Genome wide screen of RNAi molecules against SARS-CoV-2 creates a broadly potent prophylaxis

Abstract

Expanding the arsenal of prophylactic approaches against SARS-CoV-2 is of utmost importance, specifically those strategies that are resistant to antigenic drift in Spike. Here, we conducted a screen with over 16,000 RNAi triggers against the SARS-CoV-2 genome using a massively parallel assay to identify hyper-potent siRNAs. We selected 10 candidates for in vitro validation and found five siRNAs that exhibited hyper-potent activity with IC50<20pM and strong neutralisation in live virus experiments. We further enhanced the activity by combinatorial pairing of the siRNA candidates to develop siRNA cocktails and found that these cocktails are active against multiple types of variants of concern (VOC). We examined over 2,000 possible mutations to the siRNA target sites using saturation mutagenesis and identified broad protection against future variants. Finally, we demonstrated that intranasal administration of the siRNA cocktail effectively attenuates clinical signs and viral measures of disease in the Syrian hamster model. Our results pave the way to development of an additional layer of antiviral prophylaxis that is orthogonal to vaccines and monoclonal antibodies.

Figure 1:. Genome wide screen of SARS-CoV-2 by Sens.AI

(A) Oligo design. Each oligo had a unique RNAi trigger in the context of a miR-30 backbone along with a 50 nucleotide stretch from the viral genome that includes the RNAi target site (B) Library design. Each plasmid contained a specific RNAi trigger and its matching target- site in cis as part of the 3’UTR of a Venus reporter (C-E) Scheme of a screen in human cells (C) The RNAi machinery is turned on. Dicer^−/− 293FT cells were infected by the plasmid library and sorted for Venus^high expression, which formed the T₀ read out (D) The RNAi machinery is turned on. We restored the RNAi machinery by ectopic expression of a destabilised Dicer (ddDicer) (E) Modulating the activity of the RNAi machinery. Two biological replicates were subjected to seven different conditions, designed to titer Dicer expression either down, by using anti-Dicer siRNA, or up, by using Shield-1. Upon FACS sorting, Venus^low and Venus^dark cells were collected and their oligo constructs were sequenced (F) The resulting data matrix of the screen. Each row vector represents a specific combination of a treatment condition (the far left column) and a FACS gate (denoted as D, for Dark, and L, for Low). These row vectors describe the enrichment of each tested RNAi trigger (blue) compared to negative (green highlighted) and positive (red highlighted) controls. The AUC column was calculated best on the ability to separate the positive from the negative controls (G) Selecting the two best conditions. The two row vectors with the highest AUC were selected. Precision-recall (left) and receiver operating characteristics (right) curves for the intrinsic controls are shown. The screen score of each RNAi trigger is the average of these two row vectors.

Figure 2:. Validation of screen results

(A) The screen has high internal consistency. The results show a Pearson correlation of 72.2% between the screen scores in the two top performing conditions (B) Enrichment of features associated with potent shRNAs. The probability of not finding an Adenine in position 20 of the RNAi triggers targeting SARS-CoV-2 increases with screen scores, recapitulating previous studies indicating that when adenine occupies position 20 it weakens shRNA maturation (C) The screen’s resultant scores are highly correlated with bioinformatic predictions. The screen scores, representing enrichment statistics as calculated by DESeq2, are highly correlated with a published machine learning algorithm predicting the potency of shRNAs (D) Screen results and the selected candidates. The graph shows the resultant screen scores of each RNAi trigger that passed our quality threshold, as a function of position along the viral genome. Orange: RNAi triggers that target conserved regions between SARS-CoV and SARS-COV-2. Green: The selected 10 candidates. Blue: all other RNAi riggers (E) The dose-response curves of the selected 10 candidates. siRNAs were tested using a reporter assay. The plot represents the standardised median ratio between the expression of mCherry (reporter gene) and GFP (control gene) (F) Potency scores of each of the 10 selected candidates. The IC50 value of each of the 10 candidates was calculated based on the dose-response curves in (E).

Figure 3.. siRNAs repress live SARS-CoV-2 replication in VeroE6 cells.

Blue and orange: qPCR results of the E and RdRP transcripts, respectively. Unless noted othewise, the 100% viral load was calibrated to viral level after treatment with an anti-GFP siRNA (A) The viral load of SARS-CoV-2 (ancestral strain) after treatment with the top siRNA candidates from the Sens.AI screen (B) TCID50 levels of the SARS-CoV-2 (ancestral strain) following treatment with four of the top siRNA molecules. In each batch of the experiments, an siRNA against eGFP was used as a negative control (C) The viral load of SARS-CoV-2 (ancestral strain) after treatment with the top siRNA rom the bioinformatic pipeline (D) Testing the effect of various siRNA cocktails against the ancestral strain. The results were calibrated to the repression of S3 at the same concentration (E) The viral load against Delta versus the ancestral strain (F) The viral load of Omicron versus the ancestral strain after treatments with S5/Hel14 and other types of anti-virals (G) DeSEQ2 analysis of SARS-CoV-2 replicon treatment with the S3/S5 siRNA cocktail. We observed a sharp coverage decrease around the S3 (~23.5kbase) and S5 cleavage (~28kbase) sites (FDR values 4×10–83 and 6×10–22, respectively) along the replicon sequence.

Figure 4.. Saturation mutagenesis

(A) The Sensi.AI oligo design. The library consisted of the S5 as an shRNA trigger with 2143 mutations, exhausting every possible single- and double-mismatch possibilities in the siRNA target sites (B) Effect of mutation type on the S5 potency. The analysis shows all single mutations stratified by the mismatch type (Y-pyrimidine, R-purine; e.g. first letter: guide, second letter: target) (C) The effect of mismatches by position. The analysis shows all single mutations stratified by position, where position is based on guide strand orientation. Blue: means, yellow: smoothed mean effect based on position (D) The distribution of effect of mutations on S5. The vertical lines signifies no difference than the screen score of a target site without any mutation. Black: the distribution of effects of all 2143 single and double mutations. Grey: the distribution of all single mutations. Orange: the distribution of all double transition mutations. Red: the distribution of all double transversion mutations.

Figure 5.. Prophylactic treatment of SARS-CoV-2 infection in Syrian hamsters

(A) Dosing schedule. Syrian hamsters pre-treated with a non-targeting siRNA (negative control, grey), the LY-CoV555 antibody (positive control, yellow) or our lead siRNA cocktail (treatment, red) were infected with 4×10³ PFUs of the ancestral SARS-CoV-2 virus (B) Weight change after infection. The box plot presents the change in weight by treatment group relative to the infection day (C-D) Viral load (VL) by qPCR. All measurements are based on qPCR of the RdRP gene five days post infection from either homogenised (C) lungs (D) nares. In all panels p value is presented as: *<0.05,**<0.001.