Identification of sequence motifs significantly associated with antisense activity

Abstract

Background: Predicting the suppression activity of antisense oligonucleotide sequences is the main goal of the rational design of nucleic acids. To create an effective predictive model, it is important to know what properties of an oligonucleotide sequence associate significantly with antisense activity. Also, for the model to be efficient we must know what properties do not associate significantly and can be omitted from the model. This paper will discuss the results of a randomization procedure to find motifs that associate significantly with either high or low antisense suppression activity, analysis of their properties, as well as the results of support vector machine modelling using these significant motifs as features.

Results: We discovered 155 motifs that associate significantly with high antisense suppression activity and 202 motifs that associate significantly with low suppression activity. The motifs range in length from 2 to 5 bases, contain several motifs that have been previously discovered as associating highly with antisense activity, and have thermodynamic properties consistent with previous work associating thermodynamic properties of sequences with their antisense activity. Statistical analysis revealed no correlation between a motif's position within an antisense sequence and that sequences antisense activity. Also, many significant motifs existed as subwords of other significant motifs. Support vector regression experiments indicated that the feature set of significant motifs increased correlation compared to all possible motifs as well as several subsets of the significant motifs.

Conclusion: The thermodynamic properties of the significantly associated motifs support existing data correlating the thermodynamic properties of the antisense oligonucleotide with antisense efficiency, reinforcing our hypothesis that antisense suppression is strongly associated with probe/target thermodynamics, as there are no enzymatic mediators to speed the process along like the RNA Induced Silencing Complex (RISC) in RNAi. The independence of motif position and antisense activity also allows us to bypass consideration of this feature in the modelling process, promoting model efficiency and reducing the chance of overfitting when predicting antisense activity. The increase in SVR correlation with significant features compared to nearest-neighbour features indicates that thermodynamics alone is likely not the only factor in determining antisense efficiency.

Figure 1

Thermodynamic Distribution of Experimental Sequences. Empirical thermodynamic distributions of experimental antisense sequences. The curves represent the distribution of Gibbs free energy (dG) for the effective (green) and ineffective (red) antisense sequences. The average dG for the effective sequences (-24.04 kcal/mol) was significantly more negative than the average dG for the ineffective sequences (-22.80 kcal/mol).

Figure 2

Thermodynamic Distributions of (top) Significant Motifs and (bottom) Submotif-unique Significant Motifs. The distributions of Gibbs free energy values (dG) for sequences associated with effective antisense activity (green) and those associate with ineffective antisense activity (red). The difference in average dG between "good" and "bad" motifs in the subword-unique motifs (-3.45 vs. -1.66 kcal/mol) is greater than the difference in means in the entire population of significant motifs (-3.59 vs. -2.105 kcal/mol). This can be attributed to the removal of motifs from each population that contain submotifs in the opposing group.

Figure 3

Subword Structure of Effective Subword-unique Motifs. A subword tree illustrating the makeup of the subword-unique motifs associated with antisense effectiveness. A motif is linked with an arrow to another if the motif at the tail of the arrow is a submotif of the one at the head that differs only by the addition of one base to the beginning or the end. For example, "CC" would be at the tail of an arrow connecting it to either "CCG" or "GCC", but there would be no arrow connecting it to "CGC" or "GGCC". For motifs associate with antisense effectiveness, the majority of significant motifs are linked in a tree with "C" as the root node.

Figure 4

Subword Structure of Ineffective Subword-unique Motifs. A subword tree constructed in the same manner as Figure 3, but for motifs associated with ineffective antisense suppression. In this population of motifs, nodes are grouped into two main trees, one rooted with "A", one rooted with "T". Surprisingly, despite the lower Gibbs free energy associated with its presence, a third tree rooted at "G" was found within this population of motifs as well.