TL;DR
This paper introduces CopyMix, a novel method that jointly clusters single cells and profiles their copy number variations (CNVs) using a Bayesian mixture model and variational inference. By tackling these tasks simultaneously, CopyMix avoids artifacts introduced by sequential methods and provides more accurate clonal copy number profiles, which is crucial for understanding tumor evolution and heterogeneity.
What Problem Does This Paper Address?
Cancer is a complex disease driven by genetic mutations, and understanding tumor heterogeneity is key to developing effective treatments. Single-cell sequencing technologies allow us to examine the genomic landscape of individual cells within a tumor. Copy number variations (CNVs), which are changes in the number of copies of DNA segments, are a significant source of genomic variation and play a role in cancer progression and therapy resistance.
However, analyzing single-cell CNV data presents several challenges:
- Sequential Analysis: Existing methods often perform CNV profiling and cell clustering in separate steps. This sequential approach can introduce artifacts due to ad-hoc pre- and post-processing steps, potentially leading to inaccurate results.
- Clonal Profiling: Identifying copy number profiles that represent entire clonal populations (groups of cells with a common ancestor) has been challenging.
- Parameter Tuning: Some popular CNV profiling tools, like HMMcopy, require extensive manual parameter tuning, which can be labor-intensive and prone to user error.
CopyMix aims to address these limitations by offering a unified, probabilistic framework that jointly infers cell clusters and their underlying clonal CNV profiles without the need for manual parameter tuning.
Methods
CopyMix employs a mixture model where each component represents a distinct cell cluster. Each cluster is characterized by a Hidden Markov Model (HMM) that captures the clonal copy number profile across the genome. The core of CopyMix lies in its use of Variational Inference (VI), a Bayesian approach for approximating complex probability distributions.
Here's a breakdown of the key methodological aspects:
- Joint Inference: Unlike previous methods, CopyMix simultaneously infers cell cluster assignments and their corresponding copy number profiles. This joint approach helps to mitigate artifacts that can arise from sequential processing.
- Mixture Model: The model assumes that cells originate from a finite number of distinct clonal populations, each with a characteristic CNV profile.
- Hidden Markov Models (HMMs): Within each cluster, the copy number profile is modeled as a sequence of latent states (copy number values) that follow a Markov chain. This captures the genomic dependency of copy number states.
- Variational Inference (VI): VI is used to approximate the posterior distributions of the model's latent variables (cluster assignments, copy number states) and parameters. VI offers advantages over methods like Expectation-Maximization (EM) by providing full posterior distributions, which can help prevent overfitting and facilitate principled model selection.
- Two Emission Models: CopyMix offers two options for modeling the observed read counts:
- Poisson Emission: Suitable for raw read count data.
- Gaussian Emission: Used for GC-corrected read ratios, which are often employed in CNV analysis.
- SNV-CopyMix Extension: The authors also propose an extension, SNV-CopyMix, which incorporates single nucleotide variant (SNV) data alongside CNV data to further improve clustering accuracy.
Datasets
The authors evaluated CopyMix using:
- Simulated Data: They generated simulated datasets to assess CopyMix's performance against ground truth, covering various scenarios with different numbers of clusters and copy number patterns.
- Biological Data:
- Leung et al. (2017) Data: Single-cell whole-genome sequencing data from 18 primary colon cancer cells and 18 metastatic cells from a patient (CRC2). This dataset was used to analyze chromosomes 18, 19, 20, and 21.
- Laks and McPherson (2019) Data (DLP technology): Data from three cell lines from ovarian cancer, totaling 891 cells.
Key Insights and Ideas
- Joint Inference is Superior: The most significant insight is that jointly inferring cell clusters and copy number profiles leads to more accurate and artifact-free results compared to sequential approaches. This is a fundamental shift in how single-cell CNV data should be analyzed.
- Clonal CNV Profiling is Achievable: CopyMix successfully provides clonal copy number profiles, offering a more biologically meaningful representation of tumor subpopulations than cell-specific profiles.
- Parameter-Free Approach: By avoiding manual parameter tuning, CopyMix democratizes the analysis of single-cell CNV data, making it more accessible and reproducible.
- Robustness to Data Complexity: CopyMix demonstrates robustness across various simulated scenarios, including different numbers of clusters, sequence lengths, and copy number patterns, highlighting its versatility.
Key Results
- High Clustering Performance: CopyMix achieved near-perfect V-measures (close to 1) on simulated data across various configurations, indicating excellent clustering accuracy.
- Low Copy Number Profiling Errors: The method demonstrated low discordant positions between true and inferred copy number states, signifying high accuracy in CNV profiling.
- Biological Relevance: On the Leung et al. (2017) dataset, CopyMix successfully clustered cells into primary and metastatic groups, aligning with the known origins of these cells.
- Comparison to Existing Methods: CopyMix outperformed existing methods like HMMcopy and Ginkgo in terms of clustering performance and provided more comprehensive clonal CNV profiles. For instance, on the DLP data, CopyMix achieved a V-measure of 98% without needing to remove a large fraction of cells, unlike Ginkgo.
- SNV-CopyMix Improvement: The extension incorporating SNV data showed potential for further enhancing clustering, especially in scenarios with limited CNV signal.
Limitations and Future Work
- Unidentifiability with Ploidy: The paper notes a potential limitation in distinguishing ploidy when genomes are whole-genome duplicated (WGD), which can lead to unidentifiability issues in unsupervised learning.
- Overlapping CNV Patterns: When copy number profiles of different clusters are highly overlapping, CopyMix's performance might be affected, though it still performs well.
- Allele-Based CopyMix: The authors suggest developing an allele-specific copy number model as a future direction to enrich DLP results and potentially improve ploidy estimation.
- Modeling Other Genomic Events: Extending the model to incorporate biallelic deletions and aneuploidy is also proposed as a biologically desirable future task.
- Phylogenetic Inference: CopyMix can be extended for phylogenetic inference, which could further enhance understanding of tumor evolution.
Summary updated