Specific inhibition of diverse pathogens in human cells by synthetic microRNA-like oligonucleotides inferred from RNAi screens

Abstract

Systematic genetic perturbation screening in human cells remains technically challenging. Typically, large libraries of chemically synthesized siRNA oligonucleotides are used, each designed to degrade a specific cellular mRNA via the RNA interference (RNAi) mechanism. Here, we report on data from three genome-wide siRNA screens, conducted to uncover host factors required for infection of human cells by two bacterial and one viral pathogen. We find that the majority of phenotypic effects of siRNAs are unrelated to the intended "on-target" mechanism, defined by full complementarity of the 21-nt siRNA sequence to a target mRNA. Instead, phenotypes are largely dictated by "off-target" effects resulting from partial complementarity of siRNAs to multiple mRNAs via the "seed" region (i.e., nucleotides 2-8), reminiscent of the way specificity is determined for endogenous microRNAs. Quantitative analysis enabled the prediction of seeds that strongly and specifically block infection, independent of the intended on-target effect. This prediction was confirmed experimentally by designing oligos that do not have any on-target sequence match at all, yet can strongly reproduce the predicted phenotypes. Our results suggest that published RNAi screens have primarily, and unintentionally, screened the sequence space of microRNA seeds instead of the intended on-target space of protein-coding genes. This helps to explain why previously published RNAi screens have exhibited relatively little overlap. Our analysis suggests a possible way of identifying "seed reagents" for controlling phenotypes of interest and establishes a general strategy for extracting valuable untapped information from past and future RNAi screens.

Keywords: antimicrobials; high-throughput RNAi screening.

Fig. 1.

Off-target effects in RNAi pathogen infection screens. (A) Experimental setup. HeLa cells were screened for host factors required for pathogen entry. Microscopy images from two separate wells of a typical perturbation experiment are shown (DAPI-stained HeLa cell nuclei in blue; successful pathogen infection in green from B. abortus expressing GFP). All three pathogens were screened using a genome-wide library (Qiagen), and Brucella and Salmonella additionally with two kinome-wide libraries (Ambion, Dharmacon). (B) Intended on-target mechanism of siRNA action. Below, in the correlation plots, each data point represents one gene, whereby the infection phenotypes (infection index) were averaged over all of the oligos designed for a given gene by a given library vendor. (C) Unintended off-target mechanism of siRNA actions. Here, each data point represents one seed sequence, with phenotypes averaged over all oligos that happen to contain that seed sequence in a given library. For all plots in B and C, pairs of oligos that happened to share the same seed sequence and the same on-target gene (in any of the three libraries) were excluded. Note that intervendor comparisons are based on the subset of genes screened with all three libraries (i.e., the kinome subset). Both correlations in C are highly significant (P ≤ 10⁻⁵⁰).

Fig. 2.

Genome-wide screening data aggregated by shared seed sequences. (A) Visualization of the entire genome-wide data of the infection screen for B. abortus, aggregated by the seed sequences found in the various siRNA oligos. Each data point represents one heptameric seed sequence, showing the averaged phenotypes over all siRNA oligos that happen to share that seed. The color code indicates the statistical significance of the observed infection phenotypes. For the negative control, data were plotted in exactly the same way, but the position of the seed in each siRNA oligo was incorrectly assumed to be at positions 12–18. (B) Visualizations for all three pathogens screened here; blue dots mark the seeds that have been selected for experimental follow-up.

Fig. 3.

Experimental confirmation of predicted seed phenotypes. (A) Detailed phenotypes measured for B. abortus (six replicates per oligo). (B) Summary of phenotypes measured for each of the three pathogens. The siRNA oligos predicted to block infection are shown in blue (dark blue for those that were designed not to have any on-targets), and oligos predicted to enhance infection are shown in orange (again, dark orange if lacking on-targets by design). The full sequences of all oligos in this experiment are given in Fig. S3.

Fig. 4.

Human miRNA overexpression phenotypes. (A) Based on the B. abortus genome-wide siRNA screen, specific seeds were selected that happened to occur also in known, endogenous human miRNAs. Eight of these seeds were predicted to reduce infection, eight were predicted to enhance infection, and eight were predicted to be neutral. To be selected, seeds had to be represented at least 10 times in the siRNA library and had to correspond to a single known human miRNA only. The figure shows the infection outcomes of transfecting these known miRNAs (as molecular mimics), compared with their predicted phenotypes as inferred from the seed analysis. (B) Tabulated details of the eight human miRNAs that were predicted, and confirmed, to block infection.

Fig. 5.

Specificity of seed effects. (A) Effects of single-point mutations located in the seed regions. For each of the three pathogens, one seed was chosen that was predicted to block infection (data shown in blue). Shown in gray are data for the corresponding seeds that have been mutated at one position. For both the standard inventory oligos as well as for oligos designed to have no full-length on-target sequence match, the infection phenotype is abolished upon mutating the seed sequence. (B) Principal component analysis (PCA) over the entire space of seed phenotypes observed for the three pathogens. (Left) Projection of the first two components of the PCA (each data point represents one seed; only seeds observed in at least 10 independent siRNA oligos are included). The seed effects on the cell numbers are virtually identical for all three pathogens, and align well with the first PCA dimension, which explains about 50% of the variance. (Right) Dimensions #2 and #3 separate the three pathogens (seeds are color-coded according to the pathogen for which they show the most significant infection-index phenotype).