A CAR RNA FISH assay reveals functional and spatial heterogeneity of chimeric antigen receptor T cells in tissue

Chimeric antigen receptor (CAR) T cells are engineered cells used in cancer therapy and are studied to treat infectious diseases. Trafficking and persistence of CAR T cells is an important requirement for efficacy to target cancer and HIV sanctuary sites. Here, we describe a CAR RNA FISH histocytometry platform combined with a dnnRRS image analysis algorithm to quantitate spatial distribution and in vivo functional ability of a CAR T cell population at a single cell resolution. In situ, CAR T cell exhibited a heterogenous effector gene expression and this was related to the distance from tumor cells, allowing a quantitative assessment of the potential in vivo effectiveness. The platform offers the potential to study immune functions engineered cells in situ with their target cells in tissues with high statistical power and thus, as an important tool for preclinical and potentially clinical assessment of CAR T cell effectiveness. One Sentence Summary We developed a CAR RNA FISH assay to study chimeric antigen receptor T cell trafficking and function in human and mouse tissue. Impact statement We developed an imaging platform and analysis pipeline to study large populations of engineered cells on a single cell level in situ.

at its x y position. The circles were than expanded approximately 16 times the radius of the circle 236 and connective groups were created by overlapping the expanded circles. The largest connective 237 group was denoted as CD19high area whereas the remainder of the tissue section was defined as 238 CD19low area. The spatial density of CAR+ GZMB-, CAR+ GZMB+, CAR+ IFNγ -, CAR+ IFNγ 239 + was than calculated for the CD19high and CD19low area of each individual tissue section. 240 The RNA and protein expression levels as well as the location of the cells was exported as a 241 comma-separated values file and loaded into the flow cytometry data analysis software FlowJo 10 242 (Becton Dickinson). In FlowJo, the imaging data was analyzed with the gating and plotting tools. 243 To compare cell counts of the dnnRRA and manual operators, manual operators used the cell 244 counter function in imageJ to track cell counts. 245

Statistical analysis 246
The data were analyzed with graphpad prism software with the statistical tests stated in the figure 247 legends. One-way ANOVA combined with a Tukey test was used to compare multiple groups. *, 248 **, ***, **** correspond to p < 0.05, p < 0.01, p < 0.001 and p < 0.0001, respectively. Paired t 249 test was used to compare the spatial density within the CD19high and CD19low areas. *, **, ***, 250 **** correspond to p < 0.05, p < 0.01, p < 0.001 and p < 0.0001, respectively. 251 Box and whiskers plots show the median, the 5th to 95th percentile for the whiskers and the 25th 252 to 75th percentile for the box. 253 254

Rationale for CAR RNA FISH probe design 256
We designed the CAR RNA probes to be directed against the wPRE in the 3'UTR and the scFv of 257 the anti-CD19 CAR FMC63 and the anti-HIV env CAR VRC07-523 (Fig. 1A, Fig. 1A suppl.). 258 The scFv of the CAR is based on the variable region of the two mRNA species that translate to the 259 Ig heavy and light chain of an antibody. To rule out that the probes against the scFv cross-hybridize 260 with a large number of B cells in potential target tissues we stained human tonsil sections with a 261 pan-IGHG RNA FISH or a CAR FMC63 15ZZ probes. IGHG RNA FISH showed an ubiquitous 262 and bright staining throughout the tissue section ( Fig. 2A

CAR RNA FISH detects CAR T cells in in vitro cultures 272
For in vitro experiments, T cells were co-cultured with -irradiated CD19-expressing TM-LCL 273 lymphoblastoid cells for 12 days and purity was assessed by flow cytometry (~90%) (Fig. 3 suppl.). 274 To validate the CAR FMC63 26ZZ probes, anti-CD19 CD8 CAR T cells or Jurkat cells were 275 processed for cytospins and stained for CAR mRNA (Fig. 1B). The anti-CD19 CD8 CAR T cells 276 stained positive with the CAR FMC63 26ZZ probe set. By contrast, Jurkat cells stained with the 277 CAR FMC63 26ZZ probe set were negative for CAR RNA spots. RNA FISH against housekeeping 278 gene PPIB (Peptidyl-prolyl cis-trans isomerase B) and the bacterial gene DapB as positive and  279 negative controls, respectively, for RNA integrity and probe specificity, yielded characteristic 280 RNA FISH staining only in cells with relevant genes (Fig. 1B). 281 RNA FISH-based imaging techniques detect the target mRNA mostly in the cytosol that surrounds 282 the nucleus as demonstrated with the CAR mRNA (Fig. 4A suppl.). The density of tissue impedes 283 automated image segmentation to define the quantity of T cells expressing the majority of the 284 extranuclear CAR mRNA and cannot be defined based on the nuclear DAPI staining. Traditional 285 algorithms rely on nuclei detection and assume a slightly extended perimeter for the cytoplasm to 286 detect the RNA. To define cellular boundaries and to facilitate computational image segmentation, 287 we used fluorescently labeled wheat germ agglutinin as a second counterstain to label N-acetyl-D-288 glucosamine and sialic acids on cell surface glycoproteins (Fig.1C). This allowed to analyze the 289 images with a modified version of a random reaction seed algorithm recently published by our 290 group (18) (Fig. 1D). In brief, the acquired images are partitioned and, several 1000 seeds are 291 distributed on the WGA plasma membrane stain images ( Fig 1D). The algorithm searches for the 292 plasma membrane high fluorescence intensity signal, follows the signal using hidden Markov 293 model statistics and tries to close a loop, which would represent a cell and thus, creates a membrane 294 map(18). Once the cell circumference has been determined, the algorithm uses classical Laplacian 295 of Gaussian local maxima detection for nuclear and RNA staining within this perimeter (Fig. 1D). 296 The output data contains the location and RNA expression levels of each individual cell of the 297 input image (Fig. 1D). 298 To validate the CAR RNA FISH approach in combination with the dnnRRS algorithm, we stained 299 anti-CD19 CD8 CAR T and Jurkat cells mixed at different ratio (1, 0.75, 0.5, 0.25 and 0) with the 300 CAR FMC63 26zz probe set and acquired images. The number of cells stained positive by CAR 301 RNA FISH and the number of cells detected with dnnRRS increased with incrementing number 302 of input anti-CD19 CD8 CAR T cells (Fig. 1E). The input CAR T cells were 90% EGFRt+ by 303 flow cytometry (Fig 1A suppl.), hence, we expected to detect ~ 90%, 67.5%, 45%, 22.5% and 0% 304 CAR+ cells for the corresponding ratio of 1, 0.75, 0.5, 0.25 and 0. The dnnRRS algorithm detected 305 ~5 -60% CAR+ cells in the CD19 CAR T cell data set. There was a strong correlation between 306 detection of the algorithm and the expected values (R 2 =0.95) (Fig. 1F) and a significantly higher 307 number of CAR RNA dots/cells compared to Jurkat cells alone (Fig. 1G). We chose the same 308 approach to validate CAR RNA FISH for a VRC07-523 CAR in a VRC07-523 CAR Jurkat cell 309 line and confirmed that it specifically stained these VRC07-523 CAR Jurkat cells but not control 310 Jurkat cells (Fig. 5A suppl.). In addition, in serial dilution experiments an increasing number of 311 cells stained positive for the VRC07-523 CAR RNA and this was detectable by dnnRRS ( Fig. 5B  312 suppl.). The ratio of detected cells by dnnRRS (7-68%) correlated with the expected values by 313 flow cytometry (0-97%) (R 2 =0.98) (Fig. 5C suppl.) and we observed a significantly higher number 314 of CAR RNA dots/cell between the VRC07-523 CAR Jurkat cells at each ratio and non-transduced 315 Jurkat cells (Fig. 5D suppl.). Combined, these data indicate that our CAR RNA FISH probe design 316 and dnnRRS image analysis is highly specific and quantitative for detection of CAR T cells in cell 317 culture samples. 318

CAR RNA FISH detect CAR T cells in situ in mouse xenografts 319
To validate CAR RNA FISH for in situ applications in preclinical animal models, we used a BE2 320 neuroblastoma xenograft model of anti-CD19 CAR T cell-infused mice or control mice without 321 CAR T cell infusion(20). The BE2 cell line was transduced with multiple lentiviruses to express 322 CD19 and luciferase and may contain the common wPRE sequence. We performed digital droplet 323 PCR with primers directed against the signaling domain and the wPRE in genomic DNA. We 324 detected the wPRE portion in both CAR+ and CAR-BE2 tumor tissue whereas the CAR signaling 325 domain was only present in CAR+ BE2 tumor tissues (Fig. 6 suppl.). Therefore, to exclude false-326 positive cell detection due to cross-hybridization artifacts during the in situ setup of the CAR RNA 327 FISH on mouse xenograft tissues, we only used a reduced probe set of total 15 zz probes, coined 328 FMC63-15ZZ, and omitted the wPRE probes (Fig. 1A suppl.). 329 We observed a characteristic RNA FISH pattern with the CAR RNA FISH in CD19+ BE2 330 xenograft tissues from mice infused with anti-CD19 CAR T cells but not from control mice ( We analyzed image data sets of a CAR RNA FISH or PPIB RNA FISH combined with 365 immunohistochemistry for CD19 from CAR T cell-positive and negative BE2 xenograft tissues 366 with the dnnRRS algorithm and used FlowJo for data analysis. We detected 32.6% and 0.13% 367 CAR+ in the CAR+ (Fig. 2D) and CAR-BE2 ( Fig. 2E) tumor, respectively. CD19 expression was 368 below 1% and above 30% in CAR+ BE2 tumor and CAR-BE2 xenograft (Fig. 2D, E), 369 respectively. In the CAR-BE2 tumor, PPIB and CD19 expression showed homogenous 370 distribution (Fig. 2F). By contrast, the areas that stained positive for CD19 were almost devoid of 371 PPIB expression in the CAR+ BE2 tumor and showed signs of recent tissues disruption or lysis 372 ( Fig. 3G). 373 The distribution of CAR T cells in vivo differed spatially, and a substantial number of cells were 374 negative by CAR RNA FISH and CD19IHC in CAR+ BE2 tumor sections. To better characterize 375 this heterogeneity of the tumor composition, we performed additional combinational RNA FISH 376 for the CAR, CD19 and CD171, another highly expressed protein in neuroblastomas and in the 377 BE2 neuroblastoma cell line (20, 26). We found up to 38% CD171RNA+/CD19-and less than 378 0.01% CD171RNA+/CD19+ cells in the CAR+ BE2 tumor and in stark contrast ~75% 379 CD171RNA+/CD19+ in the CAR-BE2 tumor (data not shown). 380 These data indicated that the CAR RNA FISH detects specifically CAR T cells in tissues from 381 preclinical in vivo models and that the dnnRRS algorithm can be used to analyze CAR T cell 382 abundance, spatial distribution and efficacy. 383 CD4 and CD8 CAR T cells are widely distributed in the tumor and display differential 384 functional and spatial gene expression 385 To further characterize the functionality of the CAR T cells, we then set out to do combinations of 386 multiplex RNA FISH for the CAR, the T cell lineage marker CD4 and CD8α, as well as the effector 387 genes IFN and GZMB and analyzed the image data sets with the dnnRRS algorithm to phenotype 388 the tissue. 389 Independent CAR RNA FISH in combination with CD4 or CD8α showed that the CD4 and CD8 390 CAR T cells are evenly distributed within the CAR+ foci (Fig. 3A, B). We then subjected the 391 image data to computational analysis with the dnnRRS algorithm. For the CAR and CD8α RNA 392 FISH, 20.6% of the total cells were CAR+/CD8α+ and 19.9% CAR+/CD8-(presumably 393 CAR+/CD4+ T cells). For the CAR and CD4 RNA FISH, 17.2% of the total cells were 394 CAR+/CD4+ and 24.4% CAR+/CD4-(presumably CAR+/CD8+ T cells) (Fig 3C, D). Thus, in 395 this model the tissue-based distribution was relatively similar between the CD4+ and Cd8+ CAR 396 T cells. Around 2500 or 2800 CD19+ cells were detected in the GZMB and IFN datasets, respectively 410 ( Fig. 5A). We detected approximately >31000 CAR+ GZMB-and 700 CAR+ GZMB+ cells/tissue 411 section as well as >36000 CAR+IFN-and 40 CAR+ IFN+ cells/tissue section (Fig. 5B). 412 Therefore, the frequency and spatial density of CAR+ GZMB+ was approximately 2% or 37 413 cells/mm2 whereas for CAR+IFN+ cells it was below 0.2% or 1.8 cells/mm2 (Fig. 5C, D). The 414 frequency and spatial density of CAR+ cells that were negative for either GZMB or IFN was 415 above 98% or around 1500 CAR+/mm2 (Fig. 5C, D).
We divided the spatial datasets in a CD19high and CD19low area to further analyze the effect of 417 tumor antigen expression on the spatial expression of the CAR T cell effector genes GZMB and 418 IFN (Fig. 5E). The spatial density of CD19+ cells/mm2 in the CD19high and CD19low area in 419 the GZMB and IFN dataset were significantly different (Fig. 5F, G) indicating that CAR T cells 420 had a higher probability for antigen exposure in the CD19high area. For the GZMB dataset, the 421 number of total CAR+ (Fig. 5H), CAR+ GZMB- (Fig. 5I) and CAR+GZMB+ cells/mm2 (Fig. 5J) 422 was not significantly different between the CD19high and CD19low area, however, there was a 423 trend towards higher spatial density of total CAR+ and CAR+ GZMB-cells/mm2 in the CD19low 424 area and an opposite trend of higher spatial density for CAR+GZMB+ cells/mm2 in the CD19high 425 area. The ratio of CAR+ GZMB+ to CAR+ GZMB-cells/mm2 was significantly higher in the 426 CD19high area (Fig. 6K). For the IFN dataset, the number of total CAR+ (Fig. 6L), CAR+ IFN-427 ( Fig. 6M) and CAR+ IFN+ cells/mm2 (Fig. 6N) was significantly higher in the CD19low area. 428 The ratio of CAR+ IFN+ to CAR+ IFN-cells/mm2 was not significant in a comparison of the 429 CD19high and CD19low area (Fig. 6O). In summary, these data indicate that the dnnRRS CAR 430 RNA FISH approach in combination with RNA FISH for T cell lineage and effector genes can 431 provide important insights into the spatial distribution and functionality of CAR T cells on a single 432 cell level and indicate that there is a high heterogeneity and differential spatial effector gene 433 expression in CAR T cells with respect to the tumor target cells.                  Supplementary Figure 11