PhylEx: Accurate reconstruction of clonal structure via integrated analysis of bulk DNA-seq and single cell RNA-seq data

Seong-Hwan Jun, Hosein Toosi, Jeff E. Mold, Camilla Engblom, Xinsong Chen, Ciara H. O’Flanagan, Michael Hagemann-Jensen, Rickard Sandberg, Samuel Aparício, Johan Hartman, Andrew Roth, Jens LagergrenbioRxiv 2021 · DOI

TL;DR

This paper introduces PhylEx, a novel computational method that integrates bulk DNA sequencing and single-cell RNA sequencing (scRNA-seq) data to reconstruct the evolutionary history (clonal tree) of cancer. PhylEx not only accurately identifies cancer cell clones and their genetic makeup but also assigns gene expression profiles to these clones, enabling a deeper understanding of cancer's "phylo-phenotypic" landscape.

What Problem Does This Paper Address?

Cancer is a complex evolutionary process where cells accumulate mutations and diversify into distinct clonal populations. Understanding this evolutionary tree (clonal tree) is crucial for diagnosing and treating cancer. Traditional methods relied on bulk DNA sequencing, which can be ambiguous due to the mixture of cells. While single-cell DNA sequencing (scDNA-seq) improved resolution, it still faced challenges. The advent of scRNA-seq offered a way to study gene expression within individual cells, but integrating this with DNA-based clonal inference was difficult. Existing methods often treated these as separate steps, leading to inefficiencies. This paper addresses the need for a unified approach that leverages both bulk DNA and scRNA-seq data to simultaneously infer clonal structure, genotypes, and functional (expression) profiles, providing a more comprehensive view of cancer evolution.

Methods

PhylEx employs a Bayesian statistical framework. Its core innovation lies in integrating bulk DNA and scRNA-seq data within a single probabilistic model.

  • Probabilistic Model: PhylEx models raw read counts from scRNA-seq using a mixture of Beta-Binomial distributions. This is key to handling the inherent sparsity and "burstiness" of single-cell gene expression, particularly accounting for cases where a gene is present but not expressed (monoallelic expression or dropout).
  • Tree Inference: It uses a tree-structured stick-breaking process (TSSB) as a prior for inferring the clonal tree, a Bayesian non-parametric approach that learns the tree structure directly from the data.
  • SNV Assignment: SNVs are assigned to clones based on a Polya's urn scheme, encouraging SNVs to cluster within clones.
  • Cell-to-Clone Assignment: PhylEx infers cell-to-clone assignments and clonal genotypes.
  • Integration: Crucially, PhylEx integrates information from both bulk DNA (for variant allele frequencies and copy number information) and scRNA-seq (for gene expression patterns) to improve the accuracy of clonal tree reconstruction and cell assignment. It bypasses the need for explicit variant calling in scRNA-seq data, which can be error-prone.

Datasets

The paper evaluates PhylEx on:

  • Simulated Datasets: Used to benchmark performance against existing methods under various evolutionary scenarios, including those with copy number variations and different tree topologies (binary and multifurcating).
  • Biological Datasets:
    • A well-characterized high-grade serous ovarian cancer (HGSOC) cell line (OV2295 and OV2295R2) with previously determined clonal structure.
  • Multi-region HER2+ breast cancer tumor data.

Key Insights and Ideas

  • Synergy of Data Types: The most significant insight is that integrating scRNA-seq data with bulk DNA-seq data dramatically improves the accuracy of clonal tree reconstruction, especially in the presence of copy number variations, which often confound bulk DNA-only methods.
  • Beyond DNA: Functional Phenotypes: PhylEx moves beyond just reconstructing the genetic lineage. By assigning expression profiles to clones, it enables "phylo-phenotypic" analysis, revealing how evolutionary history correlates with cellular function. This offers a richer interpretation than clustering based on expression alone.
  • Robustness to Sparsity: The Beta-Binomial mixture model for scRNA-seq data is a clever way to handle the inherent noise and missing data in single-cell expression, making the method more robust.
  • Single-Region Advantage: PhylEx demonstrates that combining single-region bulk DNA-seq with scRNA-seq can achieve comparable or superior results to multi-region bulk DNA-seq alone, which is a significant practical advantage as multi-region sampling is often difficult or impossible.

Key Results

  • Superior Clonal Tree Reconstruction: PhylEx consistently outperformed state-of-the-art methods (Canopy, TSSB, B-SCITE, ddClone) on simulated datasets in terms of both tree reconstruction error and V-measure clustering metrics, often by a wide margin (e.g., V-measure of 0.866 ± 0.0018 for PhylEx vs. 0.494 for Canopy on one metric).
  • Improved Mutation Clustering: PhylEx also showed superior performance in correctly clustering mutations to their respective clones.
  • Real-World Validation: On the HGSOC cell line, PhylEx accurately reconstructed the known clonal tree and demonstrated distinct gene expression profiles for different clones, supporting its phylo-phenotypic analysis capabilities.
  • Breast Cancer Insights: Analysis of HER2+ breast cancer data revealed clone-specific mutations and pathways, highlighting the method's utility in identifying driver mutations and understanding spatial heterogeneity.

Limitations and Future Work

  • Cell-to-Clone Assignment Correctness: While PhylEx's co-clustering of SNVs and cells suggests correctness, the paper acknowledges that direct evaluation of cell-to-clone assignment accuracy is challenging without ground truth.
  • Computational Cost: While not explicitly detailed as a limitation, Bayesian inference methods can be computationally intensive, especially with large datasets.
  • Future Directions: The authors suggest extending PhylEx to incorporate subclonal copy number information, which is currently difficult to infer from bulk sequencing alone. They also envision its use in characterizing clones by both somatic mutations and copy number profiles, further enhancing the reconstruction of the full evolutionary trajectory of cancer.

Summary updated

Abstract

Abstract We propose PhylEx: a clonal-tree reconstruction method that integrates bulk genomics and single-cell transcriptomics data. In addition to the clonal-tree, PhylEx also assigns single-cells to clones, which effectively produce clonal expression profiles, and generates clonal genotypes. By analyzing scRNA-seq integrated with bulk DNA-seq, PhylEx can take advantage of co-occurrences of the mutations found in the cells. In the probabilistic model underlying PhylEx, the raw read counts from scRNA-seq follow a mixture of Beta-Binomial distributions, which accounts for the sparse nature of single-cell gene expression data; the mixture lessens the penalty caused by mutations not observed due to mono-allelic expression. We rigorously evaluated PhylEx on simulated datasets as well as a biological dataset consisting of a previously well-characterized high-grade serous ovarian cancer (HGSOC) cell line. PhylEx outperformed the state-of-the-art methods by a wide margin both when comparing capacity for clonal-tree reconstruction and capacity for correctly clustering mutations. By analyzing HGSOC and HER2+ breast cancer data, we also show that PhylEx clears the way for phylo-phenotypic analysis of cancer, i.e., that the clonal expression profiles, induced by the cell-to-clone assignments, can be exploited in a manner beyond what is possible with only expression-based clustering.
BibTeX
@misc{jun2021phylex,
  author = {Seong-Hwan Jun and Hosein Toosi and Jeff E. Mold and Camilla Engblom and Xinsong Chen and Ciara H. O’Flanagan and Michael Hagemann-Jensen and Rickard Sandberg and Samuel Aparício and Johan Hartman and Andrew Roth and Jens Lagergren},
  title = {{PhylEx: Accurate reconstruction of clonal structure via integrated analysis of bulk DNA-seq and single cell RNA-seq data}},
  year = {2021},
  doi = {10.1101/2021.02.16.431009},
  howpublished = {bioRxiv},
}
Figure