Comprehensive evaluation of composite gene features in cancer outcome prediction

Discoidin Domain Receptor-Driven Gene Signatures as Markers of Patient Response to Anti-PD-L1 Immune Checkpoint Therapy

BACKGROUND

Anti-programmed cell death 1 (anti-PD-1) and PD ligand 1 (PD-L1) immune checkpoint therapies (ICTs) provided durable responses only in a subset of cancer patients. Thus, biomarkers are needed to predict nonresponders and offer them alternative treatments. We recently implicated discoidin domain receptor tyrosine kinase 2 (DDR2) as a contributor to anti-PD-1 resistance in animal models; therefore, we sought to investigate whether this gene family may provide ICT response prediction.

METHODS

We assessed mRNA expression of DDR2 and its family member DDR1. Transcriptome analysis of bladder cancer (BCa) models in which DDR1 and 2 were perturbed was used to derive DDR1- and DDR2-driven signature scores. DDR mRNA expression and gene signature scores were evaluated using BCa-The Cancer Genome Atlas (n = 259) and IMvigor210 (n = 298) datasets, and their relationship to BCa subtypes, pathway enrichment, and immune deconvolution analyses was performed. The potential of DDR-driven signatures to predict ICT response was evaluated and independently validated through a statistical framework in bladder and lung cancer cohorts. All statistical tests were 2-sided.

RESULTS

DDR1 and DDR2 showed mutually exclusive gene expression patterns in human tumors. DDR2high BCa exhibited activation of immune pathways and a high immune score, indicative of a T-cell-inflamed phenotype, whereas DDR1high BCa exhibited a non-T-cell-inflamed phenotype. In IMvigor210 cohort, tumors with high DDR1 (hazard ratio [HR] = 1.53, 95% confidence interval [CI] = 1.16 to 2.06; P = .003) or DDR2 (HR = 1.42, 95% CI = 1.01 to 1.92; P = .04) scores had poor overall survival. Of note, DDR2high tumors from IMvigor210 and CheckMate 275 (n = 73) cohorts exhibited poorer overall survival (HR = 1.56, 95% CI = 1.20 to 2.06; P < .001) and progression-free survival (HR = 1.77 95%, CI = 1.05 to 3.00; P = .047), respectively. This result was validated in independent cancer datasets.

CONCLUSIONS

These findings implicate DDR1 and DDR2 driven signature scores in predicting ICT response.

De novo pathway-based biomarker identification

Gene expression profiles have been extensively discussed as an aid to guide the therapy by predicting disease outcome for the patients suffering from complex diseases, such as cancer. However, prediction models built upon single-gene (SG) features show poor stability and performance on independent datasets. Attempts to mitigate these drawbacks have led to the development of network-based approaches that integrate pathway information to produce meta-gene (MG) features. Also, MG approaches have only dealt with the two-class problem of good versus poor outcome prediction. Stratifying patients based on their molecular subtypes can provide a detailed view of the disease and lead to more personalized therapies. We propose and discuss a novel MG approach based on de novo pathways, which for the first time have been used as features in a multi-class setting to predict cancer subtypes. Comprehensive evaluation in a large cohort of breast cancer samples from The Cancer Genome Atlas (TCGA) revealed that MGs are considerably more stable than SG models, while also providing valuable insight into the cancer hallmarks that drive them. In addition, when tested on an independent benchmark non-TCGA dataset, MG features consistently outperformed SG models. We provide an easy-to-use web service at http://pathclass.compbio.sdu.dk where users can upload their own gene expression datasets from breast cancer studies and obtain the subtype predictions from all the classifiers.