Moduli Spaces of Phylogenetic Trees Describing Tumor Evolutionary Patterns

Sakellarios Zairis, Hossein Khiabanian, Andrew J. Blumberg, Raul RabadanBrain Informatics and Health 2014 · OpenReview

TL;DR

This paper introduces a novel mathematical framework for analyzing and comparing the evolutionary histories of tumors, represented as phylogenetic trees. By applying concepts from geometric topology to these tree spaces, the authors enable statistical inference and visualization of aggregate tumor evolution across patient cohorts, offering new insights into cancer progression and personalized treatment strategies.

What Problem Does This Paper Address?

Cancer is understood to evolve through a process of clonal Darwinian selection, where cells with advantageous genetic alterations outcompete others, leading to tumor progression. High-throughput sequencing technologies now allow us to reconstruct the evolutionary history of a tumor from multiple samples taken over time or from different locations within a patient. These histories can be represented as phylogenetic trees, and studying large collections of these trees (forests) from many patients is crucial for understanding general patterns of cancer evolution, identifying mechanisms of drug resistance, and predicting patient outcomes. However, existing methods lack a robust mathematical framework for statistically comparing and visualizing these complex, high-dimensional tree structures, hindering our ability to extract meaningful insights from large genomic datasets.

Methods

The core of this paper is the application of the geometric theory of phylogenetic tree spaces, specifically the work by Billera, Holmes, and Vogtmann, to unrooted trees representing intra-individual cancer evolution. They define an "evolutionary moduli space" ($\Sigma_m$) for $m$ samples, which is a mathematical space where each point corresponds to a unique phylogenetic tree. This space is endowed with a natural metric that allows for the calculation of distances between trees, enabling meaningful comparisons.

Key methodological contributions include:

  • Defining the Evolutionary Moduli Space ($\Sigma_m$): This space is constructed as a product of a space for internal branches and an orthant for external branches, allowing for the representation of both tree topology and edge lengths (representing evolutionary time or divergence).
  • Establishing CAT(0) Geometry: The paper proves that these evolutionary moduli spaces are CAT(0) spaces. This property is crucial because it guarantees the existence and uniqueness of geodesic paths between any two points and ensures that triangles in this space behave "nicely" (similar to Euclidean triangles), which is fundamental for statistical inference.
  • Introducing Projective Evolutionary Moduli Spaces ($P\Sigma_m$): To facilitate visualization, the authors introduce a projective space by rescaling tree lengths. This allows for the representation of evolutionary modes in lower dimensions, particularly for small numbers of samples (e.g., 3 or 4).
  • Statistical Inference Framework: Leveraging the CAT(0) property, the paper outlines how to define statistical concepts like the Fréchet mean (a generalization of the mean for metric spaces) and variance for populations of phylogenetic trees. This provides a foundation for performing statistical inference on tumor evolutionary patterns.

Datasets

The paper utilizes publicly available genomic data from existing studies on human malignancies. Specifically, they analyze:

  • Acute Myelogenous Leukemia (AML): Triplet data (normal, diagnosis, relapse) from 8 patients.
  • Follicular Lymphoma (FL) to Diffuse Large B-cell Lymphoma (DLBCL) Transformation: Triplet data from multiple patients, analyzed using different data curation strategies (WGS, WES, curated gene panel).
  • Metastatic Pancreatic Ductal Adenocarcinoma (PDAC): Quadruplet data (normal, primary tumor, liver metastasis, non-liver metastasis) from 13 patients.

These datasets are used to demonstrate the application of their framework and visualize aggregate evolutionary behaviors.

Key Insights and Ideas

  • Geometric Structure of Tumor Evolution: The paper reveals that the space of tumor evolutionary histories has a well-defined geometric structure (nonpositive curvature, specifically CAT(0)). This is a significant insight, as it means we can apply powerful mathematical tools from geometry and statistics to understand cancer evolution.
  • Visualizing Evolutionary Modes: The introduction of projective tree spaces ($P\Sigma_m$) provides a powerful way to visualize and categorize different evolutionary patterns (e.g., frozen, branched, linear, divergent evolution) across patient cohorts. This is crucial for gaining intuition about complex biological processes.
  • Distinguishing Evolutionary Trajectories: The framework allows for the quantitative comparison of evolutionary histories, enabling researchers to identify distinct patterns associated with different cancer types, treatment responses, or disease progression stages.
  • Potential for Personalized Medicine: By enabling statistical inference on tumor evolution, this work lays the groundwork for understanding how treatments perturb evolutionary trajectories and for developing personalized treatment strategies based on a patient's specific tumor evolution.

Key Results

The paper demonstrates the application of their framework to real biological data, yielding several key qualitative findings:

  • AML: Patients' evolutionary histories cluster near the "frozen evolution" vertex in the projective space ($P\Sigma_3$), suggesting that relapse in AML may be driven by mechanisms beyond the accumulation of new mutations at the DNA level, possibly epigenetic deregulation.
  • Follicular Lymphoma: The data suggests a highly branched evolutionary process, where the dominant clone in FL is not necessarily the direct ancestor of the dominant clone after transformation to DLBCL.
  • Pancreatic Cancer: Metastases appear to derive linearly from the primary tumor, but liver and non-liver metastases branch from a common ancestor, indicating a complex metastatic spread pattern.

While quantitative results like specific accuracy metrics are not the primary focus, the paper successfully visualizes and characterizes the aggregate evolutionary behavior of these cancer types using their proposed framework.

Limitations and Future Work

  • Computational Complexity: While the framework is powerful, explicitly constructing and analyzing these spaces for a large number of samples ($m \ge 6$) becomes computationally challenging due to the increasing dimensionality of the space.
  • Constructing Biologically Meaningful Distributions: The paper acknowledges that defining biologically realistic distributions on these tree spaces is a difficult problem.
  • Statistical Inference Algorithms: While the CAT(0) property enables statistical inference, developing practical and efficient algorithms for tasks like hypothesis testing and machine learning on these spaces is an active area of ongoing research for the authors.
  • Future Directions: The authors plan to develop statistical and machine learning methods for extracting biological and clinical information from these point cloud data in $\Sigma_m$. They also aim to characterize how drug treatments perturb evolutionary trajectories and to correlate regions of the space with clinical variables. The increasing density of genomic sampling in the future will necessitate creative visualization and analysis techniques.

Summary updated

Abstract

Cancers follow a clonal Darwinian evolution, with fitter subclones replacing more quiescent cells, ultimately giving rise to macroscopic disease. High-throughput genomics provides the opportunity to investigate these processes and determine specific genetic alterations driving disease progression. Genomic sampling of a patient’s cancer provides a molecular history, represented by a phylogenetic tree. Cohorts of patients represent a forest of related phylogenetic structures. To extract clinically relevant information, one must represent and statistically compare these collections of trees. We propose a framework based on an application of the work by Billera, Holmes and Vogtmann on phylogenetic tree spaces to the case of unrooted trees of intra-individual cancer tissue samples. We observe that these tree spaces are globally nonpositively curved, allowing for statistical inference on populations of patient histories. A projective tree space is introduced, permitting visualizations of evolutionary patterns. Published data from four types of human malignancies are explored within our framework.
BibTeX
@inproceedings{zairis2014moduli,
  author = {Sakellarios Zairis and Hossein Khiabanian and Andrew J. Blumberg and Raul Rabadan},
  title = {{Moduli Spaces of Phylogenetic Trees Describing Tumor Evolutionary Patterns}},
  booktitle = {Brain Informatics and Health},
  year = {2014},
  url = {https://openreview.net/forum?id=AIt0NEufSK},
}
Figure