismatching characters, specificity appears to increase with the length of the oligonucleotide. Simulated data based on 100 oligonucleotides of each length. The y-axis in both A and B depicts arbitrary units, reflecting that any choice of x and y will result in the same overall shape of the distribution, although actual counts will differ.

Figure 7.

Examples of modifications to gapmers with LNA to increase binding affinity for optimal potency (here chosen as ΔG° = −19.5 kcal/mol, which is approximately the average of the optimal affinities identified Pedersen et al. (48) and matches our own experience as well). For the 14 nt gapmer, from atcgccgtactatg to ATCGccgtactATG (lowercase: DNA, uppercase: LNA); for the 16 nt gapmer, from tcagaagaccgctact to TCagaagaccgctACT; for the 18 nt gapmer, from ggcaagactgaatatgaa to GGcaagactgaataTGAA; and for the 20 nt gapmer, from taagcaaattagcgcgtatg to TaagcaaattagcgcgtaTG. Approximated thermodynamic parameters for LNA/DNA-RNA binding were used to calculate ΔG° values (48).

Figure 8.

The relationship between the length of oligonucleotides with equal affinity to an intended target and specificity according to computational predictions based on an energy model. Simulated data of 100 oligonucleotides of each length used to search a target space of 250 kb using RNAhybrid (123). A (partially mismatched) target site was included in the count if it had a free energy of binding to the oligonucleotide no more than 3 kcal/mol higher (weaker) than for the oligonucleotide to the fully matched intended target site (black bars). About half of all possible single mismatches will be within this range. The grey bars capture the effect of oligonucleotides binding 2 kcal/mol more strongly than what is needed for optimal potency. In this case, all partially mismatched sites with free energy of binding to the oligonucleotide of no more than 5 kcal/mol higher (weaker) than that of the oligonucleotide to the fully matched intended target site were included in the count. The numeric simulation was performed using thermodynamic parameters for RNA–RNA interaction (123), but the qualitative conclusions will be the same with parameters for DNA and modified nucleotides.

Figure 9.

Specificity profile for gapmers targeting the pre-mRNA for PCSK9. (A) For each 12 nt (gray) and 20 nt (black) gapmer that can be designed, the number of fully matching unintended target sites in the transcriptome are shown as a function of the starting position for that oligonucleotide in PCSK9. At the top, introns (line) and exons (box) are shown for PCSK9. For exons, untranslated regions are in white, and coding regions in grey. (B) All possible gapmer sequences of each length complementary to the target are shown (black), along with the subset that are unique to PCSK9, i.e. have no fully matched target regions anywhere else in the transcriptome (gray bars), and the number of unique gapmers with no unintended targets in the human transcriptome with a binding energy within 3 kcal/mol of the fully matching intended target site.