Natural variation of macrophage activation as disease-relevant phenotype predictive of inflammation and cancer survival

Although mouse models exist for many immune-based diseases, the clinical translation remains challenging. Most basic and translational studies utilize only a single inbred mouse strain. However, basal and diseased immune states in humans show vast inter-individual variability. Here, focusing on macrophage responses to lipopolysaccharide (LPS), we use the hybrid mouse diversity panel (HMDP) of 83 inbred strains as a surrogate for human natural immune variation. Since conventional bioinformatics fail to analyse a population spectrum, we highlight how gene signatures for LPS responsiveness can be derived based on an Interleukin-12β and arginase expression ratio. Compared to published signatures, these gene markers are more robust to identify susceptibility or resilience to several macrophage-related disorders in humans, including survival prediction across many tumours. This study highlights natural activation diversity as a disease-relevant dimension in macrophage biology, and suggests the HMDP as a viable tool to increase translatability of mouse data to clinical settings.

C linical and pharmaceutical researchers are concerned with the lack of relevance and low reproducibility of findings obtained in standard mouse models [1][2][3] . The translation of mouse data from bench to bedside is challenging [4][5][6][7][8] and clinical trial success rates continue to be low 1,9 . Among the possible reasons, the large inter-individual variation of immune system variables in human populations is cited 10,11 . This natural variation has been shown to broadly impact on pathophysiology, for example, disease resilience, tolerance [12][13][14] and vaccination responses 10,15 . System level analyses of identical twins find both environmental and genetic components of variability 16 . Recently, inbred mouse strains were placed in a more natural, 'dirty' environment leading to greater variation of immune cell populations 17 . Here, we focus on how introducing genetic diversity into mouse research reveals a much broader spectrum of innate immune responses.
Macrophages are widely distributed throughout the body and are hence one of the first cells to react to a perturbation of homoeostasis. Their functional programs are highly contextdependent and mostly influenced by pathogens, cellular origin (monocyte-derived or embryonic) and tissue cues (including cytokines) 18,19 . Among a continuum of stimulus-dependent polarization states 18,[20][21][22][23] , most with unknown function, M1 reflects a pro-inflammatory phenotype with pathogen killing abilities and can be induced by lipopolysaccharide (LPS), and M2 reflects the default state of tissue macrophages that promotes homeostasis and wound healing. These macrophages metabolize arginine to citrulline and nitric oxide through inducible nitric oxide synthase (iNOS, Nos2). The production of the macrophage cytokine Interleukin-12 (IL-12) is a hallmark of M1 and promotes a Th1 response 24,25 . M2 macrophages can metabolize arginine to ornithine, a precursor of polyamine and hydroxyproline, and urea through Arginase (Arg1), which promotes wound healing, angiogenesis and tissue homoeostasis 24,25 . iNOS is a crucial regulator in mouse M1 macrophages, but its relevance in human macrophages is still debated 26,27 . Importantly, genetic variation has been shown to affect macrophage activation by hierarchical functions of lineage-determining transcription factors 28 . While significant differences in Th1/2 polarization propensity in healthy individuals has been shown 29 , little is known about natural diversity of macrophage polarization in health and disease.
The hybrid mouse diversity panel (HMDP) is a panel of about 100 inbred mouse strains developed for performing association studies with adequate statistical power and resolution for mapping of complex traits 30,31 . It has been successfully used for investigating gene-environment interaction in activated macrophages 32 , insulin sensitivity 33 , and susceptibility to atherosclerosis 34 . Here, we employ the HMDP as a surrogate model for human immune diversity and investigate how meaningful gene signatures of macrophage activation can be extracted from a heterogeneous population. We demonstrate that these signatures are highly robust in predicting disease susceptibility and outcomes in humans, suggesting that immune diversity is a critical parameter in translational medicine.

Results
Natural variation of LPS activation in macrophages. LPS is found in gram-negative bacterial membranes and elicits strong inflammatory immune responses, mainly via Toll-like receptor (TLR) 4 that induces activation of the NF-kB pathway 35 . To estimate the inter-individual variation of macrophage LPS responses in a diverse population we compared the transcriptional activation patterns in humans and mice. In human alveolar macrophages exposed to LPS in vivo we observed large intrinsic variation in inter-individual genetic responses, for example, NF-kB and TLR pathway activities ( Fig. 1a,b). In thioglycollate-elicited peritoneal macrophages of HMDP strains exposed to LPS, we found a similarly large range of pathway activations (Fig. 1c,d). Notably, LPS incubation was short (4 h), and reflects an early state of macrophage activation. Variation of gene expression in technical and biological replicates was low and did not explain the observed findings ( Supplementary Fig. 1, and Orozco et al. 32 ). Moreover, RNA deconvolution indicates that the peritoneal cavity was repopulated with inflammatory macrophages after thioglycollate treatment in all strains ( Supplementary Fig. 2).
We analysed the gene expression levels of key polarization genes in human alveolar macrophages and found considerable variation among healthy volunteers under homoeostatic conditions (Fig. 1e, left column; an additional data set of 70 humans is shown in Supplementary Fig. 3). After LPS treatment, iNOS, arginase and IL-12b were variably upregulated among individuals. Strikingly, IL-12b showed a wide range of upregulation between 1.4-fold and 64-fold (Fig. 1e, bottom right). Similarly, using 26 classical inbred strains of the HMDP, we found a comparable diversity in peritoneal macrophage transcriptome data. At baseline, arginase expression showed a large variation across all strains, whereas iNOS and IL-12b were expressed at low levels ( Fig. 1f, left column). After LPS treatment, the range of IL-12b upregulation between strains was large (0.4-fold to 64-fold). Other key macrophage genes such as MRC1 (CD206), MGL1, MGL2, FIZZ1 and IL-10 as well the housekeeping gene ACTB (b-actin) did not change in response to LPS ( Supplementary Fig. 4). IL-6 was variably upregulated across the HMDP, and highly correlated with IL-12b ( Supplementary  Fig. 4). We confirmed expression differences between strains at the protein level showing great variation in IL-12 p70 levels in supernatants of LPS-stimulated peritoneal macrophages and varying iNOS and arginase induction patterns using intracellular flow cytometry ( Supplementary Fig. 5). For the protein level data, we developed a representative panel of 13 strains that largely resemble gene expression profiles of the entire HDMP with regard to Arg1, iNOS and IL-12b transcription ( Supplementary  Fig. 5). Altogether, these data suggest that inter-individual differences in macrophage activation responses can be found both in human populations and strains of the HMDP.
Global map of LPS responsiveness in 83 mouse strains. Most bioinformatics tools are not designed to dissect a spectral distribution in a heterogeneous population. Analysis of commonly regulated genes after LPS treatment among all 83 strains fails ( Supplementary Fig. 6), suggesting inter-strain variations of the macrophage LPS response. To overcome this problem, we established a gene expression based factor that represents the degree of LPS-induced polarization (polarization factor). Historically, the LPS-induced polarization was defined by the arginine metabolism: M2 is Arg1 high iNOS low ; M1 is Arg1 low iNOS high36 . We calculated polarization factors based on both iNOS and IL-12b. By dividing iNOS or IL-12b by arginase-1 gene expression values averaged to the mean of the HMDP (for details see Method section), 83 mouse strains were ranked to show their LPS responsiveness. We found that there is a continuous spectrum of LPS-induced activation strength; thus, identifying inherent LPS non/low-and -high responders (for example, FVB/NJ, CE/J and KK/HIJ, PL/J, respectively) (Supplementary Data 1). IL-12b-and iNOS-based rankings of mouse strains are highly correlated (Supplementary Fig. 7). In agreement with this ranking, analysis of amino acid levels in supernatant before and after LPS stimulation showed LPS high-responder strains producing less ornithine and more citrulline ( Supplementary  Fig. 8). Since the IL-12b based polarization factor has a higher resolution due to a more robust upregulation (Fig. 2c,d; Supplementary Fig. 7) and iNOS relevance is controversial in human macrophages 26,27 , the IL-12b based polarization factor was used in subsequent analyses.
Gene signatures correlated with LPS responsiveness. Next, genes of the peritoneal macrophage transcriptomes of all HMDP strains were correlated with the polarization factor, resulting in a ranked list of 1,276 LPS-positive responder genes and 2,619 LPS negative-responder genes at a false discovery rate (FDR) o5% (Benjamini Hochberg) and Po0.0001 (Pearson) (Fig. 3a  Activation states of murine tissue-resident macrophages. The diversity of gene expression profiles of murine tissue-resident macrophages has been reported previously 19,40 . We extended this data by determining the degree of activation at steady state using gene set enrichment analysis (GSEA) 41 . As expected, all tissueresident macrophages are not predominantly M(LPS) þ enriched.
However, they show differences in the strength of M(LPS) À gene expression (Fig. 4a). Microglia have the strongest M(LPS) À enrichment, whereas lung-resident macrophages also express     (Fig. 4a,b). This likely reflects the natural exposure of lung-resident macrophages but not microglia to organisms of the commensal microbiota.
Prediction of inflammatory disease in humans. To apply our mouse findings to humans, we first tested whether the M(LPS) þ and M(LPS) À gene lists sufficiently mapped to the human transcriptome. Human alveolar macrophages expressed about 70% of the mouse-derived gene lists (135 of the top 200 M(LPS) þ genes; 142 of the top 200 M(LPS) À genes). As expected, resting human alveolar macrophages showed no significant M(LPS) þ gene enrichment under baseline conditions and but a strong shift to M(LPS) þ after LPS treatment ( Fig. 4c; Supplementary Fig. 10). Similar enrichment was observed using human CD14 þ monocyte-derived macrophages with Listeria monocytogenes infection or mouse microglia with LPS challenge ( Supplementary  Fig. 10).
As polarized macrophages are known to critically shape pathophysiology of many inflammatory diseases 18 , we tested whether LPS responsiveness indicated by these gene signatures correlates with disease in humans. Peripheral blood leukocytes of systemic inflammatory response syndrome (SIRS) and sepsis patients both at day 1 and 3 after clinical diagnosis demonstrate gradually increasing M(LPS) þ enrichment scores demonstrating sensitivity to infection severity (Fig. 4d). Isolated healthy human synovial macrophages were enriched in M(LPS) À genes, whereas macrophages from rheumatoid arthritis patients were strongly skewed towards an M(LPS) þ phenotype (Fig. 4e). In lupus erythematodes, macrophages accumulate in the glomerula of the kidney and fuel disease progression 42 . Laser-dissected glomerula were M(LPS) þ enriched in lupus nephritis, but not in healthy conditions (Fig. 4f). Importantly, in these data sets each individual patient showed varying degrees of gene enrichment. These findings suggest that the HMDP-derived gene signatures are applicable across species and in many tissues.
Inter-individual activation state predicts tumour survival. Tumour-associated macrophages (TAM) are of key importance in the tumour microenvironment that is known to reinforce M2  VCAM1  RND3  GADD45B  EXT1  JARID2  PPM1B  CASP7  ACTN1  CALCRL  RBBP8  CST7  BCL2L11  IL15  PDCD10  UBE2D3  KLF6  EIF2AK2  TAF7  CDKN1A  CD44  CFLAR  PFKP  MTHFD2  AQP9  CCND2  CCND2  PDCD10  ZC3H15  IL12A  ITGAL  CFLAR  MYO10  IL18  ITGA4  ADAR  CASP7  CRLF3  IL15RA  CALCRL  MYD88  TLR7  CST7  SOD2  TRIM21  ACTN1  AQP9  CD4A  EIF2AK2  FAM129A  TNFRSF1B a b e f c d and suppress M1 polarization 24,25,43 . Specific pharmaceutical modulation of polarization states has emerged as a new anti-cancer therapy 44,45 . The monocyte/macrophage content in solid tumours can exceed more than 50% of all leukocytes (Fig. 5a, Supplementary Fig. 11). Using RNA deconvolution, the M(LPS) þ / À phenotypes are readily detectable in macrophage transcriptomes under controlled conditions ( Supplementary  Fig. 11). In the tumour microenvironment using bulk tumour biopsies, we demonstrate that most macrophages show enriched expression of M(LPS) À genes, which substantially varies between patients ( Fig. 5b; Supplementary Fig. 11). To address the question whether M(LPS) þ signatures can predict survival, we employed the PRECOG database that ranks genes by overall tumour survival 46 . In collapsed data from 18,000 biopsies across 39 tumours, we find that the M(LPS) þ signature correlates with survival, whereas the M(LPS) À signature correlates with cancer death (Fig. 5c). This pattern was found in many cancers of different ontologic origin, for example, osteosarcoma, melanoma, chronic lymphocytic leukaemia, Burkitt lymphoma and large-cell lung carcinoma (Supplementary Fig. 12). Patients with high expression of a disease-specific set of M(LPS) þ or M(LPS) À genes show increased or decreased survival in multiple tumour entities, respectively (Fig. 5d-f). A comparison with several published M1 macrophage gene lists ( Supplementary Fig. 9) indicates that survival prediction by M(LPS) þ / À signatures is more robust in all data sets (Supplementary Fig. 13).  XRN1  RBBP8  CRLF3  TRA2A  PARP12  BATF  SDC4  SAV1  TLR7   DDX24  MX1  ADORA2B  IFI35  STX6  ST3GAL3  SNX10  RCSD1  AQP9  CST7  GHITM  SNX9  RNF11  CCL22   VDAC1  MKNK2  AFG3L2  ARL6IP4  NAPG   OGFOD2  GTF2I  GDI2  DHCR7  PRUNE  FAM63A  GMNN  UCHL5  PRKRA  PLA2G6  TPI1  DNMT1  Here, we showed how a spectrum of macrophage phenotypes in many different inbred strains can be used to extract gene signatures of LPS responsiveness. Since the analysis of commonly LPS regulated genes in all strains failed, we established a surrogate marker based on IL12b and Arg1 gene expression. Compared to published signatures, the resulting gene lists are unique, yet more robustly predictive of many human inflammatory and malignant disorders, suggesting that accounting for immune diversity in a heterogeneous population increases the translatability of mouse data. The robust correlation of a large number of genes with the polarization factor across the HMDP reveals that the M(LPS) axis is a major player in macrophage biology. Out of 12,980 genes expressed, 1,276 (9.8%) and 2,619 (20.2%) are positively and negatively correlated with the IL-12b/Arg1 polarization factor, respectively. This is not surprising per se. However, it is surprising that only three M(LPS) þ genes (Dpep1, Gkn2 and Hoxd4) are found in resting (thioglycollate-elicited) peritoneal macrophages, and only 11 M(LPS) À genes (Arg1, Tpi1, Gpi1, Mif, Pkm, Zmynd8, Pcbp4, Myo6, Grk6, Egln1, Slc44a2) suggesting that the LPS challenge exposes the activation propensity. Thus, it is more accurate to speak of individuals with an M1-skewed response to TLR challenge, rather than a priori M1-skewed individuals.
Inter-individual variability of the LPS response in healthy human PBMCs has been linked to profoundly different cytokine responses 47 . We speculate that mice and human macrophages evolved to either fight (LPS positive responders) or tolerate and heal (LPS negative responders) infections 48 . TLR polymorphisms 49 and differences in TLR signalling pathways are candidates for a genetic basis underlying M(LPS) þ and M(LPS) À macrophage phenotypes. Of course, this is an oversimplification, because it is well known that macrophage polarizations also vary along the time axis of the infection or disturbance 18 . For example, myocardial infarctions are characterized by an initial abundance of M1 macrophages, followed by an M2-dominate healing phase 50,51 . Our data could implicate that M(LPS) þ individuals may be susceptible to autoimmune diseases and M(LPS) À individuals may have unfavourable outcomes if afflicted with cancer. Similarly, individual differences in T helper cell differentiation in healthy individuals has been linked to disease susceptibility 29,52 . Multiple susceptibility genes and loci to infectious disease 12 , autoimmunity 53,54 and cancer 55,56 have been described and our data suggest that inter-individual variation of macrophage activation propensity might be a confounding variable involved.
While the human immune system is shaped by both environmental factors and genetics 16 , only the latter plays a role in laboratory inbred mice. It remains to be investigated what exact mechanisms account for the inter-strain differences. The HMDP consists of 29 classic parental inbred strains, and about 80 recombinant inbred strains (BXD, CXB, BXA/AXB and BXH panels) that are crosses between C57BL/6J, C3H/HeJ and DBA/2J 31 . The TLR4-insensitivity of the C3H/HeJ strain is therefore represented to variable degrees in its recombinant progeny. Inter-strain differences in the Interleukin-6/Interleukin-10/STAT3 pathway may also affect macrophage activation. After thioglycollate, inflammatory monocyte-derived macrophages dominate the cellular content of the peritoneal cavity in all mouse strains (Supplementary Fig. 2). Future studies on peritoneal and other tissue-resident macrophages in response to various stimuli will more fully characterize macrophage immune diversity in mice and humans. The polarization factor tool developed here will be useful for such studies.
A striking finding of this study is that the M(LPS) þ / À signatures are robust enough to predict cancer survival or death from mixed-cell biopsy material, containing cancer, stromal and inflammatory cells. This makes these gene panels well suited for predictive tests in personalized medicine. Particularly, with new generation cancer treatments that manipulate tumour-associated macrophage polarization, new diagnostics are necessary to monitor treatment efficacy 43,44 . A simple multiplex PCR or RNA-sequencing run could harbour enormous predictive value, matching or exceeding the value of traditional biomarkers, for example, prostate specific antigen or BRCA1/2, or histopathological cancer staging. Of note, in contrast to an increasing number of disease-specific genetic tests that were developed by gene-outcome association statistics, for example, in breast cancer 57 , we extracted genes in a 'bottom-up' approach centring around macrophage biology. Given the ubiquitous diseaserelevant role of macrophages 18 , this yielded transcriptomic signatures with predictive value in a number of different disease entities. Furthermore, in contrast to conventional flow cytometryor PCR-based estimations of macrophage polarization in ordinal scale, the developed gene panels allow a gradual classification, thus enabling a population-based assessment in high resolution necessary for clinical applications.
Picking as few as 13 mouse strains is sufficient to qualitatively represent the diversity in macrophage activation responses to LPS. Re-evaluating the robustness of a new biological finding in a selected, representative panel of HMDP strains may be cost effective and prudent before embarking on a clinical drug development programme. Of note, all HMDP strains are fully inbred (homozygous at all loci) and commercially available 30,31 , providing immediate access for research facilities and allowing reproducibility studies. Therefore, this approach shows great promise to improve the translation of immunological research findings from mice to humans.
In conclusion, our population-based approach yields (1) improved gene signatures with higher predictive power; (2) a methodological framework to extract meaningful signatures of data sets from heterogeneous populations, and demonstrates (3) the usefulness of the HMDP as a valuable surrogate for human diversity in translational research.

Methods
Mice. Male mice 6-10 weeks of age were obtained from Jackson Laboratories (Bar Harbor, ME, USA) and housed in pathogen-free conditions on chow diet (Ralston-Purina Co, St. Louis, MO, USA). Details on the HMDP have been reported before 30 and data is accessible online (http://systems.genetics.ucla.edu/ data). All mouse strains used in this study are listed in Supplementary Data 1. Experimental procedures were approved by the Institutional Care and Use Committee (IACUC) at the University of California, Los Angeles.
Macrophage culture and activation. Murine peritoneal macrophages were elicited with thioglycollate (BD, Sparks, MD, USA; same batch for all strains) for 4 days. Cells from up to 4 mice were pooled and plated at 4 Â 10 6 cells ml À 1 in DMEM þ 20% FBS þ 1% streptomycin/penicillin in duplicates or triplicates. After overnight culture, adherent cells were selected (adherence assay for macrophage enrichment). RNA deconvolution shows two main populations of inflammatory macrophages (F4/80 þ , MHCII À and F4/80 À , MHCII þ ) and a small population of neutrophils ( Supplementary Fig. 2). Cells were incubated with 2 ng ml À 1 LPS (List Biological, Campbell, CA, USA) or control in DMEM þ 1% FBS for 4 h before harvest. The viability of cultured macrophages was determined for some strains 32 . Briefly, macrophages were stained with 2 mM calcein AM (Molecular Probes) and the absorbance was measured at 530 nm. Viability was 490% with no difference between LPS-treated and untreated conditions 32 . Multiple replicates of some strains allowed determination of experimental variability 32 .
Gene expression profiling. Total RNA was obtained using RNeasy columns (QIAGEN, Valencia, CA, USA) with DNA digest according to manufacturer's instructions and subsequently hybridized to Affymetrix HT MG-430A chip arrays. Chip signals were transformed to robust multichip average (RMA). Raw microarray data for peritoneal macrophages of the HMDP has been published before 32 and is deposited under the NCBI GEO accession number GSE38705.
Amino acid detection. For 13 selected mouse strains, peritoneal macrophages were cultured in the presence of 2 ng ml À 1 lipopolysaccharide in DMEM þ 1% FBS media for 0, 4 or 24 h. Supernatant was collected and stored at À 80°C. Supernatant was analysed for amino acid concentration using a Hitachi L-8900 Amino Acid Analyzer.
Polarization factor and correlated genes. The polarization factor ratio (PFR) describes the macrophage polarization state in a diverse population based on Affymetrix-RMA gene expression values. For each mouse strain, iNOS (Nos2) or IL-12b was divided by Arginase-1 (Arg1) and averaged to the population's baseline average (Equations 1 and 2). This was performed for both baseline and LPS-treated conditions. An increase or decrease indicates a shift to M(LPS) þ or M(LPS) À , respectively.
Equation (2) Gene signature analysis. The gene set enrichment analysis (GSEA) tool 41 allows to investigate whether a list of genes (for example, signature) is represented in differentially expressed genes of two given conditions (for example, control versus disease). We used the GSEA tool embedded in the GenePattern 2.0 framework 58 with standard settings (weighted, 100 iterations). For enrichment in a single data set, the pre-weighted GSEA algorithm was used. Statistical parameters as standard GSEA output of the data sets used in this manuscript are: GSEA (Fig. 4c) Ingenuity's pathway analyzer (IPA, Qiagen) was used to analyse pathway enrichment in M(LPS) signatures. Transcription factors were extracted from these signatures and the P value overlap and activation z score was calculated according to pathway overlap and gene activity (inhibition versus activation) using IPA.
For bulk RNA deconvolution into cellular subsets, the CIBERSORT algorithm was used 59 . Core signatures included either the provided LM22 signature of several leukocyte subsets (for example, naive B cells, memory B cells, plasma cells, naive CD4 T cells, CD4 memory cells, follicular helper T cells, gd T cells, NK cells, monocytes, macrophages, dendritic cells, mast cells, eosinophils, neutrophils) or the HMDP-derived gene lists (top 200 genes). In survival analyses, we performed gene set enrichment analysis on the PRECOG (PREdiction of Clinical Outcomes from Genomic Profiles) database that ranks genes by clinical survival either in a collapsed pan-cancer or in tumour-specific data set 46 . Top 10 or 30 leading edge genes of M(LPS) þ or M(LPS) À lists were subsequently used to define a tumourspecific gene set that was used to retrospectively predict survival in published data sets. Survival data was analysed using PROGgene2 (ref. 60) and plotted as median without sub-cohort division in a Kaplan-Meier format. Significance was calculated using the log rank test.
Statistics. Statistical analysis was performed using GraphPad Prism (GraphPad Software, San Diego). Affymetrix gene chip data was normalized using the robust multi-array average (RMA) method ( ¼ log 2 , background-corrected, quantile normalized). Correlation analyses were based on Pearson's correlation unless otherwise indicated. The threshold for significant M(LPS) þ or M(LPS) À genes was set as Po0.0001 (Pearson) and FDR o5% (Benjamini Hochberg) for the correlation between gene expression (RMA) and PFR.
Data availability. The microarray data of the HMDP mouse strains are available in a public repository from the NCBI website under the accession number GSE38705. Other pre-published data sets referenced in this study can be found under the accession numbers GSE21257 and GSE39055 (Osteosarcoma), GSE22762 (Chronic lymphocytic leukemia), GSE11969 (lung large-cell carcinoma), GSE4475 (Burkitt lymphoma), GSE15907 (Immgen project), GSM1151665 (BMDM þ LPS), GSM1150712 (BMDM þ M-CSF), GSE10500 (rheumatoid arthritis), GSE32591 (lupus nephritis). The SKCM-TCGA (Melanoma) data set is available in a public repository from the cancer-genome atlas website (https:// cancergenome.nih.gov). The authors declare that all the other data supporting the findings of this study are available within the article, its Supplementary Information files or from the corresponding author upon reasonable request.