We obtained peripheral blood mononuclear cells (PBMCs) from patients with NSCLC from two clinical sites: University of California San Diego in San Diego (UCSD), and Institut Hospital del Mar d’Investigacions Mèdiques (IMIM) in Barcelona. Patient samples were collected under a UCSD ImmunoScape Immune Monitoring Protocol (HRPP# 150348; PI: S. Patel) and a CEIC-PSMAR-approved protocol at IMIM (2017/7174; PI: E. Arriola). All studies were approved by the IRB at La Jolla Institute for Immunology. The average patient age was 69.3 years, and 65% percent of the subjects were male. Approximately 16% of patients had advanced NSCLC (stage IV). All patients were treated with either Pembrolizumab or Nivolumab as the sole therapy. Patient characteristics are shown in Supplementary Table 1 . Response status was assessed after 6 months of treatment, with responders defined as patients who showed clinical benefit (decrease in tumor size or stable disease) for at least the first 6 months of treatment. Non-responders include patients who progressed or discontinued therapy (due to clinical deterioration or death) within the first 6 months of treatment. Blood from melanoma patients was obtained from the Biospecimen Repository Core Facility at The University of Kansas Cancer Center.

Directly-conjugated antibodies were purchased from Fluidigm and purified antibodies were ordered from the companies listed below. Conjugations were performed with the Maxpar X8 Multi-Metal Labeling Kit (Fluidigm). Cells were stained with 5mM Cisplatin (Fluidigm; live/dead), a metal-conjugated antibody cocktail, and 125nM Ir-Intercalator (Fluidigm). Cells were then resuspended at 1×10⁶cells/ml in 0.1×EQ™ Four Element Calibration Beads (Fluidigm), and the acquisition was performed on the CyTOF Helios2 (Fluidigm). Batch effects and signal drifting were minimized by tuning every 6hrs during long runs. To test for consistency between CyTOF acquisitions, a single healthy donor patient sample was included in each run. 100ml of heparinized blood was drawn from a healthy control donor, PBMCs were isolated and aliquots were frozen (90%FBS+10%DMSO) and stored at -80°C until use. Between every CyTOF run, one vial was taken for staining and acquisition.

CyTOF data were processed using a CyTOF workflow, and analyses were implemented using the CATALYST package in R (18). CyTOF data were arcsinh (inverse hyperbolic sine) transformed with a cofactor of 5, and batch-corrected using the quantile normalization method for the pooled distribution of each batch (a pair of sample and spike-in control) implemented in the function normalizeBatch from the cydar Bioconductor package (19). Self-Organizing-Map-based method FlowSOM (20) was used for clustering of cells. Visualization and high-dimensional reduction were performed using the UMAP method (21). Differentially expressed marker analysis was carried out using linear modeling, limma (22), and differential cell population abundance was determined by a Mann-Whitney test (p < 0.05). The Destiny package was used to perform diffusion map pseudotime for trajectory analyses (23).

To analyze peripheral immune cell populations, 0.5ml heparinized blood samples from additional healthy volunteers, NSCLC patients from UCSD and IMIM, and melanoma patients were stained and analyzed by flow cytometry. The following antibodies were used: Live/Dead Fixable Yellow (Invitrogen), CD14 (63D3), CD163 (GHl/61), CD274 (29E.2A3), CXCR3 (6025H7), CD33 (WM53), CD16 (3G8), Slan (M-DC8), CD273 (24F.10C12), CD9 (HI9a), HLADR (L243), CD64 (10.1), CD1c (L161), CD56 (5.1H11), CD3 (OKT3), CD19 (HIB19), CD66b (G10F5), Siglec-8 (7C9), CD279 (EH12.2H7), CCR7 (G043H7), CD45RA (HI100), CD4 (RPA-T4), CD45RO (UCHL1), CD69 (FN50), CD25 (BC96), CD127 (A019D5). Samples were acquired on a BD LSRII. Frequencies of monocyte subsets were correlated with regulatory T cell subset frequencies in the blood of the same patients.

TCGA survival data was linked to CD33 and CD9 expression in patients with lung adenocarcinoma, lung squamous cell carcinoma, and skin cutaneous melanoma by downloading data from http://www.oncolnc.org/ (24). Survival data up to 1500 days for patients with skin cutaneous melanoma, lung adenocarcinoma, and lung squamous cell carcinoma were used for survival analysis. Log-rank tests were performed using the survival package (25) and Kaplan-Meier curves were generated using the survminer package (26).

We performed ligand:receptor interaction analysis of T cells and monocytes in recently published single-cell RNA-sequencing data from normal lung tissue and adenocarcinoma tissue (27). We utilized the CellChat (28) algorithm to assess interactions between monocytes and regulatory T cells. To perform this analysis, we used normalized data and followed the official CellChat workflow for comparative analysis of multiple conditions.

To assess monocyte heterogeneity in response to anti-PD-1 therapy, we performed mass cytometry on peripheral blood mononuclear cell (PBMC) samples collected from non-small cell lung cancer (NSCLC) patients prior to treatment with anti-PD-1 therapy (n=26 patients), as well as 10 matched patient samples collected after treatment ( Figures 1A, B ). Detailed characteristics of subjects are described in the Methods section and listed in Supplementary Table 1 . PBMCs from patients were stained with a comprehensive monocyte-focused CyTOF panel ( Figure 1C ) derived from our study analyzing monocyte heterogeneity in healthy humans (29). To evaluate the interplay between monocytes and T lymphocytes, we added markers for quantifying the frequencies of CD4⁺ and CD8⁺ T cells, including the chemokine receptor CXCR3 (30). To examine the impact of immune checkpoint blockade on the phenotypes of the immune compartment overall, we also incorporated antibodies for PD-1, PD-L1, and PD-L2. We used healthy PBMC samples as spike-in controls and for batch correction of the data using cydar ( Supplementary Figure 1A ).

To define major cell types, we performed clustering on 5,359,566 live CD45⁺ cells with all surface markers via FlowSOM (20). Classical (cMo), intermediate (iMo), and nonclassical (nMo) monocytes grouped together ( Figure 1D ), and were distinguishable from other blood leukocytes based on both their expression of CD14 and CD16 and their lack of the DC-lineage marker CD1c ( Figure 1E and Supplementary Figure 1B ). Examination of the total frequencies of cMo, iMo, and nMo cells revealed no significant changes in these major compartments ( Figure 1F ), which motivated us to focus our studies on subsets of monocytes and monocyte-derived cells.

We thus sub-clustered monocytes (excluding dendritic cells) from the CD45⁺ cells ( Figure 1D ) for all samples ( Figure 2A ). After refining monocyte clusters based on median surface marker expression using consensus clustering, we identified 8 cMo subsets, 2 iMo subsets, and 5 nMo subsets ( Figures 2A–C ; Supplementary Figures 2A, B ). In particular, we discovered the presence of new nMo subsets (nMo_CD43^lo, nMo_CD9⁺, and nMo_CXCR3⁺PDL1⁺; Figures 2A–C ) that we did not observe in PBMCs from healthy humans (29). Although we observed these subsets in all patients, we noted variability in their frequencies amongst individual NSCLC patients ( Supplementary Figure 2C ), similar to what we’ve seen before (29). Nonclassical monocytes displayed enrichment of CD16, CD86, CD11a, and CD11c, but lower levels of CD14, CD11b, CD64, CCR2, and CD36 ( Figure 2B and Supplementary Figure 2A ). Notably, nMo clusters expressed less CD33 than most cMo clusters, except for cMo_CD33^lo. Four nMo subsets expressed Slan, a carbohydrate modification of P-selectin glycoprotein-1 (PSGL-1), and were further distinguished by differential expression of CD9. One nMo cluster was quite high for Slan enrichment (nMo_Slan^hi). Intermediate monocytes displayed dual expression of classical and nonclassical monocyte proteins including high expression of CD14, CD16, CD36, CD64, CD86, CD11a, CD11b, and CD11c, but lower expression of CCR2 and CD64 ( Figure 2B and Supplementary Figure 2A ). This profile is consistent with our previous analysis of healthy human PBMCs demonstrating that exclusion gating based on CCR2, CD36, HLADR, and CD11c improves the purity of intermediate monocytes (31).

Classical monocytes expressed CD14, CCR2, CD36, and CD11b ( Figure 2B and Supplementary Figure 2A ), yet certain subsets could be distinguished based on expression of CD33, a siglec that is preferentially expressed on peripheral myeloid cells (32), and the tetraspanin CD9, which is involved in lipid raft clustering. As CD33 is often used to identify cells on the entire monocyte lineage, the presence of CD33^lo monocytes indicates that this marker should not be used in isolation. cMo_CD9^-HLADR^lo and cMo_CD9⁺ appeared similar to CD9⁺ and CD9^- classical monocyte subsets that we previously reported (29). Two cMo subsets (cMo CD11a^-CD64^- and cMo CD9^-HLADR^lo had lower expression of the activation markers CD64, CD11a, and CD86 ( Figure 2B and Supplementary Figure 2A ), suggesting that these are either immature monocyte phenotypes or the newly emigrated monocyte precursors found in COVID-19 (33, 34). Ginhoux’s group recently identified the inflammatory DC3 subset, which expresses CD14, CD163, and binds IgE (35). Based on marker expression analysis, we observe DC3 cells in our NSCLC patient cohorts ( Figures 2A, B and Supplementary Figures 2A, B ). A dendrogram showing the relationships amongst all monocytes present in the NSCLC patients is presented ( Figure 2C and Supplementary Figure 2B ), highlighting increased levels of diversity amongst cMo compared to nMo.

Trajectory analyses using diffusion maps indicated two trajectory components: a monocyte maturation component and an immune regulatory component ( Figure 2D ). The first diffusion component correlated with monocyte maturing, in concordance with a continuous differentiation trajectory from cMo through iMo and into nMo (36, 37). The second diffusion component correlated with the expression of CXCR3 and PD-L1, distinguishing CXCR3⁺ monocytes from CXCR3^- subsets. These data suggest that CXCR3⁺ classical monocytes give rise to CXCR3⁺ nonclassical monocytes in a process parallel to the differentiation of CXCR3^- monocytes. To establish that CXCR3⁺ monocytes are single-nucleated cells, we confirmed that all CXCR3⁺ monocyte subsets had a similar signal intensity in the DNA intercalator channel (data not shown) and expressed low levels of lineage markers CD3, CD19, CD66b, CD161, and Siglec-8. Our trajectory analysis indicates that nMo_Slan^- monocytes give rise to nMo_Slan^hi monocytes ( Figures 2A, D ).

Given the importance of myeloid-specific expression of antigen-presenting molecules in anti-tumoral immunity (38), we examined myeloid expression of surface molecules that modulate adaptive immune responses in our NSCLC cohorts (7). HLADR was differentially expressed between monocyte subsets ( Figure 2E ). As expected, an iMo subset expressed the highest levels of HLADR (31); yet nMo_Slan^- also showed very high HLADR expression, suggesting that these monocyte subsets may be involved in antigen presentation. The high expression of HLADR in iMo is consistent with work in mice demonstrating that Ly6C^int monocytes express MHCII-associated genes CD74 and H2-aa (36). PD-L1 has been shown to be exclusively expressed by nonclassical monocytes in mice (39). However, we found that PD-L1, together with PD-L2 and their receptor PD-1, were expressed across different monocyte subsets, including cMo and iMo in NSCLC patients ( Figures 2C–E ). Our clustering analysis showed that they formed distinct, separate clusters from other monocytes (yellow island in bottom of UMAP in Figure 2A ).

Total CD14⁺ cMo frequencies are predictive of anti-PD-1 responses in melanoma patients (14), but we did not observe this in our cohorts of NSCLC patients ( Figure 1F ). To determine if specific monocyte subsets distinguished non-responders from responders amongst NSCLC patients treated with anti-PD-1, we split our monocyte UMAP between non-responders and responders ( Figure 3A ). We also quantified each monocyte subset amongst all patients ( Figure 3B and Supplementary Figure 2C ). Responders displayed significantly reduced frequencies of the cMo_CD33^lo subset and significantly increased frequencies of cMo_CD33^hi and nMo_CD9⁺ monocytes ( Figure 3C ). cMo_CD33^hi cells were significantly higher in responders, and cMo_CD33^lo cells were higher in non-responders ( Figure 3C ). There was no significant difference in either cMo_CD9⁺CD33⁺ or cMo_CD9^hiCD62L⁺ cells between NR and R ( Figure 3C ). cMo_CD33^hi cells were significantly higher in males than females ( Figure 3D ); this trend held true for both responders and non-responders ( Supplementary Figure 3 ). Males exhibited higher levels of DC3, regardless of ICB responsiveness, but this did not reach statistical significance ( Supplementary Figure 3 ).

To support the importance of CD33 and CD9 in determining anti-PD-1 ICB responses, and isolate other markers of interest in cMo and nMo subsets, we employed linear modeling via the limma and diffcyt packages, which apply high-resolution clustering and Bayesian statistical modeling (40). Consistent with our findings above, CD33 was the marker best able to distinguish responder and non-responder patients ( Figure 4A ). As a second independent analysis method, we utilized machine learning to examine monocyte markers that were most associated with responses to anti-PD-1 in our NSCLC cohorts. The random forest model was trained with the normalized marker expression from randomly selected cells over 20 iterations, and variable importance scores of the markers were calculated and scaled to 100% (41). Using this second method, we again found that CD33 was the highest ranked marker in terms of importance for predicting positive responses to anti-PD-1 ( Figure 4B and Supplementary Figure 4A ). Other key markers identified were IgE, Slan, CD86 and CD62L. We validated CD33 expression on cMo in PBMCs from responders and non-responders by flow cytometry ( Figure 4C ). While not statistically significant, the average expression of CD33 was higher in responders compared to non-responders ( Figure 4D ).

We next examined whether anti-PD-1 therapy altered the frequency of CD33^hi cMo. We compared baseline (pre-) and post-treatment samples collected from the same patients between 2 and 12 weeks after their first dose of anti-PD-1. Responders had increased cMo_CD33^hi levels at baseline, before treatment ( Figure 4E ). We found that anti-PD-1 therapy caused no change in CD33 expression on cMo, as post-treatment levels were similar to pre-treatment levels in patients ( Figure 4E ). To validate the utility of our findings in other cancer types, we analyzed an independent cohort of melanoma patients treated with anti-PD-1. We observed a confirmatory increase in CD33 expression on cMo in responders to anti-PD-1 therapy by flow cytometry ( Supplementary Figure 4B ). As CD33 was by far the most significant marker identified in terms of predicting anti-PD-1 responses, we next examined correlations between CD33 overall survival (OS) in our NSCLC cohorts. CD33 was associated with OS out to >24 months ( Figure 4F ). We also observed a significant correlation between PFS and CD33 expression on cMo in our cohorts ( Figure 4G ). Examination of lung adenocarcinoma data from the TCGA cohort ( Figure 4H ) showed that high CD33 expression was associated with increased survival. In contrast, CD33 expression was not associated with improved survival in lung squamous cell carcinoma data from the TCGA ( Figure 4I ). Survival amongst patients with skin cutaneous melanoma was associated with high CD33 expression in the TCGA cohort ( Supplementary Figure 4C ). These observations suggest that CD33 expression may be strongly associated with anti-PD-1 responses in specific cancer subtypes.

We next directed our attention towards CD9 because our findings linked nMo_CD9⁺ monocytes with anti-PD-1 responder status ( Figure 3C ). CD9 was moderately associated with anti-PD-1 responsiveness via our machine learning analysis ( Figure 4B and Supplementary Figure 4A ). Additionally, CD9 expression was linked to improved survival in patients with lung adenocarcinoma, but not in patients with lung squamous cell carcinoma, from the TCGA cohort ( Figure 5A ). Thus, CD9 may serve as an indicator of anti-PD-1 responsiveness in an NSCLC subtype-dependent manner.

CD86 was enriched in cMo_CD9⁺CD33⁺ monocytes, cMo_CD33^hi monocytes, and DC3 cells from responders ( Figure 5C ). CD86 expression was highest in cMo_CD33^hi cells, but was also elevated amongst cMo_CD9⁺CD33⁺ monocytes and DC3 ( Figure 5D ) in responders compared to non-responders. Our machine learning analysis highlighted CD86 as another marker associated with responsiveness to ICB therapy ( Figure 4B ), and we found that CD86 expression was higher amongst all monocytes from responders versus non-responders ( Figure 5B ). As CD86 contributes to antigen-presentation, these data imply an increased immune response by these cellular subsets in responders to anti-PD-1 therapy. This model is consistent with the finding that CD86 expression by some monocytes and DCs is linked to benefit from anti-PD-1 therapy in melanoma patients (42). These data also support the notion that CD9 plays a role in determining the anti-PD-1 response, as CD86 expression on nMo_CD9⁺ monocytes, although high, was not higher in responders than non-responders ( Figure 5D ).

To link monocyte subset functions with T cells in NSCLC, we performed flow cytometry for the 3 monocyte subsets shown in Figure 3C and regulatory T (CD4⁺FOXP3⁺CD25⁺; T_Reg) cells within the same patients. T_Reg cells are linked to tumor progression via their promotion of immune suppression (43), and we thus hypothesized that cMo_CD33^hi and nMo_CD9⁺ monocytes would negatively correspond with T_Reg frequencies. Both cMo_CD33^hi and nMo_CD9⁺ monocyte frequencies were significantly correlated with reduced numbers of T_Reg cells in the same NSCLC patients ( Figure 6A ).

We next analyzed a recently published scRNA-seq dataset from patients with lung adenocarcinoma (27) to explore how T cells interact with monocytes during NSCLC. In the Kim data set, scRNA-seq was performed on tumor biopsies from patients with early (Stage I) or advanced (Stage III) LUAD or adjacent normal lung tissues (Normal). We thus used the CellChat algorithm to quantitatively evaluate T_Reg cells cell and monocyte communication within this scRNA-seq dataset. The number of interactions between monocytes and T_Reg cells decreased in early cancer tissue compared to normal lung tissue, and diminished during progression from early to advanced cancer ( Figure 6B ). We also observed an increase in TNF signals from monocytes to T_Reg cells in lung tissue of advanced cancer versus healthy lung. Finally, HLA interactions were reduced from early to advanced disease, with loss of CD4 expression in advanced cancers.

PSGL-1 and CD86 signals from monocytes were diminished during cancer progression, suggesting a less immunologically active tumor ( Figure 6C ). Beta 2 integrin interactions with ICAM2 were also reduced, as are galectin9-CD44 interactions, which are important for T cell activation. CXCL16 is a putative biomarker for responses of NSCLC patients to Bevacizumab (anti-VEGF) (44), which is consistent with the reduced CXCL16-CXCR6 interactions during cancer progression herein ( Figures 6B, C ). Together, these data show significant loss of monocyte:T cell interactions during cancer progression that leads to a reduction in anti-tumoral phenotypes over time.