TL;DR
This paper introduces phyC, a novel method for clustering cancer evolutionary trees. By developing a "tree registration" process that standardizes tree topologies and edge lengths, phyC can effectively group patients based on their cancer's evolutionary patterns, revealing insights into tumor heterogeneity and potential therapeutic responses.
What Problem Does This Paper Address?
Cancer is a complex disease characterized by significant genetic diversity that evolves over time. Understanding this evolution within and between tumors is crucial for personalized medicine, as it influences treatment response and drug resistance. Researchers reconstruct "cancer evolutionary trees" from multi-regional sequencing data to visualize these evolutionary histories. However, a major challenge has been comparing these trees across different patients to identify common evolutionary patterns or patient subgroups. Existing phylogenetic tree comparison methods struggle with the unique characteristics of cancer trees, such as varying numbers of sub-clones and mutations, and non-binary branching structures. This paper addresses the need for a robust method to cluster and interpret these diverse cancer evolutionary trees.
Methods
The core innovation of phyC lies in its "tree registration" method, which transforms individual cancer evolutionary trees into a comparable format. This process involves:
- Reference Tree Encoding: Mapping each observed tree onto a standardized, large, bifurcated "reference tree." This ensures that all trees are represented within a consistent structural framework.
- Normalizing Edge Lengths: Adjusting edge lengths (representing accumulated mutations) to account for variations in the total number of detected mutations across patients. This normalization helps focus on the relative accumulation of mutations rather than absolute counts.
After registration, the trees are represented as vectors in Euclidean space, allowing for standard clustering algorithms. The authors employed hierarchical clustering and used the gap statistic criterion to automatically determine the optimal number of clusters. For interpreting the resulting clusters, phyC provides two visualization tools:
- Multidimensional Scaling (MDS): A dimension-reduction technique to visualize the relationships between registered trees in a 2D space.
- Sub-clonal Diversity Plot: An adaptation of lineage-through-time plots, showing the number of sub-clones versus the fraction of accumulated SSNVs. This plot helps visualize the rate and pattern of sub-clonal expansion.
Datasets
The phyC method was evaluated using:
- Simulated Data: The authors generated simulated cancer evolutionary trees across three scenarios to test the performance of
phyCin classifying trees based on topology, edge lengths, or both. These simulations involved creating datasets with varying degrees of "variance parameter" to assess robustness. - Real-World Datasets:
- Clear Cell Renal Cell Carcinoma (ccRCC) dataset: Consisting of 8 evolutionary trees derived from multi-regional sequencing data.
- Non-Small Cell Lung Cancer (NSCLC) dataset: Consisting of 11 evolutionary trees.
These real-world datasets were used to demonstrate the interpretability of phyC's clustering results in identifying phenotype- or cancer type-related subgroups.
Key Insights and Ideas
- The Importance of "Registration": The paper strongly emphasizes that standard tree comparison methods fail for cancer evolutionary trees due to their inherent complexities. The "tree registration" process is crucial for making these diverse trees comparable, enabling effective clustering.
- Topology and Edge Lengths Both Matter: Simulations showed that
phyC's ability to consider both tree topology and edge lengths (after registration) significantly outperforms methods that focus on only one aspect. - Interpretable Clusters:
phyCsuccessfully identified biologically meaningful clusters in real patient data, distinguishing between drug-sensitive and drug-resistant tumor evolution patterns, and even differentiating between cancer types. - Sub-clonal Diversity Plots as a Diagnostic Tool: The sub-clonal diversity plot provides a novel way to visualize and interpret evolutionary dynamics within clusters, offering insights into how sub-clones acquire mutations and potentially indicating drug response.
Key Results
- Simulation Performance: In simulations,
phyCconsistently outperformed existing methods (like tree edit distance and shortest path distance for topology, and BScore for edge lengths) in accurately clustering trees, especially when considering both topology and edge attributes. The NMI (Normalized Mutual Information) scores forphyCwere notably higher. - ccRCC Dataset:
phyCclustered the 8 ccRCC trees into 3 subgroups. One cluster was interpreted as representing drug-sensitive evolution, correlating with pretreatment with everolimus, while another cluster showed patterns consistent with parallel evolution. - NSCLC Dataset:
phyCclustered the 11 NSCLC trees into 2 subgroups, which were interpreted as reflecting non-recurrent and recurrent cancer groups. The method showed good agreement with previous findings linking sub-clonal SSNV acquisition to recurrence. - Combined ccRCC and NSCLC: When applied to a combined dataset,
phyCwas able to cluster trees based on cancer type (NSCLC vs. ccRCC), demonstrating its utility in comparing evolutionary patterns across different cancer types.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Ignoring SSNV Content: The current method focuses on tree structure and edge lengths (mutation counts) but does not consider the specific types of SSNVs (e.g., driver vs. passenger mutations). Incorporating this information could refine comparisons, especially since driver mutations have a more significant impact on evolution.
- Excluding Other Aberrations: The analysis currently focuses on SSNVs and does not account for other crucial genetic alterations like copy number variations or epigenetic changes, which also contribute to tumor heterogeneity. Extending
phyCto handle multivariate edge attributes could address this. - Sampling Bias: The method does not explicitly account for potential biases introduced by sampling strategies or individual patient variations. Developing methods to reduce sampling bias could improve clustering accuracy and interpretation.
- Driver Gene Identification: The authors suggest that distinguishing between driver and passenger mutations could simplify the algorithm and improve its running time. Future work could explore how to incorporate the effect of driver genes.
Summary updated