TriMouNet: An Algorithm for Inferring Level-1 Phylogenetic Networks from Multi-Locus Gene Tree Distributions.

TriMouNet is a novel algorithm that infers level-1 phylogenetic networks from multi-locus gene tree distributions by statistically selecting best-fitting trinets to assemble a full network, thereby outperforming concatenation-based methods like TriLoNet in accurately identifying reticulations while minimizing false positives.

The Gist

Imagine you are trying to draw a family tree for a huge group of animals, like birds or yeast. Usually, scientists assume life is a simple branching tree: a grandparent splits into two parents, who split into two children, and so on.

But in reality, evolution is messier. Sometimes, two different families mix back together (like a hybrid dog), or lineages get confused because they didn't sort out their genes fast enough before splitting. This is called reticulate evolution (think of it as a net rather than a tree).

The paper introduces a new tool called TriMouNet to draw these messy "family nets" more accurately. Here is how it works, using simple analogies:

The Problem: The "Single Snapshot" vs. The "Movie"

Previously, a tool called TriLoNet tried to figure out these family nets by looking at a single, long strand of DNA (a sequence alignment) for just three animals at a time.

• The Analogy: Imagine trying to figure out the plot of a complex mystery movie by looking at one single frozen frame. You might see two people standing next to each other, but you don't know if they are friends, enemies, or if one is hiding something. If the lighting is bad (noise in the data), you might guess the wrong relationship.

TriMouNet changes the game. Instead of looking at one frozen frame, it looks at thousands of different "snapshots" (gene trees) from the same three animals, taken from different parts of their genome.

• The Analogy: Instead of one photo, TriMouNet watches a movie made of thousands of frames. It sees that in 60% of the frames, Animal A and B are best friends, but in 40% of the frames, Animal A and C are best friends. This "movie" tells a much clearer story about their messy history.

How TriMouNet Works: The Detective's Toolkit

1. The "Three-Person" Clue (Trinets)
The algorithm breaks the whole group down into tiny groups of three animals (called "trinets"). It asks: "Who is related to whom in this trio?"

• The Old Way: It guessed based on a single DNA strand.
• The New Way: It looks at the distribution of thousands of gene trees. If the genes are arguing (some say A+B, some say A+C), it knows there was a "mix-up" (reticulation) in their history.

2. Measuring the "Confusion"
TriMouNet uses math to measure how "bumpy" the gene history is.

• The Analogy: Imagine you are trying to guess the weight of a bag of apples.
  • If you weigh it once and get 5kg, you guess 5kg.
  • If you weigh it 1,000 times and the scale jumps between 4kg and 6kg, you realize the bag contains two different things mixed together.
  • TriMouNet sees this "jumping" in the gene data. It knows that if the data is split between two stories, it's not just a simple tree; it's a network with a hybrid event.

3. Building the Net
Once it solves the puzzle for every trio of animals, it stitches all those tiny puzzles together to build the full family net.

• The Analogy: It's like solving a giant jigsaw puzzle. Instead of trying to force pieces together based on a blurry picture, TriMouNet looks at the shape of the pieces (the gene data) to see exactly where they fit.

Why It's Better (The Results)

The authors tested TriMouNet on real data (yeast, cypress trees, and birds) and compared it to the old method (TriLoNet).

• The Yeast Test: The old method got confused and mashed several distinct yeast species into one big, blurry blob. TriMouNet, looking at the thousands of gene "snapshots," correctly separated them and found the specific hybrid yeast that was a mix of two parents.
• The Bird Test: Bird evolution is notoriously messy. The old method gave up and drew a single, unhelpful circle (a "cactus") because the data was too confusing. TriMouNet successfully untangled the knots, identifying clear groups like parrots and falcons, and even spotting where different groups had mixed in the past.

The Bottom Line

TriMouNet is a smarter way to draw evolutionary family trees when the history is messy.

• Old Method: "Let's look at one photo and guess."
• TriMouNet: "Let's watch the whole movie, count the frames, and see the pattern."

By using data from many genes instead of just one, it avoids the "hallucinations" (wrong guesses) that happen when you try to force a complex, mixed-up history into a simple, straight line. It helps scientists see the true, tangled web of life.

Technical Summary

1. Problem Statement

Phylogenetic analyses increasingly utilize multi-locus datasets (thousands of genes) to reconstruct evolutionary histories. While the Multi-Species Coalescent (MSC) model explains gene tree discordance via Incomplete Lineage Sorting (ILS), it does not account for reticulate evolution (e.g., hybridization, introgression).

• Limitation of Existing Methods: Traditional methods like TriLoNet construct Level-1 networks (networks where no two cycles share a node) by inferring "trinets" (3-taxon networks) directly from single sequence alignments. This approach suffers from:
  • Lack of parsimony-informative sites in 3-way alignments.
  • Sensitivity to violations of model assumptions (e.g., molecular clock, site independence).
  • Inability to distinguish between ILS and reticulation when signal strength is low.
  • High false positive rates for reticulations when applied to concatenated data or single loci.
• Goal: Develop a method that leverages the distribution of topologies and branch lengths across many gene trees (multilocus data) to robustly infer reticulate events, distinguishing them from ILS-induced discordance.

2. Methodology: TriMouNet

TriMouNet (Trinet Multilocus Network) follows a "divide-and-conquer" strategy similar to TriLoNet but fundamentally changes the input and inference mechanism for trinets.

A. Core Workflow

1. Gene Tree Reconstruction: Reconstruct individual gene trees for all loci in the dataset (using tools like IQ-TREE 2).
2. Trinet Inference (Step A): Instead of analyzing a single alignment, TriMouNet analyzes the distribution of splits and branch lengths across all gene trees for every set of three taxa (plus an outgroup).
3. Network Assembly (Step B): Merge the inferred trinets into a global Level-1 network using a TriLoNet-like merging algorithm (Tarjan's algorithm for strongly connected components).

B. Statistical Framework

TriMouNet models branch lengths using an Exponentially Modified Gaussian (EMG) distribution to account for both coalescent waiting times (exponential) and estimation errors (Gaussian).

Key Statistical Steps:

1. Classification into S-type vs. NT-type:

   • For a triplet $\{x, y, z\}$ with outgroup $o$, the algorithm counts the frequency of the three possible topologies ($xy|z$, $xz|y$, $yz|x$).
   • A Binomial Test determines if the two least frequent topologies are significantly asymmetric.
     • Symmetric (NT-type): If the two discordant topologies are roughly equal, the signal is likely due to ILS or a specific reticulation pattern where direction is unknown.
     • Asymmetric (S-type): If one discordant topology is significantly more frequent, it suggests a specific reticulation event (hybridization) where one lineage is the descendant.

2. Branch Length Analysis (Normalized Variables):
   To distinguish specific trinet topologies, TriMouNet calculates scale-free ratios from the gene trees:

   • $V_1$ (Cherry Ratio): Normalized distance between sister taxa.
   • $V_2$ (Pending Ratio): Mean normalized distance from the third taxon to the sister pair.
   • $V_3$ (Common Edge Ratio): Used for S-type trinets to measure the shared edge length relative to the path to the outgroup.

3. Model Fitting (EMG Mixture):

   • The algorithm fits EMG distributions to the observed $V_1, V_2, V_3$ values across loci.
   • It compares a single-mode model (unimodal distribution, implying a standard tree-like structure or specific NT-type) vs. a two-mode mixture model (bimodal distribution, implying a reticulation event).
   • Objective Function: Uses the Negative Log-Likelihood of the EMG mixture. The difference in likelihood between the single and mixture models determines the trinet type (e.g., $T_1, N_2, N_3, N_4, S_1, S_2$).

4. Network Construction:

   • Uses the inferred trinet types and their statistical support scores (weights) to guide the merging process.
   • If a cactus structure (a complex of reticulations) is supported by S-type trinets, it is constructed; otherwise, standard binets or trees are used.
   • Dashed Edges: Used to represent parts of the network where the signal is ambiguous or insufficient to resolve a specific ordering, preventing false positives.

3. Key Contributions

• Shift from Sequence to Distribution: Moves trinet inference from single-alignment patterns to the statistical distribution of gene tree topologies and branch lengths, leveraging the power of genomic-scale data.
• Robustness to ILS: Explicitly models the expectation that under pure ILS, discordant triplets should be symmetric. Asymmetry is treated as a signal for reticulation.
• Statistical Support Scores: Assigns quantitative weights ($w_1$) to reticulation events based on the strength of the bimodal signal, allowing for the visualization of uncertainty.
• Improved Topological Accuracy: By including an outgroup in the 4-taxon analysis (3 ingroups + 1 outgroup) for every trinet, the method anchors the topology better than 3-taxon-only methods, reducing errors caused by long-branch attraction or clock violations.

4. Results

A. Simulated Data

• Trinet Assignment: TriMouNet successfully identified bimodal signals (indicating reticulation) in simulated datasets. Detection power increased with the number of loci and the strength of the reticulation signal (shorter branch length ratios).
• Network Reconstruction:
  • Strong Signal: TriMouNet exactly recovered the true network topology, including reticulated cherries and cactus structures.
  • Weak Signal: As the signal weakened (shorter internal branches, closer coalescent times), TriMouNet correctly avoided inferring false reticulations (low false positive rate), whereas it failed to resolve the specific topology, defaulting to a tree-like structure or ambiguous edges.
  • Comparison: TriLoNet applied to concatenated data frequently predicted incorrect reticulations due to model violations.

B. Empirical Datasets

1. Yeast (Saccharomycotina):
   • TriMouNet resolved the Saccharomyces clade and the relationship between S. kudriavzevii and its parents, consistent with previous literature.
   • It correctly identified a reticulation involving S. kudriavzevii with asymmetric parental contributions (67.5% vs 32.5%).
   • TriLoNet collapsed distinct clusters into a single cactus and failed to resolve known sister relationships (e.g., T. delbrueckii and Z. rouxii).
2. Cupressaceae (Cypress family):
   • TriMouNet resolved the Thuja clade and identified a reticulated cherry between T. standishii and T. sutchuenensis.
   • It captured complex discordance in the Cupressus-Juniperus-Xanthocyparis complex, visualizing it as a cactus structure supported by conflicting gene tree signals.
   • TriLoNet failed to resolve internal splits, collapsing the entire group into an unresolved cactus.
3. Birds (Neoaves):
   • TriLoNet failed completely, collapsing all 11 bird taxa into a single unresolved cactus due to short internal edges and high variance.
   • TriMouNet successfully recovered all major known clusters (e.g., Psittacopasseres, Australaves) and identified reticulated cherries representing ancient admixture.

5. Significance and Conclusion

• Accuracy: TriMouNet demonstrates significantly higher accuracy and lower false positive rates compared to TriLoNet, particularly in datasets with high levels of ILS or complex reticulation.
• Interpretability: By distinguishing between "tree-like" discordance (ILS) and "reticulate" discordance (hybridization) using statistical distributions, it provides a more biologically realistic view of evolutionary history.
• Future Directions: The authors suggest future work on improving gene tree quality control (filtering outliers) and extending the framework to Level-2 networks (networks with overlapping cycles), which are computationally more challenging but biologically necessary for complex histories.

In summary, TriMouNet represents a significant advancement in phylogenetic network inference by integrating multilocus statistical power with the structural efficiency of trinet-based assembly, offering a robust tool for detecting reticulate evolution in the genomic era.