Graphical abstract

   graphic file with name fx1.jpg
   [89]Open in a new tab

Highlights

     * •
       Single-cell analysis uncovers immune ecosystems in early-stage LUAD
     * •
       Immune detection of GGO-associated neoantigens in lung cancer
     * •
       Activated CXCL13^+CD8^+ T cells recognize GGO-derived neoantigens
       at the early stage
     * •
       Immune modulation of lymphatic endothelial cells conditions LUAD
       progression
     __________________________________________________________________

   The immune microenvironment in early-stage lung cancer remains elusive.
   Deng et al. unveil six immune ecotypes in early-stage lung
   adenocarcinoma, showing early GGO neoantigen detection by CD8^+ T cells
   versus solid nodule suppression, highlighting paths for early
   therapeutic interventions.

Introduction

   Lung adenocarcinoma (LUAD), the predominant subtype of lung cancer,
   boasts a meager 5-year survival rate of 50% for patients in advanced
   stages despite contemporary first-line therapies.[90]^1 The advent of
   low-dose computed tomography (CT) has dramatically improved survival
   rates by enabling early detection, particularly in the form of
   ground-glass opacity (GGO). Indeed, a retrospective study of stage I
   LUAD patients who underwent surgery underlines the significance of
   early detection through low-dose CT, revealing an impressive 5-year
   overall survival rate of 94.9%.[91]^2

   GGOs, evident on chest CT, manifest as hazy opacity without obscuring
   the underlying bronchi and pulmonary vascular structures.[92]^3 While
   GGOs typically exhibit low malignant potential, they can persist for
   years,[93]^4 with nodular GGOs suggesting early-stage malignancy if
   they enlarge or develop a solid component over time.[94]^5^,[95]^6 In
   contrast, patients may present with a complete solid pattern initially,
   indicating distinct malignant progression routes. Whether the
   manifestation of early development of lung malignancy can originate
   from different patterns and the underlying biological mechanisms that
   would be involved if this were true remain unclear. The histological
   examination of early-stage LUAD with GGO features reveals thickened
   alveolar walls with atypical cuboidal pneumocytes, observed in
   preinvasive lesions such as atypical adenomatous hyperplasia,
   adenocarcinoma in situ, minimally invasive adenocarcinoma, or
   lepidic-predominant invasive adenocarcinoma (LPA),[96]^7 showing almost
   100% 5-year disease-free survival after surgery.[97]^8 Notably, GGOs
   associated with LPA pose a higher risk of metastasis and disease
   progression.[98]^9

   These histological features of LUAD do not consistently correlate with
   mutational profiling, complicating our understanding of GGO
   progression. EGFR mutations are more frequent in smaller peripheral
   adenocarcinoma with a diameter less than 3 cm and GGO ratio over
   50%.[99]^10 However, no significant correlation exists for KRAS
   mutations or ALK rearrangements.[100]^11^,[101]^12^,[102]^13 As a
   result, researchers have focused on understanding the biological
   characteristics of the microenvironment within GGOs. Recently, the
   advent of single-cell RNA sequencing (scRNA-seq) has enabled several
   studies on the early stages of lung malignancies.[103]^14^,[104]^15 For
   example, one study revealed metabolic reprogramming exhibited by
   malignant cells in subsolid GGO lesions and dominant enrichment of
   natural killer T (NKT) cells in the subsolid GGO
   microenvironment.[105]^14 Another study identified an early-stage lung
   cancer activation module (LCAM) by using scRNA-seq data of 35 LUAD
   lesions.[106]^15 LCAM-high lesions contain high abundance of
   PD1^+CXCL13^+ T cells, IgG^+ plasma cells, and SPP1^+ macrophages and
   respond better to anti-PD1 immunotherapy independent of the tumor
   mutation burden (TMB).

   Although such studies provided exclusive insights into clinically
   relevant cellular heterogeneity associated with early-stage LUAD, we
   are still far from understanding the co-operations among subpopulations
   of cells within the malignancies. Growing evidence shows that malignant
   cells and their neighbors behave as communities, and increasing
   attention is now being directed toward the behavior of diverse cell
   types as an ecosystem.[107]^16 Here, we integrated EcoTyper, a machine
   learning framework, to delineate a high-resolution atlas of
   multicellular ecotypes depicting the progression of LUAD from GGO
   toward advanced stages. We generated and analyzed both scRNA-seq and
   whole-exome sequencing (WES) data from 58 patient samples and
   identified 42 cell states consisting of malignant cells and other
   associated cell types within the GGO microenvironment. We uncovered six
   multicellular ecotypes that extend beyond previous knowledge of
   GGO-associated malignancies, in which early detection of
   tumor-associated neoantigens represented a main barrier for the
   malignant progression. We further showed that computationally predicted
   EGFR-mutant-related neoantigens could be detected in an additional
   cohort by using human leukocyte antigen (HLA) immunopeptidome and the
   corresponding peptides could potentially induce a T cell-specific
   reactivity. These six ecotypes thus may participate in the progression
   of early-stage LUAD in a cross-community-dependent manner and have
   significant potential for therapeutic interventions.

Results

Discovery of cell states and multicellular ecotype portraits of GGO and
advanced LUAD

   To reveal the multicellular ecotypes and distinct cell states present
   within both early-stage GGO lesions and advanced LUADs, we utilized
   EcoTyper, a unified chassis for scRNA-seq datasets.[108]^17 Our
   analysis was based on scRNA-seq data obtained from 55 surgical biopsies
   of patients ([109]Figure 1A; [110]Table S1).

Figure 1.

   [111]Figure 1
   [112]Open in a new tab

   Characterization of GGO and solid lung adenocarcinoma cell states and
   LMEs

   (A) Schematic depicting the analytical framework and stratifications of
   patients based on the radiological patterns. Yellow dashed lines
   highlight the malignant lesions. Scale bars, 1 cm (identical for all
   the CT images).

   (B) Overview of clinical characteristics of primary untreated
   individuals with lung adenocarcinoma.

   (C) The t-SNE (t-distributed stochastic neighbor embedding) by major
   cell partitions (left), tissue types (middle), or specimens from
   different patients (right).

   (D) Lung malignant ecotypes detected in untreated individuals with lung
   adenocarcinoma. Top: LME compositions are depicted as network diagrams.
   The width of each edge represents the Jaccard index across tumor
   samples. Bottom: cell state compositions in the corresponding LMEs.

   (E) Characteristics of LMEs in the discovery cohort. Top: proportions
   of major cell types (averaged and scaled). Bottom: mutation frequencies
   of driver genes.

   (F) The fraction of LMEs in different radiological clinical stages.
   Endo, endothelium; Epi, epithelium; Fib, fibroblast; NK, natural killer
   cell; DC, dendritic cell; PMN, polymorphonuclear cell (mainly
   neutrophil in our datasets); Mo/Mφ, monocyte and macrophage.

   We conducted droplet-based scRNA-seq using 10X Genomics on a total of
   35 GGO samples from untreated patients and 14 solid nodule LUAD samples
   from treatment-naive patients ([113]Figure S1A; [114]Table S1). The GGO
   samples were categorized into six different patterns based on their
   radiological features, including pure GGO (pGGO, n = 6, Hounsfield unit
   [HU] < −500), high-density GGO (dGGO, n = 6, with a density measurement
   HU between −500 and −300) ([115]Figures 1A and [116]S1B), GGO with 25%
   solid component (GGO25, n = 4), GGO with 50% solid component (GGO50,
   n = 6), GGO with 75% solid component (GGO75, n = 7), and GGO with 100%
   solid component (GGO100, n = 6). The solid nodule subgroup was further
   divided based on size and lymph node metastasis, including 1–2 cm solid
   nodule (Solid1, n = 4), 3–5 cm solid nodule (Solid3, n = 3), and solid
   nodule with regional lymph node metastases (SolidN, n = 7). Among these
   samples, 47 were also subjected to WES analysis ([117]Figure 1B). This
   classification is clinically relevant and covers most of the
   radiological heterogeneity observed in outpatients ([118]Figure 1B). We
   also included 6 normal lung tissue samples from patients with benign
   lung diseases (pulmonary bullae or sequestration) to ensure the
   specificity of our analyses ([119]Table S1). We pooled all cells
   together and performed batch correction ([120]Figure S1C) and corrected
   for read depth and mitochondrial read counts ([121]Table S2). Our
   scRNA-seq analysis yielded 33,336 unique transcripts from a total of
   462,784 single cells ([122]Figure 1C). Of these, 294,507 cells (63.64%)
   were from GGO samples, 127,204 cells (27.49%) were from solid nodules,
   and the remaining 41,073 cells (8.88%) were from the normal samples
   ([123]Figures 1C and [124]S1D). Cells were then subjected to
   non-supervised cell clustering using Seurat,[125]^18 which led to the
   identification of 11 major cell subtypes, including epithelial cells,
   endothelial cells, fibroblasts, B cells, CD4^+ lymphocytes, CD8^+
   lymphocytes, natural killer (NK) cells, dendritic cells (DCs), mast
   cells, polymorphonuclear leukocytes (PMNs), and monocytes/macrophages
   ([126]Figures 1C and [127]S1D; [128]Table S2). There were no
   significant differences in terms of cell subtype distributions among
   radiologically classified samples, although a trend of lower myeloid
   cell compartment was observed in advanced solid nodules
   ([129]Figures S1E and S1F).

   To elucidate the cell states of each cell type in patient-derived
   samples, we applied non-negative matrix factorization to the gene
   expression profiles of 11 major cell types. This allowed us to
   reconstruct a weighted combination of discrete transcriptional matrices
   that represent specific cell states within each cell type (see
   [130]STAR Methods). We chose the number of cell states per cell type to
   maximize sensitivity of cell state discovery and ensure the stability
   of clustering results ([131]Figure S2A). Using this approach, we
   identified 42 distinct cellular states among different GGOs and
   advanced solid nodular samples ([132]Table S2). These included (1)
   three B cell states; (2) five CD4^+ lymphocyte states, which covered a
   TIGIT-expressing regulatory T cell state, a naive CD4 T cell state, a
   CXCL13-expressing CD4 T cell state, and a helper T cell state; (3) four
   CD8^+ lymphocyte states, including a GZMB^+ effector CD8 T cell state,
   a PD1^+CXCL13^+ exhausted CD8^+ T cell state, and a transitional CD8
   T cell state; (4) two major DC states, including a CCR7-expressing DC
   and a CD1A^+ conventional DC state; (5) five endothelial cell states,
   including a tumor-educated endothelial cell state, a tip-like
   endothelial cell state, a stalk-like endothelial cell state, a
   scavenging endothelial cell state, a lymphatic endothelial cell state,
   an artery endothelial cell state, and an alveolar type I endothelial
   cell state; (6) three epithelial cell states representing two
   alveolar-derived adenocarcinoma cell states; (7) four fibroblast cell
   states, including pericytes, collagen-expressing fibroblast cell
   states, and myofibroblasts; (8) five monocyte/macrophage cell states,
   including a SPP1^+ macrophage cell state, tissue-resident macrophages,
   monocytes, early-state macrophages, and DC-like monocytes; (9) four NK
   cell states, including an NKT cell state and a toxic XCL1^+ NK cell
   state[133]^19; (10) four mast cell states; and (11) three major PMN
   cell states ([134]Figure S2B). Ecotype-determined cell states have been
   also validated in an independent in-house cohort comprising 76 samples
   sequenced by bulk RNA sequencing (RNA-seq) ([135]Figure S2C).

   We then clustered these cell states into communities by maximizing the
   co-association patterns,[136]^17 revealing six lung multicellular
   ecotypes (LMEs) ([137]Figure 1D). The LMEs varied considerably in their
   numbers of elemental cell states. For example, LME03 contained the most
   heterogeneous cell states among all the ecotypes, in which epithelial
   cells harbored mostly EGFR mutations, histone demethylase KMT2D
   mutations, and tumor suppressor TP53 mutations ([138]Figure 1E). We
   observed a significant increase in LME01, LME02, and LME04, while a
   decrease in LME03 and LME05 accompanying the progression from normal
   lung tissue to advanced solid nodule ([139]Figure 1F) and similar
   trends of LME changes were also observed in our bulk RNA-seq dataset
   ([140]Figures S2D and S2E).

   Overall, our data revealed extensive multicellular ecotypes in GGO and
   advanced solid nodules, which may potentially indicate different stages
   of LUAD development.

Malignant cell states in GGO and advanced LUAD

   GGOs can remain dormant for decades but eventually progress toward
   advanced solid nodules as the proportion of the solid part
   increases.[141]^20^,[142]^21 However, in some cases, patients are
   diagnosed with pure solid nodule without any preceding GGO pattern. The
   differential molecular features of these radiological patterns and
   their respective cellular progression routes remain unknown. To gain
   insight into their underlying molecular characteristics, we first
   focused on the malignant cell population in our patient cohort.

   We identified malignant cells by inferring large-scale copy-number
   variations (CNVs) using their normal immune, endothelial, and
   fibroblast cells as references ([143]Figures 2A and [144]S3A). To