HyphAeon Logo
The Phylogenetic Foundation Model

Evolution.
Amortized.

Natural selection is just the beginning.
HyphAeon amortizes episodic diversifying selection in milliseconds—and then goes further. From the learned geometry of deep time emerge 3D epistatic contact networks, digital mutational scans, and directional phenotype attribution.

>1,000×
Faster Selection Scans
Sub-millisecond inference per codon site across dense trees.
4 Dimensions
Beyond Classical Selection
Epistatic sectors, artifact repair, phenotype attribution & velocity.
17,186
Mammalian Genes
Pre-trained on 9.77M codons from 742 chromosome-level assemblies.
0.942
MEME Parity (ROC-AUC)
25.8× precision-recall lift with rank concordance $\rho = 0.983$.
Act I — The Scalability Barrier

When Classical CTMCs
Can't Keep Up.

For three decades, continuous-time Markov models (CTMCs) served as the gold standard of evolutionary biology. But as comparative genomics expanded from dozens of sequences to hundreds of chromosome-level species and millions of pathogen genomes, numerical likelihood optimization hit an inescapable computational and dimensional bottleneck.

The Computational Complexity Scaling
$$\text{Classical Numerical Likelihood: } \mathcal{O}\left(L \cdot M \cdot K^3\right) \quad \gg \quad \text{HyphAeon Neural Amortization: } \mathcal{O}\left(L \cdot M \cdot d\right)$$

Where $L$ is sequence length, $M$ is species lineages, $K = 61$ sense codons ($K^3 = 226,981$ transition matrix operations per site-step), and $d = 384$ partitioned latent dimensions.

The Classical CTMC Bottleneck

Numerical Likelihood

  • ×
    The Scaling Barrier: Evaluating $\mathcal{O}(L \cdot M \cdot K^3)$ transition matrices across hundreds of species requires tens of thousands of CPU core-hours for a single mammalian proteome scan.
  • ×
    The Site-Independence Assumption: Codon sites are forced to evolve independently, leaving classical models blind to 3D macromolecular contacts and epistatic interactions.
  • ×
    Alignment Noise Sensitivity: A single alignment frameshift or sequencing error in one taxon triggers spurious false-positive selection bursts across whole genes.
  • ×
    Limited to Static Historical Rates: Infers time-averaged rates rather than tracking active temporal sweep velocities or attributing continuous organismal traits.
The New Foundation

HyphAeon Geometric Foundation Model

  • Amortized Inference: Replaces numerical matrix tuning with a single forward pass (<1 ms per site, >1,000× speedup per locus, >10,000× proteome-wide).
  • Emergent Epistatic Contacts: Cross-site axial attention correlations ($\mathbf{K}^{\text{APC}}$) recover physical 3D contacts and multi-residue functional sectors (CESI).
  • Surgical Counterfactual Masking: Pinpoints single-taxon sequencing errors via $\Delta\text{LRT}$ attribution, masking corrupted codons without discarding valuable biological sites.
  • Multi-Dimensional Latent Space: Directly projects continuous and discrete traits ($\bar{\Psi}$) into lineage space and tracks real-time temporal sweep velocities.
Act II — The Reveal

The Architecture of
Deep Time.

A compact 1.91-million-parameter backbone that internalizes the dual geometry of comparative genomics: multiple sequence alignments paired with continuous phylogenetic metric spaces.

Dual-Geometry Continuous Markov Attention Kernel
$$\mathbf{Z} = \operatorname{MDS}_4(\mathbf{D}) \in \mathbb{R}^{M \times 4}, \qquad \mathbf{A}_{ij} = \operatorname{Softmax}\left(\frac{\mathbf{Q}_i \mathbf{K}_j^T}{\sqrt{d_k}} - \lambda D_{ij}\right)$$

Patristic tree distances $D_{ij}$ are projected into 4D Classical Metric MDS space $\mathbf{Z}$, while cross-species attention is modulated by an analytical Markov substitution kernel $e^{-\lambda D_{ij}}$, guaranteeing topological invariance to phylogenetic noise.

Step 01 Dual Input
Alignment + Phylogeny
Codon sequence columns across $M$ lineages paired with continuous metric trees $(\mathcal{S}, d_{\mathcal{T}})$ or patristic distance matrix $\mathbf{D} \in \mathbb{R}^{M \times M}$.
$\mathbf{X} \in \{1,\dots,61\}^{M \times L}$
Step 02 Geometric Embedding
4D MDS & Markov Kernel
Projects patristic tree distances into 4D Classical Metric MDS space $\mathbf{Z} \in \mathbb{R}^{M \times 4}$, while cross-taxon attention is decayed via analytical Markov substitution kernel $e^{-\lambda D_{ij}}$.
$\mathbf{A}_{ij} \propto \frac{\mathbf{Q}_i \mathbf{K}_j^T}{\sqrt{d}} - \lambda D_{ij}$
Step 03 1.91M Backbone
Partitioned dN / dS Tracks
Synonymous ($dS$) and non-synonymous ($dN$) evolutionary signals are isolated into dual 192-dimensional tracks via BlockLinear projections, completely insulating positive selection from neutral rate variation.
$\mathbf{H} = [\mathbf{H}^{(dN)} \parallel \mathbf{H}^{(dS)}] \in \mathbb{R}^{384}$
Step 04 Zero-Shot
Multi-Task Emergence
A single forward pass yields episodic selection ($\widehat{\text{LRT}}_s$), epistatic contact networks ($\mathbf{K}^{\text{APC}}$), single-taxon counterfactual masking ($\Delta\text{LRT}$), and selective velocity ($v_s$).
$\widehat{\text{LRT}}_s, \, \mathbf{K}_{ij}^{\text{APC}}, \, \Delta\text{LRT}, \, v_s(t)$

Dual-Geometry Tensor

Codon sequence columns across $M$ lineages paired with a continuous metric tree $(\mathcal{S}, d_{\mathcal{T}})$ or tree-free empirical pairwise distance matrix $\mathbf{D} \in \mathbb{R}^{M \times M}$.

4D Metric MDS Space

Projects continuous patristic tree distances into 4-dimensional Classical Multidimensional Scaling coordinates $\mathbf{Z} \in \mathbb{R}^{M \times 4}$, providing natural robustness against phylogenetic noise and incomplete lineage sorting (ILS).

Partitioned dN / dS Tracks

Synonymous ($dS$) and non-synonymous ($dN$) signals are isolated into dual 192-dimensional tracks via BlockLinear projections, completely insulating positive selection from neutral synonymous rate variation.

Radical Efficiency • Pure Browser Inference

You Don't Have to Be Big
To Succeed.

Modern machine learning often assumes that bigger is always better. HyphAeon takes the opposite path: by embedding the exact geometry of evolutionary divergence into transformer attention, it achieves state-of-the-art inference in just 1.91 million parameters—and was trained from scratch for <$200 in Google Cloud credits.

Inductive Bias > Parameter Scale

1.91M Parameters. That's not a typo.

General protein language models rely on massive parameter counts to learn statistical associations across unaligned sequences. But molecular evolution has rigorous mathematical structure that shouldn't have to be learned by brute force.

HyphAeon embeds tree metric geometry (4D Classical MDS), continuous-time Markov substitution kernels ($e^{-\lambda D_{ij}}$), and strict synonymous/non-synonymous block-diagonal linear projections directly into its attention layers.

The result? A complete foundation model that fits in a 7.6 MB checkpoint and required <$200 in Google Cloud compute credits to train from scratch on commodity hardware. True democratized foundation modeling without the supercomputing budget.

1.91M Parameters
7.6 MB Model Weight
<$200 Training (GCP)
0 Cloud At Inference
Zero Install • 100% Client-Side

Run Live in Your Browser at primaeon.org

WebAssembly & WebGPU • Private & Instantaneous

Because HyphAeon is on a strict architectural diet, you don't need a compute cluster, a cloud subscription, or a complex Python environment to use it.

At primaeon.org, the entire neural inference engine executes client-side inside your browser tab. Paste your multi-sequence alignment and phylogenetic tree, and get selection scans and epistatic sectors in seconds.

🔒 Zero Cloud Upload: Your sequences never leave your device. Safe for sensitive pathogen surveillance and clinical cohorts.
Sub-Second Scans: Instant WebAssembly/ONNX inference across hundreds of species.
Zero Waiting: Get selection results before your espresso finishes dripping.
Launch Primaeon Browser App →
Act III — The Implementation

Complexity,
Distilled.

What once required a compute cluster allocation, Slurm array scripts, and days of queue waiting now runs natively with a single interactive CLI command.

> hyphaeon phenotype -a rh1_vertebrates.fa -t species.nwk --dim-light
[1/3] Ingesting 38 vertebrate rhodopsin coding sequences ($L = 330$ codons)... Done (0.04s)
[2/3] Projecting blue-shifted phenotype vector $\hat{\mathbf{y}}$ into 4D metric attention space... Done (0.07s)
[3/3] Evaluating directional attribution ($\rho_s$) across 330 codons at FDR $q \le 0.05$...
    • Site 292 (A→S): $\rho = +0.537$, $q = 0.006$ [Yokoyama $\Delta\lambda = -10\,\text{nm}$, Sector 2]
    • Site 83 (D→N): $\rho = +0.612$, $q = 0.001$ [Yokoyama $\Delta\lambda = -6\,\text{nm}$, Sector 3]
    • Site 277 (T→C): $\rho = +0.846$, $q < 10^{-5}$ [Deep-sea parallel tuning, Sector 2]
✔ Recovered 7 of 9 variable in vitro switches in 0.51 seconds (5.5× PPV lift; PAML found 0).
Act IV — Biological Reality

Where Classical $dN/dS$ Failed:
Three Grounded Realities.

Theory is cheap. Likelihood surfaces are expensive. Here is what happened when we took HyphAeon out of the textbook and tested it against real evolutionary puzzles that broke classical continuous-time Markov chains.

Case Study 01 • Ground-Truth Validation

The Spectral Tuning Switches
PAML Couldn't See.

A foundational debate in molecular evolution centers on whether statistical $dN/dS$ tests can reliably guide experimental functional biology. In a landmark study, Yokoyama et al. (2008) resurrected 11 ancestral vertebrate rhodopsins (RH1) in vitro, proving that dim-light spectral adaptation ($\lambda_{\max} \in [480, 526]\,\text{nm}$) across deep-sea fish and nocturnal vertebrates is driven by 15 specific amino acid switches across 12 sites.

The Classical $dN/dS$ Failure: Standard branch-site models (PAML M2a/M8 Bayes Empirical Bayes) detected 0 sites across the 38-species phylogeny (0.0% PPV). Classical models assume positive selection requires a loud explosion of non-synonymous substitutions. Because nature tuned dim-light vision with surgical parsimony—substituting 1–2 key residues in the retinal pocket while preserving the tightly constrained GPCR scaffold—traditional $dN/dS$ declared the gene functionally inert.

HyphAeon evaluated directional phenotype attribution ($\bar{\Psi}$) in 0.51 seconds on a laptop CPU. Of the 12 experimentally assayed sites, 9 vary across these lineages: HyphAeon recovers 7 of these 9 at FDR $q \le 0.05$ (77.8% sensitivity on variable drivers), achieving a 5.5× PPV enrichment lift over locus baseline (rising to 7.3× lift among the top 15 candidates).

Unsupervised attention covariance ($\mathbf{K}^{\text{APC}}$) clustered these positions into 4 physical epistatic sectors: Sector 2 (Schiff base tuning pocket, $C=0.582$), Sector 1 (extracellular cap/roof, $C=0.630$), Sector 3 (TM-II trigger, $C=0.481$), and Sector 4 (hydrophobic pressure clamp, $C=0.655$).

RH1 In Vitro Ground-Truth Inspector 38 Vertebrates • 0.51s
Site In Vitro $\Delta\lambda$ PAML Result HyphAeon ($\rho_s$) Status
A292S $-10\,\text{nm}$ Blue Missed ($p=0.48$) $\rho = +0.537$ ($q=0.006$) Recovered
D83N $-6\,\text{nm}$ Blue Missed ($p=0.31$) $\rho = +0.612$ ($q=0.001$) Recovered
E122Q $-15\,\text{nm}$ Blue Missed ($p=0.72$) $\rho = +0.584$ ($q=0.003$) Recovered
T277C $-8\,\text{nm}$ Blue Missed ($p=0.29$) $\rho = +0.846$ ($q < 10^{-5}$) Recovered
H278N $-4\,\text{nm}$ Blue Missed ($p=0.64$) $\rho = +0.732$ ($q < 10^{-5}$) Recovered
F261Y $+10\,\text{nm}$ Red Missed ($p=0.55$) $\rho = +0.491$ ($q=0.018$) Recovered
7 / 9 (77.8%)
Variable Drivers Recovered
5.5× Lift
PPV (14.9% vs 2.7% Baseline)
4 Sectors
Schiff, Roof, TM-II, Clamp
0.51 s
Complete CPU Runtime
Case Study 02 • High-Throughput Exome Deployment

15,868 Mammalian Families in 109.86 Minutes on a Laptop.
Without Ingesting Alignment Noise.

Comparative genomic scans across 190 mammalian species ($10,270,293$ codons in OrthoMaM v12) have traditionally required over 50,000 CPU core-hours under numerical continuous-time Markov chains. But the deeper hazard is sequencing and annotation noise: draft assemblies (e.g. platypus, koala, monito del monte with contig N50 $\sim 11.5\text{ kb}$) harbor automated gene-prediction frameshift slips that dump dense runs of spurious non-synonymous mutations into alignments.

The Classical Dilemma: Traditional pipelines face an impossible choice: either run aggressive column trimmers like Gblocks to discard thousands of alignment positions (destroying legitimate evolutionary signal across all 190 species), or let numerical CTMCs run unfiltered, falsely declaring that draft marsupials and monotremes underwent a miracle of positive Darwinian selection.

HyphAeon completed whole-database inference across all 15,868 mammalian gene families ($10.27\text{M}$ codons across mean depth $171.8$ species) in 109.86 minutes on a single Apple M5 Max laptop GPU ($1,558\text{ codons/s}$). Furthermore, evaluating each family across 5 major mammalian subtrees (Primates, Chiroptera, Cetartiodactyla, Carnivora, Rodentia) completed in minutes without tree re-optimization.

Its automated single-taxon counterfactual error filter identified and surgically masked 72,832 localized artifact tracts (mean span 33.5 codons, matching automated single-exon frameshifts). It masked only the corrupted codons in that single draft taxon, keeping the other 189 mammals intact. Spurious sitewise likelihood spikes collapsed from an artificial $\text{LRT} = 14.8$ back to baseline $0.12$, with zero columns dropped.

OrthoMaM Deployment Audit 109.86 Min • Apple M5 Max GPU
109.86 min
Single Laptop GPU (vs >50k CPU hrs)
15,868
Gene Families (10.27M Codons)
72,832
Frameshift Patches Masked
14.8 → 0.12
Spurious LRT Collapse (0 Cols Lost)
Multi-Order Adaptive Breadth
Core Pan-Mammalian: 37.7%
Multi-Order Shared: 40.5%
Dual-Order Clades: 9.6%
Order-Private Burst: 7.2%
Figure 4 from Paper • Simplified Web Version

The Selection Distribution Landscape Across 15,868 Mammalian Gene Families

Decomposing continuous selection density, subcellular compartmentalization, physical micro-clustering, and burst decoupling across $10,270,293$ codons.

A Continuous Selection Density Across Functional Systems % Codons $p \le 0.05$
Meiotic Conflict & Centromere
CENPH, KNSTRN, DSN1
10.6%
Reproduction & Gametes
ZP3, SEMG1, CRISP
9.9%
Innate Immunity & Defense
OAS1, TRIM5, BST2
9.2%
Xenobiotic & Lipid Metabolism
CYP, APO, FADS
8.9%
Mitochondrial OXPHOS
NDUF, COX, ATP5
7.4%
Neuronal Synapse & Channels
GRIN, GABR, SCN
6.2%
Cell Surface Adhesion
ITGA, CAD, EPH
5.6%
Chromatin & Epigenetics
HIST, HNRNP, HDAC
4.2%
0% 2% 4% 6% 8% 10% 12%
Key Insight: Red Queen genetic conflict systems (meiotic drive $+2.48\%$, reproduction $+1.67\%$, viral defense $+1.45\%$) exhibit $>2 imes$ the adaptive flux of core epigenetic machinery ($-1.39\%$).
B Subcellular Compartmentalization Gradient Inward Attenuation
Extracellular & Secreted
Cytokines, ECM, Venom, Defensins
11.2%
Plasma Membrane & Receptors
GPCRs, Viral Entry Gates, Transporters
9.4%
Cytosol & Metabolic Enzymes
Glycolysis, Signaling Kinases
7.8%
Nucleus & Transcription
Transcription Factors, DNA Pol
5.3%
Ribosome Core
40S & 60S Ribonucleoprotein Core
4.1%
0% 2% 4% 6% 8% 10% 12%
Monotonic Outside-In Attenuation: $\Delta = -1.17 ext{ percentage points}$ ($p = 1.6 imes 10^{-5}$) from extracellular boundary to translational core.
Key Insight: Selection pressure decays monotonically toward the cell center, mirroring direct pathogen and environmental exposure.
C Physical Inter-Site Spatial Micro-Clustering Median = 7 Codons
≤ 3 codons: 29.4% (Micro-Cluster) ≤ 10 codons: 63.0% (Single Loop) 0.12 0.06 0.00 0 10 20 30 40 50 Distance Between Selected Sites (codons)
Key Insight: Positive selection concentrates into physical micro-clusters: $63.0\%$ of adaptive codons sit within $\le 10$ codons on solvent-exposed secondary loops. Only $4.5\%$ are separated by $>50$ codons.
D Peak Episodic Bursts vs. Baseline Constraint 38% Conserved Burst
Conditional Median Max LRT IQR (25th–75th percentile) 30 20 10 0 0 1 2 3 4 Baseline Constraint (Mean LRT across gene)
Key Insight: Strong baseline purifying constraint on a gene does not prevent episodic bursts: over 38% of the most conserved quartile ($ ext{mean LRT} < 0.74$) experience localized peak bursts ($ ext{max LRT} > 10.0$).
Case Study 03 • Pathogen Surveillance

Measuring the Gas Pedal,
Not the Odometer.

During the COVID-19 pandemic, public health genomic surveillance analyzed >9.34 million viral sequences across 78 months ($1,863$ unique haplotypes). Standard surveillance relied on two metrics: mutant allele frequency curves, or static multi-year $dN/dS$ scans. Both methods fail in opposite ways.

The Frequency Lag & Stasis Blindspot: Allele frequency only peaks after the selective sweep has already fixed—you discover a variant had a fitness advantage only when the entire population already has it. Conversely, static 78-month $dN/dS$ scans dilute acute adaptive bursts under years of subsequent neutral stasis (e.g. D614G looks flat after 2020 because its adaptive non-synonymous drive occurred in early 2020).

HyphAeon calculates instantaneous positive selection velocity:

Instantaneous Positive Sweep Velocity
$$v_s(t) = \frac{\partial\,\operatorname{logit}\left(\widehat{\text{LRT}}_s(t)\right)}{\partial t}$$

Detects accelerating adaptive lineage expansion weeks before circulating frequency peaks, and drops to zero ($v_s \to 0$) upon fixation to confirm post-sweep neutral stasis.

Across SARS-CoV-2 Spike ($1,274$ codons), sweep velocity detected the emergence of Delta, Omicron BA.1, and JN.1 6–10 weeks before frequency plateaus. In large clinical cohorts, duplicate haplotype pruning collapses 32,768 field isolates into $\approx 150$ unique trees in memory, evaluating the full Spike glycoprotein in <6 seconds on a laptop CPU.

Spike Sweep Velocity vs Frequency 78 Months • 9.34M Genomes
Variant Site Epidemic Wave Peak Velocity ($v_s$) Lead-Time vs Frequency Post-Fixation
D614G Initial Expansion (2020) $+4.82\,\text{yr}^{-1}$ 8 Weeks Ahead Neutral Stasis ($v_s \to 0$)
N501Y Alpha & Omicron (2020/21) $+6.12 \parallel +7.45$ Bimodal Pre-Peak Recurrent Sweep
F486P XBB.1.5 Sweep (2023) $+5.34\,\text{yr}^{-1}$ 6 Weeks Ahead Immune Evasion
L455S JN.1 Emergence (2024) $+6.88\,\text{yr}^{-1}$ At <2% Frequency Early Detection
6–10 Wks
Lead-Time Before Peak Frequency
< 6.0 s
1,274 Codons on Laptop CPU
32k → 150
Duplicate Haplotype Compression
129 Sweeps
Confirmed Episodic Waves
Figure 13 from Paper • Simplified Web Version

Longitudinal Selective Velocity vs. Allele Frequency Lag in SARS-CoV-2

Decomposing continuous sweep velocity $v_s(t) = \partial\,\operatorname{logit}(\widehat{\text{LRT}})/\partial t$ across 78 months, $>9.34\text{M}$ genomes, and 129 confirmed sweeps.

D614G B.1 Origin
Lead: 2.5 Months Ahead
Peak $v_s$: 0.21×10⁻³
100% 50% 0% 2020 2022 2024 2026
• Cov-Spectrum Freq $f(t)$ • Selection Velocity $v_s(t)$ • Lead Gap: 2.5 mo
L452R Delta Surge
Lead: 3.5 Months Ahead
Peak $v_s$: 0.64×10⁻³
100% 50% 0% 2020 2022 2024 2026
• Cov-Spectrum Freq $f(t)$ • Selection Velocity $v_s(t)$ • Lead Gap: 3.5 mo
F486P XBB.1.5
Lead: 3.8 Months Ahead
Peak $v_s$: 2.92×10⁻³
100% 50% 0% 2020 2022 2024 2026
• Cov-Spectrum Freq $f(t)$ • Selection Velocity $v_s(t)$ • Lead Gap: 3.8 mo
L455S JN.1 "FLip"
Lead: 4.4 Months Ahead
Peak $v_s$: 0.91×10⁻³
100% 50% 0% 2020 2022 2024 2026
• Cov-Spectrum Freq $f(t)$ • Selection Velocity $v_s(t)$ • Lead Gap: 4.4 mo
The Fundamental Distinction: Measuring the Gas Pedal vs. the Odometer

Standard surveillance curves (blue) only confirm an adaptation after it reaches high population prevalence ($>80\%$). In contrast, HyphAeon selective velocity $v_s(t)$ (red) peaks 2.5 to 4.4 months ahead of frequency plateaus, detecting exponential lineage drive when the variant is still at $<2\%$ circulating frequency. Once a sweep fixes, selective velocity collapses back to zero ($v_s \to 0$), accurately distinguishing active positive selection from passive neutral stasis.

Act V — Rigorous Validation

Engineered for
Empirical Rigor.

Trained to inherit the statistical foundations of classical maximum likelihood, validated against continuous-time null simulations, and stress-tested across empirical genomes and experimental mutagenesis.

От HyPhy к HyphAeon: Факел эволюционной науки!
HyPhy passing the flame of rigorous hypothesis testing to HyphAeon — «От HyPhy к HyphAeon: Факел эволюционной науки!»
Inheriting the Flame

Passing the Torch: Trained to Be a Faster HyPhy

For over two decades, Sergei Kosakovsky Pond and Spencer Muse’s HyPhy (Hypothesis Testing using Phylogenies) has anchored statistical molecular evolution, formalizing continuous-time Markov models (CTMC), codon substitution matrices, and landmark selection tests like MEME and BUSTED. But numerical likelihood optimization over hundreds of species is computationally intensive, requiring tens of thousands of CPU hours.

HyphAeon was created not to replace HyPhy's statistical theory, but to inherit its torch and accelerate it. We trained HyphAeon’s neural geometric tensor to directly predict HyPhy-derived site-level Likelihood Ratio Tests (LRT) with continuous numerical fidelity.

By learning the geometric manifold of codon transitions across trees, HyphAeon yields the exact same hypothesis testing power in milliseconds on a single GPU—scaling twenty years of empirical evolutionary rigor to whole-proteome comparative genomics.

01 • Type-I Error Control
Continuous-Time Null Simulations & False Positive Auditing
To guarantee that inference speed does not induce hallucinated selection, we evaluated HyphAeon across 1,400 synthetic alignments (280,000 codons across 14 distinct regimes) generated under classical continuous-time Muse-Gaut 1994 (MG94) codon models. Across neutral drift and purifying selection controls, the model demonstrated tight Type-I error calibration: empirical false positive rates remained well-controlled (mean $\text{FPR}_{0.05} = 3.07\%$ across all regimes; $1.47\%$ under purifying selection and $5.54\%$ under neutral drift at nominal $\alpha = 0.05$).
Calibrated Type-I Error (Mean FPR = 3.07% at α = 0.05)
02 • Likelihood Concordance
Continuous Rank Concordance with Numerical HyPhy
HyphAeon does not output crude binary classifications; it predicts continuous site-level LRT statistics that preserve relative evolutionary rankings. Evaluated on empirical viral benchmarks like SARS-CoV-2 Spike against full numerical HyPhy MEME, HyphAeon achieves a continuous Spearman rank correlation of ρ = 0.983 ($p < 10^{-300}$). At the gene family level, mean selection scores tightly match PAML M0 baseline rates (ρ = 0.8285), demonstrating seamless fidelity with numerical optimization.
ρ = 0.983 continuous rank correlation
03 • Out-of-Distribution Transfer
Cross-Taxon Generalization Across 100M+ Years of Divergence
To test whether the model overfits to specific clade features, we deployed HyphAeon zero-shot across radically diverse evolutionary lineages without retraining. It generalized seamlessly to 39 avian genomes (8,699 gene families, 4.67M codons), deep vertebrate phylogenies (rhodopsin), and viral datasets where entire viral families were held out during training. In every cohort, HyphAeon reliably separated genuine episodic selection from neutral phylogenetic divergence.
Zero-shot transfer to 39 avian genomes
04 • Counterfactual Auditing
Automated Sequencing Artifact & Reference Error Masking
Real comparative alignments are plagued by sequencing errors, pseudogenes, and assembly gaps that routinely trigger false positives in traditional software. We tested 72,832 artifact tracts across mammalian genomes, corroborated against gold-standard PacBio HiFi chromosome-level assemblies. Using single-taxon counterfactual perturbation leaves, HyphAeon surgically isolates spurious spikes (collapsing false LRT from 14.8 to 0.12) without discarding entire alignment columns.
72,832 artifact tracts validated via PacBio HiFi
05 • Wet-Lab Reality Checks
Experimental Deep Mutational Scanning & In Vitro Ground Truth
Statistical inferences were validated against wet-lab functional assays. On 26 human Deep Mutational Scanning (DMS) datasets in ProteinGym (119,000 experimentally measured single-amino-acid variants), HyphAeon scored each variant in just 0.24 ms while matching empirical functional fitness landscapes. In vertebrate rhodopsin, HyphAeon successfully pinpointed 7 out of 9 experimentally proven spectral-tuning switches—sites where traditional time-averaged $dN/dS$ failed completely because episodic bursts were diluted across deep lineages.
119k DMS variants scored • 7/9 spectral switches recovered in rhodopsin
3.07%
Empirical Type-I Error
Mean false positive rate at nominal α = 0.05 across 1,400 synthetic continuous-time MG94 alignments.
ρ = 0.983
Continuous Rank Preservation
Spearman rank concordance with numerical HyPhy MEME on SARS-CoV-2 Spike ($p < 10^{-300}$).
25.8×
Precision-Recall Lift
Selective enrichment over random baseline across 17,186 mammalian coding gene families.
< 1 ms
Latency Per Codon Site
>1,000× faster than numerical CTMC likelihood optimization across hundreds of species.