📄 evolutionary biology

A method for massively scalable phylogenetic network inference

The paper introduces InPhyNet, a novel method that achieves linear scalability and high accuracy in inferring phylogenetic networks for large datasets by merging independently inferred sub-networks, thereby overcoming the computational limitations of existing model-based approaches while providing biologically meaningful insights into complex evolutionary histories.

Original authors: Kolbow, N., Kong, S., Solis-Lemus, C.

Published 2026-04-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kolbow, N., Kong, S., Solis-Lemus, C.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to draw the family tree of a massive group of relatives. In the old days, scientists thought every family was a perfect, branching tree: one parent splits into two children, who split into four, and so on. This works great for simple families.

But in the real world of evolution, families are messy. Sometimes, two distant cousins decide to start a family together (hybridization), or a family adopts genes from a completely different family living next door (horizontal gene transfer). When you try to force these messy, interconnected relationships into a simple tree, the picture breaks. You need a family web (a phylogenetic network) instead of a tree.

The problem? Drawing these webs for thousands of species is like trying to solve a giant 3D puzzle while blindfolded. It's so computationally heavy that current methods can only handle about 30 species at a time. If you try to do 1,000, your computer crashes.

Enter InPhyNet, a new method introduced by Kolbow, Kong, and Solís-Lemus that solves this by using a "divide and conquer" strategy. Here is how it works, explained simply:

1. The "Small Groups" Strategy (Divide and Conquer)

Imagine you are trying to organize a reunion for 1,000 people, but you can only fit 20 people in a small room to figure out their specific family connections.

Step 1: You split the 1,000 people into 50 small groups of 20.
Step 2: In each small room, you use a super-accurate (but slow) method to figure out exactly how those 20 people are related, including any messy "cousin marriages." You get 50 little, perfect family webs.
Step 3: Now, you have 50 little webs, but you need one big web for everyone. This is where the magic happens.

2. The "Glue" (Merging the Groups)

Usually, trying to glue 50 complex webs together is a nightmare. If you get one piece wrong, the whole thing falls apart.

The Innovation: InPhyNet acts like a smart glue. It looks at the 50 little webs and a "distance map" (a list of how different every person is from every other person).
The Process: It starts with all 1,000 people as separate dots. It looks at the little webs to see who must be neighbors. It then slowly merges dots together, building a giant skeleton tree first.
Adding the Mess: Once the tree skeleton is built, it goes back and inserts the "messy" parts (the hybridizations) that it recorded from the small groups. It's like building a sturdy highway system first, and then adding the complex interchanges and bridges later.

3. Why It's a Game Changer

Speed: Old methods were like trying to solve a Rubik's cube by moving one square at a time. InPhyNet is like solving it by doing 50 small cubes at once and snapping them together. It scales linearly, meaning if you double the number of species, it only takes roughly double the time, not 1,000 times longer.
Accuracy: It doesn't just guess; it uses the most accurate methods available for the small groups and then combines them without losing the biological truth.

The Real-World Test: The Plant Family Reunion

To prove it worked, the authors took a massive dataset of 1,158 land plants (a group that includes everything from mosses to giant pine trees).

The Problem: Scientists had been arguing for years about where certain groups, like the Gnetales (a weird group of plants that look like pines but act like flowers), fit in the family tree. Some said they were cousins to pines; others said they were cousins to cypress trees. The data was conflicting.
The Result: InPhyNet didn't just pick one side. It drew a web that showed the Gnetales have a "mixed heritage," connecting to both groups. It also found hidden family secrets (hybridizations) in ferns and pine trees that previous tree-based models missed.

The Bottom Line

Think of InPhyNet as a master architect who can build a skyscraper by first building 50 perfect, detailed models of the floors, and then assembling them into one giant, stable structure. It allows scientists to finally map the complex, messy, and beautiful "web of life" for thousands of species at once, revealing evolutionary stories that were previously impossible to see.

1. Problem Statement

Phylogenetic analyses have traditionally relied on bifurcating trees. However, evolutionary histories often involve non-treelike events such as hybridization, horizontal gene transfer, and introgression. Phylogenetic networks are required to model these reticulate events.

The Bottleneck: Current state-of-the-art methods for inferring phylogenetic networks (e.g., SNaQ, PhyloNet-MPL, PhyloNet-ML) suffer from high computational complexity (often polynomial or worse), limiting them to datasets with fewer than ~30–80 taxa.
The Trade-off: Methods that scale to thousands of taxa (e.g., NeighborNet) infer "implicit" networks that summarize conflicting signals but lack biological interpretability and do not model the multispecies network coalescent (MSNC) process.
The Goal: Develop a method that achieves linear scalability (handling hundreds to thousands of taxa) while maintaining statistical consistency and high biological accuracy under the MSNC model.

2. Methodology: InPhyNet

The authors introduce InPhyNet, a divide-and-conquer framework that merges locally inferred networks into a global topology. The approach is inspired by the TreeMerge algorithm for species trees but adapted for networks.

The Framework Pipeline

Decomposition: The set of taxa $X$ is partitioned into disjoint subsets $S = \{S_i\}$ .
Local Inference: A level-1 (simple) phylogenetic network $C_i$ is inferred for each subset $S_i$ using existing, accurate (but slow) methods like SNaQ or PhyloNet.
Dissimilarity Matrix: A pairwise dissimilarity matrix $D$ is computed for all taxa in $X$ (using the Average Gene Tree Internode Distance, AGID).
Merging (The Core Innovation): InPhyNet merges the local networks $C = \{C_i\}$ and the matrix $D$ into a single global network $\hat{N}$ .

The InPhyNet Algorithm

The merging process is iterative and resembles the Neighbor-Joining (NJ) algorithm but with constraints:

Constraint Networks: The local networks $C_i$ act as "constraints." They dictate which pairs of nodes can be merged.
Neighbor Definition: Two nodes are eligible to merge if they are neighbors in every constraint network where they both appear, or if they do not appear together in any constraint network.
Reticulation Recording: As nodes are merged, the algorithm tracks the topology of the subnetworks being collapsed. If a subnetwork contains a hybrid node (reticulation), the algorithm records the direction of the hybrid edges (incoming/outgoing) and inheritance probabilities ( $\gamma$ ).
Post-Processing: Once the backbone tree is built, the recorded reticulation information is used to insert hybrid nodes and directed edges back into the network, creating a semi-directed phylogenetic network.

Handling Conflicts

When merging more than two constraint networks ( $|C| > 2$ ), conflicts may arise where no valid pairs of nodes can be merged due to contradictory topological constraints. The authors employ a recursive pairwise merging strategy (Algorithm 5) to resolve these conflicts by merging networks two at a time until a single network remains.

3. Key Contributions

Scalability: The method achieves linear scalability with respect to the number of taxa ( $O(N)$ ) in practice, provided the local inference methods are efficient. Theoretical runtime is $O(N^3 + Nr)$ , but the dominant cost is the local inference, which scales linearly with $N$ when using small, fixed-size subsets.
Statistical Consistency: The authors prove that if the local inference methods ( $M_C$ ) and the distance metric ( $M_D$ ) are statistically consistent, and the subset decomposition satisfies specific conditions (ensuring hybrid nodes are "covered" by the subsets), then InPhyNet is statistically consistent for the true species network under the MSNC model.
Biological Interpretability: Unlike implicit networks, InPhyNet produces a semi-directed network with explicit hybrid nodes and inheritance probabilities, allowing for biological interpretation of reticulate events.
Open Source: The method is implemented as a Julia package (InPhyNet.jl).

4. Results

The authors evaluated InPhyNet through extensive simulations and a real-world biological application.

Simulation Study

Setup: Simulated datasets with up to 200 taxa, varying levels of incomplete lineage sorting (ILS), and different constraint network sizes ( $m=10, 20$ ).
Accuracy:
- InPhyNet showed high accuracy, comparable to running SNaQ or PhyloNet on the full dataset (which was computationally infeasible for large $N$ ).
- Accuracy was strongly correlated with the accuracy of the input constraint networks.
- The method was robust to errors in the distance matrix but sensitive to errors in the local network inference.
Runtime:
- Total runtime scaled linearly with the number of taxa.
- Using $m=10$ (smaller subsets) was significantly faster than $m=20$ while maintaining similar accuracy.
- PhyloNet-MPL was the fastest local inference method, followed by SNaQ and PhyloNet-ML.

Empirical Application: Green Plants (1,158 Taxa)

The authors re-analyzed the "One Thousand Plant Transcriptomes" dataset.

Approach: They treated well-resolved clades (non-gymnosperms/non-ferns) as a single constraint and manually decomposed the gymnosperms and ferns (areas of known discordance) into subsets.
Findings:
- Gnetales: The inferred network provided partial support for both the "Gnetifer" (sister to conifers) and "Gnepine" (sister to Pinaceae) hypotheses, reconciling previous conflicting tree-based analyses.
- Pinaceae: Detected reticulation events consistent with known polyploidy in Juniperus and related genera.
- Ferns: Identified reticulation events within Polypodiidae and Woodsia, consistent with known allotetraploidy and hybridization.
- Angiosperms: Confirmed that deep relationships among major clades (e.g., asterids, rosids) remain unresolved, highlighting the limits of current data rather than method failure.

5. Significance

Bridging the Gap: InPhyNet bridges the gap between small-scale, high-accuracy network inference and large-scale, low-accuracy implicit network analysis.
Feasibility: It makes it possible to infer phylogenetic networks for hundreds or thousands of species, a scale previously restricted to tree-based methods.
Evolutionary Insight: By enabling large-scale network inference, the method allows researchers to detect and quantify reticulate evolution (hybridization, introgression) across entire kingdoms of life, providing a more realistic view of evolutionary history than strictly bifurcating trees.
Theoretical Foundation: The proof of statistical consistency provides a rigorous mathematical basis for using divide-and-conquer strategies in phylogenetic network inference.

In conclusion, InPhyNet represents a major advancement in phylogenomics, offering a scalable, statistically consistent, and biologically interpretable solution for reconstructing the complex, reticulate evolutionary histories of large groups of organisms.