Arborist: Prioritizing Bulk DNA Inferred Tumor Phylogenies via Low-pass Single-cell DNA Sequencing Data

The paper introduces ARBORIST, a novel method that leverages low-pass single-cell DNA sequencing data to prioritize and refine tumor phylogenies inferred from bulk DNA sequencing, thereby overcoming solution non-uniqueness and improving the accuracy of cancer evolutionary reconstruction.

The Gist

The Big Picture: Solving the "Family Tree" Mystery

Imagine you are trying to figure out the family tree of a large, chaotic family reunion. You have two types of clues:

1. The "Crowd Photo" (Bulk DNA): You have a high-quality photo of the whole crowd. You can see everyone clearly, but everyone is mixed together. You can tell there are different groups of people, but you can't easily tell who is related to whom because they are all standing in a pile.
2. The "Blurry Selfies" (Single-Cell DNA): You also have thousands of individual selfies taken by the guests. These are very low-quality, blurry photos (low-pass sequencing). In many of them, you can't even see the person's face clearly, but you can spot a few unique features, like a red hat or a specific tattoo.

The Problem:

• If you only look at the Crowd Photo, you can guess the family tree, but there are too many possibilities. It's like trying to guess the order of a deck of cards just by looking at the pile; you might get it right, or you might be completely wrong.
• If you only look at the Blurry Selfies, you have too much missing information. The photos are so sparse that you can't build a reliable tree on your own.

The Solution: ARBORIST
The researchers created a tool called ARBORIST (which stands for Prioritizing Bulk DNA Inferred Tumor Phylogenies via Low-pass Single-cell DNA Sequencing Data).

Think of ARBORIST as a super-smart detective who combines the two clues. It doesn't try to build the family tree from scratch. Instead, it takes a list of possible family trees (generated from the Crowd Photo) and uses the Blurry Selfies to vote on which one is the most likely to be true.

---

How ARBORIST Works (Step-by-Step)

1. The "Guessing Game" (The Candidate Set)

First, ARBORIST asks other computer programs to look at the high-quality Crowd Photo (Bulk DNA). These programs generate a "shortlist" of possible family trees.

• Analogy: Imagine a detective asking three different experts to draw a sketch of the suspect. Expert A draws a tall guy with a hat. Expert B draws a short guy with a beard. Expert C draws a tall guy with a beard. Now you have three possible suspects.

2. The "Voting Booth" (The Scoring)

This is where ARBORIST shines. It takes the thousands of Blurry Selfies (Single-Cell DNA) and asks: "If this family tree were true, would these blurry selfies make sense?"

• It uses a mathematical technique called Variational Inference. Think of this as a "probability calculator." It doesn't just say "Yes" or "No." It calculates a score (called an ELBO) for every tree on the shortlist.
• Analogy: Imagine you have a lineup of suspects. You show them a blurry photo of a witness.
  • Suspect A (Tall, Hat): The blurry photo shows a tall shape. Score: High.
  • Suspect B (Short, Beard): The blurry photo shows a tall shape. Score: Low.
  • ARBORIST calculates the score for every single tree in the shortlist and picks the winner.

3. The "Cleanup Crew" (Refining the Clusters)

Sometimes, the initial experts made mistakes. Maybe they thought two different people were twins when they weren't.

• ARBORIST doesn't just pick the tree; it also fixes the labels. It re-examines the data and says, "Actually, this person belongs to Group A, not Group B."
• Analogy: While picking the best family tree, the detective also realizes, "Wait, the guy with the red hat isn't actually related to the guy with the blue hat, even though they look similar." It reorganizes the groups to make them cleaner and more accurate.

---

Why This Matters

1. It's a "Best of Both Worlds" approach.
Previously, scientists had to choose between the high-quality but confusing "Crowd Photo" or the detailed but incomplete "Blurry Selfies." ARBORIST proves you don't have to choose. By using the blurry selfies to check the crowd photo, you get a much clearer picture of the tumor's history.

2. It handles "Noise" better.
Cancer cells are messy. They mutate, lose parts of their DNA, and mix together. The paper shows that ARBORIST is better at ignoring the "static" (noise) in the data than previous methods.

• Analogy: If you are trying to hear a whisper in a noisy room, previous methods might just shout louder (trying to force an answer). ARBORIST is like a noise-canceling headphone that filters out the background chatter to hear the whisper clearly.

3. Real-World Success
The team tested this on a real patient with a rare cancer (a nerve sheath tumor).

• Before ARBORIST: The experts were confused about how different parts of the tumor were related.
• After ARBORIST: It picked one specific family tree that made perfect sense. It even confirmed its choice by checking the "copy number" (like checking if the family members have the same number of chromosomes), which acted as a second witness confirming the story.

The Takeaway

ARBORIST is a bridge. It connects the broad, high-quality view of a tumor with the detailed, individual view of its cells. By using the individual cells to "vote" on the best family tree, it helps doctors and scientists understand exactly how a cancer started, how it grew, and how it might spread. This is a crucial step toward designing better treatments that target the specific history of a patient's cancer.

Technical Summary

1. Problem Statement

Tumor phylogeny reconstruction aims to model the evolutionary history of cancer by ordering somatic mutations (SNVs) into a clone tree. Current approaches face a trade-off between data modalities:

• Bulk DNA Sequencing (bulk DNA-seq): High coverage and cost-effective for detecting SNVs, but requires deconvolution of mixed cell populations. This leads to a non-unique solution space (many possible trees explain the data equally well) and ambiguity in clonal ordering.
• Low-pass Single-cell DNA Sequencing (scDNA-seq): Provides single-cell resolution but suffers from extreme sparsity (0.01×–0.05× coverage). This makes it difficult to detect SNVs directly, as many loci lack read coverage. Existing scDNA-seq methods often rely heavily on copy number alterations (CNAs) rather than SNVs.
• The Gap: While methods exist to integrate bulk and single-cell data (e.g., B-SCITE, PhISCS), they were designed for older, low-throughput technologies (hundreds of cells/SNVs) and cannot scale to modern datasets (thousands of cells, tens of thousands of SNVs). Furthermore, no method currently exists to prioritize a candidate set of trees generated from bulk data using low-pass scDNA-seq data.

Core Challenge: How to select the most accurate clone tree from a candidate set (generated by bulk methods) using sparse, low-pass scDNA-seq data to resolve evolutionary ambiguities.

2. Methodology: ARBORIST

The authors propose ARBORIST, a two-step framework that treats the problem as Clone Tree Selection (CTS).

A. Input and Workflow

1. Candidate Generation: Existing bulk sequencing methods (e.g., CONIPHER, SAPLING) are used to generate an initial SNV clustering ($\psi$) and a candidate set of clone trees ($\mathcal{T}$).
2. Data Integration: ARBORIST takes the candidate trees, the initial clustering, and the raw scDNA-seq data (variant read counts $A$ and total read counts $D$ for $n$ cells and $m$ SNVs).
3. Ranking: ARBORIST evaluates each candidate tree $T \in \mathcal{T}$ against the scDNA-seq data to find the tree $T^*$ that maximizes the posterior probability $P(T | A, D, \psi)$.

B. Generative Model

The method models the data generation process with the following assumptions:

• Infinite Sites Assumption: Each mutation occurs once, though copy number aberrations can cause loss.
• Latent Variables:
  • $z_i$: The clone from which cell $i$ originated (including a "normal" clone).
  • $y_j$: The true cluster assignment for SNV $j$ (allowing for errors in the initial bulk clustering).
• Likelihood: The observed variant reads $a_{i,j}$ follow a Binomial distribution given the total reads $d_{i,j}$ and the effective variant allele frequency (VAF). The VAF is adjusted for sequencing error rates.
• Prior: A prior $\pi$ is defined for the SNV-to-cluster labels to incorporate confidence in the initial bulk clustering.

C. Variational Inference

Direct calculation of the marginal likelihood is computationally intractable due to the exponential number of latent configurations ($(k+1)^n k^m$). ARBORIST uses Mean-Field Variational Inference to approximate the posterior:

• Objective: Maximize the Evidence Lower Bound (ELBO) for each candidate tree.
• Factorization: The posterior is approximated by factorized distributions $q(z)$ (cell-to-clone) and $q(y)$ (SNV-to-cluster).
• Optimization: The ELBO is optimized using Coordinate Ascent Variational Inference (CAVI). The algorithm iteratively updates the variational parameters $\phi$ (for cells) and $\theta$ (for SNVs) until convergence.
• Output:
  • The tree $T^*$ with the highest ELBO (best fit).
  • Maximum A Posteriori (MAP) estimates for cell-to-clone labels ($z$) and SNV-to-cluster labels ($y$).
  • Model Selection: The ELBO naturally penalizes model complexity (via entropy and KL divergence terms), allowing ARBORIST to compare trees of different sizes without ad-hoc regularization.

3. Key Contributions

1. Novel Framework: Introduction of the Clone Tree Selection (CTS) problem and the ARBORIST algorithm, the first method to jointly utilize bulk and low-pass scDNA-seq data for prioritizing tumor phylogenies at scale.
2. Scalability: Unlike previous integration methods, ARBORIST is designed to handle modern high-throughput datasets (thousands of cells, tens of thousands of SNVs).
3. Refinement of Clustering: Beyond selecting a tree, ARBORIST refines the initial SNV clustering ($\psi$) by re-assigning SNVs to clusters based on single-cell evidence, effectively "denoising" the bulk-derived clusters.
4. Principled Integration: Uses variational inference to provide a principled lower bound on the marginal likelihood, balancing goodness-of-fit with model complexity.

4. Results

A. Benchmarking on Simulated Data

• Setup: Simulated tumors with $k=7$ and $k=10$ clones, bulk coverage (50×/100×), and ultra-low scDNA-seq coverage (0.02×/0.05×).
• Comparison: ARBORIST (paired with CONIPHER or SAPLING) was compared against bulk-only methods (CONIPHER, SAPLING, GraphYC) and scDNA-seq-only methods (Phertilizer, SBMClone).
• Performance:
  • Accuracy: ARBORIST significantly outperformed all baselines in Ancestral Pair Recall (APR), Incomparable Pair Recall (IPR), and Clustered Pair Recall (CPR).
  • Robustness: The improvement was most pronounced at the lowest coverage (0.02×), where scDNA-seq-only methods failed to resolve ancestral relationships.
  • Clustering: ARBORIST improved the Adjusted Rand Index (ARI) for SNV clustering compared to the initial bulk-only clustering (PYCLONE-VI).

B. Biological Validation: Malignant Peripheral Nerve Sheath Tumor (MPNST)

• Dataset: Multi-region bulk sequencing (5 samples) and 3,190 single cells from a patient (GEM2.3).
• Process: Generated 100 candidate trees from CONIPHER and SAPLING; ARBORIST selected the top tree.
• Findings:
  • Tree Resolution: ARBORIST resolved evolutionary relationships between SNV clusters that were ambiguous in bulk data.
  • Denoising: It refined the initial DPClust clustering, redistributing SNVs to better-fitting clusters and identifying sample-specific (private) mutations.
  • Orthogonal Validation:
    • VAF Separation: The selected tree showed clear separation between "within-clade" and "outside-clade" VAFs, confirming correct assignment.
    • Copy Number Proxy: Using normalized binned read counts as a proxy for copy number, the ARBORIST-selected tree achieved the lowest Davies-Bouldin Index (DBI), indicating superior cluster coherence compared to other candidate trees.

5. Significance

• Bridging Modalities: ARBORIST provides a practical solution to leverage the high sensitivity of bulk sequencing for SNV detection and the high resolution of scDNA-seq for clonal ordering, overcoming the limitations of sparsity in single-cell data.
• Downstream Impact: By providing a single, high-confidence clone tree and refined SNV clusters, it enables more accurate downstream analyses, such as studying metastatic migration, therapeutic resistance, and copy number evolution.
• Future-Proofing: As low-pass scDNA-seq becomes more accessible and affordable, ARBORIST offers a scalable, open-source framework (Python) that avoids the computational intractability of fully joint probabilistic modeling while maintaining high accuracy.

In summary, ARBORIST transforms the challenge of ambiguous tumor phylogenies by using sparse single-cell data as a "tie-breaker" for bulk-inferred trees, significantly improving the reconstruction of cancer evolutionary histories.