A total of 49 archival, formalin-fixed, paraffin-embedded (FFPE) specimens, taken from Asian patients with advanced HCC who received ICB therapy between January 2015 and December 2018 at the Division of Medical Oncology, National Cancer Centre, Singapore, were evaluated. All samples were intrahepatic resection or biopsy samples and were obtained prior to immunotherapy at the Department of Anatomical Pathology, Division of Pathology, Singapore General Hospital. Clinicopathological parameters are summarized in table 1, and further details for individual patients are presented in online supplementary table 1.

We further categorized the patients into three groups based on their treatment regimen: patients treated with (1) PD-1 or PD-L1 inhibitors only (n=30), (2) combination PD1 or PD-L1 inhibitors with CTLA-4 inhibitors (n=13) and (3) combination of PD-1 or PD-L1 inhibitors with other treatments (n=6). We analyzed the data based on all patients from the aforementioned groups (groups 1–3, n=49), termed ‘patients treated with PD-1/PD-L1 ICB therapy’. We also analyzed the data based on a subgroup (groups A, n=30), termed ‘patients treated with anti-PD-1/PD-L1 single agent therapy’.

For all patients, tumors were staged and graded according to BCLC staging system28 and Edmondson-Steiner grading system.29 The Centralized Institutional Review Board of SingHealth provided ethical approval for the use of patient materials in this study (CIRB ref: 2014/590/B).

IHC was performed on the FFPE tissue samples as previously described.30–32 Tissue sections (4 µm thick) were labeled with antibodies targeting CD38 and PD-L1, as listed in online supplementary table 2. Appropriate positive and negative controls were included. To score the antibody-labeled sections, images were captured using an IntelliSite Ultra-Fast Scanner (Philips, Eindhoven, the Netherlands). The percentage of cells displaying unequivocal staining of any intensity for CD38 and PD-L1 were determined by two pathologists blinded to clinicopathological and survival information (JY and SS). PD-L1 tumor proportion score (TPS) was calculated as previously reported.7 Where discordant, the cases were reviewed and a consensus score was given.

mIHC/IF was performed using an Opal Multiplex fIHC kit (PerkinElmer, Inc, Waltham, Massachusetts, USA), as previously described by our group and in other studies.25 31–52 Tissue sections (4 µm thick) were labeled with primary antibodies against CD38, CD8 and CD68, followed by appropriate secondary antibodies. All antibodies used are listed in online supplementary table 2. This was followed by the application of a fluorophore-conjugated tyramide signal amplification buffer (PerkinElmer, Inc) and the nuclear counterstain DAPI. A Vectra three pathology imaging system microscope (PerkinElmer, Inc) was used to obtain images, and these were analyzed using inForm software (V.2.4.2; PerkinElmer, Inc)34 47 53 54 and HALO TM (Indica Labs, Albuquerqe, New Mexico, USA).55–59

The density of CD38⁺CD68⁺ macrophages and CD8⁺ T cells were determined as follows: cell count per predefined, high-powered field (334 μm × 250 μm) represents the density of CD38⁺CD68⁺ macrophages and CD8⁺ T cells in the TME. Samples were then categorized as ‘high’ or ‘low’ according to whether the CD38⁺CD68⁺ macrophage and CD8⁺ T cell count was above the cut-off points (best thresholds) that produced the lowest p value, which were determined using previously described methods.25 31 32 36 37 43–52 60 61

Gene expression data were obtained from The Cancer Genome Atlas (TCGA; https://www.cancer.gov/tcga). Gene expression and survival data were obtained from cBioPortal after filtering for HCC samples.62 HCC samples were divided into ‘high’ or ‘low’ CD38 expression groups using the optimal thresholds that produced the lowest p value. IPA core analysis was performed to identify differentially expressed genes (DEGs) between HCC samples with high and low CD38 expression, using the entire list of genes from the TCGA data as the background. P<0.05 was the threshold used to identify significant gene enrichment.

Single cell CD38 gene expression levels in the human liver were obtained from MacParland et al,63 using the author-provided web application hosted at http://shiny.baderlab.org/HumanLiverAtlas/. PDFs of CD38 expression levels, visualized using T-distributed stochastic neighbor embedding (t-SNE) dimension reduction, were generated using this web application. Cell populations identified by the authors are labeled on the plots.

Liver tissue was cut into fine pieces and digested with 0.5 mg/mL Collagenase IV (Gibco; Thermo Fisher Scientific, Inc, Waltham, Massachusetts, USA) and 0.05 mg/mL DNAse I (Sigma-Aldrich; Merck KGaA, Darmstadt, Germany) in complete RPMI 1640 (Gibco; Thermo Fisher Scientific, Inc) for 30 min at 37^°C. Digested tissue was filtered using a 70 µm cell strainer. Cells were then pelleted and treated with red blood cell lysis buffer (G-Biosciences, St Louis, Missouri, USA) for 5 min at room temperature. Cell debris was removed using Debris Removal Solution, according to the manufacturer’s protocol (Miltenyi Biotec, Ltd, Bergisch Gladbach, Germany).

Cells were incubated with Fixable Viability Dye eFluor 455UV (eBioscience, Thermo Fisher Scientific, Inc) for 30 min at 4^°C for live/dead cell discrimination. Fc receptors were blocked using Human TruStain FcX (BioLegend, San Diego, California, USA) for 5 min at room temperature. Cell surfaces were labeled with antibodies targeting markers of interest (refer to online supplementary table 3 for antibody panel) for 30 min at 4^°C. Matched isotype controls were included for antibodies against CD38. Single color compensation controls were prepared using Ultracomp eBeads (eBioscience, Thermo Fisher Scientific, Inc). Samples were read in a Spectral flow cytometer Cytek Aurora (Cytek Biosciences, Fremont, California, USA). Data analysis was performed using FlowJo V.10 software (FlowJo LLC, Ashland, Oregon, USA).

THP-1 cells were cultured and polarized based on a previously published protocol.27 64 In brief, cells were cultured in complete RPMI 1640 (Gibco, Thermo Fisher Scientific, Inc) supplemented with 10% fetal bovine serum (Gibco, Thermo Fisher Scientific, Inc), 1% penicillin/streptomycin (Gibco, Thermo Fisher Scientific, Inc), and 50 µM 2-mercaptoethanol (Gibco, Thermo Fisher Scientific, Inc). Cells were then plated in 24-well plates at a density of 5×10⁵ cells/mL, differentiated with 50 ng phorbol myristate acetate (Sigma-Aldrich, Merck KGaA) for 24 hours, then washed and cultured in fresh medium for 48 hours. Macrophages were polarized to the M1 state with 20 ng/mL IFN-γ (R&D Systems, Inc, Minneapolis, Minnesota, USA) and 100 ng/mL lipopolysaccharide (Sigma-Aldrich, Merck KGaA), or to the M2 state with 20 ng/mL IL-4 (R&D Systems, Inc). After 24 hours, cells were harvested for flow cytometry analysis. Culture supernatant was collected for the analysis of cytokines using Luminex kits (Luminex Corporation, Austin, Texas, USA).

A total of 65 analytes from the Immune Monitoring ProcartPlex Panel (Thermo Fisher Scientific, Inc) were measured. Assays were performed according to the manufacturer's protocol. Plates were washed using a Tecan Hydrospeed Washer (Tecan Group AG, Männedorf, Switzerland) and read with a Flexmap 3D system (Luminex Corporation). Data were analyzed using Bio-Plex manager V.6.2 software (Bio-Rad Laboratories, Hercules, California, USA) with a 5-parameter curve-fitting algorithm applied for standard curve calculations. Logarithmically transformed, averaged Luminex concentrations were standardized (mean-centered and standard deviation-scaled) by analyte and depicted as a heat map for CD38hi and CD38lo macrophage samples as defined and gated with FLOWJO (Becton, Dickinson&Company, USA) as previously described.27 Euclidean distance-based complete linkage clustering was performed for both analytes and samples. Numerical computations and heat maps were then generated using the ComplexHeatmap package in R V.3.3.1.

RNA was extracted from 14 unlabeled FFPE sections (10 µm thick) using an RNeasy FFPE kit (Qiagen GmbH, Hilden, Germany) on a QIAcube automated sample preparation system (Qiagen GmbH), and was quantified using an Agilent 2100 Bioanalyzer system (Agilent Technologies, Santa Clara, California, USA). A total of 100 ng of functional RNA (> 300 nucleotides) was assayed on the nCounter MAX Analysis System (NanoString Technologies, Inc, Seattle, Washington, USA). The NanoString counts were normalized using positive control probes and the housekeeping genes, as previously reported.31 47 The count data were then logarithmically transformed prior to further analysis. P<0.05 was considered to indicate a statistically significant difference.

Patient follow-up data were obtained from medical records, and overall survival (OS) and progression-free survival (PFS) was calculated using time periods defined as from the start of treatment to either death, progression, or date of of last follow-up, respectively. We used an unpaired Mann-Whitney U test to investigate whether biomarker expression in patients, as determined by mIHC/IF, predicted their reponse to immunotherapy. The associations between clinicopathological parameters and biomarker expression were analyzed using χ² and Fisher’s exact tests. Cox regression was performed to evaluate the effects of these factors on PFS and OS. Statistical analysis was performed using RStudio 1.1.456 running R V.3.5.053, V.3.5.054 (R-core Team, R Foundation for Statistical Computing, Vienna, Austria) and GraphPad Prism V.8.0.0 for Windows (GraphPad Software, Inc, San Diego, California, USA). P<0.05 was considered to indicate a statistically significant difference.

We initially analyzed the gene expression data of a cohort of patients with HCC (n=321), taken from a publicly available database (TCGA),62 and found that increased CD38 gene expression levels were associated with an improved disease-free survival of 38 months compared with 20 months, as presented in figure 1A (p=0.005).

IPA was then used to decipher the biological functions of the DEGs. As shown in figure 1B and online supplementary table 4, IPA core analysis identified significant functional enrichment in the expression of genes associated with the following canonical pathways: ‘Th1 and Th2 activation pathway’ (p<0.001), ‘Communication between innate and adaptive immune cells’ (p<0.001) and ‘iCOS-iCOSL Signaling in T Helper Cells’ (p<0.001). Furthermore, IPA upstream regulator analysis also revealed that the TNF, IL-10, IL-2 and IL-4 genes were enriched (p<0.001). This can be attributed to the Th1 and Th2 activation pathways, as well as communication between innate and adaptive immunity. Other canonical pathways, ‘Granulocyte Adhesion and Diapedesis’, ‘Th1 Pathway’ and ‘Th2 Pathway’ were also found to be enriched (p<0.001), warranting further investigation in future studies.

Notably, as presented in the volcano plot (figure 1C and online supplementary figure 1), the CD3 and CD8 genes, along with PD-L1 (CD274), PD-L2 (PDCD1LG2), IFNG, GZMK, IL2R and TIGIT, were enriched. These genes reflect the level of T cell infiltration. Genes that indicated myeloid cell accumulation, such as CD68, CD163, SINGLEC5, HLA-DRB1, HLA-DQA2, IDO, BATF3, CD1C, and CD1A, were also highly expressed in the high CD38 gene expression group.

We then sought to verify the expression of CD38 by tumor cells and different immune infiltrates in our HCC samples (n=49). First, we visualized the HCC samples using conventional IHC, and found that while some HCC samples had a relatively low number of CD38⁺ cells (figure 2A), these cells were abundant in other samples (figure 2B). Next, we examined the samples that harbored relatively high numbers of CD38⁺ cells using mIF/mIHC, and found a high level of colocalization between CD38 and CD68 (figure 2C), indicating that CD38 is expressed by macrophages.

Notably, using a publicly available, single cell RNA sequencing human liver dataset,63 we also found that the CD38 gene is primarily expressed by macrophages and immune cells such as plasma and natural killer (NK) cells, compared with hepatocytes and stromal cells (online supplementary figure 2). This finding is in line with those from recent reports by our group.27 We then used flow cytometry to validate CD38 expression by immune cells in our HCC samples, as shown in figure 3. As expected, CD38 was expressed by monocyte, macrophage and dendritic cell populations, including CD14⁺ cells, CD16⁺ cells, CD11b⁺ cells, CD11c⁺ cells, CD68⁺ cells and CD123⁺ cells. These results are in line with those of previous publications,25 65 including our recent findings.27

Next, we investigated whether the presence of CD38⁺ cells on the FFPE tissues taken from our patients with advanced HCC treated with ICB were predictive of response. The clinicopathological parameters of the patient cohort are listed in table 1. Responses were determined according to RECIST V.1.1.66 Patients who achieved a best response of Complete Response (CR)/Partial Response (PR) according to RECIST V.1.1 were termed responders and patients who achieved a best response of Stable Disease (SD)/Progressive Disease (PD) according to RECIST 1.1 were termed non-responders.

First, we studied the expression of CD38 at the protein level using conventional IHC. Using traditional manual scoring methods, as described in the Materials and methods section, we found that responders had a significantly higher proportion of intratumoral total CD38⁺ cells compared with non-responders (p=0.019; figure 4A). The majority of the intratumoral CD38⁺ cells morphologically resembled macrophages or lymphocytes, but no convincing expression of CD38 by tumor cells was observed.

For patients who underwent ICB therapy (n=49), an optimal cut-off for total CD38⁺ cell proportion can be accurately defined using receiver-operating characteristic analysis (online supplementary figure 3A). The cut-off used was 5% positivity among total immune infiltrates, and this cut-off achieved 75.5% accuracy, 71.1% specificity and 90.9% sensitivity. The area under curve was 0.785. With this cut-off, patients with a high proportion of intratumoral CD38⁺ immune cells achieved an overall response rate (ORR) of 43.48%, compared with 3.85% for those with a low proportion as described in the following. Patients with a high total CD38⁺ cell proportion experienced a superior median PFS (mPFS) of 8.21 months compared with 1.64 months in patients with low total CD38⁺ cell proportion (p=0.0065; HR=0.383; figure 4B; table 2). Similarly, patients harboring a high proportion of total CD38⁺ cells experienced a significantly better median OS (mOS) of 19.06 months, compared with 9.59 months in patients with low CD38⁺ cell proportion (p=0.0295; HR=0.463; figure 4C; table 3). Importantly, no other clinicopathological parameters predicted for PFS or OS.

Importantly, similar survival advantage was observed in patients with high intratumoral total CD38⁺ cell proportion who received anti-PD-1/PD-L1 single agent treatment in this cohort (n=30/49; PFS: p=0.0253, HR=0.397; OS: p=0.0483, HR=0.418; ORR=38.46% (high), 5.88% (low); online supplementary tables 5 and 6, figure 4A,B).

Our group recently reported that a large proportion of the CD38⁺ cells present in HCC tissues are CD38⁺CD68⁺ macrophages (62.9%±19.2%).27 Thus, we investigated whether quantification of CD38⁺CD68⁺ macrophages in patients with HCC could be a useful predictive biomarker of response to ICB (figure 4). Using mIHC/IF, we found that a higher density of CD38⁺CD68⁺ macrophages within the tumor was significantly associated with response to ICB therapy (n=49; p=0.021; figure 4D). Notably, we found that a high CD38⁺CD68⁺ macrophage density was associated with better mPFS, 3.88 months compared with 1.61 months in patients with low CD38⁺CD68⁺ macrophage density (p=0.0072; HR=0.368; figure 4E, table 2). Most strikingly, patients harboring a high density of CD38⁺CD68⁺ macrophages experienced a significantly greater mOS of 34.43 months compared with 9.59 months in patients with low CD38⁺CD68⁺ macrophage density (p=0.0336; HR=0.416; figure 4F; table 3). However, both CD38⁻CD68⁺ macrophage density and CD38⁺CD68⁻ cells density do not show any association with PFS (table 3; figure 4G,I) and OS (table 4; figure 4H,J). Furthermore, for patients who underwent ICB therapy (n=49), an optimal cut-off for CD38⁺CD68⁺ macrophages density can be accurately defined using receiver operating characteristic analysis (online supplementary figure 3B). The cut-off used was 1 cell, which achieved 71.4% accuracy.

In addition, in the subgroup that received anti-PD-1/PD-L1 single agent treatment in this cohort (n=30/49), the association of CD38⁺CD68⁺ macrophage density with survival is also observed (PFS: p=0.0325, HR=0.381; OS: p=0.0422, HR=0.374; online supplementary tables 5 and 6, online supplementary figure 4C,D). Whereas the association is not seen for both CD38⁻CD68⁺ macrophage density and CD38⁺CD68⁻ cells density (online supplementary tables 5 and 6, online supplementary figure 4E–H), suggesting the non-macrophages CD38⁺ immune cells (such as plasma and NK cells) as well as the CD38⁻ macrophages might play negligible predictive and prognostic role in this cohort treated with ICB or anti-PD-1/PD-L1 single agent.

The association between the BMS 4-gene inflammatory signature and melanoma or gastroesophageal cancer treated with immunotherapy has been investigated in previous literature.67–69 Thus, to investigate this in HCC, we compared the use of CD38 as a predictive marker to two of the genes included in this signature. We assessed PD-L1 TPS with IHC (figure 5A) and CD8⁺ T cell density with mIHC/IF (figure 5B), and found that the ORR of the high CD38 expression group was the highest, at 43.48%. Furthermore, low CD38⁺CD68⁺ macrophage density group conferred the lowest response rate, 0.07%. This underlines the utility of CD38 as a predictive biomarker (figure 5C).Neither PD-L1 TPS nor CD8⁺ T cell density alone predicted PFS or OS in this cohort. However, as presented in table 4, the addition of PD-L1 TPS and CD8⁺ T cell density to total CD38⁺ cell proportion significantly increased the predictive value of this biomarker for PFS (∆LRχ²=7.86; p=0.0197), compared with total CD38⁺ cell proportion alone. Meanwhile, the inclusion of PD-L1 TPS and CD8⁺ T cell density to CD38⁺CD68⁺ macrophage density increased the predictive value for PFS (∆LRχ²=8.82; p=0.0121) compared with CD38⁺CD68⁺ macrophage density alone. However, there was no statistical significance observed in terms of OS when PD-L1 TPS and CD8⁺ T cell density was added to either total CD38⁺ cell proportion or CD38⁺CD68⁺ macrophage density.

Our recent publication concerning CD38^hi macrophages revealed the differential expression of CD80 and DC-SIGN compared with CD38^lo macrophages, as well as the production of higher levels of proinflammatory cytokines, such as IL-6 and TNF-α.27 In the present study, we performed a Luminex experiment to simultaneously detect 65 cytokines in supernatant collected from both CD38hi and CD38^lo macrophages. We not only validated that CD38^hi macrophages produce more IL-6 and TNF-α than CD38^lo macrophages but also found that these cells also produced more IFN-γ and related cytokines, including CXCL9, CXCL10 and CXCL11 (figure 6). Similarly, utilizing NanoString technology, we also demonstrated that the IFN-γ gene level trended higher in the patients who harbored high CD38⁺CD68⁺ macrophages (n=7), compared with the patients who harbored low CD38⁺CD68⁺ macrophages (n=7, online supplementary figure 5). However, the statistical significance was not reached.