Abstract

Background

   While highly efficacious for numerous cancers, immune checkpoint
   inhibitors (ICIs) can cause unpredictable and potentially severe
   immune-related adverse events (irAEs), underscoring the need to
   understand irAE biology.

Methods

   We used a multidimensional approach incorporating single-cell RNA
   sequencing, mass cytometry, multiplex cytokine assay, and antinuclear
   antibody (ANA) profiling to characterize the peripheral immune
   landscape of patients receiving ICI therapy according to irAE
   development.

Results

   Analysis of 162 patients revealed that individuals who developed
   clinically significant irAEs exhibited a baseline proinflammatory,
   autoimmune-like state characterized by a significantly higher abundance
   of CD57^+ T and natural killer (NK) T cells, plasmablasts,
   proliferating and activated CXCR3^+ lymphocytes, CD8^+ effector and
   terminal effector memory T cells, along with reduced NK cells and
   elevated plasma ANA levels. In irAE cases, we identified distinct
   baseline proinflammatory gene signatures, including markedly higher
   expression of IL1B and CXCL8 in monocytes and CXCR3, TNF, and IFNG in
   T/NK cells. TNF signaling was the most enriched pathway, while
   immunosuppressive genes SIGLEC7 and CXCR4 were downregulated. Following
   ICI initiation, these patients exhibited an enhanced shift toward an
   activated and inflammatory immune phenotype, including monocyte
   reprogramming characterized by upregulation of IL18 and elevated gene
   expression levels of CXCL10. Conversely, post-treatment levels of CXCL8
   were decreased in irAE patients. Notably, in patients who did not
   develop clinically significant irAE, we identified increased baseline
   abundance of a TGFBI^high myeloid cluster enriched in immunosuppressive
   markers such as STAB1. In addition, patients without irAE exhibited
   upregulation of TNF and AIRE, accompanied by distinct myeloid
   protumorigenic reprogramming.

Conclusions

   A pre-existing activated, autoimmune-like proinflammatory state drives
   the development of irAE during ICI therapy through three key axes:
   increased plasmablast/ANA, heightened interferon-gamma/CXCL10/CXCR3
   axis, and amplified TNF signaling. These findings may serve as
   potential peripheral immune biomarkers for predicting irAE and provide
   biological insights into the mechanisms governing and mitigating irAE.

   Keywords: Immune related adverse event - irAE, Immunotherapy, Immune
   Checkpoint Inhibitor, Autoimmune, Antibody
     __________________________________________________________________

WHAT IS ALREADY KNOWN ON THIS TOPIC

     * After tens of thousands of patients have been treated with
       immunotherapy, the only established risk factor for immune-related
       adverse events (irAE) is pre-existing autoimmune disease.
     * While several studies have reported on immune parameters associated
       with irAE development, the underlying biology of irAEs remains
       elusive, as do clinically useful predictive biomarkers.

WHAT THIS STUDY ADDS

     * This multiomic biomarker analysis identifies a preexisting
       activated, autoimmune-like proinflammatory state in patients who
       subsequently develop irAE.
     * Three key axes contribute to this irAE risk phenotype: increased
       plasmablasts/antinuclear antibody, heightened interferon
       (IFN)γ/CXCL10/CXCR3 axis, and amplified TNF signaling.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

     * Identification of the molecular mechanisms underlying irAE may
       serve as potential peripheral immune biomarkers for predicting
       irAE.
     * These findings also provide possible targets for specific
       therapeutic interventions affecting T cells, B cells, and
       IFN-mediated inflammation.

Introduction

   Immune checkpoint inhibitors (ICIs) are highly effective therapies that
   have revolutionized the treatment of numerous cancers. However, in some
   cases, these agents cause autoimmune toxicities termed immune-related
   adverse events (irAEs). In contrast to the predictability and limited
   scope of chemotherapy and molecularly targeted therapy toxicities,
   irAEs may occur at any point during—and occasionally even well
   after—ICI therapy.[71]^1 2 They may affect almost any organ system,
   including potentially fatal cardiovascular and neurologic events in
   rare cases.[72]^3 In recent years, ICI indications have expanded from
   refractory, advanced malignancies to earlier-stage, curable
   tumors.[73]^4 At the same time, combination ICI regimens, which are
   associated with greater irAE occurrence and severity, have received
   approval for melanoma, kidney cancer, lung cancer, liver cancer, and
   mesothelioma.[74]^5 Taken together, these trends have only heightened
   the clinical importance of irAEs.

   Further complicating ICI administration are profound challenges in
   monitoring and diagnosing irAEs. Although irAEs may affect a plethora
   of organs, the rarity of certain types such as myocarditis renders
   longitudinal, extensive multiorgan monitoring costly and impractical.
   Indeed, expert guidelines generally limit routine surveillance to
   standard chemistries and thyroid function tests, which may detect
   kidney, liver, and certain endocrine toxicities but little else.[75]^6
   7 Once irAEs do occur, accurate and timely recognition, evaluation, and
   treatment remain elusive. Like non-ICI-related autoimmune
   diseases—which may require incorporation of clinical history, physical
   examination findings, serologies, radiographic studies, and
   occasionally histologic sampling to render a diagnosis—it appears that
   irAEs are far more difficult to diagnose than chemotherapy-related or
   targeted therapy-related toxicities.[76]^8 Failure to detect and treat
   irAE quickly could result in high-grade events. Although lower-grade
   irAEs have been associated with ICI efficacy,[77]^9 10 the acute
   clinical deterioration caused by severe irAEs may overshadow any
   potential benefit.[78]^11 Given the expanding use of ICI therapy across
   cancer types and stages, the ability to predict and accurately diagnose
   irAEs represents a key clinical priority in the field of
   immuno-oncology.

   Several studies have reported on immune parameters associated with irAE
   development, including cytokines, chemokines, pre-existing or induced
   autoantibodies, T cell receptor (TCR) repertoire, immune cell subsets,
   and gut microbiomes.[79]12,[80]24 However, the underlying biology of
   irAEs remains elusive, as do clinically useful predictive biomarkers.
   To address these challenges, we performed systematic, longitudinal, and
   comprehensive multidimensional analysis incorporating single-cell RNA
   sequencing (scRNA-seq), mass cytometry (CyTOF), antinuclear antibody
   (ANA) profiling, and multiplex cytokine assays of longitudinal blood
   samples from ICI-treated patients.

Materials and methods

Study design, clinical data sources, and clinical specimen collection

   As described previously, we collected clinical data and longitudinal
   specimens from patients enrolled in an institutional prospective
   immunotherapy cohort. Eligible patients had any type of cancer planned
   for but not yet started on ICI (including anti-programmed death 1
   (PD1), PD1 ligand (PDL1), and cytotoxic T lymphocyte antigen 4 (CTLA4))
   monoclonal antibodies, administered as single-agent therapy or in
   combinations). At the time of case selection, anti-lymphocyte antigen 3
   (LAG3) antibodies had not been Food and Drug Administration approved
   and were therefore not included in the patient sample. Due to
   challenges in determining the occurrence, type, timing, and severity of
   irAE,[81]^8 two oncologists (MSvI, DH) experienced in ICI
   administration and monitoring reviewed each case for irAE, with
   discrepancies reviewed and adjudicated by a third experienced clinician
   (DEG). As performed in other studies,[82]^10 25 we categorized irAE as
   clinically significant (Common Terminology Criteria for Adverse Events
   (CTCAE) grade≥2) or not clinically significant (CTCAE grade ≤1).
   Additional details on data collection, review, and characterization are
   provided in [83]online supplemental methods.

   Enrolled patients underwent initial blood collection at pretreatment
   baseline (BL) immediately prior to ICI therapy, after one treatment
   cycle (2–4 weeks), and at 6–8 weeks. We used 6–8 weeks as the main
   post-ICI initiation time point because[84]^1 (1) previous studies
   suggested that it provided more useful comparison with pretreatment BL
   than do earlier post-ICI initiation time points,[85]^12 and[86]^2
   mosti(2) rAEs occur after this time,[87]^10 26 thereby rendering this
   time point potentially useful for prediction of future events.

   Blood samples were centrifuged at 3,000 rpm at 4°C for 15 min to obtain
   plasma and isolate peripheral blood mononuclear cells (PBMCs). For the
   selection of cases for analysis by varying laboratory methodologies, we
   used the following guiding principles: (1) including the greatest
   number of cases possible within financial constraints; (2) maximizing
   overlap of cases across methodologies, and (3) minimizing overlap of
   cases when validating across methodologies.

Single-cell RNA sequencing and data analysis

   We produced scRNA-seq data on a total of ∼400,000 cells across 72
   samples. All 72 PBMC patient samples in this study were processed for
   scRNA-seq on a 10X Chromium instrument. Integrating 72 samples for
   batch correction to identify shared cell states across samples was
   performed using Harmony. We then performed normalization and
   dimensionality reduction analysis using SCTransform on our integrated
   dataset to eliminate any technical variation and identify different
   cell states. We performed unsupervised clustering and identified 22
   different cell clusters in our dataset across 72 samples. Unsupervised
   clustering of scRNA-seq data was performed using Seurat packages in R,
   Cellenics and dseqr. We further used the Azimuth PBMC dataset as a
   reference to identify immune cell subsets at a much higher resolution
   using the top 30 principal components and performed pseudobulk
   differential gene expression analysis using dseqr with false discovery
   rate (FDR) <0.05. We further performed additional unsupervised
   sub-clustering analysis of B cells, T/natural killer (NK) and myeloid
   cells. Additional details on scRNA-seq and data analysis are provided
   in [88]online supplemental methods.

Cytokine/chemokine and ANA analysis

   Determination of cytokine and chemokine levels was performed using
   Bio-Plex Pro Human Chemokine 40-plex Panel (Bio-Rad Laboratories,
   Hercules, California, USA) according to the manufacturer’s instructions
   using a Luminex 200 System. The list of cytokines and chemokines is
   provided in [89]online supplemental table 1. Bio-Plex Manager V.6.1
   software was used for data analysis. Concentrations of cytokines and
   chemokines (pg/mL) were determined based on the fit of a standard curve
   for mean fluorescence intensity versus pg/mL. We assessed the plasma
   levels of antinuclear IgG antibodies using QUANTA lite ANA assay (INOVA
   diagnostics). All plasma samples were run in duplicates, and analysis
   was performed using GraphPad Prism. Results are expressed as the
   mean±SE.

Mass cytometry

   A panel of 35 antibodies (metal isotope-labeled conjugates, Maxpar
   Direct Immune Profiling Assay Panel by Fluidigm) was used for staining
   Cryopreserved PBMCs derived from patients receiving ICI therapy. Cells
   were analyzed on a Helios mass cytometer (Fluidigm). Data were
   normalized and analyzed with gating on live, singlet CD45^+ cells using
   the OptSNE and FlowSOM using OMIQ.ai (Omiq, Santa Clara, California,
   USA). We then performed unsupervised consensus meta-clustering using
   Euclidean as a distance metric and Max Equal sampling of all 110
   samples. Statistical analysis of cell abundance was performed using
   EdgeR in OMIQ or GraphPad Prism software. CyTOF data was also analyzed
   using a manual gating strategy.

Statistical analysis

   Statistical significance was calculated using IBM SPSS Statistics
   software (V.29) and GraphPad Prism Software (GraphPad Software, La
   Jolla, California, USA). Student’s t-test, Welch’s t-test, the
   Mann-Whitney U test, Wilcoxon matched-pairs signed-rank test, and Edge
   R were used when applicable. For toxicity analyses, because irAEs may
   occur throughout the course of treatment with ICI, patients were
   considered not to have developed irAEs only if they had been followed
   without evidence of toxicity for at least 6 months. Results are
   expressed as the mean±SE. To perform multimarker analysis, we conducted
   logistic regression analysis using the ‘glm’ function, and calculated
   the area under the receiver operating curve using the ‘pROC’package in
   R (V.4.3.2) programming. Spearman correlation analysis and calculation
   of Spearman rank-weighted composite scores were performed in R using
   dplyr, pROC and ggplot packages. Unless specified, for all statistical
   analysis, the level of significance was set at p<0.05. For all gene
   expression analysis results presented the level of significance was set
   to adjusted p<0.05 and minimum fold change ≥1.5.

Results

Clinical characterization of patients and study design

   A total of 162 patients were included in the study. [90]Figure 1
   outlines the overall experimental strategy and patient cohorts included
   in each analysis, including cytokines/chemokines (n=146), ANA (n=65),
   CyTOF (n=55), and scRNA-seq (n=36). The overall numbers and selection
   of cases for each assay reflect cost/resources, specimen availability,
   and an effort to maximize the overlap of cases across assays. [91]Table
   1 provides demographic, cancer, and treatment information for included
   cases, of which 93 (57%) had grade ≥2 irAE.

Figure 1. Study design. The study included 162 patients with cancer receiving
ICI therapy (either anti-PD1, anti-PDL1 or combination anti-PD1 plus
anti-CTLA4). Blood was collected at baseline before the initiation of ICI
therapy and post-immunotherapy (2–3 weeks and 6–8 weeks). PBMCs were used for
single-cell RNA sequencing (scRNA-seq) and mass cytometry assays. Plasma was
used to assess levels of antinuclear antibodies (ANA) and 40 cytokines. The
number of patients used for each assay is shown in the brackets. BL,
baseline; CTLA4, cytotoxic T lymphocyte antigen 4; ICI, immune checkpoint
inhibitor; PBMC, peripheral blood mononuclear cell; PD1, programmed death 1;
PDL1, programmed death ligand 1; t-SNE, t-distributed stochastic neighbor
embedding.

   [92]Figure 1
   [93]Open in a new tab

Table 1. Characteristics of patients included in the study (N=162).

             Characteristic           Number (%)
   Sex
    Female                            58 (36)
    Male                              104 (64)
   Race—ethnicity
    Asian                             4 (2)
    Black or African American         12 (7)
    Hispanic white                    12 (7)
    Non-Hispanic white                126 (78)
    Unknown                           8 (5)
   Cancer type
    Non-small cell lung cancer        81 (50)
    Melanoma                          38 (23)
    Head and neck                     11 (7)
    Kidney                            9 (6)
    Other[94]^*                       23 (14)
   ICI therapy
    PD1/PDL1                          136 (84)
    CTLA4                             1 (1)
    PD1/PDL1+CTLA4                    21 (13)
    Other PD1/PDL1 combinations[95]^† 4 (2)
   [96]Open in a new tab
   ^*

   Includes breast cancer (n=1), cholangiocarcinoma (n=2), cutaneous
   squamous cell carcinoma (n=1), glioblastoma multiforme (n=2),
   hepatocellular carcinoma (n=2), mesothelioma (n=3), pancreas cancer
   (n=2), sarcoma (n=1), small cell lung cancer (n=6), and
   urothelial/bladder cancer (n=2).
   ^†

   Includes CSF1R (n=3) and NKG2A (n=1).

   CTLA4, cytotoxic T lymphocyte antigen 4; ICI, immune checkpoint
   inhibitor; PD1, programmed death 1; PDL1, programmed death ligand 1.

Baseline abundance of immune cell subsets associated with irAEs

   We performed 5' scRNA-seq using the 10X chromium immune profiling assay
   on paired BL and 6–8 week samples from 40 patients, of whom 30 had
   experienced various types of grade ≥2 irAE. After applying quality
   control filters, 36 matched sample pairs (BL and 6–8 weeks) yielded
   high-quality transcriptomes, of which four were technical replicates (1
   no-irAE and 3 irAE) ([97]figure 2A). Following data integration,
   normalization, and dimensionality reduction, an initial unsupervised
   clustering analysis at a lower resolution identified 23 distinct cell
   clusters across samples. These clusters represented all major immune
   cell lineages ([98]figure 2B). The proportion of these 23 unsupervised
   Louvain clusters is shown in [99]figure 2C, and the expression of
   canonical marker genes is presented in [100]figure 2D. Cell annotations
   based on the expression of top five marker genes and a heat map of top
   marker genes for each unsupervised cluster are provided in [101]online
   supplemental figures 1 and 2.

Figure 2. Single-cell RNA sequencing (scRNA-seq) analysis of PBMCs at
baseline and post-ICI therapy. (A) scRNA-seq experimental design for 72 blood
samples (36 pairs) from patients with cancer at baseline (BL) before the
initiation of ICI therapy and 6–8 weeks after ICI initiation. (B) UMAP of
23 cell states identified by unsupervised clustering shown according to cell
type. (C) Frequency plot showing the proportion of 23 clusters identified by
unsupervised clustering in irAE versus no-irAE group at BL and 6–8 weeks
post-ICI therapy. (D) Dot plot showing canonical marker gene expression for
23 unsupervised clusters. Arrows point to cluster 2 (NK) and cluster 18
(Mki67+). (E) The abundance of cluster 18 (Mki67+), shown as a percentage of
total PBMCs in two groups at baseline (BL) identified by unsupervised
clustering. (F) UMAP of PBMCs annotated by Azimuth Reference shown by major
immune cell types and 30 immune cell subsets. The abundance of NK (CD56dim)
cells (G), CD4 proliferating T cells (H), plasmablasts (I), and combined
proliferating lymphocyte cluster (CD4, CD8 and NK) (J) shown as a percentage
of total PBMCs in No irAE versus irAE group at BL guided by Azimuth
Reference. ASDC, AXL+SIGLEC6+ dendritic cell; cDC, conventional dendritic
cell; CTLA4, cytotoxic T lymphocyte antigen 4; DC, dendritic cell; dnT,
double-negative T cell; gdT, gamma delta T cell; HSPC, hematopoietic stem and
progenitor cell; ICI, immune checkpoint inhibitor; irAE, immune-related
adverse event; MAIT, mucosal-associated invariant T cell; NK, natural killer
cell; PBMC, peripheral blood mononuclear cell; PD1, programmed death 1; pDC,
plasmacytoid dendritic cell; PDL1, programmed death ligand 1; TCM, central
memory T cell; TEM, effector memory T cell; Treg, T regulatory cell; UMAP,
uniform manifold approximation and projection.

   [102]Figure 2
   [103]Open in a new tab

   Frequency calculations as a percentage of total PBMCs revealed a
   significant increase in the abundance of proliferating lymphocyte
   cluster (cluster 18, characterized by Mki67^+ cells) ([104]figure 2E).
   Conversely, the NK cell cluster (cluster 6) was reduced in irAE
   patients. In addition, cluster 22, which resembled plasma
   cell/plasmablast-like clusters showed increased frequency in irAE
   patients at BL.

   To confirm these findings, we performed a reference-guided cell
   annotation analysis with Azimuth on the integrated datasets of 72
   patient samples ([105]figure 2F). Consistent with the unsupervised
   clustering results, we identified three cell clusters significantly
   correlated with the irAE development. The NK cell cluster—specifically
   CD56^dim CD16^+ NK cells—was significantly less abundant at BL in
   patients with irAE, whereas CD4 proliferating T cells and plasmablasts
   were significantly more abundant ([106]figure 2G–I). When combining
   Azimuth-guided CD4, CD8, and NK proliferating clusters, the BL
   aggregated proliferating lymphocyte cluster was most strongly
   correlated with irAE ([107]figure 2J), mirroring the unsupervised
   clustering data and demonstrating an activated immune state in patients
   predisposed to irAE. Next, we evaluated whether the BL abundance of
   this proliferating lymphocyte cell cluster might serve as a candidate
   biomarker. Receiver operating characteristic (ROC) analysis yielded an
   area under the curve (AUC) of 0.78 (p<0.0001) for the CD4 proliferating
   cluster, indicating its potential clinical utility in predicting future
   irAE occurrence.

Mass cytometry analysis reveals distinct immune subsets associated with irAEs

   We performed CyTOF analysis on PBMC samples from an expanded cohort of
   55 patients (43 irAE, 12 no‐irAE) with paired BL and 6–8 weeks post-ICI
   time points using a panel of 36 antibodies ([108]figure 3A).
   Metaclustering analysis yielded 30 distinct immune cell
   clusters—including subsets of B cells, CD4^+ T cells, CD8^+ T cells, NK
   cells, and myeloid cells ([109]figure 3B,C). A corresponding heat map
   of cell surface marker expression is shown in [110]figure 3D.

Figure 3. Mass cytometry (CyTOF) analysis of PBMC samples at baseline and
post-ICI therapy. (A) CyTOF experimental design of 110 blood samples (55
pairs) from 55 patients with cancer (n=12 no-irAE and n=43 irAE) at baseline
(BL) before the initiation of ICI therapy and 6–8 weeks post immunotherapy.
Opt-SNE of 30 cell clusters identified by unsupervised consensus
meta-clustering of CYTOF data using Euclidean as a distance metric is shown
according to major immune cell types (B) and 30 immune cell subsets (C). (D)
Heatmap showing expression of 35 cell surface immune markers used in the
CyTOF panel in 30 metaclusters. (E) Volcano plot of CyTOF data showing
significant metaclusters (aqua circles) differentially abundant in irAE
versus no-irAE at baseline with (p<0.05). The abundance of significant
clusters; K10, plasmablast (F); K2, NK (G); K4, CD8 terminal effector (H);
K8, TCRγδT (I) and K13, CD8 terminal effector (J), identified by unsupervised
clustering of CyTOF data using Edge R (p<0.05) shown as a percentage of total
PBMCs in no-irAE versus irAE cases at BL. The abundance of significant
clusters, including NKT (K) and CD57 positive clusters (L) identified by
manual gating strategy in irAE versus no-irAE cases. (M) The abundance of
CD8+ KLRG1+ CCR7 CD8 effector memory cluster identified by unsupervised T/NK
subclustering analysis of scRNA-seq data. Significant increase in abundance
of cluster K21 (CD4 memory) in irAE and no-irAE cases postimmunotherapy in
all cases (N) and in melanoma cases (O) by Wilcoxon matched-pairs signed-rank
test of CyTOF data. CTLA4, cytotoxic T lymphocyte antigen 4; FC, fold change;
FDR, false discovery rate; ICI, immune checkpoint inhibitor; irAE,
immune-related adverse event; NK, natural killer cell; PBMC, peripheral blood
mononuclear cell; PD1, programmed death 1; pDC, plasmacytoid dendritic cell;
PDL1, programmed death ligand 1; TCM, central memory T cell; TCR, T cell
receptor; scRNA-seq, single-cell RNA sequencing.

   [111]Figure 3
   [112]Open in a new tab

   We next calculated the abundance of each cell cluster as a percentage
   of total CD45^+ cells. EdgeR analysis comparing irAE and no‐irAE groups
   at BL identified 11 significant clusters in OMIQ with p<0.05
   ([113]figure 3E). Specifically, seven clusters (K4-CD8, K8-TCRγδ, K9-B
   cell, K10-plasmablast, K13-CD8, K17-TCRγδ, K25-CD8) were increased in
   irAE patients, whereas four clusters (K2-NK, K16-CD123^+, K28-CD8,
   K29-CD4) were decreased. Consistent with our scRNA-seq findings,
   plasmablasts (cluster K10) were elevated (p=0.01) while NK cells
   (cluster K2, CD56^+CD57^−) were reduced (p=0.007) in the irAE group
   ([114]figure 3F,G). Further stratification by cancer and therapy type
   revealed that clusters K4, K8, and K13 exhibited the most significant
   differences in irAE patients ([115]figure 3H–J). Clusters K4
   (CD8^+CCR7^−CD45RA^+CD27^−CD28^−CD57^+) and K13
   (CD8^+CCR7^−CD45RA^+CD27^−CD28^−CD57^−) resembled terminal effector
   memory CD8^+ T cells, while cluster K8 comprised double-negative
   TCRγδ^+CD57^+ cells. Extended CyTOF analyses by therapy and cancer type
   are presented in [116]online supplemental figure 3 (non-small cell lung
   cancer, melanoma, and PD1/PDL1 therapy). Although the limited number of
   patients receiving combination immunotherapy precluded detailed
   analysis of that subgroup, the directionality and magnitude of subgroup
   findings largely mirrored those of the overall cohort shown in
   [117]figure 3.

   For additional rigor, we performed manual gating analysis of the CyTOF
   data. We identified a greater BL abundance of the NKT cell population
   ([118]figure 3K) and a striking increase in CD57^+ T cells, including
   CD57^+CD8^+ and CD8^+CD57^+ NKT populations in irAE patients
   ([119]figure 3L). The manual gating strategy is shown in [120]online
   supplemental figure 4. Of note, the two highly correlated clusters
   identified by metaclustering were CD8 terminal effector cell cluster K4
   and TCRγδ cluster K13, which were both CD57^+. As reported by other
   groups,[121]^13 we also observed a greater abundance of
   HLA-DR^+CD38^+CD4^+ memory T cells in irAE patients at BL. To validate
   the increase in CD8^+ effector cells in scRNA-seq data, we performed
   high-resolution subclustering of CD8^+ and NK cells, which revealed a
   significantly higher abundance of CD8^+KLRG1^+CCR7^− effector memory
   clusters in irAE patients ([122]figure 3M). Collectively, our CyTOF
   analysis at BL revealed a T cell profile skewed toward an activated
   effector/terminal effector phenotype (CD57^+), increased plasmablasts,
   and reduced CD56^dim CD16^+ NK cells in patients with irAEs.

   Next, we investigated treatment-induced changes by comparing the
   abundance of metaclusters between irAE and no‐irAE groups at 6–8 weeks
   post-ICI initiation. Following ICI initiation, the abundance of CD8^+
   effector clusters (K4 and K13) and TCRγδ cluster K8 remained
   significantly higher in irAE patients, whereas the abundance of NK cell
   cluster K2 remained significantly lower. CD57^+ T and NKT cell
   populations remained significantly higher in the irAE group
   post-therapy. Using the Wilcoxon matched-pairs signed-rank test, we
   further assessed longitudinal changes in the abundance of CyTOF
   metaclusters relative to BL within each group. This analysis revealed a
   significant increase in the CD4^+ memory cluster (K21,
   CD4^+CD45RO^+CD45RA^−CD27^+CD28^+CD57^−) exclusively in irAE cases
   (p=0.003 vs p=0.38 in no‐irAE), particularly in patients with melanoma
   ([123]figure 3N,O). After ICI initiation, several populations decreased
   including memory B cells (cluster K6), exhausted T cells
   (CD57^+PD1^+CD4^+ cluster K22 and CD57^+PD1^+CD8^+ cluster K5), NKT
   cells (CD8^+CD161^+CD57⁻ cluster K28), and an intermediate naïve/memory
   CD8^+ cluster (K18) ([124]online supplemental figure 5A–E).
   Interestingly, these decreases were significant only in irAE cases,
   suggesting an enhanced activated state post-therapy. In summary, our
   CyTOF analysis revealed an activated and skewed T cell compartment in
   irAE patients at BL characterized by an effector/terminal effector
   phenotype with increased plasmablasts and reduced NK cells.

Baseline proinflammatory autoimmune-like state associated with irAEs

   Because our scRNA-seq and CyTOF analyses revealed a high BL abundance
   of plasmablasts in patients with irAEs, we further examined the B cell
   compartment by performing differential gene expression and pathway
   analysis between irAE and no-irAE groups for naïve and memory B cells
   but observed no significant changes in major pathways before or after
   therapy. Consistent with earlier work by Das and colleagues,[125]^17 we
   also observed a significant increase in plasmablasts in patients with
   irAEs treated with combination (CTLA4+PD1) therapy. Additionally, we
   found that increased BL abundance of plasmablasts was most strongly
   associated with high-grade (grade ≥3) irAEs ([126]figure 4A).

Figure 4. Baseline autoimmune-like, proliferative and inflammatory state
associated with irAE development. (A) Abundance of plasmablasts as a
percentage of total PBMCs at baseline (BL) according to the grade of irAE.
(B) Comparison of plasma concentration of antinuclear antibodies (ANA) (N=65
total; n=22 no-irAE and n=43 irAE) at BL and postimmunotherapy (at 2–4 weeks
and 6–8 weeks). (C) Comparison of plasma concentration of ANA in 65 patients
based on irAE grade at BL and postimmunotherapy. G0-1, Grade 0–1; G2, Grade
2; G3, Grade 3. (D) Abundance of CXCR3+Mki67+ cell cluster as a percentage of
total PBMCs at BL based on the grade of irAE. (E) Pseudobulk scRNA-seq
differential gene expression analysis of T and NK cells in the irAE group
compared with the no-irAE group at baseline is shown as a volcano plot.
Significantly upregulated genes are shown in blue dots, significantly
downregulated genes in red dots, and genes with no significant change are
shown in gray (F) Heat map depicting logFC and p values for the 3-gene
signature in T and NK cells comparing irAE versus No irAE group at baseline
(adjusted p<0.05 and ≥2 fold change). (G) MsigDB pathway enrichment analysis
for T/NK cells at baseline. Red arrow shows the TNFα signaling as the highly
enriched pathway in irAEs at baseline. (H) GSEA enrichment analysis for T/NK
cells showing TNF signaling as the most enriched pathway at baseline.
Abundance of CXCR3+Mki67+ cluster (I) and CXCR3+CD4+ memory cluster (J)
identified by high-resolution unsupervised clustering of scRNA-seq data. (K)
Abundance of CXCR3+CD8+ memory clusters in the scRNA-seq data at baseline.
(L) CXCR3 gene expression in T and NK cells at baseline. Abundance of
CXCR3+CD8+ clusters (M) and CXCR3+CD4+ (N) memory clusters at baseline
identified in the CyTOF data. FC, fold change; GSEA, gene set enrichment
analysis; irAE, immune-related adverse event; NK, natural killer cell; PBMC,
peripheral blood mononuclear cell; scRNA-seq, single-cell RNA sequencing.

   [127]Figure 4
   [128]Open in a new tab

   Given these findings, we next assessed plasma ANA levels—a non-specific
   clinical biomarker for systemic lupus erythematosus and other
   autoimmune diseases[129]^27—in 65 patients across time points
   ([130]figure 4B). BL ANA levels were significantly higher in patients
   who subsequently developed irAEs compared with those without irAEs, and
   levels were highest for grade ≥3 irAE cases at all time points
   ([131]figure 4C). To determine whether this autoimmune-like state
   correlates with a more proliferative and proinflammatory phenotype, we
   performed a grade-based analysis of the proliferating lymphocyte
   cluster (Mki67^+) identified in the scRNA seq data. This highly
   proliferative lymphocyte cluster, which was CXCR3^+HLADR^+, was most
   significantly associated with higher-grade irAEs ([132]figure 4D).

   Next, we evaluated gene expression changes in T and NK cells at BL.
   Pseudobulk differential analysis between irAE and no irAE cases
   revealed significant changes (adjusted p<0.05 and fold change ≥1.5) in
   over 300 genes at BL ([133]figure 4E). Notably, IFNG, TNF, and CXCR3
   were among the top upregulated genes in patients who developed irAEs
   ([134]figure 4F). Gene set enrichment analysis (GSEA) and MsigDB
   Pathway enrichment identified TNF signaling as the most significantly
   upregulated pathway in T cells of irAE patients at BL ([135]figure
   4G,H). In contrast, genes associated with T cell inhibition and
   immunosuppression (CXCR4 and SIGLEC7) were downregulated in irAE
   patients. Interestingly, the autoimmune regulator (AIRE) gene was
   significantly upregulated (fold change >2; adjusted p=0.04) after ICI
   treatment only in the no irAE group, suggesting a potential role in
   enhanced self‐tolerance and prevention of irAEs.

   Because CXCR3 is a homing receptor for CXCL9, CXCL10, and CXCL11 and is
   expressed on activated/effector T cells with a propensity to migrate to
   inflammatory sites, we examined our scRNA-seq data at a higher
   resolution and performed additional manual gating of the CyTOF data.
   This higher-resolution unsupervised analysis of PBMC scRNA-seq
   replicated our earlier finding of an increased
   Mki67^+CXCR3^+HLADR^+cells in irAE patients and also identified a
   higher abundance of the CXCR3^+CD4^+IL32^+ memory cluster associated
   with irAEs, thereby supporting the increased CXCR3 gene expression data
   ([136]figure 4I,J). Subclustering analysis of CD8/NK/Mki67 clusters
   also showed an increased abundance of activated CXCR3-expressing CD8
   memory clusters expressing CXCR3 in the irAE patients at BL
   ([137]figure 4K). An overall increase in CXCR3 gene expression in
   scRNA-seq data at BL is depicted in [138]figure 4L. Manual gating of T
   cell subsets in the CyTOF data likewise revealed a higher abundance of
   CXCR3^+CD4^+ and CXCR3^+CD8^+ effector memory cells in irAE patients at
   BL ([139]figure 4M,N).

Distinct features of myeloid cells and cytokine/chemokines associated with
irAEs

   Unsupervised clustering analysis of myeloid cells in PBMC scRNA-seq
   data ([140]figure 2B) revealed eight distinct myeloid subsets (Clusters
   6, 9, 10, 11, 13, 17, 19, and 21) defined by the expression of
   canonical myeloid markers. These included six monocyte clusters—two
   CD16^+ clusters (10 and 19), three CD14^+ clusters (6, 9, and 13), one
   intermediate monocyte cluster,[141]^11 a plasmacytoid dendritic cell
   (DC) cluster[142]^21 and a conventional DC cluster.[143]^17 Pseudobulk
   differential single-cell gene expression analysis revealed significant
   changes in over 250 genes (adjusted p<0.05 and fold change ≥1.5),
   including upregulation of IL1B and CXCL8 and downregulation of the
   immunosuppressive gene SIGLEC-10 and complement genes C1QA, C1QB, and
   C1QC in irAE patients at BL ([144]figure 5A). GSEA and MsigDB pathway
   enrichment identified TNF signaling and inflammatory response as the
   two significantly upregulated pathways in irAE patients at BL
   ([145]figure 5B,C). A heatmap of selected upregulated genes (adjusted
   p<0.05 and fold change ≥1.5) involved in TNF signaling and inflammatory
   responses is shown in [146]figure 5D.

Figure 5. Single-cell RNA sequencing analysis of myeloid cells in patients
with and without irAE. (A) Pseudobulk differential expression analysis of
unsupervised myeloid cell clusters of scRNA-seq data in irAE versus no-irAE
group at baseline is shown as a volcano plot (adjusted p<0.05 and fold change
≥1.5). Significantly upregulated genes are shown in blue dots, significantly
downregulated genes in red dots, and genes with no change are shown in gray.
Gene set enrichment analysis (GSEA) in panel B and MsigDB pathway enrichment
in panel C of myeloid cells at baseline in irAE patients. The red arrow in
panel C shows the TNF signaling and inflammatory response pathways as highly
enriched in myeloid cells in the irAE group at baseline. (D) Heatmap of
selected upregulated genes in myeloid cells in irAE versus no-irAE at
baseline (adjusted p<0.05 and fold change ≥1.5). (E) Gene expression changes
of CXCL10, CCL3, CCL4, SPARC and SIGLEC1 in myeloid cells in irAE versus
no-irAE cases post-therapy (adjusted p<0.05 and fold change ≥1.5). (F) Gene
expression changes of TNF and IRF4 in monocytes in no-irAE cases post-therapy
compared with baseline (adjusted p<0.05 and fold change ≥1.5). (G) Heatmap of
selected upregulated and downregulated genes in Azimuth-guided CD14+
classical monocytes in irAE versus no-irAE at baseline (adjusted p<0.05 and
fold change ≥1.5). (H) Gene Ontology analysis in CD14+ classical monocytes in
irAE versus no-irAE at baseline. (I) Heatmap of selected upregulated genes in
Azimuth-guided CD14+ and CD16+ classical monocytes in irAE versus no-irAE at
post-therapy (adjusted p<0.05 and fold change ≥1.5). (J) Heatmap of selected
upregulated genes in Azimuth-guided CD14+ and CD16+ classical monocytes in
no-irAE at post-therapy compared with baseline (adjusted p<0.05 and fold
change ≥1.5). (K) Unsupervised subclustering analysis of all myeloid cells
(without pDCs). (L) Dot plot showing gene expression markers for 16
unsupervised Louvain clusters for myeloid cells. (M) Significantly
upregulated cytokines/chemokine ligands and their receptors at baseline in
irAE patients compared with no-irAE patients at baseline (adjusted p<0.05 and
fold change ≥1.5). (N) Trajectory analysis of myeloid cells using Monocle 3.
(O) Continuous embedding depicting the fraction of myeloid cells expressing
CXCL8 and IL1B genes and their corresponding gene expression changes between
irAE and no-irAE patients at baseline in myeloid cells. (P) Abundance of
TGFBIhigh cluster 13 in irAE versus no-irAE patients at baseline shown as a
percentage of total myeloid cells. BL, baseline; FC, fold change; IFN,
interferon; irAE, immune-related adverse event; pDC, plasmacytoid dendritic
cell; scRNA-seq, single-cell RNA sequencing; UMAP, uniform manifold
approximation and projection.

   [147]Figure 5
   [148]Open in a new tab

   At the post-ICI initiation time point, we observed significant
   upregulation of chemokine genes (CXCL10, CCL3, and CCL4) in CD16^+
   monocytes from irAE patients, whereas the immunosuppressive gene SPARC
   was downregulated ([149]figure 5E). Intriguingly, TNF was significantly
   upregulated at 6–8 weeks post-therapy compared with BL exclusively in
   no irAE patients ([150]figure 5F). To further validate these findings
   and elucidate the role of monocytes, we performed an Azimuth
   reference-guided analysis of peripheral blood monocytes, which
   identified two clusters: CD14^+ classical monocytes and CD16^+
   non-classical monocytes. Differential single-cell gene expression
   analysis between irAE and no‐irAE cases in these monocyte clusters
   revealed significant changes in over 250 genes in classical monocytes
   at BL, including upregulation of proinflammatory genes IL1B, TNF, and
   CXCL8. TNF signaling emerged as the top enriched pathway in irAE
   patients, mirroring our unsupervised clustering results. Gene Ontology
   analysis further showed significant upregulation of genes involved in
   inflammatory responses, leukocyte/monocyte chemotaxis, and migration in
   CD14^+ monocytes, while type I interferon (IFN) signaling and
   complement genes were downregulated in irAE patients at BL ([151]figure
   5G,H). Post-ICI initiation, we detected over 450 differentially
   expressed genes in classical monocytes and more than 150 genes in
   non-classical monocytes, including upregulation of several chemokines
   involved in recruiting T cells and monocytes to sites of organ
   inflammation (among them CCL3, CCL4, and CXCL10) in irAE patients
   following ICI treatment. A heatmap depicting significant gene
   expression between irAE and no irAE patients at BL is shown in
   [152]figure 5I. Notably, patients without irAE displayed elevated
   expression of a distinct set of genes involved in macrophage
   migration/chemotaxis post-ICI treatment including SLAMF8, CCR2, MST1,
   PTK2, C5, and CCL5 in CD14^+ monocytes, and TNF, HES1, CCL5, KLF13, and
   NOTCH1 in CD16^+ monocytes, indicative of a pro-tumorigenic myeloid
   reprogramming ([153]figure 5J).

   To gain a deeper understanding of myeloid cell involvement in irAEs, we
   performed higher-resolution unsupervised subclustering analysis of
   myeloid cells, which revealed 16 distinct clusters ([154]figure 5K,L).
   As before, the irAE patients showed significant BL upregulation of IL1B
   and CXCL8, with increased CCL3, CCL4, and CXCL10 post-therapy.
   Examining all cytokines/chemokine ligands and their receptors at BL
   revealed IL1B, CCL3L1, CCR2, CXCL2, CXCL3, CXCL8, IL6R, IL6ST, TNFSF14
   and TNFRSF10D significantly upregulated in irAE patients ([155]figure
   5M). [156]F[157]igure 5N–[158]P displays trajectory analyses, which
   indicated that (1) CD14^+ monocytes differentiate into three distinct
   branches: DC-like (C6), monocyte/macrophage-like (C5, C13), and the
   most differentiated CD16^+CD14^− state (C0, C14, C15); (2) differential
   expression of IL1B and CXCL8 in myeloid clusters according to irAE
   development; and greater abundance of the TGFBI^high cluster (C13) in
   no-irAE patients.

   Given this enhanced cytokine and chemokine gene expression profile in
   irAE patients, we next evaluated plasma levels of 40 cytokines and
   chemokines using a multiplex cytokine panel in 146 patients at three
   time points: pretreatment BL, 2–3 weeks, and 6–8 weeks post-ICI
   initiation. As previously reported by our group and others,[159]^12 28
   IFNγ-inducible chemokines CXCL9 and CXCL10 were elevated in both irAE
   and no-irAE groups following ICI therapy relative to BL. However, irAE
   patients showed a significantly greater increase in plasma CXCL10 at
   6–8 weeks compared with the no-irAE group ([160]figure 6A). Similarly,
   CXCL9 levels displayed a more pronounced fold induction in irAE
   patients at 6–8 weeks ([161]figure 6B). This heightened induction of
   CXCL9 and CXCL10 was observed in patients receiving both single-agent
   and combination therapy ([162]figure 6C–F). In contrast, serum CXCL8
   levels at 6–8 weeks post-ICI were significantly lower in irAE patients
   than in those without irAEs ([163]figure 6G–I). Notably, changes in
   CXCL8, CXCL9, and CXCL10 in serum reflected their respective gene
   expression levels in myeloid cells ([164]figure 6J), implying that
   circulating monocytes may partially account for the source of these
   serum cytokines.

Figure 6. Multiplex cytokine and chemokine analysis in patients with and
without irAEs. Comparison of plasma concentrations of chemokines at baseline
(BL) and postimmunotherapy (6–8 weeks). CXCL10 (A) and CXCL9 (B) in all
patients. CXCL10 (C) and CXCL9 (D) in PD1/PDL1-treated patients. CXCL10 (E)
and CXCL9 (F) in combination (CTLA4+PD1)-treated patients. CXCL8 in all
patients (G), in PD1/PDL1-treated patients (H) and in combination
(CTLA4+PD1)-treated patients (I). Dot plot showing comparison of relative
gene expression values of chemokines CXCL10, CXCL9 and CXCL8 at baseline (BL)
and postimmunotherapy (6–8 weeks) in myeloid cells in scRNA-seq data (J).
Number of patients for comparison of plasma concentration of chemokines: All
patients (N=146 total: n=79 irAE and n=67 no-irAE); PD1/PDL1-treated patients
(N=122 total: n=59 No irAE and n=63 irAE); patients treated with combination
CTLA4+PD1 therapy (N=18 total: n=6 No irAE and n=12 irAE). CTLA4, cytotoxic T
lymphocyte antigen 4; irAE, immune-related adverse event; PD1, programmed
death 1; PDL1, programmed death ligand 1; scRNAseq, -seq, single-cell RNA
sequencing.

   [165]Figure 6
   [166]Open in a new tab

Multimarker analysis

   For both scRNA-seq and CyTOF, we integrated the individual markers with
   strongest association with irAE and performed logistic regression
   analysis. Because only 17 patients had both scRNA seq and CyTOF data,
   we were not able to combine both datasets into a single multimarker
   analysis. Incorporating (1) CD4-CD8-NK proliferating, (2) NK, and (3)
   plasmablast markers, the scRNA seq multimarker had an AUC of 0.78
   ([167]online supplemental figure 6A). Incorporating CD57+NKT, CD57+CD4,
   CD57+CD8, K4 (CD8^+CCR7^−CD45RA^+CD27^−CD28^−CD57^+), K8
   (double-negative TCRγδ^+CD57^+), and K13
   (CD8^+CCR7^−CD45RA^+CD27^−CD28^−CD57^−) populations, the CyTOF
   multimarker had an AUC of 0.79 ([168]online supplemental figure 6B).
   For 17 patients that had both ScRNA-seq and CyTOF datasets, we
   performed Spearman correlation analysis and ranked the immune variables
   by their Spearman correlation with irAE development. We then calculated
   Spearman rank-weighted composite score using the immune features that
   showed the strongest associations with irAEs and ran AUC/ROC analysis.
   Spearman rank-weighted composite scores combining CD4-CD8-NK
   proliferating, NK, plasmablast and CD57NKT markers showed AUC of 0.865
   and a p value of 0.03 ([169]online supplemental figure 6C,D).

Validation in external datasets

   We accessed publicly available external datasets by Lozano et al and
   Nuñez et al[170]^13 29 to validate our scRNA seq and CyTOF findings,
   respectively. Interpreting these analyses requires consideration of
   their features and limitations. The Lozano dataset is small (n=13),
   limited to melanoma cases, and observed irAE only in patients that
   received combination ICI. The Nuñez dataset lacks PBMC scRNA seq data
   and lacks a CD20 marker in the CyTOF panel (thereby precluding
   identification of plasmablasts). For scRNA-seq markers, we validated
   the association of lower NK cells, higher CD4 proliferating cells, and
   higher CD8 terminal effector cells at BL in patients with severe (grade
   ≥3) irAEs ([171]online supplemental figure 7A–C). For CyTOF, we
   validated the association of higher TCRγδ T cells, higher CD8 terminal
   effector cells, and lower NK cells at BL with severe (grade ≥3) irAEs
   ([172]online supplemental figure 8A–D).

Discussion

   Despite widespread recognition and attention to irAE, the mechanisms
   underlying these autoimmune ICI toxicities remain poorly understood, as
   do reliable means of predicting and even diagnosing them. To address
   these issues and knowledge gaps, we conducted a comprehensive
   multidimensional cellular and molecular analysis of the host peripheral
   immune system in a large and diverse cohort of ICI-treated patients.
   Using various assays on longitudinal peripheral blood samples, our
   study revealed several key findings. Patients who eventually developed
   irAEs had (1) an increased BL abundance of plasmablasts and
   significantly elevated plasma ANA levels; (2) a skewed BL T cell
   profile characterized by increased proliferative and activated CXCR3^+
   T cells, CD57^+ T cells and NKT cells along with a lower BL abundance
   of CD56^dim CD16^+ NK cells; (3) distinct upregulated gene signatures
   at BL in T/NK cells (CXCR3, TNF and IL1B) and in myeloid cells (IL1B
   and CXCL8) coupled with enhanced TNF signaling and higher fraction of
   IL1B-positive myeloid cells; and (4) a significant shift post-therapy
   towards an activated and inflammatory immune state, marked by increased
   CD4 memory populations, upregulated IL18 and significantly elevated
   IFN-γ inducible proinflammatory chemokines CXCL9 and CXCL10 coupled
   with reduced CXCL8 levels in myeloid cells. In contrast, patients
   without irAE exhibited an immunosuppressive profile characterized by
   increased BL abundance of a TGFBI^high myeloid cluster and
   post-treatment myeloid protumorigenic reprogramming. Together, these
   findings may suggest that underlying subclinical autoimmunity and a
   proliferative activated proinflammatory immune state predispose
   patients to irAEs.

   Our study is notable for the large size of the study cohort (which
   permitted subset analyses for some assays) and the comprehensive nature
   of the correlative studies performed (allowing cross-validation of
   findings across orthogonal methodologies). Additional strengths include
   the quality of clinical data (particularly irAE occurrence, grade, and
   type), the use of real-world cases, the focus on pretreatment BL blood
   samples (the results of which could be incorporated into the upfront
   selection of patients, treatment regimens, and monitoring strategies),
   and validation of certain assay results in external datasets. The
   numerous novel observations in our present study could reflect the
   relatively large sample size, the diversity of cancer types, and the
   inclusion of grade 2 irAE (whereas some prior reports incorporated only
   grade ≥3 toxicities[173]^13). We chose to incorporate grade 2 events
   because (1) they are medically significant, requiring adjustment of ICI
   and/or administration of immunosuppression or other therapies; (2) in
   some cases, they may become permanent; and (3) if not recognized and
   acted on, they could progress to higher-grade toxicities.

   Recently, efforts aimed at elucidating the biological underpinnings of
   irAE have emerged. A scRNA-seq analysis of ICI-treated patients with
   melanoma identified an activated CD4 effector memory subset in those
   who developed high-grade irAE.[174]^13 A separate CyTOF study found
   early on-treatment expansion of Ki67^+CD8^+ T cells and T regulatory
   cells in patients with lung cancer and melanoma who experienced
   irAE.[175]^29 The current analysis extends these findings, revealing
   associations between an increased BL abundance of plasmablasts,
   proliferating activated Mki67^+CXCR3^+HLADR^+ T cells, and high ANA
   levels, suggesting a pre-existing autoimmune-like state predisposes
   patients to irAEs.

   Multiple previous studies have identified CD8 effector and terminal
   effector T cells in irAE-affected tissues in colitis,[176]^30 31
   arthritis,[177]^32 myocarditis,[178]^33 and thyroiditis.[179]^34 In the
   present analysis, we found a higher abundance of circulating CD8
   effector and terminal effector T cells in irAE patients at BL, raising
   the possibility that these circulating CD8 T cells may act as
   precursors to the tissue-infiltrating CD8+ T cell subsets observed in
   irAE target organs. We also found elevated levels of CD57^+ T cells,
   NKT cells, and TCRγδ cells in irAE patients prior to ICI initiation.
   CD57 is a marker of terminally differentiated cells with high cytotoxic
   potential. Multiple autoimmune diseases, including Graves’ disease,
   ankylosing spondylitis, and rheumatoid arthritis, are characterized by
   increased circulating CD57^+ T cells.[180]^35 Furthermore, BL T/NK
   cells from irAE cases exhibited enrichment for the TNF signaling
   pathway and displayed a distinct three-gene signature (TNF-CXCR3-IFNG)
   that could have a potential role in irAE prediction. Conversely,
   patients without eventual irAE exhibited a more immunosuppressive and
   immune-evasive T/NK profile at BL, characterized by upregulation of
   CXCR4 and SIGLEC7.[181]36,[182]38

   scRNA-seq and cytokine analyses highlighted the central role that
   peripheral blood myeloid cells may have in mediating irAEs. In irAE
   cases, we identified a proinflammatory monocyte phenotype at BL, a
   significant upregulation of CXCL9 and CXCL10 gene expression in
   monocytes, and a marked increase in their serum levels following ICI
   initiation. After ICI initiation, monocytes in irAE cases exhibited
   upregulated IL18 expression, notable because IL18 activates T and NK
   cells and potently induces IFNγ production, thereby contributing to
   hyperinflammation.[183]^39 40 Together, these findings suggest that
   bidirectional interactions between monocytes and T cells may support
   irAE pathogenesis. In contrast, a more immunosuppressive and immune
   evasive phenotype was associated in non-irAE cases, including increased
   abundance of TGFB1^high myeloid cluster enriched for immunosuppressive
   markers such as CCR2 and STAB1 at BL and upregulation of
   immunosuppressive SPARC after ICI initiation.

   This study has multiple limitations, including the grouping of diverse
   irAE types into a single category. Although we initially intended to
   perform scRNA-seq on discrete populations according to organ-specific
   toxicities, the frequent occurrence of multiple irAE types within a
   single patient precluded that approach. Indeed, multiple other studies
   have reported irAE co-occurrence in up to 50% of irAE cases.[184]^41 42
   Relatively small case numbers for some analyses also prevent meaningful
   subset analysis according to irAE type. The inclusion of multiple
   cancer types, stages, and ICI treatments in the present study prohibits
   analysis of efficacy outcomes, which have been shown to correlate with
   irAE.[185]^24 43 Certainly, any link between proposed irAE biomarkers
   and efficacy would be critical to identify, especially if resulting
   clinical decisions could include modifying or avoiding ICI therapy.
   While that represents one approach to reducing irAE risk, other
   strategies such as personalizing irAE monitoring might mitigate the
   risk of the most severe autoimmune toxicities while not detracting from
   potential therapeutic benefits.

   Other constraints include limited sample sizes in certain cancer groups
   and among combination ICI-treated patients, as well as a lack of TCR
   and B cell receptor sequencing data. Additionally, while scRNA-seq and
   CyTOF provide valuable insights, their complexity and cost limit
   widespread clinical implementation. Future studies could validate our
   identified biomarkers, cell subsets and molecular signatures using more
   accessible methods such as flow cytometry and bulk RNA sequencing.
   Although ANA testing is clinically available, its sensitivity appears
   insufficient for standalone risk stratification. Finally, variations in
   clustering methods, cohort sizes, and cell numbers used for assays may
   influence results, underscoring the need for further validation across
   diverse cancer types and treatment regimens in larger patient cohorts
   to facilitate clinical translation.

   This multidimensional analysis identified distinct molecular signatures
   and circulating immunological determinants predictive of irAEs in
   ICI-treated patients. These include elevated BL ANA levels; increased
   plasmablasts; greater abundance of activated/proliferative CXCR3^+
   lymphocytes, CXCR3^+ T memory, and effector memory cells; and an
   upregulated 3-gene signature—IFNG, TNF, and CXCR3—in T/NK cells.
   Overall, patients who developed irAEs exhibited an activated,
   autoimmune-like proinflammatory state at BL, predisposing them to
   excessive inflammatory responses after ICI therapy, whereas non-irAE
   cases exhibited an immunosuppressive immune profile. These novel
   insights into irAE biology may contribute to future predictive
   biomarkers, open new avenues for irAE monitoring and mitigation, and
   provide a valuable resource for the growing research community focused
   on understanding irAEs.

Supplementary material

   online supplemental figure 1
   [186]jitc-13-9-s001.pdf^ (1.7MB, pdf)
   DOI: 10.1136/jitc-2025-012414
   online supplemental file 1
   [187]jitc-13-9-s002.docx^ (26.9KB, docx)
   DOI: 10.1136/jitc-2025-012414

Acknowledgements