A reference profile-free deconvolution method to infer cancer cell-intrinsic subtypes and tumor-type-specific stromal profiles

Abstract

Background: Patient stratification based on molecular subtypes is an important strategy for cancer precision medicine. Deriving clinically informative cancer molecular subtypes from transcriptomic data generated on whole tumor tissue samples is a non-trivial task, especially given the various non-cancer cellular elements intertwined with cancer cells in the tumor microenvironment. Methods: We developed a computational deconvolution method, DeClust, that stratifies patients into subtypes based on cancer cell-intrinsic signals identified by distinguishing cancer-type-specific signals from non-cancer signals in bulk tumor transcriptomic data. DeClust differs from most existing methods by directly incorporating molecular subtyping of solid tumors into the deconvolution process and outputting molecular subtype-specific tumor reference profiles for the cohort rather than individual tumor profiles. In addition, DeClust does not require reference expression profiles or signature matrices as inputs and estimates cancer-type-specific microenvironment signals from bulk tumor transcriptomic data. Results: DeClust was evaluated on both simulated data and 13 solid tumor datasets from The Cancer Genome Atlas (TCGA). DeClust performed among the best, relative to existing methods, for estimation of cellular composition. Compared to molecular subtypes reported by TCGA or other similar approaches, the subtypes generated by DeClust had higher correlations with cancer-intrinsic genomic alterations (e.g., somatic mutations and copy number variations) and lower correlations with tumor purity. While DeClust-identified subtypes were not more significantly associated with survival in general, DeClust identified a poor prognosis subtype of clear cell renal cancer, papillary renal cancer, and lung adenocarcinoma, all of which were characterized by CDKN2A deletions. As a reference profile-free deconvolution method, the tumor-type-specific stromal profiles and cancer cell-intrinsic subtypes generated by DeClust were supported by single-cell RNA sequencing data. Conclusions: DeClust is a useful tool for cancer cell-intrinsic molecular subtyping of solid tumors. DeClust subtypes, together with the tumor-type-specific stromal profiles generated by this pan-cancer study, may lead to mechanistic and clinical insights across multiple tumor types.

Systematic Assessment of Tumor Purity and Its Clinical Implications

“…Furthermore, previous studies have reported varying levels of concordance in purity estimates inferred from DNA- and RNA-based methods. 2 , 23 For instance, Aran et al 2 show much stronger concordance between ESTIMATE 24 (RNA-based purity estimator) and ABSOLUTE 13 (DNA-based purity estimator) compared with the RNA- and DNA-based methods in our study. This has significant implications because many genomic algorithms require tumor purity as an input parameter, and selection of the right algorithm for the right tumor type remains challenging.…”

“… 2 , 7 , 15 Because bulk tumor profiles are heterogeneous compositions of tumor cells and TACs featuring complex interplay, it is crucial to interpret the clinico-genomic profiles in the context of the underlying heterogeneity. 25 Many in silico deconvolution techniques have been developed to estimate relative abundance of different cell types, 24 , 26 , 27 as well as techniques that explicitly generate residual transcriptomic 11 , 12 , 15 , 18 , 23 and genomic 14 profiles of tumor-only and stromal-only cells. Use of these residual profiles has generated optimism 4 , 18 , 23 ; however, their applicability in routine bioinformatics analyses remains less popular.…”

“… 25 Many in silico deconvolution techniques have been developed to estimate relative abundance of different cell types, 24 , 26 , 27 as well as techniques that explicitly generate residual transcriptomic 11 , 12 , 15 , 18 , 23 and genomic 14 profiles of tumor-only and stromal-only cells. Use of these residual profiles has generated optimism 4 , 18 , 23 ; however, their applicability in routine bioinformatics analyses remains less popular. Herein, we recommend researchers to consider deconvolution of bulk profiles into individual component profiles (e.g., cancer and stromal profiles) to improve sensitivity and specificity of downstream analyses.…”

BLADE: Bayesian Log-normAl DEconvolution for enhanced in silico microdissection of bulk gene expression data

“…This prior knowledge significantly reduces the search space of solutions for both tasks, which leads to enhanced accuracy and coverage, especially for gene expression purification. The efficient variational inference of BLADE allowed to handle a large number of cell types (> 20 cell types) which was not possible by previous statistical approaches 17,18 . Furthermore, BLADE may be beneficial in handling cell types without a precise prior knowledge, for instance, cancer cells with highly variable gene expression profiles across the subject, unlike the nonmalignant cells 24 .…”

“…Currently, there are few probabilistic deconvolution approaches that take skewed variability into account, but these methods handle only a restricted number of cell types due to difficulties in optimization (e.g. three cell types in DeClust 17 and Demix/DemixT 18 ). Recently, CIBERSORTx introduced a two-step approach to address variable gene expression profiles across the samples: first estimate cellular fraction (deconvolution) and then reconstruct gene expression per cell type in each sample (purification).…”