A Novel Method for Across-Chromosome Phasing without Relative Data

This paper introduces a novel window-based SNP-similarity method that enables accurate across-chromosome phasing in unrelated individuals without requiring parental or relative data, achieving up to 95% accuracy when within-chromosome phase errors are minimized.

The Gist

Imagine your DNA is like a massive library containing two complete sets of books (one set from your mom, one from your dad). Each "book" is a chromosome.

The Problem: The Mixed-Up Shelves
When scientists look at your DNA, they can see the words (genes) on the pages, but they often can't tell which page belongs to Mom's book and which belongs to Dad's. They know you have a "blue eye" word and a "brown eye" word, but they don't know if those two words are sitting next to each other on Mom's page or if they are split up (one on Mom's, one on Dad's).

• Within-Chromosome Phasing: This is like sorting the pages inside a single book. We know which words go together on Mom's page vs. Dad's page for that specific book. Scientists are already pretty good at this.
• Across-Chromosome Phasing: This is the harder puzzle. It's like trying to figure out if Page 1 of the "Eye Color Book" (Chromosome 1) and Page 1 of the "Hair Color Book" (Chromosome 2) both came from Mom, or if one came from Mom and the other from Dad.

Usually, to solve this, you need to see the parents' books to compare them. But in big studies (like the UK Biobank), we often only have the child's data, not the parents'. Without the parents, it's like trying to sort a mixed-up library without seeing the original owners.

The Old Way: Finding Long Lost Cousins
Previous methods tried to solve this by looking for "Identical by Descent" (IBD) segments. Think of this as looking for long, identical stretches of text that you share with a distant cousin. If you and a cousin share a long, identical paragraph on Chromosome 1 and a long, identical paragraph on Chromosome 2, you can guess those paragraphs came from the same grandparent.

• The Flaw: This requires a huge library (millions of people) or very close relatives to find those long, matching paragraphs. If you don't have close relatives in the dataset, or if the matching paragraphs are too short, this method fails.

The New Method: The "Similarity Score" Detective
The authors (Sapin, Kelly, and Keller) invented a new way to solve this puzzle without needing parents or long cousin matches. They call it a window-based SNP-similarity metric.

Here is the analogy:

Imagine you are trying to figure out which of your two friends (Friend A and Friend B) is more similar to you.

1. The Window: Instead of looking at your whole life story, you break it down into small "windows" or chapters (e.g., "Childhood," "High School," "College").
2. The Comparison: For every single window, you compare your story to the stories of thousands of other people in the room.
   • You ask: "For the 'High School' window, whose story looks most like my 'High School' story?"
   • You do this for every window on every chromosome.
3. The Pattern:
   • If your "High School" story (Chromosome 1) and your "College" story (Chromosome 2) both look most like the same person's stories, it's highly likely those two chapters came from the same parent.
   • If your "High School" story looks like Person X, but your "College" story looks like Person Y, they likely came from different parents.

How the Algorithm Works (The "Magic" Step):
The computer doesn't just look at one window; it looks at the pattern of similarities across the whole genome.

• It calculates a "Similarity Score" for every window against everyone else.
• It then checks the correlation. Do the windows that look like "Mom's side" tend to appear together?
• If Window 1 and Window 50 both have high similarity scores with the same group of people, the algorithm says, "Aha! These two windows are on the same side of the family!"

The Results: How Well Did It Work?
The team tested this on the UK Biobank (a massive database of 500,000 people).

• The Gold Standard: They used a group where they did have the parents' data to check the answer key.
• The Score: When the initial sorting of the books was perfect, their new method got 95% accuracy. Even with some initial sorting errors, it still hit 83% accuracy.
• Comparison: It beat the old "cousin-matching" methods, especially for people who didn't have close relatives in the dataset.

Why This Matters
This is like upgrading from a magnifying glass to a high-tech scanner.

• No Parents Needed: You can now figure out the "Mom vs. Dad" origin of your DNA even if your parents aren't in the database.
• Smaller Datasets: You don't need 10 million people to make it work; 500,000 is enough.
• Better Science: Knowing which genes came from Mom and which from Dad helps scientists understand things like why some diseases only happen if inherited from the mother, or how parents' traits mix to create a child's traits.

In a Nutshell:
The authors built a smart detective that looks for subtle patterns of similarity across your entire genome to guess which chromosome chunks came from Mom and which from Dad, without needing to see Mom or Dad's DNA. It's faster, works on smaller groups of people, and is much more accurate than previous methods.

Technical Summary

1. Problem Statement

Context: Genetic phasing involves separating diploid genotypes into two haploid sets (haplotypes).

• Within-chromosome phasing: Determines which alleles are co-inherited on the same chromosome. This is a mature field with high accuracy (e.g., tools like Beagle, Eagle2, Shapeit2).
• Across-chromosome phasing (ACP): Determines which haplotypes from different chromosomes originate from the same parent (e.g., matching the maternal haplotype of Chromosome 1 with the maternal haplotype of Chromosome 2).

The Challenge:

• ACP is straightforward when parental data is available but remains a significant challenge for samples of unrelated individuals where parental or close relative data is missing.
• Existing methods rely heavily on detecting Identical-by-Descent (IBD) segments shared between individuals.
  • Some methods require "surrogate parents" (relatives linked to one biological parent).
  • Others (e.g., Noto et al., Cole et al.) require large cohorts (millions of individuals) to find sufficient long IBD segments (>5–10 cM) across multiple chromosomes.
• Current IBD-based methods perform poorly in smaller datasets (<500,000 individuals) or populations with low relatedness, often failing to phase chromosomes that lack shared IBD segments.

2. Methodology

The authors propose a novel approach that eliminates the need for explicit IBD segment calling or close relatives by leveraging window-based SNP-similarity metrics and their correlations.

A. Data Preparation

• Dataset: UK Biobank (European ancestry), focusing on 978 trio families (offspring + both parents) for ground-truth validation, and ~435,000 unrelated individuals as the reference panel.
• Pre-processing: Standard within-chromosome phasing (using Shapeit2) was applied. A multi-stage error-correction algorithm was used to minimize switch errors.
• Window Definition: The genome was divided into 78 non-overlapping windows (approx. 44 cM average length) based on recombination hotspots to ensure independence.

B. The $\hat{\psi}$ Metric (Haplotype Similarity)

Instead of detecting long IBD segments, the method calculates a similarity score between the focal individual's haplotypes and all other individuals in the sample within each window.

1. Haploid Similarity: A modified version of the standard SNP similarity metric ($\hat{\pi}$) is used to compare haploid sequences.
2. Exponentiation: The similarity score is raised to the power of $1/5$ to dampen the noise from rare SNPs, then squared (or raised to the 4th power in the final metric) to amplify the signal of shared ancestry.
3. Selection: For a focal individual's haplotype $A$ in window $w$, the algorithm compares it against both haplotypes of every non-focal individual and selects the maximum similarity value. This makes the method robust to phasing errors in the reference panel.

C. The Correlation-Based Algorithm

The core innovation is using the correlation of these similarity vectors across windows to infer phase.

1. Vector Construction: For a focal individual, two vectors are created for each window $w$: $\hat{\psi}^*_{A,w}$ and $\hat{\psi}^*_{B,w}$. These vectors contain the max-similarity scores against all other individuals.
2. Correlation Matrix: The algorithm computes the Pearson correlation between these vectors for pairs of windows ($w_g$ and $w_h$) across the genome.
   • Hypothesis: If window $w_g$ and window $w_h$ are inherited from the same parent (e.g., both maternal), their similarity profiles against the population will be highly correlated.
   • Metric ($\lambda$): A unified measure $\lambda$ is calculated:
     $$\lambda_{w_h, w_g} = r(A_g, A_h) - r(A_g, B_h) - r(B_g, A_h) + r(B_g, B_h)$$
     • A positive $\lambda$ implies $A_g$ and $A_h$ are from the same parent.
     • A negative $\lambda$ implies $A_g$ and $B_h$ are from the same parent.
3. Iterative Clustering: The algorithm iteratively selects window pairs with the strongest absolute $\lambda$ values, merges them, and repeats until all windows across all chromosomes are phased into two complete parental sets.

3. Key Contributions

1. No-Relative Requirement: The method successfully performs across-chromosome phasing without requiring parental data or close relatives in the dataset.
2. Scalability: It is effective in datasets with **<500,000 individuals**, whereas competing IBD-based methods often require >10 million individuals to achieve similar accuracy.
3. Robustness to Phasing Errors: By selecting the maximum similarity between the focal haplotype and either reference haplotype, the method mitigates the impact of switch errors in the reference panel.
4. Population Stratification Utility: The method leverages allele frequency differences (population stratification) as a signal. If a focal individual's parents come from different sub-populations, the similarity profiles naturally cluster by ancestry, aiding phasing even without long IBD segments.

4. Results

The method was validated using 978 trio offspring from the UK Biobank, where the "ground truth" (actual parental transmission) was known.

• Scenario A: Error-Free Within-Chromosome Phasing
  • When the input data had no switch errors (simulated perfect within-chromosome phasing), the method achieved a mean accuracy of 95% and a median of 100%.
  • 53% of individuals were phased perfectly.
• Scenario B: Realistic Data (with Shapeit2 errors)
  • When using standard computationally phased data (containing switch errors), the mean accuracy dropped to 83.1% (median 85.9%).
  • Key Finding: The primary limitation of the method is the accuracy of the initial within-chromosome phasing, not the across-chromosome algorithm itself.
• Comparison with Existing Methods:
  • vs. Noto et al. (2022): The proposed method significantly outperformed the IBD-based Noto method, particularly in individuals without close relatives. The Noto method's accuracy dropped sharply in the absence of sufficient IBD sharing, while the new method remained stable.
  • vs. Cole et al. (2022): The new method achieved a slightly higher median accuracy (85.66%) compared to Cole et al. (83.4%) on the same subset of 998 individuals.
• Generalizability: Validation on an independent Parent-Offspring (P/O) sample showed only a ~1% decrease in accuracy compared to the trio sample, indicating the results were not due to overfitting.

5. Significance and Future Directions

• Enhanced GWAS Power: Accurate across-chromosome phasing allows researchers to infer the "parent-of-origin" of alleles without genotyping parents. This boosts power in Genome-Wide Association Studies (GWAS) by proxy and enables the study of imprinting effects.
• Assortative Mating: It facilitates the analysis of assortative mating patterns by correlating polygenic scores across phased haplotypes within an individual.
• Future Improvements:
  • Integrating within-chromosome continuity constraints (penalizing switch errors within the same chromosome).
  • Incorporating explicit data from distant relatives (e.g., aunts/uncles) to further constrain phase.
  • Recursive application: Using improved across-chromosome phase to refine within-chromosome phasing.

Conclusion: This paper presents a computationally efficient, scalable, and highly accurate method for across-chromosome phasing that overcomes the data density requirements of previous IBD-based approaches, making it applicable to standard biobank-sized datasets.