A total of 41 patients with early stage (I–III) NSCLC from subtypes adenocarcinoma (LUAD) and squamous cell carcinoma (LUSC) consented to tissue collection. A portion of their resected tumor and tumor-free lung tissue (same lobe for paired analysis) was collected. None of the patients received neoadjuvant chemoradiotherapy. Of note, from the 41 patients we isolated TAMs from 40, and collected tissue for non-tumor-associated macrophages (NTAMs) from matched lung from 34 (one of the patients did not yield enough NTAMs, so paired analysis was performed with n=33). Sample size based on previous studies,2 16 and demographic characteristics are presented in online supplementary table S1.

NSCLC tumor specimens and specimens from matched normal lung from the same patients were collected after general anesthesia and surgical resection. A representative portion of each specimen was immediately formalin-fixed and used for paraffin-embedding and generation of tissue microarrays (TMA), for determination of tumor infiltrating lymphocytes (TILs) score by CD8⁺ staining, as described in the Immunohistochemistry and TIL score determination. The rest was immediately processed for tissue disaggregation and macrophage isolation as follows. Briefly, samples were collected in cold RPMI (supplemented with penicillin/streptomycin, L-glutamine and Na⁺-pyruvate) and processed immediately after resection. Tissue was minced with a scalpel and digested enzymatically with 0.15 WU/mL of D-Liberase (Roche) and 800 U/mL of DNase I (Sigma-Aldrich) for 15 min at 37°C. Then it was disaggregated into a single-cell suspension by passing it through a 70 µm strainer and rinsing with cold buffer (1× phosphate-buffered saline (PBS), 2 mM EDTA, 0.5% bovine serum albumin (BSA)). Red blood cells were lysed using an ammonium chloride-based lysing reagent (BD Pharm Lyse solution, BD Pharmingen) for 15 min on ice. Cells were counted for viability using Trypan Blue (Sigma-Aldrich). Around 5 million cells, when available, were stained for Fluorescent Activated Cell Sorting (FACS) as follows: FcR non-specific binding was blocked using FcR block (Miltenyi) for 30 min on ice. Then a cocktail of prelabeled antibodies was added: glycophorin-A (Pacific Blue, BioLegend #349108, dilution 1/50), CD45 (fluorescein isothiocyanate, BioLegend #304038, dilution 1/66), CD14 (APC-H7, BD Biosciences #641394, MφP9, dilution 1/20), HLA-DR (Allophycocyanin-APC, BD Biosciences #347403, L243, dilution 1/80). Cells with antibodies were incubated on ice for 20 min. For FACS sorting, cells were stained with DAPI (Sigma) for live/dead gating. Gating strategy used was as follows: singlets ->glycophorin-A^– (to remove possible residual red blood cells)/DAPI^- (live cells) ->CD45⁺ ->HLA-DR⁺/CD14⁺ (online supplementary figure S1B). Sorting was performed using a BD FACS Aria II Cell Sorter (BD Biosciences). Purified macrophage populations were then collected in Trizol-LS (Sigma-Aldrich) and stored at −80°C until RNA extraction was performed.

Isolated macrophages in the range of 10,000–50,000 were subjected to total RNA extraction using miRNeasy RNA extraction micro kit (Qiagen) following manufacturer’s instructions. Absolute quantification using real-time quantitative PCR (qPCR) method was used to measure RNA concentration. A standard curve was created with serial dilutions of RNA from a known concentration obtained from monocytes isolated from fresh human blood by Ficoll gradients using CD14⁺-isolating MACS columns. Quantification was performed by beta-2-microglobulin presence using SyBR-Green primers (Sigma). RNA quality was assessed in the bioanalyzer using RNA 6000 pico chips (Agilent) and RNAs with a RIN quality value ≥8.50 were used for the pre-amplification and library preparation procedures.

ERCC spike-in mix (Ambion, Life Technologies) was added to 15 ng of starting RNA material. Ribosomal RNA was eliminated from total RNA by poly-A RNA positive selection using Poly(A)Purist MAG kit (Ambion, Life Technologies) following manufacturer’s instructions. Poly-A RNA was then pre-amplified using SeqPlex RNA Amplification kit for Whole Transcriptome Amplification (WTA, Sigma-Aldrich). Briefly, annealing mix was added to the poly-A RNA that was then denatured at 70°C for 5 min. Then the library synthesis buffer and RT enzyme were added, and RT was performed. The cycles of optimum amplification were then optimized using a small aliquot of the sample and amplification-mix and enzyme in a real-time qPCR using ROX dye and GelGreen. The optimal number of amplification cycles were achieved by proceeding two cycles into the amplification plateau. The rest of the sample was then amplified during the optimized number of cycles, in a mix of amplification mix and DNA amplification enzyme (WTA). Amplified product was cleaned up with ZR96 DNA clean up kit (Zymo Research) following manufacturer’s instructions. Final DNA was measured with nanodrop, double stranded DNA (dsDNA) was measured with dsDNA Quantifluor (Promega). Ratio of dsDNA/DNA was >75%. Postadaptor removal by enzymatic digestion was performed to 1 µg of DNA during 60 min at 37°C following by enzyme deactivation. Size selection and clean up was performed using Agencourt AMPure XP beads (Beckman Coulter). Beads were added to the samples in a ratio 1:1 (beads:sample) followed by three washes with 80% ethanol and eluted in Tris-EDTA buffer (TE). Final amplified DNA samples were measured by Nanodrop and size was evaluated in the bioanalyzer using High Sensitivity DNA chips (Agilent). More than 75% of amplified samples had a size of 150–450 bp. Quantification of dsDNA by Nanodrop (total DNA) and dsDNA Quantifluor to assure that >85% of the DNA amplified is dsDNA. Aliquots were taken in every step of the process and different quality checks were performed.

Amplified dsDNA was subjected to library preparation using TruSeq Nano DNA library prep kit (Illumina). Samples were subjected to end-repair during 30 min at 30°C. Size selection was performed using Agencourt AMPure XP beads in two steps: removal of large DNA fragments (>450 bp) with a ratio 1.6:1 of diluted beads:sample; removal of small DNA fragments (<150 bp) with a ratio 1:3.1 of undiluted beads:sample. DNA of interest was recovered in TE; 3’ ends of 150–450 bp DNA fragments were adenylated using A-Tailing Mix during 30 min at 37°C followed by enzyme deactivation. Then, Illumina adapters were ligated using Ligation Mix 2, Resuspension Buffer and the appropriate barcoded adapter in each case. As a last step, two sets of clean ups were performed using AMPure XP Beads: first in a 1.2:1 ratio and next in a 1.1:1 ratio. DNA was eluted with TE and PCR amplified using PCR Primer Cocktail and Enhanced PCR Mix as follows: 3 min at 95°C, and 8 cycles of 20 s at 98°C, 15 s at 60°C and 30 s at 70°C. As a final step, the final amplified libraries were cleaned up using Sample Purification Beads. The size of the library was determined by HS-DNA Agilent Bioanalyzer (150–450 bp). The amount of library was determined by Nanodrop and dsDNA Quantifluor. Quality checks were performed comparing with aliquots taken from the >450 bp and <150 bp fractions.

The single-end reads (50 bp length) generated by HiSeq2500 that passed Illumina filters were filtered for those aligning to tRNA, rRNA, adapter sequences and spike-in controls. The reads were then aligned to UCSC human genome (hg19) using TopHat (V.1.4.1).17 DUST scores were calculated with PRINSEQ Lite (V.0.20.3)18 and low-complexity reads (DUST >4) were removed from the BAM files. The alignment results were parsed via the SAMtools19 to generate SAM files. Read counts to each genomic feature were obtained with the HTSeq-count program (V.0.6.0)20 using the ‘union’ option. After removing absent features (zero counts in all samples), the raw counts were then imported to R/Bioconductor package DESeq221 to identify differentially expressed genes (DEGs) among samples. DESeq2 normalizes counts by dividing each column of the count table (samples) by the size factor of this column. The size factor is calculated by dividing the samples by geometric means of the genes. This brings the count values to a common scale suitable for comparison. P values for differential expression are calculated using negative binomial test for differences between the base means of two conditions. These p values are then adjusted for multiple test correction using the Benjamini-Hochberg algorithm to control the false discovery rate (V.1.14.1). We considered genes differentially expressed between two groups of samples when the DESeq2 analysis resulted in an adjusted p value (FDR) ≤0.05 and the fold-change (FC) in gene expression was ≥2 or ≤1/2 (log₂FC≥|1|), and only genes filtered by mean expression ≥10 normalized counts across the samples were considered. Cluster analyses including principal component analysis (PCA) and hierarchical clustering were performed using standard algorithms and metrics. The t-SNE was generated using the top 3000 hypervariable genes, as calculated in DESeq2 (V.1.16.1); this allowed for unbiased visualization of the DESeq2 normalized count data, using package Rtsne (V.0.13). In order to apply log₂ base transformation, a pseudocount (+1) was added to all the countable before log transformation. This is a usual procedure so genes with value 0 return 0 after log. Adding one does not bias the initial non-zero counts since we are expressing RNA-sequencing (RNA-seq) data as proportions. Weighted correlation analysis was completed using WGCNA (V.1.61)22 from the Log₂DESeq2 normalized count data matrix and the function TOMsimilarityfromExpr (beta=5) and exportNetworkToCytoscape, weighted=true, threshold=0.05. Networks were generated in Gephi (0.92)23 using Fruchterman Reingold and Noverlap functions. The size was scaled according to the average degree as calculated in Gephi. The color was then curated given a significant (defined by WGCNA) correlation with STAT1. Purified population analysis of bulk CD8⁺ cells was previously reported2 and converted to transcripts per million (TPM).16 Data were completed as above using the TPM, the TPM values are provided in online supplementary table S5. Cell cycle and cytotoxicity gene signatures were taken from murine signatures (cluster I and cluster III),24 and our human lung cancer studies.3 Murine transcripts were converted to a direct homolog if the gene symbol was identical, otherwise genes were discarded. Signatures are provided in online supplementary table S5. Differential expression was completed as above, correcting for a batch using DESeq2. Significant genes were selected based on adjusted p value (FDR) ≤0.05 and the FC in gene expression was ≥2 or ≤1/2 (log₂FC ≥ |1|).

RNA-seq data have been deposited on Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/geo/) under accession number GSE116948.

Analysis of molecular pathways affected by DEGs was performed using Ingenuity Pathway Analysis tool (IPA, Qiagen). Genes filtered by basemean ≥10, FDR ≤0.05 and log₂FC ≥ |1| were loaded into IPA software. A total of 1038 DEGs were loaded in the TAMs/NTAMs comparison and 222 DEGs were loaded in the chemokine (C-X-C motif) ligand 9 (CXCL9) high versus CXCL9 low TAMs comparison. No extra filters were applied. Canonical pathways and upstream regulators were analyzed. Upstream regulators were ordered only among cytokines and growth factors by lowest p value (Fisher’s exact t-test). Canonical pathways were ordered by lowest p value (Fisher’s exact t-test). Gene set enrichment analysis was completed as previously described,16 using Qlucore (V.3.2), using the SNR setting to compare between two biological groups.

For the M1/M2 staining with CXCL9 and matrix metallopeptidase 12 (MMP12) in TAMs, disaggregated cell suspensions from tumor and normal lung specimens were plated in several wells of μ-slide 8-well tissue culture-treated sterile chambers (#IB-80826, Ibidi). Cells were incubated in free-serum RPMI medium at 37°C for 3 hours to allow the macrophages to attach. After washing, they were fixed and permeabilized with a solution of 0.1% Triton X-100, 4% paraformaldehyde (ethanol-free) in PBS for 10 min at 4°C. Cells were stained with CD68 (mouse monoclonal antihuman CD68, clone PG-M1, #M0876, Dako; dilution 1/100), MMP12 (rabbit polyclonal antihuman MMP12 #NBP1-31225, Novus Biologicals; diluted 1/100) and CXCL9 (goat polyclonal antihuman CXCL9 #AF392, Novus Biologicals; diluted 1/20) overnight at 4°C and after washing they were incubated with secondary (donkey antirabbit Alexa405, donkey antimouse Alexa488, donkey antigoat Alexa647; all at 1/250) for 1 hour at room temperature in darkness. Washed preparations were briefly incubated with Sytox Orange for nuclear staining in tris-buffered saline and preserved in glycerol:antifade PBS (8:2) until visualization.

For the CXCR3 and CXCL9 staining in T cells, disaggregated cell suspensions from tumor and normal lung specimens were mildly centrifuged in slides using a cytospin and fixed and permeabilized as above. Staining was performed with CXCR3 (mouse monoclonal antihuman CD183/CXCR3, #557183, clone 1C6/CXCR3, BD Biosciences, dilution 1/500) and CXCL9 (goat polyclonal antihuman CXCL9 #AF392, Novus Biologicals; diluted 1/20) followed by secondary (donkey antimouse Alexa-Fluor 488, donkey antigoat Alexa-Fluor 647; all at 1/250) and nuclear staining with DAPI. Washed preparations were then mounted in Mowiol and left dry overnight at room temperature in darkness. Confocal images were taken using a SP8 Leica Confocal Microscopy.

Density of TILs was assessed by immunohistochemistry (IHC) using formalin-fixed paraffin-embedded TMAs. Triplicate 1 mm areas representative from each tumor were selected by pathologist review and processed into TMAs using the Aphelys Minicore 2 system (Mitogen, UK). TMAs were performed for tumors from both the cohort of patients of NSCLC adenocarcinoma and squamous cell lung carcinomas source of the TAM’s RNA-seq data in the manuscript (n=40) and for the archive-cohort of patients of NSCLC adenocarcinoma (2007–2011, n=460, data available for n=393), both from Southampton University Hospital. Serial 4 µm TMA sections were stained for the immunological markers CD8 (clone C8/144B, # IR623, Dako, Denmark) and CD14 (clone EPR6353, #114-R15 Merck, USA), the chemokine CXCL9 (clone 49106, #MAB392, R&D systems, Bio-Techne), and associated ligand CXCR3 (clone 49801, #MAB160, R&D systems, Bio-Techne). Antibody detection and visualization was performed with either high pH (CD8), or low pH (CXCL9, CXCR3) heat induced epitope retrieval using EnVision FLEX+ system and DAB as the chromogenic substrate. Optimal antibody conditions were determined in a diagnostic IHC laboratory using automated Dako Link 48 platforms and standardized protocols. For CD14 IHC, the standard protocol for OptiView DAB IHC Detection Kit by Ventana Medical Systems (USA) was used. For the RNA-seq cohort (n=40), TILs (CD8⁺), CXCL9⁺ and CXCR3⁺ cells were quantified using an Olympus DotSlide, as positive cells per field using the average of nine high-power (×400) fields across representative areas of each tumor, to allow for intratumoral heterogeneity. For the larger adenocarcinoma archive cohort TILs score was assessed, instead of counting cells, as TIL^high (CD8⁺ was diffuse, present in >80% of tumor/stroma); TIL^intermediate (CD8⁺ patchy, present in 20%–80% of tumor/stroma) or TIL^low (CD8⁺ weak/absent, present in <20% of tumor/stroma). CXCL9⁺ cells were assessed using the same criteria. Pictures were taken on a Zeiss AxioCam MRc5 microscope (Zeiss, Cambridge, UK) and Zeiss Axiovision software (V.4.8.1.0; Zeiss).

Two independent lung adenocarcinoma patient cohorts consisting of n=460 patients with available data for n=393 (2007–2011 archive-cohort from Southampton University Hospital) and n=511 patients with available survival data for n=495 (The Cancer Genome Atlas (TCGA)-LUAD) patients were analyzed retrospectively for survival stratifying them by their CXCL9 expression. Expression of CXCL9 in Southampton archive cohort was assessed at the protein level, by IHC and scored, as explained above. Expression of CXCL9 or CXCR3 for the TCGA-LUAD cohort is referred to their RNA-seq data from whole tumor available from TCGA. In both cases, the patients in the top 10% percentile and bottom 10% percentile of expression (of CXCL9 or CXCR3) were categorized as high and low, respectively. Percentile of 10% was selected after visualization of the data distribution in a plateau where the top 10% of the data points are outside the plateau, as well as the bottom 10% of the data points.

Validation of findings was assessed in the dataset of lung adenocarcinoma (LUAD, RNA-seq data n=511) from TCGA consortium. The RNA-seq methodology and processing have been described by TCGA25 26; mRNA data for CXCR3 and CXCL9 expression were plotted and statistical significance was evaluated by Spearman’s correlation analysis.

For single-cell transcriptomics, tumor and lung samples were first dispersed (as above) and cryopreserved in freezing media (50% complete RMPI (Fisher Scientific), 40% human decomplemented AB serum, 10% dimethyl sulfoxide (both Sigma)) for two patients (online supplementary table S1). Cryopreserved samples were thawed, washed and prepared for staining. Cells were first incubated at 4°C with FcR block (Miltenyi Biotec) for 10 min, prior to staining with a combination of anti-CD45-AlexaFluor700 (HI30; BioLegend, 1/20), anti-CD14-BV421 (M5E2, BioLegend, 1/20), anti-HLA-DR-BB515 (G46-6, BD Bioscience, 1/20), anti-CD3-APC-Cy7 (SK7; BioLegend, 1/20), anti-CD56-Pe-Cy7 (HCD56; BioLegend, 1/20), CD19/20-PE/Dazzle (HIB19/2H7; both 1/40, BioLegend), for 30 min at 4°C. Live/dead discrimination was completed with potassium iodide immediately prior to acquisition. Samples (online supplementary table S1) were sorted as described in online supplementary figure S3A into low retention tubes (Thermo Fisher Scientific #3434) containing 500 µL suspension containing 50% MACS buffer (PBS containing 2% fetal bovine serum (FBS) 2 mM EDTA), 50% FBS and 200 U of Recombinant RNase Inhibitor (Takara) using a 100 micron nozzle on a FACS Aria-II (BD Biosciences). Samples were then gently vortexed and maintained at 4°C. Samples were processed using 10× v2 chemistry as per manufacturer’s recommendations.16 Barcoded RNA was collected and processed following manufacturer’s recommendations, as described previously. Libraries were sequenced on a HiSeq2500 (Illumina) to obtain 100 and 32 bp paired-end reads using the following read length: read 1, 26 cycles; read 2, 98 cycles and i7 index, 8 cycles (online supplementary table S1). Raw 10× data (online supplementary table S1) was processed as previously described,16 merging multiple cell types with cellranger aggr (V.2.0.2). The merged data were transferred to the R statistical environment for analysis using the package Seurat16 27 (V.2.3.0). Only cells expressing >200 genes and genes expressed in at least 3 cells were included in the analysis. The data were then log-normalized and scaled per cell and variable genes were detected. Transcriptomic data from each cell was then further normalized by the number of UMI-detected and mitochondrial genes. A PCA was then run on variable genes, and the first six principal components (PCs) were selected for further analyses based on the SD of PCs, as determined by an ‘elbow plot’ in Seurat. Outlier cells representing contaminating cells were removed and the analysis recalculated. Cells were clustered and visualized using the FindClusters function in Seurat with default settings, resolution=0.6 and 6 PCs. Differential expression between clusters was determined by converting the data to counts per million and analyzing cluster-specific differences using MAST (q<0.05, V.1.2.1).16 28 29 A gene was considered significantly different, only if the gene was commonly positively enriched in every comparison for a singular cluster.16 30 Further visualizations of exported normalized data were generated using the Seurat package and custom R scripts. Average expression across a cell cluster was calculated using the AverageExpression function, and downsampling was achieved using the SubsetData function (both in Seurat). Average expression data were clustered using average linkage and heatmaps were visualized in Qlucore as above.

Monocytes and CD3⁺ T lymphocytes were extracted from peripheral blood mononuclear cell (PBMC) cones obtained from the blood bank, by sequential positive selection using CD14⁺ MACS beads (130-050-201) followed by CD3⁺ MACS selection (130-050-101) on the CD14^- fraction. Monocytes were resuspended in RPMI containing 10% FBS at 1 million per mL and cultured for 12 days in granulocyte-macrophage colony-stimulating factor (GM-CSF) at 50 U/mL. CD3⁺ isolated cells were frozen and stored at −80°C until the day before use. Macrophage activation and polarization was completed using either IFN-γ and LPS (20 and 20 ng/mL) or IL-4 (20 ng/mL) on days 13 and 15 and uptake experiments completed on day 16 (equating to 72 hours postinitial activation). CD3⁺ cells were defrosted and rested in RPMI containing 10% FBS for 24 hours before use.

Bodipy-FL C16 (ThermoFisher #D3821) was used in conjunction with flow cytometry to assess competitive uptake. Macrophage culture media was removed to limit cytokine carry over. CD3⁺ lymphocytes and differentiated macrophages were combined in a 1:1 ratio in fresh RPMI containing 10% FBS. Cells were treated at a final concentration of 500 nM Bodipy FL C16 (reconstituted in FA Free BSA) for 30 min at 37°C. Uptake was quenched using ice cold 10 μM fatty acid free BSA. Supernatants were transferred to 4 mL FACS tubes and adherent cells detached by trypsinization and gentle scraping. Cells were spun down and washed once in PBS.

Live/Dead staining was completed using zombie violet viability kit (#423113) and cells labeled for CD3 (#344827), CD8 (#344710), CD103 (#350216), HLA-DR (#307616) (all purchased from BioLegend) and CD14 (BD Biosciences: #557831) before proceeding to flow cytometry. T cells with T_RM markers were gated as CD14^-/HLA-DR^-, CD3⁺/CD8⁺/CD103⁺ as outlined in online supplementary figure S6.

Monocytes extracted from blood were differentiated into macrophages using GM-CSF over 5 days; 100 nM of either Scramble (Qiagen 1027287) or 33 nM each of FABP3, FABP4, FABP5 (Qiagen 1027415) were transfected using Hyperfect (Qiagen). Bodipy treatments were completed 24 hours post-transfection using 10 μM Bodipy FL C16 (ThermoFisher; D3821) for 30 min. Uptake was quenched using ice cold PBS containing fatty acids free BSA and washes were completed using PBS. Cells were detached and Bodipy uptake assessed by flow cytometry using zombie violet for the exclusion of dead cells. Mean fluorescence intensities under each treatment were compared with small interfering RNA (siRNA) scramble control to assess the percentage knockdown.

Comparison between two groups were assessed with a Mann-Whitney U test for non-parametric samples. Multigroup comparisons were assessed by one-way analysis of variance with Kruskal-Wallis test. Wilcoxon signed-rank two-tailed test was used for paired non-parametric analysis. Kaplan-Meier survival curves were tested statistically using the Mantel-Cox log-rank test for two groups comparison or log-rank test for trend when more than two groups were compared. GraphPad PRISM 7 was used for all statistical analysis. For RNA-seq analysis, p values for differential expression are calculated using the negative binomial test for differences between the base means of two conditions, and adjusted for multiple test correction using the Benjamini-Hochberg algorithm to control the false discovery rate using DESeq2. Visualizations including PCA, gene set enrichment analysis (GSEA), t-distributed Stochastic Neighbor Embedding (t-SNE), hierarchical clustering and heatmaps were performed using R V.3.3.2 and Qlucore (V.3.2). Flow cytometry data were calculated using Flowjo (V.10.4).

We compared the genome-wide transcriptional profile of CD45⁺CD14⁺HLA-DR⁺ cells that includes predominantly macrophages (confirmed by single cell sequencing in figure 3) and to a lesser extent, monocytes and neutrophils as well as a recently discovered small population of mature cDC2 dendritic cells.31 We will refer to these cells as TAMs (we acknowledge that we may be excluding rare populations of macrophages that do not express any CD14), isolated from lung tumor and NTAMs from adjacent non-involved lung tissue samples (NTAMs) obtained from patients (paired analysis, n=33, online supplementary table S1) with treatment-naïve NSCLC (figure 1A and online supplementary figure S1A,B). This paired comparison is important to identify the molecular features specific to TAMs as they share with NTAMs the same genetic background, residency in the same end organ as well as any other potentially confounding factors (ie, gender, age, smoking, comorbidities, etc), which might otherwise bias our evaluation.

Bulk RNA-seq analysis revealed major differences in the TAM molecular program; >1000 transcripts (1038) were expressed differentially in TAMs compared with NTAMs (figure 1B and online supplementary table S2), and t-SNE plots showed that TAMs clustered separately from NTAMs (figure 1C, online supplementary figure S1C). IPA revealed strong activation of both protumor and antitumor functions linked to the M2-like and M1-like macrophage gene profiles, respectively32 (figure 1D and online supplementary table S2). Transcripts associated with M2 protumor functions, such as angiogenesis (vascular endothelial growth factor signaling), metalloprotease activity, fibrosis and cancer metastasis, were expressed at higher levels in TAMs compared with NTAMs (figure 1E), which was further confirmed at a more global level, by GSEA (figure 1F and online supplementary table S2). As expected, M2-inducing cytokines like IL-4, IL-10, IL-13 and transforming growth factor-β were among the top upstream regulators of TAMs (compared with NTAMs) but surprisingly, so were typical M1-inducing cytokines such as IFN-γ and tumor necrosis factor (TNF)-α33 (IPA, figure 1G and online supplementary table S2). This finding might explain the simultaneous activation of both pro-inflammatory and antitumor pathways in TAMs, such as granulocyte adhesion and diapedesis and signaling of leukocyte extravasation, that were predicted by IPA (figure 1D). Together, these data indicate that TAMs in lung cancer expressed higher levels of transcripts linked to both an M2-gene profile, associated with protumor functions, and an M1-gene profile, involved in promoting antitumor T-cell responses.

We wondered whether bulk RNA-seq allowed stratification of TAMs in tumor samples according to relative enrichment of M1-like or M2-like signatures (‘hot or cold’), in order to establish differential protumor or antitumor roles. The expression pattern of known M2 genes (n=122 Methods' section34) that were significantly upregulated in the TAM population (~40% of known M2 genes), which included CD209, IL10, WNT5A and MMP12, was homogeneous across all samples, indicating a uniform M2 signature in TAMs (figure 2A and C and online supplementary table S3). In contrast, the expression pattern of known M1 genes (n=116 Methods' section34) that were significantly upregulated in TAMs (~31% of known M1 genes), which included CXCL9, CXCL10, CXCL11, CXCL12, STAT1 and AIM2, was highly heterogeneous and defined a clear subgroup of patients that exhibited ‘M1 hot’ features in their TAMs (figure 2B and C and online supplementary table S3).

We wondered whether in tumors with these M1^hot TAMs, there would be a reciprocal relationship with M2 TAMs which would show weaker M2 signatures. We selected CXCL9 to define M1^hot TAMs. CXCL9 is a chemokine previously known as monokine induced by IFN-γ, overexpressed by macrophages after stimulation by IFN-γ (the key cytokine behind M1 signature activation) and is involved in recruitment of antitumor T cells as well as having anti-angiogenic roles and has been suggested as a novel target for cancer therapy.35 We thus correlated the expression of CXCL9, as one of the strongest M1 marker genes in macrophages, with key M2 marker genes (CD209, ADORA3, STAT6, SOCS3, IL10 and IRF4) (figure 2D, online supplementary figure S2). As before, the M2 markers were homogeneously distributed across tumors, and showed no correlation to the CXCL9 (M1 marker gene) expression. The expression pattern of CXCL9 and MMP12 (a top DEG in figure 2A, strongly associated with M2 signatures36) in TAMs was not linked to their expression in matched NTAMs (figure 2E). Overall, these results indicate that TAMs in all lung cancer tumors show a strong M2 gene profile. IPA analysis suggests that this is likely due to the presence of pro-M2/Th2 cytokines such as IL-4 and IL-13, in the tumor microenvironment (figure 1G). In some tumors (25%–35%), however, the concomitant presence of pro-M1/Th1 cytokines (TNF-α and IFN-γ, figure 1G) may also be driving the M1 gene profile overexpression in TAMs.

Next, we wished to ascertain whether the TAM population was a mixture of different proportions of cells manifesting either an M1 or M2 phenotype (as suggested by most of previous literature) or whether individual TAM cells might express both the M1 and M2 signatures. Therefore, we performed single-cell RNA-seq analysis in purified macrophage populations (online supplementary figure S3A, table S1, Methods' section) isolated both from tumor and adjacent normal lung tissue from two additional patients with early stage lung cancer.

Analysis of the single-cell transcriptomic profile of ~9000 cells (Methods' section), revealed 8 distinct clusters (figure 3A and online supplementary figure S3B,C). As expected, transcripts with increased expression in bulk populations of NTAMs compared with TAMs (shown in figure 1B) were expressed at higher levels by single cells in the NTAM-enriched clusters (cluster 0 and 4), and the transcripts with increased expression in bulk populations of TAMs compared with NTAMs were expressed at higher levels by individual cells in the TAM-enriched clusters (cluster 2 and 3) (figure 3B,C and online supplementary figure S3C). While M2-like signature genes were expressed at higher levels (and in a higher proportion of cells) in both the TAM-enriched clusters (cluster 2 and 3), the M1 signature genes were expressed at higher levels (and in a higher proportion of cells) only in cells in cluster 3 (figure 3D–F and online supplementary table S4). The transcripts enriched in cluster 3 were also seen in our bulk population data, tightly co-expressed with STAT1, a known master regulator mediating the response to IFN-γ and activating the expression of M1-related genes (figure 3G). Overall, our single-cell transcriptome data confirm our bulk RNA-seq analysis but additionally, demonstrate that M1-related and M2-related genes can be strongly co-expressed in the same individual cells.

Confocal microscopy showed co-expression in TAMs of CXCL9 and MMP12 proteins, which we used as canonical markers of M1-like and M2-like phenotypes, respectively.34 NTAMs, expressed very low levels of both markers (figure 3H, online supplementary figure S3D). Together, these findings confirmed the existence of a TAM subset with dual M1-like and M2-like phenotype, overexpressing both gene profiles (M1^hot), in addition to a TAM subset with exclusive M2-like phenotype, only overexpressing the M2 gene profile (M1^cold).

To understand the potential functional impact of ‘M1^hot TAMs’ on antitumor immunity, we classified our cohort of cancer subjects based on the expression of M1 marker genes such as CXCL9, CXCL10, CXCL11, CXCL12, STAT1, FAM26F (heatmap, figure 4A), choosing CXCL9 as the candidate gene to classify tumors into M1^hot (top 25th percentile) M1^intermediate (25th–75th percentile) and M1^cold (bottom 25th percentile) tumors (figure 4A). The expression levels of CXCL9, CXCL10 and additional M1 marker genes were significantly higher in M1^hot TAMs (but not in M1^cold TAMs) than in NTAMs, while the transcript levels of CD209, MMP12 and additional M2 marker genes were higher in both M1^hot and M1^cold TAMs, when compared with NTAMs (figure 4B,C and online supplementary figure S4A). In our patients, M1 status of TAMs (M1^hot or M1^cold) was not related to gender, tumor stage or histological subtype of NSCLC and M1^hot status was not more frequent in ‘never-smokers’ (online supplementary figure S4B).

We next compared M1^hot TAMs with M1^cold TAMs to understand whether transcriptomic differences relate to outcomes in patients with lung cancer. We found that 222 transcripts were differentially expressed between M1^hot versus M1^cold TAMs (figure 4D); transcripts expressed at higher levels in M1^hot TAMs were involved in pathways linked to antitumor T-cell immune responses such as: the recruitment of T_H1 T cells, by expressing more CCL5, CXCL9, CXCL10 and CXCL11, antigen presentation, expansion of T cells, cytotoxicity and differentiation of effector T cells (IPA, figure 4E and online supplementary table S5). We found that M1^hot tumors had a significantly higher density of CD8⁺ TIL compared with M1^cold tumors (figure 4F, n=40). Furthermore, GSEA showed that expression of M1 signature genes in TAMs was strongly associated with a higher CD8⁺ T-cell density in tumors (q=0.04), whereas the expression of M2 signature genes in TAMs was not associated with either a higher or lower CD8⁺ T-cell density (q=0.75) (figure 4G). Taken together, our data indicate that while performing M2 functions, M1^hot TAMs appear to be involved in the recruitment and proliferation of T cells and therefore could shape the quality or magnitude of the antitumor response.

We next sought to determine the mechanism by which M1^hot TAMs could influence TIL density and hence, survival outcomes. Key defining transcripts of M1^hot TAMs encode for potent chemokines (CXCL9, CXCL10, CXCL11, CXCL12) that attract activated T_H1 T cells.37 38 Therefore, we asked whether the presence of transcriptomically M1^hot TAMs was reflected in CXCL9 protein expression and whether the density of CXCL9 protein-expressing TAM correlated with a CD8⁺ T-cell density. Immunostaining on TMA of tumor samples (figure 5A) showed that high CXCL9 protein expression in TAM was indeed associated with higher expression of CXCL9 mRNA by TAMs (ie, presence of M1^hot TAMs) (figure 5B) and a greater CD8⁺ T-cell density (figure 5C). CXCL9 is a known chemoattractant for T cells expressing CXCR3.37 We observed that a higher expression of CXCL9 transcripts in TAMs was associated with a higher infiltration of CXCR3⁺ cells in the tumor (CXCR3^high, figure 5D). Confocal microscopy on cells disaggregated from lung tumors showed that CXCR3 co-localizes with CXCL9 in the membrane of T cells (figure 5E). Taken together, these results suggested that there was a strong CXCL9-dependent infiltration of CXCR3-expressing CD8⁺ T cells in tumors (figure 5A and C–E) and that M1^hot TAMs were the main source of CXCL9 (figure 5B).

We hypothesized that M1^hot TAMs are associated with better prognosis for patients with lung cancer, since they link with higher CD8⁺ T-cell tumor infiltration. We assessed the survival outcome for patients whose tumors were classified by the density of cells expressing CXCL9 (a representative M1-like gene) into M1^hot, M1^intermediate and M1^cold (Methods' section). This assessment was done in an independent, large cohort of predominantly early stage patients with lung cancer (total n=393; a majority of them stage I to IIIA, online supplementary table S6) who had been diagnosed from 2007 to 2011 and followed up until 2016, a cohort with similar characteristics to the one used to generate the transcriptomic analysis of this study. Consistent with our finding that M1^hot TAMs are linked to a stronger antitumor immune response, M1^hot tumors were associated with a 48% survival outcome at 10 years, compared with a 24% 10-year survival in M1^cold tumors. The two-group comparison (M1 hot vs M1 cold) is shown in figure 5F (n=138, p=0.0006, Methods' section; Kaplan-Meier plot with log-rank test p value), and the three group comparison (including intermediate group) is shown in online supplementary figure S5A (n=393, Methods' section; Kaplan-Meier plot with log-rank test p value).

In order to broaden the validity of this important finding, we interrogated the RNA-seq dataset published by TCGA. We classified tumors based on the expression of CXCL9 transcripts (Methods' section), as M1^hot and M1^cold; we then compared the survival of patients classified as having M1^hot or M1^cold tumors from both the TCGA lung cancer adenocarcinoma dataset (TCGA-LUAD, total n=495 patients with complete survival information; M1^hot=49, M1^cold=49) and from data available from our Southampton cohort (n=393). Importantly, the M1^hot tumors showed better survival outcomes in both TCGA and Southampton datasets (figure 5F,G and online supplementary table S6 including a univariate and multivariate analysis of overall survival). Using the TCGA database, we confirmed that in TCGA-LUAD, CXCL9 transcript levels correlated strongly with CXCR3 transcript levels, reflecting the presence of higher numbers of T cells (n=511, figure 5H), which validated our observation in figure 5D. Higher expression of CXCR3 transcript levels was also linked to better survival (figure 5I). Overall, these results suggested that an M1^hot status of a tumor is positively linked to better survival outcomes in lung cancer through recruitment of a better antitumor TIL response.

In patients with solid tumors, it has recently been shown that the density of intratumoral CD8⁺ T_RM is associated with better survival outcomes.2 Therefore, we asked if M1^hot tumors were also associated with a higher density of T_RM cells in tumors. In our historical cohort of patients with lung cancer (n=393), whose tumors were preclassified on the basis of the density of cells expressing CXCL9 (M1^hot, M1^intermediate, M1^cold), we determined the density of cells expressing CD103, a marker of T_RM cells (T_RM ^high, T_RM ^int or T_RM ^low). We found a clear positive association in that T_RM ^high tumors were frequently also M1^hot tumors (34% vs 5% among M1^cold tumors). Similarly, T_RM ^low tumors were mainly also M1^cold tumors (80% vs 28% among M1^hot tumors) (figure 6A and online supplementary table S6). To directly assess the association between M1^hot TAMs and T_RM cells, we compared the transcriptome of tumor-infiltrating CD8⁺ T cells2 from the same patients with cancer in whom we had assessed the TAM phenotype as M1^hot and M1^cold (figure 6B and online supplementary table S5). Confirming our previous data, we found that in tumors with M1^hot TAMs compared with M1^cold TAMs, the tumor-infiltrating CD8⁺ T cells showed increased expression of genes strongly enriched for T_RM signature genes (eg, ITGAE, RBPJ) in addition to cell cycle and cytotoxic/cytokine signature genes (figure 6B,C). This result suggested that M1^hot TAMs were tightly linked to robust T_RM responses in the tumor.

We next asked how M1^hot TAMs might influence or modulate T_RM responses in the tumor microenvironment. M1^hot TAMs, through expression of CXCL9, 10 and 11, may enhance the local mobility of T_RM cells expressing CXCR3. However, further processes are required to sustain the antitumor activities of the recruited T_RM cells and a possible mechanism was suggested by our observation that the M1^hot TAMs showed reduced expression of transcripts encoding for the fatty acid binding proteins FABP3, FABP4 and FABP5, when compared with M1^cold TAMs (figures 4D and 6D). It is known that for survival in the tissues, T_RM cells depend on uptake of essential nutrient fatty acids through FABP4/5.39 Also, it has been shown that macrophages treated with IL-4, (‘M1^cold’), increase their fatty acid uptake.40 Hence, we hypothesized that M1^cold TAMs will be more efficient than M1^hot TAMs in competing for fatty acids in the tumor microenvironment (ie, due to higher expression of FABP3, 4, 5), and hence will outcompete T_RM cells for this essential nutrient, thus compromising long-term maintenance of T_RM cells in tumors.

We tested this hypothesis experimentally by treating blood-derived macrophages (BDMs) with IFN-γ and LPS or with IL-4 in order to mimic M1^hot and M1^cold TAMs, respectively.34 M1^hot-like BDMs (IFN-γ and LPS) showed a reduction in fatty acid uptake when compared with M1^cold-like BDMs (IL-4) (figure 6E); similar results were observed when FABP3, 4 and 5 were knocked down using siRNA to mimic the expression pattern observed in M1^hot TAMs (online supplementary figure S6C). We then co-cultured M1^hot-like BDMs or M1^cold-like BDMs with T cells isolated from blood of the same donors and observed that CD8⁺CD103⁺ T cells showed a significantly increased uptake of fatty acids when co-cultured with M1^hot-like BDM, compared with co-culture with M1^cold-like BDM (figure 6F). CD8⁺CD103^- T cells and the rest of CD3⁺ T cells used as controls, did not show significant differences in lipid uptake, regardless of the type of BDM with which they were co-cultured.

Our experiments suggested that M1^cold TAMs may reduce the survival and maintenance of T_RM cells in tumors (80% of M1^cold TAM tumors are T_RM ^low) and that a possible mechanism may be intercellular competition for lipid nutrients.