Using Variable Window Sizes for Phylogenomic Analyses of Whole Genome Alignments

This study introduces a splitting-and-merging approach that extends previous fixed-window methods to allow for variable window sizes across whole genome alignments, demonstrating improved accuracy in recovering gene tree topologies on simulated data and providing nuanced insights into recombination-driven variation in *Heliconius* butterflies and great apes.

The Gist

Imagine you are trying to understand the history of a massive, ancient library (the genome) by reading its books. But here's the catch: this library isn't written by a single author. It's a patchwork quilt stitched together by many different storytellers over millions of years. Sometimes, the stories change abruptly because of "recombination"—a biological process where DNA from different ancestors gets shuffled, like shuffling two decks of cards together.

To figure out the true family tree of the species in this library, scientists need to read the books in chunks. This is where the paper comes in.

The Old Way: The "One-Size-Fits-All" Ruler

In the past, scientists tried to solve this by cutting the library's books into fixed-size chunks (like cutting a loaf of bread into slices that are all exactly 2 inches thick). They would then analyze each slice to see what story it told.

The problem? Nature doesn't follow a ruler.

• Sometimes, a chunk of DNA is a long, unbroken story that should be read as one big piece.
• Other times, the story changes every few pages.

If you use a fixed 2-inch slice:

• Too big: You might cut through a story change, mixing two different histories together. It's like trying to read a mystery novel and a cookbook glued together; the result is confusing nonsense.
• Too small: You might cut a single story into tiny, meaningless fragments that are too short to make sense of.

The New Idea: The "Smart Scissors"

The authors of this paper, Jeremias Ivan and Robert Lanfear, proposed a new method. Instead of using a rigid ruler, they invented "Smart Scissors" that can cut and paste the DNA based on what the story actually looks like.

They call this a "Splitting-and-Merging" strategy. Here is how it works, using a simple analogy:

1. The Starting Point: Imagine you have a whole loaf of bread (the whole chromosome).
2. The Splitting (Cutting): The algorithm looks at a slice. If the story inside that slice seems to have two different endings (meaning the DNA history changed in the middle), the "Smart Scissors" cut it in half. It checks if the two smaller pieces tell a clearer story than the big piece.
3. The Merging (Gluing): Conversely, if two neighboring slices are telling the exact same story, the algorithm glues them back together into one bigger piece.
4. The Scorecard (AIC): How does the computer know if a cut or a glue-job is good? It uses a mathematical score called AIC (Akaike Information Criterion). Think of this as a "Story Clarity Score." If cutting or gluing makes the story clearer (higher score), the computer keeps the change. If it makes it messier, it reverts.

What They Found

The team tested this "Smart Scissors" method on two groups: Butterflies (Heliconius) and Great Apes (Humans, Chimps, Gorillas, etc.).

1. It's More Accurate:
On simulated data (where they knew the "true" answer), the variable window sizes were much better at finding the correct family tree than the old fixed-size method. It was like using a tailor who custom-fits a suit versus someone using a generic "Medium" size for everyone.

2. The Results for Butterflies:

• The "Smart Scissors" found that some DNA chunks were tiny (as small as 18 base pairs), while others were huge (up to 100,000 base pairs).
• The old method tried to force everything into a tiny 250-base-pair box, which was too small for the big chunks and too big for the tiny ones.
• Despite the different chunk sizes, the overall family tree of the butterflies looked very similar to what scientists had found before, giving them confidence that the new method is reliable.

3. The Results for Great Apes:

• Here, the difference was striking. The new method found that the chunk where humans and chimps are grouped together (the "major story") was supported by 79% of the DNA.
• The old fixed-size method had only found 60%.
• Why the difference? The old method was likely mixing up different stories because its "slices" were too big, blurring the clear signal that humans and chimps are closest relatives.

The "Aha!" Moments

The new method didn't just give better numbers; it found specific biological "treasures":

• The Wing Color Spot: In butterflies, they found a specific 400,000-base-pair region on a chromosome that controls wing colors. The "Smart Scissors" realized this whole region had a unique history (an inversion) that the old fixed method missed. It's like finding a secret chapter in a book that explains why the hero has blue wings, which was previously hidden because the book was cut in the wrong place.
• The Mitochondrial Mystery: In apes, they looked at mitochondrial DNA (the tiny DNA in our cell's power plants). The method found that most of it told the "Human-Chimp" story, but two small chunks told a different, confusing story. This helped scientists realize those two chunks were likely just "noisy" or hard to read, rather than a sign of a different evolutionary history.

The Bottom Line

This paper is about flexibility.

Imagine trying to assemble a puzzle. The old way was to force every piece into a grid, even if the pieces were different shapes. The new way is to let the pieces snap together naturally, regardless of their size.

By letting the window sizes vary, scientists can now see the true "recombination breakpoints"—the exact spots where the evolutionary story changes. This leads to a much clearer, more accurate picture of how species like humans, apes, and butterflies are related to one another. It's a move from a rigid, one-size-fits-all approach to a flexible, custom-fit solution for reading the book of life.

Technical Summary

1. Problem Statement

Phylogenomic studies often analyze whole genome alignments by partitioning them into non-overlapping windows to infer gene trees and assess evolutionary processes like incomplete lineage sorting (ILS) and introgression. A critical challenge in this approach is selecting an appropriate window size:

• Too large: Windows may span multiple recombination breakpoints, leading to "concatenation" where distinct evolutionary histories are averaged, obscuring the true tree topology.
• Too small: Windows lack sufficient phylogenetic signal, leading to high gene tree estimation error.

Previous studies (including Ivan et al., 2025) proposed using the Akaike Information Criterion (AIC) to select a single, fixed optimal window size per chromosome. However, this approach has two major limitations:

1. Stochasticity: Even under constant recombination rates, the lengths of non-recombining blocks (c-genes) vary stochastically.
2. Heterogeneity: Recombination rates vary significantly along chromosomes (e.g., higher at telomeres, lower at centromeres), meaning a single fixed window size cannot accurately represent the entire genome.

2. Methodology

The authors propose a Splitting-and-Merging algorithm that allows window sizes to vary dynamically across the genome to better approximate underlying recombination breakpoints.

The Algorithm:

1. Initialization: The genome alignment is partitioned into non-overlapping windows of a user-defined starting size. The authors tested two extremes: the AIC-selected fixed size and full concatenation (the entire chromosome as one window).
2. Splitting Phase:
   • Each window is iteratively tested for splitting into two sub-windows.
   • Three split ratios are tested: 0.25:0.75, 0.50:0.50, and 0.75:0.25.
   • Gene trees are inferred for the original window and the potential splits using IQ-TREE (Maximum Likelihood, Jukes-Cantor model).
   • Decision Rule: If the sum of the AIC scores of the two sub-windows is lower (better) than the AIC of the original window, the split is retained. This repeats until no further splits improve the AIC.
3. Merging Phase:
   • Neighboring windows are iteratively tested for merging.
   • The AIC of the merged window is compared to the sum of the AICs of the two separate windows.
   • Greedy Strategy: Merges are prioritized based on the magnitude of AIC improvement. Conflicts (e.g., a window merging with both left and right neighbors) are resolved by selecting the merge with the highest AIC gain.
   • This repeats until no further merges improve the AIC.
4. Implementation: The method uses a greedy algorithm (rather than a genetic algorithm) to ensure scalability to whole-genome datasets. It differs from similar tools like GARD by using Maximum Likelihood (ML) instead of Neighbor-Joining and AIC instead of AICc.

Validation:

• Simulations: 70 simulated chromosomes (10 replicates of 7 scenarios) were generated using ms and AliSim. Scenarios included homogeneous recombination rates and heterogeneous rates (mimicking empirical patterns like high rates at ends/low in center).
• Metrics: Accuracy was measured via Site Accuracy (proportion of sites recovering the true topology) and RMSE (Root Mean Squared Error of the tree topology distribution).
• Empirical Data: Applied to whole-genome alignments of the Heliconius butterfly erato-sara clade and Great Apes (Human, Chimpanzee, Gorilla, Orangutan).

3. Key Contributions

• Methodological Innovation: Developed a flexible, AIC-driven splitting-and-merging framework that generates variable window sizes, moving beyond the constraint of fixed-window analyses.
• Optimization Strategy: Demonstrated that starting the algorithm with full concatenation (the whole chromosome) is superior to starting with small fixed windows. This avoids getting trapped in local optima and better captures long non-recombining regions.
• Computational Efficiency: The greedy algorithm allows the method to scale to whole-genome alignments in a reasonable timeframe, unlike more computationally intensive genetic algorithms.

4. Results

Simulation Results:

• Superior Accuracy: The variable window approach consistently outperformed fixed-window methods.
  • In homogeneous recombination scenarios, site accuracy increased by 3.0–4.7%.
  • In heterogeneous scenarios, site accuracy increased by 4.7–8.3%.
  • RMSE decreased significantly in heterogeneous scenarios (0.57–0.81% reduction).
• Starting Size Sensitivity: While results were generally robust, starting with full concatenation yielded the best results in complex scenarios (e.g., Scenario 3), recovering long non-recombining blocks that were fragmented when starting with small windows.

Empirical Results (Heliconius Butterflies):

• Window Size Variation: Window sizes varied drastically, ranging from 18bp to 106kb across chromosomes. The average window size (538–808bp) was significantly larger than the fixed window sizes (250bp) previously selected by AIC.
• Topology Distribution: Despite the change in window sizes, the overall distribution of gene tree topologies remained consistent with previous studies. However, the study highlighted that topology frequency differs significantly when calculated by sites vs. windows, as longer windows (supporting specific topologies) contribute more sites.
• Biological Insights: The method identified specific genomic regions dominated by minor topologies, including a ~400kb region on chromosome 15 spanning the cortex locus (wing color pattern), confirming the presence of structural variants (inversions) affecting phylogeny.

Empirical Results (Great Apes):

• Window Size Variation: Ranged from 17bp to 4.5Mb. Average window sizes were 4.2–6.2kb, significantly larger than the 500bp–1kb fixed sizes used in prior studies.
• Topology Distribution: The proportion of the major topology (Human + Chimpanzee clade) increased to ~79% of sites (compared to ~60% in fixed-window analyses).
• Mitochondrial DNA: The mtDNA was partitioned into 8 windows. Six recovered the major topology, while two recovered minor topologies, likely due to systematic errors in specific regions (e.g., rRNA genes and control regions with low informative sites).

5. Significance

• Addressing Recombination Heterogeneity: The study proves that a single fixed window size is insufficient for whole-genome phylogenomics due to the stochastic and heterogeneous nature of recombination. Variable windows provide a more biologically realistic partitioning of the genome.
• Improved Topological Accuracy: By better aligning windows with non-recombining blocks, the method reduces both concatenation bias and gene tree estimation error, leading to more accurate estimates of species trees and introgression patterns.
• Practical Utility: The method is computationally feasible for large datasets and provides a robust framework for identifying genomic regions of interest (e.g., inversions, introgression hotspots) that might be missed by fixed-window approaches.
• Future Directions: The authors suggest that while this method improves window selection, future work should integrate these variable windows with more complex substitution models (e.g., GHOST) or site-specific history inference methods (e.g., MAST, HMMs) to further refine phylogenetic accuracy.

In conclusion, this paper presents a significant advancement in phylogenomic methodology, offering a data-driven, flexible approach to windowing that adapts to the specific recombination landscape of each chromosome, thereby enhancing the accuracy of evolutionary inferences.