siRNA off-target effects in genome-wide screens identify signaling pathway members

Abstract

We introduce a method for analyzing small interfering RNA (siRNA) genetic screens based entirely on off-target effects. Using a screen for members of the Wnt pathway, we demonstrate that this method identifies known pathway components, some of which are not present in the screening library. This technique can be applied to siRNA screen results retroactively to confirm positives and identify genes missed using conventional methods for on-target gene selection.

Figure 1. siRNA on-targeting and off-targeting to genes in a hypothetical pathway.

(a) On-target model correctly infers gene B as a pathway member due to on-target effects, depicted by the solid arrow from the siRNA (blue) to gene B. Extensive base-pairing between the siRNA and target gene B results in silencing. (b) A false-positive result incorrectly infers non-pathway gene C as a pathway member by neglecting off-targeting effects, depicted by dashed gray arrows from the siRNA (red) to pathway genes B and F. (c) Haystack explains screen results as a linear combination of the predicted off-targeting effects, depicted by dashed gray arrows from the siRNA (red) to pathway genes B and F. Imperfect base-pairing between siRNA (red) and 3′UTR region of off-target genes results in down-regulation.

Figure 2. Training an siRNA off-target model.

The seed sequence for an siRNA is defined here as the reverse complement of the heptamer at the 5′ end of the guide strand of the siRNA (bases 2–8), appended with an “A”. Four orthogonal match types are defined between the seed sequence of the guide (antisense) strand and a given 3′UTR: PM (octamer; perfect match), M1 (heptamer, mismatch on base 1 of the guide seed), M8 (heptamer, mismatch of base 8 of the guide seed), M18 (mismatch of bases 1 and 8 of the guide seed). The sequence of these match types are defined for two example siRNAs, PIK3CB-6338 and PIK3CB-6340. The length of a 3′UTR is also used as a predictive feature, as it has been empirically observed to be correlated with up-regulation when there are no matches of an siRNA to the 3′UTR (Figure 1). These features were then used in a linear regression against the mean log ratio of the transcript from gene expression profiles in which the siRNAs were transfected into cells (profiles GSM134511 and GSM134512 respectively, downloaded from GEO). The linear models trained from these two data sets were then cross-validated on each other, to demonstrate models derived from one siRNA can be successfully applied to another. Each model was used to predict significantly (p-value < .01) down-regulated transcripts in the data set/siRNA that it was not derived from, and the results of this cross-validation were displayed as ROC curves. The dashed line in each graph corresponds to the expected performance of a random model (AUC = 0.5). Finally, the data sets were merged to generate a final off- target model.

Figure 3. Reproducibility of t-statistics.

T-statistics calculated in the first step of the method, used to evaluate the null-hypothesis that the predicted off-target effects of the library on a given transcript are not correlated with the screening results. Large t-statistics in either the positive or negative direction indicate that there is significant correlation or anti-correlation respectively between the off-target effects on the transcript and the screening results. These correlations are reproducible across subsets of the data, in this case, across the data from each of the three single siRNAs designed against each gene.

Figure 4. Validation of Statistical Framework.

(a) A q-q plot of t- statistics generated by random permutation of the druggable singles data (in red) and non- randomized data (in blue).(b) Incidences and number of false positives in 1000 random permutations the druggable singles data. (c) 3′UTR length distribution for false positives (red) and positives (blue) compared to the overall distribution of 3′UTR lengths (black).