2.1 Single nucleus cell profiling of osimertinib-sensitive and osimertinib-resistant LUAD tissues

The single-cell atlas of the LUAD tissues was characterized using nine surgical frozen samples comprising six osimertinib-sensitive LUAD samples and three osimertinib-resistant LUAD samples. Each sample was processed to isolate a single nucleus without prior selection of cell types, and then single-nucleus RNA sequencing (snRNA-seq) was performed using a 10× Genomics Chromium platform to generate RNA-seq data. After quality filtering, 12,061 high-quality cells with a median of 1006–1367 genes per cell were analyzed (Figure 1A, Table S1). These cells were further identified to be nine separate cell types, including tumor cells (22%), alveolar type 2 cells (21%), endothelial cells (19%), lymphoid endothelial cells (10%), fibroblasts (7%), epithelial cells (7%), aerocyte endothelial cells (6%), alveolar type 1 cells (6%), and macrophage (2%) (Figure 1B). The t-SNE plot also showed distinct clustering according to the tumor origin (Figure 1C), and the heatmap depicted the differentially expressed marker genes in nine clusters (Figure 1D).

The cell type composition and their proportions showed a high degree of variation between osimertinib-sensitive and osimertinib-resistant LUAD tissues. Immune cells mainly included macrophages (CD68⁺), nonimmune cells mainly included tumor cells (ROS1⁺), fibroblasts (COL1A2⁺), alveolar type 2 epithelial cells (SFTPC⁺), alveolar type 1 epithelial cells (AGER⁺), epithelial cells (EPCAM⁺), endothelial cells (CD31⁺), lymphoid endothelial cells (CCL21⁺), and aerocyte endothelial cells (SOSTDC1⁺) (Figure 1E–G). With Cell chat analysis, we established a global cell-to-cell communication network among most of the cell types, including CAFs, endothelial cells, epithelial cells, macrophages, tumor cells, etc (Figure S1A). The number of cell-to-cell interactions (ligand-receptor pairs) was significantly higher in the OR group compared with the OS group (383 and 128, respectively), whereas the strength of these interactions was almost similar (6.554 and 6.377, respectively) (Figure 1H). Moreover, the number of interactions between different cell populations, especially the CAF–tumor cell interactions, was significantly higher in osimertinib-resistant tissues compared with sensitive ones, including the CSF2-CSF2RA pair, CCL20-CCR6 pair, etc. (Figure 1I; Figure S1B, C). Collectively, these data suggest that CAF–tumor cell interactions might participate in the development of osimertinib resistance.

2.2 CAFs were correlated with osimertinib resistance in LUAD

Previous research has identified three types of CAFs, including inflammatory CAFs (iCAFs), myofibroblastic CAFs (myCAFs), and antigen-presenting CAFs (apCAFs). iCAFs are primarily distinguished by the release of inflammatory substances such as interleukin-6 (IL-6), leukemia inhibitory factor (LIF), and interleukin 1 (IL-1). MyCAFs are characterized by high actin alpha 2 (α-SMA), transforming growth factor-β (TGF-β), and extracellular matrix expression, while apCAFs commonly express major histocompatibility complex class II (MHC class II) and class II transactivator (CD74) molecules.^{23, 24} In this study, we performed a re-clustering of 812 fibroblasts and assigned each of these subclusters based on known markers (Figure S2). Consistent with previous research, we also identified myCAFs (COL1A2⁺αSMA⁺), apCAFs (COL1A2⁺MHCII⁺), and iCAFs (COL1A2⁺IL6⁺) (Figure 2A). Interestingly, iCAFs were significantly enriched in the osimertinib-resistant tissues compared with sensitive ones (Figure 2B). Moreover, immunofluorescence analysis also confirmed that the proportion of iCAFs was significantly higher in resistant tissues, indicating that CAFs assume an iCAF phenotype after osimertinib resistance. (Figure 2C, D).

To determine whether CAFs contribute to osimertinib resistance, they were isolated from osimertinib-sensitive (CAF^OS) or osimertinib-resistant (CAF^OR) LUAD tissues and cultured in vitro. Then, LUAD cell lines (PC-9 and HCC827) were treated with conditioned medium from CAF^OS (CAF^OS-CM) and CAF^OR (CAF^OR-CM). Interestingly, compared with CAF^OS-CM, CAF^OR-CM treatment significantly increased the viability and colony-formation capacity of LUAD cells after osimertinib exposure (Figure 2E–G). Furthermore, CAF^OR-CM dramatically abolished the osimertinib-induced cell cycle arrest, apoptosis, and senescence in LUAD cells (Figure 2H–J). Taken together, our data suggest that CAF^OR-CM confers osimertinib resistance in LUAD cells.

2.3 CSF2 was required for CAF^OR-induced osimertinib resistance in LUAD cells

To determine which “ligand-receptor pair” mediates the interaction between CAF^OR and LUAD cells, we conducted the human cytokine array to explore the cytokine profiles of CAF^OR-CM and CAF^OS-CM and identified higher levels of CSF2 in CAF^OR-CM (Figure 3A; Figure S3A, B). The enzyme-linked immunosorbent assay was further performed to measure the concentration of CSF2 in the culture medium of CAF^OR and CAF^OS. As expected, CAF^OR secreted much more CSF2 than CAF^OS, but primary tumor cells isolated from osimertinib-sensitive and osimertinib-resistant LUAD tissues showed no difference in CSF2 expression and secretion (Figure 3B; Figure S3C, D). To assess the clinical significance of CSF2, we detected the level of CSF2 in the plasma samples of osimertinib-sensitive and osimertinib-resistant LUAD patients and found that the CSF2 level was significantly correlated with the development of osimertinib resistance, showing about 10.32-fold higher in resistant patients than that in sensitive ones (Figure 3C; Figure S3E, F; Table S2). These data together show that CSF2 might be a potential monitoring biomarker for the acquired resistance to osimertinib in LUAD.

Then, by adding a neutralizing antibody against CSF2, we found that blocking CSF2 successfully abrogated CAF^OR-CM-induced increase in cell viability following osimertinib exposure in a dose-dependent manner (Figure 3D, E). In addition, the anti-CSF2 antibody also significantly rescued the influences of CAF^OR-CM on colony formation, cell cycle, apoptosis, and senescence in LUAD cells under osimertinib treatment (Figure 3F–I). Collectively, these results indicate that CSF2 plays a critical role in CAF^OR-induced osimertinib resistance in LUAD cells.

2.4 CAF^OR-derived CSF2 facilitated osimertinib resistance by promoting the expression of lnc-CSRNP3 in LUAD cells

To molecularly dissect how CAF^OR-derived CSF2 induces osimertinib resistance, whole-transcriptome profiling of LUAD cells was compared between hIgG1-Fc (control group) and recombinant CSF2-Fc (rCSF2 group) treatment. This analysis revealed that both protein-coding mRNAs and lncRNAs were differently expressed, and the differences were more apparent in terms of lncRNAs than mRNAs (Figure 4A; Figure S4). Of the 112 common differentially expressed lncRNAs in two cell lines (Figure 4B), we selected the top five upregulated lncRNAs after CSF2 exposure for further validation by qRT-PCR, and the results showed that only lnc-CSRNP3-6:1 (briefly renamed as lnc-CSRNP3) showed a marked increase in both cell lines (Figure 4C). Moreover, like CSF2, CAF^OR-CM also increased the expression level of lnc-CSRNP3 in LUAD cells, which was attenuated by the anti-CSF2 neutralizing antibody (Figure 4D). More importantly, CSF2 exposure significantly increased the IC₅₀ values of osimertinib in LUAD cells, which could be dramatically reversed by lnc-CSRNP3 knockout (Figure 4E). These data indicate that CAF^OR-derived CSF2 facilitates osimertinib resistance by promoting the expression of the lnc-CSRNP3 in LUAD cells.

Next, we established a xenograft model of PC-9 and CAF^OR co-culture cells by subcutaneous injection in nude mice (Figure 4F). The anti-CSF2 antibody or lnc-CSRNP3 knockout significantly reduced the CSF2 content in mouse plasma and the lnc-CSRNP3 expression in xenograft tissues, as well as markedly decreased the tumor volume and tumor weight (Figure 4G–L). In addition, the characteristics of xenograft tissues were also studied. The H&E staining demonstrated that the xenograft tumors exhibited both the increased size and abnormal arrangement of nuclei. Immunohistochemical staining exhibited that both anti-CSF2 antibody and lnc-CSRNP3 knockout reduced the expression of proliferation marker Ki-67 while increasing the levels of cell cycle inhibitor P27^KIP1 and cell apoptosis marker cle-caspase3 (Figure 4M). Taken together, these data confirm the promoting roles of both CAF^OR-derived CSF2 and lnc-CSRNP3 in osimertinib resistance in vivo.

2.5 CAF^OR-derived CSF2 upregulates lnc-CSRNP3 through activating the JAK2/STAT3 pathway

To further reveal the mechanism by which CAF^OR-derived CSF2 upregulates lnc-CSRNP3 in LUAD cells, the KEGG pathway analysis was performed using differentially expressed mRNAs between the control and rCSF2 groups. Both the enrichment score and gene numbers indicated that the JAK2/STAT3 signaling pathway was enriched after CSF2 treatment (Figure 5A). Besides, GSEA analysis also showed these differentially expressed mRNAs were enriched in the JAK2/STAT3 signaling pathway (Figure S5A). Simultaneously, we analyzed the promoter region of lnc-CSRNP3 and found three potential binding sites of STAT3 (Figure 5B). Treatment with CAF^OR-CM or CSF2 successfully activated the JAK2/STAT3 pathway in LUAD cells, which could be rescued by anti-CSF2 antibodies (Figure 5C; Figure S5B). Additionally, the phosphorylated STAT3 was upregulated in osimertinib-resistant LUAD tissues, but not in sensitive ones (Figure 5D). These results indicate that CAF^OR-derived CSF2 could activate the JAK2/STAT3 pathway in LUAD cells.

Then, we investigated whether STAT3 could directly bind to the promoter of lnc-CSRNP3, and upregulate its expression. Luciferase reporter assays showed that CSF2 treatment significantly enhanced the transcriptional activity of the lnc-CSRNP3 promoter in PC-9 cells, which could be abrogated by STAT3 knockdown (Figure 5E). The chromatin immunoprecipitation (ChIP) assay further revealed that STAT3 directly bound to the fragment P1 of lnc-CSRNP3 promoter (Figure 5F), and only the STAT3-fragment P1 interaction could promote the transcription of lnc-CSRNP3 in CSF2-treated PC-9 cells (Figure 5G). STAT3 inhibitor, Stattic, significantly reversed CSF2-induced lnc-CSRNP3 upregulation (Figure 5H).²⁵ Altogether, these findings support that CAF^OR-derived CSF2 upregulates lnc-CSRNP3 in LUAD cells by activating JAK2/STAT3 pathway.

2.6 Lnc-CSRNP3 promotes osimertinib resistance by cis-regulating nearby gene CSRNP3

To elucidate the potential molecular mechanisms through which lnc-CSRNP3 promotes osimertinib resistance, we first observed its subcellular localization in LUAD cells, because the functional mechanisms of lncRNAs are dependent on their subcellular distribution.²⁶ Our results showed that lnc-CSRNP3 was prevailingly localized in the nuclear compartment (Figure 6A, B). Given that regulating the expression of nearby genes is one of the important ways in which the nuclear lncRNAs exert their biological effects,²⁷ we observed the chromatin localization of lnc-CSRNP3 and found that it was in the vicinity of three protein-coding genes, UDP-GalNAc transferase 3 (GALNT3), cysteine/serine-rich nuclear protein 3 (CSRNP3) and tetratricopeptide repeat protein 21B (TTC21B) (Figure 6C). Among them, only the level of CSRNP3 could be significantly modulated by lnc-CSRNP3 in LUAD cells (Figure 6D; Figure S6). Moreover, the results of the CCK-8 assay showed that lnc-CSRNP3-induced decrease in osimertinib sensitivity could be rescued by CSRNP3 knockdown (Figure 6E), indicating that lnc-CSRNP3 promotes osimertinib resistance by upregulating CSRNP3 expression.

To investigate the process by which lnc-CSRNP3 controls the expression of CSRNP3, we utilized a biotin-labeled RNA pulldown experiment, followed by mass spectrometric analysis, to identify the proteins that potentially interact with lnc-CSRNP3. (Figure 6F). Our results revealed that chromodomain helicase DNA binding protein 9 (CHD9) was the highest-ranked interactor. Then, the interaction of lnc-CSRNP3 with CHD9 was confirmed via biotin-labeled RNA pull-down along with western blotting assays and anti-CHD9 RIP assays (Figure 6G, H). According to previous studies, CHD9 exerts its biological effects by directly binding to the promoter of downstream genes and acts as transcriptional coactivator or chromatin remodeling factor.²⁸ Remarkably, we have discovered three CHD9 binding sites in the promoter region of CSRNP3 using the JASPAR database. (Figure 6I). Then we performed dual luciferase assay and anti-CHD9 ChIP assay, confirming that CHD9 bound directly to the BS1 and BS2 fragments of CSRNP3 promoter and promoted its transcription (Figure 6J, K). Taken together, these observations suggest that lnc-CSRNP3 enhances CSRNP3 expression through directly interacting with CHD9.

2.7 CSRNP3 promotes ribosome biosynthesis by directly binding to PP1α

To further explore the downstream molecules regulated by CSRNP3, we used mass spectrometry to identify its potential interacting proteins in LUAD cells (Figure 7A). As shown in Figure 7B, a total of 100 proteins were screened out. Among them, 51 proteins were co-located in the nucleus with CSRNP3, including the catalytic subunit of Ser/Thr phosphatase-1 (PP1) (Figure S7A, B, Table S3). A previous study has shown that CSRNP3 contains an RVxF motif, which binds to the backside of PP1.²⁹ In addition, using the BioGRAD database (https://thebiogrid.org/), we found that there might exist a protein–protein interaction (PPI) between CSRNP3 and a PP1 catalytic subunit, PP1α (Figure 7C). Therefore, we performed proximity ligation assay (PLA) and co-immunoprecipitation (Co-IP) assays to investigate their interaction, and found that CSRNP3 indeed directly interacted with PP1α through its RVxF motif (Figure 7D–G).

PP1α, which belongs to the phosphoprotein phosphatase family, is a ubiquitous eukaryotic enzyme that dephosphorylates a broad range of phosphoproteins,³⁰ and its interacting proteins could change its expression, subcellular distribution or directly inhibit its enzyme activity.³¹ Consistent with previous results, we found that the binding of PP1α by CSRNP3 inhibited PP1α activity, but had no effect on its expression or subcellular distribution in NSCLC cells (Figure 7H; Figure S7C, D). As a phosphatase, PP1α has many nuclear substrates such as phosphorylated mouse double minute 2, myocyte enhancer factor 2A, retinoblastoma (p-Rb), and cAMP-response element binding protein, playing a pivotal role in tumor progression.²⁹ To determine the downstream molecules of the CSRNP3/PP1α axis, we performed a western blotting assay and found that the protein levels of p-Rb were dramatically upregulated after CSRNP3 overexpression in LUAD cells, while were downregulated after CSRNP3 knockdown, which could be abolished by PP1α inhibitor, calyculin A (Figure 7I).³² According to previous research, p-Rb was involved in ribosome biosynthesis by promoting ribosomal RNA (rRNA) transcription.^33-35 By using qRT-PCR, we found that the levels of 5S, 18S, 28S, and 45S rRNA transcripts were markedly increased after CSRNP3 overexpression in PC-9 cells, but were decreased after CSRNP3 knockdown, which could be rescued by calyculin A (Figure 7J). Puromycin incorporation assays demonstrated that CSRNP3 could promote protein synthesis in PC-9 cells (Figure 7K). Furthermore, immunofluorescence of fibrillarin (FBL), which is participated in pre-rRNA processing and used as a ribosome biosynthesis marker,³⁶ showed that CSRNP3 could significantly enhance the ribosome biosynthesis (Figure 7L). Altogether, these data demonstrate that CSRNP3 can promote ribosome biosynthesis through directly binding to PP1α.

2.8 Inhibition of CSF2 pathway overcomes osimertinib resistance in LUAD PDX mice model

Finally, we sought to validate the function of the CSF2 pathway in vivo and intended to find potential therapeutic strategies for osimertinib resistance in LUAD. For these purposes, we generated osimertinib-resistant PDX models (PDX^OR) (Figure S8) and observed the therapeutic efficacy of anti-CSF2 neutralizing antibody and ribosome biogenesis inhibitor CX5461 against osimertinib resistance (Figure 8A). Our findings indicate that by targeting CSF2 or ribosome biogenesis, it is possible to drastically reduce both the volume and weight of tumors. However, mice in various groups had similar average body weights (Figure 8B–E). In addition, both anti-CSF2 and CX5461 decreased the expression of Ki-67 but increased the level of P27^KIP1 and cle-caspase3 (Figure 8F). According to previous reports, a greater number of AgNOR dots per nucleus and a proportion of NOR-occupied nuclear area indicate more active ribosome biogenesis.³⁰ Our results showed that both the AgNOR dots number per nucleus and NOR sliver-staining area were significantly decreased after targeting lnc-CSRNP3 or ribosome biogenesis (Figure 8G). These results suggest that targeting the CSF2 pathway can efficiently overcome the resistance to osimertinib in vivo.

4.1 Cell lines

HEK293 cells and human LUAD cell lines PC-9 and HCC827(EGFR19del) were obtained from the National Collection of Authenticated Cell Culture of China. The cells were grown in RPMI 1640 medium (Gibco) with 10% fetal bovine serum at 37°C in a humidified environment with 5% CO₂ and 95% air. All cell lines were mycoplasma-screened and verified by STR profiling.

4.2 Ethics statement

This study was approved by the Ethics Committee Board of Chongqing Medical University (no. 2017009) and was registered with the Chinese Clinical Trial Registry on January 27, 2018 (no. ChiCTR1800014660). A total of 30 LUAD patients who underwent first-line osimertinib therapy (15 osimertinib-sensitive patients and 15 acquired osimertinib-resistant patients) were enrolled. Written informed consent was obtained from all of them. Details of patients are listed in Table S4.

4.3 RNA extraction and quantitative real-time PCR

Total RNA was extracted from whole-cell lysate using TRIzol. The miRNeasy Serum/Plasma Kit (QIAGEN) extracted total RNA from clinical plasma samples. A PARIS kit (Thermo-life) separated nuclear and cytoplasmic RNA, and the percentages in fractions were calculated using the formula: 2^Ct(nuclear)/2^{Ct (nuclear)} + 2^{Ct (cytoplasm)} or 2^{Ct(cytoplasm)}/ 2^{Ct (nuclear)} + 2^{Ct (cytoplasm)}. Then, RNA was reverse transcribed into cDNA using the PrimeScript RT Reagent Kit (Takara Bio). qRT-PCR was performed using SYBR Premix Ex Taq II (Takara Bio) according to the manufacturer's instructions. The primer sequences used in this study are shown in Table S5.

4.4 Proteins purification and western blotting assay

Total protein was extracted and quantified with the Pierce BCA protein assay kit (Thermo Fisher Scientific). Transferring proteins to PVDF membranes following 10% SDS-PAGE. After 1 h of blocking with 5% BSA, the membrane was incubated with primary antibodies overnight at 4°C (Table S6). After 1 h of incubation with HRP-conjugated secondary antibody, immunoreactive bands were identified using ECLTM Prime (GE Healthcare) and LAS-3000 imager.

4.5 Lentivirus transfection and stable cell line construction

Lentiviruses overexpressing lnc-CSRNP3 or CSRNP3 and the CRISPR/Cas9 system targeting it were transfected into LUAD cells for 48 h with 5 mg/mL polybrene. Stable cell clones were selected for 1 week with puromycin (5 µg/mL). The overexpression or knockdown efficiency was measured by qRT-PCR. The sequences of siRNA, shRNA, and gRNA used are provided in Table S7.

4.6 ChIP assay

Following manufacturer instructions, Pierce Magnetic ChIP kit (Thermo Fisher Scientific) was used for the ChIP experiment. Approximately 4 × 10⁶ cells were collected and crosslinked in 1% formaldehyde. Next, MNase (ChIP grade) and sonication produced 200−1000 bp DNA fragments. The antibody-target protein-DNA complex was formed by adding anti-CHD9 or anti-STAT3 (Abcam) and immunoprecipitating it with Protein A/G magnetic beads. Washing and reversing cross-links purified the concentrated DNA for qRT-PCR analysis. The primer sequences are listed in Table S8.

4.7 Statistical analysis

The statistical analyses were done using SPSS 20.0 (IBM) and GraphPad Prism 6.0. Data were validated in three different experiments and presented as means ± SD, with comparable variance among groups. A one-way ANOVA and two-tailed Student's t-test were used to compare data sets. A p-value < 0.05 indicated statistical significance.