Systematic Assessment of Tumor Purity and Its Clinical Implications

Abstract

Purpose: The tumor microenvironment is complex, comprising heterogeneous cellular populations. As molecular profiles are frequently generated using bulk tissue sections, they represent an admixture of multiple cell types (including immune, stromal, and cancer cells) interacting with each other. Therefore, these molecular profiles are confounded by signals emanating from many cell types. Accurate assessment of residual cancer cell fraction is crucial for parameterization and interpretation of genomic analyses, as well as for accurately interpreting the clinical properties of the tumor.

Materials and methods: To benchmark cancer cell fraction estimation methods, 10 estimators were applied to a clinical cohort of 333 patients with prostate cancer. These methods include gold-standard multiobserver pathology estimates, as well as estimates inferred from genome, epigenome, and transcriptome data. In addition, two methods based on genomic and transcriptomic profiles were used to quantify tumor purity in 4,497 tumors across 12 cancer types. Bulk mRNA and microRNA profiles were subject to in silico deconvolution to estimate cancer cell-specific mRNA and microRNA profiles.

Results: We present a systematic comparison of 10 tumor purity estimation methods on a cohort of 333 prostate tumors. We quantify variation among purity estimation methods and demonstrate how this influences interpretation of clinico-genomic analyses. Our data show poor concordance between pathologic and molecular purity estimates, necessitating caution when interpreting molecular results. Limited concordance between DNA- and mRNA-derived purity estimates remained a general pan-cancer phenomenon when tested in an additional 4,497 tumors spanning 12 cancer types.

Conclusion: The choice of tumor purity estimation method may have a profound impact on the interpretation of genomic assays. Taken together, these data highlight the need for improved assessment of tumor purity and quantitation of its influences on the molecular hallmarks of cancers.

FIG 1.

Purity landscape in The Cancer Genome Atlas (TCGA) prostate cancer cohort (PRAD). (A) Distribution of TCGA prostate tumor purity estimates (n = 333) using in silico methods and consolidated multiobserver pathology reviews; (B) Patient-wise purity estimates grouped by Gleason score. Gray represents missing data, including both failed estimates and missing molecular profiles (see Methods for details). Columns were clustered using Ward hierarchical clustering method. Data from INTEGER were available for 107 samples using the low-pass DNA sequencing data; (C) Pearson correlation between purity estimates inferred using in silico methods and pathology reviews. Rows and columns were clustered using Ward hierarchical clustering method.

FIG 2.

Deviation of pathologist-inferred tumor purity from in silico estimates. Difference between pathology estimates of tumor purity and in silico estimates from DNA and mRNA abundance profiles. P, pathology estimates; R, Pearson’s correlation coefficient; P_R, statistical significance of observed correlation.

FIG 3.

Molecular correlates of tumor purity. Genomic correlates of tumor purity as summarized using androgen receptor (AR) signature score (A), percent genome altered ([PGA], B), and mutation burden (C). Correlation statistic was estimated using Pearson correlation. (D) Purity estimates stratified by prostate cancer–specific driver mutations and ERG fusions. log₂FC represents difference in mean purity (log₂ scale) between mutant and wild-type samples (ERG represents ERG fusions). Statistical significance was estimated using Wilcoxon rank sum test, and P values were adjusted for multiple comparisons using the Benjamini–Hochberg method. Statistical tests were performed for genes with more than three mutant samples. Therefore, IDH1, RB1, AKT1, and CHD1 (displayed with “x”) were deemed inappropriate for statistical testing. (E) Correlation between purity estimates and variant allele frequency of mutant samples. Correlation statistic was estimated using Pearson correlation, and P values were adjusted for multiple comparisons using the Benjamini–Hochberg method. For reliable correlation estimates, genes (in panel 3D) with more than 10 mutant samples were considered for estimating correlation with tumor purity. FDR, false discovery rate; miRNA, microRNA.

FIG 4.

Deconvolved prostate cancer profiles, and DNA- and mRNA-derived purity estimates across The Cancer Genome Atlas (TCGA) cancer types. (A) Correlation between purity estimates derived using pathology, DNA, mRNA, and microRNA (miRNA) profiles and molecular profiles (mRNA.naive = bulk mRNA abundance, mRNA.ISOpure = deconvolved mRNA abundance, miRNA.naive = bulk miRNA abundance, miRNA.ISOpure = deconvolved miRNA abundance, and CNA = bulk copy number data; deconvolved RNA profiles were generated using ISOpure). Each feature (genes for mRNA and copy number aberration [CNA] profiles, miRNAs for miRNA profiles) was correlated with tumor purity estimators (pathology, DNA, RNA, miRNA) separately. The x-axis represents number of purity estimators where a feature was found to be significantly correlated (Spearman’s |ρ| > 0.3, false discovery rate–adjusted P < .01). (B) Distribution of tumor purity estimates across 13 TCGA tumor types (4,830 tumors) using an in silico DNA-based (ASCAT) and mRNA-based (ISOpure) method. “Mean” estimate indicates combined mean of purity estimates from ASCAT and ISOpure. “Pearson’s R” indicates correlation between ASCAT and ISOpure estimates. “n” shows total number of samples with valid estimates available for both ASCAT and ISOpure.