Phylogenetic estimation of diversity-dependent biogeographic rates using deep learning

This paper introduces DDGeoSSE, a deep learning-based generative model that enables the estimation of diversity-dependent biogeographic rates and reveals how local species richness shapes diversification dynamics in Caribbean Anolis lizards and Viburnum plants.

The Gist

Imagine you are a detective trying to solve the mystery of why some places on Earth are bursting with life, while others are relatively empty.

For a long time, biologists have known two big things:

1. Life isn't spread out evenly. Some islands or forests are packed with species; others have very few.
2. Life doesn't grow forever. Even in paradise, there's a limit to how many species can fit before things start crashing down.

This paper introduces a new, super-smart tool called DDGeoSSE to figure out exactly how nature hits that "brake pedal."

The Problem: The "Traffic Jam" of Evolution

Think of an ecosystem like a busy city.

• Speciation (Birth): New species are born (like new families moving in).
• Extinction (Death): Species die out (like families moving away or passing away).
• Dispersal (Travel): Species move to new areas (like people commuting to a new neighborhood).

In the past, scientists had models that assumed these rates were constant, like a car driving at a steady speed on an empty highway. But in reality, as the city gets crowded, things change.

• If too many people move in, it gets hard to find a house (resources), so birth rates slow down.
• If the city is too crowded, accidents happen more often, so death rates go up.
• If a neighborhood is full, it's hard for new people to move in, so travel rates drop.

This is called Diversity-Dependence. The problem is, calculating the exact math for how these rates change as the crowd grows is incredibly difficult. It's like trying to predict traffic patterns in a city where every driver is reacting to every other driver in real-time. The math gets so messy that standard computer methods can't solve it.

The Solution: The "AI Traffic Cop"

The authors of this paper built a new model (DDGeoSSE) that simulates this crowded city perfectly. But since the math is too hard to solve with a calculator, they used Deep Learning (a type of Artificial Intelligence) to act as a "Traffic Cop."

Here is how they did it:

1. Training the AI: They didn't try to solve the math directly. Instead, they built a video game simulator. They ran thousands of simulations where they created fake evolutionary histories (trees of life) under different rules (e.g., "What if crowding stops birth rates?" or "What if crowding increases death rates?").
2. Teaching the Pattern: They fed these thousands of fake trees into a neural network (the AI). The AI learned to look at the shape of the tree and say, "Ah, this tree looks like it was created when crowding slowed down births," or "This one looks like crowding increased deaths."
3. The "Black Box" Solver: Now, the AI can look at a real tree of life and guess the rules that created it, even though the underlying math is too complex for humans to write down in a simple formula.

The Detective Work: Two Real-Life Cases

The team took their AI detective and applied it to two real-world groups of animals and plants to see what was actually happening in nature.

Case 1: Caribbean Anoles (Lizards)

• The Scene: These lizards live on Caribbean islands. They are famous for evolving into different shapes to live on tree trunks, twigs, or grass.
• The Verdict: The AI found that as the islands got crowded with lizards, new species stopped being born (birth rates dropped) and extinction rates went up.
• The Metaphor: It's like a party that got too full. No new guests are invited (low birth), and people are getting kicked out or leaving because it's too cramped (high extinction). The lizards hit a "carrying capacity" where the island can't support any more new types.

Case 2: Viburnum (Cloud Forest Plants)

• The Scene: These are shrubs and trees living in mountain cloud forests from Mexico down to South America.
• The Verdict: The AI found that as these plants got crowded, new species stopped being born, and new plants stopped moving into crowded areas (dispersal dropped). However, extinction didn't seem to change much.
• The Metaphor: Imagine a popular hiking trail. Once the trail is full of hikers, no new hikers want to join the group (low dispersal), and the hikers already there aren't splitting up to make new groups (low birth). But nobody is getting hurt or dying; they just stop growing and moving.

Why This Matters

Before this paper, scientists had to guess how crowding affects evolution, or they had to use very simple models that missed the details.

• The Old Way: Like trying to guess the weather by looking at a single cloud.
• The New Way: Using a supercomputer to analyze millions of weather patterns to predict the storm.

This new tool allows scientists to finally ask: "Is the reason we have so many species in the tropics because they are born faster, or because they die slower?" The answer, according to this study, is often that crowding acts as a brake, slowing down the creation of new species and limiting how far they can spread.

In a Nutshell

Nature has a limit. When a place gets too crowded, evolution slows down. The authors built a super-smart AI that learned to read the "fossil record" of DNA trees to figure out exactly how that crowding changes the rules of life. They proved that for lizards and plants, too many neighbors means fewer new neighbors.

Technical Summary

1. Problem Statement

Ecological theory posits that local species richness influences biogeographic rates (speciation, extinction, and dispersal), often leading to an equilibrium or "carrying capacity" where diversity stabilizes. However, existing phylogenetic models face significant limitations:

• Unbounded Growth: Most standard diversification models (e.g., GeoSSE, birth-death models) allow species richness to grow indefinitely, contradicting equilibrium theories like the Theory of Island Biogeography.
• Lack of Flexibility: Existing diversity-dependent models (e.g., DAISIE, modified MuSSE) often assume a single global carrying capacity parameter or rely on deterministic, non-phylogenetic frameworks. They struggle to model complex interactions where diversity affects speciation, extinction, and dispersal differently across multiple regions.
• Intractable Likelihoods: As models become more complex to accommodate diversity-dependence, deriving analytical likelihood functions becomes mathematically intractable, preventing the use of standard Maximum Likelihood or Bayesian inference methods.

2. Methodology

A. The DDGeoSSE Model

The authors introduce DDGeoSSE (Diversity-Dependent Geographic State Speciation-Extinction), a fully generative, event-based phylogenetic model.

• Mechanism: It extends the GeoSSE framework by allowing local species counts ($n_i$) in specific regions to modulate evolutionary rates.
• Logarithmic Formulation: The model uses a "Log-DDG" approach where rate factors are governed by the natural logarithm of species richness ($\log(n_i) + 1$). This introduces a diminishing effect, meaning diversity dependence acts on orders of magnitude rather than raw counts, preventing extreme sensitivity to parameter values.
• Rate Modulation: The absolute rate for any process $p$ (extinction $e$, within-region speciation $w$, dispersal $d$, between-region speciation $b$) in region $i$ is defined as:
  $$Rate(t) = \rho_p \times (\log(n_i(t)) + 1)^{p^D}$$
  Where $\rho_p$ is the base rate and $p^D$ is the diversity-dependent effect parameter.
  • $p^D < 0$: Diversity reduces the rate (e.g., competition limiting speciation).
  • $p^D > 0$: Diversity increases the rate (e.g., diversity-dependent extinction).
  • $p^D = 0$: No diversity dependence.
• Equilibrium Diversity: Unlike previous models that treat carrying capacity as a fixed parameter, DDGeoSSE treats equilibrium diversity ($n^*_i$) as an emergent property derived from the balance between incoming (speciation + dispersal) and outgoing (extinction) rates. The authors derive analytical solutions for $n^*_i$ under specific constraints and numerical solutions for general cases.

B. Deep Learning Inference (Phyddle)

Since the likelihood function for DDGeoSSE is intractable, the authors employ likelihood-free inference using deep learning via the software phyddle.

• Simulation: They generated 50,000 training trees for each of 8 distinct submodels (combinations of diversity dependence on speciation, extinction, and dispersal) across 5 regions.
• Neural Network Architecture: Convolutional Neural Networks (CNNs) were trained to map phylogenetic tree structures (encoded via Compact Diversity Vectors and tip states) directly to model parameters.
• Tasks:
  1. Parameter Estimation: Predicting base rates ($\rho$) and diversity effect parameters ($p^D$).
  2. Model Selection: Classifying which submodel generated a given tree (i.e., determining which processes are diversity-dependent).

3. Key Contributions

1. Novel Model Framework: DDGeoSSE is the first birth-death model to simultaneously account for species diversity, historical biogeography, and state-dependent diversification rates with high generality (allowing different diversity effects on different processes and regions).
2. Theoretical Derivation: The paper provides mathematical proofs for the existence of local equilibrium diversity and derives analytical formulas for equilibrium levels under various constraints (Lemma 1–3).
3. Likelihood-Free Inference Pipeline: It demonstrates a robust pipeline using deep learning to perform parameter inference and model selection for complex biogeographic models where traditional likelihood methods fail.
4. Tree Statistics Analysis: The authors characterize how diversity-dependent processes alter tree shape statistics (e.g., Aldous' $\beta$ for balance, Pybus & Harvey's $\gamma$ for branching times), providing diagnostic tools for model adequacy.

4. Results

Simulation Studies

• Tree Shape Signatures: Different diversity-dependent mechanisms produce distinct phylogenetic signatures:
  • Diversity-dependent speciation ($w^D < 0$): Increases tree balance ($\beta$) and mean range sizes but reduces the number of extant taxa.
  • Diversity-dependent extinction ($e^D > 0$): Increases tree imbalance and turnover (positive $\gamma$), reducing both range sizes and extant taxa.
  • Diversity-dependent dispersal ($d^D < 0$): Primarily affects range sizes and extant taxa counts without drastically altering tree topology.
• Inference Accuracy:
  • Deep learning networks achieved high accuracy in model selection (73–87% accuracy in detecting the presence/absence of diversity dependence).
  • Parameter estimation was precise for base rates ($\rho$) but less certain for diversity effect parameters ($p^D$), particularly extinction effects ($e^D$), which are notoriously difficult to estimate in phylogenetics.

Empirical Applications

The model was applied to two real-world datasets:

1. Caribbean Anolis Lizards (158 species):
   • Best Fit: Submodel 7 (diversity dependence on speciation, extinction, and dispersal).
   • Findings: Strong evidence that high local species richness reduces within-region speciation and immigration rates while increasing extinction rates. This supports the hypothesis of ecological saturation in island radiations.
2. Neotropical Viburnum Plants (Oreinotinus clade, 38 species):
   • Best Fit: Submodel 6 (diversity dependence on speciation and dispersal).
   • Findings: Strong evidence that richness limits within-region speciation. There was moderate support for diversity-dependent dispersal (incumbency effect) but no evidence for diversity-dependent extinction. This suggests the clade may not have reached equilibrium in all regions or that extinction dynamics differ from lizards.

5. Significance

• Bridging Ecology and Phylogenetics: DDGeoSSE provides a rigorous framework to test how ecological competition and niche saturation shape macroevolutionary patterns over deep time.
• Methodological Advancement: By successfully applying deep learning to complex biogeographic models, the paper overcomes the "likelihood bottleneck," opening the door for analyzing increasingly realistic evolutionary scenarios that were previously computationally impossible.
• Biological Insight: The application to Anolis and Viburnum confirms that diversity-dependence is a major driver of biogeographic radiation, but the specific mechanisms (e.g., whether extinction or speciation is the primary regulator) vary significantly between clades.
• Future Utility: The derived tree statistics and the open-source phyddle pipeline offer new tools for researchers to test model adequacy and infer complex diversification histories without relying on simplifying assumptions.