SC-BIG: A Hierarchical Bayesian Model for Bulk-Informed Single Nucleotide Variant Calling in Single Cells

The paper introduces SC-BIG, a hierarchical Bayesian model that leverages bulk sequencing data to jointly estimate cancer cell fractions and improve the accuracy and uncertainty quantification of somatic SNV detection in single cells, outperforming existing methods in the presence of copy-number alterations and clonal admixtures.

The Gist

The Big Picture: Finding a Needle in a Haystack (That's on Fire)

Imagine you are trying to find a specific, rare type of needle in a giant haystack. But there are two problems:

1. The Haystack is messy: The needles are scattered, some are broken, and the hay is constantly moving.
2. Your eyes are blurry: You are looking at the needles through a foggy window.

In the world of cancer research, the "haystack" is a tumor, the "needles" are specific genetic mutations (changes in DNA), and the "foggy window" is Single-Cell DNA Sequencing.

Scientists want to know exactly which cells in a tumor have a specific mutation. This helps them understand how the cancer is evolving. However, looking at just one cell at a time is incredibly noisy. It's like trying to hear a whisper in a hurricane. Sometimes the machine misses the mutation (it's there, but the signal drops out), and sometimes it thinks it hears a mutation that isn't there (a false alarm).

The Old Way: Guessing with a Crude Map

Previously, scientists had two main ways to solve this:

1. The "Naive" Approach: Just look at the data from the single cell and guess. "If I see the mutation, it's there." This is like trying to find the needle by squinting hard. It works sometimes, but often you get it wrong.
2. The "ProSolo" Approach: This was a smarter tool that used a "bulk" sample (a mix of all the cells) as a reference map. It was better, but it had a major flaw: it assumed every cell was a perfect, standard copy. It couldn't handle it if a cell had lost a chromosome or gained extra copies of DNA. It was like trying to use a map of a flat city to navigate a mountain range; the terrain was too weird for the map.

The New Solution: SC-BIG (The Smart Detective)

The authors of this paper created SC-BIG (Single-Cell Bulk-Informed Genotyping). Think of SC-BIG as a super-smart detective who doesn't just look at the crime scene (the single cell) but also has a detailed police report (the bulk sample) and a deep understanding of how the city is built (the biology of cancer).

Here is how SC-BIG works, using a simple analogy:

1. The "Crowd Report" (The Bulk Data)

Imagine you are trying to figure out if a specific rumor (a mutation) is true in a crowd of 1,000 people.

• The Bulk Sample is like taking a microphone to the whole crowd and asking, "Who knows this rumor?" If 80% of the crowd says "Yes," you know the rumor is very common.
• The Problem: You don't know exactly who in the crowd is lying or who is just confused.

2. The "Individual Interrogation" (The Single Cell)

Now, you go to one specific person in the crowd (a single cell) and ask, "Do you know this rumor?"

• The Noise: Because the person is whispering (low data quality), you might not hear them clearly. They might say "Yes" when they mean "No," or stay silent when they actually know the rumor.

3. How SC-BIG Connects the Dots

SC-BIG combines the Crowd Report with the Individual Interrogation using a special mathematical trick called a Hierarchical Bayesian Model.

• Step 1: The "Vibe Check" (Estimating Prevalence): SC-BIG looks at the "Crowd Report" (bulk data) to estimate how common the rumor is. It asks: "Is this rumor in 10% of people? 50%? 90%?" It accounts for the fact that the crowd might be messy (some people have extra copies of the rumor, some have lost them).
• Step 2: The "Probability Score": Instead of giving a simple "Yes" or "No" answer for the individual, SC-BIG gives a confidence score.
  • Old way: "Yes, this cell has the mutation." (Or "No.")
  • SC-BIG way: "There is a 92% chance this cell has the mutation, but I'm 8% unsure because the signal was fuzzy."

Why is SC-BIG Better?

The paper tested SC-BIG against the old methods using computer simulations (creating fake tumors to test the tools). Here is what they found:

1. It Handles "Weird" Cells: Cancer cells often have weird numbers of chromosomes (like having 3 copies of a shoe instead of 2). The old tool (ProSolo) got confused by this and made mistakes. SC-BIG understands that cells can be weird and adjusts its math accordingly.
2. It's Honest About Uncertainty: SC-BIG is great at saying, "I'm not sure." It produces well-calibrated probabilities. If it says "70% chance," it really means 70%. This is crucial for scientists who need to know how much they can trust the data before making life-or-death decisions about treatment.
3. It Wins the Race: In the tests, SC-BIG was much more accurate at finding the "needles" (mutations) than the old methods, especially when the mutations were common in the tumor or when the tumor had complex genetic changes.

The Takeaway

SC-BIG is a new, smarter way to read the genetic code of individual cancer cells.

Instead of just guessing based on noisy data, it uses the "big picture" data from the whole tumor to help interpret the "small picture" data from a single cell. It acknowledges that cancer is messy, that cells are weird, and that sometimes we just can't be 100% sure. By admitting uncertainty and quantifying it, it gives doctors and researchers a much clearer, more reliable map of how cancer evolves.

In short: It turns a blurry, confusing whisper into a clear, trustworthy conversation.

Technical Summary

1. Problem Statement

Single-cell DNA sequencing (scDNA-seq) is a powerful tool for studying tumor heterogeneity and evolution but suffers from significant technical noise, including:

• Sparse coverage: Low read depth per cell.
• Amplification bias: Stochastic deviations in whole-genome amplification (WGA).
• Allelic Dropout (ADO): Failure to amplify one allele, leading to false negatives.

While existing de novo callers (e.g., Monovar, SCcaller) address these issues, they often struggle with accuracy. A promising strategy is to use bulk whole-genome sequencing (WGS) data from the same tumor sample to inform single-cell variant calling. However, existing bulk-informed methods, specifically ProSolo, have critical limitations:

1. Diploid Assumption: ProSolo assumes all sites are diploid ($\theta_s \in \{0, 0.5, 1\}$), failing to account for somatic copy-number alterations (CNAs) common in cancer.
2. Simplified Genotype Mixture: It assumes bulk variants arise from a mixture of at most two genotypes differing by a single mutational step, which fails to model complex subclonal structures or copy-number states accurately.
3. Miscalibration: These assumptions lead to poorly calibrated posterior probabilities, especially for subclonal variants in regions with CNAs.

2. Methodology: SC-BIG

The authors propose SC-BIG, a hierarchical Bayesian model that leverages bulk sequencing data to improve SNV detection in single cells while explicitly modeling copy number, purity, and clonality.

Core Workflow

The inference proceeds in three steps:

1. Empirical Prior Construction: An empirical prior on the Cancer Cell Fraction (CCF) is constructed from the aggregate fraction of single cells showing variant reads.
2. CCF Posterior Estimation (Bulk Inference): Using Markov Chain Monte Carlo (MCMC), the model estimates the posterior distribution of the CCF ($P(\text{CCF}|D_b)$) by marginalizing over:
   • Bulk tumor purity ($\pi$).
   • Local copy number ($C$).
   • Variant multiplicity ($m$).
   • Bulk read counts ($k_b, n_b$).
3. Per-Cell Variant Probability: The CCF posterior acts as a prior for the second step, calculating the probability that a specific mutation is present in a single cell ($z_i \in \{0, 1\}$) given single-cell read counts ($k_{sc}, n_{sc}$) and the error model.

Key Mathematical Formulations

• Variant Presence Probability:
  $$P(z=1 | k_{sc}, n_{sc}, D_b) = \int_0^1 P(z=1 | k_{sc}, n_{sc}, \text{CCF}) \cdot P(\text{CCF}|D_b) \, d\text{CCF}$$
  The model assumes $P(z=1|\text{CCF}) = \text{CCF}$, meaning a malignant cell harbors the mutation with probability equal to the variant's cellular prevalence.

• Error Models: SC-BIG supports two error models for single-cell data:

  1. Simple Beta-Binomial: A technology-agnostic model using germline heterozygous SNPs (hSNPs) to estimate amplification bias parameters ($\alpha, \beta$). This preserves the expected mean Variant Allele Frequency (VAF) while capturing overdispersion.
  2. Lodato Model (MDA-specific): A coverage-dependent beta-binomial mixture model (used in ProSolo) that accounts for amplification artifacts. SC-BIG adapts this by mapping true allele frequencies ($m/C$) to the nearest supported diploid state for the likelihood calculation, ensuring consistency with ProSolo for fair comparison.

• Priors:

  • Multiplicity ($m$): Uses a geometric prior ($P(m|C) \propto 2^{-m}$) favoring heterozygous variants.
  • Copy Number ($C$) & Purity ($\pi$): Modeled as normal distributions centered on user-provided estimates.

3. Key Contributions

1. Relaxation of Diploid Assumption: Unlike ProSolo, SC-BIG supports arbitrary copy number states ($C \in \{1, \dots, 10\}$) and multiplicities ($m \le C$), making it robust to CNAs.
2. Explicit Subclonal Modeling: By marginalizing over CCF as a continuous latent variable, the model avoids forcing variants into discrete clonal/subclonal categories, better capturing complex tumor architectures.
3. Well-Calibrated Uncertainty: SC-BIG produces posterior probabilities that are well-calibrated, meaning a probability $p$ accurately reflects the true frequency of cells harboring the variant. This is crucial for downstream analyses like phylogenetic reconstruction.
4. Open-Source Implementation: The tool is implemented in Python (using Pyro/PyTorch) and is available as an open-source package with CLI support for simulation, inference, and comparison.

4. Results

The authors evaluated SC-BIG against ProSolo and a Naive VAF thresholding method using simulated datasets with varying CCFs, copy numbers, and tumor purities.

• Overall Performance:

  • ROC-AUC: SC-BIG achieved 0.942, outperforming ProSolo (0.877) and Naive VAF (0.869).
  • Calibration: SC-BIG's posterior probabilities closely tracked the ideal diagonal ($R^2 = 0.96$), whereas ProSolo showed significant attenuation and scatter ($R^2 = 0.77$).

• Stratified by CCF:

  • Low CCF (0–0.3): All methods performed similarly, with SC-BIG showing perfect calibration.
  • Medium/High CCF (0.3–1.0): SC-BIG significantly outperformed ProSolo. ProSolo's performance collapsed in these ranges (ROC-AUC dropped to ~0.70 for high CCF), likely because its restricted genotype mixture model incorrectly interpreted strong bulk signals as near-clonal presence, misclassifying non-carrier cells as ADO events.
  • Impure Tumors ($\pi=0.5$): SC-BIG still outperformed ProSolo, though the performance gap narrowed, supporting the hypothesis that ProSolo's errors are driven by strong bulk signals in pure samples.

5. Significance and Limitations

Significance:

• Improved Accuracy: SC-BIG provides a more accurate method for calling SNVs in single cells, particularly in the presence of copy-number alterations and subclonal heterogeneity.
• Downstream Utility: The well-calibrated posterior probabilities allow for direct integration into probabilistic downstream tools (e.g., CellPhy for phylogenetic tree reconstruction), enabling more accurate modeling of tumor evolution without relying on hard binary calls.
• Robustness: By marginalizing over uncertainty in copy number and purity, the model avoids the pitfalls of fixed assumptions that plague current state-of-the-art tools.

Limitations:

• Simulation Bias: The comparison used priors set to the "true" values of the simulation, which may not reflect the uncertainty present in real-world data estimates.
• Circularity: The empirical CCF prior is derived from the same single-cell data being analyzed, introducing a degree of circularity (though the authors argue this is minimal).
• Error Model Constraints: The adaptation of the Lodato model for non-diploid states involves discretizing allele frequencies, which loses information. The alternative beta-binomial model assumes a 1:1 allelic ratio for hSNPs, which may not hold in regions with CNAs.
• Lack of Real Data Validation: The study relies on simulated data; validation on empirical datasets is required to assess performance against real-world systematic biases.

In conclusion, SC-BIG represents a significant advancement in single-cell genomics by integrating bulk data through a flexible hierarchical Bayesian framework that explicitly accounts for the complexities of cancer genomes, offering superior accuracy and interpretability over existing methods.