Comprehensive characterization of the inflammatory ecosystems in immunotherapy-induced adverse events versus chronic inflammatory diseases

Abstract

Background

Immune-related adverse events (irAEs) induced by immune checkpoint inhibitors (ICIs) share symptomatic and therapeutic similarities with chronic inflammatory diseases (CIDs). However, the molecular distinctions between irAEs and CIDs remain unclear. Herein, we systematically compared irAEs and CIDs across multiple tissues using large-scale multi-omics profiling.

Methods

We compiled a transcriptomic compendium of over 4.3 million cells from 1137 samples and 24 spatial transcriptomes across diverse inflammatory sites in irAEs and CIDs. In addition, we collected 526 pre-ICI and post-ICI treatment-matched blood transcriptomes coupled with germline exomes, alongside 75 serum proteomes with irAE information.

Results

Hallmark gene signatures and cell states that characterize the inflammatory milieu of various irAEs and CIDs were defined. Across tissues, irAEs involve a distinct inflammatory ecosystem enriched with myeloid-derived components, including NLRP3⁺ mononuclear phagocytes, LAMP3⁺ dendritic cells, and CXCL8/10 cytokine signaling, alongside TNFRSF18⁺ regulatory T cell, T/natural killer interferon, and HAVCR2⁺ exhausted T cell. These cell states exhibit unique interaction patterns, forming spatially localized niches specific to irAEs but absent in CIDs. Severe irAEs are accompanied by elevated CXCL10 cytokine signaling post-ICI treatment, with a higher prevalence among HLA-B*46:01 carriers.

Conclusions

Our multimodal analyses highlight key differences between irAEs and CIDs, indicating that irAEs constitute a distinct inflammatory ecosystem that differentiates them from CIDs. Our study provides insights into potential biomarkers and therapeutic targets for improved irAE management.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Immune checkpoint inhibitors (ICIs) can induce immune-related adverse events (irAEs) that clinically resemble chronic inflammatory diseases (CIDs) across multiple organs.

• Current management guidelines for irAEs largely parallel those for CIDs, advocating corticosteroids as first-line therapy and biologics for corticosteroid-refractory cases.

• Despite these clinical and therapeutic similarities, the molecular mechanisms that distinguish irAEs from CIDs across affected organs remain poorly understood.

WHAT THIS STUDY ADDS

• This study provides the first multidimensional atlas of irAEs and CIDs, integrating transcriptomic data from 4.3 million single cells across 1,137 samples and 24 spatial transcriptomes from diverse inflammatory sites. This compendium also includes matched pre-ICI and post-ICI blood transcriptomes, germline exomes, and clinical data from 263 patients, as well as serum proteomes from 75 individuals with irAE information.

• We identified 92 distinct immune cell states shaping the inflammatory landscape and discovered an irAE-specific ecosystem characterized by NLRP3⁺ mononuclear phagocytes, LAMP3⁺ dendritic cells, CXCL10 cytokine signaling, TNFRSF18⁺ regulatory T cells, T/natural killer interferon, and HAVCR2⁺ exhausted T-cell states within spatially localized niches absent in CIDs. We further uncovered immune cell crosstalk mediated by CXCL9/10/11 and CCL3 pathways that further differentiates irAEs from CIDs.

• Severe irAEs were linked to marked upregulation of CXCL10 cytokine signaling following ICI therapy and a higher prevalence of the HLA-B*46:01 carriers, suggesting a genetic predisposition to exacerbated immune responses.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• This comprehensive compendium delineates the molecular differences between irAEs and CIDs, offering potential biomarkers and therapeutic targets for preventing and managing irAEs.

• Targeting the CXCL9/10/11 signaling axis may represent a promising strategy for mitigating irAE severity and improving patient outcomes.

• For patients carrying HLA-B*46:01, personalized therapeutic approaches, including ICI dose adjustments, may help minimize the risk of severe irAEs.

Background

Cancer immunotherapy based on immune checkpoint inhibitors (ICI) has transformed the treatment landscape for various malignancies, delivering unprecedented clinical outcomes and resulting in numerous Food and Drug Administration approvals across multiple cancer types.^{1 2} Despite these remarkable successes and their potent antitumor efficacy, ICI therapy can trigger immune-related adverse events (irAEs), which may cause significant morbidity and often necessitate treatment discontinuation.^{3 4} irAEs occur in 70–90% of patients receiving ICIs and can affect any organ system,^{3 5} occasionally leading to life-threatening complications such as cardiac, pulmonary, or neurologic toxicities,^{6 7} thereby presenting substantial clinical challenges.

Notably, irAEs exhibit phenotypic similarities to chronic inflammatory diseases (CIDs) across various affected organs.8,10 Moreover, the treatment approaches for irAEs closely resemble those for CIDs, as irAEs often respond to systemic corticosteroids. Current guidelines also recommend the use of biologics, such as tumor necrosis factor (TNF)-α inhibitors and interleukin (IL)-6 inhibitors, for cases resistant to corticosteroids.11,13 While acute toxicities are more common, up to 40% of patients with irAEs may develop chronic inflammatory conditions that persist even after ICI discontinuation,^{14 15} evolving into chronic irAEs and further strengthening the link between irAEs and CIDs.

Although irAEs and CIDs share symptomatic and therapeutic similarities, irAEs typically have an abrupt onset,¹⁶ can occur in the absence of autoantibodies,^{5 17} and have been associated with autoreactive T cells targeting antigens shared by tumors and affected tissues,^{18 19} suggesting distinct mechanisms underlying these two inflammatory conditions. To date, several studies have compared irAEs and CIDs within specific organs20,22; however, a large-scale, systematic comparison encompassing multiple affected tissue sites and peripheral blood remains lacking. Furthermore, while previous irAE research has primarily focused on the role of T cells,^{20 22 23} the interactions between various cell types that contribute to a complex inflammatory ecosystem and their roles in disease pathogenesis are not yet fully understood. A more comprehensive characterization of the intricate interplay among diverse cellular states within the inflammatory microenvironments that distinguish irAEs from CIDs is needed in both blood and affected tissues. Such insights could broaden our understanding of irAE pathogenesis and support the development of more effective treatment strategies beyond the current reliance on systemic corticosteroids and biologics used for CIDs.

In this study, we present a systematic multidimensional analysis of irAEs and CIDs, integrating transcriptomes of 4.3 million single cells from 1,137 samples, spatial transcriptomes for both conditions, and 526 paired blood transcriptomes collected before and after ICI treatment along with germline exomes and clinical irAE data. We aimed to identify hallmark gene signatures that are consistent across different irAE and CID inflammation sites in both peripheral blood and affected tissues. We also sought to characterize a diverse array of cell states encompassing the inflammatory microenvironments. Our results suggest that the NLRP3⁺ mononuclear phagocyte (MNP), LAMP3⁺ dendritic cell (DC), CXCL8/10 cytokine signaling, TNFRSF18⁺ regulatory T cell (Treg), T/natural killer (NK) interferon (IFN), and HAVCR2⁺ exhausted T cell (Tex) cell states form an irAE-specific inflammatory ecosystem in spatially localized niches absent in CIDs. Additionally, distinct cellular interactions mediated by CXCL9/10/11 and CCL3 further distinguish irAEs from CIDs. We also found that severe irAEs are associated with the activation of CXCL10 cytokine signaling following ICI treatment especially among HLA-B*46:01 carriers. Our comprehensive multidimensional compendium highlights key differences between irAEs and CIDs, providing valuable insights into potential biomarkers and therapeutic targets for the effective prevention and management of irAEs. For the research community, we have developed a web portal (https://sites.google.com/view/omics-cia) that enables researchers to interactively explore, download, and apply code tutorials to their own datasets.

Methods

Dataset collection and construction of a single-cell compendium of inflammation

We searched for single-cell transcriptome datasets that included irAEs, CIDs, and paired normal tissue samples from both peripheral blood and affected tissues (eg, organ tissues, bronchoalveolar lavage, cerebrospinal fluid, and synovial fluid). To minimize technical variability, we selected studies conducted using 10x genomics technology. Datasets were downloaded from publicly available repositories, including Gene Expression Omnibus, Single Cell Portal (https://singlecell.broadinstitute.org/), Gut Cell Atlas (https://www.gutcellatlas.org/), European Molecular Biology Laboratory (https://www.ebi.ac.uk/), Genome Sequence Archive for Human (https://ngdc.cncb.ac.cn/). In total, we collected and analyzed 38 single-cell transcriptome datasets comprising 1,137 samples and over 4.3 million cells (figure 1a and online supplemental data 1).

Figure 1. Scheme of the study. (a) Overview of the single-cell transcriptomes of irAEs, CIDs, and healthy control collected in this study. (b) Representative UMAP of the collected single-cell transcriptomes colored by cell types. (c) Dot plot for the expression of cell type-specific marker genes. CID, chronic inflammatory disease; DC, dendritic cell; irAE, immune-related adverse event; pDC, plasmacytoid dendritic cell; NK, natural killer; UMAP, Uniform Manifold Approximation Projection. Created with Biorender https://biorender.com.

Single-cell RNA sequencing data processing and analysis

The gene columns of each dataset were re-aligned according to the GRCh38 human reference genome (official Cell Ranger reference, V.2020-A). For each dataset, cells with fewer than 1,000 unique molecular identifier (UMI) counts and 500 detected genes were considered empty droplets and removed for further analysis. Cells with more than 7,000 detected genes were considered potential doublets and were removed from analysis. After quality control measures, patient samples with fewer than 30 cells were excluded. The Python Scanpy (V.1.10.1) package was used to load and analyze the count matrices.²⁴ Scrublet (V.0.2.3) was used for doublet detection.²⁵

Since the computation burden limited us from merging all datasets and proceeding with downstream analysis, we performed the remaining analysis by each dataset: Filtered count matrices underwent count-normalization and log transformation, and highly variable genes were identified using the default settings in Scanpy. Next, the expression matrices were scaled and principal component analysis (n=50 components) was conducted. The principal components were adjusted using the Harmony algorithm (V.0.0.10) to account for each patient’s technical variations.²⁶ The Uniform Manifold Approximation Projection (UMAP) coordinates were then determined by using the neighborhood graphs. The Leiden algorithm was employed for the clustering of cells.

Cell type annotation was performed using the CellTypist (V.1.6.2) model from count-normalized and log-transformed expression matrix for each dataset, and was refined manually based on cell type-specific marker genes.²⁷

Sample metadata curation and data harmonization

Patient sample metadata was searched in the Gene Expression Omnibus (for datasets obtained from Gene Expression Omnibus), corresponding research articles, and supplementary datasets for each dataset. For every patient sample, disease type (eg, ulcerative colitis, irAE colitis), disease category (eg, irAE, CID, and Healthy), tissue site (peripheral blood or organ), affected tissue site (eg, thyroid, synovium, colon), sequencing chemistry (3’ or 5’ chemistry), treatment status (treatment naive or treated), and available treatment regimen information were searched and collected.

For analysis that requires single-cell transcriptome data integration (figures 1b,c and 5a), data integration was performed using the anndata.concat function in the anndata package (V.0.12.2). To harmonize single-cell transcriptome datasets and obtain a representative UMAP at a global scale (figure 1b), a subset of cells was selected for each dataset (up to 20,000 cells per dataset) and subsequently, the SCANVI algorithm was implemented with default parameters,²⁸ using ‘Sample’ as batch key and sequencing chemistry as a covariate key to adjust for technical variations across patient samples.

Identification and analysis of hallmark gene signatures

To identify differentially expressed genes between each irAE and CID type, a pseudo-bulk matrix was compiled for each sample with at least 10 cells and 1,000 accumulated counts. Subsequent differential expression analysis was performed between irAEs and CIDs for each inflammation and tissue sampling site (eg, irAE colitis blood sample vs CID colitis blood sample) with PyDESeq2 (V.0.4.11).²⁹ We selected genes with a p value<0.05 with log2 fold-change values >0.5 (online supplemental data 2) for each tissue site, and the genes that are identified from the majority tissue sites (four or more out of eight tissue sites) were defined as hallmark gene signatures of each cell type.

We utilized Enrichr implemented in the gseapy (V.1.1.4) package to perform pathway enrichment analysis and identify the biological functions of the hallmark gene signatures of irAEs and CIDs across diverse cell types.^{30 31} We employed Gene Ontology (GO) terms from MsigDB Hallmark 2020, GO Biological Process 2021, and Kyoto Encyclopedia of Genes and Genomes (KEGG) 2021 Human. Only the terms with an adjusted p value<0.05 were considered as significant (figure 2b).

Figure 2. Characterization of the hallmark gene signatures of irAEs and CIDs across various inflammation sites. (a) Heatmap of the hallmark genes of irAEs and CIDs across inflammation sites in both peripheral blood and affected tissues. Each color in the heatmap represents log2 fold-change values with positive values (red) and negative values (blue) indicating upregulation in irAEs and CIDs, respectively. (b) Pathway enrichment analysis of the hallmark genes of irAEs and CIDs across diverse cell types using EnrichR. The color and size of the dot represent the OR and p value from the two-sided Fisher’s exact test, adjusted with the Benjamini-Hochberg method, respectively. Each dot color represents an OR with positive values (red) and negative values (blue) indicating enrichments in irAEs and CIDs, respectively. adj. p, adjusted p value; CID, chronic inflammatory disease; DC, dendritic cell; IL, interleukin; irAE, immune-related adverse event; MNP, mononuclear phagocyte; NK, natural killer; TGF, transforming growth factor; TNF, tumor necrosis factor.

NMF pre-processing and analysis

For each major cell type category, we performed non-negative matrix factorization (NMF) for each individual sample, accounting for the tissue sampling site, to identify gene modules that constitute the inflammatory environment in each sample. NMF analysis was performed through the following steps.

1. From the log-normalized scaled gene expression matrices for each dataset, the negative values were set to zero.

2. Gene expression matrices were selected for each patient.

3. The sklearn.decomposition.NMF was applied with default parameters as implemented in the scikit-learn Python package (V.1.6.1). Considering that NMF requires a K parameter, we ran NMF using different values (K=5, 6, 7, 8, 9), thereby generating 35 modules for each sample.³²

4. Gene modules derived from a single-cell type were integrated, and each gene module was scaled using sklearn.preprocessing.MinMaxScaler, thereby creating a scaled module x gene matrix.

5. Highly variable genes were selected in the module x gene matrix, and low-quality modules (modules with an NMF weight less than 30~50 or greater than 250, depending on each cell type) were removed.

6. Dimensional reduction was performed with principal component analysis, and technical variations between datasets were adjusted using the Harmony algorithm.

7. Gene modules were visualized with UMAP and clustered with the Leiden algorithm. Gene module clusters that were present in fewer than 5% of patients or observed exclusively in single datasets, likely representing technical artifacts rather than true biological cell states, were excluded.

8. A list of the top 50 genes with the highest NMF weight was derived for each gene module cluster. Subsequently, we removed clusters enriched with either ribosomal protein genes or mitochondrial-encoded genes (containing 3 or more genes among 50 genes).

9. Clusters composed of gene modules derived from a single study were removed for further analysis to only account for robust gene modules that commonly occur across datasets.

Cell state annotation and analysis

After visualizing and clustering gene modules for each cell type as listed above, cell state annotation was conducted through the following steps.

1. Gene module clusters (cell state) were initially identified using a Leiden resolution of 0.7, and the top 50 genes with the highest NMF weights were then determined for each cluster.

2. For instances of overlapping genes between cell states, the genes were orthogonally assigned to the cell state with the higher NMF weight.

3. Gene module clusters comprising fewer than five genes were regarded as low-quality clusters and were excluded from further analysis.

4. To annotate cell state marker genes, we performed a comprehensive literature search including the Celltypist database (https://www.celltypist.org/), and the resulting reference gene signatures were systematically compared with our cell state genes (online supplemental data 3).

5. When previously established signatures suggested a potential heterogeneity within a cluster, additional subclustering was iteratively performed to achieve a finer-grained cell state annotation.

6. To assess the robustness of our defined cell states, we calculated cell state scores and gene signature scores from previous studies using the sc.tl.score_genes function for each cell type. For each sample, mean gene signature scores were computed, and our cell state scores were compared with the reference gene signatures using the Pearson correlation (online supplemental figure S4).32,37

To identify the inflammatory disease category (irAE, CID, and healthy control) and blood or affected tissue preferences for each cell state, we used the ratio of observed to expected (Ro/e) analysis.^{32 36} We first selected organs (colon, joint, and thyroid) that include all disease categories (irAE, CID, and healthy control) to quantify tissue enrichments for each cell state. Subsequently, a contingency table was created by counting the gene modules from various tissue origins for each disease category (ie, irAE blood, irAE affected tissue, CID blood, CID affected tissue) in the cell state of interest versus others. An expected number was derived using the χ² analysis, and Ro/e was calculated as the ratio of observed to expected values. We considered a cell state enriched or depleted in a specific cell state if Ro/e >1 or Ro/e <1, respectively.

Construction of cell state networks

We constructed cell state networks to visualize co-occurrence patterns across various inflammatory conditions. To begin, we calculated cell state scores within the corresponding cell type categories using the sc.tl.score_genes function from the Scanpy package. Next, we determined the mean cell state scores for each sample. To exclude rare cell states from influencing the network architecture, we applied an Ro/e filtering threshold across sample sources (eg, blood vs affected tissue) and disease states (eg, inflammatory disease vs healthy control) using the following formula:

$${\displaystyle {\displaystyle \mathrm{R}\mathrm{o}/\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{l}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{i}\mathrm{n}\mathrm{g}\mathrm{t}\mathrm{h}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{h}\mathrm{o}\mathrm{l}\mathrm{d}=\mathrm{m}\mathrm{e}\mathrm{a}\mathrm{n}(\mathrm{R}\mathrm{o}/\mathrm{e})-1\mathrm{s}.\mathrm{d}(\mathrm{R}\mathrm{o}/\mathrm{e})}}$$

Cell states were classified as rare if their Ro/e values did not exceed this filtering threshold. From the sample cell state score matrix, we computed an adjacency matrix with “signed” mode using PyWGCNA (V.2.2.0).³⁸ To minimize spurious effects, adjacency values for cell state pairs with p values>0.05 were set to 0. The resulting adjacency matrix was then imported into the NetworkX (V.3.2.1) package, using the Kamada Kawai layout to construct a connected network.³⁹ The node colors represent each cell type category, edge thickness reflects the magnitude of adjacency between cell states, and node size corresponds to the weighted adjacency of a given cell state with others.

Ligand-receptor interaction analysis

To explore the interaction patterns of NLRP3⁺ MNPs and LAMP3⁺ DCs with other cell types, we analyzed their potential cellular interactions using cell–cell interaction inference tools, such as CellPhoneDB (V.3.1.0),⁴⁰ with a focus on the ligand genes associated with CXCL8/10 cytokine signaling cell states. Interaction intensity was determined by multiplying the normalized expression values of ligands and receptors for each cell–cell pair.

Spatial transcriptome analysis

Spatial transcriptome analysis was conducted using cell2location (V.0.1.4) using default parameters to quantify the spatial distribution of cell types across individual spatial spots⁴¹ (online supplemental data 4). For each spatial transcriptome dataset, the corresponding single-cell transcriptome dataset from the relevant disease was used as a reference to deconvolute each spatial spot (eg, the irAE colitis single-cell transcriptome dataset as a reference for the irAE colitis spatial transcriptome). Spatial spots with fewer than 100 UMI counts or with more than 50% of counts in mitochondrial genes were considered low-quality spots and removed from further analysis. For cell states that were not identified as distinct cell types (eg, CXCL8/10 cytokine signaling and IFN cell states), we employed the sc.tl.score_genes function from the Scanpy package to assess the spatial enrichment of these cell states. To conduct spatial ligand-receptor analysis, we used the liana (V.1.5.1) package to identify spatial hotspots involved in CXCL9/10/11:CXCR3-mediated interactions.⁴²

To identify cell types significantly enriched in niches with high CXCL10 cytokine signaling cell state activity (figure 5b and online supplemental figure S8a–c), we first selected the top 5% of spots most enriched or depleted in this cell state for each spatial sample. We then calculated the mean cell type abundances for both CXCL10 cytokine signaling cell state-enriched and depleted spots. The statistical significance of enrichment or depletion for each cell type was determined using a two-sided Wilcoxon rank-sum test, with p values adjusted using the Benjamini-Hochberg method.

IrAE cohort data analysis

We processed the blood transcriptome and exome data of 263 ICI-treated patients as previously described.⁴³ As for the transcriptome data, raw FASTQ files underwent adapter sequence trimming using Trimmomatic (V.0.39), followed by ribosomal RNA filtering using SortMeRNA (V.2.1b).^{44 45} Filtered sequences were aligned to the reference genome by STAR (V.2.7.6a) with gencode.V.37.annotation.gtf.⁴⁶ The read count was obtained using HTSeq (V.0.12.4), and the normalized expression level was calculated by the transcripts per million normalization method.⁴⁷ The blood transcriptomes were subjected to cell state deconvolution based on the single sample gene set enrichment analysis method as implemented in the gseapy package.³⁰ For HLA typing, we used raw FASTQ files from whole-exome sequencing to extract HLA genotypes at four-digit resolution (second field level) by running HLA-HD (V.1.3.0).⁴⁸

Protein level validation

To validate our findings at the protein level, we analyzed serum proteomic data generated using the Olink Proteomics Inflammation Panel, collected from patients before and after ICI treatment, together with clinical information on irAE occurrence and severity.⁴⁹ For each cell state of the myeloid-derived inflammatory ecosystem, we calculated the mean fold change in the associated serum protein expression between the pretreatment and post-treatment samples and compared these values across groups stratified by irAE severity.

Construction of interactive web portal

The interactive web portal was developed using the ShinyCell2 package in R (https://github.com/the-ouyang-lab/ShinyCell2), to provide a user-friendly interface for single-cell data exploration.⁵⁰ The portal supports multifaceted visualization, including dimension reduction plots (UMAP), gene expression profiles (violin, dot, and feature plots), and metadata distributions. To facilitate comparative analyses, backend statistical modules allow users to perform hypothesis testing between selected groups, including pairwise comparisons using various statistical tests. All visualizations and processed data subsets can be downloaded to support the generation of high-resolution figures.

Results

Construction of a single-cell transcriptome compendium for irAEs and CIDs

To delineate the distinct molecular characteristics of irAE and CID inflammation, we collected single-cell transcriptomes of blood, affected tissue, and paired normal tissue samples in both conditions (online supplemental data 1). For stringent curation and annotation of each dataset, we applied a standardized pipeline to exclude low-quality cells and doublets, followed by cell type annotation using the CellTypist model (see Methods).²⁷ This process yielded an inflammatory landscape derived from 38 studies, encompassing 1,137 samples (blood; n=467, affected tissue; n=670) from 801 donors (figure 1a and online supplemental data 1).

For a systematic comparison between irAEs and CIDs, we reclassified conditions based on the site of inflammation (eg, inflammatory bowel disease as CID colitis, immune-related thyroiditis as irAE thyroiditis; online supplemental table 1). Consequently, our compendium covered nine different inflammatory organs (figure 1a), with paired blood and affected tissue samples from arthritis, colitis, dermatitis, neuroinflammation, pneumonitis, and thyroiditis cases. The representative UMAP for all datasets demonstrated a well-integrated embedding of 38 studies with 20 global cell types (figure 1b,c and online supplemental figure S1a,b).

Hallmark gene signatures defining the distinct inflammatory landscapes of irAEs and CIDs

To characterize hallmark gene signatures consistently expressed across multiple inflammatory sites, we identified differentially expressed genes (DEGs) between irAEs and CIDs for each cell type in blood and affected tissue samples (figure 2a and online supplemental data 2). Intriguingly, DEG patterns remained consistent across different inflammation sites and sample sources (ie, blood or affected tissues) (figure 2a and online supplemental figure S2). Across multiple immune cell types, IFN-related genes (eg, IFITM1, CD74, PTPN6, and IRF5) were consistently elevated in irAEs (shown in red in the heatmap), whereas TNF-α signaling genes (eg, BTG1, CCNL1, GPR183, and IER5) were frequently elevated in CIDs (shown in blue in the heatmap).

Certain inflammation-related DEGs were overexpressed in irAEs (eg, FLOT2 and CNN2), whereas others were overexpressed in CIDs (eg, CXCR4 and GPR183) (figure 2a). Flotillin-2 (FLOT2), a membrane scaffolding protein that negatively regulates inflammatory responses and modulates immune cell activation, was elevated in irAEs commonly in B cells, DCs, NK cells, and CD8⁺ T cells. This elevation may represent a compensatory mechanism to mitigate excessive inflammation, consistent with its role as a negative regulator that suppresses hyperactive immune responses across various immune cell types.^{51 52}

Elevated CNN2 expression in B cells, DCs, NK cells, and CD4⁺ T cells was also observed in irAEs (figure 2a). CNN2, an actin cytoskeleton-associated gene, regulates key cellular processes such as proliferation, migration, and phagocytosis.⁵³ Notably, CNN2 has also been implicated in the pathogenesis of inflammatory arthritis,⁵³ and its elevation in multiple immune cell subtypes suggests increased immune activation and cytokine secretion, contributing to exacerbated inflammation in irAEs.

CXCR4 and GPR183 are critical regulators of immune cell trafficking, activation, and inflammation, playing similar roles in directing DCs, MNPs—including monocytes and macrophages but not DCs—and T cells to inflamed tissues.54,57 CXCR4, a chemokine receptor, binds to its ligand CXCL12 and activates MAPK and NF-κB signaling pathways,^{54 55} while GPR183 responds to its ligand 7α,25-dihydroxycholesterol to facilitate immune cell migration and enhance inflammatory responses.^{56 57} Elevated expression of both genes is linked to increased production of pro-inflammatory cytokines, including TNF-α, interleukin (IL)-6, and IL-1β.^{54 56} Their elevated expression in DCs, MNPs, and CD4⁺/CD8⁺ T cells in CIDs underscores their role in driving inflammation and sustaining immune activation, distinguishing CIDs from irAEs (figure 2a).

In line with the patterns of DEGs, pathway enrichment analysis also revealed the upregulation of IFN response in irAEs and TNF-α signaling via NF-κB pathways in CIDs across multiple cell types (figure 2b and online supplemental figure S2).⁵⁸ Additionally, irAEs were characterized by enrichment of genes associated with cytokine response, homotypic cell adhesion, oxidative phosphorylation, and IL-12-mediated signaling pathways (figure 2b). In contrast, pathways related to apoptosis, hypoxia, and myc targets were predominantly enriched in CIDs (figure 2b).

Several pathways were activated in different cell types between irAEs and CIDs (figure 2b). For instance, genes involved in allograft rejection were significantly upregulated in B cells, CD4/CD8⁺ T, and NK cells in irAEs, whereas they were activated in DCs and MNPs in CIDs (figure 2b). Adipogenesis was active in CD8⁺ T cells in irAEs, but in DCs and MNPs in CIDs (figure 2b).

CD8⁺ T cells in irAEs exhibited enrichment for pathways related to E2F targets, glucose metabolic process, glycolysis, and mTORC1 signaling (figure 2b). These metabolic and proliferative features in CD8⁺ T cells likely reflect an augmented inflammatory process characterized by cellular activation, proliferation, and increased energy demands. Collectively, these findings highlight the distinct mechanisms underlying the dysregulated inflammatory responses observed in irAEs and CIDs.

Diverse cell states constituting the inflammatory milieu of irAEs and CIDs

We next aimed to systematically dissect the complex inflammatory milieu of irAEs and CIDs in terms of various cell states. We first identified gene modules by applying NMF to each sample.^{32 34 59} These gene modules collected from different samples were integrated for each cell type and subjected to graphical visualization and clustering³² (figure 3a–c and online supplemental figure S3; see Methods). Despite being derived from diverse inflammatory conditions, these gene modules clustered together, reflecting shared gene expression programs conserved across heterogeneous inflammatory contexts (figure 3a–c and online supplemental figure S3).

Figure 3. Deconvolution of the diverse cell states constituting the inflammatory milieu. (a–c) UMAP visualization of gene modules in (a) (b) myeloid, (c) and T/NK cell states. Gene modules were graphically clustered and colored by cell states. (d–e) Ro/e analysis of enrichments in (d) blood and (e) affected tissue cell states. The horizontal line represents where Ro/e=1. The statistical significance was calculated using the two-sided χ² test and adjusted with the Benjamini-Hochberg method; *FDR <0.05, **FDR <0.01 and ***FDR <0.001. DC, dendritic cell; FDR, false discovery rate; GC, germinal center; HEV, high endothelial venules; ILC, innate lymphoid cell; inf., inflammatory; MAIT, mucosal-associated invariant T cell; NK, natural killer; pDC, plasmacytoid dendritic cell; Ro/e, ratio of observed to expected; TCM, central memory T cell; TEM, effector memory T cell; Tex, exhausted T cell; Tfh, follicular helper T cell; Th17, T helper 17 cell; Trm, tissue-resident memory T cell; Treg, regulatory T cell; UMAP, Uniform Manifold Approximation Projection.

As a result, we identified a total of 92 clusters of gene modules (hereafter referred to as “cell states”) that constitute the inflammatory milieu of irAEs and CIDs (figure 3a–c, online supplemental figure S3 and online supplemental data 3). Although derived from non-tumor tissues, many of the cell states reflected gene signatures previously reported in cancer studies,^{32 34 59} including the HAVCR2⁺ Tex (HAVCR2 and LAYN), CCL19⁺ fibroblast (CCL19 and RBP5), WNT5A⁺ inflammatory fibroblast (WNT5A and MMP1), epithelial senescence (SERPINB3 and SULT2B1), and epithelial stress (JUN, FOSB, and HSPA2) cell states (figure 3c and online supplemental figure S3). Our cell states also strongly correlated with previously reported gene signatures,32,37 supporting their robustness and biological relevance (online supplemental figure S4).

The Ro/e analysis (see Methods) revealed differential enrichment of cell states between irAEs and CIDs (figure 3d,e and online supplemental figure S5). Multiple DC cell states (DC1, DC2, and plasmacytoid dendritic cell) were under-represented in irAEs compared with CIDs or healthy controls in both blood and affected tissue samples, in line with previous irAE studies that reported DCs as under-represented in irAEs^{7 60} (figure 3d–e). Compared with healthy normal blood samples, the FCRL5⁺ atypical B (FCRL5 and FCRL2), TCL1A⁺ naive B (TCL1A and FCRLA), T/NK IFN (ISG20 and PSMB8) and myeloid IFN (IFIT1 and IFIT2) cell states were over-represented in both irAEs and CIDs. In contrast, naive T (IL7R and CCR7) and XCL1⁺ NK (XCL1 and XCL2) cell states were under-represented in both irAEs and CIDs relative to healthy normal blood samples (figure 3d).

In affected tissues, various cell states were overrepresented in irAEs, including CXCL10 cytokine signaling (CXCL10 and CXCL9), FGD2 inflammatory response (FGD2 and LPCAT2), HAVCR2⁺ Tex, and TNFRSF18⁺ Treg (TNFRSF18 and TNFRSF4) cell states (figure 3e). In contrast, FOLR2⁺ macrophage (FOLR2 and F13A1), myeloid cell cycling (MKI67 and TOP2A), naive T, germinal center B (RGS13 and SUGCT) cell states were under-represented in irAEs (figure 3e).

To investigate the immune cell state compositions across blood and affected tissues in two distinct inflammatory conditions, we performed Ro/e analysis to quantify compartment-specific enrichment (online supplemental figure S6). Both irAEs and CIDs demonstrated broadly similar patterns of cell state distribution, with macrophages, HAVCR2⁺ Tex, follicular helper T cell (Tfh) (CXCL13 and TOX2), and germinal center B-cell states preferentially accumulated in affected tissues. Conversely, CD16⁺ NK (FCGR3A and GNLY), ZNF683⁺ TEM (ZNF683 and GZMH), and CD14⁺/CD16⁺ monocytes exhibited enrichment within the blood compartment (online supplemental figure S6). Importantly, certain cell states showed differential enrichment patterns between irAEs and CIDs. For example, PAX5⁺ naive B (PAX5 and AFF3) cell states were predominantly enriched in the blood samples of irAEs but not in CIDs, while activated B (CD69 and NFKBID) and CD4⁺ TEM (AQP3 and ITGB1) cell states demonstrated an opposite pattern between irAEs and CIDs (online supplemental figure S6). These results suggest that the compartmental organization of immune cell states may be influenced by the underlying type of inflammation.

Distinct ecosystems in the inflammatory milieu of irAEs and CIDs

We then aimed to delineate distinct ecosystems composed of co-regulated cell states in blood and affected tissue samples of irAEs and CIDs. To achieve this, we constructed an undirected network based on correlations between the cell states, which revealed different cellular ecosystems in irAEs and CIDs across both blood and affected tissues (figure 4a–d).

Figure 4. Co-occurrence network between cell states in blood and affected tissues of irAEs and CIDs. (a–d) Co-occurrence network of cell states in (a) blood of irAE samples, (b) blood of CID samples, (c) affected tissues of irAE samples, (d) affected tissues of CID samples. The color and the thickness of the edges correspond to the cell type category and the magnitude of the adjacencies between cell states, respectively. CID, chronic inflammatory disease; DC, dendritic cell; GC, germinal center; inf., inflammatory; irAE, immune-related adverse event; MAIT, mucosal-associated invariant T cell; NK, natural killer; pDC, plasmacytoid dendritic cell; TCM, central memory T cell; TEM, effector memory T cell; Tex, exhausted T cell; Tfh, follicular helper T cell; Th17, T helper 17 cell; TNF, tumor necrosis factor; Trm, tissue-resident memory T cell; Treg, regulatory T cell.

The blood networks of irAEs and CIDs exhibited distinct co-occurrence patterns of the IFN and CXCL10 cytokine signaling cell states (figure 4a,b). In irAEs, CXCL10 cytokine signaling and myeloid IFN cell states formed an ecosystem in connection with the PAX5⁺ naive B, and T/NK cell cycling cell states (figure 4a). In contrast, the CID network showed a weaker connectivity within these cell states (figure 4b). Also, the TNF-α signaling cell state (TNF and NFKBIZ) was closely adjacent to NLRP3⁺ MNPs, stress/heat shock from T/NK, and FGD2 inflammatory response in irAEs. By contrast, in CIDs, the TNF-α signaling cell state exhibited limited connectivity and was more distantly located with these cell states (figure 4a,b). These findings suggest differential crosstalk between innate inflammatory responses and T cell-mediated pathways between irAEs and CIDs. Especially, myeloid-derived cell states are concurrently active in irAE blood, including the myeloid IFN-CXCL10 cytokine signaling and NLRP3⁺ MNP-TNF-α signaling. Furthermore, these cell states form integrated ecosystems with adaptive immune cell states, emphasizing the critical role of myeloid-driven inflammatory pathways in shaping the unique inflammatory nature of irAEs. This stands in contrast to CID blood ecosystems, which exhibit a more compartmentalized architecture of myeloid-derived cell states (figure 4b).

Comparison between the cell state networks of affected tissues in irAEs and CIDs revealed distinct differences in cell state distribution (figure 4c,d). Notably, we identified one ecosystem enriched with myeloid-derived cell states, including NLRP3⁺ MNP, CXCL8/10 cytokine signaling, and LAMP3⁺ DC. These cell states are more centralized and interconnected within the irAE network, whereas the CID network exhibits a more diffuse organization with less pronounced clustering of these components (figure 4c,d). This ecosystem in irAEs was also closely adjacent to T/NK cell states over-represented in irAEs, including the TNFRSF18⁺ Treg, IFN, and HAVCR2⁺ Tex. This organization indicates that inflammatory myeloid pathways are more actively co-regulated with T/NK cell states in irAEs than in CIDs. In addition, the Trm (DOCK10 and PARP8) cell states in irAEs were more tightly interconnected with FGD2 inflammatory response, germinal center B, and FCRL4⁺ atypical B (figure 4c,d), whereas the CID network displays a more compartmentalized structure, with these ecosystems showing weaker connectivity (figure 4d).

Cytokine-mediated immune cell crosstalk differentiates irAEs from CIDs

A myeloid-derived inflammatory ecosystem emerged as one of the most prominent distinguishing factors between irAEs and CIDs. The CXCL10 cytokine signaling cell state, along with specific T-cell states such as the TNFRSF18⁺ Treg and HAVCR2⁺ Tex cell states, were overrepresented in irAEs (figure 3d,e). Numerous genes associated with CXCL10 cytokine signaling (eg, CXCL9/10/11 and IL1RN) and CXCL8 cytokine signaling (eg, CXCL2/3/8), NLRP3⁺ MNP (ICAM1), and TNF-α signaling (eg, CCL3/4, CCL3L1, CCL4L2) cell states contained ligand genes that mediate immune cell recruitment and inflammatory responses61,63 (online supplemental data 3).

For further investigation, we conducted ligand-receptor interaction analyses to compare irAEs and CIDs in affected tissue samples from colitis, thyroiditis, and arthritis (figure 5a and online supplemental figure S7). We first searched for cell types within the myeloid-derived inflammatory ecosystems that express these ligands or corresponding receptor genes. This led us to LAMP3⁺ DCs and NLRP3⁺ macrophages, which express many of these ligand genes, along with their cell type-specific markers (online supplemental figure S7a).

Figure 5. Distinct cellular interactions and spatial architectures between irAEs and CIDs in affected tissue samples. (a) Ligand-receptor analysis of LAMP3⁺ DCs and NLRP3⁺ macrophages with other cell types. The interaction intensity was calculated by multiplying the normalized expression values of ligands and receptors in each cell–cell pair. (b) Barplot depicting the spatial co-localizations with CXCL10 cytokine signaling cell state with other cell types. The statistical significance of the enrichment or depletion was calculated using the two-sided Wilcoxon rank-sum test and adjusted with the Benjamini-Hochberg method. Only the cell types that reached statistical significance are presented. (c) Spatial co-localization of irAE-specific inflammatory ecosystems in irAE and CID colitis and dermatitis, with colors representing cell abundance. CID, chronic inflammatory disease; DC, dendritic cell; EC cell, enterochromaffin cell; inf., inflammatory; irAE, immune-related adverse event; NK, natural killer; TEM, effector memory T cell; Tex, exhausted T cell; Tfh, follicular helper T cell; Treg, regulatory T cell.

This finding prompted us to examine ligand-receptor interactions for these two cell types (figure 5a). IrAEs and CIDs shared CCL3:CCR5-mediated interactions of NLRP3⁺ macrophages with GZMK⁺ CD8⁺ TEM and HAVCR2⁺ Tex in colitis, thyroiditis, and arthritis (figure 5a and online supplemental figure S7b,c). Notably, the crosstalk between NLRP3⁺ macrophages and TNFRSF18⁺ Treg via CCL3:CCR5 was stronger in irAEs compared with CIDs (figure 5a). Both irAE-colitis and CID-colitis involved interactions between LAMP3⁺ DCs and ADAMDEC1⁺ fibroblasts via CXCL11:ACKR3, as well as NLRP3⁺ macrophages and SOX6⁺/WNT5A⁺ fibroblasts via CXCL11:ACKR3 and IL1RN:IL1R1 (figure 5a).

In contrast, CXCL9/10/11-mediated irAE-specific immune cell crosstalk was absent in CIDs. CXCL9/10/11, a group of IFN-γ-inducible chemokines, promote pro-inflammatory cytokine release and recruit effector cells to inflamed tissues.⁶² For example, in irAE colitis, LAMP3⁺ DCs interacted with GZMK⁺ CD8⁺ TEM, HAVCR2⁺ Tex, Tfh, and TNFRSF18⁺ Treg via CXCL9/10/11:CXCR3 (figure 5a). However, these interactions occurred through CD70:CD27 in CID colitis. Similarly, NLRP3⁺ macrophages demonstrated CXCL9/10/11:CXCR3-mediated interactions with T cells in irAEs, which were absent in CIDs, across various inflammation sites (figure 5a and online supplemental figure S7b,c). These pro-inflammatory interactions of CXCL9/10/11-expressing LAMP3⁺ DCs and NLRP3⁺ macrophages with CXCR3-expressing T cells suggest an amplified IFN-γ-driven inflammatory pathway, distinguishing irAEs from CIDs.

Cellular crosstalk mediated by CCL3 and ITGAL further highlighted the differences between irAEs and CIDs. Specifically, NLRP3⁺ macrophages and HAVCR2⁺ Tex were engaged primarily via CCL3:CCR5 in CIDs but via CXCL9/10/11:CXCR3 in irAEs (figure 5a). Similarly, HAVCR2⁺ Tex and NLRP3⁺ macrophages interacted through ITGAL:ICAM1 in CIDs but through CCL3:CCR1 in irAEs (figure 5a). These differences were consistently observed across both colitis and thyroiditis. Although both ITGAL:ICAM1 and CCL3:CCR1 interactions promote immune cell infiltration and inflammation,^{64 65} CCL3 and CCR1 are involved in acute inflammation development,⁶⁶ whereas ITGAL and ICAM1 are associated with chronic inflammatory conditions.^{67 68} This distinction highlights the acute inflammatory nature of irAEs, in contrast to the persistent inflammation observed in CIDs. Taken together, these findings underscore how differential ligand-receptor engagements shape the unique inflammatory landscapes of irAEs and CIDs.

Myeloid-derived inflammatory ecosystem creates spatially restricted pro-inflammatory niches in irAEs

Having identified the cell states characteristic of irAEs (ie, NLRP3⁺ MNP, LAMP3⁺ DC, CXCL10 cytokine signaling, TNFRSF18⁺ Treg, T/NK IFN, HAVCR2⁺ Tex), we set out to explore distinct spatial configurations of these cell state components between irAEs and CIDs (figure 5b,c and online supplemental figure S8). To this end, we conducted spatial transcriptome analysis of colitis and dermatitis in both irAEs and CIDs (figure 5b,c and online supplemental data 4).

The myeloid-derived inflammatory ecosystem, comprising components such as CXCL10 cytokine signaling, NLRP3⁺ macrophages, LAMP3⁺ DCs, TNFRSF18⁺ Treg, and HAVCR2⁺ Tex, and T/NK IFN cell states exhibited spatial co-localization in irAE colitis and dermatitis (figure 5b,c and online supplemental figure S8). Remarkably, this spatial organization was absent in CID colitis and dermatitis. Given that CXCL9/10/11:CXCR3-mediated crosstalk distinguished irAEs from CIDs (figure 5a), we identified spatial regions with high chemotactic activity of the CXCL9/10/11:CXCR3 axis in irAEs by using spatial ligand-receptor interaction analysis. These regions showed a notable enrichment of the components of the myeloid-derived inflammatory ecosystem (online supplemental figure S9). Together with the single-cell transcriptome-based analysis (figure 5a), these findings suggest that CXCL9/10/11, expressed by LAMP3⁺ DCs and NLRP3⁺ macrophages, may recruit CXCR3-expressing TNFRSF18⁺ Treg and HAVCR2⁺ Tex into the local inflammatory environment. This mechanism may play a role in the formation of irAE-specific niches with localized pro-inflammatory responses.

These results underscore the fundamental differences in inflammatory mechanisms between irAEs and CIDs. In irAEs, CXCL9/10/11-driven chemotactic activity establishes a tightly organized pro-inflammatory niche that likely plays a pivotal role in irAE pathogenesis (figure 5b,c). In contrast, the loss of co-localization in CIDs reflects the absence of CXCL9/10/11:CXCR3 chemotactic activity, potentially due to differences in underlying inflammatory mechanisms or the chronicity of inflammation, which may influence the spatial proximity of these cell populations.

Severe irAEs are associated with CXCL10 cytokine signaling and HLA-B*46:01

Given multiple lines of evidence showing that blood and affected tissue samples share similarities in the context of inflammation, and that tissue inflammation can be mirrored in peripheral blood,69,72 we postulated that the myeloid-derived inflammatory ecosystem would also be reflected in peripheral blood during ICI treatment. Therefore, we used data from an irAE cohort, which included pre-ICI and post-ICI treatment-matched blood transcriptomes along with germline exomes and clinical information on irAE occurrence for 263 patients.⁴³ As a result, post-treatment samples from the irAE group exhibited significant increases in the CXCL10 cytokine signaling, myeloid IFN, GZMK⁺ TEM, HAVCR2⁺ Tex, and Tfh cell states in their blood transcriptomes (figure 6a). Although the CXCL10 cytokine signaling and HAVCR2⁺ Tex cell states also increased in the control group, the increase was substantially greater in the irAE group (figure 6a).

Figure 6. ICI-induced changes of cell states and genetic predispositions of irAE occurrence. (a) Dot plot showing the cell states changes on ICI treatment. Upregulated and downregulated cell states after ICI treatment are colored in red and blue, respectively. The statistical significance was calculated using the paired two-sided Wilcoxon rank-sum test and adjusted with the Bonferroni method. (b) Line plot depicting the CXCL10 cytokine signaling cell state scores before and after ICI treatment according to irAE severity. The vertical whiskers display the SE with the mean represented by the center dot. The statistical significance was calculated using the two-sided Wilcoxon rank-sum test; *p<0.05. (c) Boxplot displaying the mean fold change of CXCL10 cytokine signaling-associated proteins following ICI treatment across irAE severity groups. The box represents the interquartile range (25th–75th percentiles), the central line indicates the median, and whiskers extend to the most extreme values within 1.5 times the IQR. Statistical significance was assessed using a two-sided Kruskal-Wallis test. (d) Stacked barplot illustrating the proportion of HLA-B*46:01 carriers according to the severity of irAE. The statistical significance was calculated using the two-sided χ² test. (e) Boxplot showing the fold change of CXCL10 cytokine signaling on ICI treatment by HLA-B*46:01 carrier status. The box represents the interquartile range (25th–75th percentiles), the central line indicates the median, and whiskers extend to the most extreme values within 1.5 times the IQR. The statistical significance was calculated using the one-sided Wilcoxon rank-sum test. adj. p, adjusted p value; DC, dendritic cell; ICI, immune checkpoint inhibitor; inf., inflammatory; irAE, immune-related adverse event; NK, natural killer; TCM, central memory T cell; TEM, effector memory T cell; Tex, exhausted T cell; Tfh, follicular helper T cell; Treg, regulatory T cell.

Next, we sought to identify cell states associated with severe irAEs (grade ≥3). Among those significantly increased following ICI treatment, the CXCL10 cytokine signaling cell state was markedly upregulated in post-treatment samples from the severe irAE group compared with both the mild irAE (grade 1–2) and control groups (figure 6b). Other cell state scores in the severe irAE group were comparable to those in the mild irAE group but were higher than those in the control group (online supplemental figure S10). This finding was further confirmed at the protein level, showing the most significant elevation of CXCL10 cytokine signaling cell state proteins in the severe irAE group on ICI administration⁴⁹ (figure 6c and online supplemental figure S11a–d).

Considering the crucial role of HLA alleles in shaping immune responses and irAE occurrence,^{43 73 74} we hypothesized that specific HLA types could influence the development of severe irAEs. Since the cohort data included germline exomes, we searched for HLA types associated with irAE severity. As a result, a significantly higher proportion of HLA-B*46:01 carriers were found in the severe irAE group compared with the mild irAE and control groups (χ² test p=6.8 e⁻⁰³, control: 6.9%, mild: 9.8%, severe: 28%) (figure 6d). This finding suggests a potential genetic predisposition to the development of severe irAEs.

To examine how HLA-B*46:01 status affects the CXCL10 cytokine signaling cell state (the cell state associated with severe irAEs), we compared the magnitude of changes in the cell state score on ICI treatment between HLA-B*46:01 carriers and non-carriers. HLA-B*46:01 carriers exhibited an increase in the CXCL10 cytokine signaling scores in their blood transcriptomes after ICI treatment compared with non-carriers (p=6.8 e⁻⁰², Cohen’s d=0.3, figure 6e), supporting our hypothesis that certain HLA alleles may contribute to an exacerbated inflammatory response following ICI treatment due to overactive CXCL10 cytokine signaling.

Discussion

In this study, we compiled a comprehensive gene expression compendium of irAEs and CIDs, comprising 4.3 million single cells from 1,137 samples. This compendium was analyzed in combination with additional data, including spatial transcriptomes, bulk blood transcriptomes, germline exomes, and clinical information. We identified hallmark gene signatures differentially expressed between irAEs and CIDs, consistently appearing across various inflammatory sites in both peripheral blood and affected tissues. Among the 92 distinct cell states underlying the inflammatory processes, we discovered that the NLRP3⁺ MNP, LAMP3⁺ DC, CXCL10 cytokine signaling, TNFRSF18⁺ Treg, T/NK IFN, and HAVCR2⁺ Tex cell states collectively establish an irAE-specific inflammatory ecosystem within spatially localized niches that are unique to irAEs but absent in CIDs. Additionally, we uncovered immune cell crosstalk mediated by CXCL9/10/11 and CCL3, which further differentiates irAEs from CIDs. Patients who experienced severe irAEs exhibited significant activation of CXCL10 cytokine signaling following ICI treatment. Furthermore, a higher frequency of HLA-B*46:01 was observed among these patients, suggesting a genetic predisposition linked to severe immune responses. We have created a web portal (https://sites.google.com/view/omics-cia), where researchers can interactively visualize, download, and apply code tutorials to their own cohorts.

Although irAEs and CIDs share similar symptoms and clinical manifestations,8,10 we identified hallmark gene signatures that are differentially expressed between these conditions in various cell types, shared across diverse inflammation sites in both peripheral blood and affected tissues. Despite tissue-specific heterogeneity embedded in various organs, our findings suggest the existence of pan-irAE and pan-CID gene expression programs across diverse cell types. The hallmark gene signatures of irAEs and CIDs also exhibited distinct pathway enrichments. Across multiple cell lineages, irAE gene signatures were enriched in pathways related to IFN and cytokine responses, oxidative phosphorylation, and IL-12-mediated signaling.⁵⁸ In contrast, CID gene signatures were enriched in pathways associated with TNF-α signaling, hypoxia, apoptosis, and Myc targets. These results highlight the distinct molecular mechanisms driving the dysregulated inflammatory responses in irAEs and CIDs, which are shared across multiple cell types. While current treatment guidelines for irAEs recommend the use of TNF-α inhibitors particularly for corticosteroid-resistant cases,11,13 the enrichment of TNF-α signaling in CIDs compared with irAEs suggests that TNF-α inhibitors may be more effective for treating CIDs than irAEs.

Our spatial transcriptome analysis revealed the co-localization of myeloid-derived inflammatory cell states components in irAEs in contrast to CIDs. Ligand-receptor analysis revealed irAE-specific interactions between LAMP3⁺ DCs and NLRP3⁺ macrophages with T-cell subtypes, mediated by CXCL9/10/11, a group of IFN-γ-inducible chemokines that drive pro-inflammatory responses. In irAEs, IFN-γ-induced inflammatory responses promote the release of CXCL9/10/11 from LAMP3⁺ DCs and NLRP3⁺ macrophages. This leads to the recruitment and activation of CXCR3-expressing GZMK⁺ CD8⁺ TEM, TNFRSF18⁺ Treg, HAVCR2⁺ Tex, and Tfh, thereby creating a localized niche and fostering the development of an inflammatory ecosystem not observed in CIDs. While many studies on irAEs focus on the role of T cells,^{20 22 23} our findings highlight the critical role of myeloid cells in driving the inflammatory cascade and contributing to irAE pathogenesis. Future research targeting these myeloid cell states or the CXCL9/10/11 axis should be pursued for effective management of irAE patients and to mitigate their symptoms.

Our irAE cohort analysis identified CXCL10 cytokine signaling as markedly upregulated on ICI treatment in severe irAEs. This finding was validated at the protein level and aligns with previous studies reporting increased circulating levels of CXCL9 and CXCL10 proteins in severe cases.^{7 49} In patients undergoing ICI treatment who develop symptoms indicative of side effects, changes in the CXCL10 cytokine signaling cell state may serve as a reliable indicator for identifying those with severe irAE. We also identified a significantly higher frequency of HLA-B*46:01 carriers among patients with severe irAEs and demonstrated their potential to stimulate CXCL10 cytokine signaling. While prior studies have highlighted the role of HLA variants in autoimmune diseases and the development of irAEs,^{43 73 74} our findings provide new insights into the association between HLA variants and irAE severity. Furthermore, for HLA-B*46:01 carriers undergoing ICI treatment, personalized therapeutic strategies, such as ICI dose adjustments, should be considered to prevent the development of severe irAEs.

Our study provides comprehensive multidimensional analyses, integrating single-cell and spatial transcriptomes of irAEs and CIDs, along with pre-ICI and post-ICI treatment-matched blood transcriptomes with germline exomes and clinical information on irAE occurrence. This approach enabled us to define pan-irAE and pan-CID hallmark gene signatures and identify an irAE-specific inflammatory ecosystem enriched with myeloid components. These components exhibit distinct cellular interactions that promote the formation of spatially localized pro-inflammatory niches, which are absent in CIDs. Additionally, we highlight the association of severe irAEs with CXCL10 cytokine signaling and HLA-B*46:01, both of which may serve as valuable biomarkers to guide precision medicine strategies for the prevention and management of irAEs. Our findings offer critical insights into the unique inflammatory features distinguishing irAEs and CIDs, laying the foundation for future research.

Data availability statement

Data are available in a public, open access repository.