Robust prediction of response to immune checkpoint blockade therapy in metastatic melanoma

Abstract

Immune checkpoint blockade (ICB) therapy provides remarkable clinical gains and has been very successful in treatment of melanoma. However, only a subset of patients with advanced tumors currently benefit from ICB therapies, which at times incur considerable side effects and costs. Constructing predictors of patient response has remained a serious challenge because of the complexity of the immune response and the shortage of large cohorts of ICB-treated patients that include both 'omics' and response data. Here we build immuno-predictive score (IMPRES), a predictor of ICB response in melanoma which encompasses 15 pairwise transcriptomics relations between immune checkpoint genes. It is based on two key conjectures: (i) immune mechanisms underlying spontaneous regression in neuroblastoma can predict melanoma response to ICB, and (ii) key immune interactions can be captured via specific pairwise relations of the expression of immune checkpoint genes. IMPRES is validated on nine published datasets^1-6 and on a newly generated dataset with 31 patients treated with anti-PD-1 and 10 with anti-CTLA-4, spanning 297 samples in total. It achieves an overall accuracy of AUC = 0.83, outperforming existing predictors and capturing almost all true responders while misclassifying less than half of the nonresponders. Future studies are warranted to determine the value of the approach presented here in other cancer types.

Figure 1.

(A) Boxplots showing IMPRES of high vs low immune response in test and validation datasets of non-ICB treated melanoma patients; P-values are computed via a one-sided Rank-sum test. Boxplots centre lines indicate medians, box edges represent the interquartile range, whiskers extend to the most extreme data points not considered outliers, and the outliers are plotted individually using the ’+’ symbol. (B) Kaplan-Meier survival curves of patients with high versus low IMPRES (computed over the combined test and validation datasets). The median IMPRES is used to define the “Low IMPRES” and “High IMPRES” subgroups. The P-value is computed via a two-sided log-rank test. (C) Upper Panel: Heatmaps showing the enrichment P-values for CDPs that are up (orange) or down (purple) regulated in responders versus non-responders across the anti-PD-1 (encapsulated in the left rectangle) and the anti-CTLA-4 melanoma datasets^,,, (right rectangle). The lower Panel displays the enrichment P-values for these CDPs in high immune response vs other subtypes in non-ICB treated melanoma, and in spontaneous regression vs non-spontaneous regression in the NB dataset. (D) Heatmaps showing the rank correlation ρ between expression levels of each CDP (vertical axis) and each of the IMPRES features ratios (horizontal axis), computed separately over the anti-PD-1 datasets, the anti-CTLA-4 datasets, the non-ICB treated melanoma datasets and the neuroblastoma dataset. White-colored entries denote non-statistically significant associations.

Figure 2.

(A) Receiver Operating Characteristic (ROC) curves quantifying IMPRES prediction AUC across numerous publically available ICB response datasets^–. (B) ROC curves for the MGH dataset of ICB response (with 10 patients treated with anti-CTLA-4 and 31 patients treated with anti-PD-1) and for the aggregate datasets including all 297 samples, the 216 samples of patients treated with anti-PD-1 and 81 with anti-CTLA-4. (C) Bar plots showing the prediction accuracy and error types for different IMPRES thresholds (where a positive label corresponds to a ‘responder’ prediction) on the aggregate compendium of 297 patients included in all 11 datasets studied. The dashed line represents the total number of responders. (D) Precision/recall evaluation of IMPRES on the same aggregate compendium. The Y-axis displays the precision/recall as a function of the number of ‘responder’ predictions made (shown on the X-axis, obtained by decreasing the classification threshold, whose value is also displayed in italic font). Prediction performance in terms of specificity and sensitivity values is provided in Supp. Table 5. (E)-(F) Kaplan Meier survival curves for the ICB treatment datasets^,, with high vs. low IMPRES scores (using the median IMPRES as a threshold differentiating between the high and low groups). The P-values are computed via a two-sided log-rank test. (G)-(H) Boxplots comparing progression free survival between low vs. high IMPRES in the ICB^, datasets (using the median IMPRES as a differentiating threshold). P-values are computed via a one-sided Rank-sum test. Boxplots centre lines indicate medians, box edges represent the interquartile range, whiskers extend to the most extreme data points not considered outliers, and the outliers are plotted individually using the ’+’ symbol.

Figure 3.

(A) AUC of IMPRES and other published predictors across 9 publicly available ICB treatment datasets grouped by treatment type and stage (pre and on stands for before and during ICB treatment). The one-sided Rank-sum P-values comparing the performance of each predictor evaluated to that of IMPRES over all datasets are presented (P-value of 0.002 is achieved when IMPRES AUC is larger than that obtained by the other predictor for all 9 datasets, and 0.004 when it is larger for 8/9 datasets). Bar centre is defined by the mean and error bars via SD. (B) The empirical P-values comparing IMPRES performance to that of each of the other predictors in the three different ICB treatment classes and for the aggregate of all datasets (using n=1000 permutations, the value of ‘<1e-3’ denotes that IMPRES’ prediction performance was superior to that of the predictor with which it was compared in all 1,000 repetitions). (C) A network representation of the 15 pairwise features comprising IMPRES. Each node represents an immune checkpoint gene and each edge describes a pairwise relation (an IMPRES feature). The direction of edge A -> B denotes that the higher expression of A vs. that of B is associated with better patients’ response. The color of the outline of each node denotes if it is inhibitory or activating and its fill color denotes whether it belongs to the PD1 or CTLA-4 pathways. (D) Clustogram (with average linkage function) of the individual predictive power of the 15 IMPRES features (based on their expression ratios) in each of the melanoma treatment datasets studied (the color scaling denotes the AUC obtained using each individual ratio as a response predictor, ranging from 0 to 1). (E) Scatter plots showing the correlation between CIBERSORT-inferred CD8+ T cells abundance (X-axis) and the gene expression ratios of two IMPRES features that are significantly associated with it (Y-axis); CD40/PD1 (upper panel) and PD1/OX40L lower panel). The Spearman ρ and associated P-values are shown for each ICB response data^,,, individually (on the right) and for all four datasets together (in the plot)