Local gene editing of fibroblasts in tumors reveals a new cancer-associated fibroblast state

Abstract

Fibroblasts play critical roles in regulating cellular relationships during tissue homeostasis, immunity, and tumor biology at multiple sites. However, tools to perturb fibroblasts at just one site in vivo are limited, restricting our understanding of how these cellular relationships act locally. We optimized local gene editing of fibroblasts in mouse tumor models to investigate how fibroblast perturbations affect the tumor microenvironment (TME). By knocking out receptors Osmr, Tgfbr2, or Il1r1 on cancer-associated fibroblasts (CAFs), we uncover that TGFBR2 signaling loss induces the emergence of a new Col18a1^hi CAF cell state that is associated with worse survival in pancreatic cancer patients. Combinatorial gene KOs in CAFs reveals a circuit where these Col18a1^hi CAFs reshape the TME by recruiting Siglec-F^hi neutrophils via Cxcl5 expression, and where this Col18a1^hi CAF cell state is dependent on TNFR1 and canonical Wnt signaling. Together, a fast, affordable, and modular engineering method is demonstrated, allowing discovery of modified fibroblast identities and local intercellular relationships in the TME.

Introduction

Fibroblasts play essential roles as both mediators of homeostasis and drivers of transitions in tissues, for example, during repair (Buechler et al., 2021b; Foster et al., 2021; Guerrero-Juarez et al., 2019; Hu et al., 2023; Mascharak et al., 2021; Plikus et al., 2021). However, fibroblast dysregulation has been associated with many chronic diseases, including fibrosis, autoimmunity, and cancer (Henderson et al., 2020; Kinchen et al., 2018; Korsunsky et al., 2022; Sahai et al., 2020; Tsukui et al., 2020). In cancer specifically, the term “cancer-associated fibroblast” (CAF) refers to a cell that is defined by the absence of epithelial, endothelial, and immune cell lineage markers and lack of cancer cell mutations (Sahai et al., 2020). Recent work has greatly expanded our understanding of the heterogenous CAF cell states in human cancer and mouse tumor models (Grout et al., 2022; Helms et al., 2022; Krishnamurty et al., 2022; McAndrews et al., 2022), where CAF subset identity is informed by their transcriptional state and ability to influence tumor, endothelial, and immune cell function ex vivo in experimental co-culture conditions (Biffi et al., 2019; Elyada et al., 2019; Öhlund et al., 2017).

However, studying the interaction of CAFs with other cells in biological settings has been challenging, as in vitro models do not fully model complex three-dimensional tumor microenvironments (TMEs), and testing gene functions in mice has suffered from a lack of methods to modify fibroblast gene expression locally at scale (Sahai et al., 2020). In some cases and to circumvent these limitations, tumor cells have been co-injected with defined CAF subsets, which revealed different CAF subset functions in these tumor models (Hutton et al., 2021). However, the injected CAFs are often outcompeted by host fibroblasts, leading to poor engraftment of the co-injected CAFs (Sahai et al., 2020). Mouse models using Cre-loxP provide a highly control-lable method to perturb individual genes in fibroblasts by using Cre-driver lines such as Pdgfra-Cre, Col1a2-Cre, or others that have been developed recently (Plikus et al., 2021). These models require mouse breeding of Cre-driver lines with floxed genes of interest, are time- and cost-intensive, and, when combined with inducible cell-depletion genes, such as diphtheria toxin receptor, typically deplete large populations of fibroblasts globally (Roberts et al., 2013), which limits the ability to draw conclusions about their specific roles in a defined environment. Thus, we set out to develop an in vivo fibroblast perturbation platform that is localized, fast, affordable, and modular.

Viral vectors have been extensively used to modify cells in vivo, with recombinant adeno-associated viruses (AAVs) being a popular choice due to their low immunogenicity and cytotoxicity (Wang et al., 2019). We hypothesized that local delivery of a guide RNA (gRNA) via AAV in combination with conditional expression of Cas9 in fibroblasts could generate local KOs of target genes in a defined area of the skin, allowing subsequent study of tumor growth within that tissue area of fibroblast perturbation. Similar strategies of local or organ-specific cell perturbations have been employed to target a variety of cells, such as brain cells (Challis et al., 2022), muscle cells (Tabebordbar et al., 2021), or hepatocytes (Wang et al., 2021). Additionally, screening efforts designed around varying the AAV capsid sequence have identified AAV serotypes with altered cell tropisms (Bryant et al., 2021; Maheshri et al., 2006; Nyberg et al., 2023). Among those, the evolved variants 7m8 and DJ (Westhaus et al., 2020), together with the natural variant AAV serotype 1 (AAV1) (Ellis et al., 2013), have emerged as candidates for efficient fibroblast targeting.

To test our approach, we targeted the following three receptors known to shape CAF biology: (1) transforming growth factor β type II receptor (Tgfbr2) encoding a surface receptor that upon binding to its ligand TGF-β1 together with TGFBR1 has been demonstrated to drive profound phenotypic changes in fibroblasts (Massagué and Sheppard, 2023). In tumors, TGF-β1–mediated activation of fibroblasts leads to a contractile phenotype that is associated with increased extracellular matrix (ECM) production, which leads to an overall remodeling of the ECM in the growing tumor (Batlle and Massagué, 2019; Chakravarthy et al., 2018; Laklai et al., 2016). As TGFBR2 is expressed on a multitude of epithelial, stromal, and immune cells, deciphering how TGFBR2 signaling affects tumor growth requires targeted deletion (Krishnamurty et al., 2022). (2) Oncostatin M (OSM) receptor (Osmr) encodes a surface receptor that has been implicated in stimulating a pro-inflammatory gene program in fibroblasts (Araujo et al., 2022; Hu et al., 2023; Lee et al., 2021). (3) Interleukin 1 receptor type 1 (Il1r1), which—in colorectal cancer lesions—stimulates CAFs to adopt an inflammatory cell state that is associated with increased tumor infiltration of immuno-suppressive macrophages (Koncina et al., 2023) and results in therapy-induced senescence of CAFs (Nicolas et al., 2022), both contributing to increased tumor growth.

Here, we sought to develop tools that are adaptable for knocking out genes, singly or in combination. Focusing on these three genes that have produced data indicating intratumoral roles for them in fibroblasts, we sought to develop both the method for localized fibroblast engineering and to answer questions about the localized role of these genes and pathways in intratumoral populations of fibroblasts.

Results

Targeted local KO of gene of interest in fibroblasts

We hypothesized that two elements were necessary for an approach to locally genetically engineer fibroblasts. First is a vector system that delivers a gRNA to the fibroblast. Together with preexisting fibroblast-specific alleles that regulate expression of Cas9, the choice of gRNA provides modular flexibility to target any gene of interest for gene editing. Second is identification of a highly efficient AAV serotype that delivers necessary elements to fibroblasts locally in vivo. Toward the first, we designed a self-complementary AAV (scAAV) (McCarty et al., 2001; McCarty et al., 2003; McCarty, 2008; Wang et al., 2003) cargo cassette comprised of the human U6 promoter (hU6) driving RNA polymerase III–mediated expression of a gRNA and a hybrid form of the CBA promoter (CBh) (Gray et al., 2011) driving mCherry expression as a transduction readout (Fig. 1 A). Cargo plasmids encoding gRNAs against Thy1 (gene encoding Thy1/CD90) or Pdpn (gene encoding podoplanin, PDPN) were packaged into scAAVs with the AAV1 capsid—from here on referred to as AAV-gRNA—and used to transduce the mouse fibroblast cell line NIH/3T3 expressing Cas9 (NIH/3T3.Cas9) (Fig. 1B). The scAAV was transduced at an efficiency of >75% and induced KO noticeable by loss of surface expression of Thy1 or PDPN (Fig. 1, C and D). A comparison of the AAV1 capsid with the recently developed fibroblast-tropic synthetic capsids 7m8 (Isgrig et al., 2019; Westhaus et al., 2020) and DJ (Grimm et al., 2008) showed similar transduction and KO efficiency in vitro (Fig. S1 A).

Figure 1. Local in vivo gene KO in fibroblasts.

(A) scAAV cassette for single-gene KO (sKO). ITR, inverted terminal repeat; Δtrs, no terminal resolution site; CBh, chicken β-actin hybrid promoter; pA, SV40 late polyadenylation signal. (B) Experimental scheme for testing transduction and KO efficiency of scAAV in vitro. NIH/3T3.Cas9 cells were transduced with AAV1 at an MOI of 1e5. 72 h later, cells are analyzed via flow cytometry. (C) Flow cytometry plots showing mCherry expression in NIH/3T3.Cas9 cells after transduction with AAV as outlined in B. One of two representative experiments is shown. (D) Flow cytometry overlay plots of (left) Thy1 or of (right) PDPN protein surface expression in NIH/3T3 cells expressing Cas9 (green) or not (black) after transduction with AAV as outlined in B. Only Cas9-expressing NIH/3T3 cells exposed to AAV-gRNA lose surface expression of target protein. One of two representative experiments is shown. (E) Experimental scheme for gene KO in tumor fibroblasts using scAAV in mice challenged with 5e5 YUMM5.2 melanoma cells. (F) Representative histograms showing Thy1 surface expression analyzed by flow cytometry of all EGFP⁺ tumor fibroblasts from mice treated as outlined in E and analyzed on day 7. One of two representative experiments is shown. (G) Quantification of KO efficiency in all EGFP⁺ tumor fibroblasts on day 7 or 14 after tumor challenge (n = 4–6 mice/group, pooled from two experiments). (H) Same as (F) in CD45⁺ CD3e⁺ T cells. (I) Quantification of H on day 7 (n = 3/group). Significance was tested using Welch’s t test. (J) Experimental scheme for testing local KO in tumor fibroblasts of mice challenged with 5e5 YUMM5.2 melanoma cells. (K) Left: Representative flow cytometry plots showing mCherry expression as a readout for AAV transduction. Cells gated on EGFP⁺ tumor fibroblasts. Right: Quantification of mCherry⁺ EGFP⁺ tumor fibroblasts in paired flank tumors from same mouse (n = 3 mice). (L) Quantification of Thy1 KO efficiency in all EGFP⁺ fibroblasts from paired flank tumors on same mouse (n = 3 mice). (M) Quantification of mCherry expression in fibroblasts of brachial and inguinal LNs pooled from left or right flanks (n = 3 mice). (N) Quantification of mCherry expression in splenic fibroblasts after flank AAV injection (n = 3 mice). (O) Overlay histograms of TGFBR2 surface expression by flow cytometry on all tumor fibroblasts from mice treated with 3e10 vg AAV-gTgfbr2 or AAV-gSCR (control) as outlined in E. Tumor fibroblasts of Col1a2^CreERT2/+;Tgfbr2^fl/fl mice treated with tamoxifen on five consecutive days prior to tumor engraftment (see Materials and methods for details on dose and application) were included as negative staining control. GeoMFI value listed for each sample. FMO, fluorescence minus one from Pdgfra-Cas9-EGFP mouse. One of two representative experiments is shown. (P) Quantification of TGFBR2 (GeoMFI) surface expression on all tumor fibroblasts from O relative to FMO (n = 2–4 mice/group). One of two representative experiments is shown. (Q) Experimental scheme for KO assessment in orthotopically transplanted pancreatic tumors. (R) Overlay histograms of TGFBR2 surface expression by flow cytometry on all tumor fibroblasts isolated from orthotopically grown HY19636 pancreatic tumors from Pdgfra-Cas9-EGFP mice after intrapancreatic co-injection of 3e10 vg AAV-gTgfbr2 and 2.5e5 HY19636 cells. GeoMFI values listed for each sample. One of two representative experiments is shown. All data are mean ± SD, and significance was tested using one-way ANOVA with Tukey’s multiple comparisons test, unless otherwise noted. *P < 0.05 and ****P < 0.0001. fl, floxed; GeoMFI, geometric mean fluorescence intensity; SCR, scramble; ns, nonsignificant. MOI, multiplicity of infection.

This prompted us to find a solution to the second requirement to identify a highly efficient AAV serotype for local delivery and execution of gene editing elements to fibroblasts in vivo. For this, Cas9 expression in host fibroblasts was achieved by combining a Pdgfra-driven Cre allele (either Pdgfra-Cre or Pdgfra-CreERT2 [Chung et al., 2018; Roesch et al., 2008]) with the Rosa26-LSL-Cas9-EGFP allele (Platt et al., 2014), hereafter referred to as Pdgfra-Cas9-EGFP mice (Materials and methods for details). In single-cell suspensions of tumors, fibroblasts were gated using the markers: DAPI⁻ CD31⁻ CD45⁻ E-Cadherin⁻ MCAM⁻ Thy1⁺ PDPN⁺ (Fig. S1 B). Of those, 97.3 ± 1.3% (mean ± SD) were GFP⁺ in Pdgfra-Cre;R26-LSL-Cas9-EGFP mice and 71.2 ± 11.7% in Pdgfra-CreERT2;R26-LSL-Cas9-EGFP mice (Fig. S1 C), indicating a high percentage of Cas9 expressing CAFs. Conversely, gating on EGFP⁺ cells revealed that 93–97% of all recombined cells express fibroblast markers Thy1⁺ and PDPN⁺ (Fig. S1 D). PDPN staining of EGFP⁺ cells in tumors was confirmed by microscopy (Fig. S1 E). Testing local delivery of AAV-gRNA was performed by injecting 1e10 viral genomes (vg) of AAV-gThy1-CBh-mCherry s.c. on the dorsal flank of mice and marking the injection site to allow subsequent injection of YUMM5.2 mouse melanoma tumor cells at the same site on the next day (Fig. S1 F). We excluded the 7m8 capsid from further in vivo experiments because s.c. injection of 7m8-serotyped scAAV transduced the panniculus carnosus (Fig. S1 G). When comparing AAV1 versus DJ capsids, measuring mCherry expression in the CAFs that populate the growing tumor mass was used as a readout for transduction efficiency, and surface expression of Thy1 was used as a readout for KO efficiency of the AAV. AAV1- and DJ-serotyped scAAVs transduced CAFs at different efficiencies (day 7 time point: 78.5 ± 8.5% [mean ± SD] and 53.6 ± 11.9% [mean ± SD] mCherry⁺, respectively, Fig. S1 H), leading to higher Thy1-negative levels by scAAV1 versus scDJ (90.9 ± 1.8% [mean ± SD] and 78 ± 4.9% [mean ± SD] Thy1-negative, respectively, Fig. S1 H). AAV1 also predominantly infected CAFs in the Pdgfra-Cas9-EGFP mice (93.3 ± 1.7% [mean ± SD]) (Fig. S1, I and J). On this basis, our future experiments used scAAV1.

Next, we tested efficiency of local gene KO in these CAFs by targeting Thy1 for deletion (Fig. 1 E). More than 85% of all CAFs were Thy1-negative 7 days after tumor challenge, compared with ∼10% without AAV (Fig. 1, F and G). Targeted Thy1 KO was stable in CAFs, as the same percentage of about 85% of all CAFs were still Thy1-negative when sampled 14 days after tumor challenge (Fig. 1 G), while the percentage of mCherry⁺ CAFs expectedly decreased in the same timeframe (Fig. S1 K), due to the non-integrative, episomal AAV genome being lost over time in mitotic cells. When extending this analysis to T cells, another cell type in the TME besides CAFs that express Thy1, AAV-gThy1 injection in Pdgfra-Cas9-EGFP mice did not induce loss of surface Thy1 levels (Fig. 1, H and I), indicating targeted gene KO predominantly in CAFs. This efficient and persistent KO of a target gene in CAFs was consistent at higher doses of AAV (Fig. S1 L). To test if the KO was truly local, the same Pdgfra-Cas9-EGFP mouse was injected on only one flank s.c. with AAV-gThy1 but challenged with YUMM5.2 tumor on both flanks on the following day (Fig. 1 J). CAF transduction and Thy1 KO were only detected in the tumor that was grown at the site of prior AAV-gThy1 injection (Fig. 1, K and L), not in tumor-draining LNs (Fig. 1 M) nor in the spleen (Fig. 1 N). This suggests that s.c. injection of AAV-gRNA allows local KO and limits systemic spread of AAV, minimizing KO at unintended locations.

As part of a goal to understand localized roles for specific genes in fibroblasts, we next generated AAV-gTgfbr2 to target the Tgfbr2 gene in fibroblasts for KO (Fig. S1 M). AAV-gTgfbr2 decreased TGFBR2 levels on all CAFs in Pdgfra-Cas9-EGFP mice to similar levels as conventional Col1a2-CreERT2–mediated KO of the floxed Tgfbr2 gene in Col1a2^CreERT2/wt;Tgfbr2^fl/fl mice (Fig. 1, O and P), demonstrating that AAV-delivered gRNAs have an in vivo KO efficiency almost resembling conventional Cre/loxP-based approaches, without having to rely on lengthy mouse breeding to get to the desired genotype for experimental analysis. Instead, AAV-delivered gRNAs required only 2 wk of cloning and AAV generation until injection of AAV-gRNA for targeted, local KO of any gene of interest. This allows subsequent functional studies of any gene of interest expressed by CAFs via local gene KO.

To assess the KO efficiency of our AAV-based system at (A) a different site and (B) within a different setting, we (A) co-injected the pancreatic cancer cell line HY19636, which is derived from a mouse pancreatic ductal adenocarcinoma (PDAC) KPC mouse (Kras^LSL–G12D/+; Trp53^lox/+; p48/Ptf1a-Cre) (Yamamoto et al., 2020), together with AAV-gTgfbr2 orthotopically in the pancreas of Pdgfra-Cas9-EGFP mice (Fig. 1 Q); or (B) tested it in a skin fibrosis (Błyszczuk et al., 2019) model (Fig. S1 N). Fibroblasts isolated from orthotopically grown pancreatic tumors or in skin fibrosis also lost TGFBR2 surface expression when AAV-gTgfbr2 was co-injected (Fig. 1 R and Fig. S1 O), demonstrating the adaptability of this approach to different tissue sites and disease models.

Local Osmr or Tgfbr2 KO in CAFs repolarizes them toward homeostatic state but differentially affects TREM2⁺ tumor-associated macrophage phenotype

As a proof of concept and test of the specificity achievable with selective gene KOs, we set out to compare how blocking of a putatively pro-inflammatory gene program in CAFs via Osmr KO (Fig. S2, A-C) or blocking of a putatively pro-fibrotic gene program via Tgfbr2 KO (Fig. 1 P) can affect the fibroblast cell states in growing YUMM5.2 melanoma or HY19636 pancreatic cancer tumors (Fig. 2 A). The HY19636 PDAC model represents a more desmoplastic tumor with high CAF involvement (Fig. S2 D). To assay changes in fibroblast identity, we initially measured surface markers that are typically indicative of different activity states: expression of SCA-I⁺ Ly6C⁺ are associated with homeostatic fibroblasts in skin (Buechler et al., 2021b), skin repair (Hu et al., 2023) (Fig. S2, E and F), and tumor growth (Krishnamurty et al., 2022) (Fig. S2, H and 2 I), whereas loss of SCA-I and Ly6C with concomitant upregulation of Ncam1 marks activated fibroblasts in these settings (Fig. S2, G and J). We thus distinguished double-positive (DP) SCA-I⁺ Ly6C⁺, intermediate (Int), double-negative (DN) SCA-I⁻ Ly6C⁻, and DN NCAM1^± CAF subsets (Fig. S2, K and L). We corroborated these cell states as defined by the aforementioned markers by measuring increased expression of homeostatic marker Dpt in the DP CAFs and, inversely, increased expression of Acta2 and Postn in DN NCAM1⁺ CAFs (Fig. S2 M). We also noted that DN CAFs had the highest expression of Cthrc1, a recently described marker of a pro-fibrotic fibroblast subset (Fig. S2, N and O) (Tsukui et al., 2020; Tsukui et al., 2024).

Figure 2. Local Osmr or Tgfbr2 KO in CAFs repolarizes them toward homeostatic state and subtly affects TREM2⁺ TAM MHC-II expression.

(A) Different AAV-gRNA cassettes used in outlined experiment for local KO in tumor fibroblasts. Each AAV-gRNA was injected s.c. in HY19636 pancreatic cancer model at a dose of 1–3e10 vg AAV in Pdgfra-Cre;R26-Cas9-EGFP^+/+ or Pdgfra-CreERT2^+/wt;R26-Cas9-EGFP^+/+ mice. (B) Flow cytometry plots of (left) EGFP⁺ skin fibroblasts from Pdgfra-Cas9-EGFP mice and (right) the DN SCA-I⁻ Ly6C⁻ subset. (C) Flow cytometry plots of (left) EGFP⁺ tumor fibroblasts and (right) DN tumor fibroblasts from control-, AAV-gOsmr, or AAV-gTgfbr2–injected Pdgfra-Cas9-EGFP mice on day 14 after HY19636 tumor challenge. (D) Quantification of C. (E) Stacked bar chart of main cell populations identified in tumors from C via flow cytometry plotted as proportion of all live cells. Mean values of n = 6–13 tumors/group, pooled from four independent experiments, are plotted. (F) Experimental scheme for ex vivo stimulation and analysis of CAFs isolated from Pdgfra-Cas9-EGFP mice challenged with HY19636 tumors and prior injection of AAV-KIKO-Thy1 (control), AAV-KIKO-Osmr, or AAV-KIKO-Tgfbr2. (G) Left: Representative immunofluorescence images of sorted EGFP⁺ CAFs from F stimulated with OSM. Cells were fixed and counterstained with DAPI prior to staining for pSTAT3. Filled arrowheads mark nuclear location of pSTAT3; open arrowheads mark absence of nuclear pSTAT3. Scale bar, 50 μm. Right: Quantification of EGFP⁺ CAFs with nuclear pSTAT3. Each dot represents CAFs sorted from one tumor (n = 3 tumors/group). Data are mean ± SD, and significance was tested using Welch’s t test. (H) Left: Representative immunofluorescence images of sorted EGFP⁺ CAFs from F stimulated with TGF-β1. Cells were fixed and counterstained with DAPI prior to staining for pSMAD2/3. Filled arrowheads mark nuclear location of pSMAD2/3; open arrowheads mark absence of nuclear pSMAD2/3. Scale bar, 50 μm. Right: Quantification of EGFP⁺ CAFs with nuclear pSMAD2/3. Each dot represents CAFs sorted from one tumor (n = 2–3 tumors/group). (I) Representative flow cytometry plots of (left) CD11b⁺ MonoMacs and (right) F4/80⁺ TREM2⁺ TAMs from HY19636 tumors with Osmr or Tgfbr2 KO in fibroblasts as outlined in A. (J) Quantification of F. Data represent three (B) or four (C and I) independent experiments. Data in D and J are mean ± SD, and significance was tested using one-way ANOVA with Tukey’s multiple comparison test (n = 6–13 tumors/group, pooled from four independent experiments). CBh, chicken β-actin hybrid promoter; pA, polyadenylation signal; SCR, scramble; ns, nonsignificant. *P < 0.05, **P < 0.01, and ***P < 0.001.

While homeostatic skin fibroblasts expectedly had mostly DP fibroblast and low NCAM1 activation-marker expression (Fig. 2 B), CAFs exhibited a more DN phenotype (Fig. 2 C), which was decreased in both AAV-gOsmr and AAV-Tgfbr2 groups when compared with controls (Fig. 2 D), suggesting that Osmr or Tgfbr2 KO in CAFs is sufficient to shift the subset balance away from the activated DN CAF population. Tgfbr2 KO also decreased the frequency of the NCAM1⁺ DN population (Fig. 2 D). We did not observe a significant effect on tumor weight or fibroblast, endothelial, and immune cell numbers (Fig. 2 E and Fig. S3 A). These findings were replicated in the YUMM5.2 melanoma model (Fig. S3, B-D), demonstrating the robust adaptability of this local fibroblast gene editing tool to other settings.

For the control group, we decided to pool tumor samples that had received AAV with a scrambled gRNA sequence (gSCR) that does not recognize any sequence in the mouse (Suzuki et al., 2016); AAV with a gRNA targeting the T cell receptor α constant (Trac) locus—which is not expressed in fibroblasts and, thus, serves as a genome cutting control (Fig. S3 E); and samples that had received no AAV. These three control groups all showed no significant differences on measured fibroblast phenotypes when compared with each other (Fig. S3 F) and so are presented as one “control” group for the rest of the study.

To confirm gene inactivation of Osmr and Tgfbr2 in the targeted CAFs, we initially attempted to stably knockin (KI) mCherry in the Osmr or Tgfbr2 locus, respectively, which would allow enrichment of knocked-out CAFs by sorting mCherry⁺ cells and then use them for further studies to assess gene inactivation. We designed KI-KO (KIKO) AAV constructs that combine a U6-gRNA cassette to induce DNA double-strand breaks with a homology-directed repair template (HDRT) to KI genetic payloads into regions of interest in target cells (Fig. S3 G). The homology arms of the HDRT and the payload are designed to generate an in-frame KI of mCherry at the gOsmr or gTgfbr2 target sites, thereby disrupting endogenous Osmr or Tgfbr2 expression (Fig. S3 H). In vivo KI efficiencies in CAFs of 3–12% and 18–45% were detected for targeting Osmr and Tgfbr2, respectively (Fig. S3 I). These differences in mCherry reporter expression between different gene targets potentially reflect differences in gene level expression after KIKO and/or differences in KI efficiencies of the HDRTs at their unique genomic loci.

To ascertain Osmr or Tgfbr2 gene inactivation after in vivo gene editing, CAFs were sorted from HY19636 challenged Pdgfra-Cas9-EGFP mice that received AAV-KIKO-Osmr or AAV-KIKO-Tgfbr2 and were then exposed ex vivo to OSM (the primary ligand for OSMR in mice) or TGF-β1 (ligand for TGFBR2) to measure pathway activity in KIKO CAFs (Fig. 2F). Here, all EGFP⁺ CAFs were sorted, instead of just the mCherry⁺ CAFs, to maximize cell yield. CAFs from AAV-KIKO-Thy1–treated mice served as controls. Decrease in nuclear location of pSTAT3 or pSMAD2/3 in KIKO-Osmr or KIKO-Tgfbr2 CAFs, respectively, upon stimulation demonstrated stable functional in vivo KO of Osmr or Tgfbr2 (Fig. 2, G and H). Thus, AAV-based in vivo KO of target genes allows functional gene inactivation.

A benefit of engineering fibroblasts specifically in the TME is the ability to study the effect of their modification upon other local cells. Given the relationship between macrophages and fibroblasts in a multitude of tissue functions and pathology (Buechler et al., 2021a; Franklin, 2021), we sought to apply our method to understand modulation in tumor-associated macrophages (TAMs, gated as DAPI⁻ CD45⁺ Ly6G⁻ CD90.2⁻ B220⁻ NK1.1⁻ SiglecF⁻ non-[CD64⁻ CD24⁺] CD11b⁺ F4/80⁺ TREM2⁺) (Fig. S3 J). While TAMs did not change in their frequency within all immune cells, Tgfbr2 KO in CAFs led to a slight but statistically significant increase in the frequency of those that expressed MHCII⁺ (Fig. 2, I and J; and S3 K)—a phenotype associated with immunostimulatory myeloid cells (Kilian et al., 2023; Wang et al., 2011). This increase in MHCII⁺ TAMs was reproduced in the YUMM5.2 model again without an increase in total intratumoral CD90⁺ lymphocytes or monocyte/macrophages (Macs) (Fig. S3, L-N). These findings highlight the utility of modulating genes of interest in CAFs to induce changes in their phenotype, with secondary effects on other cells of the TME, such as TAMs.

Extending the analysis of CAF pathway perturbations beyond OSM and TGFβ signaling, we tested how Il1r1 KO in CAFs (Fig. S3, O and P) in the PDAC model affects their subset polarization because IL-1 pathway activity in CAFs has been shown to promote disease progression (Koncina et al., 2023; Nicolas et al., 2022). Here, we found no changes compared with control (Fig. S3 Q), as well as no changes when measuring TAM frequencies in the tumor, although MHCII⁺ TAMs were again more abundant in CAF Il1r1 KO tumors (Fig. S3 R). Thus, Osmr, Tgfbr2, or Il1r1 KO in CAFs repolarizes CAFs differently, prompting us to investigate those differences and their effect on the networks of gene expression in the TME in more detail.

Local Tgfbr2 KO in CAFs reveals emergence of unique Col18a1^hi CAF cell state

To study the effect of Osmr, Tgfbr2, or Il1r1 KO in CAFs on a transcriptional level and to capture its effect on other cells of the TME, immune, stromal, and tumor cells of s.c. grown KPC tumors were collected for single-cell RNA-sequencing (scRNA-seq) analysis after targeted, local KO by AAV in Pdgfra-Cas9-EGFP mice (Fig. 3A and Fig. S4 A). A total of 115,637 high-quality cells were captured from a total of 16 tumors (n = 4 per AAV-gRNA group), covering all major cell populations in the TME: neutrophils, MonoMacs, dendritic cells (DCs), mast cells, T cells, natural killer (NK) cells, endothelial cells, pericytes, and fibroblasts (Fig. S4, B-D).

Figure 3. Local Tgfbr2 KO in CAFs reveals emergence of unique Col18a1^hi CAF cell state.

(A) Experimental layout. (B) Uniform manifold approximation and projection (UMAP) of CAF subsets. (C) Marker gene expression of identified CAF subsets. (D) UMAP of CAFs highlighted by AAV-gRNA condition. (E) Bar charts of frequency of CAF subsets (scRNA-seq) as a fraction of all fibroblasts by AAV-gRNA condition plotted as mean ± SEM (n = 4 per group). *P < 0.05 one-way ANOVA with Dunnett’s multiple comparison test. (F and G) Similarity of DEGs between CAF subsets (columns) and (F) mouse fibroblasts in perturbed states by Buechler et al. (2021b) (rows) or (G) mouse PDAC CAF subsets by Elyada et al. (2019). Statistical significance of DEG overlap (−log10 [P value], hypergeometric test, P value adjusted using Bonferroni correction) for each pairwise comparison is plotted as heatmap. Yellow indicates high statistically significant overlap of DEG lists. (H) UMAP visualization of select genes (homeostatic Ly6c1 and Ly6a; myofibroblast-associated Acta2 and Ncam1; AAV-gTgfbr2–associated Col18a1 and Itga1) in CAFs. (I) Flow cytometry contour plots of tumor fibroblasts on day 14 after 2.5e5 HY19636 s.c. cell injection in Pdgfra-Cas9-EGFP mice with or without 1e10 vg AAV-gTgfbr2 injection on day −1. ITGA1⁺ NCAM1⁻ gate (top left) highlights AAV-gTgfbr2–specific CAF population. One of three representative experiments is shown. (J) Quantification of I. Data are mean ± SD, and significance was tested using Welch’s t test (n = 4/group). One of two representative experiments is shown. *P < 0.05, **P < 0.01, and ns, nonsignificant. (K) Gene expression of CAFs sorted as in I analyzed by quantitative real-time PCR. Data are mean ± SEM (n = 4 tumors/group), and significance was tested using one-way ANOVA with Tukey’s multiple comparison test.

Sub-clustering of the CAF compartment revealed eight different subsets (Fig. 3, B and C; and Table S1). Homeostatic fibroblasts expressed Pi16, Ly6c1, and Ly6a, genes associated with “universal” fibroblasts (Fig. S2 K); inflammatory CAFs expressed Il6, Lif, and Saa3, genes previously described in pro-inflammatory settings (Öhlund et al., 2017; Tsukui et al., 2024); myofibroblastic CAFs (myCAFs) were characterized by myofibroblast-associated genes, such as Acta2, Lrrc15, and Ncam1 (Dominguez et al., 2020; Hinz et al., 2001; Ye et al., 2024) and could be further subdivided into four groups based on proliferative markers (Mki67; myCAF_prolif), Slit1 and Fabp5 (myCAF_2), or Wif1 and Lamc3 (myCAF_3). The Int cluster harbored genes of both homeostatic and inflammatory fibroblasts, whereas the Col18a1 cluster was named after one of its most differentially expressed gene (DEG), Col18a1, and noticeably has low expression of the other homeostatic, inflammatory, or myofibroblast-defining genes (Fig. 3 C).

The Col18a1 cluster was only present in the AAV-gTgfbr2 group (Fig. 3 D) and represented the most abundant CAF subset within those samples at about one half of all AAV-gTgfbr2 CAFs (Fig. 3 E). Conversely, Tgfbr2 KO in CAFs leads to a decrease in all myCAF subsets (Fig. 3 E). This could be explained by the lack of a recently published fibroblast TGF-β1 response signature specifically in AAV-gTgfbr2 CAFs (Fig. S4 E), a prominent driver of myofibroblast differentiation (Mariathasan et al., 2018). To assess if the emergence of the Col18a1-expressing CAFs in the mouse PDAC model was unique to our AAV-based Tgfbr2 KO, we re-analyzed a published mouse PDAC scRNA-seq data set using conditional Tgfbr2 KO in universal fibroblasts (Krishnamurty et al., 2022) and found that a population with the Col18a1 CAF cluster signature can also be identified specifically when Tgfbr2 was knocked out using Cre/LoxP-based mouse genetics (Fig. S4 F), demonstrating the methodological robustness of using AAV-gRNA or Cre/LoxP-based gene KOs.

To gain a better understanding of how Osmr, Tgfbr2, or Il1r1 KO affects CAF transcriptional polarization, we compared the overlap of DEG lists from each KO group and noticed that the majority of DEGs were unique to their respective KO group (Fig. S4, G and H; and Table S2). For example, AAV-gOsmr and AAV- gTgfbr2 only shared 20 and 24 genes that were up or down-regulated, respectively, when compared with the unperturbed control CAFs (Fig. S4 H). Gene set enrichment analysis (GSEA) revealed that AAV-gOsmr CAFs were enriched for inflammatory response genes (Ifitm1, Il1a, Irf1, Lyn, and P2rx4), but depleted of hypoxia-related genes that are involved in carbohydrate metabolic processes (Slc2a1, Pfkl, Pdk1, and Gys1), whereas AAV-gTgfbr2 CAFs were enriched for hypoxia-related (Ndrg1, S100a3) and glycolytic genes (Pgk1, Ldha, and Ugp2) but depleted of epithelial-to-mesenchymal transition genes (Acta2, Tagln, Lrrc15, and Ccn1) (Fig. S4, I and J), a known consequence of absent TGFBR2 signaling in fibroblasts (Buechler et al., 2021b; Elyada et al., 2019; Hinz et al., 2001). Together, these results demonstrate that the AAV-mediated KO of individual receptors can induce distinct transcriptional changes in CAFs.

With the Col18a1 cluster uniquely emerging in the AAV-gTgfbr2 condition and it not being characterized by homeostatic, inflammatory, or myofibroblastic gene signatures, we also compared its transcriptional profile with published mouse scRNA-seq fibroblast data sets (Buechler et al., 2021b; Elyada et al., 2019). As expected, the DEGs of the homeostatic and Int clusters had the most overlap with the Pi16⁺ and Col15a1⁺ homeostatic/universal state; the inflammatory CAF cluster had the most overlap with the colitis-associated Adamdec1⁺ state, and the myCAF_1 and myCAF_2 clusters were most similar to Lrrc15⁺ myofibroblasts (Fig. 3 F) and PDAC myCAFs (Fig. 3 G). However, the Col18a1 cluster did not uniquely map to any previously identified mouse fibroblast states (Fig. 3, F and G). Even when comparing the Col18a1 gene signature with published human fibroblast profiles (Gao et al., 2024; Hwang et al., 2022), the Col18a1 cluster did not have the highest overlap with any identified human fibroblast subset (Fig. S4, K-M), further suggesting that Tgfbr2 KO CAFs adopt a unique transcriptional cell state.

The myCAF_1 cluster exhibited the highest similarity with human PDAC myofibroblasts and non-small cell lung cancer FAP⁺ CAFs (Fig. S4, K and L), suggesting it is correct to refer to these as myofibroblasts. While the myCAF_2 cluster had some similarity with human PDAC myofibroblasts, they also shared transcriptional similarity with previously described inflammatory, “stress-associated” HSPA6 fibroblasts (Fig. S4 M). The myCAF_3 cluster, which emerged in the AAV-gOsmr condition (Fig. 3 E), was also most similar to stress-associated HSPA6 fibroblasts (Fig. S4 M). In summary, the myCAF_1 cluster transcriptionally resembles myCAFs, whereas the myCAF_2 and myCAF_3 clusters exhibit more mixed transcriptional phenotypes of inflammatory and myofibroblastic states.

To identify the Col18a1 CAFs orthogonally via flow cytometry, we identified surface expressed proteins that were transcriptionally expressed at high levels within the Col18a1 cluster. While Itga1 (encoding integrin α 1/ITGA1) was not exclusively expressed in the Col18a1 cluster (Fig. 3 H), when combining ITGA1 staining with SCA-I, Ly6C, and NCAM1, we identified a ITGA1⁺ NCAM1⁻ CAF population that specifically emerged in AAV-gTgfbr2 tumors (Fig. 3, I and J) and was low for the homeostatic markers SCA-I and Ly6C (Fig. S4 N). qPCR analysis of sorted CAF populations based on NCAM1 and ITGA1 markers further validated their usefulness in identifying CAFs expressing homeostasis-associated gene Dpt (ITGA1⁻ NCAM1⁻ DN), myofibroblast-associated Acta2 (ITGA1⁻ NCAM1⁺), and Col18a1 (ITGA1⁺ NCAM1⁻) (Fig. 3 K).

Tgfbr2 KO-induced Col18a1^hi CAFs are distinct from iCAFs and myCAFs and correlate with worse survival in human pancreatic cancer patients

The emergence of the Col18a1^hi CAF state prompted us to investigate the possibility that local Tgfbr2 KO in fibroblasts during tumor growth led to the polarization of a distinct CAF cell state. In silico pseudotime-based trajectory analysis using Monocle 3 (Cao et al., 2019) of all captured CAFs, sans proliferating my-CAFs, placed Col18a1 CAFs at the end of the trajectory (Fig. 4 A). The analysis suggested that CAFs, in the setting of Tgfbr2 KO, progressed from homeostatic to inflammatory through the my-CAF cell state until they eventually arrived at the Col18a1 cell state (Fig. 4 B). This progression along the pseudotime axis was characterized by a loss of the universal marker Pi16 and sequential emergence of inflammatory markers Saa3 and Cthrc1, prior to upregulation of myofibroblast markers Lrrc15, Acta2, and Ncam1, which then were lost when the Col18a1 state was reached (Fig. S5 A).

Figure 4. Tgfbr2 KO-induced Col18a1^hi CAFs are distinct from iCAFs and myCAFs and correlate with worse survival in human pancreatic cancer patients.

(A) UMAP of (top) CAF subsets without myCAF_prolif subset and (bottom) pseudotime trajectory starting at the homeostatic fibroblast subset and ending at cluster myCAF_3 or Col18a1. (B) Violin plot of CAF subsets plotted according to their distribution in pseudotime. (C) Pairwise Pearson correlation coefficient of CAF NMF factor expression across all CAFs. Rows and columns are ordered by hierarchical clustering. (D) UMAP visualization of select NMF factor expression in CAFs. (E) Violin plots of (top) NMF factor_6 and (bottom) NMF factor_9 expression in the CAF subsets. (F) Enriched gene sets (Gene Ontology: biological process [BP]) based on top contributing genes in (top) NMF factor_6 and (bottom) NMF factor_9. (G) Kaplan–Meier survival plot of PAAD patients from TCGA excluding PNET (pancreatic neuroendocrine tumor) diagnosed patients categorized into upper and lower quartiles by the ratio of “AAV-gTgfbr2 up” DEG signature to the fibroblast progenitor signature (see Materials and methods for details). The mouse-derived AAV-gTgfbr2 up DEG signature was generated using human orthologs. Hazard ratio (HR) and P value of Cox regression fit are shown. UMAP, uniform manifold approximation and projection.

By focusing on the AAV-gTgfbr2 fibroblast object only, which predominantly contains the Col18a1^hi subset but also homeostatic, inflammatory CAFs, myCAF_1, and myCAF_2 subsets (Fig. 3 E), we noticed a branchpoint in cell trajectory (Fig. S5 B) that separated the homeostatic and inflammatory CAFs (prebranchpoint) from the myCAF_1 subset (branch 1) and Col18a1^hi subset (branch 2) (Fig. S5, C and D). Differential gene expression comparing the pre-branchpoint versus branch 1 or branch 2 (Fig. S5, E and F) revealed increased expression of AP-1 transcription factors (Fos, Fosb, Jun, and Junb) in cells along branch 2 (Col18a1^hi subset), which inversely decreased in expression along branch 1 (myCAF_1 subset) (Fig. S5 G). AP-1 transcription factors have been shown to interact with and cross-regulate Smad3/Smad4 proteins from the TGF-β signaling pathway (Massagué and Chen, 2000), suggesting that in this context of Tgfbr2 KO in CAFs, the Col18a1^hi CAF subset emerges because of aberrant transcriptional regulation.

With the Col18a1 cluster not previously classified, we therefore identified de novo signatures across all identified CAF states using nonnegative matrix factorization (NMF) (Stein-O’Brien et al., 2018). This approach allows non-cluster–based analysis of transcriptional cell states by decomposing the cell-by-gene scRNA-seq count matrix into two smaller matrices, where one matrix represents the factors, i.e., groupings of co-expressed genes, and the other matrix represents the expression of those individual factors across each cell. Especially for cell populations with mixed cell states or without clearly defined subsets, such as fibroblasts, this approach can uncover heterogeneous gene expression patterns that are masked by clustering analysis.

We identified a total of 18 different factors within the CAF population (Fig. S5 H and Table S3). While most averaged factor weights did not differ between the AAV-gRNA groups, factor 6 and factor 9 were elevated in AAV-gTgfbr2 CAFs, and factor 14 was elevated in the AAV-gOsmr CAFs (Fig. S5 I). Correlation of factor weights across all CAFs further revealed that there were three major groupings of gene factors: (1) factors 7 and 13 that were enriched in inflammatory CAFs; (2) the factors 5, 11, and 14 that were enriched in myCAFs; (3) and the AAV-gTgfbr2–enriched factors 6 and 9 (Fig. 4, C and D). Noticeably, the inflammatory factors 7 and 13 were anticorrelated with both other groupings, suggesting that the inflammatory cell state is mutually exclusive with other CAF cell states detected here. CAFs within the Col18a1 cluster were enriched for factors 6 and 9 (Fig. 4 E), which were identified as gene signatures of various developmental states (factor_6) and ossification/osteoblast differentiation (Fzd1, Pth1r, and Spp1) (factor_9) (Fig. 4 F and Table S3). With osteoblasts representing a unique mesenchymal lineage within the bone tissue (Zhu et al., 2024), we speculated whether a Tgfbr2 KO in CAFs reprograms them to an osteoblast-like cell. However, comparison of Col18a1 DEGs with a recently published mouse bone marrow stromal cell atlas (Baryawno et al., 2019) did not suggest that the Col18a1 cluster adopted a specific bone marrow stroma program (Fig. S5 J). This suggested that the AAV-gTgfbr2–induced CAFs harbored a cell state that is distinct of myofibroblasts, homeostatic, and inflammatory CAFs, instead presenting as a mix of ECM-producing CAFs with an osteoblast-like gene signature.

While both a high stromal gene signature (Combes et al., 2022) and a high fibroblast TGF-β1 response signature (Mariathasan et al., 2018) in human cancers are associated with worse survival (Fig. S5, K and L), we were curious how signatures derived from the DEGs from AAV-gRNA CAFs would associate with survival in pancreatic adenocarcinoma (PAAD) patients in the TCGA cohort (excluding pancreatic neuroendocrine tumor patients). To identify genes for inclusion in DEG-based signatures, we applied the following four steps (more details in Materials and methods): (1) the upregulated DEGs from the pseudobulk analysis of CAFs were selected for each AAV-gRNA group (Table S2); (2) the mouse genes were transferred to their human orthologs; (3) genes were kept if they were expressed most highly in fibroblasts or stellate cells (resident fibroblasts of the pancreas) in human pancreatic cancer scRNA-seq data sets (Tyler et al., 2025); (4) genes were excluded if they were shared between >1 AAV-gRNA group (Table S4). By categorizing PAAD patients based on the ratio of the AAV-gRNA DEG signature to a progenitor signature, we found that a high “AAV-gTgfbr2 up DEG signature/progenitor signature” ratio was associated with worse survival in PAAD, reaching a significant hazard ratio of 1.98 (P = 0.0175) (Fig. 4 G). Notably, survival analysis using the AAV-gOsmr, AAV-gIl1r1, or the TGF-β response signature did not yield significant differences in survival of PAAD patients (Fig. S5 M). Together, this suggests that a relative increase of the AAV-gTgfbr2–induced CAF cell state compared with progenitor fibroblasts, characterized by an osteoblast-like gene signature, is pro-tumorigenic in PAAD.

Tgfbr2 KO-induced Col18a1^hi CAFs recruit Siglec-F^hi neutrophils via high Cxcl5 expression

To gain a better understanding of how the AAV-gTgfbr2–induced Col18a1^hi CAFs interact with their TME, the percentage of all identified major immune cell subsets (Fig. S5, N-Q) were correlated across all samples (Fig. 5 A). This revealed several cell networks, with Col18a1 CAFs positively correlating with two populations: the TAM.Ndrg1 subset and the neutrophil.4 subset. The neutrophil.4 subset was most abundant in AAV-gTgfbr2 tumors (Fig. 5B) and characterized by high Siglecf expression (Fig. 5C). The TAM.Ndrg1 subset had elevated MHC-II gene (H2-Ab1, H2-Aa, and H2-Eb1) expression in AAV-gTgfbr2 tumors (Fig. S5 R), mirroring our previous findings of increased MHCII⁺ TAMs, as analyzed by flow cytometry (Fig. 2 J).

Figure 5. Tgfbr2 KO-induced Col18a1^hi CAFs recruit Siglec-F^hi neutrophils via high Cxcl5 expression.

(A) Heatmap showing the pairwise Pearson correlation of cell type frequencies among AAV-gRNA samples. Box highlights grouping of Col18a1 CAF-associated cell types. *P < 0.05. (B) Bar charts of frequency of neutrophil subsets (scRNA-seq) as a fraction of all neutrophils by AAV-gRNA condition plotted as mean ± SEM (n = 4 per group). *P < 0.05 one-way ANOVA with Dunnett’s multiple comparison test. (C) Violin plot of Siglecf expression among all neutrophil subsets. (D) Bubble plot of scaled expression of all detected Ccl and Cxcl genes in all CAF subsets. Cxcl5 highlighted in red as most highly expressed in Col18a1 cluster. (E) Flow cytometry contour plots of DAPI⁻ CD45⁺ CD11b⁺ Ly6G⁺ neutrophils on day 14 after 2.5e5 HY19636 s.c. cell injection in Pdgfra-Cas9-EGFP mice with single or combinatorial 1e10 vg AAV-gRNA injection on day −1. (F) Quantification of E: (left) total neutrophil numbers normalized by tumor weight and (right) percentage of Siglec-F⁺ neutrophils among all neutrophils. Data are mean ± SD, and significance was tested using one-way ANOVA with Tukey’s multiple comparison test (n = 10–12 tumors/group, pooled from four independent experiments). *P < 0.05; ns, nonsignificant.

To discover whether Col18a1^hi CAFs mediated recruitment of the neutrophil.4 subset, we analyzed transcriptional expression of all detected CCL/CXCL chemokine family members in the CAF subsets and noted the elevated expression of Cxcl5 in Col18a1^hi CAFs (Fig. 5 D), a chemokine binding to the constitutively expressed chemokine receptor CXCR2 on neutrophils (Griffith et al., 2014). Cancer cell–derived CXCL5 has been described to promote neutrophil infiltration into tumors (Zhou et al., 2014), with Tgfbr2 deletion in breast cancer cells resulting in increased CXCL5 production (Yang et al., 2008). To test if Cxcl5 expression in AAV-gTgfbr2–induced Col18a1^hi CAFs is recruiting the neutrophil.4 subset to the PDAC TME, we used Siglec-F staining as a marker for the neutrophil.4 cluster (Fig. 5 C) and employed combinatorial AAV-gRNA administration (AAV-gCxcl5 plus AAV-gTgfbr2) to KO Cxcl5 together with Tgfbr2 in CAFs (Fig. S5 S). While Siglec-F staining and analysis by flow cytometry did not show a significant increase of Siglec-F⁺ neutrophils in AAV-gTgfbr2 tumors, dual Cxcl5 and Tgfbr2 KO locally in CAFs via combinatorial AAV-Tgfbr2 plus AAV-gCxcl5 delivery did reduce Siglec-F⁺ neutrophil accumulation in the TME (Fig. 5, E and F), implying that CAF-derived CXCL5 contributes to Siglec-F⁺ neutrophil accumulation in PDAC tumors.

Combinatorial KO shows that emergence of Tgfbr2 KO-induced Col18a1^hi CAFs is dependent on TNFR1 and canonical Wnt signaling

Finally, to nominate pathways that could potentially positively amplify the generation of the osteoblast-like AAV-gTgfbr2–initiated Col18a1^hi CAF subset, we used two computational approaches: ligand–receptor analysis using CellChat (Jin et al., 2021) and transcription factor activity inference using decoupleR (Badia-I-Mompel et al., 2022).

Ligand–receptor-based analysis using CellChat to identify potential signaling pathways between the Col18a1 CAF cluster and the TAM.Ndrg1 cluster, as well as the neutrophil.4 cluster, nominated TNF-TNFR1 (Tnf-to-Tnfrsf1a) signaling as a major contributor to the Col18a1 CAF cell state (Fig. 6 A). Both the TAM.Ndrg1 and the neutrophil.4 clusters were most abundant in AAV-gTgfbr2 tumors (Fig. 6 B) and were transcriptionally enriched for the GO hallmark term “TNFa signaling via NFkB” (Fig. 6 C), suggesting that they, together with the Col18a1 CAFs, exist in a TME with high TNFα signaling. Thus, we hypothesized that TNFR1-mediated signaling in AAV-gTgfbr2 CAFs mediates the emergence of the Col18a1/factor 9 osteoblast-like cell state (Fig. 4 F).

Figure 6. Combinatorial KO shows that emergence of Tgfbr2 KO-induced Col18a1^hi CAFs is dependent on TNFR1 and canonical Wnt signaling.

(A) Chord diagrams of ligand–receptor pairs between (left) TAM.Ndrg1 cells (senders) and Col18a1 CAFs (receivers) or (right) neutrophil.4 subset (senders) and Col18a1 CAFs (receivers). Tnf/Tnfrsf1a ligand–receptor pairing highlighted as bold. (B) Left: Proportion of MonoMac subsets (scRNA-seq) in each AAV-gRNA group. Right: Proportion of neutrophil subsets (scRNA-seq) in each AAV-gRNA group. (C) Top: Enriched gene sets (Gene Ontology hallmark) based on DEGs from TAM.Ndrg1 cluster. Bottom: Enriched gene sets (Gene Ontology hallmark) based on DEGs from neutrophil.4 cluster. (D) Heatmap of average activity score calculated by PROGENy (Badia-I-Mompel et al., 2022) of osteoblast-associated transcription factors (Zhu et al., 2024) in CAF subsets. (E) Flow cytometry contour plots of tumor fibroblasts on day 14 after 2.5e5 HY19636 s.c. cell injection in Pdgfra-Cas9-EGFP mice with single or combinatorial 1e10 vg AAV-gRNA injection on day −1. (F) Quantification of E. Data are mean ± SD, and significance was tested using one-way ANOVA with Tukey’s multiple comparison test (n = 2–5 tumors/group, pooled from two independent experiments). *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

For transcription factor activity inference, we focused on osteoblast-associated transcription factors (Zhu et al., 2024) because NMF gene program analysis revealed that the Col18a1 cluster was enriched for osteoblast-related genes (Fig. 4 F). Ctnnb1 and Lef1—two regulators of the Wnt pathway—were nominated as highly active in Col18a1 and all myCAF subsets (Fig. 6 D). This prompted us to hypothesize that Wnt pathway activity—a known contributor to the myofibroblast phenotype in cancer (Mosa et al., 2020) and skin fibrosis (Akhmetshina et al., 2012)—in the context of TGFBR2 signaling loss, produces the Col18a1 osteoblast-like CAF phenotype. To test the contributions of TNFR1 or Wnt pathway signaling in CAFs, we generated AAV-gTnfrsf1a and AAV-gCtnnb1 (Fig. S5, T and U) and found that combinatorial KO of Tgfbr2 with Tnfrsf1a or Ctnnb1 in CAFs (Fig. S5, V and W) decreased the ITGA1⁺ NCAM1⁻ CAF population compared to AAV-gTgfbr2 alone (Fig. 6, E and F), indicating that both pathways are contributing to ITGA1⁺ Col18a1 CAF cluster generation. Furthermore, AAV-gCtnnb1 alone decreased the myofibroblast NCAM1⁺ population to similar levels as AAV-gTgfbr2, while simultaneously maintaining the ITGA1⁻ NCAM1⁻ DN, homeostatic-like CAF population (Fig. 6, E and F).

Using AAV-gRNA–mediated KOs in fibroblasts, we present an experimental toolbox that allows local in vivo gene editing of fibroblasts to study gene functions—both on an individual as well as on a combinatorial level—in the context of mouse tumor models.

Discussion

The modularity of the localized in vivo CAF gene editing approach presented here—relying only on delivering a gRNA packaged in an AAV and conditional Cas9 expression in PDGFRA⁺ fibroblasts—allows customizable, user-defined investigation of individual or combinatorial gene perturbations, while the local application allows spatially confined perturbations that limit potentially confounding effects induced by systemic gene perturbations, such as conventional Cre-loxP–based mouse models. We believe this approach will be beneficial to the advancement of understanding the complexity of fibroblast biology within different disease settings, as gene candidates can be more quickly and cost-effectively evaluated by not fully relying on complex mouse breeding to generate genetically engineered single or multiple KO mouse strains.

Previous to our study, AAV-mediated gene delivery in fibroblasts has been demonstrated in vivo, resulting in transient target gene expression from a classically introduced promoter (Francisco et al., 2021; Rezvani et al., 2016; Song et al., 2016). However, our study goes further in achieving true genome editing, plus improving selectivity for fibroblast lineages as well as the local environment. As an alternative approach to our CRISPR/Cas9 method, another recent study reported the knockdown of a target in fibroblasts in vivo using shRNA (Rajendran et al., 2023). While this will be beneficial in situations when Cas9 expression cannot be achieved in the target cell population, downregulation of mRNA in this method is transient, whereas CRISPR/Cas9-mediated KO allows sustained gene KO and more long-term analysis of gene perturbations—we observed stable KO of up to 14 days in the tumor fibroblast population (Fig. 1). Generating a sustained gene KO is especially valuable when studying long-term processes, excluding confounding effects via residual gene expression in knockdown scenarios. Additionally, the AAV-mediated gene delivery in fibroblasts allows editing in adult animals of genes that can cause embryonic lethality when knocked out constitutively.

Importantly, transduction of fibroblasts with a non-cutting gRNA (gSCR) or a gRNA cutting an irrelevant, non-expressed gene (gTrac) in vivo did not affect polarization of the SCA-I⁺ Ly6C⁺ states noticeably compared with non-transduced fibroblasts (Fig. S3 F). This indicated that at least in the context of tumor growth, the use of AAV1-serotyped scAAV does not affect fibroblast biology on its own and can be considered a fairly innocuous transduction method (Shirley et al., 2020) that does not mask effects of targeted gene KO.

One clear opportunity afforded here is the possibility of sculpting fibroblasts through a series of edits. Fibroblasts receive a multitude of signaling input during tumor growth, as demonstrated by their expression of different receptor families, as well as their transcriptional and functional polarization when receiving these signals (Biffi et al., 2019; Dominguez et al., 2020; Plikus et al., 2021; Wu et al., 2021). Sequential gene edits via subsequent injection of multiple sKO-AAVs or simultaneous edits via co-injection of multiple sKO-AAVs allow uncovering of redundant, synergistic, compensatory, or agonistic effects of gene functions in CAF biology. For example, inflammatory and myofibroblast CAF states have been described as two distinct fates fibroblasts can adopt in PDAC (Biffi et al., 2019; Krishnamurty et al., 2022). While inhibition of either the JAK/STAT or the TGFBR signaling pathway via systemic drug application polarized CAFs to opposite cell states (Biffi et al., 2019), we noticed adoption of a separate osteoblast-like transcriptional cell state when TGFBR2 signaling was specifically inhibited in CAFs only, as opposed to pathway inhibition in any cell that a systemically administered drug could reach. Furthermore, KO of members of the TNF pathway (via Tnfrsf1a KO) or the canonical Wnt pathway (via Ctnnb1 KO) in combination with TGFBR2 diminished CAF polarization toward both myofibroblast and osteoblast-like phenotypes, demonstrating how multiple gene edits via AAV-gRNA delivery can uncover hierarchies of pathway activities in CAFs (Fig. 6).

Lack of TGFBR2 signaling in FSP1⁺ fibroblasts with a resulting increase in secreted hepatocyte growth factor has been described as a transforming factor in epithelial cells, specifically contributing to squamous cell carcinoma of the forestomach, intraepithelial neoplasia in the prostate of mice, and developmental abnormalities in mammary duct development (Bhowmick et al., 2004; Cheng et al., 2005; Eikesdal et al., 2018). Other tissues appeared histologically normal, emphasizing the importance of how organ-specific and temporally controlled fibroblast perturbations (e.g., constitutively throughout development or acute perturbation in the adult animal) can have distinct effects on their local tissue environments. The local fibroblast-specific gene editing afforded us the possibility of temporally controlled perturbation, where Tgfbr2 KO induced a unique CAF cell state that led to the recruitment of distinct neutrophil and TAM subsets. This observation was similar to recent findings of conditional fibroblast-specific KO of Tgfbr2 in either alveolar fibroblasts in the setting of lung fibrosis or in brain border fibroblasts in stroke, where an increase in monocytes or neutrophils, respectively, was observed when fibroblast-specific TGFBR2 signaling was disrupted (Tsukui et al., 2024; Ewing-Crystal et al., 2025). Intriguingly, mice fared worse in both settings when Tgfbr2 was knocked out in fibroblasts, mirroring the worse survival of PAAD in patients with a high Tgfbr2 KO-associated, osteoblast-like gene signature (Fig. 4 G). While the higher morbidity in the lung fibrosis and stroke models could be attributed to a disrupted resolution of inflammation, further work is required to determine how a lack of Tgfbr2 signaling in CAFs contributes to worse survival. In particular, the functional importance of the Col18a1 fibroblasts in the tumor setting remains to be established, likely across multiple settings. Application of this local gene editing platform in orthotopic cancer models and other models of fibrosis in future studies can shed light on their importance, especially by targeting the TNFR1 and/or Wnt signaling pathways, which are required for their formation. Targeting of Col18a1 fibroblast marker genes, such as Col18a1 itself, provides another intriguing strategy to assess the function of Col18a1 fibroblasts in TME formation.

Single-cell studies of tissue ecosystems have highlighted the relationship of immune, stroma, and parenchymal cells in settings of disease and homeostasis (Bonnardel et al., 2019; Combes et al., 2022; Luca et al., 2021; Pelka et al., 2021). Observing these processes over space and time allow generation of experimentally and computationally inferred cell state trajectories within individual cells, as defined by transcriptional changes occurring throughout the process (Hu et al., 2023; Qiu et al., 2022). These types of changes of transcriptional states by individual cells are linked through paracrine signaling with their neighboring cells and have been formulated in two-cell circuits between heterotypic cells, such as macrophages and fibroblasts (Zhou et al., 2018). By manipulating fibroblasts locally with high genetic precision, we were able to extend these findings to test how perturbations in fibroblasts alter the relationships between different cell types and which cell states are linked during multicellular tissue processes. Crucially, this goes beyond computationally inferred cell–cell interactions and affords rapid validation of which mediators orchestrate the observed multicellular changes, which are not necessarily observed on a compositional level of changed major cell type proportions but instead are hidden within linked transcriptional cell state changes. One previously underappreciated linkage discovered here is between TGFBR2 signaling inhibited CAFs recruiting Siglec-F^hi neutrophils via increased Cxcl5 expression (Fig. 5). Siglec-F^hi neutrophils in the TME have been described as CD8 T cell inhibitory (Simoncello et al., 2022), inversely correlated with immunotherapy response (Gungabeesoon et al., 2023), and being recruited by cancer cell–derived Cxcl5 (Zhou et al., 2014). Here, we extend that cell–cell relationship to include Cxcl5⁺ CAFs, demonstrating how their altered state can negatively affect the cellular TME by recruiting these immune inhibitory neutrophils. This highlights how fibroblasts occupy an important niche within the cellular hierarchy in the TME (Mayer et al., 2023), and how their local perturbation allows discovery of linked constellations of cell states at site.

The method defined here thus provides a rapid and site-selective approach to modulate fibroblasts and record the resulting changes within them and the pathways that are modified via single or combinatorial KOs of target genes. We envisage that with this quicker turnaround time from gene target identification to combinatorial gene perturbations, we can learn more, discover faster, and describe better fibroblast functions in their native tissue context.

Limitations

The outlined approach requires the use of a transgenic mouse line expressing Cas9 driven by Cre expression in fibroblasts. While the choice of Cre driver can be further modified to control conditional Cas9 expression in a desired cell population to potentially investigate more defined fibroblast subpopulations or other cell types (Plikus et al., 2021), it does prevent this technology from being directly adopted for use in genetically engineered mouse models (GEMMs) of cancer that rely on Cre expression in the cancer-initiating cell. GEMMs of cancer generally recapitulate aspects of tumor initiation, progression, metastasis, and outgrowth more faithfully than transplanted tumor models, as used in this study. Fibroblasts have been described as being involved in all of these processes (Sahai et al., 2020), emphasizing the need to develop more sophisticated approaches. It is conceivable to combine the Cre-driven Cas9 expression in the fibroblast with alternative recombinase systems, such as the Flp/Frt or Dre/rox systems (Chen et al., 2021; Han et al., 2021; Schönhuber et al., 2014), or non-recombinase–dependent tumor models such as the MMTV-PyMT mammary carcinoma model (Guy et al., 1992), which would allow studying these gene perturbations in GEMMs of cancer.

We also note that when targeting one gene for KO in a cell population, compensatory expression of the targeted gene in the non-edited fraction of the cell population may limit our ability to draw conclusions about the absolute requirement for any one gene in any given cell (e.g., Il1r1 in this study).

The dual KO approach of co-injecting two different AAVs delivering gRNAs targeting two different genes has the disadvantage of potentially generating nonuniform KO efficiencies between the two targeted genes, as seen in this study (Fig. S5, S-W). This can lead to a low percentage of double KO observed within the same cell if the two AAV transduction events are independent and if one transduction is significantly lower than the other, which was seen in some mice, but not all. Future studies targeting more than one gene should use multiplexed designs, wherein only one AAV is used carrying multiple gRNAs in its genome (McCarty et al., 2020). This should increase multiple gene KO efficiencies in the target cell by only requiring efficient transduction of one AAV (Fig. S1 H), circumventing the need for one cell being transduced by multiple different AAVs, and thereby increasing the robustness of this combinatorial local gene editing approach in fibroblasts.

This study investigated CAF cell states as collected from s.c. transplanted pancreatic cancer cells. The heterotopic growth of the tumor in the skin does not fully recapitulate the TME in pancreatic cancer. Thus, aspects of pancreas-specific fibroblasts, such as pancreatic stellate cells, which contribute unique aspects of CAF biology (Sherman and di Magliano, 2023), are not considered here. Orthotopic injection of AAVs into the organ of interest, such as the pancreas, is conceivable, albeit requiring invasive or ultrasound-guided surgical procedures (Camara Serrano, 2022). While we demonstrate feasibility of local CAF editing in transplantable orthotopic pancreatic tumor models using this system (Fig. 1, Q and R), future and extensive studies may apply this approach to isolating disease-specific functions and networks of fibroblast models, beyond those in the skin. We also note that fibroblasts and their gene networks may operate differently in different organs.

Materials and methods

Animals

Mouse strains and breeding

All mice were housed in an American Association for the Accreditation of Laboratory Animal Care–accredited animal facility and maintained in specific pathogen-free conditions. All animal experiments were approved and performed in accordance with the Institutional Animal Care and Use Program protocol number AN200424. WT C57BL/6J (#000664), Pdgfra-Cre (#013148), and Rosa26-LSL-Cas9-EGFP (#026175) mice were purchased from The Jackson Laboratory. Pdgfra-CreERT2 (stock #032770) mice were a kind gift from Dr. Tien Peng (University of California, San Francisco [UCSF], San Francisco, CA, USA). Cthrc1-CreERT2;Ai14 mice were a kind gift from Drs. Tatsuya Tsukui and Dean Sheppard (UCSF, San Francisco, CA, USA). Transgenic Pdgfra-Cre mice were crossed with Rosa26-LSL-Cas9-EGFP^+/+ mice to generate F1 Pdgfra-Cre;Rosa26-LSL-Cas9-EGFP^+/wt mice. F1 Pdgfra-Cre;Rosa26-LSL-Cas9-EGFP^+/wt mice were crossed with Rosa26-LSL-Cas9-EGFP mice to generate F2 Pdgfra-Cre;Rosa26-LSL-Cas9-EGFP^+/+ mice, which were used for experiments. KI Pdgfra-CreERT2 mice were bred in the same manner to generate Pdgfra-CreERT2^+/wt;Rosa26-LSL-Cas9-EGFP^{+/+ or +/wt} mice and used in experiments after tamoxifen treatment (see below) only in Figs. S1 and 2. Col1a2-CreERT2 (MGI 6721050 from Bin Zhou [Shanghai Tech University, Shanghai, China]) (He et al., 2017) were crossed with Tgfbr2^{Exon2–fl/fl} mice (MGI 2384513, kind gift from Dr. Ari Molofsky [UCSF, San Francisco, CA, USA]) (Chytil et al., 2002) to generate Col1a2-CreERT2^+/wt;Tgfbr2^{Exon2–fl/fl} mice. All mice were housed at the UCSF animal facility with typical light/dark cycles and standard chow.

Tamoxifen treatment

For tamoxifen-induced Cre recombination in Cthrc1-CreERT2, Pdgfra-CreERT2, and Col1a2-CreERT2 mice, adult mice aged 6–12 wk were injected i.p with 100 μl of tamoxifen (#T5648; Sigma-Aldrich) dissolved in corn oil or ChremophorEL:EtOH:PBS (ratio 1:1:2) at 20 mg/ml, yielding a dose of 2 mg per mouse, on five consecutive days 1 wk prior to tumor engraftment.

Tumor models

For tumor growth studies and engraftment, YUMM5.2 mouse melanoma cancer cells derived from male Tyr-Cre;Braf^V600E/wt; Trp53^−/− mice (Meeth et al., 2016) (5 × 10⁵ cells/50 μl PBS) or HY19636 mouse pancreatic cancer cells derived from female KPC mice (Kras^LSL–G12D/+;Trp53^lox/+;p48/Ptf1a-Cre) (Bardeesy et al., 2006) (2.5 × 10⁵ cells/50 μl PBS) were transplanted into the s.c. mouse flank of male or female mice, respectively. For tumor growth studies, tumor size was monitored twice a week using the formula 0.5 × (length × width²). Once reaching an endpoint of size >1,000 mm³, mice were euthanized. For indicated analyses, mice were sacrificed at indicated time points, tumors were excised, and processed for downstream analysis.

For intrapancreatic orthotopic growth of HY19636 cells, mice were anesthetized with 3% isoflurane and s.c. injected with 50 μl of 0.25% bupivacaine and 50 μl of 50 μg/ml buprenorphine for analgesia. The hair was removed on the left side flank, and the skin was prepared by three alternating rounds of applying ChloraPrep swab stick (2% wt/vol chlorhexidine gluconate and 70% vol/vol isopropyl alcohol) and alcohol wipes. A small incision was made in proximity to the spleen, both in the skin and the peritoneum, and the pancreas was carefully exposed by gently pulling on the spleen with forceps. For injection, 2.5e5 HY19636 cells and 3e10 vg of AAV were co-injected in a total volume of 50 μl of PBS into the tail of the pancreas using a 31G insulin syringe. After injection, the peritoneal incision was closed using 4-0 PDS II sutures (#Z823G; Ethicon), and then the skin incision was closed using the BD AutoClip Wound Closing System. Mice were monitored for 48 h after surgery for any signs of distress and given second dose of buprenorphine 4–6 h after surgery, third dose the following morning, fourth dose in the afternoon, and additional doses as needed.

Bleomycin-induced skin fibrosis model

For induction of skin fibrosis, 6–20-wk-old male C57BL/6 mice Pdgfra-Cre;Rosa26-LSL-Cas9-EGFP^+/+ mice were injected s.c. with 100 μl of 1 U/ml bleomycin (#16714090801; NorthStar Rx) dissolved in PBS three times per week for 3 wk into a 1 × 1 cm² area on the dorsal skin of the animal, rotating injection site at every injection.

Cell lines

The YUMM5.2 mouse melanoma cell line (CRL-3367; ATCC), the HEK293T cell line (CRL-3216; ATCC), and the mouse fibroblast cell line NIH/3T3 (CRL-1658; ATCC) were purchased from ATCC. The HY19636 pancreatic tumor cell line was a kind gift from Dr. Hiaoqing Ying (MD Anderson Cancer Center, Houston, TX, USA). The Lenti-X 293T cell line was purchased from Takara (#632180). All cell lines were cultured at 37°C in 5% CO₂ in cell growth media: DMEM (11995065; Gibco) supplemented with 10% FBS (Foundation C, 900-308; Gemini Bio), 2 mM L-glutamine, 100 U/ml penicillin, 100 μg/ml streptomycin, and 50 μM β-mercaptoethanol (21985-023; Gibco). For all tumor injections, cells were used within the first two passages after thawing. The HY19636 cell line was kept below passage #10, and the YUMM5.2 cell line was kept below passage #20 for experiments.

To generate NIH/3T3 cells expressing Cas9 (NIH/3T3.Cas9.EGFP), NIH/3T3 cells were lentivirally transduced with the viral supernatant produced in Lenti-X 293T cells after transfection with the following plasmids: cargo plasmid containing Cas9 and EGFP (pL-CRISPR.EFS-GFP, #57818; Addgene), VSV-G envelope–containing plasmid (pMD2.G, #12259; Addgene), and envelope plasmid (pCMV-dR8.91, #OVT2971; Creative Biogene) at a molar ratio of 1:0.44:0.8, respectively, and added polyethylenimine (PEI), linear, MW 25000 (#23966; Polysciences) at a PEI-to-DNA ratio of 4-to-1. Two days after Lenti-X 293T transfection, viral supernatant was filtered through a 0.45-μm filter (Whatman) and added to NIH/3T3 cells with supplemented polybrene (#TR-1003-G; Sigma-Aldrich) at a final concentration of 8 μg/ml. Cas9-expressing NIH/3T3 cells were enriched by sorting EGFP⁺ cells.

AAV production

The helper plasmid (pHelper, #6234; Takara) was used in conjunction with one of three different RepCap plasmids (7m8, #64839; Addgene [Isgrig et al., 2019]; pAAV-DJ [Grimm et al., 2008] Cell Biolabs, VPK-420-DJ; pAAV2/1, #112862; Addgene), and a scAAV cargo vector (see design below) to package vector genomes into double-stranded AAV particles. For AAV production, on day 1, three 15-cm dishes were seeded, each with 9e6 HEK293T cells (ATCC, CRL-3216) in 23 ml of cell growth media DMEM (11995065; Gibco) supplemented with 10% FBS (Foundation C, 900-308; Gemini Bio), 2 mM L-glutamine, 100 U/ml penicillin, 100 μg/ml streptomycin, and 50 μM β-mercaptoethanol (21985-023; Gibco). On day 2, the HEK293T cells were transfected using PEI at a PEI-to-DNA ratio of 8-to-1 (660 μg total PEI), with the DNA comprising 20.25 μg cargo plasmid, 26 μg RepCap plasmid, and 36.36 μg pHelper, all in a total volume of 7.5 ml of a 150 mM NaCl solution. Each plate of HEK293T received 2.5 ml of the PEI-DNA transfection reagent by dropwise adding it on top. Transfected HEK293T cells were cultured for 3 more days, and then AAV was collected using the AAVpro Purification Kit Midi (6675; Takara) following the manufacturer’s instructions. After purification, aliquots were taken to determine titers by qPCR after DNase I (#B0303S; NEB) treatment and proteinase K (#1114886; Qiagen) digestion. For qPCR, primers amplifying a 138-bp region of the hU6 promoter on the cargo plasmid within the ITRs were used (titer_U6_Forward, 5′-GAGGGCCTATTTCCCATGATTCC-3′; titer_U6_Reverse, 5′-CCCAAGAAATTATTACTTTCTACGTCACG-3′) with the SsoAdvanced Universal SYBR Green Supermix (#1725270; Bio-Rad) on CFX384 BioRad Real-Time PCR System with the following cycling protocol: 95°C for 30 s, 95°C for 5 s, 58°C for 20 s, and repeat steps 2 and 3 40 times. Relative quantity was determined by comparison with a serial dilution of the cargo plasmid standard of known concentration. AAVs were stored in autoclaved protein low-binding tubes (3207; Costar) and handled with protein low-binding pipets (4151; Corning).

AAV cargo plasmid design

To generate the pscAAV-hU6-gSCR-CBh-mCherry plasmid (#249124; Addgene) for generation of scAAVs carrying a gRNA cassette plus mCherry as a reporter gene for transduction efficiency, the pscAAV-CAG-EGFP plasmid (83279; Addgene) was cut using AvrII and NotI, and then Gibson assembled (Gibson et al., 2009) with a PCR-amplified fragment from pAAV-U6-gRNA-CBh-mCherry (forward primer, 5′-GTTCCTGGAGGGGTGGAGTCG TGACCTAGGGAGGGCCTATTTCCCATGATTCCTT-3′; reverse primer, 5′-AAAGCATCGAGATCGCAGGTGAGGCCTAGCGGC CGCTTACTTGTACAGCTCGTCCATGCC-3′). This yielded the intermediate plasmid pscAAV-hU6-gRNA-CBh-mCherry, which was never used for scAAV production because it had a 77 bp sequence between the RNA polymerase III termination sequence at the 3′ end of the gRNA scaffold sequence and the KpnI cut site at the 5′ end of the CBh promoter sequence. The pscAAV-hU6-gRNA-CBh-mCherry plasmid was then cut with AvrII and KpnI and Gibson assembled with a 402-bp long double-stranded DNA (dsDNA) oligo synthesized by Twist Biosciences containing the scrambled gSCR sequence (5′-GCTTAGTTACGCGTGGACGA-3′) (Suzuki et al., 2016), which does not exist in the murine genome; gRNA scaffold; and shortened sequence between the gRNA scaffold and CBh promoter, yielding the pscAAV-hU6-gSCR-CBh-mCherry plasmid with a 2053 bp insert from 5′ ITR to 3′ ITR sequence. This plasmid was then used to clone all other plasmids containing a single gRNA by cutting it with AvrII and KpnI and then inserting a dsDNA oligo (synthesized by Twist Biosciences) containing the U6 promoter sequence plus a gRNA sequence of interest. These plasmids were then used for scAAV packaging.

To generate the KIKO plasmid pscAAV-KIKO-U6-gThy1-p2a-mCherry-STOP-pA-400 (#249129; Addgene), the pscAAV-hU6-gSCR-CBh-mCherry plasmid was cut with AvrII and SpeI and Gibson assembled with two dsDNA oligos synthesized by Twist Bioscience containing the hU6-gThy1 sequence, gRNA scaffold, 400-bp left homology arm complementary to the sequence upstream of the gThy1 cut site, P2A element, mCherry ORF with stop codon, SV40late polyadenylation sequence, and a 380-bp right homology arm complementary to the sequence downstream of the gThy1 cut site. All other KIKO constructs were generated using the same cloning strategy with adjusted homology arms and gRNA sequences depending on KI site. All plasmid and gRNA sequences used in this study are found in Table S5 and have been deposited on Addgene (#249119-249131).

AAV transduction in vitro

For validation of AAV-mediated KO, NIH/3T3.Cas9 cells were used as target cells and analyzed by flow cytometry to assess KO efficiencies. On day 1, 1e4 NIH/3T3.Cas9 cells were seeded in 500 μl culture media per well in a 24-well cell culture plate. On day 2, 16 h later, AAV was added at specified multiplicity of infection. After 3–5 days of incubation, culture media was aspirated; cells were washed once with PBS and collected using trypsin (0.05% Trypsin-EDTA, #25300054; Gibco) to detach the cells. Cells were further processed for flow cytometry analysis of surface protein expression of target gene KO and mCherry reporter gene expression transduction efficiency. See below for details on flow cytometry analysis.

AAV application in vivo

All AAV injections in this study were done s.c. on the dorsal lateral flank of mice 24 h prior to tumor engraftment at the same site. The hair at the site of injection was removed using clippers and cleaned using alcohol wipe pads. The doses used were either 1e10-3e10 vg AAV per injection in 50 μl PBS using a 31-G insulin syringe (163311564B; Sol-M). The actual dose used for each experiment can be found in the figure legends. AAV injection sites were demarcated using green tattoo paste (2420101; Thermo Fisher Scientific), allowing engraftment of tumor cells at the same site the next day (see Fig. S1 F).

Indel frequency analysis

Indel frequencies within target cell populations at gRNA target sites were estimated using genomic DNA as input for amplicon PCR surrounding the gRNA target sites, followed by Sanger sequencing and ICE analysis (Synthego Performance Analysis, ICE Analysis. 2019. version 3.0. Synthego). Primers used for amplicon PCRs are listed in Table S6.

Tissue collection and processing for flow cytometry

Tumor and skin samples

At time of analysis, mice were euthanized, and tumor mass was carefully excised from the s.c. region without collecting the adjacent skin and surrounding adipose tissue. For skin collection, a 3 × 3 cm² region on the dorsal flank, cleared from hair using clippers, was excised using scissors, and excess fat was removed. On ice, samples were finely minced with scissors and then placed in a 2-ml tube containing 1.5 ml of digestion medium (2 mg/ml collagenase XI [C7657; Millipore Sigma], 0.5 mg/ml hyaluronidase [LS005477; Worthington], and 0.1 mg/ml DNase [10104159001; Roche] in DMEM [11995065; Gibco] supplemented with 10% FBS [Foundation C, 900–308; Gemini Bio], 2 mM L-glutamine, 100 U/ml penicillin, 100 μg/ml streptomycin, and 50 μM β-mercaptoethanol [21985-023; Gibco]). The tube was placed horizontally in a bacterial shaker for 45 min at 37°C and 225 rpm. The tumor samples were then filtered through a 40-μm filter, and skin samples were filtered through a 100-μm filter. Samples were then washed with 10 ml cold PBS. The generated single-cell suspension was then processed further depending on analysis method.

Spleen and LNs

Inguinal and brachial LNs were collected, and the capsule was mechanically torn open using two 25-G needles prior to digestion. LNs and spleens were then separately digested in 1 ml digestion media (2 mg/ml collagenase D, collagenase IV [50 μg/ml] in DMEM with 2% FBS, and 10 mM HEPES [Sigma-Aldrich, H0887]) for 30 min at 37°C without agitation. Every 10 min, samples were pipetted up and down five times using a P1000 pipette. Samples were filtered through a 35-μm filter and washed with 3 ml of cold PBS. After pelleting the spleen sample at 500 g for 5 min in a centrifuge, the sample was resuspended in 1 ml of RBC lysis buffer (11814389001; Roche) and incubated for 10 min on ice to lyse red blood cells. The lysis was quenched by adding 4 ml of fluorescence-activated cell sorting (FACS) buffer (staining buffer [PBS, 2% FBS, and 2 mM EDTA]). Cells were pelleted again at 500 g for 5 min in a centrifuge and resuspended in FACS buffer and used for further analysis.

Pancreas tumor processing

To collect orthotopically grown tumors after intrapancreatic tumor cell injection, mice were euthanized, and tumor mass was carefully excised from the pancreatic tissue. Tumor tissue was minced with scissors on ice in 200 μl of digestion buffer (1× PBS, 4% bovine serum albumin [A4503; Sigma-Aldrich], 0.2 mg/ml trypsin inhibitor [T6522; Sigma-Aldrich], 0.2 mg/ml DNase I, 2 mg/ml collagenase XI, and 0.5 mg/ml hyaluronidase), then incubated for 30 min in 5 ml total digestion buffer on an orbital shaker at 37°C and 180 rpm in a 15-ml conical tube. Sample was pulse vortexed once every 10 min. Digestion was stopped by adding 5 ml of Stop buffer (1× PBS, 5% FBS [Foundation C, 900-308; Gemini Bio], and 0.2 mg DNase I). Single-cell suspension was filtered through 70-μm nylon filter, washed with 5 ml of 1× PBS, and pelleted at 350 g for 5 min at 4°C. Red blood cells were lysed with RBC lysis buffer (11814389001; Roche) for 10 min on ice. Cells were washed and pelleted again at 350 g for 5 min at 4°C and then used for subsequent processing and analysis.

Flow cytometry

AAV validation in vitro

NIH/3T3.Cas9 cells transduced with AAVs in vitro were collected, washed, and pelleted at 500 g for 5 min centrifugation in FACS staining buffer (PBS, 2% FBS, and 2 mM EDTA) and resuspended in 50 μl of staining buffer plus 0.5 μg purified anti-mouse CD16/32 antibody (clone 2.4G2, Tonbo Biosciences). Samples were incubated for 10 min on ice to block FC receptors. Following this incubation, the cells were stained with antibodies against surface receptors that were targeted by AAV. Antibodies were diluted in 50 μl staining buffer prior to addition of cells to yield a final volume of 100 μl staining volume. Antibodies used for staining and their dilutions are listed in Table S6. The cells were incubated for a further 10 min on ice in the dark. After this staining incubation, cells were washed with 1 ml of staining buffer, pelleted again at 500 g for 5 min, and resuspended in staining buffer containing DAPI (1 μg/ml) as a live/dead stain. Cells were analyzed on a BD LSRFortessa Cell Analyzer (BD Biosciences) at the Parnassus Flow Cytometry core at UCSF and gated on live (DAPI-negative) singlets. Loss of surface receptor expression compared with naive (not exposed to AAV) NIH/3T3.Cas9 cells was used as a readout of KO efficiency, and expression of mCherry was used as a readout for transient AAV expression in the KO constructs.

Staining of tumor and tissue samples

Single-cell suspensions from tumors and other tissues were generated as described above, pelleted by centrifugation at 500 g for 5 min, resuspended in 5e6 cells/50 μl FACS staining buffer containing 1 μg purified anti-mouse CD16/32 antibody (clone 2.4G2; Tonbo Biosciences). Samples were incubated for 10 min on ice to block FC receptors. Following this incubation, 50 μl FACS staining buffer containing antibodies for surface staining was added and incubated for 20 min at room temperature in the dark. Antibodies used for staining and their dilutions are listed in Table S6. Cells were washed with 1 ml of FACS staining buffer, pelleted, and resuspended in FACS staining buffer containing DAPI (1 μg/ml) for live cell discrimination prior to analysis on a BD LSRFortessa Cell Analyzer (BD Biosciences).

FACS of fibroblast subsets

For sorting of tumor fibroblast subsets, tumors were processed as described above to generate a single-cell suspension and stained as described in the previous section. After staining, cells were filtered on a 35-μm filter right before being sorted on a BD FACSAria II cell sorter (BD Biosciences) Parnassus Flow Cytometry core at UCSF. Cells were collected based on the gating scheme outlined in Fig. S1 B to yield live singlet fibroblasts as defined by DAPI⁻ CD31⁻ CD45⁻ E-Cadherin⁻ MCAM⁻ Thy1⁺ PDPN⁺ EGFP⁺, and then further subsetted as outlined in Fig. S2 K to sort four fibroblast subsets: DP SCA-I⁺ Ly6C⁺, Int SCA-I^mid Ly6C^mid, DN SCA-I⁻ Ly6C⁻ NCAM1⁻, and DN SCA-I⁻ Ly6C⁻ NCAM1⁺. Sorted cells were collected in 4°C cold cell culture media (DMEM with 10% FCS, 2 mM L-glutamine, 100 U/ml penicillin, 100 μg/ml streptomycin, and 50 μM β-mercaptoethanol). For each sample, 2,000–24,000 cells were collected per fibroblast subset. Individual samples were immediately pelleted by centrifugation at 500 g for 5 min at 4°C and processed for qPCR analysis as described below.

Ex vivo stimulation and immunofluorescence imaging of CAFs

For ex vivo culturing, stimulation, and subsequent immunofluorescence imaging (Fig. 2 F), CAFs from day 14 HY19636 tumors that had received 3e10 vg AAV-KIKO-Thy1 (control), AAV-KIKO-Osmr, or AAV-KIKO-Tgfbr2 to generate local KIKOs on day 1 were isolated and processed for sorting as described above using the following markers: samples were stained with anti-mouse CD31-PerCP-Cy5.5 (clone 390; BioLegend) and anti-mouse CD45-BUV395 (clone 30-F11; BD Biosciences) and sorted as singlets DAPI⁻ CD31⁻ CD45⁻ EGFP⁺ cells on a BD FACSAria II cell sorter (BD Biosciences) Parnassus Flow Cytometry core at UCSF. 2,000–3,000 EGFP⁺ CAFs were collected from each tumor in 500 μl of DMEM (11995065; Gibco) supplemented with 10% FBS (Foundation C, 900-308; Gemini Bio), 2 mM L-glutamine, 100 U/ml penicillin, 100 μg/ml streptomycin, and 50 μM β-mercaptoethanol (21985-023; Gibco) and then directly transferred into one well of a 8-well Lab-Tek ChamberSlide with removable gasket (177445; Lab-Tek). CAFs were allowed to adhere to slide for 14 h during incubation at 37°C in 5% CO₂. CAFs from AAV-KIKO-Thy1 (control) and AAV-KIKO-Osmr–treated tumors were stimulated with recombinant mouse OSM (762802; BioLegend) at a concentration of 50 ng/ml for 20 min at 37°C in 5% CO₂. CAFs from AAV-KIKO-Thy1 (control) and AAV-KIKO-Tgfbr2–treated tumors were stimulated with recombinant mouse TGF-β1 (763102; BioLegend) at a concentration of 1 ng/ml for 20 min at 37°C in 5% CO₂. After removing media and washing in 1× PBS, cells were then fixed in 4% formaldehyde (18814; Polysciences, Inc.) for 15 min at room temperature. Fixative was aspirated, and fixed cells were rinsed three times in 1× PBS for 2 min each. Cells were then permeabilized with ice-cold 100% methanol (179337; Millipore Sigma) for 10 min at room temperature. Methanol was aspirated, and cells were rinsed three times in 1× PBS for 2 min each. Cells were then blocked in blocking buffer (1× PBS, 5% mouse serum, and 0.3% Triton X-100) for 60 min at room temperature and then stained with anti-human/mouse phospho-STAT3-AF647 (clone 4/P-STAT3, BD Biosciences, 1:100) or anti-human/mouse phospho-SMAD2/3-AF647 (clone 072-670, BD Biosciences, 1:100) plus DAPI (1 μg/ml) in antibody staining buffer (1× PBS, 1% bovine serum albumin [A4503; Sigma-Aldrich], and 0.3% Triton X-100 [T8787; Millipore Sigma]) for 16 h at 4°C in the dark. Cells were then washed three times in 1× PBS, the gasket was removed from the Lab-Tek slide, and cells were mounted using VECTASHIELD PLUS (H-1900; Vector Laboratories, Inc.). For imaging, samples were acquired on a Leica Stellaris 5 (Leica) inverted laser scanning confocal microscope with a water immersion 25 × HC Fluotar VISIR (506375; Leica) objective using the 405, 488, and 638 nm laser lines to excite DAPI, EGFP, and AF647, respectively. A 4 × 4 tiled image was acquired, and images were analyzed using the Imaris software suite (Bitplane). For quantification, CAFs were identified based on EGFP signal, and all DAPI⁺ nuclei were counted, and a percentage of DP DAPI⁺ AF647⁺ nuclei of all DAPI⁺ nuclei was calculated per sample.

Tissue collection and processing for microscopy

Tissue collection

S.c. tumors were excised from flank skin, including adjacent skin layers and then fixed in 4% paraformaldehyde (PFA) (16% PFA, Electron Microscopy Sciences, diluted in 1× PBS) at 4°C overnight on a rotator. This was followed by a 15% and 30% wt/vol sucrose in PBS incubation step at 4°C for 8 and 16 h, respectively. Samples were then embedded and frozen in OCT (4583; Sakura). 10- or 200-μm thick sections were then made using a cryostat and placed into PBS to wash OCT residue away. 200 μm samples were mounted using VECTASHIELD PLUS (H-1900; Vector Laboratories, Inc.) and then used for second-harmonic generation (SHG) imaging if no additional staining was required. For additional staining of PDPN on 10-μm sections, sections were preprocessed by washing once for 5 min in deionized H₂O at room temperature, refixed with 4% PFA for 10 min at room temperature, and then washed again for 5 min in deionized H₂O at room temperature. Sections were incubated with blocking buffer (1× PBS, 5% hamster serum, and 0.3% Triton X-100) for 1 h at room temperature in the dark and stained with anti-mouse PDPN-APC (clone 8.1.1., 127409; BioLegend; 1:100) plus DAPI (1 μg/ml) for 16 h at 4°C. Samples were then washed three times in deionized H₂O prior to being mounted using VECTASHIELD PLUS.

Immunofluorescence imaging of 10 μm sections

For 10-μm sections, samples were imaged using a Leica SP8 laser scanning confocal microscope with a white light laser (for detection of EGFP, mCherry, and PDPN-APC) and 405 nm diode (for detection of DAPI), using a 20× 0.75NA (HC PL APO 20×/0.75 IMM CORR CS2, Leica) objective. After acquisition, individual tiled images were stitched together using the LAS X software (Leica) and then analyzed using the Imaris software suite (Bitplane).

SHG and immunofluorescence imaging using 2-photon microscopy of 200-μm sections

For 200-μm sections, a custom-built 2-photon setup equipped with two infrared lasers (8W Mai Tai Ti:Sapphire, Spectra Physics; 18W Chameleon Vision II, Coherent) was used. The Mai Tai laser was tuned to 920 nm for excitation of EGFP. The Chameleon laser was tuned to 780 nm for simultaneous excitation of mCherry and detection of SHG. The two lasers were exciting in alternating sequences. Emitted light was detected using a 25 × 1.05-NA water lens (Olympus, XLPlan N) coupled to a six-color detector array (custom; utilizing Hamamatsu H9433MOD detectors). Emission filters used were: violet 417/50, green 510/42, red 607/70 to detect SHG, EGFP, and mCherry emission, respectively. The microscope was controlled using the Micro-Manager software suite (Pinkard et al., 2016) to acquire 16 z-stacks at z-depth of 10 μm to cover 150 μm total z-dis-tance of the 200-μm thick sample. Multiple adjacent z-stacks were acquired to cover a larger XYZ tissue section. The individual z-stacks were stitched together using the Imaris-Micro Magellan compiler plugin from the Micro-Manager software suite. Data analysis was performed using the Imaris software suite (Bitplane).

Reverse transcription real-time quantitative PCR of sorted fibroblast subsets

Fibroblasts from tumors were isolated and sorted as described using the gating strategy outlined in Fig. S1 B and Fig. S2 K. Cells were washed and pelleted in 1× cold PBS, and RNA was isolated using the Qiagen RNeasy Micro Kit (#74004) according to the manufacturer’s instructions. All isolated RNA was used as input for cDNA generation using the iScript Reverse Transcription Supermix (#1708840; Bio-Rad) following the manufacturer’s instructions. The reaction mix was used for qPCR analysis using the SsoAdvanced Universal SYBR Green Supermix (#1725270; Bio-Rad) on a Bio-Rad thermocycler with the following cycling conditions: 95°C for 30 s, 95°C for 5 s, 60°C for 10 s, repeat of steps 2 and 3 39 times, and then melt-curve analysis 65–95°C with 0.5°C increments at 5 s each. Each sample was run in technical triplicates for each assayed gene. 18 s expression was used for the reference gene normalization. Primer sequences used can be found in the Table S6.

scRNA-seq reference data sets

The previously published data sets by Hu et al. (2023) (mouse wound fibroblasts) and Krishnamurty et al. (2022) (mouse pancreatic tumor fibroblasts) were used as reference data sets for scRNA-seq analysis. Deposited data was downloaded from Gene Expression Omnibus (GEO) (Hu et al., 2023; GSE204777) or ArrayExpress (Krishnamurty et al., 2022; E-MTAB-12028). To generate the wound fibroblast Seurat object from Hu et al. (2023) the data was processed as described (Hu et al., 2023). To generate the tumor fibroblast Seurat object, the ReadMtx function from the Seurat R package (version 5.0.1) (Hao et al., 2024) was used to read in matrix.mtx, genes.tsv, and barcodes.tsv files. The CreateSeuratObject function was then used to generate the tumor fibroblast Seurat object. Cells were annotated using the meta.txt file, and fibroblasts were extracted based on rows matching “fibroblast” in the “CellType” column using the subset function. Cells were removed according to the feature count (<300) and mitochondrial count (>5%) parameters outlined in the original paper. The remaining fibroblasts were then further processed using the following Seurat functions in the listed order: NormalizeData (log-normalize RNA transcript counts), FindVariableFeatures (selection.method “vst,” nfeatures = 2000), ScaleData (default parameters), RunPCA (default parameters), FindNeighbors (15 PCs), and RunUMAP (13 PCs) were used for 2D visualization. Clustering was performed using the FindClusters function (resolution of 0.3), and cluster 5 was removed as it was identified as a contaminating immune cell cluster (~1% of total object, 66 of 6,589 cells). The remaining tumor fibroblasts and wound fibroblasts were used to plot gene expression of homeostatic and activation marker genes using the DotPlot, FeaturePlot, and VlnPlot functions in the Seurat package.

Tumor collection and processing for scRNA-seq

For tumor collection and subsequent analysis by live cells from day 14 HY19636 tumors (after s.c. injection of 2.5e5 HY19636 on day 0 into Pdgfra-Cas9-EGFP mice) that had received 3e10 vg scAAV1-gTrac, scAAV1-gOsmr, scAAV1-gTgfbr2, or scAAV1-gIl1r1 to generate local KOs on day 1 were isolated and processed for sorting as described above on a BD FACSAria II cell sorter (BD Biosciences, Parnassus Flow Cytometry Core at UCSF) with minor modifications: live cells were identified as negative for Zombie-NIR (BioLegend, 1:200) and positive for Calcein blue AM (C1429; Thermo Fisher Scientific, 10 μM working concentration). Samples were additionally stained with anti-mouse CD45-BUV395 (clone 30-F11; BD Biosciences, 1:500), anti-Thy1.2/CD90.2-BV786 (clone 53–2.1; BD Biosciences, 1:500, for Thy1 detection), and anti-E-Cadherin/CD324-PE-Cy7 (clone DECMA-1; BioLegend, 1:200). Four live cell populations were sorted: (1) CD45⁺ (immune cells), (2) CD45⁻ EGFP⁺ Thy1⁺ (fibroblasts), (3) CD45⁻ EGFP⁻ E-Cadherin⁻ (miscellaneous), and (4) CD45⁻ EGFP⁻ E-Cadherin⁺ (tumor cells) (see Fig. S4 A for gating scheme). The four populations from each sample were then pooled in a 15-ml Falcon tube and pelleted for 7 min at 500 g. The supernatant was removed, and cells were further processed using the Chromium Fixed RNA Profiling (Gene Expression Flex) pipeline. In brief, each sample was first fixed using the Single Cell Fixed RNA Sample Preparation Kit (PN-1000414; 10x Genomics) following the manufacturer’s instructions. Samples were fixed for 18 h at 4°C and, after addition of quenching buffer and storage buffer, stored at −80°C until further processing. After thawing of samples, samples were processed using the Chromium Fixed RNA Kit, Mouse Transcriptome (PN-1000496; 10x Genomics), following the manufacturer’s instructions to generate a fixed RNA gene expression library. For multiplexing, each sample (4 replicates × 4 AAV groups = 16 total) was hybridized with a separate Probe Barcode library including added custom probes (Egfp, 8casd09, 5′-GGTAGTGGTCGGCGAGCTGCA CGCTGCCGTCCTCGATGTTGTGGCGGATC-3′; and Egfp, 8casd10, 5′-AGGGTGTCGCCCTCGAACTTCACCTCGGCGCGGGTCTTGTAG TTGCCGTC-3′). Approximately 2.5e5 total cells were encapsulated and sequenced on two lanes of a NovaSeq X flow cell (1.25 B reads/lane) to reach about 10,000 reads/cell. Sequencing output was processed using cellranger-7.2.0 and the mouse transcriptome probe set v1.0.1_mm10-2020-A with added custom Egfp probes from above. After demultiplexing the 16 samples using the provided Probe Barcodes, the samples were integrated in Seurat version 5.1.0 (Hao et al., 2024), and low-quality cells were filtered out if they had <150 genes or >20% mitochondrial genes. Features with counts in <3 cells were also filtered out. After merging all samples using the Seurat merge() function, the gene expression matrix was normalized using Seurat NormalizeData() function, with default parameters. The top 2,000 variable genes were identified, and the gene expression matrix was scaled prior to principal component analysis (PCA). The first 23 PCs were used for UMAP visualization based on the elbow plot cutoff, and the FindClusters() function at a resolution of 0.2 was used to identify individual clusters. The following markers were used to identify major cell types: fibroblasts (Pdgfra, Col1a2, and Egfp), pericytes (Rgs5), epidermal basal cells (Krt14; Krt5), endothelial cells (Pecam1), tumor cells (Cdh1; Msln), T cells and NK cells (Trac; Ncr1), neutrophils (S100a8), mast cells (Mcpt4), and myeloid cells (H2-Ab1; Lyz2). Major cell types were subsetted and further processed to remove doublets. DCs were further subsetted from the “myeloid cells” object. This generated nine separate objects: fibroblasts (n = 9,514), pericytes (n = 745), endothelial cells (n = 3,895), tumor cells (n = 81,588), T and NK cells (n = 5,209), neutrophils (n = 43,824), mast cells (n = 1,581), MonoMacs (n = 35,810), and DCs (n = 5,059). Tumor cells were subsampled to a total of 10,000 cells, and then all objects were remerged for Fig. S4 B. The individual fibroblasts, MonoMac, DCs, neutrophils, and T/NK cell objects were further sub-clustered, and the DEGs of the individual sub-clusters can be found in Table S1.

Pseudobulk differential gene expression analysis

For identification of DEGs between CAFs of different AAV-gRNA groups, we used a pairwise pseudobulk approach comparing the control group (AAV-gTrac) with the three different experimental groups (AAV-gOsmr, AAV-gTgfbr2, or AAV-gIl1r1) individually. Each group was represented by four replicates, and DEGs were identified using MAST with random effects (Finak et al., 2015; Zimmerman et al., 2021). The number of DEGs are grouped by cell type with a threshold of P value <0.05 after false discovery rate correction using the Benjamini–Hochberg procedure. The list of DEGs for each experimental group can be found in Table S2.

Pseudotime analysis

Pseudotime analysis of the fibroblast scRNA-seq object

Prior to pseudotime analysis of the fibroblast object, the “myCAF proliferating” cluster was removed. After rerunning PCA, a UMAP representation was generated using the first 18 PCs (Fig. 4 A) in Seurat. The Seurat object was converted to a cell_data_set object using Monocle 3 (Trapnell et al., 2014) (version 1.3.7). The pseudotime trajectory was calculated using Monocle 3 with the homeostatic fibroblast cluster used as the root cell state. To plot the distribution of cells in pseudotime as a violin plot in Seurat, the position of each cell within the pseudotime graph was extracted from the cell_data_set object and then added as metadata to the Seurat object.

Pseudotime trajectory analysis of AAV-gTgfbr2 CAF object

The Seurat object containing AAV-gTgfbr2 CAFs only was analyzed independently to identify potential branchpoints in fibroblast polarization trajectory when Tgfbr2 is knocked out. The monocle developmental trajectory was constructed using Monocle 2 (Qiu et al., 2017) (monocle R package, version 2.34.0). Raw count matrices were extracted and used to generate the CellDataSet object. Size factors and dispersions were estimated using estimateSizeFactors() and estimateDispersions() functions. Ordering genes were identified based on differential expression across cell identity clusters using differentialGeneTest(), and the top 1,000 genes ranked by q value were selected as ordering genes. Pseudotime was computed using reduceDimension() (method = “DDRTree”) and orderCells() functions. Trajectories were visualized using plot_cell_trajectory(). Branch-dependent gene expression changes were analyzed using Branch Expression Analysis Modeling at branch_point = 1, and the top 50 branch-specific genes were visualized using plot_genes_branched_heatmap() with num_clusters = 2 to highlight distinct gene modules associated with divergent trajectories.

Comparison of gene expression similarity between CAFs and existing data sets

To measure the similarity of DEGs between CAF clusters in this study to published data sets, we performed the hypergeometric test to find overlapping genes using the phyper() function from the stats package in R. When comparing to published human data sets, the human genes from those data sets were label transferred to mouse genes using the homologene (version 1.4.68.19.3.27) (Mancarci, 2018) R package. DEGs from each individual data set were filtered based on a P value <0.05 and an average log2 fold change >0.25.

NMF decomposition

NMF was performed on the normalized gene-by-cell count matrix of the fibroblast object using the nmf function from the RcppML (version 0.3.7) R package (DeBruine et al., 2021, Preprint). All genes expressed in >2% of all cells were used as input. Mitochondrial (starting with “mt-”) and ribosomal (starting with “Rps” or “Rpl”) genes were excluded. To select the number of factors used for factorization (Fig. S5 H, top), first, the mean squared error (MSE) of an NMF model was plotted versus its factorization rank for NMF models of rank 1–50. Second, an inflection point was identified by plotting the difference between successive MSE values and then selecting the rank where the MSE values plateau (Fig. S5 H, bottom). K = 18 was the rank where the MSE values plateaued and was thus chosen as the K value in the nmf() function of the RcppML package using 1,000 maximum iterations (maxit = 1,000) and tol = 1e-5 as the stopping criteria. The factor-weight-by-cell matrix (h matrix output) was added to the fibroblast Seurat object as a new “NMF” assay using the Seurat CreateAssayObject function. The factor weight values from the NMF assay were used to visualize NMF factor weights per cell in UMAP representation (Fig. 4 D). The AverageExpression function in Seurat was used to calculate the average NMF factor weights per sample.

NMF factor weight correlation across CAFs

The NMF factor weights of each cell were correlated in pairwise fashion using the cor() function and “method = pearson” from the stats package. Statistical significance was measured using cor.mtest() function from the corrplot (version 0.95) R package (Wei and Simko, 2010).

GSEA

GSEA for fibroblast NMF factors was done using the hypeR (version 2.0.0) package (Federico and Monti, 2020) using R (version 4.3.2). The top 50 genes of each NMF factor were inputted for GSEA, and statistically significant enrichment among the gene ontology term “Biological Process” from the Molecular Signatures Database (MSigDB version 7.2.1) was computed using the hypergeometric test.

Survival analysis of TCGA data based on gene signatures

Data download and processing

The TCGA transcriptomic TPM data were downloaded from the Toil recompute (Vivian et al., 2017) in the TCGA Pan-Cancer cohort on the University of California, Santa Cruz, Xena browser (http://xena.ucsc.edu) and imported into R. The tumor samples were filtered down to only include primary solid tumors (sample type code = 01), and tumor types with <100 samples were excluded from further analysis. Prior to calculation of gene signature scores derived from mouse genes (i.e., the top 25 upregulated genes from AAV-gTgfbr2 DEG, AAV-gOsmr DEG, and AAV-gIl1r1 DEG analysis), the mouse gene names were transferred to their human orthologs using the gprofiler2 (version 0.2.3) R package (Kolberg et al., 2019). Mouse genes without human orthologs or human orthologs not present in the gene-by-sample matrix of TCGA data set were dropped from further analysis. Then, to select genes that are predominantly specific to fibroblasts in the TME, we only included genes if they were most highly expressed in fibroblasts or stellate cells (resident fibroblasts of the pancreas) within at least one of The Curated Cancer Cell Atlas (Tyler et al., 2025) pancreatic cancer objects (Hwang et al., 2022; Lin et al., 2020; Moncada et al., 2020; Peng et al., 2019; Raghavan et al., 2021; Steele et al., 2020). Finally, to maintain gene signature specificity, genes that were shared between the three “AAV-gTgfbr2 up DEG”/“AAV-gOsmr up DEG”/“AAV-gIl1r1 up DEG” signatures were excluded. For the “stromal signature,” the identified genes by Combes et al. (2022) were used. For the “fibroblast TGFβ response signature,” the genes by Mariathasan et al. (2018) were used. For the “progenitor signature,” the top5 nonoverlapping marker genes of the progenitor/homeostatic fibroblast clusters c03-COL15A1 and c05-PI16 by Gao et al. (2024) were used. The individual genes for each signature are listed in Table S4.

Gene signature score calculation

Gene signature scores were calculated on log10(TPM+1) normalized data, and the expression of each gene was converted to percentile ranks across the samples. The final gene signature score of each sample was calculated as the average percentile rank from all genes belonging to one gene signature.

Survival analysis

For pan-TCGA survival analysis, the Cox proportional hazards model from the survival (version 3.8-3) R package was used to test overall survival differences between patients in the top and bottom quartiles based on gene signature scores “stromal” and “fibroblast TGFbeta response,” with age, sex, and cancer type as covariates. For survival analysis in the TCGA PAAD cohort, the same analysis was done on gene signatures after normalizing them to the “progenitor signature” because a high stromal signature is already associated with worse survival (Fig. S5 K). Kaplan–Meier survival curves were plotted using the survminer (version 0.5.0) R package. Forest plots of all analyzed cancer types were generated after meta-analysis using a random effects model using the meta (version 8.1-0) R package.

Identification of cell frequency correlation between fibroblasts and immune cells

To identify potential co-occurring cell populations of CAF subsets and captured immune cells in the scRNA-seq data set, the proportion of each cell subset within its corresponding major cell compartment (fibroblasts, MonoMacs, DCs, neutrophils, T, and NK cells) was calculated. The proportion of each cell subset was correlated in pairwise fashion with all other cell subsets across all 16 samples (4 replicates × 4 AAV groups) using the cor() function and “method = pearson” from the stats package. Statistical significance of correlation was measured using cor.mtest() function from the corrplot package (Wei and Simko, 2010), and results were plotted as a heatmap with samples arranged by hierarchical clustering.

Candidate identification of ligand–receptor interactions

The CellChat (Jin et al., 2021) (version 2.1.2) package was used to identify putative ligand–receptor interactions between two target cell populations. First, all cell subsets belonging to one experimental AAV group were subsetted from the full Seurat object. Then a CellChat object was created and processed using default parameters. Source and target cells were defined prior to plotting the top six inferred ligand–receptor interactions.

Transcription factor activity prediction with decoupleR

To infer osteoblast-related transcription factor activity in the CAF subsets, the decoupleR (version 2.10.0) package was used to run a univariate linear model (function “ulm”) with default parameters. All genes were used as input, and all mouse transcription factors were initially considered. For plotting osteoblast-related transcription factor activity (Zhu et al., 2024), the mean activity of each transcription factor per CAF subset was used.

Quantification and statistical analysis

Statistical analyses were done using R version 4.3.2. Details of individual experiments can be found in their respective figure legends.

Supplementary Material

Online supplemental material

Table S1 shows DEGs for each cell subtype identified within major cell types. Major cell types grouped by sheet. Table S2 shows DEGs of AAV-gRNA groups using pairwise pseudobulk differential gene expression analysis comparing control (AAV-gTrac) with experimental groups (AAV-gOsmr, AAV-gTgfbr2, or AAV-gIl1r1). Table S3 shows weighted gene output ordered by NMF factors from fibroblast object. Table S4 shows signature genes used for survival analysis. Table S5 shows DNA sequences of plasmids and gRNAs used in this study. Table S6 shows table of reagents used in this study.

Data availability

The raw scRNA-seq data are available under GEO accession number GSE315708. The processed scRNA-seq data generated for this study and presented in Fig. 3, 4, 5, 6 and Fig. S4 and S5 have been deposited in the Synapse.org data repository under Project SynID syn71058181. Companion code is deposited on Github (https://github.com/NickKuhn/Kuhn_AAV-gRNA_CAF). The generated plasmids have been deposited at Addgene (#249119-249131). All other data supporting the findings of this study are available from the corresponding author on reasonable request.