A single-cell map of antisense oligonucleotide activity in the brain

Abstract Antisense oligonucleotides (ASOs) dosed into cerebrospinal fluid (CSF) distribute broadly throughout the central nervous system (CNS). By modulating RNA, they hold the promise of targeting root molecular causes of disease and hold potential to treat myriad CNS disorders. Realization of this potential requires that ASOs must be active in the disease-relevant cells, and ideally, that monitorable biomarkers also reflect ASO activity in these cells. The biodistribution and activity of such centrally delivered ASOs have been deeply characterized in rodent and non-human primate (NHP) models, but usually only in bulk tissue, limiting our understanding of the distribution of ASO activity across individual cells and across diverse CNS cell types. Moreover, in human clinical trials, target engagement is usually monitorable only in a single compartment, CSF. We sought a deeper understanding of how individual cells and cell types contribute to bulk tissue signal in the CNS, and how these are linked to CSF biomarker outcomes. We employed single nucleus transcriptomics on tissue from mice treated with RNase H1 ASOs against Prnp and Malat1 and NHPs treated with an ASO against PRNP. Pharmacologic activity was observed in every cell type, though sometimes with substantial differences in magnitude. Single cell RNA count distributions implied target RNA suppression in every single sequenced cell, rather than intense knockdown in only some cells. Duration of action up to 12 weeks post-dose differed across cell types, being shorter in microglia than in neurons. Suppression in neurons was generally similar to, or more robust than, the bulk tissue. In macaques, PrP in CSF was lowered 40% in conjunction with PRNP knockdown across all cell types including neurons, arguing that a CSF biomarker readout is likely to reflect ASO pharmacodynamic effect in disease-relevant cells in a neuronal disorder. Our results provide a reference dataset for ASO activity distribution in the CNS and establish single nucleus sequencing as a method for evaluating cell type specificity of oligonucleotide therapeutics and other modalities.


INTRODUCTION
Antisense oligonucleotides (ASOs) can, in principle, modulate the expression of nearly any gene in the central nervous system (CNS) ( 1 ). Bolus injected into cerebrospinal fluid (CSF) ( 2 ), ASOs are internalized by cell surface proteins, escape from endosomes ( 3 , 4 ) and become durably acti v e in the cytoplasm and nucleus ( 5 ). Since the 2016 approval of nusinersen, an ASO modulator of pre-mRNA splicing in spinal muscular atrophy ( 6 ), 23 ASOs have entered trials for CNS disorders ( 1 , 7 ), with several advancing to Phase III, along with others administered on an 'N-of-1' basis (8)(9)(10)(11). The majority of CNS ASOs in trials today are 'gapmers' --ASOs with 2' sugar modifications in the wings (typically 5 base pairs on either side) and a 'gap' in the middle with no modifications except for a phosphorothioate backbone ( 2 ) --designed to lower the expression of a target RNA by recruiting the enzyme RNase H1 to cleave it (12)(13)(14). For CNS diseases caused by a toxic gain of function, gapmer ASOs offer a rational approach to target the root cause of disease by lowering the toxic RNA or protein ( 1 ).
This rational mechanism is only useful, howe v er, if the drug can be deli v ered to the right tissues and the right cell types ( 15 ). Following the seminal discovery that bolus injection into CSF is an efficient means of deli v ering ASOs to the CNS ( 16 ), se v eral studies hav e demonstrated drug uptake and durable target RNA silencing for gapmers in rodent and non-human primate (NHP) brain (17)(18)(19)(20)(21). Meanwhile, the limited available autopsy data have confirmed that ASOs distribute to multiple regions of spinal cord and brain in humans ( 11 , 22 ). Jafar-nejad and Powers et al. in a thorough pharmacology atlas spanning 35 brain regions in NHP, demonstrated a drug concentration gradient, with more accumulation in superficial than in deep brain structures, but at least some tar get RNA lo wering was observed across all regions studied ( 21 ). At the bulk tissue le v el, the above studies demonstrate widespread CNS biodistribution and activity of centrally delivered ASOs.
Availab le e vidence also suggests that ASO activity is br oadly distributed acr oss cell types within the CNS. ASOs have been successfully employed to ameliorate CNS disease in animal models with pathologies specific to various cell types including astr ocytes, oligodendr ocytes, micr oglia, and se v eral neuronal sub-types ( 18 , 23-29 ). Histological analysis has been used to qualitati v ely demonstrate ASO uptake and target RNA modulation across di v erse CNS cell types in both rodent and NHP ( 21 ). Mor eover, dose-r esponse r elationships were determined for an ASO in four cell types isolated from mouse cortex at two weeks post-dose ( 21 ). Thus, whereas rather marked differences in ASO uptake or activity between cell types have been observed in liver and lung ( 30 , 31 ), the above findings suggest relatively broad ASO activity across cell types within the CNS.
Ne v ertheless, a mor e compr ehensi v e understanding of how ASO activity is distributed at the cell type and at the single cell le v el would be a valuable asset. Important knowledge gaps include: how bulk tissue knockdown is distributed across individual cells; quantitative assessment of knockdown across detectable cell types; duration of action by cell type; and ability to translate cell type-specific activity profiles between relevant model systems, such as from mouse to NHP, and from low-potency tool compounds used at high doses versus high-potency compounds used at low doses. These knowledge gaps are particularly salient when considering the interpretation of CSF-based target engagement biomarkers in ASO trials. Cell type-specific differences in ASO uptake or activity, combined with drug concentr ation gr adients, could gener ate variability in the degree of target engagement among relevant CNS cells, yet biomarker values from a single sampling compartment can underpin choices to advance or halt clinical programs. A deeper profiling of ASO activity across cell types should help to inform such crucial decisions.
We hypothesized that single nucleus RNA sequencing (snRNA-seq) could begin to fill the above knowledge gaps. In particular, transcriptomic information would allow us to assign cell type for each nucleus, providing a relati v ely unbiased sampling of cells in a tissue. Comparing the number of target RNA counts within those transcriptomes for ASO v ersus v ehicle-treated animals could then be used to quantify target engagement for any group of cells, such as a cell type or subtype, and could e v en provide information about the distribution of activity across single cells. Here, we employed snRNA-seq to refine our understanding of ASO activity in the mouse and cynomolgus macaque CNS. Our results illuminate the broadness of RNase H1 ASO target engagement across individual cells and across cell types, reveal cell type-specific differences in extent of target RNA lowering and in duration of action, link neuronal target engagement to a CSF biomarker outcome, and establish a reference dataset and a methodology for assessing the cell type specificity of oligonucleotide therapies.

Mice
All mice were female C57BL / 6N. Animals for 3 week postdose harvest were dosed at the Broad Institute (IACUC protocol 0162-05-17) and were 16 weeks old at the time of dosing. Animals for 2 and 12 week post-dose harvest were dosed at Ionis Pharmaceuticals (IACUC protocol 2021-1176) and were 8-12 weeks old at dosing. Mice were Nucleic Acids Research, 2023, Vol. 51, No. 14 7111 dosed via intracerebroventricular injection as described ( 32 ). ASOs were deli v ered as a single bolus injection of 500 g ( Prnp ASOs) or 50 g ( Malat1 ASO) formulated in a 10 l volume of dPBS. Mice were perfused with HEPESsucrose solution (110 mM NaCl, 10 mM HEPES, 25 mM glucose , 75 mM sucrose , 7.5 mM MgCl 2 , 2.5 mM KCl, pH 7.4) and brains harvested as described ( 33 , 34 ). , and ICH-S8 (Immunotoxicity Studies for Human Pharmaceuticals). Animals were 2-4 years old at injection, mixed sex (2M / 2F per cohort), and were of Asian origin. Lumbar punctures were perf ormed on da ys 1, 29, 57 and 85. The procedure was performed fasting under ketamine / medetomidine anesthesia with a pencil-point pediatric needle at a position between L2 and L6. First, ≥0.5 ml of CSF was collected, then, 20 mg ASO was deli v ered in a 1 ml volume of artificial CSF (aCSF) injected over 1 minute, followed by a flush of 0.25 ml aCSF. 15 minutes after the procedur e, animals wer e awakened with atipamezole. CSF was ejected into Protein LoBind tubes and snap frozen in liquid nitrogen. The CSF samples analyzed here were collected at day 85, just prior to the fourth dose, while brain tissues were collected at day 92. Because the majority of CSF volume was used for regulated studies, the aliquots available for analysis in this study varied from 120-300 l and 0.03% CHAPS was added only after fr eeze / thaw; these pr e-anal ytical factors likel y contribute additional variability between samples ( 35 ).

Tissue dissection
For mouse brains, cryostat (Leica CM3050 S) dissection was performed as described ( 36 ): after storage in optimal cutting tempera ture (O.C .T.) compound (Tissue-Tek 4583) a t -80 • C , mouse brains were mounted by the frontal cortex onto cryostat chucks with O.C.T. leaving the entire posterior half of the brain exposed. A ∼2.5 mg piece of tissue was then excised using a pre-chilled ophthalmic microscalpel (Feather P -715) and placed into a pre-chilled PCR tube. For mice, a piece of somatosensory cortex was used for snRNAseq, while an adjacent piece of visual cortex was used for bulk qPCR; thalamus was cut along the fiber tract and the dorsal half was used for snRNA-seq while the ventral half was used for qPCR; cerebellum was cut through the ansiform lobule and a piece of simple / ansiform lobule was used for snRNA-seq while a piece of ansiform / paramedian lobule was used for qPCR. Cynomolgus brains were coronally sectioned at a thickness of 4 mm, and cylindrical tissue punches of 2 mm diameter were taken for RN A anal ysis and of 6 mm diameter for protein analysis. The 2 mm diameter by 4 mm length cylindrical tissue punch was then sectioned lengthwise into quarters on the cryostat and one quarter was used for single cell analysis. From the most rostral section containing frontal cortex, punches were taken from middle frontal gyrus across all histological cortical layers. From section 13-14, where both cerebellum and medulla are visible, punches were taken from the posterior lobe of the cerebellum across the granular, ganglionic and molecular layers.

Bulk tissue qPCR
Tissue pieces dissected on the cryostat were placed in RN Alater-ICE (Invitro gen AM7030) and allowed to thaw overnight at -20 • C. Once samples were thawed, tissue was homogenized in 1 ml QIAzol lysis reagent, using 3 × 40 s pulses on a Bertin MiniLys homogenizer in 7 ml tubes preloaded with zirconium oxide beads (Precellys CK14, Bertin KT039611307.7 / P000940-LYSK0-A). RNA was isolated from homogenate using RNeasy Lipid Tissue Mini Kit (Qiagen 74804) per the manufacturer protocol. RNA was eluted with 40 l RNase-free water. RT-PCR samples were pr epar ed using Taqman 1-Step RT-PCR master mix (Invitrogen) and Taqman gene expression assays (Invitrogen) for mouse Prnp (Mm00448389 m1; spanning exons 1-2) and mouse Tbp (Mm00446971 m1) and for cynomolgus TBP (Mf04357804 m1). The following gene-specific primerprobe sets were custom ordered from IDT: Malat1 (mouse), Forward: A GGCGGGCA GCTAA GGA, Reverse: CCC-CA CTGTAGCATCA CAT CA, Probe: TT CCT CTGCCG-GT CCCT CGAAAG; PRNP (cynomolgus; spanning intron 1-exon 2), Forward: CCT CT CCT CA CGA CCGA, Re v erse: CCCAGTGTT CCAT CCT CCA, Probe: CCA-CAAA GA GAACCA GCATCCA GCA. Samples were run on a QuantStudio 7 Flex system (Applied Biosystems) using manufactur er's r ecommended cycling conditions. Each biological sample was run in duplicate and the le v el of all targets were determined by Ct whereby results were first normalized to the housekeeping gene Tbp and then to PBSor aCSF-treated animals.

Single cell sequencing
After cryostat dissection, samples were batched in groups of eight, chosen to include treated and control animals in e v ery run. Single nucleus suspensions were prepared as described ( 37 , 38 ). Briefly: tissue samples were triturated, by pipetting, in an extraction buffer containing Kollidon VA64, Triton X-100, bovine serum albumin, and RNase inhibitor, then passed through a 26-gauge needle, washed and pelleted, then passed through a cell strainer. Nuclei positi v e for DAPI signal were isolated by fluorescence-activated cell sorting with a Sony SH800 or MA900 calibrated with a 70 m chip, with a 405 nm excitation laser and light collected with a 425-475 nm filter. Sorted nuclei were counted using a Fuchs-Rosenthal C-Chip hemocytometer and a hand tally counter. A volume chosen to target 17 000 nuclei was submitted to the Broad Institute's Genomics Platform, where 10X library construction (3' V3.1 NextGEM with Dual Indexing) was performed according to manufacturer instructions ( 39 ). Libraries were sequenced on an Illumina Novaseq 6000 S2 for 100 cycles.

Data processing and analysis
Raw binary base call (BCL) files were synced to Google Cloud and analyzed on Terra.bio. Cumulus ( 40 ) Cell Ranger ( 41 ) 6.0.1 (cellranger workflow v28) was employed, with flags -include introns and -secondary set to true, to process BCL files into unique molecular identifier (UMI) count matrices for each individual sample. Mouse samples were aligned to Cell Ranger r efer ence package mm10-2020-A and cynomolgus samples were aligned to a custom Cell Ranger r efer ence made from Ensembl Macaca fascicularis 6.0 (release 108). Matrices were then aggregated using Cell Ranger 7.0.1 (aggr with the -normalize flag set to none) to yield one UMI count matrix per species and brain region. Statistical analyses and data visualization were conducted using custom scripts in R 4.2.0.

Cell type assignment
Aggregated count matrices were examined using Loupe Bro wser. V iewing cells in 2-dimensional uniform manifold approximation and projection (UMAP) ( 42 ) space, we looked for cell type markers established or validated in se v eral prior single-cell studies ( 36 , 43-49 ). Clusters corresponding to empty droplets , doublets , debris , or mitochondria were flagged and removed based on low UMI or unique gene count, low percentage intronic reads, lack of obvious differ entially expr essed genes, high expr ession of mitochondrial genes, or location between two other clusters and expression of markers of each. Assignments were then validated by generating dot plots in Seurat V4 ( 50 ) in R. For cortical excitatory and inhibitory neurons in 3 week postdose animals, a list of barcodes was exported from R and r ecluster ed in Loupe Browser.

Pharmacokinetic studies
Quantification of ASO N in NHP tissue was performed as described ( 51 ). Briefly, tissue samples were weighed, homogenized, and extracted first using a liquid-liquid extraction (LLE) with ammonium hydroxide and phenol:chloroform:isoamyl alcohol (25:24:1), followed by a solid phase extraction (SPE) using a 96-well Strata X packed plate (Phenomonex), followed by a pass through using a Protein Precipitation Plate (Phenomonex). Eluates were dried down under nitrogen at 50 • C before reconstituting with 100 l water containing 100 M EDTA. Samples were then analyzed by ion-pairing (IP) LCMS / MS with an Agilent 6460 LCMS / MS system (Agilent), using an AC-QUITY UPLC OST C18 column (Waters) heated to 55 • C with a flow rate of 0.3 ml / min. The column was equilibrated with 400 mM HFIP / 15 mM TEA in water. A gradient from 10 to 40% MeOH over 6 min was used to elute ASO N. All mass measurements were made on-line with MRM transitions of m / z 881.6 and 773.2 both with a product ion of 94.8 for ASO N, and the internal standard, respecti v ely. Mass spectra were obtained using a spray voltage of -1500 V, a nebulizer gas flow of 25 psig, a sheath gas flow rate of 12 l / min at 350 • C, a drying gas flow rate of 5 l / min at 350 • C, and a capillary voltage of -3750 V. Chromatograms were analyzed using Agilent Mass Hunter software. ASO N concentration was determined from its calibration curve with dynamic range 0.035 g / g (0.05 M) to 176.95 g / g (25.0 M).

Statistics
For each combination of brain region, timepoint, and treatment condition, snRNA-seq data were grouped by animal and cell type and the sum of target UMIs and total UMIs was calcula ted. A nega ti v e binomial model was fit to the resulting data, with target RNA UMIs as the dependent variable; cell type and a cell type-treatment interaction term as the dependent variables, and total UMIs as the offset. This utilized the MASS package in R, with the call: glm.nb(target umi ∼ celltype + celltype:treatment +offset(log(total umi))). This returns coefficients in natural logarithm space. For the ASO-treated conditions, the coefficient for each cell type-treatment interaction term coefficient was then exponentiated to yield the mean estimate of the residual target RNA in that cell type. The 95% confidence interval was defined as that mean estimate ±1.96 of the standard errors returned by the model. Each individual animal's point estimate of residual target RNA in each cell type was obtained by adding the residual from the model to the cell type-trea tment coef ficient, and then exponentiating. To account for the different abundance of different cell types, which impacts the size of our confidence intervals on tar get knockdo wn, w e used w eighted Pearson's correlations (wtd.cor from the weights package in R) to test candidate variables and weighted standard deviations (square root of wtd.var from the Hmisc package in R) to evaluate the variability in target engagement between cell types within differ ent brain r egions. Thr oughout, all err or bars and shaded areas in figures represent 95% confidence interv als. P v alues less than 0.05 were considered nominally significant.

Generation and cell type classification of single nucleus transcriptomes
We selected 4 previously characterized ASOs: 2 Prnp ASOs that extend survival in prion-infected mice ( 32 , 52 ), 1 Malat1 ASO with e xtensi v e pharmacology data ( 21 ), and 1 human PRNP ASO sequence-matched in macaques ( 53 ) ( Table 1 ). We analyzed a total of 78 single nucleus transcriptomes from tissues of mice and macaques treated with these ASOs or with vehicle (Supplementary Table S1-2) totaling 598066 single nuclei. Samples averaged 532 million reads mapping to 7667 cells and yielding 7650 unique molecular identifiers (UMIs) per cell, corresponding to a median of 3108 detected genes per cell (Supplementary Table S1). Transcriptomes wer e aggr egated by species and brain region yielding fiv e count matrices. Distinct clusters were assigned cell types

Distribution of ASO activity at the single cell level
The long non-coding RNA Malat1 is a valuable model target for single cell assessment of ASO activity because it is highly expressed, accounting for 11.4% of all UMIs in our mouse transcriptomes, and because a potent and well-characterized tool ASO against Malat1 is available ( 21 ). The high expression means that Malat1 averages hundreds of UMIs per individual cell --an asset when evaluating knockdown in single cells --whereas most genes exhibit Poisson distributions with many zeroes (Supplementary Table S11) ( 55 ). Considering the Malat1 ASO's median effecti v e dose (ED 50 ) of ∼50 g in cerebellum ( 21 ) Tables S8-S10). snRN A-seq inherentl y yields low sequencing coverage in any one nucleus, meaning that most genes are not detected in most nuclei, e v en where they are expressed (Supplementary Table S11). Unlike the highly expressed Malat1 , most potential ASO targets will have UMI counts that are Poisson or negati v e binomial distributed in single nuclei data. For instance, Prnp averaged just 0.85 UMIs / cell in the cerebella of PBS-treated animals. Ne v ertheless, when the data from 12 weeks after a single 500 g dose of ASO 6, were fit to the same negati v e binomial model as Malat1 , ASO 6 displayed a highly similar pattern of activity across cerebel-lar cell types (rho = 0.96, P < 3.9e-8, weighted Pearson's; Figure 2 E; Supplementary Table S10).
Despite lower basal expression , we posited that examination of UMI / cell histograms for Prnp could re v eal information about the distribution of drug activity across single cells. As an example, we compared histograms for astrocytes, which had 56% residual Prnp in ASO 6-treated animals, v ersus three models. A negati v e binomial model fit to the PBS-treated animals mirrored those animals' actual distribution almost perfectly. Lowering Prnp to 56% residual by setting 44% of astrocytes' Prnp counts to zero would have yielded a histogram with far more zeroes, and fewer ones, than the observed distribution in ASO 6-treated animals. In contrast, lowering Prnp to 56% residual by lowering the negati v e binomial parameter mu by 44%, corresponding to equal knockdown in all cells, yielded a distribution nearly identical to that in ASO-treated animals ( Figure 2 F; Supplementary Table S13). Thus, for Prnp as for Malat1 , bulk tissue knockdown appears to arise from broad knockdown in all cells, albeit with a stereotypical pattern of differences across distinct cell types.

ASO activity across regions and cell types in the mouse brain
We assessed the profile of ASO target engagement across cell types in 3 brain regions in mice at 3 weeks post-dose with Prnp ASO 6 ( Figure 3 ). Because this tool compound is less potent than the Malat1 ASO, we used a 500 g dose, which modifies prion disease in mice ( 32 , 52 ) and lowers whole hemisphere PrP to an estimated 56% residual after 4 weeks ( 56 ). Whereas Malat1 localizes to the nucleus ( 57 ), Prnp is a protein-coding gene w hose mRN A reaches the cytosol, and ASO 6 targets the Prnp 3'UTR, so cytosolic activity is possible. Nonetheless, we found that the value of residual Prnp obtained by snRN A-seq, w hich will detect nuclear ASO activity onl y, agreed closel y with the value obtained by bulk tissue qPCR, which used exon junction-spanning primers and ther efor e will only detect mature mRNA (Figure 3 A-C).
Breakdown of single cell data by cell type showed broad target engagement across cell types including di v erse types of neurons and glia (Figure 3 D). As with the Malat1 ASO (Figure 2 ), cell type differences were relati v ely pronounced in the cerebellar neurons, where knockdown was deeper in Purkinje cells and MLI than in granule cells. Across regions in ASO-treated animals, endothelial stalk cells, pericytes, and fibroblasts generally had both the highest residual Prnp and the lowest count of cells sequenced, giving rise to wide confidence intervals that overlapped the PBS-treated animals. Nonetheless, point estimates for these cells generally

Potency and duration of action across ASO chemistries
Ga pmer ASOs currentl y in clinical trials are 2'MOE ga pmers similar to ASO 6, but improved chemical modifications of ASOs are a highly acti v e area of research ( 58 , 59 ), prompting us to investigate the cell type profile of an ASO incorpor ating 2'-4' constr ained ethyl (cEt) modifications ( 60 ). Prnp ASO 1 (Table 1 ), a mixed 2'MOE / cEt oligonucleotide, targets the same site as ASO 6 and is effecti v e in prion-infected mice ( 32 , 52 ). We evaluated the activity of ASO 1 and ASO 6 in mouse cortex at both 2 and 12 weeks after a single 500 g bolus dose (Figure 4 ). Again, single cell and bulk qPCR measurements of overall knockdown concurr ed (Figur e 4 A). ASO 1 had a shorter duration of action than ASO 6, with residual target rising from 47% to 91% of saline controls (by bulk qPCR) residual, a 44% recov ery, v ersus 31% to 65%, a 34% recov ery, for ASO 6 ( Figure 4 A; Supplementary Table S7). Each compound provided substantial knockdown at 2 weeks across all cell types detected, and each exhibited marked differences across cell types in the rate of recovery (Figure 4 B; Supplementary Table S16). For example, for both ASOs, microglia exhibited the most complete recovery of any cell type (+51% for ASO 6 and +70% for ASO 1), while excitatory neurons were comparati v ely steady (+29% for both;

Cell type profile and biomarker impact in non-human primates
We examined tissue from cynomolgus macaques that recei v ed ASO N. In addition to permitting us to examine ASO activity in a larger brain, the macaques also differed from our mice in being dosed intrathecally (IT) rather than ICV, and receiving 4 repeat doses at 4-week intervals. Both cortex and cerebellum exhibited substantial drug accumulation (Supplementary Figure S3, Supplementary Table S17-S18). In cortex, bulk residual PRNP measured by snRNAseq again mirrored that by qPCR (Figure 5 A), although in cerebellum, knockdown measured by snRNA-seq appeared slightly deeper (Figure 5 B). Residual PrP protein le v el quantified by ELISA ( 56 ) in ASO N-treated animals was 41% in cortex, 82% in cerebellum and 60% in CSF  Table S10). In cortex, knockdown was deepest in neurons and weakest in endothelial stalk and pericytes / fibroblasts. In cerebellum, knockdown was deepest in Purkinje cells and molecular layer interneurons (MLIs) and weakest in pericytes / fibroblasts. Because these tissues were obtained just 1 week after the animals' final dose of ASO, we compared the cell type profile of target engagement in macaques to that observed in mice 2 weeks after a single dose of ASO 6 ( Figure 5 E, F). The two datasets shared robust knockdown in MLI, Purkinje, and cortical neurons and relati v ely limited knockdown in pericytes / fibroblasts. Correlation of knockdown across cell types was positi v e, though significant only in cerebellum (rho = 0.40 and 0.80, P = 0.33 and 0.0019 for cortex and cerebellum respecti v ely, weighted Pearson's).

Comparison of cell type specificity across paradigms
To ask how br oadly the pr ofile of activity across cell types was shared among all our datasets, we defined a difference from overall residual as a cell type's residual target RNA, expressed as a percentage of control animals, minus the overall residual target RNA across all cell types (Supplementary Table S21). In cortex, the most abundant cell types , chiefly neurons , clustered near 0% (excitatory neurons, mean + 1%, inhibitory neurons, mean + 2%), reflecting the bulk tissue closely; outliers were rarer cell types with wider confidence intervals (Figure 6 A). In cerebellum, howe v er, granule cells (mean + 7%) differed considerably from the next two most abundant cell types, MLIs (-30%) and Bergmann glia (-10%). Variability across cell types was lower in cortex and thalamus (mean weighted standar d de viation 7% for both) than in cerebellum (mean weighted standar d de viation 12%; Figure 6 A). Accor dingl y, we observed mostl y positi v e but non-significant correlations between pairs of cortex datasets (Figure 6   close to the bulk tissue residual (worst case, +12% difference for granule cells in Malat1 ASO-treated mouse cerebellum at 12 weeks) or exhibited much deeper target engagement (-38% for MLIs in Malat1 ASO-treated mouse cerebellum at 12 weeks; Figure 6 D; Supplementary Table S23). We did not observe any conditions in which any population of neurons exhibited dramatically weaker knockdown than the bulk tissue.

DISCUSSION
Here we deployed single nucleus transcriptomics to quantify ASO target engagement in the CNS. We observed a broad distribution of activity across individual cells, but with differences in activity and in duration of action between cell types. The profile of activity across cell types was largely shared across different ASOs examined and between mouse and NHP.

Distribution of ASO activity across individual cells
Theoretically, 50% knockdown in bulk tissue could result, in the most extreme cases, from either 100% knockdown in half of cells or 50% knockdown in all cells. A longstanding question is where the activity of CNS ASOs falls on this spectrum.
By examining the distribution of target RNA counts per cell in single nucleus sequencing data, we provide evidence of ASO activity in every single cell in a bulk tissue, albeit with differences in degree between cell types. This should be expected based on the number of drug molecules contained in a dose of ASO. For a mouse with ∼10 8 brain cells ( 61 , 62 ), a 50 g dose of a ∼7 kDa ASO, or ∼4 × 10 15 ASO molecules, is > 10 7 molecules per brain cell. Only a small minority of ASO molecules are belie v ed to undergo producti v e uptake ( 3 ), but e v en if this figure is 1%, then 10 5 producti v e ASO molecules per brain cell is a sufficient number that it is unlikely that any cells would avoid ASO activity simply by chance. Of course, there may be cells in deep brain structures lacking any appreciable ASO activity due to limited drug distribution ( 21 ); we only analyzed tissues with robust target engagement at the bulk le v el. Under this precondition, ASO activity appears very broadly distributed across individual cells.
This property of ASOs could prov e mar kedly different from some gene thera py a pproaches to CNS diseases. In mice, engineered viral vectors for gene delivery may transduce ∼50% of CNS neurons ( 63 ), and DNA-targeted thera peutics, with onl y 2 targets per cell, could provide nearl y complete target suppression in those cells that are transduced. If so, modalities exhibiting similar levels of bulk target engagement could reflect rather distinct distributions at the single-cell le v el. These contrasting profiles might in turn present opposing challenges and opportunities for different targets.

Differences in activity and duration of action between cell types
We observed differences in target engagement and in duration of action between cell types. There exist multiple possible mechanistic explanations for these differences. Distinct cell types could di v erge at various steps in ASOs' cellular pathway ( 3 ) including differences in gross uptake, in the proportion of producti v e uptake, in the kinetics of endosomal escape, in the rate of RNase H1 cleavage, in the rate of release from the cell, in activity of nucleases that degrade the ASO, or in the presence or absence of cell division diluting out the ASO.
While the short duration of action in microglia would at first glance appear consistent with a role for dilution by cell division, the estimated microglial turnover rate (median lifetime > 15 months ( 64 )) is too slow to a ppreciabl y dilute ASO. Moreover, we observed comparatively weak knockdown in cells of the vasculature --fibroblasts , pericytes , and endothelial cells --which are described as largely quiescent in the adult brain ( 65 ).
Histolo gical anal ysis of ASO-treated brain tissue indica tes tha t the dif ference in ASO activity between granule and Purkinje cells may be due to total ASO uptake ( 21 , 27 ). It may be, howe v er, that not all differences between cell types are explained simply by gross uptake. In lungs of mice trea ted intra tracheally with divalent siRNA, fibroblasts exhibited deeper target engagement than other cell types despite lower drug accumulation ( 66 ); an LNA ASO was similarly most acti v e in lung fibroblasts ( 31 ). In our dataset, across cell types in the mouse cortex, deeper initial target engagement at 2 weeks appeared to correlate with more washout by 12 weeks. This correlation is expected to some degree, because target expression after washout should ne v er recov er to > 100% of the untreated condition, but may also suggest that deeper initial knockdown in some cell types does not necessarily indicate a longer-lasting endosomal repository of compound.
Our dataset is ill-suited to ask genome-wide questions such as which specific cell surface proteins are most important for uptake, because any two cell types differ in the expression of many markers, not just one, and in addition, the thousands of possible answers present a large multiple testing burden which cannot be overcome by analyzing the small number of distinct cell types detected here. None of the few specific hypotheses we tested appear to explain the cell type differences we observed. Cell size might be inversely related to the surface area to volume ratio, and thus to the amount of opportunity for cell surface protein binding, but UMIs / nucleus, a proxy for cell size ( 36 ), was not correlated with ASO activity in our dataset. RNase H1 expression varied little across cell types and neither RNase H1 nor target expr ession corr elated with initial target engagement or washout. In fact, this should be expected based on the number of drug molecules per cell. PrP RNA expression is on the order of hundreds of transcripts per million ( 67 ), so a cell with 10 5 mRNA molecules might have just tens of PrP mRNA molecules, not nearly enough to sa tura te 10 5 producti v ely uptaken ASO molecules.

Limitations of this study
Our study has many limitations. The expense of singlecell sequencing limited us to small cohort sizes (usually N = 4). For some rarer cell types, just a handful of cells per sample were observed. Many steps including nuclei disso-ciation, flow cytometry, and library construction, can all yield variability in number of cells and number of sequencing reads per sample. All of these factors combine to make the confidence intervals on our estimates of knockdown in many cell types rather large.
Certain key observations replicate across our datasets --particularly the broadness of target engagement across cell types, with weaker knockdown in granule cells and deeper knockdown in Purkinje and inter lay er neurons, and the generally weaker knockdown in cells of the vasculature. Howe v er, we studied only 3 brain regions, 4 ASOs, 2 targets, and 2 animal species, so it remains to be determined just how broadly these findings may generalize. We observed similar patterns of target engagement across cell types in mice treated with a high dose (500 g) of a low potency ASO targeting Prnp and with a low dose (50 g) of a high potency ASO targeting ubiquitously expressed Malat1 , howe v er, these e xperiments used different compounds against different targets, and it remains to be determined how the cell type specificity of ASOs differs as a function of dose response for a single compound.
Like wise, we observ ed dura tion of action dif ferences in certain cell types between a 20-nucleotide 2'MOE gapmer ASO and a 17-nucleotide 2'MOE / cEt ASO, but without testing a more thorough battery of compounds, it is impossible to discern whether these differences are effects of chemical backbone, of length, or simply of random chance.
We lack any method of quantifying drug concentration in the same cells that are sequenced, so are unable to answer questions about the phar macokinetic / phar macodynamic rela tionship a t the single-cell le v el.
Because we relied on purification of nuclei from frozen tissue, w e w er e only able to measur e target engagement in the nucleus. It is reassuring that snRNA-seq and qPCR generally a greed, b ut these anal yses were necessaril y performed on adjacent pieces of tissue, making it unclear whether their occasional di v ergence r epr esents discordance between cytosolic and nuclear outcomes, or simply regional gradients in target engagement.

Implications for prev entiv e trials in prion disease
Pharmacologic interventions are seldom trialed in presymptomatic individuals at risk for neurodegenerati v e disease ( 68 ). Observing clinical endpoints in such individuals may r equir e lengthy f ollow-up ( 69 ) or ma y be outright numerically infeasible ( 70 ). This has led to the suggestion that in prion disease, where the central role of PrP in disease is incontrov ertib le ( 71 ), the lowering of CSF PrP --a target engagement biomarker only --could serve as a primary endpoint in trials of at-risk individuals ( 72 ). This prospect demands that especially strong data from animal studies will be needed to certify the links between CSF PrP, target engagement in the disease-relevant cells, and disease modification ( 73 ).
In prion disease, the critical cells to engage are neurons. Although astrocytes may contribute to disease by propagating prions (74)(75)(76), only neurons degenerate in prion disease, and neurotoxicity is cell autonomous: neurons that do not express PrP are protected e v en if they are in direct contact with misfolded prions produced by neighboring cells (76)(77)(78). In contrast, neuroinflammatory responses from astr ocytes and micr o glia ( 79-83 ) a ppear to be strictl y nonautonomous, requiring neuronal prion infection ( 76 ).
That PrP-lowering ASOs e xtend survi val in prioninfected mice ( 32 , 52 , 84 ) implies that they must lower PrP in neurons; ne v ertheless, we felt it prudent to further examine this link to determine whether there might e v er e xist circumstances in which a bulk tissue readout would indicate PrP lowering despite little or no target engagement in neurons. It is reassuring, then, that across a range of experimental parameters --dosing regimens, times post-dose, ASO chemistries and gapmer configurations , targets , species and brain regions --we ne v er identified a circumstance in which bulk tissue would misinform about PrP RNA having been lowered in neurons.
Of course, gi v en the some wha t dif fering activity of ASOs in distinct CNS cell types and the potential for drug concentr ation gr adients across the br ain, no single compartment readout, such as CSF PrP, can accurately report on e v ery disease-relevant cell in this whole brain disease. Still, our findings provide one pillar of support for the expectation that lowered CSF PrP in an ASO trial is reasonably likely to predict clinical benefit in individuals at risk for prion disease.

Concluding notes
Our study, together with recent data from the mouse lung ( 31 ), establishes the feasibility of using single nucleus transcriptomics to quantify ASO target engagement across cell types and individual cells within a bulk tissue. Our results answer key questions about ASO activity in the CNS, and support the utility of a CSF target engagement biomarker. Nonetheless, our studies were of limited scope relati v e to the number of questions that could be asked. Future studies of oligonucleotide drugs, particularly of ne w deli v ery routes, formula tions, or conjuga tes, should consider single cell or single nucleus transcriptomics as tools for evaluating cell type specificity and single cell distribution.

SUPPLEMENT ARY DA T A
Supplementary Data are available at NAR Online.