TL;DR
VICTree is a novel Bayesian framework for reconstructing clonal trees from single-cell sequencing data. It introduces a new Tree-structured Mixture Hidden Markov Model (TSMHMM) and a variational inference approach to efficiently infer tumor evolution, cell-to-clone assignments, and copy number profiles, addressing limitations of existing methods.
What Problem Does This Paper Address?
Cancer evolution is a complex process driven by genetic mutations, leading to a diverse population of cells within a tumor (heterogeneity). Understanding this evolution, particularly the changes in DNA copy numbers (CNs), is crucial for diagnosing and treating cancer. Single-cell sequencing technologies allow us to analyze these changes at the individual cell level, but the resulting data is often noisy and complex. Existing methods for reconstructing clonal trees (representing the evolutionary relationships between cell populations) and inferring their CN profiles face challenges: some rely on heuristic approaches, others make simplifying assumptions that limit their accuracy, and many require extensive pre-processing or are computationally expensive. This paper aims to provide a more robust, automated, and efficient Bayesian framework for this task.
Methods
The core of VICTree is a novel Tree-structured Mixture Hidden Markov Model (TSMHMM). This model combines:
- CopyTree: A tree-structured Markov model that captures dependencies in copy number profiles along the evolutionary tree.
- Mixture Emission Model: Associates observed single-cell data with the hidden states (clones) of the CopyTree, extending existing Mixture Hidden Markov Models (MHMMs).
For inference, VICTree employs Variational Inference (VI), specifically a Coordinate Ascent Variational Inference (CAVI) approach. This is a computationally efficient alternative to traditional Markov Chain Monte Carlo (MCMC) methods, especially for complex models. Key novel algorithmic contributions include:
- LARS (Labeled Arborescence Sampler): A new algorithm for sampling directed labeled rooted trees, used as a proposal distribution for importance sampling during tree inference.
- Split-and-Merge Algorithm: A domain-specific algorithm to combat clone degeneracy (where multiple inferred clones become indistinguishable), improving the practical applicability of VI-based clonal analysis.
Datasets
The authors evaluated VICTree on:
- Simulated Data: Generated from their own model to assess performance and robustness under controlled conditions, including varying numbers of clones and mutation rates.
- Real Datasets:
- Multiple Myeloma (MM) samples (MM29, MM03, MM04).
- A xenograft breast cancer dataset (SA501X3F).
Key Insights and Ideas
- Unified Bayesian Inference: VICTree offers a comprehensive Bayesian framework that jointly infers clonal trees, cell-to-clone assignments, copy number profiles, and cell-specific emission baselines, eliminating the need for separate tools or extensive pre-processing.
- Handling CN Dependencies: The TSMHMM effectively models dependencies within copy number profiles along the evolutionary tree, a crucial aspect for accurate tumor evolution reconstruction.
- Efficient Tree Sampling: LARS provides an efficient way to sample tree topologies, which is essential for exploring the vast space of possible evolutionary histories.
- Addressing Clone Degeneracy: The split-and-merge algorithm is a practical solution to a common problem in mixture modeling, making VICTree more robust in real-world scenarios.
- ELBO for Model Selection: The Evidence Lower Bound (ELBO) serves as a principled metric for selecting the optimal number of clones (K), offering a data-driven approach to model complexity.
Key Results
- High Accuracy: VICTree achieved near-perfect scores on simulated data for clustering, CN calling, and tree reconstruction.
- Robustness: The method demonstrated robustness to high mutation rates, a scenario where other methods struggle.
- Efficiency: VICTree converges in a reasonable time, outperforming traditional MCMC methods. For example, it can infer results for a breast cancer dataset within an hour.
- Competitive Performance: On real datasets, VICTree's results were comparable to or better than state-of-the-art methods like HMMCopy and CONET, often with less pre-processing.
Limitations and Future Work
- Computational Complexity: While VICTree is efficient, the complexity of tree sampling can still become a bottleneck for very large numbers of clones. The authors suggest exploring more efficient sampling schemes for larger K.
- Model Improvements: Future work could involve enhancing the CN evolution model to explicitly capture specific CN events (e.g., whole-genome duplications) and developing an observational model that accounts for ploidy to improve the TSMHMM.
- Ensembles and Multi-modality: The authors suggest exploring ensembles or mixtures of variational distributions to better accommodate multi-modal posteriors.
Summary updated