Using Tagged mRNA to Quickly Find Lipid Nanoparticles That Target Tissues Outside the Liver for Immune Engineering

Our group and others have previously reported the use of HTS approaches based on b‐DNA for LNP delivery screening [11, 12, 13, 14]. While effective, these screening approaches suffer from the drawback that LNP physicochemical and delivery properties are strongly influenced by the nature of the encapsulated payload [23, 24, 25]. The issue of cargo‐carrier interdependence can be partially circumvented through the dual encapsulation of mRNA with low amounts of b‐DNA to produce more “mRNA‐like” LNPs, as demonstrated previously [11, 26]. Ultimately, however, b‐DNA screening methods track the b‐DNA molecule rather than the mRNA cargo of interest. To allow accurate tracking of mRNA cargo distribution in a high‐throughput manner, our group has previously reported a proof‐of‐concept b‐mRNA platform [16]. Using this platform, we demonstrated that moving the barcode sequence to the mRNA molecule was feasible and, moreover, yielded more accurate results than b‐DNA screening. In the present work, we sought to establish a next‐generation b‐mRNA platform for high‐fidelity in vivo mRNA LNP screening, sufficient for screening large pooled LNP libraries for extrahepatic mRNA delivery.

As error correction is an important consideration for pooled screening approaches, we first employed an in silico optimization approach to develop a large pool of 12 nt barcodes with error correction capabilities [27]. We sorted the resultant list of 1814 barcodes by edit distance and selected the 200 sequences with the greatest average edit distance (Table S1). Based on these optimized sequences, we created 200 distinct double‐stranded DNA (dsDNA) templates for in vitro transcription (IVT) differing only in the barcode sequence. Mature b‐mRNA contained as primary features a m⁷GpppAmG “cap 1” structure, a 5' untranslated region (UTR), a codon‐optimized enhanced green fluorescent protein (EGFP) coding sequence (CDS), a 3′ UTR, a 12 nt barcode sequence, a short “clamp” sequence, and a poly(A) tail (Figure 1a). b‐mRNA products demonstrated uniform electrophoretic mobility (Figure 1b), suggesting comparable size and degree of polyadenylation across sequences. Moreover, Sanger sequencing of selected b‐mRNAs and subsequent sequence alignment yielded a strong consensus sequence, with strong alignment to desired sequences within the fixed regions and expected polymorphism within the 12 nt barcode region (Figure 1c). To evaluate the ability of b‐mRNA to facilitate transgene expression, we treated HeLa cells with C12‐494 mRNA LNPs containing a pool of all 200 candidate b‐mRNAs (Figure S1) and compared expression of EGFP to cells treated with the same mRNA LNP formulation encapsulating barcode‐free EGFP mRNA. As expected, we observed that EGFP expression did not appear to be influenced by the presence of barcode sequences (Figure 1d).

As RNA in the cellular compartment is subject to a plethora of sequence‐dependent interactions and regulatory pathways [28, 29], we next sought to eliminate any b‐mRNAs demonstrating differential performance across tissues or cell types. In particular, as endogenous microRNA (miRNA) binding sites are generally considered to be 6 to 8 nt in length [30], we anticipated the possibility of differential miRNA binding to b‐mRNAs within the 12 nt barcode region. As the prevalence of miRNA species differs across tissues and cell types even under homeostatic conditions [31, 32, 33], differential miRNA binding to b‐mRNA could be expected to lead to bias in screening results if overlooked. To avoid screening bias arising from barcode polymorphism, we performed an in vivo study of b‐mRNA accumulation (Figure 1e). We pooled all 200 b‐mRNAs together and encapsulated this pooled library into a single C12‐494 LNP formulation, which we then administered systemically to healthy C57BL/6 mice. We collected peripheral blood, hearts, kidneys, livers, lungs, and spleens from treated mice to assess bioaccumulation of each b‐mRNA via NGS. As expected, distributions of relative b‐mRNA accumulation were largely symmetric and central (Figure 1f). We performed a formal enrichment analysis to identify and eliminate b‐mRNAs demonstrating statistically significant enrichment or depletion, suggestive of unwanted interactions with endogenous cellular machinery (Figure 1g), yielding a library of 134 b‐mRNAs suitable for pooled in vivo mRNA delivery screening (Table S2).

Having developed a platform for high‐throughput in vivo LNP screening, we set out to investigate cellular mRNA delivery facilitated by novel mRNA LNPs. Our group has previously reported that, under systemic administration, ionizable lipid structure plays a pivotal role in LNP cellular tropism [11]. We therefore opted to screen a large library of novel ionizable lipids, largely focused on a set of biodegradable lipids easily synthesized through a “plug‐and‐play” strategy from commercial reagents (Figure S2). We have observed strong liver transfection by a subset of these lipids and reasoned that a more detailed investigation of tropism would yield interesting findings in the liver and beyond [34]. We synthesized 120 “plug‐and‐play” ionizable lipids and used them to formulate b‐mRNA LNPs (Figures S3–S9). We also included several field‐standard and industry‐standard ionizable lipids as controls, including C12‐200 [35], cKK‐E12 [36], DLin‐MC3‐DMA [37], 306O_i10 [38], SM‐102 [39], and ALC‐0315 [40]. We further included the ionizable lipids previously reported by our group C14‐482, C14‐488, C16‐488, C12‐494, C14‐494, C16‐494, and C14‐c494, each of which has shown promise across various RNA delivery applications [11, 41, 42, 43]. All told, we synthesized 133 b‐mRNA LNPs for evaluation (Table S3). Typical for screening experiments using large numbers of candidate LNPs [12, 13], a small number of our formulated LNPs demonstrated poor encapsulation, likely due to ineffective complexation with certain ionizable lipids. We therefore pooled 122 resultant LNPs demonstrating sufficient mRNA encapsulation (Figure S10) with a “naked” b‐mRNA, then administered this pool systemically to both C57BL/6 and APOE^–/– mice to assess ApoE‐dependent and ‐independent mRNA delivery. We collected lungs, spleens, livers, and peripheral blood from treated mice and performed fluorescence‐activated cell sorting (FACS) to isolate cell populations of interest before preparing libraries for sequencing via NGS. As intracellular exogenous mRNA is vulnerable to enzymatic degradation, we opted to analyze b‐mRNA accumulation at a relatively short time following treatment to avoid potential screening bias introduced by extensive degradation. However, peak mRNA translation occurs relatively late after treatment, well after mRNA degradation onset would be expected [25]. Future studies may wish to analyze results at later time points to be able to correlate mRNA transfection with b‐mRNA accumulation data. Alternatively, some form of fluorescence signal amplification, such as anti‐EGFP antibodies, could help extend the fluorescence detection window to better align with the window for maximum b‐mRNA recovery. While difficult to perform in bulk, chemical modification of the b‐mRNA backbone may also enhance intracellular stability and allow for simultaneous transfection and accumulation measurements with minimal bias [44].

Using NGS data, we analyzed the biodistribution profile of b‐mRNA LNPs within our library. In wild‐type mice, several candidate LNP formulations demonstrated both hepatic and extrahepatic tropism (Figure 2a). In APOE^–/– mice, more standout LNP formulations appeared, likely due to the absence of ApoE‐mediated LNP clearance in the liver (Figure 2b) [18]. Across both genotypes, we observed only modest specificity of delivery, with uptake by a given cell type generally correlating well with uptake by other cell types, with the main exceptions of liver endothelium in wild‐type mice and lung leukocytes in APOE^–/– mice (Figure 2c,d). More interestingly, however, we observed generally strong correlation between accumulation profiles in wild‐type and knockout mice, suggestive of ApoE‐independent hepatic and extrahepatic transfection (Figure 2e). Based on NGS results, we selected LNPs 9, 37, 94, 96, 97, and 112 for further evaluation.

These LNPs contained the biodegradable ionizable lipids 2D8, 7D4, 17D8i, 17D9.2, 18D4, and 20D8i, respectively [34]. In accordance with previous findings by our group [11], we observed a nuanced relationship between ionizable lipid structure and intracellular accumulation in cell types of interest, with few structural features appearing to be universally favorable (Figure S14). This highlights a major strength of high‐throughput screening as employed here: rather than relying on a priori knowledge of difficult‐to‐establish relationships between general lipid structural features and function to narrow candidate LNP libraries to a tractable size, entire expanded libraries can instead be directly evaluated in vivo in an unbiased manner.

To confirm transfection of lead mRNA LNP candidates identified by our high‐throughput b‐mRNA screen, we performed low‐throughput validation studies. We first assessed the extrahepatic transfection capabilities of our LNPs at the organ level. We treated C57BL/6 mice with the selected LNP candidates encapsulating NanoLuc mRNA at a low dose and assessed organ transfection using bioluminescence imaging (Figure 3a). Interestingly, all novel LNP candidates tested demonstrated clear splenic transfection, with most demonstrating minimal liver transfection. To aid in evaluating tropism of candidate mRNA LNP, we quantified splenic vs. hepatic luminescence signal to classify LNPs as spleen‐ or liver‐tropic (Figure 3b). Interestingly, LNPs 9, 37, 94, 96, and 112 demonstrated splenic transfection with relatively little liver transfection, causing us to classify them as spleen‐tropic. LNP 97 demonstrated strong hepatic transfection, causing us to classify it as liver‐tropic. As expected, SM‐102 LNPs were also classified as liver‐tropic, though they did additionally demonstrate strong splenic transfection. Excitingly, LNP 112 achieved extremely strong splenic transfection, with comparable spleen luminescence to SM‐102 LNPs and over an order‐of‐magnitude decrease in liver luminescence, sufficient to shift its classification to spleen‐tropic. As such, LNP 112 emerged from first‐pass validation as our lead LNP candidate for in vivo reprogramming of splenocytes.

Having established the potency of lead mRNA LNP candidates, we next sought to investigate cellular tropism in greater detail. To this end, we treated C57BL/6 mice with fluorescently labeled EGFP mRNA LNPs and performed flow cytometry to analyze accumulation at the cellular level. In the liver, we observed generally high LNP accumulation in Kupffer cells, with lower degrees of accumulation occurring in hepatocytes and non‐Kupffer leukocytes (Figure 3c,d). For most LNPs tested, we generally observed greater transfection of Kupffer cells than other liver cell types, which agrees well with observed accumulation data (Figure 3e,f). However, SM‐102 LNPs demonstrated very strong transfection of hepatocytes—over 50% on average—and appreciable but less marked transfection of liver endothelial cells. Strikingly, LNP 97 mediated even stronger hepatic transfection than SM‐102 LNPs, attaining nearly 80% mean hepatocyte transfection and roughly 20% transfection of liver endothelial cells. These impressive hepatic transfection characteristics make LNP 97 a promising candidate for treating disorders of the liver and may further make it an interesting prospect for treating autoimmune disorders by leveraging the tolerizing environment of the liver [45].

In the spleen, our tested LNPs largely accumulated in myeloid cells (Figure 3g,h). LNP 96, LNP 97, and SM‐102 LNPs demonstrated indistinguishable accumulation in splenic macrophages, each reaching in excess of 70%. LNPs 9 and 94 each accumulated in roughly 40% of splenic macrophages, while LNPs 37 and 112 demonstrated substantially lower accumulation. LNP 96, LNP 97, and SM‐102 similarly demonstrated the greatest accumulation in dendritic cells (DCs) of LNPs tested, each reaching at least 30%, while LNPs 9 and 94 again demonstrated lower accumulation and LNPs 37 and 112 again demonstrated lower signal. Strikingly, LNP 112 demonstrated remarkable transfection of splenic myeloid cells, attaining over 10% mean transfection of splenic macrophages and over 6% mean transfection of splenic DCs (Figure 3i,j). As with the liver, despite its low accumulation in myeloid cells, LNP 112 demonstrated strong transfection of these cell types in the spleen, an exciting finding for vaccine applications. Tissue‐resident macrophages have previously been shown to mediate transfer of mRNA LNP cargo to other cell types, which may partially explain the observed discrepencies between LNP accumulation and mRNA translation [46]. Future studies may wish to leverage approaches such as radioisotope labeling to precisely track lipid and mRNA components of LNPs to better understand mechanisms of cargo transport and lipid clearance [47]. Altogether, validation experiments demonstrated the potential of lead LNPs for both hepatic and extrahepatic mRNA delivery, identifying LNP 112 as a lead candidate for the in vivo engineering of splenocytes.

It is important to note that our in vivo data are collected in C57BL/6 mice. It is possible that tropism and immune responses to vaccination with lead immunotropic LNPs formulations may differ in other strains of mouse and in higher animals. Previous studies evaluating LNP biodistribution in primatized and humanized mice have demonstrated differential cellular responses to LNP treatment [48]. To translate HTS findings to the clinic, it will be important to evaluate LNP candidates in higher species. Our pooled screening approach based on b‐mRNA is well‐suited for mRNA LNP evaluation in higher species such as non‐human primates (NHPs) with minimal resource requirements and could help to facilitate the clinical translation of immunotropic mRNA LNPs [49].

Based on the strong performance of LNP 112 for transfection of splenic antigen‐presenting cells (APCs), we sought to evaluate its potential for use in a cancer vaccine application. To this end, we established a syngeneic mouse melanoma model by inoculating C57BL/6 mice of mixed sex with B16‐OVA melanoma cells. After allowing tumors to engraft, we treated mice with one of four treatments at a 5 d prime‐boost interval (Figure 4a): (i) phosphate‐buffered saline (PBS) (negative control), (ii) LNP 112 encapsulating antigen‐irrelevant firefly luciferase (FLuc) mRNA (vehicle control), (iii) LNP 112 encapsulating ovalbumin (OVA) antigen mRNA, or (iv) SM‐102 LNPs encapsulating OVA mRNA (clinical control). We closely monitored mouse body weight (Figure S15), tumor volume (Figure S16), and survival over the course of 60 d. Excitingly, LNP 112 encapsulating OVA mRNA demonstrated a strong therapeutic effect, substantially reducing tumor burden compared to any other treatment group over the course of 21 d (Figure 4b) and significantly extending survival (Figure 4c), with a median survival time of 31 d. Notably, of the 12 mice treated with LNP 112 containing OVA mRNA, 2 mice entirely cleared their tumors following the boost dose.

To characterize the immune response to vaccination with LNP 112, we collected mouse sera 7 d following the boost dose. To characterize the humoral immune response, we quantified reciprocal endpoint anti‐OVA antibody titers using enzyme‐linked immunosorbent assays (ELISAs). Interestingly, despite the clearly superior performance of LNP 112 to SM‐102 LNPs, total anti‐OVA immunoglobulin G (IgG) appeared to be comparable between the groups (Figure 4d). However, analysis of antibody isotypes showed significantly higher endpoint antigen‐specific immunoglobulin G 1 (IgG1) titers for the LNP 112‐treated group (Figure 4e), with comparable endpoint immunoglobulin G 2c (IgG2c) titers (Figure 4f). As a result, quantified endpoint antigen‐specific IgG1:IgG2c titer ratios were significantly greater for mice treated with LNP 112 compared to those treated with SM‐102 LNPs (Figure 4g). As IgG1 production is indicative of a type 2 helper T (Th2) immune response while IgG2c production indicates a type 1 helper T (Th1) response, these results suggest a greater Th2 bias in mice vaccinated with LNP 112 compared to SM‐102 LNPs. The improved efficacy of LNP 112 in evoking anti‐tumor responses may therefore be due to enhanced induction of humoral immunity compared to SM‐102 LNPs.

Previous findings suggest that a Th1‐biased response may be favorable for cancer vaccination [50]; however, we observed that the more Th1‐biased SM‐102 OVA mRNA LNP produced weaker antitumor effects than the relatively Th2‐biased LNP 112. Future studies should more thoroughly investigate potential mechanisms of this phenomenon. We believe that cellular immune responses not captured here may be partially responsible for the observed potent antitumor effect of LNP 112. More detailed characterization of the production of cytokines such as type II interferon (IFN) would likely be useful as secondary confirmation of immune activation and antitumor response [51, 52]. If necessary, immune responses could be tuned through the inclusion of additional adjuvants during vaccination. Moreover, evaluation of LNP 112 and other lead LNP candidates in non‐melanoma contexts would be informative of their promise in cancer vaccination and immune modulation more broadly.

Having evaluated the organ and cellular tropism of our mRNA LNPs, we next sought to investigate possible factors driving LNP tropism. The nanoparticle protein corona has been identified as a major determinant of in vivo fate [53]. LNPs in particular are known to interact with ApoE in circulation, which interacts with LDLRs on cells in the liver to endow much of the classical hepatic tropism of LNPs [18, 19, 20, 21]. Research in recent years has identified other candidate serum proteins that may help to direct LNP tropism to extrahepatic targets such as the spleen and the lungs through their participation in the formation of a nanoparticle protein corona [22]. One of the serum proteins currently thought to be responsible for the spleen tropism of some LNP formulations is β2‐GPI [22, 43]. We therefore sought to evaluate the adsorption of both ApoE and β2‐GPI to the surface of our lead mRNA LNP formulations to better understand the mechanisms behind their tropism.

Previous studies of LNP protein corona formation have largely relied on classical proteomics approaches based on centrifugation and mass spectrometry [22]. While effective and useful, these assays are labor‐intensive and slow and rely on the skilled interpretation of complex datasets—a difficult feat given the diversity of chemical species found in both LNPs and biological fluids. Additionally, these assays only give insight into bulk properties of an aggregate of LNPs, not single‐particle resolution. To gain insight into protein adsorption with single‐particle resolution in a rapid manner, we developed a novel experimental approach based on small‐particle flow cytometry. Based on previous studies analyzing LNP properties at a single‐particle level using custom flow devices, we employed a commercially available flow cytometer designed for the detection of extracellular vesicles (EVs) to analyze protein binding to LNPs. We reasoned that leveraging (i) flow cytometry, a widely used technique in the nucleic acid delivery field, and (ii) a commercially available flow cytometer would help to maximize the accessibility of our experimental technique. We fluorescently dyed NanoLuc reporter mRNA and encapsulated it in LNPs formulated containing a small amount of dyed phospholipid. We then fluorescently dyed recombinant mouse ApoE and β2‐GPI proteins. After establishing suitable parameters for the detection of single LNPs, we incubated fluorescently dyed NanoLuc LNPs with dyed proteins of interest at a fixed molar ratio. After incubation, we diluted samples to minimize further interactions and acquired data using flow cytometry. Adsorption of protein to candidate LNPs resulted in an increase in protein fluorescence, which we used to quantify the proportion of LNPs with adsorbed protein (Figure 5a).

Upon analysis of flow cytometry data, we were able to clearly resolve three distinct event populations: free mRNA, mRNA‐loaded LNPs, and empty LNPs (Figure 5b). Using fluorescent polystyrene size calibration beads, we confirmed that the diameter of analyzed particles was generally between 100 and 200 nm, in line with dynamic light scattering (DLS) measurements of LNP size (Figure 5c). Having confirmed the validity of our measurements, we then analyzed particle fluorescence following incubation either solely with unlabeled bovine serum albumin (BSA) or with a mixture of unlabeled BSA and fluorescently labeled ApoE or β2‐GPI. As expected, we generally observed substantial adsorption of ApoE to the tested LNP formulations (Figure 5d,e). Somewhat surprisingly, however, SM‐102 LNPs demonstrated minimal adsorption of ApoE. Our group has previously reported that ApoE readily binds to 1,2‐distearoyl‐sn‐glycero‐3‐phosphoethanolamine (DOPE)‐containing LNPs, which may help explain this finding [17]. Importantly, as some leukocytes express LDLRs, we do not expect adsorption of ApoE to contribute solely to hepatic tropism: instead, it is likely that increased adsorption of ApoE enhances trafficking to both the spleen and the liver, to possibly differential extents [54, 55].

The mechanism(s) by which β2‐GPI adsorption leads to spleen tropism, if any, remain unclear, though this effect has been postulated to stem from the presence of externalized anionic phospholipids in the spleen [22]. However, our group has previously shown that increased β2‐GPI binding capacity may improve transfection of splenocytes, with more modest increases in hepatic transfection [43]. We therefore anticipated increased binding affinity for β2‐GPI for our spleen‐tropic LNP formulations. Interestingly, however, we observed the strongest β2‐GPI adsorption for LNP 97, our strongest liver‐tropic LNP, which demonstrated over fivefold greater β2‐GPI adsorption than any of our spleen‐tropic LNP formulations (Figure 5f,g).

To validate our findings from single‐protein experiments, we next performed competitive binding experiments wherein we incubated dyed LNPs with equimolar amounts of dyed ApoE and β2‐GPI and assessed the adsorption of both proteins to the LNP surface. Competitive binding results generally recapitulated those from single‐protein investigations, though we observed an expected decrease in protein binding in some cases (Figure 5h,i). We also analyzed competitive adsorption of ApoE and β2‐GPI to previously‐reported selective organ targeting (SORT) LNP formulations, observing consistency between our findings using small particle flow cytometry and published results (Figure S17). In all, our results suggest a nuanced relationship between serum protein adsorption and transfection of hepatic and extrahepatic targets and support previous findings of differential adsorption of serum proteins based on ionizable lipid structure [56].

The insights into protein adsorption to LNPs offered by our small particle flow cytometric method would seem to challenge existing dogma that greater degrees of ApoE adsorption bolster hepatic transfection, whereas greater amounts of adsorbed β2‐GPI confer stronger splenic transfection [22]. This model is based on the idea that a greater density of adsorbed protein affords more opportunity for ligand‐receptor interactions, leading to a greater probability of localization to target cells and/or receptor‐mediated endocytosis. However, our group and others have previously demonstrated that an intermediate amount of targeting moieties on the nanoparticle surface yields the greatest cellular targeting effect [57, 58]. It is conceivable that, similarly, an intermediate density of adsorbed protein may be maximally beneficial for cellular uptake and organ tropism, perhaps due to steric crowding or similar phenomena. It is also likely that additional serum proteins other than ApoE and β2‐GPI are involved in LNP interactions with cells, and there may be additional “cofactors” necessary for effective targeted transfection of particular organs and cell populations. In all, our findings do not conclusively refute existing models of LNP protein corona formation and its influence on organ and cellular tropism, but they do draw into question the universality of these models and identify a need for greater detail in analysis of protein adsorption and subsequent cellular interactions. We expect that the new small particle cytometry‐based method we report here will facilitate such detailed analyses in future work. Furthermore, though we did not observe a significant time dependency of protein corona adsorption in our experiments, adopting this technique to analyze time‐dependent protein adsorption (i.e., hard versus soft protein corona formation) in a more complex environment may provide additional insight into the nuances of LNP protein corona formation in the future. Ultimately, future studies will need to evaluate new models of protein corona formation in vivo to establish causal relationships between protein adsorption and LNP fate through such means as conditional gene knockout or targeted sequestration of candidate proteins [59, 60].

LNPs were produced via microfluidic mixing essentially as described previously [61]. Briefly, all lipid components were combined in ethanol, mRNA was placed in acidic citrate buffer, and the phases were combined using chaotic mixing in a microfluidic device containing staggered herringbone micromixers at an aqueous:organic flow rate ratio of 3:1. Resultant LNPs were dialyzed against PBS for 2 h and stored at 4 for later use.

mRNA entrapment was measured using a RiboGreen assay essentially as described previously [25]. Briefly, LNPs were diluted 100‐fold in tris‐EDTA (TE) buffer or TE buffer containing 0.1% Triton X‐100. RNA standards of known concentration were plated along with diluted LNP samples in a well plate and fluorescence detection was used to determine RNA concentration in each sample. Encapsulated mRNA concentration was determined by subtracting the concentration in TE buffer (free mRNA) from that in buffer containing detergent (total mRNA). mRNA entrapment efficiency was determined as the ratio of encapsulated RNA concentration to total RNA concentration.

LNP size distributions were assessed via DLS using a DynaPro plate reader (Wyatt Technology Corporation, Santa Barbara, CA) following 10‐fold dilution in PBS.

All animal use was in accordance with the guidelines of and with approval from the Institutional Animal Care and Use Committee (IACUC) at the University of Pennsylvania (protocol number 806540). C57BL/6J (strain 000664) or APOE^–/– (strain 002052) mice were purchased from the Jackson Laboratory and bred according to standard procedures. For all animal experiments, the ionizable lipids used were at least 90% pure by liquid chromatography‐MS (LC‐MS).

NGS data were analyzed and visualized similarly to previously described using the R statistical programming language with a number of packages from the Comprehensive R Archive Network (CRAN) and Bioconductor [11, 27, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74]. Briefly, reads were demultiplexed and trimmed of adapter sequences. Barcode and UMI sequences were then extracted from reads using the UMI‐tools Python package [75]. Barcode sequences were matched to candidate sequences using the DNABarcodes R package [27], and UMIs were collapsed to yield read counts for each barcode. Normalized accumulation was calculated as the ratio of output read fraction to input read fraction [13]. Specifically, read fractions were first computed by dividing the read count for each barcode by the total number of barcode counts within each sample. Each sample read fraction (output) was then divided by the corresponding mean read fraction from the uninjected pool samples (input) to yield normalized accumulation.

For enrichment analysis, Wilcoxon rank‐sum tests were performed to compare normalized accumulation for each barcode to the normalized accumulation of all other barcodes using the normal approximation with continuity correction. False discovery rate was controlled using the method of Benjamini and Hochberg [76].

All data were analyzed using the R statistical programming language with several packages from the Comprehensive R Archive Network (CRAN) [62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 81, 82, 83, 84, 85, 86]. The Nix package manager (with pinned Nixpkgs revision 3ff0e34b) was used for all data and dependencies to maximize reproducibility [87]. Unless otherwise noted, statistical inference was performed on mean responses using one‐way analysis of variance (ANOVA) with post hoc t tests using the Bonferroni‐Holm correction for multiple comparisons. For comparisons involving only two groups, simple unpaired t tests were used for inference. For melanoma model survival analysis, pairwise log‐rank tests were employed to compare survival curves to PBS‐treated mice, and the Bonferroni–Holm correction was applied [88].