henhao
Sign in
Cancer BiologybioRxiv
Preprint

Invasion status stratifies the composition of the human pancreatic cancer perineural niche

Zainab HussainiD, Dig Vijay Kumar YarlagaddaiD, Qianzi Li, Nuray TezcaniD, Pavel StupakoviD, Golabahar Sadatrezaei, Richard J. WongiD, Joan MassaguéiD, Zeynep Cagla TarcaniD, Olca BaştürkiD, Sylvie DebordeiD, Christina S. LeslieiD, Mara H. ShermaniD

Abstract

The peripheral nervous system innervates the pancreatic ductal adenocarcinoma (PDAC) microenvironment, and perineural invasion (PNI), the invasion of cancer cells in and around nerves, correlates with metastatic burden and poor outcomes. Though PDAC innervation is near universal and the majority of PDAC patients harbor PNI, the cellular and molecular composition of the heterocellular perineural niche is largely unknown, obscuring functional significance. Here we provide a deeply phenotyped spatial and single-cell atlas of the human PDAC perineural niche, enabling a high-resolution comparison of invaded versus non-invaded nerve neighborhoods. This atlas leverages a novel vision-centric artificial intelligence model for imaging-based spatial transcriptomics, coupled with pathology-guided single-nucleus RNA-seq. These analyses revealed that invaded nerve neighborhoods harbor cancer cells of the classical subtype, together with myofibroblastic cancer-associated fibroblasts (CAFs) and lipid-associated macrophages. Non-invaded nerve neighborhoods, in contrast, harbor inflammatory CAFs and infiltration of B and T lymphocytes. These findings raise the possibility that PNI and not innervation itself is immune-suppressive in this setting, motivating functional studies. Statement of SignificanceHuman PDAC is innervated and almost invariably harbors PNI, but the composition and functions of perineural niches remain unclear. Here we provide a spatial and single-cell atlas of the human PDAC perineural niche, defining cancer and stromal cell states in the context of PNI and nominating this process as potentially immune-suppressive.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
1
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (6 Aug 2026) · CC BY 4.0

Xenium and single-nucleus profiling of 14 human pancreatic tumors shows invaded nerve neighborhoods enriched for classical cancer cells, myCAFs and lipid-handling macrophages but depleted of lymphocytes.

Perineural invasion occurs in roughly 70 percent of PDAC patients and tracks with recurrence and poor outcome, yet the cell types and states forming the human perineural niche remain poorly defined, and segmentation-based imaging spatial workflows misassign transcripts and drop low-count cells. Pairing a discrete-codebook spatial annotation model with matched snRNA-seq, the work maps nerve neighborhoods by invasion status at single-cell resolution.

Key results

  1. ShownInvaded nerve neighborhoods are enriched for classical malignant cells, myofibroblastic cancer-associated fibroblasts (myCAFs) and lipid-metabolic ALOX5AP+/OLR1+/MARCO+ macrophages, whereas non-invaded ones hold tumor-infiltrating lymphocytes (TILs), inflammatory CAFs and CD163-high macrophages.

    352 pathologist-annotated nerves (68 invaded, 284 non-invaded) with 250μm neighborhoods across 11 sections; per-nerve mixed model with patient random intercept plus permutation and bootstrap checks. myCAFs and T cells peaked at the boundary (0μm), B cells across 0-200μm; NN TIL fraction exceeded off-nerve tissue in all PNI-low and 3 out of 8 PNI-high samples. · Spatially resolved cellular composition shifts…; Manual nerve annotation and definition of nerve…; Paired within versus outside NN TIL fraction; Cell state enrichment forest at nerves

  2. ShownInvaded neighborhoods upregulate tumor epithelial and invasion genes, the neuroactive factors NMU and BDNF, and matrix remodeling genes, while losing canonical neurotrophic factors, myelination genes and inflammatory-CAF and quiescent stromal genes, suggesting nerve injury.

    Mann-Whitney U per nerve over 480 panel genes (68 vs 284 nerves), BH-corrected, with patient-level pseudobulk mixed model as primary statistic. Up: NMU (peaking 75μm and 125μm from boundary), BDNF, MMP1, MMP11, COL11A1, LAMC2, ITGB4, OLR1, MARCO. Down: GFRA1, NGFR, NRN1, GDNF, MPZ, PLP1, PMP22, IGF1, C7, GZMK, CD79B. · Uncovering gene expression profiles and spatial…; Gene program to nerve boundary distance heatmap; Perineural transcript-level differential…

  3. SuggestedNeighborhood codes define 16 spatial niches; invaded nerves share an LRRC15+ desmoplastic myCAF niche and are depleted of inflammatory-CAF stroma, splitting into mutually exclusive classical-tumor niches, one with a myCAF rim and BDNF/SEMA3A expression that encircles or enters the nerve sheath.

    Per-nerve fractions of the 16 niches, Ward-clustered within invaded (n=68) and non-invaded (n=284) strata: invaded nerves were depleted of iCAF reactive stroma and CD8-marginalized myCAF barrier niches, while non-invaded subsets favored B cell-rich lymphocyte aggregate and SPP1+/CD163+ TAM niches; effect sizes not reported in text. · Uncovering gene expression profiles and spatial…; Codebook interpretation and downstream cell…; Per nerve niche composition clustermap

  4. SuggestedsnRNA-seq shows no PNI-linked shift in cell abundance but PNI-high-associated programs: NPSR1, NSMF, GDF15 and AP-1 (JUN, ATF3) in malignant cells; AP-1, collagen, neurotrophic receptor (NTRK2, NGFR) and NFASC adhesion genes in CAFs, with TGF-β genes in two subsets; OLR1/MARCO lipid-processing genes in TAMs.

    108,698 nuclei from 11 samples (8 PNI-high, 3 PNI-low), 13 major types and 45 subtypes; Milo found no significantly differentially abundant neighborhoods. Cell-level Wilcoxon tests (adjusted p < 0.05, |log2 fold change| > 0.32); pySCENIC (100 runs) placed NTRK2 and NGFR in CAF FOS/FOSB regulons; immune-modulatory CAFs enriched for over 15 nerve-related pathways. · Single nucleus transcriptomic analyses of PDAC…; Differential expression gene analysis; Regulon inference; Differential abundance testing

  5. SuggestedLabel transfer and ligand-receptor inference identify nCAFs_APOD as likely neural fibroblasts and nominate PNI-high CAF-malignant-TAM axes centered on VEGFA-neuropilin, laminin-integrin adhesion and lipid-associated ligands converging on CAF LRP1.

    RCTD mapped over 80% of fibroblast-like Schwann cells to nCAFs_APOD. LIANA (CellPhoneDB method, permutation p < 0.05, DEG-filtered) ranked CAF-malignant pairs most numerous, with more CAF-to-malignant pairs in PNI-high; prioritized VEGFA-NRP1/NRP2, LAMB3-ITGA6/ITGB4 and APOE/LPL-LRP1, versus LGALS2-LAG3 (malignant to CD8+ T cell) in PNI-low. · Cross-platform integration and ligand-receptor…; Label transfer to Xenium; Ligand–receptor analysis (LIANA)

Methods, data and reproducibility

PNI-high and PNI-low resections were profiled by Xenium (480-gene panel) and Chromium Flex snRNA-seq. Nicheverse, a hierarchical vector quantized autoencoder with 256 cell and 16 neighborhood codes coupled by cross-attention, annotated Xenium cells; 250μm nerve neighborhoods were compared by invasion status and integrated with snRNA-seq (RCTD, LIANA).

  • Treatment-naive Whipple or distal pancreatectomy specimens from MSKCC (2020-2024): 13 PDAC and one adenosquamous carcinoma, PNI status assigned by GI pathologist review · human · n = 14 patients (11 PNI-high, 3 PNI-low)
  • Xenium In Situ sections (4 fresh-frozen OCT, 7 FFPE) profiled with a custom 480-gene tumor, microenvironment and nerve panel · human · n = 11 sections (8 PNI-high, 3 PNI-low); 4,395,459 cells
  • 10x Chromium Flex snRNA-seq of fresh-frozen tissue; 8 samples matched to Xenium (5 PNI-high, 3 PNI-low) plus 3 additional PNI-high · human · n = 11 samples (8 PNI-high, 3 PNI-low); 108,698 nuclei
  • Pathologist-annotated nerve polygons on Xenium morphology images, each expanded by 250μm to define nerve neighborhoods · human · n = 352 nerves (68 invaded, 284 non-invaded)
  • Published PDAC malignant and fibroblast programs (Hwang, Moffitt, Collisson, Raghavan et al.) used for hypergeometric subtype mapping
Methods & tools
Xenium In Situ (10x Genomics) with custom 480-gene panel, 10x Genomics Chromium Flex snRNA-seq, Nicheverse (hierarchical vector quantized autoencoder), CellSplat (3D Gaussian splatting transcript assignment), SoupX; CellBender, Seurat v5 with SCTransform v2; scDblFinder; Harmony, scanpy v1.10; squidpy v1.6, Milo (pertpy) differential abundance, inferCNV, pySCENIC (100-run consensus regulons), RCTD (spacexr), LIANA (CellPhoneDB method); gseapy GSEA
Code
https://github.com/digvijayky/nicheverse
Data & accessions
https://nicheverse.org/
Software & environment
Python v3.10 (scanpy v1.10, squidpy v1.6, shapely v2, PyTorch v2); R v4.3 (lme4 v1.1, Seurat v5, SCTransform v2); Xenium Analyzer 1.9.2.0, 2.0.1.0 or 3.2.1.2; GENCODE v38 (GRCh38); Nicheverse seed 49.
Assay conditions & compute
Nicheverse trained for 30 epochs on a single GPU with mini-batches of 2,048 cells; CellBender run on GPU; GPU model, runtime and memory not reported.
Reproducibility gaps
Code, weights, codebooks and notebooks to be deposited on Zenodo upon acceptance (GitHub mirror); Xenium data to GEO and snRNA-seq to dbGaP, with accessions not yet listed.
Limitations (6)
  • The targeted 480-gene panel lacks several identity markers, and nuclei-based transcript assignment can redistribute transcripts in dense tumor and nerve regions, so some annotations are uncertain, rare types underrepresented, and RCTD mapping of CAF and immune cells diffuse. · stated by the authors
  • Small cohort with PNI-status imbalance (imST: 8 PNI-high vs 3 PNI-low) and pronounced interpatient heterogeneity; findings need validation in larger multicenter cohorts enriched for high-PNI disease. · stated by the authors
  • Few nerves with severe PNI were captured, so niche shifts tied to the degree of nerve involvement cannot yet be linked to PNI severity grades; stratifying patients by severity scores is needed. · stated by the authors
  • Mechanisms remain untested: whether perineural immune exclusion precedes or follows invasion, whether lipid-associated TAMs drive PNI, the proposed AP-1 regulatory hub, and the TLS-like nature of lymphocyte aggregates all require functional validation. · stated by the authors
  • Nicheverse is presented as segmentation-free, yet preprocessing retains only molecules inside vendor nucleus polygons and the Discussion acknowledges nuclei-based segmentation; no head-to-head benchmark against standard segmentation-based annotation is reported. · noted in this summary
  • Cell-level Wilcoxon and GSEA contrasts treat nuclei as replicates with only 3 PNI-low patients; reporting is also internally inconsistent (PNI-high non-invaded nerves 167 vs 168; an excluded PNI-low invaded nerve despite all PNI-low nerves described as uninvaded). · noted in this summary

Who should read:Read this if you study perineural invasion, tumor-nerve-stroma crosstalk or immune exclusion in pancreatic cancer, or want a discrete-codebook method for annotating imaging-based spatial transcriptomics.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISWang '25Heij '20Guillot '20Yıldızhan '25Tan '22Kiemen '24
77% semantic matchCancer Biology
Peter L. Wang, Nicole A. Lester, Ella N. Perrault+35 more·bioRxiv·Aug 26, 2025