Phylo-Diffusion: Trait Evolution Models

• Phylo-Diffusion is a dual paradigm that employs diffusion-based frameworks to model quantitative trait evolution along phylogenies, integrating stochastic processes with generative image synthesis.
• The framework extends traditional Brownian motion by introducing branch-specific drift via a relaxed-drift model, enabling precise Bayesian inference of evolutionary trends.
• Generative models with hierarchical embeddings allow for controlled image generation and perturbation experiments that reveal functional relationships between traits and phylogenetic structure.

Phylo-Diffusion denotes two rigorously defined modeling paradigms that exploit diffusion-based mathematical frameworks for studying the evolution of quantitative biological traits or traits observable in empirical data subject to phylogenetic structure. The first paradigm (introduced by Gill et al., 2016) extends stochastic process models for phylogenetic trait evolution to include lineage- and branch-specific drift, explicitly addressing non-neutral evolutionary scenarios in continuous traits along phylogenies (Gill et al., 2015). The second paradigm (introduced by Paik et al., 2024) applies deep conditional diffusion models, structured by hierarchical phylogenetic embeddings (HIER-Embeds), to generative modeling of images across species, thereby enabling direct visualization and analysis of evolutionary trait variation in complex observation spaces (Khurana et al., 2024).

1. Stochastic Diffusion Models for Phylogenetic Trait Evolution

The foundational Phylo–Diffusion model with relaxed drift describes the stochastic evolution of multivariate continuous traits $X(t)\in\mathbb{R}^M$ along the edges of an unknown phylogenetic tree $\tau$. The evolutionary dynamic for a branch $b$ of length $t_b$ is given by the stochastic differential equation

$dX(t) = \mu_b\,dt + \Sigma^{1/2} dW(t)$

where $\mu_b$ is a deterministic drift vector specific to branch $b$, $\Sigma$ is the diffusion covariance matrix, and $W(t)$ is a standard $M$-variate Wiener process. The transition along a branch is then

$\tau$0

allowing for explicit modeling of branch-specific evolutionary trends beyond neutral Brownian motion.

To ensure model identifiability when assigning distinct $\tau$1 to branches, a “relaxed-drift” constraint is imposed: at each internal node, at most one child branch can acquire a new drift relative to its parent, encoded via ternary indicators $\tau$2. This setup yields a uniquely invertible mapping for the drift parameters along the tree (Gill et al., 2015).

2. Bayesian Statistical Framework and Efficient Algorithms

The joint model simultaneously reconstructs molecular sequences and continuous traits, assuming: (i) a CTMC substitution model for sequences, (ii) Brownian diffusion with relaxed drift for traits, and (iii) conditional independence given the tree and parameters. The posterior distribution is

$\tau$3

where $\tau$4 includes drift vectors, increments, precision matrix $\tau$5, root state, and indicator variables.

Trait-data likelihood is evaluated efficiently via a drift-augmented Felsenstein pruning algorithm scaling as $\tau$6. MCMC sampling alternates between topology updates, Gibbs sampling of conjugate parameters (precision matrix), and stochastic-search variable selection (BSSVS) for drift increments via ternary proposals on the indicator variables $\tau$7. This formalism enables scalable, robust Bayesian inference and is implemented as an extension of BEAST v1.x (Gill et al., 2015).

3. Generative Phylogenetic Diffusion Models with HIER-Embeds

The second paradigm leverages modern diffusion probabilistic models, specifically denoising diffusion probabilistic models (DDPMs) and latent diffusion models (LDMs), for image generation conditioned on phylogenetic hierarchies. The forward process iteratively corrupts data $\tau$8 with Gaussian noise according to a schedule $\tau$9: $b$0. The reverse process is learned via U-Net architectures, with phylogenetic information injected as conditioning via hierarchical embeddings (Khurana et al., 2024).

The hierarchical embedding—HIER-Embed—partitions the tree of life into $b$1 ancestral levels (e.g., 4), and for each species $b$2 collects concatenated learned vectors for the ancestor at each level:

$b$3

where $b$4 is the embedding for the ancestor at level $b$5. This construct ensures that shared ancestry at a given phylogenetic level is reflected as identical segments in embedding space, structuring the generative process along evolutionary constraints.

Conditioning is performed by feeding the HIER-Embed through a small text encoder to produce key/value representations for cross-attention blocks within the U-Net, allowing the model to modulate generation according to evolutionary relationships.

4. Perturbation Experiments: Trait Masking and Trait Swapping

Phylo-Diffusion introduces direct intervention mechanisms in the embedding space to probe trait–phylogeny relationships, drawing analogy to gene knockout and gene swapping.

• Trait masking: For a given species, a specified level in the HIER-Embed is replaced with i.i.d. Gaussian noise. Sampling with the masked embedding typically “melts” classifier probability mass into the relevant subtree, quantifiable via a probability-shift metric $b$6 defined as the change in classifier output before and after masking, averaged within and outside the affected clade. Results demonstrated $b$7 vs $b$8, indicating specificity of trait perturbations to evolutionary subgroups (Khurana et al., 2024).
• Trait swapping: The $b$9-th level segment in the HIER-Embed for species $t_b$0 is swapped with that of a sibling lineage $t_b$1. Generation with this swapped embedding alters key phenotypic traits (e.g., disappearance of barbels, changes in stripe/fins for fish), and classifier probabilities shift accordingly, increasing for descendants of the swapped-in lineage and dropping for the original.

These experiments provide a formal mechanism for in silico functional trait analysis and lineage-specific perturbation studies.

5. Empirical Evaluation and Biological Outcomes

Phylo-Diffusion was evaluated on multispecies image datasets (5,434 fishes and 190 bird species), using OpenTree and birdtree.org phylogenies discretized to four levels, with images encoded via VQGAN and U-Nets with cross-attention conditioning.

Key quantitative outcomes:

• Fréchet Inception Distance (FID): Phylo-Diffusion 11.38 (lower is better; outperforming GAN-based Phylo-NN at 28.08).
• Classification F1 score: 82.2% (compared to 70.2% for SciName-encoding).
• Embedding distances align with phylogenetic distance, unlike baseline class label embeddings.
• Trait masking and swapping produce classifier probability shifts and image variations aligned with biological logic.

Notably, trait diffusion along specific clades enables discovery of key evolutionary transitions: appearance/disappearance of fish barbels at level-2 branching, caudal fin morphology transitions, and coloration differences in avian eye and beak shape (Khurana et al., 2024).

6. Case Studies in Phylogenetic Trait Evolution

The relaxed-drift diffusion framework was applied to several biological systems (Gill et al., 2015):

System	Trait Modeled	Insight from Phylo-Diffusion
HIV-1 in Central Africa	Longitude & latitude spread	Relaxed drift uncovers east–south migration and localized counterflows; branch-specific drifts estimated; improved ancestral location precision
West Nile Virus, North America	Longitude & latitude spread	Single westward drift suffices; relaxed drift does not improve fit, validating model selection sensitivity
HIV-1 antigenic evolution	Log-IC50 for bNAb resistance	Population-level drift toward VRC01 resistance and strong branch-specific deviations detected

This demonstrates the capacity to recover both broad and localized evolutionary trends, decoupling stochastic diffusion from persistent directional change.

7. Limitations, Extensions, and Prospects

Current limitations for generative Phylo-Diffusion include discretization to a fixed number of phylogenetic levels; a plausible extension is dynamic or continuous-time embedding assignment, allowing for variable granularity along the tree. Additional avenues include incorporating genomic or morphological covariates, detecting convergent evolution via embedding perturbation correlations, and quantifying ancestral state uncertainty (Khurana et al., 2024).

Traditional stochastic Phylo–Diffusion is constrained by the assumption of Gaussian trait evolution, which may not capture trait distributions with heavy tails or context-dependent evolvability; further generalizations could address these issues. Both paradigms open new methodological and empirical directions for evolutionary biology, macroevolution, and quantitative trait analysis by connecting trait variation, stochastic evolutionary processes, and modern generative modeling.