Elevated levels of MMP12 sourced from macrophages are associated with poor prognosis in urothelial bladder cancer

Patients included in the discovery cohort were diagnosed with UBC between 1986 and 1996 at the University Academic Hospital of Uppsala, Sweden. The study was performed in accordance with the Declaration of Helsinki and approved by the regional ethical review board of Uppsala (ID 2015/143/1). Plasma samples (collected in heparin tubes) from 90 of these patients and paired urine samples (morning urine) from 40 (33/40 passed QC) were stored at -80 °C and were not thawed until their analysis with PEA as described below. Reference pathology was performed by an experienced pathologist according to the 2004 WHO Classification of Bladder Tumours [16]. Clinical and histopathological characteristics are presented in Table 1. The clinical follow-up of the patients was conducted in accordance with the European and National Guidelines with regular cystoscopies and transurethral resection of the tumours when recurrences had been detected. For validation, we utilised PEA data derived from the pre-treatment serum samples of 79 UBC patients recruited to the West Midlands’ Bladder Cancer Prognosis Programme (BCPP), UK (ethical approval: 06/MRE04/65) [15, 17].

Levels of 92 immuno-oncology-related proteins (OLINK Immune Oncology panel I) in urine and plasma were quantified by PEA [19] (Proseek platform, OLINK) at the Clinical Biomarker facility (Science for Life Laboratory, Uppsala, Sweden). Briefly, 92 proteins were detected in 1 μl of sample in a multiplex assay utilising concomitant binding of matched oligonucleotide-conjugated antibody pairs. When in close proximity, the corresponding antibody oligonucleotide pairs hybridize, resulting in a unique DNA template. Following a polymerase amplification reaction, the sequences are quantified by microfluidic qPCR and presented as Normalised Protein eXpression (NPX), an arbitrary unit based on log₂-transformed Ct values (http://​www.​olink.​com).

To complement the small cohort used for the OLINK assay and to examine whether the observed differences could be noted in a larger cohort, made up of tissue biopsies, microarray data for the tumours from 308 UBC patients (in the form of non-normalised, raw intensity values along with their detection p-values) were retrieved from the gene expression omnibus (GEO) repository (GSE32894); clinical information regarding the patients and their samples was obtained from the associated publication [20]. Two of the samples were excluded from further analysis because the stage was not reported. Quantile normalisation, probe filtering and batch effect correction were performed according to the methods supplied by the original authors. Differential expression analysis was conducted in an R 4.1 environment (limma, v.3.48.0 [21], R). A linear model comprising a design matrix without intercept for stage of each sample, as well as a contrast matrix stating the comparisons to be made were fitted to the data to calculate fold changes, with positive values indicating higher expression in muscle invasive bladder cancer (MIBC) and negative values reflecting higher expression in non-muscle invasive bladder cancer (NMIBC). Patients with available survival data (n = 224) were selected for subsequent survival analyses (survival 3.2.7, R), and were split into highly- or lowly-expressing for each queried protein with cutoffs determined by maximally selected rank statistics (maxstat v.0.7.25, R).

Analysis of single-cell RNA-sequencing data from a murine UBC model and human UBC patients were explored with the goal of identifying the putative source of the mRNA transcripts coding for selected proteins from the OLINK assay. Two sets of single-cell RNA-sequencing data were downloaded and processed: unique molecular identifier (UMI) count matrices originating from bladders of a urothelial cancer mouse model [22] (GSE174182) and raw sequence reads derived from patient tumour material from the European Nucleotide Archive under the accession code PRJNA662018 [23]. To generate count matrices for patient samples, the raw sequences were aligned to the GRCh38 reference genome through CellRanger 2.2.0. Downstream processing followed the Seurat v4.0 pipeline [24]. Murine cells were filtered to retain higher quality cells (> 200 & < 8000 uniquely identified genes, < 25% of reads mapped to mitochondrial genes), whereas the filtering thresholds for the human samples remained identical to the original publication (< 6000 unique genes, > 1000 UMI counts per cell and < 10% reads stemming from mitochondrial DNA). SCTransform [25] was employed to normalise and cluster the murine-derived cells, with clusters annotated based on their top differentially expressed genes. As for the human-derived cells, an additional integration step was performed prior to clustering due to the highly patient-specific clusters observed.

For group comparisons, a two-sided Wilcoxon rank-sum test was used with adjustment for multiple testing by the Benjamini–Hochberg procedure (PEA data), or Bonferroni (mRNA data). Due to the internal normalisation procedures applied for the quantification of each protein, NPX values cannot be compared between proteins. Instead, the strength of association between protein markers was interrogated by Spearman rank correlation within or between respective sample types (serum and urine). Receiver operating characteristic (ROC) curves were constructed from logistic regression. Proteins important for overall survival (OS) were retrieved from a random survival forest model (RSF) [26]. The ranking of proteins was based on their importance in classifying the data in the model using their variable importance (VIMP) measurements. The assembly model used herein was built on the time-to-event data of the cohort and used the intersection of the 63 markers shared between the OLINK Immuno-Oncology panel I and the 422 OLINK markers investigated in the serum of UBC patients by Bryan et al. [15] (Supplementary Table 1). NPX values were scaled and trained on the Swedish cohort of 90 plasma samples, followed by validation against 79 serum samples from the UK cohort. The machine learning model was developed in the R milieu (RStudio Team 2019, Boston, USA) using the packages caret (v.6.0.86) and randomForestSRC (v.2.9.3). Maximally selected rank statistics (maxstat v.0.7.25, R) were used to dichotomize protein concentrations (high/low) in Kaplan–Meier log-rank test analyses. Multivariate overall survival estimates were calculated from Cox proportional hazard regression models comprising the covariates stage, grade, sex and age (survival 3.2.7, R).

Plasma and urine samples from a treatment-naïve UBC patient cohort were subjected to a highly sensitive targeted proteomic analysis (Immuno-oncology panel I, OLINK). Clinical and demographic data are presented in Table 1. From the initial screening of 92 protein panel markers, a subset was excluded from further analyses because NPX levels were below the limit of detection (LOD) in a large proportion of samples. Markers below LOD in plasma: IL-1α, IFN-γ, IL-33, IL-21, IL-35, CXCL12, FGF2, IL-2, IFN-β, IL-4, IL-13, CD28; and in urine: IL-5, IL-13, CXCL12, TNF, CD40L, IL-10, IL-4, IFN-β, NOS3, IL-21, IFN-γ, IL-2, IL-33, IL-35. Unsupervised hierarchical clustering and uniform manifold approximation and projection (UMAP) dimensional reduction analyses of the remaining proteins in plasma and urine revealed immuno-oncology profiles predominantly driven by differences among individuals as opposed to clinical variables such as stage, grade, sex or age (Fig. 2A-D). The strength of the association between protein markers was investigated by Spearman rank correlation. The analysis revealed several proteins with significantly correlated urine and plasma concentrations from the 33 patients with paired samples. Following adjustment for multiple testing, 12 proteins involved in vital immune cell functions remained, with the strongest correlations observed for the surface-bound receptor/ligands LAMP3, MICA/B, PD-1 and CD27 (Fig. 2E).

NPX values of proteins were compared between NMIBC and MIBC patients. Systemic levels of matrix metalloproteinase 7 (MMP7) and C–C motif chemokine ligand 23 (CCL23) were significantly elevated in the plasma of MIBC patients compared to patients diagnosed with NMIBC (p = 0.028 and p = 0.03, respectively) (Fig. 3A); however, these levels were not statistically different in urine (p = 0.87 and p = 0.59, respectively). Instead, protein levels of the two TNF-receptor superfamily members, CD27 and CD40, were found to be present at significantly lower levels in the urine of MIBC patients compared to urine from NMIBC patients (p = 0.044 and p = 0.040, respectively) (Fig. 3B). To be noted, following correction for multiple testing, neither of the reported proteins reached significance (q < 0.05, FDR). When assessed as stage-specific classifiers in terms of sensitivity and specificity by calculating the area under the curve (AUC), the two urine markers CD27 (AUC = 0.71) and CD40 (AUC = 0.72) performed relatively well in combination (AUC = 0.75), whereas the plasma markers CCL23 (AUC = 0.64) and MMP7 (AUC = 0.64) alone, or in combination (AUC = 0.65) performed poorly (Supplementary Fig. 1).

Compared to univariable or multivariate regression analysis with an individual protein, the RSF model included all proteins to predict survival status. To evaluate the accuracy of the model, data from a UK cohort [15] were used. The model included the 63 proteins shared between the Swedish and UK cohorts (Supplementary Table 1) and the respective patients’ overall survival status. Data from the Swedish cohort (90 plasma samples) were used as a discovery cohort to train the model with data provided by Bryan et al. [15] used as a validation cohort. Although not a perfect validation cohort (Swedish cohort = plasma, UK cohort = serum), the machine learning model presented an error rate of 32% given the different sample types. Here, MMP12 was ranked as the most important contributor to OS, as shown in the variable importance (VIMP) plot (Fig. 4A). The model’s main purpose was to identify prognostic biomarkers; thus, it did not include clinical variables. Instead, the top-ranked proteins were validated in a subsequent multivariate Cox-regression analysis comprising the covariates stage, grade, sex, and age (Supplementary Fig. 2, Supplementary Table 2A-B). When assessed as a continuous variable, MMP12 remained significantly associated with a poor outcome (HR = 1.8, CI:1.3–2.5, p < 0.001) in the training cohort (Fig. 4B) and in the validation cohort (HR = 2.20, CI:1.42–3.41, p < 0.001) (Supplementary Fig. 3, Supplementary Table 2C-D). The association between high levels of MMP12 and poor OS was maintained in the independent validation cohort (Fig. 4C). An interesting possibility with RSF models is the ability to study how MMP12 levels influence the predicted survival over time. Here, as a proof-of-concept, survival probability is presented at 6 months, 1 year and 5 years using the aforementioned model (63 markers, clinical information excluded) (Fig. 4D).

Due to the clinical importance of systemic levels of MMPs in this cohort, illustrated by stage specificity and survival differences, we sought to investigate their expression at the transcript level in datasets generated with complementary methods. We utilised a microarray-based dataset from a UBC patient cohort [20], as well as two single-cell RNA-sequencing datasets: a clinically relevant murine urothelial cancer model developed by our group [22] and a patient cohort comprising four non-invasive and four invasive tumours [23].

Analysis of the microarray gene expression data from 213 NMIBC and 93 MIBC patient tumour biopsies revealed that MMP7 and MMP12 mRNA expression levels were higher in MIBC. An approximately 1.85-fold increase for MMP7 (p-adj = 0.0018) (Fig. 5A) and a 1.89-fold increase for MMP12 (p-adj = 0.0001) (Fig. 5B) were observed for MIBC. Survival analyses in this cohort showed that patients with high expression of MMP7 had statistically significant, poorer OS (p = 0.018) (Fig. 5C). The same trend was noticeable for patients with high expression of MMP12, however this comparison marginally failed to reach statistical significance (p = 0.055) (Fig. 5D). To be noted, the cohort examined was heavily skewed towards NMIBC patients.

Thus, we proceeded to study single-cell RNA data to fully capture the cellular source of the MMP findings. Subsequent examination of a public human urothelial cancer dataset (PRJNA662018) at the single-cell RNA level determined the source of these transcripts to be epithelial cells for MMP7 and macrophages for MMP12 transcripts (Supplementary Figs. 4 and 5). Notably, weak MMP12 mRNA expression from epithelial cells was present in less than 10% of the total cells.

Single-cell transcriptomic analysis of murine tumours further corroborated the source of human tumour-derived MMP12 expression to macrophages, as also murine Mmp12 expression was mapped to macrophages (Supplementary Fig. 5). However, the expression of Mmp7 was not adequately identified in the murine dataset. A total of 2869 cells were annotated as macrophages, 1355 of which originated from NMIBC and 1514 from MIBC murine samples (Fig. 5E). Mmp12 expression was increased > sevenfold in male-derived macrophages (p-adj = 1.63^–40) in MIBC compared to NMIBC and almost sixfold in female-derived macrophages (p-adj = 1.47^–16) in MIBC compared to NMIBC (Fig. 5E). In addition to the elevated levels of Mmp12 in MIBC, the percentages of Mmp12-expressing macrophages also increased as the disease progressed (Fig. 5F).