The perceptual and spatial architecture of Mullerian mimicry in Heliconius Butterflies

By integrating deep-learning phenotyping with avian and butterfly visual modeling, this study reveals that *Heliconius* Müllerian mimicry is not a set of discrete, reciprocal rings but a perceptually structured and spatially dynamic continuum driven by local convergence, heterogeneous spatial overlap, and observer-specific sensory weighting.

The Gist

Imagine a massive, living billboard in the rainforests of Central and South America. On this billboard, thousands of butterflies are flashing bright, warning colors: "I taste terrible! Don't eat me!"

For over a century, scientists have believed these butterflies follow a strict rulebook called Müllerian mimicry. The theory goes like this: If you are a poisonous butterfly, you want to look exactly like your neighbors. If a bird eats one, it gets sick and learns to avoid that specific pattern. So, all the poisonous butterflies in a town should converge on one single, perfect "uniform" to maximize their safety. They form distinct "rings" or clubs, where everyone looks identical.

But a new study by Christopher Lawrence and his team suggests the reality is much messier, more dynamic, and far more interesting. They used AI and deep learning to take a fresh look at these butterflies, specifically two famous species: Heliconius erato and Heliconius melpomene.

Here is the story of what they found, explained simply:

1. The "Uniform" is Actually a Gradient

The Old Idea: Think of mimicry like school uniforms. Everyone in the "Postman Club" wears the exact same red and black shirt. Everyone in the "Rayed Club" wears the exact same yellow and black shirt. You are either in the club or you aren't.

The New Discovery: The researchers treated the butterflies' wing patterns not as "uniforms" but as a continuous spectrum of colors, like a giant rainbow.
Using AI (specifically deep learning), they measured the "distance" between every butterfly's pattern. They found that while there are general clusters (the "clubs"), the butterflies don't all look identical. Instead, they form a smooth gradient. A butterfly with two red bands might look slightly different from one with a double white band, even if they are in the same "club."
The Analogy: Imagine a neighborhood where everyone agrees to wear red shirts. The old theory says everyone wears the exact same shade of red. The new study shows that some wear bright crimson, others wear brick red, and some wear maroon. They are all "red," but the shade changes gradually across the neighborhood.

2. Geography Matters More Than Family

The Old Idea: You might think that if two butterflies look alike, they must be close relatives (like cousins).

The New Discovery: The researchers built a family tree and compared it to a map of what the butterflies look like. The result? Family ties barely matter.
A butterfly in the Andes mountains might look nothing like its cousin in the Amazon, even though they share DNA. Instead, where you live dictates what you look like. If you live in a valley where the local birds are used to seeing yellow patterns, you evolve to look yellow, regardless of your family history.
The Analogy: Think of it like fashion trends. A teenager in Tokyo might dress exactly like a teenager in New York, even though they have no family connection. They dress alike because they are influenced by the local "scene" (geography), not their genetics. The butterflies are copying their neighbors, not their parents.

3. The "Imperfect" Copy is Actually Perfect (Depending on Who's Looking)

The Old Idea: If a butterfly looks slightly different from its neighbor, it's a "bad" mimic and might get eaten.

The New Discovery: The researchers asked: Who is doing the looking?
They simulated the vision of birds (the predators) and butterflies (the mates).

• Birds (High Resolution): Birds have sharp eyes. To a bird, a tiny difference in a wing pattern might be obvious.
• Butterflies (Low Resolution): Butterflies have compound eyes that are a bit blurry. To a butterfly, that same tiny difference might be invisible.
  The study found that what looks like a "mistake" or an "imperfect" copy to us (or even to a bird) might be perfectly indistinguishable to a butterfly trying to find a mate.
  The Analogy: Imagine looking at a low-resolution JPEG image on your phone. It looks blurry and pixelated. If you zoom in, you see the "imperfections." But if you stand far away, it looks like a perfect photo. The "imperfection" only exists if you are looking too closely. The butterflies are evolving to look perfect for the distance and eyes of the bird hunting them, not for a scientist with a microscope.

4. The "One-Way Street" of Mimicry

The Old Idea: Mimicry is a handshake. Species A copies Species B, and Species B copies Species A. It's a fair, reciprocal deal.

The New Discovery: The AI models revealed that this "handshake" is often asymmetric.
Sometimes, Species A evolves to look like Species B, but Species B doesn't bother changing to look like A. It's more like advergence (one-way convergence). One species is the "model" (the original), and the other is the "mimic" (the copycat).
The Analogy: Think of a viral dance trend on TikTok. One person invents a dance (the model). Everyone else copies it. But the original creator doesn't start copying everyone else; they just keep doing their own thing. The study suggests mimicry often works this way: one species sets the standard, and the others chase it, rather than everyone meeting in the middle.

5. The "Patchwork" of Overlap

The Old Idea: If two species are mimics, they must live in the exact same spot, side-by-side, all the time.

The New Discovery: The researchers mapped where these butterflies live and found that while they often overlap, they don't always. Some pairs live in the exact same neighborhood; others share a "vibe" (pattern) but live in different parts of the forest.
The Analogy: Imagine a city where everyone wears a specific color to signal they are part of a gang. The old theory says all gang members must live on the same street. The new study shows that while many do, some members of the "Red Gang" live on the North side, and others live on the South side. They don't need to be neighbors to share the same "uniform"; they just need to be in the same general region where the local police (birds) patrol.

The Big Takeaway

This paper changes how we see the natural world. It tells us that nature isn't a collection of rigid, perfect categories (like a library with books in perfect order). Instead, it's a fluid, shifting landscape.

• Mimicry isn't a static ring; it's a moving gradient.
• It's driven by local geography, not just family history.
• "Imperfect" mimicry is often perfect for the eyes that matter.
• Evolution is often a one-way street, not a fair trade.

By using AI to "see" through the eyes of birds and butterflies, the researchers showed us that the warning colors of the rainforest are a complex, living language that is constantly being rewritten by the environment, the viewer, and the distance between them.

Technical Summary

1. Problem Statement

Müllerian mimicry theory posits that chemically defended species converge on shared warning signals (aposematism) to reduce predation through positive frequency-dependent selection. In Heliconius butterflies, this has traditionally been conceptualized as discrete "mimicry rings" where species form reciprocal, exact matches. However, several critical gaps remain in the literature:

• Categorical vs. Continuous: Mimicry is often treated as a binary state (present/absent) rather than a continuous trait, obscuring the degree of resemblance.
• Symmetry: It is unclear whether convergence is reciprocal (mutual adaptation) or asymmetric (one species evolving toward another, known as advergence).
• Perceptual Basis: "Imperfect" mimicry is common, but it is unknown whether these imperfections are perceptually meaningful to predators or if they fall below the discrimination thresholds of specific visual systems.
• Drivers of Variation: The relative roles of phylogenetic history (ancestry) versus geographic proximity (local selection) in shaping mimicry patterns are not fully quantified.

2. Methodology

The authors employed a multi-faceted approach combining deep learning, perceptual modeling, and spatial ecology to analyze 56 subspecies of Heliconius erato and H. melpomene.

A. Data Acquisition and Preprocessing

• Dataset: Standardized dorsal images of 56 subspecies (27 H. erato, 29 H. melpomene).
• Source: Images sourced from the Clinique Heliconius website and the Cambridge Butterfly Collection.
• Preprocessing: Images resized to 1280×1280 pixels; color patterns standardized to remove artifacts from lighting and wing wear using the pavo R package.

B. Perceptual Quantification (Deep Learning)

• Feature Extraction: A pre-trained Convolutional Neural Network (CNN) was used to generate high-dimensional perceptual embeddings. Euclidean distances in this space served as a continuous measure of mimetic similarity.
• Visualization: t-SNE was used to project these high-dimensional embeddings into 2D for visualization, though statistical analyses were performed in the original multidimensional space.
• Null Models: 1,000 random heterospecific pairs were generated to establish a baseline for similarity, compared against known mimicry pairs using Wilcoxon rank-sum tests.

C. Phylogenetic and Geographic Analysis

• Phylogenetic Signal: A phenotypic tree (UPGMA clustering of perceptual distances) was compared against a known species phylogeny. Signal strength was quantified using Blomberg's K and Pagel's λ.
• Geographic Overlap: Georeferenced occurrence data (~16,000 records) were used to calculate Schoener's D (niche overlap) for 16 mimicry pairs using kernel density estimation.

D. Visual Acuity and Classification Modeling

To simulate predator and conspecific perception, the authors trained ResNet-50 models under different visual regimes:

• Visual Filters: Images were processed using AcuityView to simulate:
  • Avian Vision: Modeled after the Belted Kingfisher (Megaceryle alcyon), a primary predator.
  • Butterfly Vision: Modeled after male and female Heliconius (accounting for sexual dimorphism in ommatidia number).
• Training Regimes:
  • Allnet: Trained on both lineages jointly.
  • Eratonet: Trained exclusively on H. erato.
  • Melpomenenet: Trained exclusively on H. melpomene.
• Asymmetry Test: Cross-lineage generalization was tested by training on one lineage and evaluating performance on the other.
• Trait Salience: INTR (a computer vision attention mechanism) was used to generate heatmaps identifying which image regions (e.g., band curvature, patch boundaries) drove classification decisions.

3. Key Results

A. Mimicry as a Continuum, Not Discrete Rings

• Subspecies assigned to the same traditional mimicry ring clustered in perceptual space but showed continuous variation rather than forming discrete, tight clusters.
• Specific visual traits (e.g., red band presence) contributed more to similarity than categorical ring assignments.
• Known model-mimic pairs were significantly more similar than random pairs ($p < 0.001$), validating the perceptual distance metric.

B. Geography Dominates Phylogeny

• Phylogenetic Signal: Weak and non-significant (Blomberg's $K = 0.0365, p = 0.193$). Closely related species did not consistently share similar patterns.
• Geographic Structure: Mimetic similarity was strongly organized by geography. Amazonian populations clustered by yellow-shifted phenotypes, while Andean populations showed cooler-toned clusters.
• Conclusion: Convergence is driven by regional selective environments and community context rather than shared ancestry.

C. Heterogeneous Spatial Overlap

• Spatial niche overlap (Schoener's $D$) among co-mimics was heterogeneous (range: 0.148 to 0.998; mean: 0.789).
• While some pairs showed near-perfect overlap (e.g., H. erato microclea and H. melpomene xenoclea), others with similar phenotypes had limited overlap.
• Implication: Frequency-dependent selection operates at variable spatial scales; shared phenotypes can persist even without strict pairwise co-occurrence in diverse communities.

D. Observer-Dependent Perception and Asymmetry

• Visual Acuity: Avian-calibrated models achieved higher classification accuracy than butterfly-calibrated models, indicating that fine-scale pattern differences are resolvable by predators but may be blurred for conspecifics.
• Sexual Dimorphism: Male-based visual models outperformed female-based models, suggesting sexual selection maintains signal fidelity within species.
• Asymmetric Convergence: Cross-lineage generalization was asymmetric. Patterns learned from H. melpomene generalized to H. erato more effectively than the reverse.
  • This suggests advergence (one species evolving toward a stable template provided by the other) rather than strictly reciprocal Müllerian mutualism.

E. Trait Salience

• Avian Models: Focused on localized, high-frequency features (band curvature, patch boundaries).
• Butterfly Models: Prioritized broader wing regions due to lower spatial resolution.
• Imperfect Mimicry: "Imperfections" often involve traits that fall below the perceptual discrimination thresholds of the relevant receiver (e.g., fine hindwing ray width differences invisible to birds at a distance).

4. Key Contributions

1. Quantitative Framework: Shifts the study of mimicry from categorical assignment to continuous, perceptual distance metrics using deep learning.
2. Perceptual Ecology: Demonstrates that "mimicry rings" are observer-dependent; what constitutes a match depends on the visual system (predator vs. conspecific) and acuity.
3. Asymmetry in Evolution: Provides empirical evidence that mimicry convergence is not always reciprocal, challenging the classical assumption of balanced Müllerian mutualism and supporting the role of advergence.
4. Spatial Dynamics: Reveals that mimicry is structured by geographic gradients and community composition rather than just pairwise species interactions.

5. Significance

This study fundamentally recasts Müllerian mimicry from a static system of discrete, mutually reinforcing rings into a dynamic, perceptually structured continuum.

• Theoretical Impact: It challenges the notion that mimicry requires exact, reciprocal matching, suggesting instead that selection operates on a continuum of "good enough" resemblance defined by the receiver's perceptual limits.
• Evolutionary Insight: The findings highlight that evolutionary trajectories in mimicry are shaped by the interplay of local ecology, community composition, and the specific sensory biases of predators and conspecifics.
• Methodological Advance: The integration of biologically calibrated visual filters with deep learning offers a robust framework for quantifying adaptive resemblance in other systems where traditional morphological metrics fail to capture perceptual reality.

Limitations Noted: The study relies on standardized dorsal images (ignoring posture and 3D structure), does not model dynamic predator learning or fitness consequences directly, and excludes UV reflectance and motion cues, which are relevant in natural contexts.