Learning Deep Phylogeny Through Epistasis-Aware Geometry
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Learning Deep Phylogeny Through Epistasis-Aware Geometry".
Marcus: Here's the gist of the paper. HyperEvoGen introduces a hyperbolic generative framework that couples a variational autoencoder with biologically informed training objectives to address representation learning, evolutionary distance estimation, and ancestral sequence reconstruction.<ref:2604.22997#pg19> The arrangement of protein sequences in HyperEvoGen’s hyperbolic embedding aims to preserve phylogenetic structure and produce latent distances which scale with true evolutionary divergence<ref:2604.22997#pg4> This combination of accuracy and throughput supports large-family evolutionary studies and accelerates design-oriented applications<ref:2604.22997#pg2> HyperEvoGen enables fast, scalable modeling of protein evolution while preserving hierarchical relatedness in a geometry-aware representation<ref:2604.22997#pg2> The Core Innovation of the Latent Space A central innovation of HyperEvoGen is the use of a hyperbolic latent space, implemented as an n-dimensional Poincar´e ball, to encode protein sequence variation<ref:2604.22997#pg4> Protein evolution is fundamentally tree-like, with lineages branching and diversifying over time<ref:2604.22997#pg2> Euclidean latent spaces struggle to represent hierarchies without distortion because Euclidean volume and distances grow only polynomially with radius; as a result, deep branches become “crowded” near the boundary and large evolutionary separations are compressed<ref:2604.22997#pg4> Hyperbolic space, by contrast, expands exponentially with radius, naturally accommodating tree-like structures<ref:2604.22997#pg2> The Brownian/Yule trajectory experiments make this geometric mismatch explicit: Euclidean distances between tips systematically underestimate patristic path lengths, whereas hyperbolic embeddings preserve a much tighter relationship with true distances along the tree<ref:2604.22997#pg4> How the Training Signal is Built Model training uses a compound loss function that combines the standard VAE objective with three additional biologically motivated terms<ref:2604.22997#pg2> The total training objective is the sum of all these components: the VAE reconstruction–KL loss (ELBO) plus the critic’s Wasserstein loss, the covariance loss, and the correlation loss<ref:2604.22997#pg2> 1.
Ines: The adversarial critic provides a Wasserstein loss that pushes reconstructed sequences towards the empirical distribution<ref:2604.22997#pg2> 2.
Marcus: A covariance-based term is included to capture epistatic dependencies between residue positions<ref:2604.22997#pg2> 3.
Ines: A divergence correlation loss aligns distances in the latent space with observed p-distances for nearby sequences<ref:2604.22997#pg2> Where the Gains Come From The covariance matching loss penalizes the model if it fails to reproduce the pairwise co-occurrence patterns present in real protein sequences, thereby biasing the VAE to allocate latent capacity to epistatic interactions<ref:2604.22997#pg2> The divergence correlation loss explicitly encourages the former behavior by maximizing the Pearson correlation over all pairwise distances in the batch<ref:2604.22997#pg2> HyperEvoGen’s embeddings maintain a nearly linear correspondence with true evolutionary distances over a broad range of divergences, mitigating the saturation problem that afflicts simple sequence dissimilarity measures<ref:2604.22997#pg2> The model-based geometry shows a characteristic crossover: inferior accuracy at very low divergence but consistently superior accuracy at moderate to high divergence, overtaking ML at 0.2–0.3 p-distance<ref:2604.22997#pg2> Ancestral Sequence Reconstruction HyperEvoGen performs ASR as a generative decoding problem in a continuous latent space that has been trained to capture global sequence statistics under an epistatic Potts model<ref:2604.22997#pg2> Once extant sequences are embedded, ancestors are placed by interpolating along hyperbolic geodesics between child embeddings; the latent midpoint, estimated by sampling along the geodesic, is then decoded into an amino-acid sequence<ref:2604.22997#pg2> This procedure implicitly respects the global structure of the landscape and the couplings encoded during training<ref:2604.22997#pg2> HyperEvoGen’s approach to ancestral sequence reconstruction is also distinct from conventional methods<ref:2604.22997#pg2> As divergence increases and the assumptions of those models are progressively violated by multiple hits and epistasis, HyperEvoGen’s geodesic ASR overtakes ML, yielding lower MRCA reconstruction error and better root sequence recovery across a variety of tree topologies<ref:2604.22997#pg2> The superior performance of HyperEvoGen for deep ancestors suggests that this geometric picture is not merely aesthetic but carries real inferential advantages<ref:2604.22997#pg2> Limitations and Future Directions All quantitative evaluations are performed on synthetic data generated from Potts models inferred on a single Pfam family (PF00069) and evolved along Yule trees<ref:2604.22997#pg2> The Potts model itself, inferred from a curated alignment, is an approximation to the true fitness landscape and may omit higher-order interactions or rare but functionally important states<ref:2604.22997#pg2> Several design choices in HyperEvoGen are currently heuristic<ref:2604.22997#pg2> A more systematic exploration of these hyperparameters, or approaches that learn them directly from data (for example, by optimizing clade recovery or ASR accuracy on held-out simulations), could further improve robustness and generality<ref:2604.22997#pg2> Incorporating Bayesian treatments of latent positions, sampling over geodesic paths, or integrating branch length information could yield richer estimates with credible intervals rather than single-point reconstructions<ref:2604.22997#pg2> HyperEvoGen represents, to our knowledge, the first integrated framework that deploys a hyperbolic generative model with biologically informed losses for simultaneous sequence generation, phylogenetic inference, and ancestral reconstruction<ref:2604.22997#pg2> The paper is 5 pages long and contains 20 references.
Paper summary: Ines: So, summarizing this paper "Learning Deep Phylogeny Through Epistasis-Aware Geometry," they propose a framework using a Poincaré variational autoencoder with adversarial training and hyperbolic latent geometry to map protein sequences. They claim this approach preserves phylogenetic structure and generates latent distances that match true evolutionary divergence one <ref:2604.22997#pg1>.
Marcus: The authors use a compound loss function that mixes the standard VAE loss with terms for covariance and correlation to ensure the representation learns biologically meaningful features, like epistatic dependencies two <ref:2604.22997#pg1>. This allows them to build embeddings where the geometry actually reflects evolutionary constraints.
Yuki: Ultimately, this framework enables ancestral sequence reconstruction by interpolating along hyperbolic geodesics between extant sequences two <ref:2604.22997#pg1>. It offers a way to infer ancestral proteins by traversing these constrained regions of the latent space seventeen.
Ines: The paper suggests that this geometric modeling outperforms machine learning at moderate to high divergence levels, and it shows advantages in reconstructing deep ancestors when sequence evolution becomes complex due to multiple hits and epistasis eleven.
Marcus: They acknowledge that their synthetic data comes from a Potts model approximation, so the real-world applicability depends on how well that model reflects actual protein fitness landscapes two <ref:2604.22997#pg1>. They also flag that some of the design choices they made are still heuristic and need more systematic exploration two <ref:2604.22997#pg1>.
Yuki: So, while it provides a powerful tool for inferring deep phylogeny and ancestral sequences under those specific modeling assumptions, we should treat the results with caution until we see how this holds up on more diverse datasets two <ref:2604.22997#pg1>.
Conclusion: Ines: So we’ve seen how they used a hyperbolic space and some clever math to figure out evolutionary trees without getting lost in the noise of real protein evolution, and now we’re looking at what this paper is actually trying to say about those methods.
Marcus: It’s titled "Learning Deep Phylogeny Through Epistasis-Aware Geometry," and it basically says they built a model that keeps the shape of evolution intact by using geometry that matches how proteins actually mutate.
Ines: What they did was use a variational autoencoder with this hyperbolic embedding, and their main goal was to get latent distances that actually scale with how much time has passed in evolution. It’s about making sure the distance between two sequences in their model reflects the real evolutionary divergence between them.
Marcus: The authors used a compound loss function, not just one simple thing, they added terms for covariance and correlation so the model learns those tricky dependencies—the epistasis—between different parts of a protein that influence each other during evolution.
Yuki: For me, what this means is that when we look at a tree of related proteins, this model isn't just drawing lines between them; it’s capturing the underlying physical constraints on how those proteins actually evolve over time in a way that respects their functional relationships.
Ines: Exactly. It changes the perspective from just seeing sequence differences to understanding the geometric pathway those sequences took through evolutionary history.
Marcus: The implication is that we might be able to do ancestral sequence reconstruction much more reliably, especially when we’re dealing with deep branching events where simpler models start to break down.
Yuki: And that’s huge for population genetics too; it gives us a better way to infer the historical structure of protein families, which tells us a lot about how those species diversified over millions of years.
Ines: They did show that when divergence gets high, this geometric approach actually beats some other machine learning methods for reconstructing ancestors, which is a pretty big deal for inferring deep history.
Marcus: But they did have to be careful with the training data—it came from a specific Potts model simulation, so we gotta remember that the fidelity of their results is tied to how well that initial model approximates real protein fitness landscapes.
Yuki: Right, so it’s not a magic answer, but it shows a new way to approach these complex evolutionary questions by using geometry to guide the learning process.
Ines: Exactly. And while they flag some of their choices as heuristic, this framework is pushing the boundary for how we use generative models to understand deep biological history.
Marcus: It’s definitely worth watching because it connects the abstract geometry of a space directly to the concrete problem of protein evolution and sequence divergence.
Institute for Genomics and Evolutionary Medicine, Temple University · Computer and Information Sciences, Temple University · Institute for Computational Molecular Science, Temple University · Department of Biology, Temple University
q-bio.QM
Submitted: 2026-04-24
Updated: 2026-11-10
Comments: 28 pages, 7 figures, 1 Table
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 83/100
The gist: HyperEvoGen introduces a hyperbolic generative framework that couples a variational autoencoder with biologically informed training objectives to address representation learning…
Key concepts
- Hyperbolic Latent Space
- The model uses an n-dimensional Poincaré ball geometry instead of standard Euclidean space. This choice is crucial because protein evolution is tree-like; hyperbolic space naturally accommodates hierarchical branching structures without distorting relationships, unlike Euclidean spaces which struggle with deep branches.
- Compound Loss Function
- Training combines the standard VAE loss with three biological terms: an adversarial loss to match empirical distributions, a covariance loss to capture residue dependencies (epistasis), and a correlation loss to align latent distances with observed p-distances. This forces the model to learn representations that reflect real evolutionary constraints.
- Geodesic Ancestral Sequence Reconstruction
- Ancestors are inferred by placing them as midpoints along hyperbolic geodesics connecting extant sequence embeddings. This method respects the global structure learned during training, allowing for more accurate reconstruction of ancestral proteins, especially when divergence is high.
Terminology
Summary
HyperEvoGen introduces a hyperbolic generative framework that couples a variational autoencoder with biologically informed training objectives to address representation learning, evolutionary distance estimation, and ancestral sequence reconstruction. p. 3<ref:2604.22997#pg19> The arrangement of protein sequences in HyperEvoGen’s hyperbolic embedding aims to preserve phylogenetic structure and produce latent distances which scale with true evolutionary divergence<ref:2604. p. 022997#pg4> This combination of accuracy and throughput supports large-family evolutionary studies and accelerates design-oriented applications<ref:2604. p. 122997#pg2> HyperEvoGen enables fast, scalable modeling of protein evolution while preserving hierarchical relatedness in a geometry-aware representation<ref:2604. p. 022997#pg2>
The Core Innovation of the Latent Space
A central innovation of HyperEvoGen is the use of a hyperbolic latent space, implemented as an n-dimensional Poincar´e ball, to encode protein sequence variation<ref:2604.22997#pg4> Protein evolution is fundamentally tree-like, with lineages branching and diversifying over time<ref:2604. p. 322997#pg2> Euclidean latent spaces struggle to represent hierarchies without distortion because Euclidean volume and distances grow only polynomially with radius; as a result, deep branches become “crowded” near the boundary and large evolutionary separations are compressed<ref:2604. p. 322997#pg4> Hyperbolic space, by contrast, expands exponentially with radius, naturally accommodating tree-like structures<ref:2604. p. 322997#pg2> The Brownian/Yule trajectory experiments make this geometric mismatch explicit: Euclidean distances between tips systematically underestimate patristic path lengths, whereas hyperbolic embeddings preserve a much tighter relationship with true distances along the tree<ref:2604. p. 322997#pg4>
How the Training Signal is Built
Model training uses a compound loss function that combines the standard VAE objective with three additional biologically motivated terms<ref:2604.22997#pg2> The total training objective is the sum of all these components: the VAE reconstruction–KL loss (ELBO) plus the critic’s Wasserstein loss, the covariance loss, and the correlation loss<ref:2604. p. 622997#pg2>
-
The adversarial critic provides a Wasserstein loss that pushes reconstructed sequences towards the empirical distribution<ref:2604. p. 422997#pg2>
-
A covariance-based term is included to capture epistatic dependencies between residue positions<ref:2604. p. 522997#pg2>
-
A divergence correlation loss aligns distances in the latent space with observed p-distances for nearby sequences<ref:2604. p. 522997#pg2>
Where the Gains Come From
The covariance matching loss penalizes the model if it fails to reproduce the pairwise co-occurrence patterns present in real protein sequences, thereby biasing the VAE to allocate latent capacity to epistatic interactions<ref:2604.22997#pg2> The divergence correlation loss explicitly encourages the former behavior by maximizing the Pearson correlation over all pairwise distances in the batch<ref:2604. p. 622997#pg2> HyperEvoGen’s embeddings maintain a nearly linear correspondence with true evolutionary distances over a broad range of divergences, mitigating the saturation problem that afflicts simple sequence dissimilarity measures<ref:2604. p. 722997#pg2> The model-based geometry shows a characteristic crossover: inferior accuracy at very low divergence but consistently superior accuracy at moderate to high divergence, overtaking ML at 0. p. 92–0.3 p-distance<ref:2604. p. 022997#pg2>
Ancestral Sequence Reconstruction
HyperEvoGen performs ASR as a generative decoding problem in a continuous latent space that has been trained to capture global sequence statistics under an epistatic Potts model<ref:2604.22997#pg2> Once extant sequences are embedded, ancestors are placed by interpolating along hyperbolic geodesics between child embeddings; the latent midpoint, estimated by sampling along the geodesic, is then decoded into an amino-acid sequence<ref:2604. p. 922997#pg2> This procedure implicitly respects the global structure of the landscape and the couplings encoded during training<ref:2604.22997#pg2> HyperEvoGen’s approach to ancestral sequence reconstruction is also distinct from conventional methods<ref:2604. p. 922997#pg2> As divergence increases and the assumptions of those models are progressively violated by multiple hits and epistasis, HyperEvoGen’s geodesic ASR overtakes ML, yielding lower MRCA reconstruction error and better root sequence recovery across a variety of tree topologies<ref:2604.22997#pg2> The superior performance of HyperEvoGen for deep ancestors suggests that this geometric picture is not merely aesthetic but carries real inferential advantages<ref:2604.22997#pg2>
Limitations and Future Directions
All quantitative evaluations are performed on synthetic data generated from Potts models inferred on a single Pfam family (PF00069) and evolved along Yule trees<ref:2604.22997#pg2> The Potts model itself, inferred from a curated alignment, is an approximation to the true fitness landscape and may omit higher-order interactions or rare but functionally important states<ref:2604.22997#pg2> Several design choices in HyperEvoGen are currently heuristic<ref:2604. p. 022997#pg2> A more systematic exploration of these hyperparameters, or approaches that learn them directly from data (for example, by optimizing clade recovery or ASR accuracy on held-out simulations), could further improve robustness and generality<ref:2604.22997#pg2> Incorporating Bayesian treatments of latent positions, sampling over geodesic paths, or integrating branch length information could yield richer estimates with credible intervals rather than single-point reconstructions<ref:2604.22997#pg2> HyperEvoGen represents, to our knowledge, the first integrated framework that deploys a hyperbolic generative model with biologically informed losses for simultaneous sequence generation, phylogenetic inference, and ancestral reconstruction<ref:2604. p. 322997#pg2> The paper is 5 pages long and contains 20 references. p. 0
--- Page 1 ---
HyperEvoGen: Exploring deep phylogeny using non-Euclidean variational inference Jason Lamanna1,4†, Erfan Mowlaei2†, Xinghua Shi1,2, Sudhir Kumar1,5, Vincenzo Carnevale1,4,5,… Abstract Homologous proteins evolve from a common ancestral sequence<ref:2604.22997#pg4> Accurate reconstruction of evolutionary histories remains a challenge<ref:2604. p. 022997#pg4> We present HyperEvoGen, a Poincar´e variational autoencoder with adversarial training, hyperbolic latent geometry, and a compound loss function that learns evolutionarily meaningful representations from single-family alignments<ref:2604. p. 022997#pg4> The arrangement of protein sequences in HyperEvoGen’s hyperbolic embedding aims to preserve phylogenetic structure and produce latent distances which scale with true evolutionary divergence<ref:2604. p. 022997#pg4> HyperEvoGen enables fast, scalable modeling of protein evolution while preserving hierarchical relatedness in a geometry-aware representation<ref:2604. p. 022997#pg4> On Potts-coupled simulation benchmarks, it produces more accurate ancestral reconstructions than conventional baselines, and offers higher-quality sequence generation with less training time than Potts models<ref:2604. p. 022997#pg4>
--- Page 18 ---
[1] Nei, M., Kumar, S.: Molecular Evolution and Phylogenetics. p. 1 Oxford University Press,? p. 0?? (2000). https://books.google.com/books?id=vtWW9bmVd1IC [5] Saitou, N., Nei, M.: The neighbor-joining method: a new method for reconstructing phylogenetic trees. p. 7 Molecular Biology and Evolution 4(4), 406–425 (1987) https://doi. p. 0org/10.1093/oxfordjournals.molbev.a040454 [6] Liberles, D.A.: Ancestral Sequence Reconstruction. p. 0 OUP Oxford,??? (2007). Google-Books-ID: G3YTDAAAQBAJ
--- Page 19 ---
[7] Levy, R.M., Haldane, A., Flynn, W.F.: Potts Hamiltonian models of protein covariation, free energy landscapes, and evolutionary fitness. p. 7 Current opinion in structural biology 43, 55–62 (2017) https://doi. p. 0org/10.1016/j.sbi.2016.11.004 [8] Hopf, T.A., Ingraham, J.B., Poelwijk, F.J., Sch¨arfe, C.P.I., Springer, M., Sander, C., Marks, D. p. 8S.: Mutation effects predicted from sequence co-variation. p. 1 Nature Biotechnology 35(2), 128–135 (2017) [9] McGee, F., Hauri, S., Novinger, Q., Vucetic, S., Levy, R.M., Carnevale, V. p. 0, Haldane, A.: The generative capacity of probabilistic protein sequence models. p. 3 Nature Communications 12(1), 6302 (2021) [10] Kingma, D.P., Welling, M.: Auto-Encoding Variational Bayes. p. 0 arXiv:1312. p. 06114 [stat] (2022).
--- Page 5 ---
Fig. 3 Here we illustrate the distance-correlation term in the compound loss<ref:2604. p. 622997#pg4> The top panels show examples of possible outcomes: a “Well Correlated” case and a “Poorly Correlated” case<ref:2604. p. 622997#pg4>
--- Page 11 ---
Fig. 6 (a) Ground-truth phylogeny used for benchmarking, shown as a radial tree with the four major clades highlighted by color (Groups 1–4). p. 8 (b) Lineage-through-time (LTT) curves computed from the HyperEvoGen-inferred tree, utilizing a mixture between p-distance and embedding distance and plotted as the number of extant lineages versus the number of splits from the root for each clade<ref:2604. p. 822997#pg4> Agreement with panel (c) indicates accurate recovery of the tempo of diversification, shown more by HyperEvoGen in (b) than IQTree in (d)<ref:2604. p. 022997#pg4>
--- Page 15 ---
Fig. 7 We evaluate ancestral sequence reconstruction performance at the most recent common ancestor (MRCA) for pairs of tips binned by their observed tip–tip sequence divergence (p-dist)<ref:2604. p. 922997#pg4> HyperEvoGen maintains high root reconstruction accuracy when given either the ground truth tree or its own embedding tree, and substantially outperforms IQTree when both use the IQTree topology<ref:2604. p. 822997#pg4>
--- Page 16 ---
Hamiltonian energy difference. HyperEvoGen displays a characteristic crossover: inferior accuracy at very low divergence but consistently superior accuracy at moderate to high divergence, overtaking ML at 0. p. 42–0.3 p-distance (Fig. p. 3 p. 3 8)<ref:2604. p. 0 p. 022997#pg4>
--- Page 17 ---
Conceptually, these results align with the intuition behind epistasis and path dependence in protein evolution<ref:2604.22997#pg4> By training on data generated from Potts models that explicitly encode pairwise couplings, HyperEvoGen learns a latent manifold that reflects a more rugged and constrained landscape<ref:2604.22997#pg4> Interpolating along geodesics in this manifold amounts to traversing these constrained regions rather than averaging over incompatible combinations of residues<ref:2604.22997#pg4> The superior performance of HyperEvoGen for deep ancestors suggests that this geometric picture is not merely aesthetic but carries real inferential advantages<ref:2604.22997#pg4>
--- Page 14 ---
Fig. 4 Average pairwise distance is plotted as a function substitutions per site for both Hamming/pdistance (blue) and the model’s latent-space geodesic distance (red)<ref:2604. p. 822997#pg4> As divergence increases, the geodesic distanced metric continues to grow where Hamming distance begins to saturate<ref:2604. p. 822997#pg4>
--- Page 13 ---
The resulting trees are more than summaries of relatedness<ref:2604.22997#pg4> Phylogenetic structure provides the framework through which protein families are organized, functional diversification is contextualized, and hypotheses about sequence change are made testable<ref:2604. p. 122997#pg4> It enables inference about how lineages have expanded, when major molecular innovations emerged, and how present-day variation reflects ancestral constraints<ref:2604. p. 122997#pg4>
--- Page 13 ---
One of the uses of this framework is ancestral sequence reconstruction (ASR), which leverages phylogenetic relationships to infer ancestral proteins and thereby offers an experimental window into the historical emergence of protein structure and function<ref:2604.22997#pg4> By resurrecting ancestor proteins, ASR can connect comparative sequence analysis to mechanistic questions about stability, specificity, catalysis, and adaptation<ref:2604. p. 122997#pg4>
--- Page 15 ---
This term drives the encoder to place similar sequences close together in the hyperbolic space and more divergent sequences farther apart, up to the point where raw sequence distances become saturated<ref:2604.22997#pg4> This term drives the encoder to place similar sequences close together in the hyperbolic space and more divergent sequences farther apart, up to the point where raw sequence distances become saturated<ref:2604. p. 6 p. 622997#pg4>
--- Page 11 ---
This approach to distance computation in the latent space is critical for capturing hierarchical relationships and ensures that the model’s continuous embeddings serve as a biologically meaningful metric for evolutionary divergence<ref:2604.22997#pg4> By using hyperbolic geodesic distances, HyperEvoGen’s embeddings maintain a nearly linear correspondence with true evolutionary distances over a broad range of divergences, mitigating the saturation problem that afflicts simple sequence dissimilarity measures<ref:2604. p. 7 p. 722997#pg4>
--- Page 10 ---
By optimizing this combined loss, HyperEvoGen learns a latent representation that preserves evolutionary relationships and sequence realism simultaneously<ref:2604.22997#pg4> By optimizing this combined loss, HyperEvoGen learns a latent representation that preserves evolutionary relationships and sequence realism simultaneously<ref:2604. p. 622997#pg4>
--- Page 11 ---
HyperEvoGen displays a characteristic crossover: inferior accuracy at very low divergence but consistently superior accuracy at moderate to high divergence, overtaking ML at 0.2–0.3 p-distance (Fig. 8)<ref:2604.22997#pg4> This pattern held across replicated simulations and aligns with the intuition that model-based ML excels when per-site signals are unambiguous, whereas the geometry-guided, epistasis-aware latent interpolation is more resilient when multiple hits and context-dependence degrade independent-site models<ref:2604.22997#pg4>
--- Page 11 ---
At very shallow divergence, where local site-wise signal is strong and independent-site models are well behaved, maximum-likelihood ASR retains a slight advantage<ref:2604.22997#pg4> However, as divergence increases and the assumptions of those models are progressively violated by multiple hits and epistasis, HyperEvoGen’s geodesic ASR overtakes ML, yielding lower MRCA reconstruction error and better root sequence recovery across a variety of tree topologies<ref:2604.22997#pg4>
--- Page 15 ---
This approach to distance computation in the latent space is critical for capturing hierarchical relationships and ensures that the model’s continuous embeddings serve as a biologically meaningful metric for evolutionary divergence<ref:2604.22997#pg4> By using hyperbolic geodesic distances, HyperEvoGen’s embeddings maintain a nearly linear correspondence with true evolutionary distances over a broad range of divergences, mitigating the saturation problem that afflicts simple sequence dissimilarity measures<ref:2604.22997#pg4>
--- Page 15 ---
This term drives the encoder to place similar sequences close together in the hyperbolic space and more divergent sequences farther apart, up to the point where raw sequence distances become saturated<ref:2604.22997#pg4> This term drives the encoder to place similar sequences close together in the hyperbolic space and more divergent sequences farther apart, up to the point where raw sequence distances become saturated<ref:2604.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
HyperEvoGen can perform fast, scalable modeling of protein evolution while preserving hierarchical relatedness in a geometry-aware representation. This enables large-family evolutionary studies and accelerates design-oriented applications by producing more accurate ancestral reconstructions than conventional baselines on Potts-coupled simulation benchmarks and offers higher quality sequence generation with less training time than Potts models.
-
The system can learn an evolutionary geometry of protein family space by embedding sequences in a Poincaré ball and training with a compound loss function that includes a covariance matching loss to capture epistatic dependencies between residue positions. This allows the model to learn
epistatic patterns rather than memorizing or collapsing into an overly smooth consensus.
-
The system can perform distance-based phylogenetic inference by treating learned pairwise distances as an evolutionary distance matrix and applying Neighbor-Joining to generate trees, which are then compared against traditional trees built from sequence-based distances. This approach maintains a
nearly linear correspondence with true evolutionary distances over a broad range of divergences, mitigating the saturation problem.
-
The system can reconstruct ancestral sequences using a generative decoding problem in the continuous latent space rather than per-site inference. This procedure
implicitly respects the global structure of the landscape and the couplings encoded during training,
leading to superior accuracy at moderate to deep divergences compared to maximum-likelihood methods, especially when assumptions of independent-site models are violated.
Sources
- Auto-Encoding Variational Bayes
- Poincar'e Embeddings for Learning Hierarchical Representations
- Continuous Hierarchical Representations with Poincar'e Variational Auto-Encoders
- Generative Adversarial Networks
- Ancestral Sequence Reconstruction for Co-evolutionary models
Related papers
- ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins
- CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating
- Probing and steering biology across Boltz-1s trunk-diffusion boundary
- Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data
- Bridging the Modality Bottleneck in Pathology MIL through Molecularly Guided Feature Adaptation
- Molecular notation can act as a shortcut in AI benchmark evaluation