Peptide-encoding mRNA barcodes for the high-throughput in vivo screening of libraries of lipid nanoparticles for mRNA delivery

We reasoned that an ideal screening system for functional mRNA delivery would be model independent so that it could be applied in any preclinical model of disease, would consist of multiple measures of protein production that are each orthogonal to any others such that multiple formulations could be tested within the same assay and would be highly sensitive and quantitative over a wide dynamic range. The simplest approach to having an orthogonal functional output for each NP within a pool would be to have each NP effect production of a unique protein. The target tissues or cell populations could then be collected and lysed, and each expressed protein could be detected and its expression level measured using LC–MS/MS, which is a selective and sensitive method for protein detection and quantification (Fig. 1b). Since the only prerequisite for this assay is that the targeted cell type be able to translate functional protein from exogenously delivered mRNA, the proposed method would inherently be model independent. Thus, the assay as described would meet all three of our design criteria for an ideal screening system for functional delivery.

To that end, we developed an assay in which NPs are formulated with unique peptide barcode encoding mRNAs, which, when functionally delivered to the cytoplasm of target cells, express a protein fused with a peptide barcode sequence that can be detected and quantified through LC–MS/MS and a short epitope tag for quantification of expression using enzyme-linked immunosorbent assay. To reduce background noise from non-barcoded proteins present within target cells, we used a monomeric variant of streptavidin (monomeric streptavidin, mSA)25,26, which can easily be enriched from crude lysates using a biotinylated scaffold, as the carrier protein for each peptide barcode (Fig. 1c). After barcode-tagged mSA is immobilized on biotinylated beads, the C-terminal end of the carrier protein is cleaved by tobacco etch virus (TEV) protease, releasing the peptide barcode into solution. Finally, the resulting peptide barcode mixture is separated from the biotinylated beads through filtration and run on LC–MS/MS to quantify the relative amounts of each peptide present in solution.

As an initial proof of concept, we wanted to determine whether this approach could be used to accurately measure functional delivery to cells cultured in vitro. To that end, we prepared a set of 24 mRNAs that, upon functional delivery to the cytosol of target cells, would be translated into functional carrier proteins tagged with unique peptide barcode sequences (Supplementary Table 1). To determine whether this assay would enable simultaneous quantification of functional delivery of each mRNA, we created two pools (F2 and F3) that were each composed of the same 24 peptide barcode mRNAs but in different relative amounts and used them to transfect human embryonic kidney 293T (HEK293T) cells (Fig. 2a). F2 was constructed by adding each peptide barcode mRNA (except bc28 and bc32, which were each added at the median dose) at a different dose ranging from 13.6 ng to 1.95 µg; F3 was constructed by forming six groups each consisting of four peptide barcode mRNAs, with the mRNAs of each group added at a relative dose four times lower than that of the previous group, such that the dynamic range covered (2.42 ng to 2.48 µg) was roughly three orders of magnitude (Supplementary Table 1). To account for natural variations in measured peptide abundance due to different extents to which peptides ionize and fragment, as well as any differences in translation efficiencies between mRNAs, we transfected a separate set of cells with an additional pool composed of each mRNA in an equal amount (equal delivery reference). By normalizing the MS2 intensity of each peptide with respect to its intensity in the equal delivery reference, we are able to obtain a relative peptide abundance with respect to the other peptides within the pool. In general, we observed a very strong correlation (R2 = 0.951 and R2 = 0.991 for pools F2 and F3, respectively) between the normalized abundance of each peptide barcode measured via LC–MS/MS to the dose of the corresponding mRNA (Fig. 2b).

Fig. 2: Validation and expansion of the peptide barcoding assay.
figure 2

a, An mRNA pool was constructed by combining different doses of each peptide barcode mRNA; the mRNA pool was then packaged into LNPs or lipoplexes and administered to mice or HEK293T cells, respectively; peptide barcode solutions purified from cell lysates were then run on LC–MS/MS, and the relative abundance of each peptide was compared to its dose in the mRNA pool. Relative mRNA dose is calculated as the amount of each mRNA normalized to the average amount of each mRNA in the pool, while relative peptide abundance is calculated as the MS2 intensity measured for each peptide normalized to the MS2 intensity of that peptide in a sample generated using equal amounts of each mRNA (equal delivery reference). b, Parity charts comparing relative peptide abundance to relative mRNA dose using 24 peptide barcode mRNAs in HEK293T cells (left two panels) and in mouse liver (rightmost panel). c, A library of IVT templates were generated, pooled together in equal amounts and transcribed to produce a pool of 114 peptide barcode mRNAs, which was then formulated into LNPs and administered to mice. d, MS1 intensity (left) and peptide dot product (right) for each of the 114 peptide barcode sequences that were screened. e, Parity chart comparing relative peptide abundance and relative mRNA dose for a subset of 65 peptide barcodes chosen from the original pool of 114 on the basis of their MS1 intensities and dot products. For in vivo experiments in b, d and e, values shown are the mean of n = 3 mice per group, with error bars representing the s.d. For the in vitro experiments of b, values shown are the result from a single experiment. UTR, untranslated region.

Encouraged by these results, we performed a similar proof of concept experiment in vivo, using previously described cKK-e12 LNPs27 to deliver an mRNA pool composed of paired peptide barcode mRNAs at doses linearly spaced from 0.001 to 0.1 mg kg−1 (Fig. 2a and Supplementary Table 2). We again found that, when normalized to an equal delivery reference, the measured peptide abundances align closely (R2 = 0.929) with the administered doses of the corresponding peptide barcode mRNAs (Fig. 2b). Altogether, the results from these proof of concept experiments demonstrate that the peptide barcoding assay can accurately measure the amount of protein translated from multiple mRNAs within a single biological sample.

We then sought to optimize and expand the peptide barcode mRNA library to increase the screening throughput of the system. Our previous proof of concept experiments were performed using HA-tagged barcode sequences; since this tag comprises over half of the amino acid sequence of each barcode, we reasoned that varying the epitope tag could substantially affect sensitivity and quantification using the assay. Thus, we explored the effect of using four different epitope tags (HA, c-myc, FLAG and 6xHis) on measured peptide abundance. As expected, the choice of epitope tag had a significant impact on peptide abundance, with myc-tagged barcodes having more than an order of magnitude increase in their measured precursor ion intensity relative to those that were HA-tagged; conversely, FLAG-tagged and His-tagged barcodes were either undetected or resulted in intensities that were several orders of magnitude lower than their HA-tagged and myc-tagged counterparts (Supplementary Fig. 1).

To increase throughput of the assay, we generated a library of transcription templates encoding 114 unique myc-tagged peptide barcode mRNAs (Supplementary Table 3). These were then combined in equal amounts and subsequently transcribed to produce a single pool of peptide barcode mRNAs; this pool was then delivered in vivo to the mouse liver using cKK-e12 LNPs, and the resulting peptide barcode mixture was analysed using LC–MS/MS to identify a subset of suitable barcode sequences (Fig. 2c). Of the 114 unique barcode sequences, we identified a set of 65 peptides with high peptide dot products in liver lysates from treated mice, low dot products in PBS injected liver lysates, high signal to noise ratio (measured as the ratio of peptide abundance in treated liver lysates to abundance in untreated liver lysates) and some degree of separation in either precursor mass or retention time from other peptides in the library (Fig. 2d and Supplementary Table 3). We then performed an in vivo validation experiment in the mouse liver using either multiple reaction monitoring (MRM) or data-independent acquisition (DIA) for quantification of peptide abundance and observed excellent correlations between mRNA dose and measured precursor ion intensity, with slightly better quantification using DIA (R2 = 0.944) than MRM (R2 = 0.825) (Fig. 2e and Supplementary Fig. 2). In vitro validation experiments in HEK293T cells yielded similar results, indicating that the expanded library of 65 peptide barcodes can be used to accurately and simultaneously measure functional delivery rates both in vitro and in vivo (Supplementary Fig. 3). We note that only a weak correlation was observed between MS1 intensities of peptides expressed from the equal delivery controls in cultured HEK293T cells and mouse liver lysates (R2 = 0.412), which may be due to differences in translation efficiency in the two contexts, as well as differences in peptide extraction efficiencies from the two matrices (Supplementary Fig. 4).

We then sought to demonstrate that peptide barcoding can be used to develop and optimize LNP formulations. To that end, we synthesized a library of 384 unique ionizable lipids, all featuring the same general branched tail structure containing several biodegradable ester linkages, using the combinatorial reaction and components shown in Fig. 3. We then screened this library for mRNA delivery efficacy by formulating each ionizable lipid into a separate LNP with 1,2-dioleoyl-sn-glycero-3-phosphoethanolamine (DOPE), cholesterol and C14-PEG2000, and assaying for hepatic protein production effected by each LNP. Using the peptide barcoding assay, this entire library was screened using only eight mice (48 unique LNPs per mouse, not including replicates). To avoid potential false positives from background peptides that may be mistaken for barcodes, we set a minimum peptide dot product cut-off of 0.8. Of the 384 LNPs evaluated, 43 resulted in peptide dot products greater than the cut-off, with most (37 out of 43) being derived from either timnodonic acid (133-) or docosahexaenoic acid (137-) (Fig. 4a). We then validated the top four hits identified in the peptide barcoding analysis using the firefly luciferase (FLuc) assay and found that all four LNPs resulted in high levels of FLuc protein production in the liver when evaluated individually (Fig. 4b).

Fig. 3: Synthesis of 384 ionizable lipids with biodegradable ester linkages and extensive tail branching.
figure 3

Full list of the esters (left) and amines (right) used to generate the library are shown. RT, room temperature. DIC, N,N’- diisopropylcarbodiimide.

Fig. 4: Evaluation of an ionizable lipid series using peptide barcoding.
figure 4

a, Results from the peptide barcoding analysis of the combinatorial library (left) and structures of the top four hits (right). Only lipids that resulted in a peptide dot product more than 0.8 are shown. As in Fig. 2, relative peptide abundance is calculated as the MS1 intensity measured for each peptide normalized to the MS1 intensity of that peptide in a sample generated using equal amounts of each mRNA delivered using a reference lipid, cKK-e12 (equal delivery reference). b, Representative luminescence image of the top four hits 6 h after intravenous (i.v.) administration of 0.5 mg kg−1 FLuc mRNA (left; from left to right, RM133-3, RM137-15, RM137-14, RM133-14) and quantification of average radiance (right). Values shown are the mean of n = 3 mice, with error bars representing the s.d. Statistical significance in b was determined using a one-way ANOVA with Dunnett’s correction.

Next, we explored whether we could further improve performance of the lead compound, RM133-3, by optimizing its formulation. Several previous ionizable lipid screens have been performed using either a constant 1,2-distearoyl-sn-glycero-3-phosphocholine (DSPC)28,29 or DOPE20,30 formulation; in addition, a systematic study of the factors important for LNP efficacy found that phospholipid identity and ionizable lipid to weight ratio were the most significant predictors of LNP performance27. Thus, we chose to vary the phospholipid identity (DOPE versus DSPC) and the ionizable lipid to mRNA weight ratio (10, 15 and 20), as well as the molar ratios of lipid components in the formulation, which we varied linearly between the constant DSPC and DOPE formulations previously mentioned (Fig. 5a and Supplementary Table 4). Using a full factorial design of these variables, we generated a library of 24 unique peptide-barcoded LNPs with RM133-3 as the ionizable lipid and administered these to the same mouse. From this library, we identified three formulations that performed at a significantly higher level than the original RM133-3 formulation (Fig. 5b). In accordance with previous findings27,31, we found that formulations containing DOPE significantly outperformed those that contained DSPC (Extended Data Fig. 1). To verify that the barcoding analysis can be used to quantitatively assess individual LNP performance in a pooled setting, we compared the top three performing LNPs (RM133-3-20, RM133-3-14 and RM133-3-21) as well as the worst performing LNP (RM133-3-04) to the original formulation (RM133-3-01) using the human erythropoietin (hEPO) assay (Fig. 5c). While the ranking of the top three LNPs identified in the peptide barcoding analysis was not preserved in the results of the hEPO assay, we found that all three top performers did, in fact, outperform the original formulation; in addition, the worst performing LNP in the barcoding analysis resulted in lower protein production than any other formulation tested. Finally, the top-performing LNP, RM133-3-21, was found using the FLuc assay to be roughly 4.5 times more potent than DLin-MC3-DMA, which is a benchmark ionizable lipid currently used in the clinic (Fig. 5d).

Fig. 5: Optimization and characterization of lead lipid, RM133-3.
figure 5

a, Overview of the formulation parameters used in the optimization of RM133-3 LNPs. b, Peptide barcoding analysis of the LNP library. Several of the formulations resulted in significantly higher protein production than the original formulation, RM133-3-01. As in earlier figures, relative peptide abundance is calculated as the MS1 intensity measured for each peptide normalized to the MS1 intensity of that peptide in a sample generated using equal amounts of each mRNA delivered using a reference lipid, cKK-e12 (equal delivery reference). c, Validation of the barcode analysis using the hEPO assay. The top three performers identified in the barcode analysis (red), as well as the worst performer (green), were compared to the original formulation (blue) using the hEPO assay. d, FLuc expression following i.v. administration of RM133-3-21 (left) and MC3 (right) LNPs. Values shown are the mean of n = 3 mice, with error bars representing the s.d. Statistical significances in b and c were determined using a one-way ANOVA with Dunnett’s correction. Statistical significance in d was determined using an unpaired, two-tailed Student’s t-test. NS, not significant.

Read more here: Source link