Exploiting convergent phenotypes to derive a pan-cancer cisplatin response gene expression signature

Abstract

Precision medicine offers remarkable potential for the treatment of cancer, but is largely focused on tumors that harbor actionable mutations. Gene expression signatures can expand the scope of precision medicine by predicting response to traditional (cytotoxic) chemotherapy agents without relying on changes in mutational status. We present a new signature extraction method, inspired by the principle of convergent phenotypes, which states that tumors with disparate genetic backgrounds may evolve similar phenotypes independently. This evolutionary-informed method can be utilized to produce consensus signatures predictive of response to over 200 chemotherapeutic drugs found in the Genomics of Drug Sensitivity in Cancer (GDSC) Database. Here, we demonstrate its use by extracting the Cisplatin Response Signature (CisSig). We show that this signature can predict cisplatin response within carcinoma-based cell lines from the GDSC database, and expression of the signatures aligns with clinical trends seen in independent datasets of tumor samples from The Cancer Genome Atlas (TCGA) and Total Cancer Care (TCC) database. Finally, we demonstrate preliminary validation of CisSig for use in muscle-invasive bladder cancer, predicting overall survival in a small cohort of patients who undergo cisplatin-containing chemotherapy. This methodology can be used to produce robust signatures that, with further clinical validation, may be used for the prediction of traditional chemotherapeutic response, dramatically increasing the reach of personalized medicine in cancer.

Fig. 1. Visual representation of convergent evolution in animals and convergent phenotypes in tumors.

a Birds and bats are genomically disparate, but both have individually evolved the ability to fly. b Two tumors may evolve cisplatin resistance independently, despite being genomically distinct from one another. Created with BioRender.com.

Fig. 2. Schematic representation of CisSig derivation.

a Description of the epithelial-origin subset of the Genomics of Drug Discovery in Cancer (GDSC) dataset (denoted with the pill icon in future figures). These data include 429 epithelial-based cancer cell lines, with drug response measurements to over 200 drugs and gene expression characterization via microarray. b Pipeline for extracting connectivity seeds. First, differential gene expression analysis between the top and bottom 20% of cisplatin responders found genes with significantly increased expression in a state of cisplatin sensitivity. These differentially expressed genes became “seed genes” in a co-expression network built using gene expression from clinical samples of epithelial-based tumors and tissue in The Cancer Genome Atlas (TCGA). Seed genes that were highly co-expressed with each other were denoted as “connectivity genes.” c Schematic of data partitioning, where GDSC epithelial-based cancer cell lines from (a) are split into 5 folds. Each fold underwent the pipeline in (b). Genes found in at least 3 of the 5 connectivity gene sets were included in the final signature, CisSig. Created with BioRender.com.

Fig. 3. Visualization of CisSig expression within GDSC dataset.

a An unclustered heatmap showing gene expression of the CisSig genes (rows) in cell lines (columns) from the top and bottom quintiles of cisplatin IC50. Color of the heatmap represents the Z-score of gene expression, normalized to each gene. Cell lines denoted as sensitive (right, yellow bar) tend to display higher expression of CisSig genes than cell lines denoted as resistant (left, green bar). Z-scores above 2.5 are denoted as 2.5, and Z-scores below −2.5 are denoted as −2.5. b Violin plots comparing the distribution of CisSig scores between the cell lines in the highest and lowest quintile of cisplatin IC50. A Wilcoxon rank-sum test found that the median CisSig scores between these two cohorts was significantly different (p < 0.0001). c Comparison of the distribution of cisplatin IC50 between cell lines in the highest and lowest quintile of CisSig score. Y-axis represents the proportion of the cohort with a cisplatin IC50 greater than the cisplatin concentration on the X-axis. A log-rank test between the two cohorts demonstrates significantly different drug response between the two cohorts (p < 0.0001). d Null distribution of hazard ratio using 1000 random gene signatures with the same length as CisSig and the model described in (c). CisSig’s performance is compared to the 95% confidence interval of the null distribution, where each signature’s performance (CisSig and nulls) is represented by the hazard ratio between two cohorts separated by the signature score. Created with BioRender.com.

Fig. 4. CisSig predicts IC50 using a variety of modeling techniques in the GDSC dataset.

a Scatterplot of the actual vs. predicted IC50 using CisSig score to predict IC50 with linear regression. Plot shows the best performing fold (measured by Spearman’s rho) from 5-fold cross validation. b Null distribution of the performance metric from (a) (Spearman’s rho), built using 1000 random gene signatures to predict IC50 as described in (a). As with CisSig, the metric of the best performing fold is used to represent each null signature. The median of the null distribution and the cutoff for the 95th percentile of the null distribution are represented by the solid and dashed gray line, respectively. CisSig’s performance, red solid line, outperforms at least 95% of the null distribution. c–d Violin plots containing the null distribution of performance metrics for 11 modeling methods, split into regression (continuous outcome) and classification (binary outcome) methods, respectively. Each distribution was created as discussed in a, b, where CisSig’s performance is compared to the performance of 1000 random gene signatures of the same length. For each violin, a shaded gray bar represents the top 5% of each null distribution and CisSig’s performance is shown with a red dot. The modeling methods, including input and output, are described in Table 2. Created with BioRender.com.

Fig. 5. Cancer subtypes with greater CisSig expression tend to have cisplatin included in standard of care guidelines.

Cancer subtypes are ranked by median CisSig Score in three datasets, GDSC (left), TCGA (middle), and TCC (right). The color of each violin plot represents the rank of the cancer subtype. The ranks of intersecting subtypes between each dataset are compared with Spearman’s rank correlation, reported with correlation ρ and p-value. Rank correlation ρ between GDSC and TCGA and GDSC and TCC datasets is 0.78 (p = 0.0002) and 0.92 (p < 0.0001), respectively. Rank correlation r between TCGA and TCC datasets is 0.93 (p < 0.0001). Violin plots display the distribution of CisSig scores for each cancer subtype. Within each violin, a boxplot denotes median signature score for each subtype (middle horizontal line) and 25th/75th percentile for signature scores (box edges). Numbers to the left of each violin plot represent sample size included in each cancer subtype. For disease sites labeled as using cisplatin in select circumstances, notes about these circumstances are included in Supplementary Table 8. Created with BioRender.com.

Fig. 6. CisSig-trained model is predictive in patients who have received cisplatin, but lacks signal in patients who have not received cisplatin.

a Schematic description of model training and testing, where a model is trained using patients who did receive cisplatin-containing treatment from Dataset A. Testing of the trained model is done using patients from the Dataset A who did not receive cisplatin-containing treatment and patients from the Dataset B who did receive cisplatin-containing treatment. b Test samples that did receive cisplatin-containing treatment are separated into groups of “high” and “low risk” based on the model’s predictions using a median cutoff. Kaplan–Meier curves show a significant separation between the two groups. c The same analysis shown in (b), using an optimal cutpoint (determined by chi-square statistic) instead of median to separate the cohorts. d, e The same analyses shown in (b, c), separating the groups into “high”, “middle”, and “low risk” groups using tertiles and the optimal two cutpoints, respectively. f, g The same analyses shown in (b) and (d), using samples from Dataset A that did not receive cisplatin-containing treatment, demonstrating no significant separation between the two groups. Created with BioRender.com.