Clinical cell-surface targets in metastatic and primary solid cancers

Abstract

Therapies against cell-surface targets (CSTs) represent an emerging treatment class in solid malignancies. However, high-throughput investigations of CST expression across cancer types have been reliant on data sets of mostly primary tumors, despite therapeutic use most commonly in metastatic disease. We identified a total of 818 clinical trials of CST therapies with 78 CSTs. We assembled a data set spanning RNA-seq and microarrays in 7,927 benign samples, 16,866 primary tumor samples, and 6,124 metastatic tumor samples. We also utilized single-cell RNA-seq data from 36 benign tissues and 558 primary and metastatic tumor samples, and matched RNA versus protein expression in 29 benign tissue samples, 1,075 tumor samples, and 942 cell lines. High RNA expression accurately predicted high protein expression across CST therapies in benign tissues, tumor samples, and cell lines. We compared metastatic versus primary tumor expression, identified potential opportunities for repositioning, and matched cell lines to tumor types based on CST and global RNA expression. We evaluated single-cell heterogeneity across tumors, and identified rare normal cell subpopulations that may contribute to toxicity. Finally, we identified combinations of CST therapies for which bispecific approaches could improve tumor specificity. This study helps better define the landscape of CST expression in metastatic and primary cancers.

Keywords: Clinical practice; Oncology.

Figure 1. Cell-surface target tumor expression.

(A) Schematic of all data used in this study. (B) Hierarchical clustering of median bulk RNA expression of 78 cell surface protein targets (rows) across 43 solid cancer types (columns: primary and metastatic). RNA expression levels are percentile rank normalized across all genes, ranging from 0 to 1. Black and white asterisks = FDA-approved indication.

Figure 2. Primary versus metastatic tumor expression.

A Rank-biserial correlation was calculated between the primary and metastatic samples in 23 tumor types (columns), with at least 10 samples in each group for each of the 78 cell-surface targets (rows). Within each histology, negative values (red) indicate enrichment of the target in metastatic tumors of that histology, while positive values (blue) indicate enrichment of the target in primary tumors. Hierarchical clustering was performed. Asterisks indicate Wilcoxon’s rank-sum FDR comparing primary versus metastatic expression: *P ≤ 0.05, **P ≤ 0.01, ***P ≤ 0.001, ****P ≤ 0.0001.

Figure 3. Drug-repositioning candidates.

Box-and-whisker plots show the metastatic (blue) and primary (gray) cancer expression in the RNA-seq data for various cell-surface targets. RNA expression levels are percentile rank normalized, ranging from 0 to 1. Green/gray/red dotted lines: 1st/50th/90th percentiles of housekeeping gene expression. Middle line: 50th percentile (median). Lower hinge: 25th percentile. Upper hinge: 75th percentile. Whiskers: minimum and maximum.

Figure 4. Selecting a cell line model system.

(A) Schematic of rationale for identifying optimal cell line based on both cell-surface target expression and correlation with high-purity clinical tumor samples of the intended patient population. (B) Number of cell lines that met the following criteria: (i) cell lines with expression that was in the 95th percentile of expression or greater for each target in the tumor RNA-seq (horizontal blue line) and (ii) 95th percentile or greater of correlation with tumor samples of each primary/metastatic tumor type (vertical blue line). Gene expression for each cell line was correlated with each high-purity primary/metastatic cancer sample, and a median was calculated for each cancer type. Each data point represents 1 cell line target expression (y axis) versus the median correlation of that cell line to a primary or metastatic tumor type (x axis). Red/green point = cell lines of the same cancer type as the metastatic/primary cancer tissue sample that meet criteria i and ii. Gray dot = cell lines of a different cancer type that meet criteria i and ii. (C) Number of cell lines (indicated by dot size) that meet criteria i and ii across tumor types (rows) and targets (columns). Green and red dots represent cell lines that are the same cancer type for primary and metastatic cancer, respectively. Gray dots represent cell lines of a different cancer type.

Figure 5. Single-cell heterogeneity.

(A) Heatmap of the proportion of the tumor cells in each scRNA-seq sample that expresses each cell-surface target. Rows are targets and columns are scRNA-seq samples. Red represents samples with a high proportion of tumor cells expressing each target, green represents samples where a more mixed proportion of tumor cells express each target, and white represents an absence of cell-surface target expression. (B) Percentage of cells positive for ERBB2, ERBB3, and TACSTD2 across all cell types and within each cell type subpopulation in replicate lung scRNA-seq samples. Middle line: 50th percentile (median). Lower hinge: 25th percentile. Upper hinge: 75th percentile. Whiskers: minimum and maximum. (C) Detection of vedotin-conjugate target expression across different pancreatic cell types showing that target expression in α cells correlates with presence or absence of hyperglycemia as a clinical toxicity.

Figure 6. Combinations of 2 targets.

(A) We created logistic regression models for every pair of cell-surface targets, comparing each cancer type and each normal tissue type. We then only retained pairs where (i) expression was higher in cancer versus normal and the combination had good discriminative power, with an F1 score of greater than 0.95; and (ii) each individual gene was contributing independently, with both having Wald test P values (corrected for multiple testing) of less than 0.05. The number of combinations as a function of the proportion of cancer-normal tissue comparisons meeting these 2 criteria is shown. The red line indicates cell surface target combinations where greater than 90% of the cancer-normal tissue comparisons met criteria i and ii, and is where the examples shown in B are drawn. (B) Example combinations of how gene expression percentiles of 2 cell-surface targets (x and y axes) can better stratify benign samples (gray) from cancer samples (red).