show Abstracthide AbstractSolid cancers like pancreatic cancer (PDAC) frequently exploit nerves for rapid dissemination. This neural invasion (NI) is an independent prognostic factor in PDAC, but insufficiently modelled in genetically-engineered mouse models (GEMM) of PDAC.Here, we systematically screened for human-like NI in Europe's largest repository of GEMM of PDAC comprising 295 different genotypes. This phenotype screen uncovered two GEMM of PDAC with human-like NI, which are both characterized by pancreas-specific overexpression of transforming-growth-factor-alpha (TGFa) and conditional depletion of p53. Mechanistically, cancer-cell-derived TGFa upregulated CCL2 secretion from sensory neurons, which induced hyperphosphorylation of the cytoskeletal protein paxillin via CCR4 on cancer cells. This activated the cancer migration machinery and filopodia formation toward neurons. Disrupting CCR4 or paxillin activity limited NI, and dampened tumor size and tumor innervation. In human PDAC, phospho-paxillin and TGFa-expression constituted strong prognostic factors. Therefore, TGFa-CCL2-CCR4-p-paxillin axis is a clinically actionable target for constraining NI and tumor progression in PDAC. Overall design: In the spatial transcriptome analyses with murine DRG, DRGs T8 to T12 were extracted bilaterally from TPAC (Ela-TGFa; Ptf1a-Cre; Trp53fl/fl;RelAfl/fl , n=6, 47-54 weeks old), KPC (Ptf1a-Cre; LSL-Kras+/G12D; Trp53+/fl , n=3, 20-21 weeks old) and age-similar tumor-free control mice (Ela-TGFa; Trp53fl/fl;RelAfl/fl , n=3, 48-50 weeks old for TPAC and LSL-Kras+/G12D; Trp53+/fl , n=1, 20 weeks old; Trp53+/fl , n=1, 21 weeks old and LSL-Kras+/G12D; n=1, 22 weeks old for KPC). Following the extraction, DRGs were briefly washed in ice-cold, fresh PBS (D5262, Sigma-Aldrich) and fixated in freshly prepared 4% PFA (15710, Electron Microscopy Sciences; diluted with PBS) for 24 hours at room temperature. The samples were consecutively dehydrated, paraffine embedded and sectioned. A whole transcriptome analysis was performed on NeuN rich regions within the DRG sections using the NanoString GeoMx® digital spatial profiler. 5µm thick unstained sections on slides were subjected to incubation for 35 minutes at 60°C. Afterwards, the slides were moved into Leica BOND Rx. Proteinase K (1ug/ml, for 5 minutes) and HIER2 (20 minutes) were used for RNA antigen retrieval. NanoString GeoMx Mouse Whole Transcriptome Atlas RNA probes were added for hybridization for 16 hours at 37°C. Following the stringent washes per manufacturer's instructions, slides were blocked with W buffer for 1 hour. The slides were then stained with AF647 anti-NeuN (1:100, ab190565, Abcam), AF594 anti-Cd68 (1:200, sc-20060, Santa Cruz) AF532 Syto83 (1:10, ThermoFisher) and AF488 anti-GFAP (1:200, NBP2-3318AF488, Novus) antibodies. After washing the slides twice with SSC buffer, slides were loaded to the DSP and scanned for the regions of interest (ROI) selection. Two NeuN rich region per mouse were selected as ROIs.For the human spatial transcriptome studies, two patients with pancreatic ductal adenocarcinoma were selected for transcriptome spatial analysis using the NanoString GeoMx Digital Spatial Profiler (DSP)/Human NGS Whole Transcriptome Atlas. Cases were selected based on the presence of perineural invasion in tissue sections. Each slide was stained with fluorescence-labeled antibodies to allow identification of tissue: pan-cytokeratin for epithelial cells and PGP9.5 for nerves. Tumor and Nerve areas of interest (AOI) were selected as follows: (i) NI Tumor: nerve-invading pancreatic (N=6); (ii) no NI Tumor: non-invading tumor cells (n=7); (iii) NI Nerve: tumor-invaded nerve (n=5); (iv) no NI Nerve: non-invaded nerve (n=4). AOIs were drawn manually around nerves and tumor cells. Following the ROI selection, UV-photocleaved oligonucleotide barcodes from each ROI were collected and prepared by the NanoString GeoMx protocol. Sequencing was performed on the NextSeq 2000. Raw basecell files were demultiplexed into fastsq files. NanoString's GeoMx NGS Pipeline v.2.0.0 was used to convert the fastsq files to Digital Count Conversion files. None of the 45 AOIs were below the warning sequencing saturation of 50%. Data was then normalized, and the downstream gene expression analyses were performed using the R package DESeq2.