Integrative modeling of read depth and B-allele frequency improves single-cell copy number calling from targeted DNA sequencing panels

This paper introduces scPloidyR, a hidden Markov model that jointly analyzes sequencing read depth and B-allele frequency from targeted single-cell DNA panels to significantly improve the accuracy of copy number variation detection compared to depth-only methods, provided that allelic information is available.

The Gist

The Big Picture: Solving the "Copy Number" Puzzle

Imagine your DNA is a massive library of books. Sometimes, cancer cells get confused and start photocopying certain pages (gains) or tossing them out (losses). These changes are called Copy Number Variations (CNVs).

To understand how a tumor grows and resists treatment, scientists need to look at these changes in individual cells, not just a blurry average of the whole tumor. However, reading the DNA of a single cell is like trying to read a book in a dark room with a flickering flashlight. It's noisy and hard to see.

The Problem: Two Signals, One Confusion

The Mission Bio Tapestri machine is a high-tech flashlight that shines on specific pages of the DNA library. It gives scientists two types of clues:

1. Read Depth (The "Volume" Clue): How much ink is on the page? If there's extra ink, maybe the cell copied the page. If there's less ink, maybe it threw the page away.
2. B-Allele Frequency (The "Version" Clue): Most people have two versions of a gene (one from Mom, one from Dad). If a cell has a "Copy Number" change, the ratio of Mom's version to Dad's version changes in a specific way.

The Old Way (karyotapR):
Existing tools mostly rely on the "Volume" clue (Read Depth). It's like trying to guess how many people are in a room just by how loud the noise is. It works okay, but if the room is noisy or if the volume doesn't change much (like when a cell swaps one version for another without changing the total amount), the tool gets confused and misses important details.

The New Solution: scPloidyR

The authors created a new tool called scPloidyR. Think of this tool as a detective who uses both clues at once.

Instead of just listening to the volume, scPloidyR also checks the "Version" ratio. It uses a mathematical framework called a Hidden Markov Model (HMM).

• The Analogy: Imagine walking through a hallway of rooms (chromosomes). Some rooms have 1 chair, some have 2, some have 3.
  • The old tool just guesses the number of chairs based on how much space they take up.
  • scPloidyR looks at the space and the color of the chairs. It also knows that rooms usually have similar numbers of chairs next to each other (spatial coherence). If one room suddenly looks weird, but the neighbors are normal, it knows it's probably just a glitch, not a real change.

The Experiments: Testing the Detective

The researchers tested their new detective against the old one using two methods:

1. The Simulation (The "Fake Crime Scene"):
They created thousands of fake cancer cells with known secrets. They tested the tools under different conditions:

• When the "Version" clue was clear: scPloidyR was a superstar. It found the hidden changes almost perfectly, while the old tool missed many of them.
  • Key Finding: Even adding just one tiny "Version" clue per page made the new tool jump from 55% accuracy to 90% accuracy.
• When the "Version" clue was noisy or missing: If the data was too messy (high noise) or if there were no "Version" clues at all (no genetic differences to measure), the old tool actually performed better. The new tool got confused by the noise.

2. The Real Data (The "Real Crime Scene"):
They applied both tools to a real mix of five different cancer cell lines.

• Both tools found the big changes.
• However, scPloidyR produced a much cleaner, more logical map. It didn't jump back and forth between "normal" and "abnormal" randomly; it created smooth, consistent patterns that made biological sense.

The Takeaway: When to Use Which Tool?

The paper concludes with a simple rule of thumb for scientists:

• Use the New Tool (scPloidyR) if your data has clear "Version" clues (heterozygous variants). It is much better at finding the tricky, hidden changes that the old tool misses. It's like having a detective with a magnifying glass and a fingerprint kit.
• Use the Old Tool (karyotapR) if your data is messy or lacks "Version" clues. In those cases, the simpler tool that just looks at "Volume" is safer and more reliable.

In short: By combining two different types of DNA signals, the new method gives us a much sharper, more accurate picture of how cancer evolves in individual cells, provided the data is clean enough to support it.

Technical Summary

1. Problem Statement

Copy number variations (CNVs) are critical drivers of cancer initiation and progression. While single-cell DNA sequencing (scDNA-seq) allows for the resolution of intra-tumor heterogeneity, accurately calling CNVs from targeted panels (like the Mission Bio Tapestri platform) remains challenging.

• Data Characteristics: Tapestri generates two complementary signals: Sequencing Depth (total DNA content) and B-allele Frequency (BAF) (allelic information from heterozygous SNPs).
• Limitations of Current Methods: Existing tools, such as karyotapR, primarily rely on read depth using Gaussian Mixture Models (GMM). These methods often ignore BAF information, leading to:
  • Inability to detect allele-specific events (e.g., copy-neutral loss of heterozygosity).
  • Failure to distinguish between different genotypes that share the same total read depth (e.g., different trisomy genotypes).
  • Lack of spatial coherence, as they often treat amplicons independently without leveraging the linear ordering of loci along chromosomes.

2. Methodology: scPloidyR

The authors introduce scPloidyR, an R package implementing a Hidden Markov Model (HMM) designed specifically for targeted single-cell data.

• Model Architecture:
  • State Space: Hidden variables represent copy number states ($k \in \{1, 2, 3, 4, 5\}$) for each amplicon.
  • Transitions: A persistence matrix governs state transitions between adjacent amplicons, enforcing spatial coherence (rare breakpoints).
  • Emission Probabilities: The model jointly models two observation types:
    1. Read Depth: Modeled as a Gaussian distribution where the mean is proportional to the copy number state ($\mu_k = k/2$).
    2. BAF: Modeled by marginalizing over possible genotypes compatible with the copy number state. The likelihood accounts for the number of minor alleles and the heterozygosity rate, using a truncated normal distribution for observed minor allele frequencies.
• Parameter Estimation:
  • Parameters (depth variance, BAF noise, heterozygosity rates) are learned from reference cells.
  • The Baum-Welch algorithm (Expectation-Maximization) is used to estimate model parameters.
  • Viterbi decoding is applied to determine the most likely sequence of copy number states for each chromosome.
• Implementation Details:
  • Independent HMMs are fitted per chromosome.
  • Amplicon-level calls are aggregated to segments using the mode.
  • Regularization is applied when heterozygosity is low to prevent depth means from collapsing.

3. Key Contributions

1. Novel Algorithm: Development of scPloidyR, the first HMM-based framework to jointly model depth and BAF specifically for sparse, targeted single-cell DNA panels.
2. Systematic Benchmarking: A comprehensive evaluation comparing scPloidyR against the established depth-only method (karyotapR) across two simulation studies and a real-world dataset.
3. Parameter Sensitivity Analysis: Identification of the specific conditions (BAF noise, variant density, heterozygosity) under which joint modeling outperforms depth-only approaches.
4. Practical Guidelines: Establishing clear criteria for when researchers should use joint modeling versus depth-only methods based on data quality.

4. Results

Simulation Study 1 (Multi-state Heterogeneity)

• Performance: scPloidyR significantly outperformed karyotapR on class-balanced metrics.
  • Macro-F1: 0.472 (scPloidyR) vs. 0.264 (karyotapR).
  • Alteration F1: 0.902 (scPloidyR) vs. 0.383 (karyotapR).
  • Sensitivity: scPloidyR achieved perfect sensitivity (1.0) for single-copy deletions (CN=1), whereas karyotapR was only 0.175.
• Conclusion: Joint modeling drastically improves the detection of non-diploid states, particularly deletions.

Simulation Study 2 (Parameter Sensitivity)

The study varied five parameters to determine the "regime of effectiveness":

• Variant Density: This was the most critical factor. Adding just one heterozygous variant per amplicon increased scPloidyR's accuracy for gains from 0.548 to 0.899. Without variants (depth-only), scPloidyR underperformed karyotapR.
• BAF Noise: scPloidyR performance degraded sharply as BAF noise increased (standard deviation > 9), while karyotapR remained stable.
• Heterozygosity Rate: scPloidyR performance improved non-linearly with heterozygosity, reaching near-perfect accuracy at 100%. At 0% heterozygosity, karyotapR was superior.
• Amplicon Density & Sample Size: Both methods improved with higher amplicon density, but sample size had minimal impact on either method.
• Deletions vs. Gains: Deletions were consistently easier to detect than gains for both methods.

Real Data Application (Five-Cell-Line Mixture)

• Applied to a public Tapestri dataset containing five cell lines.
• Concordance: Both methods detected major large-scale karyotypic structures.
• Divergence: scPloidyR produced more spatially coherent profiles.
  • On Chromosome 19, scPloidyR identified scattered gains driven by BAF signals, whereas karyotapR called them diploid.
  • On the X chromosome, scPloidyR provided more uniform calls within cell populations, suggesting better handling of allelic signals.

5. Significance and Conclusion

• Advantage of Joint Modeling: When allelic information (heterozygous variants) is available and of reasonable quality, integrating BAF with read depth via an HMM provides a clear and substantial advantage over depth-only methods. It resolves ambiguous states and improves the detection of allele-specific events.
• Critical Caveat: The method is highly dependent on the quality and quantity of BAF data. If variant density is zero or BAF noise is high, the joint model fails to outperform simpler depth-only approaches (like karyotapR).
• Practical Impact: The study provides researchers with a new tool (scPloidyR) and a decision framework:
  • Use scPloidyR when the targeted panel contains sufficient heterozygous variants (e.g., $\ge$ 1 per amplicon) and BAF noise is moderate.
  • Use depth-only methods when allelic information is sparse or unreliable.

This work establishes that leveraging the full information content of targeted single-cell sequencing panels requires moving beyond simple depth normalization to integrative statistical modeling.