A Large-Scale Meta-Analysis Reveals Positive Feedback between Macrophages and T Cells That Sensitizes Tumors to Immunotherapy

Identification of robust immunogenomic connections, particularly macrophage T-cell interactions, in a large-scale pan-cancer meta-analysis and development of a predictive model for immunotherapy response that outperformed existing models could facilitate clinical decision-making.

Abstract

Although considerable efforts have been dedicated to identifying predictive signatures for immune checkpoint inhibitor (ICI) treatment response, current biomarkers suffer from poor generalizability and reproducibility across different studies and cancer types. The integration of large-scale multiomics studies holds great promise for discovering robust biomarkers and shedding light on the mechanisms of immune resistance. In this study, we conducted the most extensive meta-analysis involving 3,037 ICI-treated patients with genetic and/or transcriptomics profiles across 14 types of solid tumor. The comprehensive analysis uncovered both known and novel reliable signatures associated with ICI treatment outcomes. The signatures included tumor mutational burden (TMB), IFNG and PDCD1 expression, and notably, interactions between macrophages and T cells driving their activation and recruitment. Independent data from single-cell RNA sequencing and dynamic transcriptomic profiles during the ICI treatment provided further evidence that enhanced cross-talk between macrophages and T cells contributes to ICI response. A multivariable model based on eight nonredundant signatures significantly outperformed existing models in five independent validation datasets representing various cancer types. Collectively, this study discovered biomarkers predicting ICI response that highlight the contribution of immune cell networks to immunotherapy efficacy and could help guide patient treatment.

Significance:

Identification of robust immunogenomic connections, particularly macrophage T-cell interactions, in a large-scale pan-cancer meta-analysis and development of a predictive model for immunotherapy response that outperformed existing models could facilitate clinical decision-making.

Graphical Abstract

Introduction

Immune checkpoint inhibitors (ICI) that target PD-1, PD-L1 or CTLA4, provide durable response in only a subset of patients with advanced-stage cancers. Therefore, it is crucial to discover predictive biomarkers of ICI response. With great effort, a number of signatures associated with ICI outcome have been identified, including the expression of target ligands such as PD-L1 (1), co-inhibitory receptors such as PD-1 (2–4) and LAG3 (5), tumor mutational burden (TMB; refs. 6, 7), tumor infiltration lymphocytes (8), T-cell repertoire (9), and secretion of IFNγ (10). Although these biomarkers showed remarkable clinical effects, the mechanism by which intercellular signalling pathways modulate the sensitivity or resistance to ICI therapy remain unclear (11). As an example, tumor infiltration lymphocyte is one of the most effective biomarkers, but its trafficking and how it contributes to ICI response is largely unknown (12). Moreover, biomarkers identified in one study generally had poor performance or even showed opposite association in other studies. For example, mutations in DNA damage pathway have been identified to be related to ICI response (13). A recent meta-analysis of >1,000 ICI-treated patients across seven tumor types, however, cannot reproduce the association (14). The poor generalizability and reproducibility of current biomarkers are mainly due to tumor heterogeneity and limited sample sizes of those discovery cohorts. Large-scale studies with sufficient number of patients are essential for addressing tumor heterogeneity and thus identifying robust biomarkers.

A growing number of multiomics studies have generated vast amount of data across various tumor types, presenting a significant opportunity for biomarker discovery (15). We have collected all the publicly available genomic data under ICI treatment and developed an immuno-genomic portal to facilitate the translation of the rich data source into biological insights and clinical applications (16).

Here, we performed the largest meta-analysis involving 3,037 patients with genetic and/or transcriptomics profiles across 27 studies and 14 types of solid tumor, who received therapy of anti-PD-1/L1, anti-CTLA4, or both (Table 1). We investigated the association of ICI outcomes with an extensive number of multidimensional and multi-omics features, including TMB, gene and pathway mutations, mutational signatures, gene and pathway expression, pairwise transcriptomic relationships involving immune cells activation and migration, and immune cell compositions (Fig. 1). On the basis of a random-effect meta-analysis model, we uncovered both well-known and novel signatures significantly associated with ICI response. These signatures include TMB and PD-1 expression, expression of IL21, GTSF1L, and ZEBD2, and notably, genes and pathways related to macrophages and T cells activation and migration, such as proportion of M1 macrophage, cytotoxic T lymphocytes (CTL), and CXCR3 binding pathways. We further verified these associations in three validation cohorts compiled from five independent datasets representing various cancer types (Fig. 1). Single-cell RNA profiling and dynamic transcriptomic alterations during ICI treatments provides further evidence, supporting the idea that positive feedbacks between macrophages and T cells play a significant role in immunotherapy response. A multivariable model, constructed based on eight nonredundant signatures, demonstrated a substantial improvement in performance compared with predictions relying solely on TMB (an FDA-approved biomarker) or EaSleR (17) in the three validation cohorts. Consistently, pairwise transcriptomic relationships associated with immune cells activation and migration is the most critical and relevant feature in the predictive model.

Table 1.

Summary of datasets in the study.

	Cancer type	Data source	No. of genetic/transcriptomic/single-cell data	No. of samples receiving anti-CTLA4/anti-PD-(L)1/both	No. of pre-/on-treatment samples	Reference
Discovery cohort	Melanoma	Auslander et al.	−/37/−	6/27/4	14/23	(18)
	Melanoma	Chen et al.	−/104/−	43/61/−	57/57	(19)
	Melanoma	Hugo et al.	38/28/−	−/40/−	36/4	(20)
	Melanoma	Prat et al.	−/25/−	−/25/−	25/−	(21)
	Melanoma	Samstein et al.	320/−/−	75/130/115	−/320	(7)
	Melanoma	Nathanson et al.	34/−/−	34/−/−	34/−	(22)
	Melanoma	Van Allen et al.	110/42/−	112/−/−	112/−	(23)
	Melanoma	Jerby-Arnon-ValCo2 et al.	−/112/−	−/112/−	112/−	(25)
	Melanoma	Gide et al.	−/91/−	−/50/41	73/18	(24)
	Non-Small Cell Lung Cancer	Hira Rizvi et al.	240/−/−	−/206/34	204/36	(26)
	Non-Small Cell Lung Cancer	Naiyer Rizvi et al.	35/−/−	−/34/−	34/1	(27)
	Non-Small Cell Lung Cancer	Samstein et al.	350/−/−	−/329/21	−/350	(7)
	Metastatic Urothelial Cancer	Mariathasan et al.	−/348/−	−/348/−	348/−	(31)
	Renal Cell Carcinoma	Samstein et al.	151/−/−	−/122/29	−/151	(7)
	Renal Cell Carcinoma	Miao et al.	35/33/−	−/93/5	98/−	(28)
	Bladder Cancer	Samstein et al.	215/−/−	−/192/23	−/215	(7)
	Glioma	Samstein et al.	117/−/−	−/114/3	−/117	(7)
	Glioma	Cloughesy et al.	−/29/−	−/29/−	29/−	(29)
	Colorectal Cancer	Samstein et al.	110/−/−	1/99/10	−/110	(7)
	Colorectal Cancer	Le et al.	31/−/−	−/31/−	31/−	(30)
	Head and Neck Cancer	Samstein et al.	139/−/−	−/131/8	−/139	(7)
	Esophagogastric Cancer	Samstein et al.	126/−/−	2/93/31	−/126	(7)
	Gastric Cancer	Kim et al.	−/61/−	−/61/−	61/−	(32)
	Breast Cancer	Samstein et al.	44/−/−	20/20/4	−/44	(7)
	Hepatocellular Carcinoma	Harding et al.	31/−/−	−/31/−	−/−	(33)
	Prostate Cancer	Kwek et al.	−/22/−	22/−/−	11/11	(32)
	Basal cell carcinoma	Yost et al.	15/−/−	−/15/−	7/8	(34)
Validation cohort	Melanoma	Snyder et al.	30/30/−	30/−/−	−/30	(35)
	Melanoma	Riaz et al.	68/117/−	−/139/−	78/61	(36)
	Melanoma	Liu et al.	144/144/−	−/144/−	143/1	(37)
	Urothelial cancer	Damrauer et al.	218/218/−	218/218/−	Unknown/Unknown	(38)
	Pan-cancer	Pleasance et al.	−/73/−	1/53/19	Unknown/Unknown	(39)
	Renal Cell Carcinoma	Bi et al.	−/−/4	−/4/−	−/4	(40)
	Breast cancer	Zhang et al.	−/−/6	−/6/−	6/−	(41)

Figure 1.

Schematic overview for comprehensive identification of ICI-related biomarkers. A total number of 3,037 samples across 14 cancer types were collected from 27 studies in the discovery cohort. Eight types of genetic and transcriptomic features were defined and their associations with ICI response were evaluated. The signatures identified in the discovery cohort were further validated by three validation cohorts, which were constructed from five independent datasets encompassing various cancer types, along with two single-cell RNA-seq datasets. R, responder; NR, nonresponders.

Materials and Methods

Datasets

The discovery cohort consists of 27 studies across 14 tumor types, including nine from melanoma (7, 18–25), three from non–small cell lung cancer (7, 26, 27), two from renal cell carcinoma (7, 28), glioma (7, 29), and colorectal cancer (7, 30), one from metastatic urothelial cancer (31), bladder cancer (7), head and neck cancer (7), esophagogastric cancer (7), gastric cancer (32), breast cancer (7), hepatocellular carcinoma (33), prostate cancer (32), and basal cell carcinoma (34), respectively. We downloaded clinical outcomes containing treatment types (anti-CTLA4, anti-PD-1/L1, or combo treatment), biopsy time (pre- or on-treatment), RECIST including complete response (CR), partial response (PR), progression disease (PD) and stable disease (SD), overall status (living or deceased), overall survival, and progression-free survival time from original literature (Supplementary Table S1). Mutation and expression profiles were also downloaded from original literature or public databases. In the case when expression profiles were not initially provided, and only raw sequencing data in FASTQ files were available, we carried out a standard bioinformatics pipeline to generate expression profiles.

All processed data can be accessed at https://bioinfo.vanderbilt.edu/database/Cancer-Immu/ (16). The criteria for defining responders and nonresponders were obtained from original studies when provided. If not explicitly specified, patients achieving CR and PR were categorized as responders, whereas those with SD and PD were classified as nonresponders.

Five datasets representing various cancer types were used for validation, including three datasets from melanoma (35–37), one dataset from urothelial cancer (38), and one pan-cancer dataset encompassing 20 different cancer types (39). Within the three melanoma datasets, there are 144 patients who received anti-PD-1 treatment (37), 30 patients who received anti-CTLA4 treatment (35), and 139 samples that include both pretreatment and on-treatment data (36). The dataset from urothelial cancer comprises 218 samples (38), and the pan-cancer dataset contains 73 samples (39). We created three validation cohorts from the five datasets: the melanoma cohort, which comprises the three melanoma datasets; the urothelial cohort, consisting of the dataset from urothelial cancer; and the pan-cancer cohort, which encompasses all five datasets (Fig. 1).

Two single-cell datasets were employed for the investigation of cell–cell interactions. The first dataset comprises four samples from renal cell carcinoma under anti-PD-1 treatment (40), whereas the second dataset includes six breast cancer samples under the same treatment (41). Normalized counts and cell annotation for single-cell datasets were obtained from the Single Cell Portal (https://singlecell.broadinstitute.org/single_cell/study/SCP1288/tumor-and-immune-reprogramming-during-immunotherapy-in-advanced-renal-cell-carcinoma#study-download) and GSE169246 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE169246).

Genetic and transcriptomics features

To explore genetic and transcriptomics data, we defined several feature categories encompassing gene-level, pathway-level, immune cell-level, and global-level. These include TMB, gene and pathway mutations, mutational signatures, gene and pathway expression, pairwise transcriptomic relationships associated with immune cell activation and migration, and immune cell compositions.

At the genetic level, we evaluated TMB, gene and pathway mutations, and mutational signatures. TMB data were obtained from the original literature when available. In cases where it was not provided, we used the mutation count for each sample as the tumor mutational load. Gene mutations were obtained from the original literature or databases, considering only nonsynonymous mutations. Pathway mutations were determined using 32,283 human molecular gene sets sourced from the Molecular Signatures Database (MSigDB; ref. 42). Any mutation associated with a pathway was considered as a mutation for that specific pathway. For mutational signatures, we employed COSMIC mutational signature types (43) to characterize combinations of mutational types in samples with over 50 somatic mutations. Initially, somatic mutations within exome regions were isolated. Trinucleotide counts were then normalized based on the frequency of each trinucleotide context observed within the exome region for each mutational signature type. We utilized the ‘deconstructSigs’ (v1.6.0) package (44) for calculating mutational signatures.

At the transcriptomics level, we examined gene and pathway expression, pairwise relationships related to immune cell activation and migration, and immune cell compositions. Gene expression profiles were acquired from original literature or databases and subsequently normalized and log2 transformed when available. In instances where expression profiles were not provided, we applied a general pipeline for sequencing alignment and expression quantification. Briefly, the pipeline began with a quality control assessment using FastQC. All reprocessed samples successfully passed the quality control in the FastQC step (Supplementary Table S1). We then performed the alignment of sequencing reads to the hg19 genome using STAR (45) and quantified expression with FeatureCounts (46). For normalization, we employed the variance stabilizing transformation (vst) function of DESeq2 (47). Finally, we computed transcripts per kilobase million (TPM) values using RSEM with default parameters (48).

Pathway expression was determined by averaging the expression of genes within each respective pathway. This was based on the evaluation of 32,283 human molecular gene sets from MSigDB (42). When assessing pairwise transcriptomic relationships associated with immune cells activation and migration, we considered 45 immune checkpoint genes with known costimulatory or co-inhibitory effects extracted from Auslander and colleagues's study (18). We also included five ligands of CXCR3 (CXCL9, CXCL10, CXCL11, CXCL13, and PF4), in addition to CXCR3 itself, to gauge immune interactions and migration. In total, we established 1,275 pairwise relationships, with each gene pair defined based on their logical relation between expression levels. For example, if A > B, ‘A_B’ pair was defined as 1, otherwise ‘A_B’ was 0. The interaction score is defined as the mean score of ICI-related gene pairs. Immune cell compositions were derived from gene expression profiles using CIBERSORT (49, 50), which provided estimates of the relative percentage of each immune cell type in each sample. LM22, a gene matrix encompassing the expression of 547 genes for 22 distinct immune cell types, was employed as signatures for distinguishing immune cell populations and activation states.

Meta-analysis

In the meta-analysis, all genomic values, whether binary (e.g., gene mutations) or continuous (e.g., TMB and gene expression), were standardized to z-scores (14, 51) to ensure comparability across different measurement values. Effect size and standard error of the effect size were computed using binomial logistic regression, with z-scores as the terms and binary ICI response as the values. The meta-analysis results were subsequently estimated by aggregating individual results from each study using a random-effects model of meta-analysis.

Data aggregation analysis

As gene and pathway mutations are genetic features that can be compared across datasets, aggregated mutation profiles were created by combining mutation profiles from multiple datasets into a single unified profile. We assigned a value of 1 for mutated genes, and 0 for nonmutated ones. Then binomial logistic regression was used to estimate the association between gene mutations and ICI response. We added TMB as a covariate into the regression formula to adjust the associations between gene mutation and ICI response. In addition, the associations between gene mutations and TMB were detected by logistic regression using tumor TMB as terms and binomial gene mutation status as values.

Dysregulated ligand–receptor interactions from single-cell transcriptomic

scLR is a statistical method for the identification of dysregulated ligand–receptor interactions between conditions (52). In comparison with CellphoneDB (53) and Cellchat (54), which are primarily designed to identify ligand–receptor interactions within a single condition, scLR aligns more closely with our analytical requirements for discovering dysregulated interactions between responders and nonresponders (52). scLR first summed gene counts in each cell cluster in each sample to estimate gene expression in the cluster and in the sample. Then the summed data were normalized by a median normalization method in DESeq2 and then log2 transformed. The mean interaction strength of ligand and receptor between cell types c_i and c_j across all the samples under responders can be denoted as , where . Similarly, for nonresponders . Finally, was calculated as the cell-type-specific dysregulated interactions between responders and nonresponders. scLR was downloaded from https://github.com/cyhsuTN/scLR.

Predictive models

We employed machine learning technique, gradient boosted tree algorithm (XGBoost; ref. 55), to exploit the large-scale immune-genomic data. R package ‘xgboost’ was used for fitting a multivariable predictive model (MVP) of ICI response. All 19 features, including TMB, expression of IFNG, IRF1, LAG3, CXCL10, PDCD1, IL21, GTSF1L, CCL19, ZBED2, and CXCL11 genes, expression of PD1 blockade pathway, CTL pathway, CD8⁺ alpha beta T-cell activation, NO2IL12 pathway, T-cell receptor binding and CXCR3 receptor binding, M1 macrophage composition and interaction score of all 3,037 samples were inputted into the XGBoost. To mitigate batch effects and facilitate the comparability of biomarkers with different measurement scales (e.g., TMB and gene expression), all feature values were standardized to z-scores. Following this standardization process, we did not observe any batch effects across different studies. All parameters of XGBoost were kept as default values. TMBM was built with the same processing steps. R package ‘ROCR’ was used for the ROC curve analysis. DeLong test was utilized to compare the performance of different models.

From a clinical standpoint, we developed a simplified model using nonredundant features. We started by measuring the correlations between all biomarkers and then calculated their average correlations to other features. Features with average correlation greater than 0.6, including expression of CXCR3 receptor binding, CTL pathway, NO2IL12 pathway, CD8 T-cell activation, and T-cell receptor binding, and gene expression of IFNG, LAG3, IRF1, CXCL11, CXCL10, and PDCD1, showed high similarities due to sharing common genes. CTL pathway and CXCR3 receptor binding, showing the greatest correlations to other features, played completely independent roles without any overlapping genes. Finally, eight nonredundant features were kept, including TMB, expression of IL21, GTSF1L, CCL19, and ZBED2, pathway expression of CXCR3 receptor binding and CTL pathway, and the interaction score. The performance of the MVP based on the eight nonredundant features (denoted as MVP), were compared against the model relying solely on TMB and EaSleR (17), which predicts ICI response based on system-wide signatures of the tumor microenvironment, immune-cell composition, and intra- and intercellular communications, in the three validation cohorts.

Data availability

All processed data can be accessed at https://bioinfo.vanderbilt.edu/database/Cancer-Immu/. The datasets analyzed in this study and the access information are provided in Supplementary Table S1.

Results

Meta-analysis identifies robust ICI-related biomarkers

We collected 3,037 ICI-treated patients with genetic and/or transcriptomics data from 27 studies across 14 types of solid tumor (Fig. 2A). We utilized a random-effects meta-analysis model to identify consistent ICI-related biomarkers across different studies (Materials and Methods). We investigated genetic and transcriptomics features in a multidimensional view, ranging from individual genes, immune cell compositions, pathways, to pairwise expression relationships involving immune cells activation and migration.

Figure 2.

The global view of the identification and validation of ICI-related biomarkers. A, We collected 875 melanoma, 625 non–small cell lung cancer, 348 metastatic urothelial cancer, 249 renal cell carcinoma, 215 bladder cancer, 146 glioma, 141 colorectal cancer, 139 head and neck cancer, 126 esophagogastric cancer, 61 gastric cancer, 44 breast cancer, 31 hepatocellular carcinoma, 22 prostate cancer, and 15 basal cell carcinoma. B, The odds ratio and significance of each biomarker in the discovery cohort as well as in the three validation cohorts, the melanoma, the urothelial cancer, and the pan-cancer. NSCLC, non–small cell lung cancer; MUC, metastatic urothelial cancer. HNSC, head and neck cancer; Ns., nonsignificant.

As expected, we found TMB is the most significant biomarker among all genetic features (OR for responders vs. nonresponders = 1.59; 95% confidence interval, 1.12–2.27; and P value = 9.50e−03; Fig. 2B). We also observed a strong relationship between TMB and survival (Supplementary Figs. S1A and S1B). Patients with lower TMB tend to have reduced overall survival (OR = 1.50; 95% CI, 1.23–1.83; P = 5.14e−05) and progression free survival (OR = 3.72; 95% CI, 1.70–8.10; P = 9.64e−04). This could be explained by increased tumor immunogenicity induced by high TMB (56).Except TMB, we did not find any other genetic features related to ICI response using the random-effect meta-analysis model. Because somatic mutations are typically rare (57), resulting in unstable or undefined effect estimates in each individual study, the random-effect model to combine effect estimates in each individual study may not work well in this situation. To address the issue, we tried an alternative strategy by aggregating multiple datasets into one (Materials and Methods). The aggregative analysis identified mutations of 103 genes related to ICI outcomes (Supplementary Fig. S2). However, all the associations became insignificant after we adjusted by TMB. We noticed that most of those genes have been reported as frequently mutated genes due to their long coding-regions, such as TTN, MUC16, and DNAH5 (58). Therefore, we believed that those associations were purely driven by TMB alone. Although the DNA damage response signalling pathway has been reported to be one of the determinants of immunotherapy response (13), we found the link is just driven by the association between the pathway and TMB (P = 7.5e−15). Its relationship with the response became insignificant after adjusting for TMB. This finding was consistent with a previous study, which discovered that tumor defective in DNA damage response would accumulate mutations, resulting in high TMB (59). TMB stands out as the most significant biomarker, surpassing other mutations in genes or pathways. This implies that variations in the DNA damage pathway are just one among several factors contributing to a high TMB, which, in turn, triggers a response to ICI.

At the transcriptomics level, we observed expression of 10 genes associated with ICI response (FDR < 0.1). The strongest predictor was IFNG expression (OR = 1.60; 95% CI, 1.35–1.91; P = 8.21e−8), followed by IRF1 (OR = 1.59; 95% CI, 1.25–2.02; P = 1.70e−4), LAG3 (OR = 1.54; 95% CI, 1.30–1.82; P = 3.64e−7), CXCL10 (OR = 1.52; 95% CI, 1.26–1.83; P = 1.14e−5), PDCD1 (PD-1; OR = 1.46; 95% CI, 1.20–1.78; P = 1.22e−4), IL21 (OR = 1.43; 95% CI, 1.21–1.67; P = 1.55e−5), GTSF1 L (OR = 1.43; 95% CI, 1.19–1.73; P = 1.6e−4), CCL19 (OR = 1.43; 95% CI, 1.18–1.73; P = 2.93e−4), ZBED2 (OR = 1.43; 95% CI, 1.17–1.76; P = 6.10e−4), and CXCL11 (OR = 1.40; 95% CI, 1.19–1.64; P = 3.24e−5; Fig. 2B). All of 10 genes are well known immune-related genes except GTSF1L. High amplification of GTSF1 L was found to be linked to higher infiltration of CD8⁺ naive T cell and resting mast cell in cervical cancer in a recent study (60), suggesting its role in immune regulation. Among the nine immune-related genes, PDCD1 (PD-1) and LAG3 are immune checkpoint proteins. Expression of PDCD1 has been demonstrated to be a predictor of response in several studies (2–4), and anti-LAG3 is expected as the foremost inhibitor next to anti-PD-1/PD-L1 and anti-CTLA4 in the development of cancer immunotherapy (5). IFNG, IRF1, IL21, and ZBED2 are all involved in interferon gamma signalling pathway. IRF1 is IFNG-induced transcription factors that regulates immune response (61), whereas ZBED2 competes with IRF1 for binding to gene promoters (62), and IL21 in synergy with IL15 and IL18 stimulates IFNγ production in T and natural killer (NK) cells (63, 64). CXCL10, CXCL11, and CCL19 are chemokines for T cells migration (65, 66). IFNγ, secreted by active T and NK cells, mediates polarization of macrophages to a “M1-like” state and induces them to release CXCL10 and CXCL11 (66), which would recruit T cells through CXCR3 binding. On the basis of these findings, we hypothesized that the positive feedback between macrophages activation and T cells recruitment might play an important role in ICI response.

To explore the association between pathway activities and ICI outcome, we defined the pathway activity as an average expression of genes in the pathway (Materials and Methods). We observed six pathways significantly associated with ICI outcome in the discovery cohort (FDR < 0.1), including PD1 blockade (OR = 1.53; 95% CI, 1.22–1.93; P = 3.02e−4), CTL pathway (OR = 1.52; 95% CI, 1.21–1.91; P = 3.84e−4), T-cell receptor binding (OR = 1.52; 95% CI, 1.20–1.93; P = 5.7e−4), CD8⁺ alpha beta T-cell activation (OR = 1.51; 95% CI, 1.18–1.92; P = 9.7e−4), NO2-dependent IL12 pathway in natural killer cells (NO2IL12) pathway (OR = 1.51; 95% CI, 1.17–1.94; P = 1.4e−3), and CXCR3 chemokine receptor binding (OR = 1.46; 95% CI, 1.17–1.83; P = 9.6e−4; Fig. 2B). There were some redundancy among the first five pathways, which had highly correlated activity scores and were all indicative of T cells activation and IFNγ secretion. CXCR3 chemokine receptor binding includes CXCL9, CXCL10, and CXCL11, which are mainly produced by macrophages after stimulation with IFNγ and key T cells chemo-attractants.

Furthermore, we investigated the relationships between immune cell compositions and ICI outcome, which were estimated by CIBERSORT (Materials and Methods). We found M1 macrophage was most associated with ICI response (OR = 1.47; 95% CI, 1.22–1.77; P = 4.36e−s5; Fig. 2B).

Pairwise expression relationships of immune checkpoint genes had been demonstrated to be good predictors of ICI response (18). We expanded pairwise relations by considering not only immune checkpoint genes, but also genes involved in T cells migration and recruitment. Focusing on 45 immune checkpoint genes, five CXCR3 binding ligands (CXCL9, CXCL10, CXCL11, CXCL13, and PF4; ref. 14) and CXCR3 itself, we generated 1,275 pairwise relationships (Materials and Methods). We found 10 relationships, including CXCL9_CD276, CXCL13_PVR, ICOS_CD200R1, CXCL13_PDCD1LG2, CXCL10_CD200, CXCL13_CTLA4, CTLA4_CD200R1, CXCL13_CD200, CXCL9_CD200, and LAG3_PVR, were significantly associated with ICI response (Fig. 2B). Defining an interaction score as the average score of the 10 pairs, we found interaction score  ≥0.7 was significantly associated with ICI response (OR = 1.63; 95% CI, 1.31–2.03; P = 1.10e−5; Fig. 2B), which achieved higher OR than any individual pair. These findings suggest that combining immune checkpoint genes with genes responsible for T-cell recruitment could provide a more comprehensive understanding of ICI-related immune interactions. In addition, the interaction score showed a significant association with overall survival (OR = 2.84; 95% CI, 1.60–5.03; P = 3.48e−4; Supplementary Fig. S3A) and progression free survival (OR = 5.76; 95% CI, 2.76–12.02; P = 3.09e−6; Supplementary Fig. S3B).

Biomarkers validated in three independent cohorts encompassing various cancer types

We conducted further investigations to validate whether the biomarkers identified in the discovery cohort could be confirmed in three independent cohorts comprising melanoma, urothelial cancer, and pan-cancer cases. These cohorts were assembled from five different datasets (Materials and Methods). We confirmed TMB was higher in responders compared with nonresponders in the melanoma cohort with an OR of 1.60 [1.14–2.25] and a P value of 7.07e−03 and the pan-cancer cohort with an OR of 1.42 [1.06–1.90] and a P value of 1.99e−02 (Fig. 2B; Supplementary Table S2). We also observed a strong relationship between TMB and progression free survival in the melanoma cohort with an OR of 2.18 [1.25–3.80] and a P value of 6.25e−3 (Supplementary Fig. S1C).

We successfully validated the association between high expression of IFNG, IRF1, LAG3, CXCL10, PDCD1, CCL19, and CXCL11 with responders in all three cohorts (Fig. 2B; Supplementary Table S2). For instance, high expression of IFNG exhibited significant links to ICI responders with an OR of 1.55 [1.12–2.15] and a P value of 7.92e−3 in the melanoma cohort. In the urothelial cancer cohort, it showed an OR of 2.16 [1.18–3.92] with a P value of 1.18e−02, and an OR of 1.52 [1.20–1.91] with a P value of 3.86e−04 in the pan-cancer cohort. In addition, we confirmed the association between high expression of IL21, GTSF1L, and ZBED2 with ICI responders in melanoma and pan-cancer cohorts (Fig. 2B; Supplementary Table S2). For instance, high expression of IL21 demonstrated significant associations with ICI responders with an OR of 1.91 [1.23–2.96] and a P value of 4.2e−3 in the melanoma cohort, and an OR of 1.64 [1.26–2.12] and a P value of 1.71e−04 in the pan-cancer cohort (Fig. 2B).

Meanwhile, we verified the association between the expression activities of six pathways and ICI response, including PD1 blockade, CTL pathway, T-cell receptor binding, CD8⁺ alpha beta T-cell activation, NO2IL12 pathway, and CXCR3 chemokine receptor binding (Fig. 2B; Supplementary Table S2). For instance, the activity of CXCR3 binding exhibited a significant association with an OR of 1.57 [1.11–2.23] and a P value of 1.08e−2 in the melanoma cohort, an OR of 3.24 and a P value of 1.56e−04 in the urothelial cancer cohort, and an OR of 1.75 [1.40–2.19] and a P value of 7.22e−07 in the pan-cancer cohort (Fig. 2B).

With regard to immune cell compositions, we substantiated the association between M1 macrophage and ICI response in all three validation cohorts, showing an OR of 1.41 [0.99–2.00] with a P value of 0.05 in the melanoma cohort, an OR of 2.76 [1.54–4.97] with a P value of 6.61e−04 in the urothelial cohort, and an OR of 1.52 [1.23–1.89] with a P value of 9.51e−05 in the pan-cancer cohort (Fig. 2B).

We observed significant associations of pairwise expression relationships of the 10 pairs with ICI outcomes in two or three cohorts (Fig. 2B; Supplementary Table S2). Particularly noteworthy is the interaction score derived from these 10 pairs, which exhibited a significant association with ICI response in all three cohorts, with an OR of 1.82 [1.35–2.44] and a P value of 7.07e−05 in the melanoma cohort, an OR of 2.27 [1.33–3.86] and a P value of 2.37e−03 in the urothelial cohort, and an OR of 1.83 [1.46–2.28] and a P value of 7.76e; 95% CI, 08 in the pan-cancer cohort (Fig. 2B). In addition, the interaction score also demonstrated a significant association with overall survival, with an OR of 1.98 [1.17–3.33] and a P value of 1.07e−02 (Supplementary Fig. S3C), and progression-free survival, with an OR of 4.99 [2.42–10.28] and a P value of 1.36e−05, in the pan-cancer cohort (Supplementary Fig. S3D).

In summary, the majority of features identified in the discovery cohort were validated in at least two independent cohorts comprising various cell types. This further underscores their robustness and potential important roles in ICI response.

Positive feedbacks between macrophages and T cells sensitize tumor to ICI response

The identified biomarkers associated with ICI response suggest the significant involvement of macrophages and T cells, particularly highlighting the role of their interaction in driving macrophage activation and the recruitment of T cells. To gain a deeper understanding of the cell–cell interactions contributing to ICI response, we reanalyzed two single-cell RNA-seq datasets obtained from renal cell carcinoma (40) and breast cancer (41).

In renal cell carcinoma, we observed higher expression of LAG3 and PDCD1 in T cells in responders than nonresponders (Fig. 3A), which verified that overexpressed co-inhibitory receptors increased the sensitivity of T cells status change from inhibition to activation. We also found elevated expression of IFNG in T cells and NK cells in responders compared with nonresponders (Fig. 3A). Expression of IRF1, associated with IFNγ response and antigen processing on MHC class I (40), was upregulated in M1 macrophages, tumor-associated macrophages (TAM), tumor cells, monocytes and dendritic cells in responders (Fig. 3A), indicating proinflammatory microenvironment. CXCL10 and/or CXCL11 exhibited higher expression in myeloid cells, including M1 macrophages, TAM, monocytes, endothelial cells, and dendritic cells in responders than nonresponders (Fig. 3A). CXCR3, a chemokine receptor of CXCL10 and CXCL11, showed higher expression in T cells, NK cells, B cells, and endothelial cells in responders than nonresponders (Fig. 3A). Notably, the two responders in renal cell carcinoma study had higher percentage of M1 macrophages in all myeloid cells than the two nonresponders (4.9% and 4.9% in responders vs. 3.3% and 0.5% in nonresponders). Most notably, we discovered dysregulated ligand–receptor interactions between macrophages and T cells in responders compared with nonresponders using scLR (Materials and Methods; ref. 52). We found the most upregulation of IFNG-IFNGR2 between T cells and M1 macrophages (Fig. 3B). In responders, IRF1 was overexpressed in T cells and CXCL11 production was induced in M1 macrophages (67). The induced inflammatory cytokine release from M1 macrophages further enhance T cells migration and recruitment, which would be confirmed by the upregulation of CXCL10–CXCR3/CXCL11–CXCR3 interactions between M1 macrophages and T cells (Fig. 3B). Consistently, previous studies discovered that the binding of CXCR3 and its ligands (CXCL9, CXCL10, and CXCL11) were associated with levels of tumor infiltrated lymphocytes (68) and macrophage-derived CXCL9 and CXCL10 were required for immune response of ICI (69).

Figure 3.

Differential expression of ICI-related biomarkers, dysregulated ligand–receptor, and dynamic expression in responders compared with nonresponders. A, Different expression of ICI-related biomarkers in responders compared with nonresponders in each cell type. Dot plot for IFNG, IRF1, LAG3, CXCL10, CXCL11, and CXCR3. The color intensity of each dot corresponds to gene expression level, with the dot's size indicating the percentage of cells expressing that gene. B, Dysregulated IFNG-IFNGR2, CXCL10–CXCR3, and CXCL11–CXCR3 interactions in responders compared with nonresponders. C, Boxplot of z-scores for LAG3, PDCD1, IFNG, IRF1, IL21, CTL, and CXCR3 binding pathways. The significance of differential expression between matched pre- and on-therapy samples was assessed using a two-sided paired t test.

The gene expression and ligand–receptor interaction analysis in breast cancer yielded similar results (Supplementary Fig. S4). When comparing responders to nonresponders, we observed higher expressions of LAG3, PDCD1, CXCR3, and IFNG in various types of T cells, upregulated CXCL10 and CXCL11 expression in M1 macrophages (labeled as macrophage-MMP9 and macrophage-CCL2 in the original study), and elevated IRF1 expression in macrophage-CCL2 cells (Supplementary Fig. S4A). Notably, we consistently identified an upregulated IFNG–IFNGR2 interaction between T cells and macrophage-MMP9, along with increased CXCL10–CXCR3 and CXCL11–CXCR3 interactions between macrophage-MMP9 and T cells (Supplementary Fig. S4B). These findings from our analysis of two distinct cancer studies further reinforce the notion that positive feedback loops between M1 macrophages and T cells contribute to their persistent and enhanced interactions, ultimately impacting ICI response.

Furthermore, we investigated dynamic transcriptomic changes during treatment in a matched melanoma study (36). Baseline transcriptional programs (pretreatment) validated the association between ICI response and eight expression signatures identified in the discovery cohort except the expression of GTSF1 L and ZBED2. More interestingly, the expression of immune checkpoint molecules LAG3 and PDCD1, IFNG, IRF1, IL21, CTL and CXCR3 binding pathways increased very significantly after treatment in responders. In contrast, they did not show a significant change in nonresponders (Fig. 3C). Upregulated IFNG, CTL, and CXCR3 binding pathways on-therapy in responders indicates the importance of dynamic interactions between macrophages and T cells contributing to response. Interestingly, responders had higher baseline M1 macrophage levels than nonresponders, but their M1 macrophage levels stay stable after treatment (Fig. 3C), suggesting preexisting M1 macrophages and its stimulation is important to ICI response.

In summary, besides costimulatory/co-inhibitory, interactions between macrophages and T cells also contribute to ICI response. Activated T cells secrete IFNγ, which induces lysis of tumor cells and stimulates M1 macrophages. Simultaneously, M1 macrophages exert antitumor function, which includes directly mediating cytotoxicity and promoting T cells recruitment by secreting CXCL10 and CXCL11.

A multivariable predictive model outperforms TMB and EaSleR in three independent cohorts encompassing various cancer types

In total, we identified 19 biomarkers, including TMB reflecting neoantigen load, IFNG, and chemokine expression signifying T cells and macrophage activation, and pairwise relationships between immune checkpoints and chemokines released by macrophage, suggesting macrophage–T cell interactions. On the basis of these 19 biomarkers, we built a multivariable predictive model (MVP_19) using a decision tree–based ensemble machine learning algorithm XGBoost (Materials and Methods; ref. 55). We compared the performance of MVP_19 and TMB-only model (TMBM) in the discovery cohort, which has been approved by the FDA as a diagnostic measure for ICI treatment. MVP_19 achieved an AUC value of 0.72, which was significantly higher than TMBM, with an AUC score of 0.59 (DeLong test, P < 2.2e−16; Fig. 4A). In addition, the model trained on the biomarkers excluding TMB (MVP_19_noTMB) also demonstrated superior performance compared with TMBM, with an AUC of 0.69 (DeLong test, P = 2.4e−10; Fig. 4A). These results underscore the prognostic significance of these biomarkers, which complement TMB in predicting ICI response.

Figure 4.

Multivariate predictive models of ICI response. A, ROC curves and AUC values comparing MVP_19, MVP_19_noTMB, and TMBM in the discovery cohort. B, Pairwise correlations between 19 ICI-related biomarkers. C, ROC curves and AUC values comparing MVP, MVP_noTMB, and TMBM in the discovery cohort. D, Importance scores of eight nonredundant biomarkers used in the MVP. E, ROC curves and AUC values comparing MVP, MVP_noTMB, TMBM, and EaSleR in the three validation cohorts. MVP, a multivariable model based on the eight nonredundant biomarkers. TMBM, the model based on TMB-only.

Some redundancy was observed among the 19 biomarkers due to gene and functional overlap, leading to highly correlated activities (Fig. 4B). After eliminating redundant biomarkers that exceeded a correlation threshold of 0.6, we retained eight biomarkers, which included TMB, the interaction score, the expression of IL21, GTSF1L, CCL19, ZBED2, the CTL pathway, and the CXCR3 receptor binding pathway. The simplified multivariable predictive model based on these eight biomarkers (MVP) achieved an AUC score of 0.68, significantly outperforming TMBM with an AUC of 0.59 (DeLong test, P = 2.9e−11; Fig. 4C). In addition, the MVP_noTMB model (AUC = 0.66) also exhibited superior performance compared to the TMBM model (DeLong test, P = 3.3e−4; Fig. 4C). Of the eight biomarkers, the interaction score, TMB, and the CTL pathway emerged as the most pivotal features, accounting for over 50% of the total importance score (Fig. 4D). Interestingly, the interaction score held an even higher importance score than TMB, underscoring the crucial importance of simultaneously considering immune activation and migration in predicting ICI response.

We further evaluated the performance of the MVP model in three independent validation cohorts representing various cancer types (Fig. 4E; Supplementary Fig. S5). In addition to comparing the MVP with TMBM, we also assessed the MVP against another existing model, EaSleR (17), which predicts ICI response based on system-wide signatures of the tumor microenvironment, immune-cell composition and intra- and intercellular communications. The MVP achieved AUC values of 0.70, 0.61, and 0.66 in the three cohorts, respectively, which were significantly higher than the AUC values of 0.56, 0.48, and 0.58 obtained by TMBM (P = 1.5e−5; P = 6.8e−3; and P = 1.7e−3 by DeLong test; Fig. 4E). Furthermore, the MVP exhibited higher AUC values than EaSleR in the three cohorts as well, with EaSleR AUC values of 0.55, 0.52, and 0.57, respectively (Fig. 4E). The MVP_noTMB had significant higher AUC values than TMBM in the melanoma and urothelial cancer cohorts (P = 0.03 and 0.01, respectively), but not in the pan-cancer cohort (P = 0.85).

In addition, we assessed the contribution of macrophage–T cell interactions to the model's predictive performance. We compared various predictive models, including the MVP, the model with only T signatures without macrophage signatures (MVP_noCXCR3), the model with only macrophage signatures without T signatures (MVP_noCTL), and the model without either T or macrophage signatures (MVP_noCTL&CXCR3). Across the three validation cohorts, we consistently observed that the MVP achieved much higher AUC values compared with the other three models, whereas the three models produced similar AUC values (Supplementary Fig. S6). These findings indicate that T or macrophage signatures alone have minimal impact on prediction, whereas the interaction between macrophages and T cells plays a substantial role in predicting ICI response. The superior performance of the MVP in the three independent validation cohorts encompassing various cancer types further underscores the prognostic value of these biomarkers.

Discussion

Here, we performed the largest meta-analysis of over 3,000 patients across 14 tumor types to identify ICI response-related features. We not only identified known features such as TMB, IFNG, PDCD1, and LAG3 expression, but also novel biomarkers such as proportions of M1 macrophage, expression of IL21, ZBED2, CCL19, CTL pathway and CXCR3 binding pathway. Multiple genes and pathways related to ICI response are involved in macrophage activity, T-cell cytotoxicity and recruitment, suggesting the potential role of interaction between macrophages and T cells in cancer immunotherapy. Single-cell RNA profiling confirmed that IFNG from T cells induced IRF1 expression of M1 macrophage, TAM, and tumor, promoting T-cell chemo-attractants CXCL10 and/or CXCL11 secretion in those cells. The secretion of CXCL10/11 further recruited T cells and enhanced antitumor immunity through CXCR3–CXCL10/11 axis. Pre and on-therapy transcriptome not only validated the upregulated baseline levels of IFNG, CTL and CXCR3 binding in responders compared to nonresponders, but also uncovered their elevated expression after treatment, further suggesting positive feedbacks between macrophages and T cells in responders. The interaction score considering not only T cells activation via immune checkpoint proteins, but also T cells recruitment through CXCR3 axis, becomes the most important feature in the multivariate predictive model. Its high importance over TMB further indicates the contribution of interaction between macrophages and T cells in immunotherapy.

The MVP model, utilizing eight nonredundant signatures, demonstrated superior performance when compared with the TMBM and EaSleR in three independent validation cohorts encompassing various cancer types. Furthermore, comparisons involving different MVP variants, such as MVP_noTBM, MVP_noCXCR3, MVP_noCTL, and MVP_noCTL&CXCR3, underscored the critical role of these identified signatures, especially the interaction between macrophages and T cells. While previous studies have linked T-cell and macrophage activity to ICI response (68, 69), our findings emphasize that individual T or macrophage signatures have minimal impact, whereas their interaction substantially predicts ICI response.

Notably, MVP, despite achieving lower AUC values than MVP_19 in the discovery cohort, exhibited improved performance in all three validation cohorts (Supplementary Fig. S7), suggesting that a model based on nonredundant features may provide better generalization. While we calculated pathway activity based on the average expression of genes within the pathway, we observed that different measurement strategies have a subtle impact on the performance. For instance, the pathway activity determined by single-sample gene set enrichment analysis (ssGSEA; ref. 42) exhibited a high correlation with average expression, and we obtained similar AUC values when utilizing pathway activity derived from ssGSEA (Supplementary Fig. S8).

Although we identified novel biomarkers underpinning ICI response and improving clinical decision, up to 30% patients remain unexplained. It might be necessary to explore more sophisticated relations beyond individual genes/pathways or dynamic of gene regulations. For example, we found the MVP model utilizing expression changes of IL21, GTSF1L, CCL19, ZBED2, CTL, and CXCR3 binding pathways obtained a higher AUC of 0.95 than MVP with static features (AUC = 0.90; P = 0.08 of DeLong test; Supplementary Fig. S9). Dynamic transcriptome was more interpretable than “snapshot” expression due to the capture of immune-genomic reaction to the treatment. However, limited data generated from matched pre- and on-treatment biopsies greatly hindered systematic analysis for dynamic regulation. Although our study was designed to identify robust pan-cancer ICI-response signatures using a random effect meta-analysis model, we did observe heterogeneous predictive effects of specific biomarkers across various cancer types. Notably, TMB emerged as the most significant feature in melanoma, whereas ZBED2 expression took precedence in renal cell carcinoma. When an adequate number of samples are accessible for each cancer type, the development of cancer-specific predictive models may lead to enhanced performance.

More importantly, complicated tumor-immune microenvironments and spatially defined cell communities have been recognized as contributing to ICI response (70). Bulk genomics and transcriptomics, although provide valuable insights into immunotherapy, obscure cell compositions, cell-type–dependent signatures, tumor heterogeneity, and tumor-immune interactions. Although we performed single-cell transcriptomics analysis and revealed cell-dependent IFNG and chemoattractants secretion and immune interactions, the study was informative but not conclusive by the limited sample size. Large-scale single-cell and spatial omics studies, with the ability to provide the resolution and richness of data, become the next frontier to decipher the mechanism of how tumor and immune systems co-evolve in concert for responses or resistances.

Supplementary Material

Supplementary Figure S1

Associations of TMB with overall survival and progression free survival across multiple data sets. A and B. The association between TMB and overall survival and progression free survival in the discovery cohort. C. The association between TMB and progression free survival in the validation datasets.

Supplementary Figure S2

Associations between gene mutations and ICI response before and after TMB adjustment.

Supplementary Figure S3

Associations of the interaction score with overall survival and progression free survival in the discovery and validation cohorts. A and B. The association between the interaction score and overall survival and progression free survival in the discovery datasets. C and D. The association between interaction score and overall survival and progression free survival in the validation datasets.

Supplementary Figure S4

Differential expression of ICI-related biomarkers and dysregulated ligand-receptor in the breast cancer study. A. Different expression of ICI-related biomarkers in responders and non-responders in each cell type. Dot plot for IFNG, IRF1, LAG3, CXCL10, CXCL11 and CXCR3. The color intensity of each dot corresponds to gene expression level, with dot’s size indicating the percentage of cells expressing that gene. B. Dysregulated IFNG-IFNGR2, CXCL10-CXCR3 and CXCL11-CXCR3 interactions in responders compared to non-responders.

Supplementary Figure S5

ROC curves and AUC values for MVP compared to TMBM and EaSleR in the five validation data sets.

Supplementary Figure S6

ROC curves and AUC values for MVP, MVP_noCTL, MVP_noCXCR3, and MVP_noCTL&CXCR3 in the three validation cohorts.

Supplementary Figure S7

ROC curves and AUC values for MVP and MVP using ssGSEA to measure pathway activity in the three validation cohorts.

Supplementary Figure S8

ROC curves and AUC values for MVP, MVP_32, MVP_19, and TMBM in the three validation cohorts.

Supplementary Figure S9

ROC curves and AUC values for MVPs with dynamic expression and static expression, and TMBM.

Supplementary Table S1

A detailed summary of the datasets used in this study, including download sources, data availability (raw or processed), and the total number of samples retained after quality control.

Supplementary Table S2

The odds ratio and significance of each biomarker in five independent validation datasets.

Authors' Disclosures

No disclosures were reported.

Authors' Contributions

J. Yang: Resources, data curation, software, formal analysis, investigation, visualization, methodology, writing–original draft. Q. Liu: Supervision, funding acquisition, methodology, project administration, writing–review and editing. Y. Shyr: Supervision, funding acquisition, investigation, methodology, project administration.