Archive 4 checkpoints (1-4)

paper-releases/pbisc-diffusion/updates/20261015T141849Z-rare-cell-7k-8k-panels/diffusion/dataset/server_1/model8/plans/rare_gene_context_expansion_20261015/research/disease/references/PMC11383589.txt

@@ -0,0 +1,85 @@
+Address correspondence to: Jessica Neely, 550 16th Street, San Francisco, California 94158, USA. Email: jessica.neely@ucsf.edu.
+Authorship note: GR and CW are co–first authors.
+Corresponding author.
+Juvenile dermatomyositis (JDM) is one of several childhood-onset autoimmune disorders characterized by a type I IFN response and autoantibodies. Treatment options are limited due to an incomplete understanding of how the disease emerges from dysregulated cell states across the immune system. We therefore investigated the blood of patients with JDM at different stages of disease activity using single-cell transcriptomics paired with surface protein expression. By immunophenotyping peripheral blood mononuclear cells, we observed skewing of the B cell compartment toward an immature naive state as a hallmark of JDM at diagnosis. Furthermore, we find that these changes in B cells are paralleled by T cell signatures suggestive of Th2-mediated inflammation that persist despite disease quiescence. We applied network analysis to reveal that hyperactivation of the type I IFN response in all immune populations is coordinated with previously masked cell states including dysfunctional protein processing in CD4+ T cells and regulation of cell death programming in NK cells, CD8+ T cells, and γδ T cells. Together, these findings unveil the coordinated immune dysregulation underpinning JDM and provide insight into strategies for restoring balance in immune function.
+Keywords: Autoimmunity
+Keywords: Autoimmune diseases, Bioinformatics, Rheumatology
+Disease heterogeneity in juvenile dermatomyositis is underpinned by coordinated dysregulation across the immune system at the cellular level.
+Received 2023 Nov 7; Accepted 2024 May 8; Collection date 2024 Jun 24.
+Juvenile dermatomyositis (JDM) is part of a broad group of childhood-onset autoimmune conditions characterized by a type I IFN gene signature and specific autoantibodies ranging from related systemic conditions such as systemic lupus erythematosus (SLE) to endocrine-specific disorders such as type 1 diabetes (1–3). Despite a shared IFN signature, JDM is associated with pathognomonic rashes and proximal muscle weakness resulting in distinct clinical phenotypes. The etiology of JDM is not fully understood, but studies have shown that JDM is autoimmune mediated and associated with a combination of genetic and environmental risk factors (4). While mortality is low with corticosteroid treatment, long-term patient follow-up studies have reported that 60%–70% of patients have cumulative tissue damage, with the risk of damage increasing almost linearly for each year after diagnosis (5–7). This finding highlights the importance of early disease intervention and the need for a personalized approach to disease management to improve upon these outcomes.
+Clinical management of JDM currently relies on compiled empirical metrics such as physician global visual analog scale (VAS) of disease activity and muscle strength quantified via the childhood myositis assessment scale (CMAS) or manual muscle testing (MMT) (8). However, how these clinically observable phenotypes are rooted in disease immunopathology remains insufficiently understood. The presence of myositis-specific antibodies (MSA) that correspond to distinct clinical phenotypes and recent work showing that MSAs may be pathogenic suggest the involvement of B cells (9–11). The expansion of naive B cells in JDM has been highlighted by 3 independent studies using flow cytometry, mass cytometry, and single-cell RNA-Seq (12–14). The adaptive arm of the immune system is further implicated in disease pathogenesis by several large immunophenotyping studies that demonstrated the expansion of extrafollicular Th2 memory cells and central memory B cells (15, 16). Additionally, the innate immune system has emerged as a contributor in JDM with increased macrophages in skin and NK cell dysfunction described peripherally (13, 17, 18). Together, these findings highlight the involvement of both the adaptive and innate immune compartments in JDM in blood and disease-affected tissues. However, it also raises the question of whether the cause of JDM lies in a single cell type or is a manifestation of broadly dysregulated cellular interactions across the immune system.
+Systems-level studies based on single-cell measurements are required to reveal how dysregulated cell populations act individually or cooperatively to produce the observed inflammation. Accordingly, several groups have turned to next-generation sequencing, as it enables unbiased profiling of tissues at a single-cell resolution. We previously described a pan–cell type IFN gene signature overexpressed in treatment-naive JDM that was most strongly correlated with disease activity in cytotoxic cell types (14). This signature has since been independently identified in the peripheral blood of treatment-naive patients (19). However, these studies have utilized small cohorts and lack pediatric controls. Thus, it has been challenging to determine which of these findings are specific to JDM compared with healthy children, how these disease-specific dysregulated cell states are coordinated with one another, and which of these states cooperatively change in response to treatment.
+In this study, we addressed this challenge by profiling JDM across several stages of disease activity using multiplexed Cellular Indexing of Transcriptomes and Epitopes by sequencing (CITEseq) of peripheral blood mononuclear cells (PBMCs) from 15 patients with JDM, totaling 22 samples, and 5 healthy controls (HC). Compositional analysis of immune populations identified a disease activity–associated imbalance of naive and mature lymphocytes, corroborated by distinct immunophenotypes in treatment-naive disease. To move beyond the identification of disease-associated cell populations and toward an understanding of immune-scale dysregulation in JDM, we applied a recently developed computational method DECIPHERseq to infer networks of coordinated cell states from large cohorts of single-cell data (20). Importantly, this unsupervised method takes advantage of the biological heterogeneity in the entire data set, improving upon previous work that relied on pairwise comparisons of subsetted disease groups. Among other signatures previously masked by traditional single-cell analysis, this approach revealed cooccurring cell states in CD4+ T and B cell populations, suggestive of extrafollicular responses. A subset of these CD4+ T signatures implicates disruption of protein targeting and immune tolerance processes; notably, these cell states persist even in patients in remission off medication. Furthermore, we show that the hyperactive type I IFN response in disease is paralleled by impaired cell death processes in cytotoxic immune cells, highlighting the functional imbalance across immune compartments that typifies this complex autoimmune disease. This broadened understanding of the underlying immune dysregulation in disease can inform precision treatment strategies for JDM.
+To gather a data set with appropriate controls and limited confounding, patients were selected according to disease activity and medication status (Figure 1A and Supplemental Table 1; supplemental material available online with this article; https://doi.org/10.1172/jci.insight.176963DS1). Of the 15 patients with JDM, serial samples were collected from 6 individuals, totaling 22 samples. Detailed information on sample numbers from patients can be found in Figure 1A. To minimize confounding by immune suppression, the study included 9 treatment-naive samples as well as 6 samples from patients with inactive disease off medication. CITEseq was performed on PBMCs to generate single-cell libraries (Figure 1B). Surface protein expression was measured using antibody-derived tags (ADT). Following preprocessing steps, we identified 29 clusters, which comprised 21 distinct immune cell populations across 105,827 cells (Figure 2A). Clusters were annotated using canonical RNA (Figure 2B) and protein markers (Supplemental Figure 1, A and B) within all major mononuclear immune cell compartments.
+(A) Overview of clinical characteristics of study cohort. Individuals are labeled by donor ID (JDM 1–15, HC 16–20). Longitudinal samples were collected from the following donors: JDM1 (n = 2), JDM2 (n = 2), JDM4 (n = 3), JDM8 (n = 2), JDM13 (n = 2), JDM15 (n = 2). Icon shapes denote disease activity group, and shades of blue denote medication regimen. (B) Analysis strategy for CITEseq data from PBMCs. n = 22 JDM, n = 5 HC. 
+(A) UMAP constructed using weighted nearest neighbors (wnn) clustering colored by cell type. pDCs, plasmacytoid DCs; cDCs = classical DCs; PBs, plasmablasts; B_mem, memory B cells. (B) Heatmap with top 2 markers per cluster. (C) Box plot shows cell type proportion by disease group, using Kruskal-Wallis test with Dunn’s post hoc test comparing TNJDM with HC, TNJDM with inactive JDM, and inactive with HC (Holm’s, Padj < 0.05; *Padj < 0.05, **Padj < 0.01). The dot plot above shows the Spearman correlation between corresponding cell type proportion in box plot and Physician Global VAS, where the size of the dot indicates the correlation, the color indicates the direction of the correlation, and the border weight indicates significance (Padj < 0.05). (D) Heatmap with selected ADT protein markers. Asterisks mark significant comparisons between TNJDM and HC per cell type with an absolute LFC > 0.5 and Padj < 0.05.
+We first characterized global changes to cell composition across disease states comparing treatment naive JDM (TNJDM), inactive JDM, and HC (Figure 2C). Within the T cell compartment, the proportion of Tregs (CD45RO+, IL2R+, FOXP3+) was increased in patients with TNJDM (P = 0.02) consistent with previous findings (14). CD4+ effector T cells (CD45RO+) and γδ T cluster 2 (TRDC, TRGC) were significantly increased in patients with inactive JDM, and the proportion of cells from these populations negatively correlated with disease activity measures (P < 0.05). There was an overall decrease in innate populations in TNJDM compared with HC and inactive JDM, and the proportion of these cell types also correlated negatively with disease activity (P < 0.05).
+Compared with HCs and patients with inactive disease, treatment-naive patients had higher proportions of multiple naive B cell populations, herein referred to as “B_naive,” and their UMAP cluster number, including B_naive1 (IgM+IgD+CD38+CD24+CD10+) corresponding to an immature naive B population, B_naive2 (IgM,+IgD+CD38loCD24lo), and B_naive3 (IgM+IgD+CD38+CD24+), and the proportion of these populations positively correlated with multiple disease activity measures (P < 0.05) (Figure 2C). The proportion of B_mem cells, characterized by TNFRSF13B expression, negatively correlated with the muscle VAS score (P < 0.05). The immature naive B population had higher expression of CD38 (both RNA and protein) and MZB1, 2 genes essential for plasma cell differentiation, compared with all other B cell clusters (21, 22).
+Given the observed imbalance of lymphocytes in treatment-naive JDM, we next sought to immunophenotype B cell and CD4+ T cell subsets in JDM at the proteomic level to gain molecular insight into cell states (Figure 2D and Supplemental Table 2). Differential protein analysis of immature naive B cells comparing treatment-naive JDM to HC identified increased expression of MICA-MICB and decreased expression of CD1C, BAFF-R, and PD-L1 (Figure 2D). Within the CD4+ T compartment, CD4+ Tregs from TNJDM had higher expression of Tim-3, ICOS, CD164, and CD38 and downregulation of CD101 a molecule which decreases proinflammatory T cell responses (23). CD4+ Teff in patients with TNJDM had higher surface expression of CD164 and PD-1 and showed downregulation of KLRG1, an inhibitory molecule (Figure 2D). The overexpression of PD-1 on the cell surface suggested that peripheral Th cells might be present in JDM (24, 25). However, while ICOS expression was higher (P < 0.05), no difference was found in surface expression of CXCR5 between CD45ROhiPD-1hiCD4+ T cells and CD45ROloPD-1loCD4+ T cells, and these cells were not significantly expanded in JDM (Supplemental Figure 2).
+We next compared gene and protein expression between treatment-naive JDM and HC samples in all cell types based on the hypothesis that certain cell types may not be altered in composition but may be functionally altered at the molecular level. Monocytes displayed the highest number of differentially expressed genes and proteins in this analysis, including CD169 (SIGLEC-1), CD107a (LAMP-1), and CD164 (Supplemental Figure 3). SIGLEC-1 is a monocyte-restricted IFN-induced protein that was recently identified as a potential biomarker in JDM (26). Both CD107a and CD164 are cell adhesion molecules involved in trafficking of activated PBMCs and adhesion to vascular endothelium (27).
+A common finding across all cell types when comparing treatment-naive JDM and HC samples was overexpression of genes enriched in type I IFN processes, which was previously reported in bulk expression data and confirmed in single-cell studies (Supplemental Figure 4) (14, 19, 28). Using an IFN gene score derived from the transcriptional data (Supplemental Figure 5), we plotted the per-patient average score in each cell type (Figure 3A). This approach did not detect IFN gene expression to persist beyond the treatment-naive state, and 2 patients with TNJDM had negligible IFN gene signatures as quantified by this method. This heterogeneity of the IFN gene signature was partly explained by disease activity level (Figure 3B), as a bulk IFN gene score correlated with disease activity (r = 0.69). However, the remaining unexplained heterogeneity of this IFN score exemplifies a limitation of utilizing gene scores identified through pairwise comparisons between subsets of the data.
+(A) Heatmap of average IFN score per cell type and sample. Hierarchical clustering was performed using Euclidean distance and the complete clustering method. IFN score was calculated based on average expression of IFN module across all cells per sample. (B) Spearman correlation between IFN score and Physician Global VAS colored by disease group. (C) Scatter plot showing Spearman correlation between CD169 (SIGLEC-1) expression in CD14+ monocytes and Physician Global VAS. (D) Scatter plot showing Spearman correlation between IFN score and Physician Global VAS. (E) Scatter plot showing Spearman correlation between CD169 expression and IFN score in CD14+ monocytes.
+Given that SIGLEC-1 is a type I IFN–induced protein, we investigated whether patterns of type I IFN–stimulated gene expression were reflected at the protein level, as protein biomarkers are more amenable for clinical lab-based testing. SIGLEC-1 expression in CD14+ monocytes also correlated with disease activity to a similar degree as the IFN gene signature (Figure 3, C and D), and SIGLEC-1 expression was itself highly correlated with the IFN gene signature in monocytes (Figure 3E). This suggests that SIGLEC-1 expression in CD14+ monocytes is a representative composite measure of the IFN gene signature in JDM. These results underscore the potential of SIGLEC-1 as a biomarker of IFN responses in JDM that may be useful for tracking disease activity.
+We next turned to a systems-level approach to better understand the coordination of immune cell gene programs in JDM relative to HCs and in relation to disease activity level. We applied an unsupervised network inference method, DECIPHERseq, to the 6 major cell types annotated in the data set: B cells, CD4 T, CD8 T, NK cells, γδ T cells, and myeloid cells (Figure 4A). DECIPHERseq relies on nonnegative matrix factorization (NMF) to first break the data set down into gene sets that represent distinct states of biological activity, or “activity programs,” and then constructs a network of gene expression programs (GEPs) based on how expression of the programs covaries across patient samples (Figure 4A) (29). After outlier filtering, NMF identified 76 activity programs (Figure 4B).
+(A) Overview of the DECIPHERseq workflow. (B) Heatmap showing 6 major clusters of GEPs identified by DECIPHERseq (Pearson). GEPs are clustered into modules, with isolated GEPs filtered out (grayscale).
+Next, a force-directed network graph from the correlation matrix of activity programs was constructed in which each node represents a program and each edge represents a statistically significant (P < 0.05) positive correlation between 2 nodes (Figure 5A). Using DECIPHERseq’s community detection algorithm, we identified 6 hubs of interconnected activity programs or “modules.” All modules contained multiple cell types, highlighting that many biological processes in JDM are coordinated across several immune cell types (Figure 4B and Figure 5A). We annotated each node using gene set enrichment analysis (GSEA) of Gene Ontology (GO) terms on each program’s ranked marker gene list (Supplemental Tables 3–4 and Supplemental Figures 6–11) (30, 31).
+(A) Network constructed from correlated GEPs in PBMCs from patients with JDM and HCs. Nodes represent programs in the given cell types, and edges represent positive significant correlations (Pearson, P < 0.05). (B) Dot plot showing selected gene sets found to be enriched within specific modules compared with the rest of the network. Color corresponds to module enrichment P value, and size corresponds to a set’s rank in the list of significantly enriched gene sets for that given module ordered by ascending module enrichment P value (network permutations, GSEA, FDR < 0.01). All gene sets shown fall in the top 10 terms for their respective modules (total gene sets: 626).
+DECIPHERseq’s module enrichment analysis identified consensus biological themes for each module in an unsupervised manner (Figure 5B and Supplemental Table 5). Module 1 was enriched for type I IFN response programs such as Response to Virus. Module 2 was enriched in ribosomal processes including Translational Initiation. Module 3 included lymphocyte programs and was significantly enriched for cell adhesion and migration. Module 4 represented cells’ steady state processes as it was enriched for gene sets like Circadian Rhythm. Module 5 was annotated as a Stress Response module and enriched for Regulation of Cell Death and Cellular Response to Chemical Stress. Module 6 contained few unique gene sets and was enriched for programs intrinsic to eukaryotic cells like DNA Packaging.
+Next, we aimed to interpret the annotated network in the context of GEPs associated with JDM compared with HCs irrespective of disease activity. We first focused on Module 1, which was enriched in type I IFN responses. Many programs in this module were increased in TNJDM, as expected (Figure 6, A and B). All 6 major cell types expressed IFN gene programs that were highly correlated to one another, as shown by the closely connected hub at the center of Module 1 (Figure 6A). This IFN hub was associated with JDM as compared with HCs (P < 0.05) (Figure 6, B and C). IFN modules identified by NMF were highly expressed in all treatment-naive patients as well as some patients with active disease, inactive disease, and a HC (Figure 6B), in contrast to the signature of IFN gene expression previously detected by differential gene expression in Figure 3A. This highlights the strength of this method to more accurately reflect the low-dimensional space of gene expression where measurement of many genes working together may be needed to detect underlying biological processes (32–34).
+(A) Zoomed-in graph of Module 1. GSEA results for Response to type I IFN GO term shown with each node colored according to FDR. Padj value of module enrichment is also shown (network permutations, Methods). (B) Heatmap showing significant differences in expression of selected programs between HC (n = 5) and patients with JDM (n = 22), with columns annotated by P values (P < 0.05) of case-control (t test) and disease activity association (4-group 1-way ANOVA). (C) Network graph showing case-control analysis of each program’s expression, with node size scaled according to P value and colored according to strength of the association between disease status and program expression (t test).
+We next identified gene programs in Modules 1–3 expressed more highly in all patients with JDM compared with HCs (P < 0.05). These included B cell (5 and 14) and CD4T (1, 10, 17) programs, and their expression persisted even in patients with inactive JDM who previously achieved remission off medication (Figure 6B and Supplemental Figure 12). Patients with JDM more highly expressed 2 B cell programs; B5 in Module 1 was enriched in mRNA metabolic processing, RNA splicing, chromatin organization and modification, and cell cycle regulation, and B14 in Module 3 was enriched in chromatin remodeling and cytoskeletal organization (Supplemental Figure 6 and Supplemental Figure 8). These enriched biological processes suggest that a subpopulation of B cells are more transcriptionally active and undergoing epigenetic regulation in JDM relative to HCs.
+In Module 3, correlated to B14, CD4T1 (enriched in cell migration, adhesion, activation, and secretion) was expressed more highly in JDM and in the region of the Uniform Manifold Approximation and Projection (UMAP) corresponding to CD4+ Teff (Supplemental Figure 13). This CD4T1 program expressed by CD4+ Teff contained genes (GATA3, CCR4, PRDM1) that indicate possible skewing toward a Th2 subset, while expression of PRDM1 (Blimp-1) suggests participation in extrafollicular reactions (Figure 6B). Th2 CD4+ T cells were previously found to be expanded in JDM and associated with extrafollicular B cell–T cell help (15, 16). We observed similar expression of Th2 genes (GATA3, CCR4, PRDM1) in CD4T10, a Treg program (FOXP3, IKZF2, IL2RA) expressed more highly in JDM (Figure 6B). CD4T1 and CD4T10 included genes for costimulatory molecules OX40 (TNFRSF4) and GITR (TNFRSF18), both of which have been described to promote survival and proliferation of CD4+ Teff and have been targets of autoimmune disease therapeutics (Supplemental Table 4) (35–39). Notably, CD4T17 (AIM2, ACTB, ACTG1, NCF1, ID3, SOX4) was negatively associated with JDM, and expression was significantly decreased in nearly all patients (Figure 6B, Supplemental Table 4, and Supplemental Figure 14A). This program was enriched in protein targeting to the membrane and endoplasmic reticulum and included several genes important in T cell regulation. NCF1 has been found to be a critical regulator of T cell tolerance in a collagen-induced arthritis mouse model (40), and coexpression of ID3 and SOX4 transcription factors has been identified as a mechanism of CAR-T cell exhaustion and dysfunction (41). Together, these results suggest multiple mechanisms by which CD4+ T cell dysfunction may occur in JDM, including participation in extrafollicular reactions, expression of costimulatory molecules, and downregulation of genes important in mediating tolerance and exhaustion.
+We next wanted to identify modules and gene programs associated with stages of disease activity in JDM (HC, inactive and active JDM, and TNJDM). To do so, we performed a 4-group 1-way ANOVA on each program in the network and post hoc pairwise analysis using the Tukey test. We identified programs in Module 1, 2, and 5 that were significantly associated with disease activity (P < 0.05) (Figure 7A). We confirmed that these biological programs moved in the direction expected within most patients with longitudinal assessments (Supplemental Figure 15). The IFN gene programs were also significantly overexpressed in treatment-naive patients with JDM, as expected (Figure 6B). Notably, expression of the central Module 1 IFN hub GEPs in all 6 major cell types more strongly correlated to the clinically evaluated PGA than the pseudobulk IFN gene score derived from pairwise DEG analysis (Supplemental Figure 16), underscoring the utility of a dimensionality reduction approach in uncovering clinically relevant gene signatures.
+(A) Network graph showing results of 4-group 1-way ANOVA of each program’s expression, with node size scaled according to P value and colored according to strength of the association between disease status and program expression. (B) Heatmap showing significant differences in expression of selected disease activity–associated programs between HC (n = 5) and patients with inactive JDM (n = 6), active JDM (n = 7), and TNJDM (n = 9). Columns are annotated by P values of case-control t test and disease activity association (4-group 1-way ANOVA). (C and D) Selected network modules colored by FDR of enrichment for indicated gene ontology set (FDR < 0.01) or gene loading similarity within Modules 2 (C) and 5 (D).
+By isolating these IFN GEPs in each cell type, we were able to determine disease activity–associated programs correlated with the IFN hub, some of which corroborate previous findings (Figure 7, A and B). This approach identified B9, an immature naive B cell program (CD24, CD38, MME; Supplemental Table 4) to be significantly associated with disease activity (Figure 7B). This gene program shared several top markers (TCL1A, SOX4, NEIL1) with the immature B cell population that was previously found to be expanded in treatment-naive JDM (15, 23) (Supplemental Table 4 and Supplemental Figure 14B). Notably, expression of this activated immature B cell program could be attributed to the B_naive1 cluster that we observed to be increased in treatment-naive JDM during the compositional analysis (Supplemental Figure 17). Similarly correlated with the IFN hub, NK12 was associated with treatment-naive JDM compared with active and inactive disease (Figure 7B and Supplemental Figure 18A). NK12 (MKI67, HIST1H1B) was enriched for gene sets related to cell proliferation and epigenetic regulation, confirming findings that a subset of NK cells in JDM are highly activated and proliferative (Supplemental Table 4 and Supplemental Figure 14C) (13, 17).
+We next focused our attention on the other disease activity–associated programs that DECIPHERseq identified as correlated with the Module 1 IFN hub. Importantly, these disease activity associations were only revealed in the lower dimensional space of gene sets identified by NMF rather than the noisier space of differential expression of individual genes. We annotated CD4T10, also significant in the case control analysis, as a proliferative Treg program (FOXP3, IL2RA, MKI67) which expressed genes implicated in extrafollicular B cell–T cell interactions (PRDM1) and genes associated with Th2-mediated inflammation (GATA3, CCR4; Figure 7B and Supplemental Table 4) (42). Notably, CD4T10 included the marker CCR4, a chemokine receptor highly expressed in Tregs that are preferentially recruited to skin under inflammatory conditions (43). Expression of both CD4T1 and CD4T10 colocalized with surface protein expression of CCR4 in the UMAP as well (Supplemental Figure 19), highlighting the advantage of this multimodal sequencing approach in identifying functional markers of transcriptomic signatures.
+This network approach also identified the program γδ T4, a cytotoxic Th1 polarized γδ T program (GZMB, CX3CR1, TBX21) that was correlated with the central IFN hub and was significantly increased in treatment-naive patients compared with both active and inactive JDM and HC (P < 0.05; Figure 7B and Supplemental Figure 14D). High expression of TRGC1 and TBX21, encoding the transcription factor T-bet responsible for regulating IFNG expression, specifically identified cells expressing this program as Th1-like TCRVd1 γδ T cells (Figure 7B and Supplemental Table 4) (44). A similar subpopulation of γδ T cells was found to be increased in synovial fluid and blood of patients with juvenile idiopathic arthritis, which expressed IFN-γ and TNF-α to the same degree as CD4+ T cells (45). This suggests this subpopulation of γδ T cells may reflect an important inflammatory cell state specific to treatment-naive disease correlated with IFN gene expression.
+The disease activity programs that were highly expressed in treatment-naive JDM were components of Module 1, which was enriched for type I IFN and its associated immune processes. The network-wide 1-way ANOVA also revealed disease activity–associated programs in Module 2 and Module 5 (Figure 7A), which were anticorrelated with Module 1 (Supplemental Figure 20), and expression was decreased in treatment-naive patients with JDM compared with HCs and other patients with JDM (Figure 7B). Module 2 was significantly enriched for GO terms “ribosome assembly” and “translational initiation” (Figure 7C), while Module 5 was enriched for terms “regulation of cell death” and “cellular response to chemical stress” (module enrichment, P < 0.005) (Figure 7D). The disease-associated programs within these modules were expressed significantly lower in treatment-naive JDM, suggesting dysfunction of cellular processes that underpin ribosomal activity and cell death regulatory processes at disease onset (Figure 7B and Supplemental Figure 18, B–D).
+Notably, disease activity–associated programs CD8T11, NK8, and γδ T15 (FOS, JUN, DUSP1, NR4A2, GADD45B) in Module 5 share a common gene signature (Figure 7B; Supplemental Figure 14, C and D; and Supplemental Table 4) and are each individually enriched in “regulation of cell death” and “regulation of cell cycle” (Figure 7D and Supplemental Figure 10). We quantified the overlap in gene expression between activity programs by Fisher’s exact test and confirmed the high gene loading similarity between programs in Stress Response Module 5 (Figure 7D). All 3 of these programs were expressed at lower levels in active and treatment-naive JDM and negatively correlated with activated disease–associated cell signatures identified in Module 1 (Supplemental Figure 18, B–D). The gene loading similarity analysis revealed that the programs CD4T9 and B10 also share top marker genes (FOS, JUN, DUSP1, NR4A2, GADD45B) (Figure 7D and Supplemental Table 4); however, these 2 programs were not associated with disease activity status. This suggests that regulatory mechanisms of cell death may be uniquely disrupted in circulating cytotoxic cell populations in patients with active disease.
+In Module 2, CD4T17 and NK9 were enriched in several gene sets related to protein processing, such as protein targeting to the ER (Supplemental Figure 7A). Interestingly, the CD4T17 program was also characterized by high expression of several genes encoding members of the actin protein family (ACTB, ACTG1; Supplemental Table 4). Given the crucial role actin filaments play in antigen recognition during the formation of the immune synapse, dysfunction in components of that protein machinery could have substantial effects on the immune system. Among other disease activity–associated programs, differential expression of CD4T17 between HCs and patients with JDM persisted even in patients who achieved remission off medication (Figure 7B). Taken together, disease activity–associated programs in Modules 2 and 5 highlight shared cellular processes that may be underactive in JDM, providing what we believe to be novel insights into potential cellular mechanisms that accompany the known signature of overactive IFN-response in JDM.
+We next investigated whether these JDM-associated signatures could be identified in an independent set of samples. Using DECIPHERseq’s marker quantification method, we subset the genes in each JDM-associated GEP that contributed the most to that program. With each gene list as input, we calculated a proxy GEP metric that quantified rank-based expression of each program as the enrichment of that subset of genes in each cell (46). This proxy NMF GEP metric recovered signatures identified by NMF in the original data set (Supplemental Figure 21).
+Using this proxy NMF GEP method, we validated key signatures in an independent set of CITEseq data from 5 JDM samples and 2 HC (Figure 8, A and B). Importantly, these samples were obtained from patients who had active disease, and 4 of 5 were being treated with medication. We compared GEP expression between cases and controls (P < 0.05), and we identified significantly higher expression of CD4T1 (CD4+ Teff program, CCR4, PRDM-1, GATA3) and CD4T10 (Treg program, FOXP3, CCR4, PRDM-1, GATA3) in JDM and lower expression of CD4T17 (AIM2, ACTB, ACTG1), replicating the original results in the initial cohort (Figure 8C). In the B cell compartment, B5 and B14 trended toward increased expression in patients with JDM. However, a single individual had low expression in both programs, indicating heterogeneity of expression of these B cell signatures (Figure 8C).
+(A) Clinical characteristics of validation cohort (n = 7). HC18 was included in original cohort, but an independent sample was collected and analyzed for this data set. Individuals are labeled by the donor ID. Immunosuppressants denoted as “+” for patients JDM22–25 were as follows: (JDM22: methotrexate), (JDM23: IVIG, cytoxan), (JDM24: hydroxychloroquine, MMF, IVIG, tofacitinib), (JDM25: methotrexate, IVIG). (B) UMAP of single-cell RNAseq data from validation cohort PBMC samples, colored by 6 major cell types corresponding to labels used in original cohort. (C) Box plots of case-control comparisons (HC = 2, JDM = 5) for selected programs queried in validation data set using AUCell (t test, *P < 0.05, **P < 0.01). (D and E) Scatter plots correlating disease activity (PGA) with AUCell scores for selected IFN programs (D) and selected disease activity programs (E) in validation data set (Spearman, P < 0.05).
+To validate disease activity–associated programs, we correlated GEP expression of disease activity–associated programs with the PGA score. It was infeasible to validate γδ T programs due to low cell numbers. IFN GEPs were most strongly correlated with disease activity in CD4T cells, B cells and NK cells (Figure 8D) and trended toward a positive correlation in myeloid cells, but there was no significant association in CD8 T cells (Supplemental Figure 22, A and B). Additionally, B9 (immature B cell signature; CD38, CD24, MME), CD4T10 (Treg program; FOXP3, CCR4, PRDM-1, GATA3), and NK12 (proliferative activated NK cell program; MKI67, CENPF, HISTH1HB) strongly correlated with disease activity (Figure 8E). A subset of the key signatures identified in the original cohort with DECIPHERseq trended with disease activity but did not significantly correlate with PGA in the validation cohort (Supplemental Figure 22, C–E). Together, these results demonstrate the robustness of many signatures quantified with a less precise method even in an independent data set with fewer samples. Cell death regulatory signatures previously identified to be negatively correlated with IFN signaling in cytotoxic populations may be more strongly associated with a treatment-naive state or more heterogeneous across healthy individuals such that significant differences in expression could not be identified in this smaller cohort.
+Multiple components of the adaptive and innate immune compartments have been implicated in the pathogenesis of JDM, consistent with its categorization as a complex autoimmune condition. However, previous studies have been unable to uncover how multiple disease-associated cell states are coordinated to produce inflammation. Here, in the largest single-cell study of JDM to our knowledge to date, we provide an unbiased, comprehensive picture of immune dysregulation in peripheral blood, including a subset of aberrant signatures that persist despite disease quiescence in individuals off medication. Through traditional analyses, we first show that immune dysregulation in JDM manifests at the level of compositional imbalance of immune populations and that these compositional changes are correlated to clinical metrics of disease activity. Next, we identified distinct disease-associated molecular signatures of lymphocyte and myeloid subsets through multimodal differential analysis and demonstrate that, of these markers, surface expression of SIGLEC-1 in CD14+ monocytes is a composite metric of disease activity and reflects the type I IFN response in JDM (28). Using DECIPHERseq to deconvolve disease-associated programs beyond the broad type I IFN response, we uncovered CD4+ T cell states that persist in JDM despite disease remission coordinated with downregulated cell death processes in cytotoxic immune populations. Together, these findings generate more nuanced hypotheses for disease etiology.
+Within the B cell compartment, we observed skewing toward an immature state in treatment-naive disease and observed the distinct transcriptomic and proteomic signature of immature naive B cells consistent with what we and others previously reported (14, 19). Given that autoantibodies are thought to play a role in disease pathogenesis, this skewing of the B cell compartment would seem counterintuitive. However, given recent findings emphasizing the importance of extrafollicular B cell differentiation pathways through which autoreactive “activated naive” B cells are precursors to antibody-secreting cells, we hypothesize that this skewing may be suggestive of extrafollicular reactions in JDM (24, 47). In fact, the expanded immature naive population had higher expression of CD38 and MZB1, genes important for plasma cell differentiation, than all other B cell clusters. The overall low expression of CD27 and CXCR5 across all B cells made it difficult to conclude whether this population matches the double-negative B cell population associated with SLE (47). However, recent immunophenotyping work in a large cohort of patients with JDM that found simultaneous expansion of CXCR5– central memory B cells and Th2 cells provides support for further investigation into extrafollicular B cell–T cell help in JDM (15). Alternatively, this skewing could represent more mature B cells homing to tissues as has been described in antisynthetase syndrome (48, 49). Functional work to support or negate the extrafollicular pathway in JDM will be critical to determine whether this is a targetable pathway therapeutically.
+Accompanying these immunophenotypic changes in B cells, we observed complementary dysregulation in the T cell compartment that lends further support to the hypothesis of extrafollicular interactions in JDM. In populations of peripheral blood FOXP3+ Tregs and CD4+ effector T cells, we identify a shared disease-associated signature composed of genes suggestive of Th2 activation (GATA3, CCR4), involved in promotion of survival and proliferation (GITR, OX40), and associated with extrafollicular T cell responses (PRDM1). Notably, this signature persisted in disease remission in patients off medication. Likewise, we identify a CD4+ T signature, decreased in all patients with JDM regardless of medication status and disease activity, containing genes crucial for tolerance (NFC1) and regulation (ID3, SOX4). This cell state could represent an inflammatory signature or a compensatory mechanism of CD4+ T cells in more long-standing disease. These findings are consistent with previous work that identified skewing of CD4+ T cells toward a Th2 phenotype in JDM and showed in vitro that peripheral Th2 cells were efficient in helping B cells, including stimulating antibody production (16). A previous study also showed that tertiary lymphoid structures are present in muscle of new-onset JDM, further supporting a role for extrafollicular reactions in JDM (50). Future work using paired blood and tissue with spatial information to immunophenotype cells interacting within these structures would further justify investigating therapeutic strategies that prevent homing of CD4+ T and B cells to sites of inflammation in tissue.
+While other studies have reported that Tregs in JDM have diminished suppressive capacity, raising the possibility of Treg exhaustion (51), the results in this study show that the expanded population of peripheral Tregs in blood are proliferative and activated (MKI67, IL2RA, IRF4), taking on an effector phenotype. Likewise, the shared signature with CD4+ Teff suggests that these Tregs are coopting the transcriptional machinery of effector T cells as has been described by others (52, 53). In this data set, JDM Tregs also upregulate transcriptomic and proteomic expression of CCR4 — paralleled by increased expression of CCR4 in CD4+ effector T cells — which is preferentially expressed in Tregs recruited to the skin (43). Thus, we speculate that this expanded population of Tregs in JDM could represent a peripheral response to site-specific Th2-mediated inflammation in disease-affected tissue. Alternatively, these Tregs could be functioning in a reparative manner at sites of tissue damage. Future functional studies of peripheral blood and tissue-specific Tregs, particularly investigation of the influence of type I IFN on Treg-suppressive capacity in JDM, would provide mechanistic insight on this population’s role in disease pathogenesis. The potential translational effect of investigating Treg dysfunction is corroborated by active development of multiple therapeutics targeting Tregs for autoimmune diseases (54).
+More broadly, we show that unsupervised approaches such as DECIPHERseq can be used to consolidate disparate findings into a systems-level understanding of how interactions among cell states could manifest in disease. Here, our network analysis revealed that a module of hyperactivated IFN response across cell types is coordinated with dysfunction in ribosomal biogenesis, protein processing, and the regulation of cell death that is also shared across many cell types. This model contextualizes recent work that has identified ribosomal dysfunction in NK cells as a disease signature in JDM but also raises the possibility that defective translational machinery is not unique to that cell population (13, 17). Given that type I IFN directly promotes the activation and proliferation of NK cells (55, 56), we speculate that NK cells in JDM are unable to properly translate cytolytic protein machinery required for effector function in response to IFN signaling, potentially perpetuating the IFN response. Similarly, the shared program between CD8T, γδ T, and NK cells that describes regulation of cell death and cellular stress response suggests a common dysfunction across cytotoxic cell populations in JDM. Given the importance of cytotoxic cells in clearing cellular debris including autoantigenic neutrophil extracellular traps shown to be pathogenic in JDM, dysfunctional cytotoxic populations could result in accumulation of such debris, thereby triggering an autoimmune response mediated by lymphocytes (11).
+Finally, the observation that type I IFN responses increase with clinical metrics of disease activity adds to the growing body of work suggesting that disease activity in JDM correlates with this transcriptional signature (57). However, given the time and cost, it remains infeasible to use transcriptomic sequencing as a lab-based clinical diagnostic tool. These data point to surface expression of SIGLEC-1 in monocytes as a composite measure of the IFN gene signature in JDM and disease activity. Together with a recent independent study of JDM, we provide external validation that SIGLEC-1 is a suitable biomarker for disease monitoring to pursue in larger immunophenotyping studies, given the lower cost and ease of implementing screening by flow cytometry (26). Importantly, we show that SIGLEC-1 directly reflects the IFN gene signature using paired gene and protein expression measurements, strengthening support for its use as a biomarker. Further study of this biomarker, and the role of SIGLEC-1 in disease, is an important step toward precision care of JDM.
+These findings should be interpreted in the context of the study’s limitations. First, despite being the largest single-cell study in JDM to our knowledge to date, sample numbers are still limited such that the study lacks statistical power to quantify the contribution of MSA status to disease heterogeneity. Furthermore, a majority of patients in the treatment-naive group are TIF1γ+, which could introduce a bias to disease activity–related programs, though it remains unknown whether MSA status is associated with distinct biological mechanisms. Some patients in the “active” disease group had relatively low disease activity, which may have prevented us from identifying more associations with disease activity. Additionally, this study lacked data from matched JDM skin and muscle, which would have enabled insight into how dysregulated cell states in blood might influence local microenvironments in disease-affected tissue. Although profiling blood limits the mechanistic insight compared with skin or muscle, it is a more suitable sample type for biomarker discovery, particularly in a pediatric disease that requires longitudinal monitoring, and future comparison with tissue data will enable us to identify populations in peripheral blood with tissue correlates. Lastly, the DECIPHERseq algorithm relies on a k selection procedure to accurately decompose the data. As a dimensionality reduction technique, NMF is distinct from principal component analysis in that there is no single solution for the number of patterns or components into which the data are segmented. Therefore, it is necessary to optimize the parameter “rank K” such that the NMF results capture the relevant biology at an appropriate granularity. We addressed this limitation of NMF by using the phylogenetic clustering–based k-selection method described by Murrow et al., where the authors demonstrate that saturation of this metric reflects the appropriate granularity of biological programs such that results are robust across multiple choices of rank K (20).
+In summary, using CITEseq to profile compositional and functional imbalance of peripheral blood immune cells and the relationship to disease activity, we provide a comprehensive map of the coordinated immune dysregulation underlying JDM. We identify persistent transcriptional changes in B and CD4+ T cells associated with JDM that persist even in patients in remission off medication and reveal cell states associated with the IFN signature that generate hypotheses for the role of extrafollicular interactions in disease pathogenesis, drawing parallels to other autoimmune diseases. Importantly, we believe these findings pose a new paradigm to how we approach JDM treatment. The dysregulation of processes simultaneously with hyperactivation of other cell states necessitates that we identify therapeutic strategies that restore balance to the dynamic interactions between immune populations rather than simply turning off a set of pathways. Taken together, our work sets the stage for improving clinical management of JDM by providing a foundation for systems-level inquiry into the cellular basis of this disease. More broadly, application of a similar analytical strategy could provide insight into the immunologic basis of other childhood-onset autoimmune diseases characterized by a type I IFN gene signature.
+This study contained samples from human males and females. Sex was not considered as a biological variable in downstream analyses.
+Patients were recruited to the Juvenile Myositis Precision Medicine Biorepository between 2018 and 2023 and underwent informed consent. The diagnosis of JDM was per clinician judgment; however, all patients included in this study met EULAR/ACR classification criteria for “definite” juvenile idiopathic inflammatory myopathy based on typical skin manifestations of either Gottron’s and or heliotrope rashes (58). This study was approved by the UCSF IRB. Clinical data were collected by study investigators and recorded in a secure REDCap database. Treatment-naive JDM was defined as a new diagnosis of JDM with no systemic immune suppressant use in the prior 4 weeks. Inactive JDM was defined as normal CK, MMT8 ≥ 78 and Physician Global VAS score < 0.5 to reflect PRINTO clinically inactive disease (59) definitions but with some modifications based on the data available. Active disease was defined as Physician Global VAS score ≥ 0.5, and all patients in this category were taking immune suppressive medications. Longitudinal samples from n = 6 patients with JDM were included separated by at least 4 months in time and accompanied by a change in disease activity. Measures of disease activity, including the Cutaneous Disease Area and Severity Index (CDASI), were collected at study visits (60). HCs were enrolled who had no prior autoimmunity; no known or suspected genetic disorders, immunodeficiency, active cancer, or history of organ or bone marrow transplantation; no infection or antibiotics in the prior 4 weeks; no chronic systemic immunomodulatory medication use; and no vaccinations in the prior 6 weeks. Peripheral blood samples were collected at each study visit and processed by the Pediatric Clinical Research Core Sample Processing Lab (UCSF). PBMCs were collected in SepMate tubes (STEMCell, 85415) (n = 9) using Ficoll separation or CPT tubes (n = 18), isolated per manufacturer’s guidelines, and cryopreserved in liquid nitrogen.
+The experimental protocol was previously published (14). These experiments were carried out using early access kits from BD Genomics before the implementation of commercially available single-cell protein/RNA assays (e.g., Feature Barcoding, 10X Genomics; BD Abseq, BD Genomics; Supplemental Table 6), and researchers are recommended to use those newer solutions for any follow-up studies as the techniques and reagents have been refined. PBMCs from 27 distinct samples were gently thawed in a 37°C water bath and resuspended using a pipette set to 1 mL. Cell counts and viability were determined using a Cellometer Vision (Nexcelcom) with AOPI staining (Nexcelcom, CS2-0106-5ML). Cells were multiplexed into 4 pools: 1 “cross pool” with all samples that consisted of only 1 time point and 3 pools consisting of longitudinal samples. Longitudinal samples from the same individual were assigned to separate pools to enable genetic demultiplexing. After pooling, cells were resuspended in 90 μL of 1% BSA in PBS and Fc blocked with 10 μL Human Trustain FcX (BioLegend, 422302) for 10 minutes on ice before being stained on ice for 45 minutes with a pool of 268 antibodies in 100 μL, for a final staining volume of 200 μL. Antibodies were pooled on ice with 2.2 μL per antibody per 1 × 106 cells (BD Genomics). Cells were quenched with 2 mL 1% BSA in PBS and spun at 350g for 5 minutes and further washed 2 more times with 2 mL of 1% BSA in PBS. After the final wash, cells were resuspended in 100 μL and strained through a 40 μM filter (SP Bel-Art, H13680-0040). Each longitudinal pool was split across two 10X Chromium lanes while the “cross pool” was split across six 10X lanes (6 wells total, 5 × 105 cells/well). The 10X Chromium controller (10X Genomics, 1000202) was run, and post-GEM reverse transcription (RT) and cleanup were done according to manufacturer’s protocol (10X Genomics 3′ Kit V3). Starting at cDNA amplification, modifications to the protocol were made: 1 μL of 2 μM additive primer (BD Genomics, beta kit) specific to the antibodies tags was added to the amplification mixture. During the 0.6X SPRIselect (Beckman Coulter, B23318) isolation (0.6X ratio of SPRIselect beads/sample reaction volume) of the post-cDNA amplification reaction cleanup, the supernatant fraction was retained for ADT library generation. Subsequent library preparation of the cDNA SPRIselect pellet was done exactly according to protocol, using unique SI PCR Primers (10X Genomics). For the ADT supernatant fraction, a 1.8X (beads/original sample reaction volume) SPRIselect was done to isolate ADTs from other nonspecifically amplified sequences, followed by sample index PCR. Sample index PCR for the ADTs was done using the cycling conditions as outlined in the standard protocol (15 cycles) but using unique SI-PCR primers such that all libraries could be mixed and sequenced together. Subsequent SPRI selection was performed, and all libraries were quantified and analyzed via Qubit 2.0 (Thermo Fisher Scientific) and Bioanalyzer (Agilent), respectively, for quality control. We sequenced the libraries on 2 lanes of a NovaSeq S4 (Illumina), aligned using CellRanger (10X Genomics) to generate feature barcode matrices.
+Data were demultiplexed using genotypes with demuxlet (61), and doublets were filtered using DoubletFinder (62). Next, cells were filtered to remove RNA transcripts expressed in < 3 cells, cells with > 60% ribosomal reads and > 15% mitochondrial reads (mtDNA). Cells were further filtered for > 5,000 ADT counts to avoid antibody aggregates, < 70 antibodies detected, and any antibody isotype control measurements > 50. To remove background ambient RNA signal, we ran SoupX separately on each of the 6 RNA libraries before merging (63). Aggregated data were log-normalized and scaled, regressing out percent mtDNA, percent ribosomal DNA, and cell cycle (S, G2M) (64). Data were then integrated with Harmony, with 20 maximum iterations of Harmony and 30 maximum clustering iterations per round of Harmony (65).
+Normalization was performed using the dsb algorithm on all 6 ADT libraries individually, using default parameters except for more stringent quantile clipping (lower threshold = 1st percentile, upper threshold = 99th percentile) (66). The background distribution of empty droplets was defined as suggested in the dsb vignette. Isotype controls were then removed from the data set, and reciprocal PCA (RPCA) was used to integrate the dsb-normalized ADT data across libraries. Following RPCA, the data were rescaled and the cell cycle scores and number of ADT counts and features were regressed out. The harmonized RNA and RPCA-corrected ADT were combined using weighted nearest neighbors, with default parameters except for the threshold for filtering edges in the Shared Nearest Neighbors (SNN) graph (‘prune.SNN’ = 1/20). Leiden clustering was run on the resulting graph (method = igraph) at a 1.4 resolution (67). Two clusters were removed with low to no expression of ADT, and the object was reclustered with the same parameters. The Seurat function FindAllMarkers was used to identify the top 5 markers per cluster.
+We removed 3 additional clusters: 2 were small clusters with a transcriptomic profile consistent with doublets (original Leiden clusters 26 and 29; Supplemental Figure 23), and 1 was a diffusely expressed cluster (original Leiden cluster 19; Supplemental Figure 23). We further subclustered 3 clusters that expressed genes representative of more than 1 cell type: original Leiden clusters 16, 17, and 23. Subclustering was performed using Seurat’s FindSubCluster function using the lowest possible resolution to divide the population into 2 clusters. Based on minimal transcriptional differences between them, original Leiden clusters 1, 5, 9, 11, and 15 were merged into a single CD4+ T naive population; clusters 3 and 10 were merged into a single naive CD8+ T population; and cluster 7 and part of the subsetted cluster 23 were merged into a CD56dim NK population. Due to interpersonal heterogeneity in monocytes, all CD14+ monocyte clusters were merged into a single CD14+ monocyte population (68).
+While annotating, we discovered that the FOXP3 signature normally attributed to Tregs was only present in a subset of the cluster, and FindSubCluster did not appropriately isolate the FOXP3+ cells. We therefore subsetted the cluster and reran FindVariableFeatures, ScaleData, RunPCA, FindNeighbours, FindClusters with the Louvain algorithm and a resolution of 0.8, and RunUMAP. This enabled us to subset a smaller group of cells with a statistically significant expression of FOXP3 compared with other clusters using FindAllMarkers, which we hence annotated Tregs. Annotation was performed using both canonical gene and protein markers. One B cell population consisted almost solely of cells from 2 donors. This was annotated as B_naive4 and was not used in downstream analysis, but it was included in UMAPs.
+Cell type proportion was calculated as the proportion of each cell type for each individual and was compared for: treatment-naive JDM compared with HC, treatment-naive JDM compared with inactive JDM, and inactive JDM compared with HC using Kruskal-Wallis test with Dunn’s post hoc test. To determine the association between cell abundance and disease activity, the Spearman correlation coefficient between cell type proportion and physician global VAS scores was calculated, and P values were adjusted using Benjamini-Hochberg (BH).
+The DGE and DPE analysis were completed using DESeq2. Size factors were set using the function computeSumFactors from the scran package. The DESeq function was used with default parameters for single cell data (69). Batch was included as a covariate using the “reduced” argument. We filtered genes and proteins that were not expressed in at least 5% of cells and analyzed only cell types where there were at least 100 cells in each group. We used cutoffs of |LFC| ≥ 1 for genes, |LFC| ≥ 0.5 for proteins, and P < 0.05. Overrepresentation analysis was performed on up- and downregulated genes per cell type using the clusterProfiler package with GO Biological Processes terms set as the reference and adjusted P < 0.05. For the PD1/CD45R0 subanalysis (70), we compared groups using Seurat’s FindMarkers with test.use = ‘MAST’, latent.vars = ‘well’, |LFC| ≥ 0.5, and P < 0.05.
+We curated a list of all genes differentially expressed between treatment-naive JDM and HC in at least 2 cell types. The average expression of each gene was calculated using Seurat’s AverageExpression function per sample for each cell type and visualized using dittoSeq’s dittoHeatmap (71) with default, unsupervised clustering settings of both rows and columns. Dendrograms were ordered using the dendsort package (72). This identified 7 distinct modules, where Module 1 consisted exclusively of type I IFN–stimulated genes. Average Module 1 scores for each cell type were then calculated using Seurat’s AddModuleScore with default settings. Correlations between disease activity and IFN score was calculated using Spearman correlation and visualized using ggplot2 (73).
+We applied NMF to the raw RNA count data as implemented in the DECIPHERseq method with default parameters and the NMF rank, k, chosen using the weighted subtrees metric based on phylogenetic clustering (20). The final choices of rank for each cell type were kB = 17, kCD4T = 17, kCD8T = 14, kγδT = 13, kMyeloid = 17, kNK = 11 according to the saturation point in the elbow plots (Supplemental Figure 24). Network clustering was performed on the per-sample averaged program scores with default parameters. Marker gene scores were calculated on the corresponding gene loading vectors for each GEP as previously described (74). GSEA was performed on the resultant ranked gene lists using the fgsea (75) package in R with GO and Hallmark gene sets. Module themes were assigned by calculating module enrichment P values using the Get_enrichment_pvals function in DECIPHERseq with default parameters. Module and gene set enrichment results were visualized using ClusterProfiler (76).
+We performed CITEseq using the same protocol as described above with PBMCs from 5 patients with JDM and 2 healthy pediatric controls. We used the same steps for data processing, with the exception that we used CellBender rather than SoupX for ambient RNA removal and clustered cells using the RNA measurements only (77). To derive proxy scores for GEP in the independent data set, we ranked the gene lists comprising each program by marker score, which quantifies how strongly a single gene contributes to that GEP. Using the top 5% of genes by marker score, that list was used as input for the rank-based gene subset enrichment method AUCell (78). Pseudobulk proxy GEP scores were calculated as the per-patient mean expression in the same way for the original NMF programs. Case versus control comparisons were done using 2-tailed t tests between patients with JDM and HCs. Given the patients per group (HC, 2; active JDM, 4; TNJDM, 1) in this validation cohort, we could not repeat the 1-way ANOVA comparisons used for disease activity association in the original data set. Instead, proxy GEP scores were correlated to VAS global using the Spearman correlation.
+All statistical analyses and visualization of results were performed using open-sourced R (version 4.2.3). Pairwise comparisons of cell proportions between patient groups were performed using a Kruskal-Wallis test with 2-tailed post hoc Dunn’s comparison, with P values adjusted for multiple comparisons by Holm’s correction. Significance of Pearson correlations between GEPs used for network construction was calculated using bootstrapping as implemented in DECIPHERseq. Analyses of disease association with GEPs was performed using 2-tailed unpaired t test or 1-way ANOVA with post hoc Tukey test. FDRs for GSEA annotation and module enrichment across programs were calculated and corrected at the cutoff FDR < 0.01. Gene loading similarity was calculated as the Pearson correlation between gene loadings for each activity program and all other activity programs in the same module, with P values calculated by permutation testing. Data are shown as mean ± SD. Correlation methods used in specific figures are described in the corresponding legends and in Methods, and significance for all statistical tests was set at the threshold P < 0.05.
+This study was approved by the UCSF IRB 17-24003. Written informed consent to participate in this study was provided by the participant or the participant’s legal guardian depending on the age of the participant. Assent was obtained when appropriate.
+The data sets presented in this study are deposited in the CZ CELLxGENE Discover resource as “CITEseq of JDM PBMCs” (https://cellxgene.cziscience.com/collections/c672834e-c3e3-49cb-81a5-4c844be4a975). The code used for this analysis has been made publicly available on Github at “github.com/grabadam-cal/jdm-DECIPHER-2024”. Values for all data points in graphs are reported in the Supporting Data Values file.
+GR and CW were assigned co–first authorship because their contributions were deemed equally essential to the study, with GR listed first because she led the writing of the manuscript. JN, SK, MS, and CJY designed the experiments. JN and SK recruited patients for the study. GCH, YS, and JN performed the experiments. GCH performed the alignment and sample demultiplexing. EF, GR, and CW performed the integration and quality control. CW performed the annotations and the cell proportion and differential analysis. GR conceptualized and wrote the pipeline for the network analyses; EF ran the pipeline. GR, CM, and GKF performed analysis of the second data set. ZJG, MS, and JN provided essential feedback on the analysis strategy and statistical methods. GR and JN wrote the manuscript; MS and ZJG revised the manuscript. JN, MS, and SK conceptualized the study. JN led the project administration and acquired funding for the study. All authors reviewed and edited the manuscript.
+The authors would first like to express gratitude for the patients with JDM and their families who contributed their samples and data. We would also like to acknowledge the Institute of Human Genetics Genomics Core at UCSF, Center for Advanced Technology at UCSF, and the UC Berkeley Vincent J. Coates Genomics Sequencing Laboratory. We thank the members of the UCSF Pediatric Rheumatology Division for helping to recruit patients: Emily von Scheven, William Bernal, Michael Waterfield, Erica Lawson, Alice Chan, Geraldina Lionetti, Nicole Ling, Tara Valcarcel, Andy Nguyen, William Soulsby, Julia Shalen, Sara Haro, Tyalor Laflam, and our research coordinators, Bhupinder Nahal, Kathy Nguyen, Katrina Gonzales, Bethany Bautista, and Zian Zheng. JN was supported by the Cure JM Foundation, the Doris Duke Physician Scientist Fellowship grant 2019124, K12 HD105250, and K08 AR081983. GR was supported by the NIH ipCBS Fellowship, NIH T32-GM008155, and the Rheumatology Research Foundation and in part by grants from the NIH (R01GM135462 and R33CA247744) and the UCSF Center for Cellular Construction (DBI-1548297), an NSF Science and Technology Center. CW was supported by the DARE Fellowship, Fabrikant Aage Lichtingers Foundation, Director Ib Henriksens award, Knud Hoejgaards Foundation, Jorcks Foundation, and the Denmark-America Foundation. GCH was supported by the NSF Graduate Research Fellowship. MS, GKF, CJY, JN, and EF were supported by P30-AR070155. ZJG is a Chan Zuckerberg BioHub San Francisco Investigator.
+In-Press Preview
+Electronic publication
+Conflict of interest: GR is a former employee of 23andMe. CJY is founder for and holds equity in ImmunAI and Survey Genomics, a Scientific Advisory Board member for and holds equity in Related Sciences and ImmunAI, a consultant for and holds equity in Maze Therapeutics, and a consultant for TReX Bio, HiBio, ImYoo, and Santa Ana. CJY has received research support from the Chan Zuckerberg Initiative, Genentech, BioLegend, ScaleBio, and Illumina. ZJG is a scientific advisor for and owns shares in Scribe Biosciences and Serotiny. MS is a scientific advisor for Exagen.
+Copyright: © 2024, Rabadam et al. This is an open access article published under the terms of the Creative Commons Attribution 4.0 International License.
+Reference information: JCI Insight. 2024;9(12):e176963.https://doi.org/10.1172/jci.insight.176963.
+Gabrielle Rabadam, Email: gabrielle.rabadam@ucsf.edu.
+Camilla Wibrand, Email: camilla.wibrand@ucsf.edu.
+Emily Flynn, Email: e.flynn@ucsf.edu.
+George C. Hartoularos, Email: george.hartoularos@ucsf.edu.
+Yang Sun, Email: Yang.Sun@ucsf.edu.
+Chioma Madubata, Email: chioma.madubata@ucsf.edu.
+Gabriela K. Fragiadakis, Email: Gabriela.Fragiadakis@ucsf.edu.
+Susan Kim, Email: Susan.Kim@ucsf.edu.
+Zev J. Gartner, Email: Zev.Gartner@ucsf.edu.
+Marina Sirota, Email: marina.sirota@ucsf.edu.
+Jessica Neely, Email: jessica.neely@ucsf.edu.
+The data sets presented in this study are deposited in the CZ CELLxGENE Discover resource as “CITEseq of JDM PBMCs” (https://cellxgene.cziscience.com/collections/c672834e-c3e3-49cb-81a5-4c844be4a975). The code used for this analysis has been made publicly available on Github at “github.com/grabadam-cal/jdm-DECIPHER-2024”. Values for all data points in graphs are reported in the Supporting Data Values file.

paper-releases/pbisc-diffusion/updates/20261015T141849Z-rare-cell-7k-8k-panels/diffusion/dataset/server_1/model8/plans/rare_gene_context_expansion_20261015/research/disease/references/PMC12049347.txt

@@ -0,0 +1,75 @@
+Corresponding author.
+Contributed equally.
+Immune checkpoint inhibitors (ICI) targeting CTLA-4 and PD-1 have shown remarkable antitumor efficacy, but can also cause immune-related adverse events, including checkpoint inhibitor-induced liver injury (ChILI). This multi-omic study aimed to investigate changes in blood samples from treated cancer patients who developed ChILI. PBMCs were sequenced for by transcriptomic and T cell receptor repertoire (bulk and single-cell immune profiling), and extracellular vesicle (EV) enrichment from plasma was analyzed by mass spectroscopy proteomics. Data were analyzed by comparing the ChILI patient group to the control group who did not develop ChILI and by comparing the onset of ChILI to pre-ICI treatment baseline. We identified significant changes in T cell clonality, gene expression, and proteins in peripheral blood mononuclear cells (PBMCs) and plasma in response to liver injury. Onset of ChILI was accompanied by an increase in T cell clonality. Pathway analysis highlighted the involvement of innate and cellular immune responses, mitosis, pyroptosis, and oxidative stress. Single-cell RNA sequencing revealed that these changes were primarily found in select T cell subtypes (including CD8 + effector memory cells), while CD16 + monocytes exhibited enrichment in metabolic pathways. Proteomic analysis of plasma extracellular vesicles showed enrichment in liver-associated proteins among differentially expressed proteins. Interestingly, an increase in PBMC PD-L1 gene expression and plasma PD-L1 protein was also found to be associated with ChILI onset. These findings provide valuable insights into the immune and molecular mechanisms underlying ChILI as well as potential biomarkers of ChILI.
+Trial registration number NCT04476563.
+The online version contains supplementary material available at 10.1007/s00262-025-04033-z.
+Keywords: Checkpoint inhibitor, ChILI, Liver injury, Biomarker, PD-L1
+Received 2025 Jan 27; Accepted 2025 Mar 24; Collection date 2025 Jun.
+Immune checkpoint inhibitors (ICI) targeting CTLA-4 (cytotoxic T lymphocyte-associated protein 4) and PD-1 (programmed cell death 1) demonstrated antitumor efficacy in preclinical models and humans across several types of cancers [1–3]. These ICIs have become the cornerstones of modern immunotherapy for a variety of cancer types that demonstrated promising durable responses. These T cell co-inhibitory pathways, or checkpoints, maintain desirable cellular immunity in the body. Disruption of these immune checkpoints can result in enhanced antitumor immune response but can also cause immune-related adverse events (irAE) affecting various organ/tissues including liver [4–6]. The incidence rate of checkpoint inhibitor-induced liver injury (ChILI) is 11.5 per 1000 person months; about 3% of patients treated with single agent develop ChILI, but this can reach as high as 32% when combination of ICIs are used [7].
+CTLA-4 is primarily expressed on CD4 + and CD8 + T cells [8]. It competes with CD28 for binding to CD80/86—the second signal required for T cell activation. PD-1 plays a vital role in keeping T cell activity in check in the peripheral tissues through interaction with one of its ligands, PD-L1 or PD-L2 [9], thereby maintaining immunologic tolerance. The exact immune mechanisms of irAEs or ChILI have not been elucidated [5]. ChILI exhibits features of acute liver injury and clinically resembles idiosyncratic DILI and acute autoimmune hepatitis (AIH). ChILI has phenotypic and morphologic characteristics that are similar to autoimmune diseases [10]. An animal model of DILI was developed by inhibiting immune tolerance; acute liver injury occurred when PD-1(−/−) mice were treated with the combination of anti-CTLA-4 antibody and amodiaquine [11]. If the mechanism of action were to drive adverse events, then the determinants of efficacy should also be similar to those with adverse effect.
+Early diagnosis and differentiation of ChILI from DILI and AIH would facilitate appropriate clinical management of ChILI. Various approaches have been explored to identify biomarkers associated with efficacy in ICI-treated patients, such as immunophenotyping [12–14], cytokine profiling [15, 16], immune-related transcriptional signature [17–23], TCR clonality [24, 25], and multiplex IHC [26–28]. In monkey and mouse models with dual inhibition of CTLA-4 and PD-1 pathways, activation, proliferation, and infiltration of T cells in affected tissues (e.g., heart or liver) were observed [21, 29]. Immune signatures in peripheral blood and tissue have been identified as associated with irAEs and ChILI in these models. In the mouse ChILI model, CD8 + T cells were essential since liver injury was abrogated when CD8 + T cells were depleted. Histological examinations revealed multi-focal infiltrates with necrotic hepatocytes surrounded by lymphocytes [29]. ChILI patients treated with ipilimumab and nivolumab displayed some similar liver pathological changes and immune cell features to this mouse model. Focal necrosis throughout the liver parenchyma with CD8 + T cell-rich mixed mononuclear infiltration was commonly observed in ChILI patients [10, 30].
+In the current study, PBMCs were analyzed by T cell receptor (TCR) sequencing and RNAseq. Single-cell RNAseq was performed on a subset of samples. Extracellular vesicle (EV) enrichment was performed on plasma samples followed by tandem mass tag (TMT) mass spectroscopy (M/S) for unbiased proteomic analysis. Samples from subjects affected by ChILI were compared to both pre-treatment baseline and control subjects receiving ICI therapies that did not develop ChILI. Distinct changes at transcriptomic and proteomic levels that were associated with the onset of ChILI were observed and are discussed.
+The experimental design is summarized in Fig. 1. Patients with cancer where immune checkpoint inhibitor (ICI) treatment was indicated as the main therapy were prospectively recruited with informed consent (trial registration: NCT04476563) at Nottingham University Hospitals NHS Trust (NUH). Blood samples were collected prior to treatment with ICI (before checkpoint inhibitor; BC). If patients developed ChILI (13–233 days following start of treatment, median 61.5 days), a blood sample was taken at the visit at the time of liver injury (TOL) with acute manifestation of liver injury meeting previously defined clinical chemistry criteria [31]. Additional blood samples were taken for a subset of subjects at follow-up (FU) visits (FU1: 3–49 days after TOL, median 19 days; FU2: 10–62 days after TOL, median 23 days). In addition, ChILI subject samples were collected at TOL for which BC were not available. Cancer patients that did not develop ChILI or any other organ toxicity were followed up after 12 weeks with a second blood collection, and this group was followed as a non-ChILI control (AC). Whole blood was preserved in PAXgene tubes for subsequent RNAseq, and plasma was isolated from EDTA tubes for proteomic analysis. The patients included 15 males and 46 females with a median age of 60 years old (ranging from 30–88 years). Patient metadata, including cancer type, ICI regimen and any corticosteroid treatment, are summarized in Suppl Table 1.
+Mixed prospective study experimental design. The various study groups are delineated by BC (before checkpoint inhibitor), AC (after checkpoint inhibitor control), TOL (time of liver injury), and FU (follow-up) timepoints 1 and 2. Under each group designation, the number of PBMC samples assessed for RNAseq is shown, followed by the number of plasma samples assessed by proteomics in parentheses. The numbers along the arrows indicate the number of patients in control and ChILI groups for whom BC samples were available. Additional ChILI subjects lacking BC samples were added to increase the power of the AC vs TOL comparison
+RNA was isolated from blood preserved in PAXgene tubes (Qiagen) using the PAXgene Blood RNA Kit (Qiagen) according to manufacturer’s protocol. Full methods are detailed in the Data Supplement.
+RNA input was 100 ng for the Lexogen QuantSeq 3’mRNA-Seq Library Preparation kit for Illumina (FWD) (Lexogen, #015.96) and libraries were prepared according to the manufacturer’s protocol. After elution, libraries were quantified using the Qubit 4 Fluorometer (ThermoFisher Scientific, #Q33238) using the Qubit 1X dsDNA HS Assay Kit (ThermoFisher Scientific, #Q33231).
+Libraries were sequenced either on an Illumina Nextseq 500 or Nextseq2K (Illumina, San Diego, CA) using a 75 cycle or 100 cycle single end flow cell, respectively.
+For bulk human TCR RNA sequencing experiments, total RNA was normalized and used for TCR libraries generation using the SMARTer Human TCR-a/b Profiling kit v2 (Takara Bio Inc.) according to manufacturer’s instructions. Libraries were sequenced on a MiSeq v3 (600 cycles) flow cell, targeting 5,000,000 read pairs per library.
+Single-cell RNAseq was performed by the 10 × Genomics single-cell 5′ Gene Expression and V(D)J library platform. Dead cells were removed using the Dead Cell Removal Kit (Miltenyi Biotec). Single-cell libraries were generated with Chromium Single Cell 5’ Gene Expression and V(D)J Reagent Kit (10 × Genomics) per manufacturer’s instruction. Purified libraries were sequenced on the Illumina NextSeq 2000 with 200 cycle kits.
+Takara NGS Immune Profiler Software was utilized to process TCRseq samples in a conda base environment with python version 3.7.4. We used the *_cdr3_clones_results.csv files for diversity and clonality analysis [32]. The Gini coefficient was calculated to quantify clonality [33].
+For bulk RNAseq data analysis, reads were pseudo-aligned to the Homo sapiens reference genome GRCh38.p14 from Ensembl v. 110 and quantified using Salmon version 1.10.0. Normalization factors were determined using the edgeR package (version 3.40.2) for expression data normalization and processing. Pairwise comparisons were run for differential expression, defined by timepoint (before and after checkpoint inhibitor) and outcome (ChILI and non-ChILI control). Benjamini–Hochberg multiple testing correction was performed on p-values. Heatmaps were generated using the internally developed, publicly available application “HotGenes” (https://github.com/pfizer-opensource/Open-Hotgenes).
+For single-cell RNAseq and V(D)J sequencing analysis, sequence reads were demultiplexed using CellRanger multi (10 × Genomics) and normalized, QC filtered and merged using the R package Seurat. Cell clusters were annotated using the human PBMC reference (HuBMAP Consortium), and differentially expressed genes within cell types in groups defined by timepoint and outcome were identified using the R package NEBULA. All RNAseq data have been uploaded to GEO (GSE287540, GSE287541).
+EV enrichment on patient plasma samples was performed using the Mag-Net method automated on a Kingfisher Apex as previously described [34]. The resulting peptides were then labeled with TMTpro reagents with a common pool of selected samples used as a bridge channel. Pooled TMT plexes were fractionated and analyzed by LC-MS3 on an Eclipse Tribrid mass spectrometer using real-time search.
+After instrument acquisition, data were analyzed on Proteome Discoverer 3.0. Raw protein abundance values were exported, bridge and global median normalization was performed, and resulting values were used for differential analysis with limma using sex as a factor. Proteins with Benjamini–Hochberg adjusted p-value of 0.1 or less were utilized for pathway analysis. All proteomic data have been uploaded to MassIVE (MSV000096928).
+The Gini coefficients for TCR samples were plotted in GraphPad Prism (v10.2.1). For samples with BC data, Gini coefficients for subsequent visits were normalized to BC, and one-way ANOVA was performed in GraphPad using Sidak multiple comparison test.
+For all differentially expressed transcripts and proteins (padj < 0.1; all raw p-values were < 0.05), gene IDs and corresponding p-values and fold changes were loaded into ingenuity pathway analysis (Qiagen). A comparison analysis was performed among all contrasts, and results were exported for enrichment of canonical pathways and upstream regulators. Common pathways were ranked by log(p-value) of enrichment and log2(fold change).
+T cell receptor repertoires were assessed by profiling transcripts encoding TCRα and TCRβ subunits. Changes in clonality of TCR sequences were evaluated by calculating the Gini coefficient for each sample. The Gini coefficient is a measurement of the distribution of TCR sequences, ranging from 0 (maximum diversity) to 1 (maximum clonality) [35]. In subjects that developed ChILI, a trend toward increased clonality was observed in both TCRα and TCRβ from BC to TOL, which was reversed by FU2 in the few available FU2 samples (Fig. 2A and B). When samples were normalized to their respective BC time point, TOL samples were significantly elevated compared with both AC and FU time points (Fig. 2C and D), indicating a transient but significant increase in clonality in the T cell population in checkpoint inhibitor treated patients at the time of liver injury.
+T cell receptor clonality. Gini index was calculated for each sample. The Gini indices for TCRA (A) and TCRB (B) sequences are shown over time for subjects that developed ChILI. *p < 0.05, ****p < 0.0001
+RNAseq data were analyzed for differences at various time points, between ChILI and control, and the impact of steroids (the full list of differentially expressed genes can be found in Suppl Table 2). The largest factor was found in subjects with PBMCs measured at TOL. A total of 554 differentially expressed genes (DEGs) were identified between TOL and BC, with 649 DEGs identified when comparing TOL with AC. A set of 261 DEGs (> 40%) were overlapping between the two contrasts. The top 50 DEGs between TOL and BC or AC are shown in Fig. 3. Most of the TOL samples cluster together, while AC and BC samples are intermixed. The majority of DEGs, as shown in red under the TOL samples, are upregulated at TOL relative to BC or AC. The top 20 DEGs between TOL or AC vs BC with fold changes and p-values are shown in Table 6 (found in Suppl). Comparing PBMC profiles in control subjects (those treated with ICI who did not develop liver injury), no changes in gene expression were observed between post-treatment (AC) to pre-treatment (BC). Therefore, the differentially expressed genes associated with TOL are likely indicative of the ChILI condition, and not the ICI treatment itself. One caveat in the TOL subjects was that a subset of these genes may also be impacted by steroid treatment; asterisks in Table 6 denote genes that were also statistically significant when comparing TOL patients on steroids to those who were not on steroids. (A total of 191 DEGs were identified for this comparison; top 20 DEGs are shown in Suppl Table 3, with the full list available in Suppl Table 2) While many checkpoint inhibitor-related transcripts were unchanged across all contrasts, CD274 (the transcript encoding PD-L1) increased 1.47–2.12-fold in TOL vs AC and TOL vs BC comparisons, with no influence by steroid treatment. This increase could represent a persistent change due to feedback regulation following ICI inhibition.
+Heatmap showing the top 50 differentially expressed genes among samples from TOL, AC and BC. Each column represents a single sample. Red represents up-regulation, blue represents down-regulation. In the top dendrogram, light blue represents TOL samples, dark blue represents AC samples, and green represents BC samples
+To better understand the pattern of gene expression changes, pathway analysis was performed in ingenuity pathway analysis. The top 20 canonical pathways and predicted upstream regulators are shown in Table 1 (the full list of pathways for all contrasts is found in Suppl Table 4). These top 20 pathways/regulators primarily contain innate immune responses such as TLRs, type I IFN, NLR, and STING, and cellular immune response including MHC I antigen presentation and T cell effector response (TNF, IL17A, and IFNγ). As with DEGs, asterisks denote pathways also impacted in TOL subjects with versus without steroids. Common pathways with potential interaction by steroids included several inflammatory pathways, such as neutrophil degranulation, interleukin-1 family signaling, class I MHC-mediated antigen processing and presentation, and pathogen-induced cytokine storm signaling. Likewise, common upstream regulators predicted to be activated by TNF, IFNA, and IFNG were predicted to be inhibited by steroids, as would be expected. Several pathways implicated by ChILI DEGs related to mitosis, pyroptosis, and oxidative stress, however, were not enriched in the steroid vs no steroid comparison. While this suggests minimal impact of the steroids on reversing such pathways, and indeed a recent publication advocates a strategy that spares use of steroids [36], the smaller N for the steroid vs no steroid comparison is one caveat to this conclusion.
+Top 20 pathways and upstream regulators in common between TOL vs BC and TOL vs AC contrasts enriched in RNAseq data
+Single asterisks denote pathways or regulators that were significantly enriched with a positive z-score; double asterisks denote pathways or regulators that were significantly enriched with a negative z-score
+Additional contrasts compared the FU visits to TOL. Only 46 DEGs were identified for FU1 vs TOL, and 62 DEGs compared FU2 to TOL. No statistically significant pathways based on z-scores were identified for these contrasts. Another contrast within the BC group compared those subjects that would develop ChILI to those who did not, to look for potential risk factors ahead of treatment. Only four DEGs were identified between these two groups at baseline (MAML1, RPL10P9, PPP1R11, and HLA-DR2), and the fold changes were negligible (1.02–1.03).
+To gain a higher resolution of the transcriptional changes following development of ChILI, PBMCs from a subset of patients (10 BC, 10 AC, 10 TOL, and 8 FU1) were evaluated by single-cell RNAseq. After separation of data from individual cell populations, most differentially expressed genes comparing post-treatment samples to BC in subjects who developed ChILI were identified as coming from CD8 + naïve T cells, CD16 + monocytes, CD8 + effector memory cells (TEM), and CD4 + central memory cells (TCM) (Fig. 4A). Many of the top pathways identified by bulk RNAseq, such as IFNg signaling, pathogen-induced cytokine storm signaling, and antigen presentation, were driven primarily by the T cell subtypes (Table 2). CD8 + TEM, CD8 Naive, and CD4 + TCM cells also showed enrichment in the T cell exhaustion signaling pathway. DEGs within CD16 + monocytes, conversely, were enriched for predominantly metabolic pathways. Despite monocytes being a prevalent cell type in PBMCs, enrichment of these metabolic pathways such as alanine biosynthesis, metabolism of water-soluble vitamins and cofactors, and GDP-L-fucose biosynthesis, were not detected in the bulk RNAseq data (Suppl Table 4).
+Single-cell RNAseq. Number of DEGs in each of the cell types in PBMC comparing post-treatment to pre-treatment for those subjects that developed ChILI (A) and those subjects that did not develop ChILI (B)
+Top 10 pathways for ChILI post-treatment vs BC identified by single-cell RNAseq in CD16 monocytes, CD4 TCM, CD8 naïve T cells, or CD8 TEM cells
+The given value is the − log(p-value), and pathways are ranked by the sum across the four cell types. Significant enrichment (> 1.3) is denoted by bold
+EV enrichment was performed on plasma for analysis of protein content by TMT LC-MS3 based proteomics (Fig S1). Over 6,900 proteins were detected, with quantification of 3,758; 325 of these quantified proteins were found to be enriched in liver (Fig. 5A). Differentially expressed proteins (DEPs) were identified for TOL vs BC (319 DEPs) and TOL vs AC (261 DEPs). Across these comparisons, over 60% of the proteins with elevated expression in ChILI plasma were enriched for expression in liver (Fig. 5B and C).
+Results of proteomic analysis of EV-enriched plasma. A Venn diagram depicting the overlap between total proteins detected, quantified, and those annotated as enriched in liver. Volcano plots highlighting the enrichment of liver-associated DEPs for B TOL vs BC and C TOL vs AC. Red circles denote increased expression, blue circles denote decreased expression of liver-associated proteins
+Like the RNAseq results in PBMCs, the plasma proteomic data showed > 40% overlap in DEPs between the TOL vs BC and TOL vs AC contrasts, and no DEPs in the AC vs BC cohort. Unlike the RNAseq data, no DEPs were found to be associated with steroid treatment. Thus, the DEPs are likely not influenced by the steroid or ICI pharmacology but instead are a result of the ChILI condition. The top 20 DEPs in common between TOL vs BC and TOL vs AC are shown in Suppl Table 7 (found in Suppl) and includes several liver-enriched proteins involved various aspects of metabolism or functions, such as ALDOB, CPS1, ALDH1L1, GPT, BHMT, RBP4, POR, TTR, and GPT. At the onset of liver injury when compared with AC or BC, in addition to some of the liver-enriched enzymes (ALDOB, ALDH, CPS1, BHMT), there were also increases in sera in those proteins associated with tissue/cell injury (intracellular non-secreting proteins), such as STAT1, NCL, POR, PSMB8, API5, RSP20, and RPLP0. The full list of DEPs for all contrasts can be found in Suppl Table 2. Many proteins involved in the ICI response, such as CD27, CD28, CTLA-4 and PD-1, were not detected in the EV proteomic data. Like the PBMC RNAseq data, however, PD-L1 was significantly increased 1.75–1.90-fold in TOL vs AC and TOL vs BC contrasts.
+Comparing the enriched pathways and upstream regulators for the proteomic data to the RNAseq data highlighted a clear role for innate immunity, interferon signaling, antigen presentation, and inflammatory responses (Table 3; Suppl Table 7). Enriched pathways predicted to be activated that were unique to the proteomic data included eukaryotic translation elongation, cellular response to heat stress, and estrogen receptor signaling. The full list of pathways enriched by DEPs is available in Suppl Table 4.
+Top 20 pathways and upstream regulators in common between TOL vs BC and TOL vs AC contrasts enriched in proteomic data
+Asterisks denote pathways or regulators in common between RNAseq and proteomic datasets
+The present study sought to profile circulating biomarkers in patients impacted by ChILI using transcriptomic and proteomic analyses. Blood samples were evaluated from prospectively enrolled cancer patients including those indicated for ICI before and during treatment, supplemented with additional post-treatment samples from ChILI patients and controls (ICI-treated subjects who did not develop ChILI). The results indicated that ChILI was accompanied by an increase in T cell clonality, and gene expression analysis revealed significant differences between ChILI and control subjects, particularly at the time of liver injury. Pathway analysis highlighted the involvement of innate and cellular immune responses, as well as pathways related to mitosis, pyroptosis, and oxidative stress. The single-cell RNA sequencing data showed that T cell subtypes primarily drove the identified pathways, while CD16 + monocytes exhibited enrichment in biosynthetic pathways, including alanine, GDP-L-fucose, L-glutamine and molybdenum cofactor biosynthesis. Proteomic analysis of plasma EVs revealed enrichment in liver-associated proteins among the differentially expressed proteins. The findings from both RNAseq and proteomics analyses highlight the activation of innate immunity, interferon signaling, antigen presentation, and associated inflammatory responses in ChILI.
+The findings of increased clonality in ChILI subjects at TOL in the present study are consistent with expansion of antigen-specific T cells which may be involved in the liver injury. A single-cell RNA and TCR sequencing study of samples from patients dosed with ICI showed similar evidence of TCR clonotype expansion in patients experiencing myocarditis [37]. Other studies that have followed ICI-treated patients with single-cell TCR sequencing have observed early increases in T cell clonality in the absence of adverse tissue injury [38–40]. Early expansion of T cell clonality in circulating T cells as well as tumor-infiltrating T cells has been reported for oral cancer patients treated with ICI [41]. If an early increase in T cell clonality is a necessary component of ICI therapy, the difference in the present study in the cases of ChILI may either be a continued expansion that is propagated by other mechanisms, and it could be a secondary expansion as part of the reaction to a liver-specific antigen, or some combination thereof. Future studies profiling clonality changes over time would be necessary to ensure to provide a more comprehensive understanding of primary or secondary responses to therapy or tissue injury.
+The majority of gene expression changes in the present study were found by comparing TOL to BC or AC; minimal changes were observed between AC and BC. A prior study in patients treated with ICI that developed myocarditis detected increases in multiple S100A family members by bulk RNAseq [39]. Aside from S100A9 (known to be involved in neutrophil chemotaxis and adhesion), which was elevated in TOL vs AC and TOL vs BC bulk RNAseq data, other S100A members were unchanged at the transcript or protein level in the present dataset. Another single-cell sequencing study also reported higher levels of effector T cells in patients experiencing myocarditis following ICI therapy [42]. Other studies have demonstrated significant impact of the ICI on PBMC immunophenotypes and gene expression. Single-cell RNAseq 1–3 weeks following treatment with ICI identified an increase in effector memory T cells and NK cells, accompanied by a reduction in B cells and CD4 + T cells, though these findings were in the absence of ChILI or myocarditis [38]. As the present data were typically collected at later time points during treatment, the T effector memory changes may be short-lived in response to treatment but may persist longer in subjects who develop ChILI. This highlights one of the limitations of this kind of study; the transcriptomic and proteomic data are collected at snapshots in time, while the underlying processes are dynamic. In the present study, however, the timing of the TOL samples (including FUs) is largely comparable to the AC control group.
+The single-cell RNAseq data in the present study provided further information on ChILI-associated changes in specific populations of circulating immune cells. The changes in the CD8 + T effector memory cell population was previously identified in an overlapping set of ChILI samples to the present study by mass cytometry time of flight (CyTOF), and comparison to drug-induced liver damage and autoimmune hepatitis demonstrated specificity for this change in ChILI (Astbury et al., submitted manuscript). An increase in the CD8 + effector memory population was also detected by CyTOF and single-cell RNAseq in a cohort of ICI-treated patients with myocarditis [37], suggesting potentially common mechanisms underlying tissue injury due to a cell-mediated immune response. In a separate study of ICI patients with myocarditis, however, single-cell and bulk RNAseq also revealed changes in NK cells and B cells [39], which were not observed in the present study. None of the patients included here developed myocarditis, which is not surprising considering the low incidence (0.06–1%) of myocarditis compared to that of ChILI (8.8%) [7, 43].
+The proteomic data generated in the present study used a new method focused on EV in plasma [34], and provided additional insight into the biology of ChILI. Specifically, a significant proportion of liver-associated proteins were found to be differentially elevated in subjects with ChILI, suggesting that proteomics identified potential biomarkers not restricted to circulating immune cells. In a recent study, some of these protein biomarkers such as Aldolase B and Carbamoyl-Phosphate Synthase 1 have been shown to be biomarkers that were able to distinguish DILI from acute non drug-induced liver injury [44]. A prior study used immuno-depletion of high abundance proteins in plasma followed by M/S proteomics in patients with ICI-associated myocarditis highlighted pathways related to innate immunity and acute inflammatory response [42], but showed minimal overlap with the differential proteins identified in the present ChILI samples. This may imply that ChILI is not simply a consequence of the pharmacological action of ICI. A targeted analysis of checkpoint related proteins in exosomes from lung cancer patients 4–6 months after treatment with ICI revealed an increase in exosomal PD-L1 protein compared with healthy controls, but a reduction in PD-L1 protein in patients responding to therapy [45]. EV-associated PD-L1 protein can derive from tumor cells, some tissues, and circulating and tissue myeloid cells. In the present study, an increase in both gene expression in PBMC and EV-associated PD-L1 protein was observed for TOL compared to BC or AC groups, but no differences were seen solely due to therapy (AC vs BC). Since the increase in circulating PD-L1 protein is accompanied by an increase in PD-L1 gene expression in PBMC, circulating myeloid cells are one of the contributors of the increased production of and EV-associated PD-L1 protein in TOL. Hepatocytes constitutively express low level of PD-L1 which can increase significantly upon IFN stimulation therefore hepatocytes also contributed to the increased EV-associated PD-L1 in ChILI. This result suggests that increased expression of blood PD-L1 gene and protein can potentially serve as a biomarker of ChILI.
+In summary, these results provide valuable insights into the immune and molecular mechanisms underlying ChILI in cancer patients. These data represent the first transcriptomic and proteomic analyses in a prospective cohort with adjudicated ChILI cases and control cancer patients. We found an increase in transcript encoding PD-L1 as well as EV-associated PD-L1 in association with TOL and may assist in the stratification of patients on ICI. One limitation to the present study is the relatively small sample size. Particularly when comparing FU1 and FU2 and when considering the steroid effects. While some impact of steroid treatment is clear, effect of steroids on certain pathways in ChILI patients cannot be completely ruled out. The small number of timepoints may not allow the capture of the full dynamic changes in T cell clonality. The use of a single pathway analysis tool simplifies analysis, but may also introduce some analytical bias. Re-analysis of these kinds of datasets and integration with other published datasets using foundational models or other similar artificial intelligence approaches may yield additional valuable insights. Another limitation is the lack of paired liver tissue evaluations to provide further insight at the tissue site and a more complete mechanistic picture of ChILI. The identification of ChILI-specific gene signatures and proteins, however, provides a source of novel potential biomarkers of ChILI that could be further validated in future exploratory clinical studies.
+Below is the link to the electronic supplementary material.
+Supplementary file1 (DOCX 158 KB)
+Supplementary file2 (DOCX 16 KB)
+Supplementary file3 (XLSX 16 KB)
+Supplementary file4 (XLSX 5261 KB)
+Supplementary file5 (XLSX 17 KB)
+Supplementary file6 (XLSX 291 KB)
+Supplementary file7 (XLSX 55 KB)
+The authors are grateful to the study participants, the melanoma & renal cancer clinical nurse specialists at Nottingham University Hospitals NHS Trust and the team at NIHR Nottingham Biomedical Research Centre. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR or the Department of Health. Figure 1 is created in Biorender (https://BioRender.com).
+Conceptualization was done by CJ and GPA; supervision administration and resources were done by SKR, GPA, JIG, MFO, and TAL; methodology was done by SK, JQ, JF, EA, and Stuart A; analysis was done by JF, MS, DC, Seda A, and EO. All authors reviewed the manuscript.
+Funding was provided by Pfizer Inc. EA, SA, JIG, and GPA are supported by NIHR Nottingham Biomedical Research Centre (NIHR203310).
+All RNAseq data have been uploaded to GEO (https://www.ncbi.nlm.nih.gov/geo/): GSE287540, GSE287541. All proteomic data have been uploaded to MassIVE (https://massive.ucsd.edu/ProteoSAFe/static/massive.jsp): MSV000096928.
+CJ, SK, JQ, JDF, MS, DC, Seda A, EO, MFO, SKR, and TAL are Pfizer employees.
+Publisher's Note
+Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
+Guruprasad P. Aithal and Thomas A. Lanz have contributed equally to this work.
+Changhua Ji, Email: jichanghua5@gmail.com.
+Thomas A. Lanz, Email: thomas.a.lanz@pfizer.com
+Supplementary file1 (DOCX 158 KB)
+Supplementary file2 (DOCX 16 KB)
+Supplementary file3 (XLSX 16 KB)
+Supplementary file4 (XLSX 5261 KB)
+Supplementary file5 (XLSX 17 KB)
+Supplementary file6 (XLSX 291 KB)
+Supplementary file7 (XLSX 55 KB)
+All RNAseq data have been uploaded to GEO (https://www.ncbi.nlm.nih.gov/geo/): GSE287540, GSE287541. All proteomic data have been uploaded to MassIVE (https://massive.ucsd.edu/ProteoSAFe/static/massive.jsp): MSV000096928.