geestaltt/ternary-codon-encoder

Ternary Codon Encoder: P-adic Hyperbolic Embeddings

The Ternary Codon Encoder is a neural embedding model that maps the 64 genetic codons into a 16-dimensional hyperbolic space. It is the first model to explicitly use 3-adic valuation as a mathematical prior to organize the genetic code's hierarchical structure.

Model Description

• Architecture: MLP-based encoder (12-dim one-hot input $ ightarrow$ 16-dim hyperbolic output).
• Mathematical Foundation: Leverages 3-adic mathematics to represent the discrete hierarchy of the codon table.
• Latent Space: Poincaré ball where radial distance encodes 3-adic valuation (conservation/variability).

Key Discoveries

• Physics Dimension: Latent dimension 13 correlates strongly ($ ho = -0.70$) with molecular mass, volume, and force constants ($k$).
• Linear Stability Manifold: Provides high-quality feature vectors for sequence-only protein stability ($\Delta\Delta G$) prediction.
• Synonymous Cohesion: Synonymous codons cluster together in hyperbolic space while maintaining clear boundaries between amino acid groups.

Performance

• DDG Spearman $ ho$: 0.614 (Sequence-only benchmarking on diverse datasets).
• Improvement: +105% over baseline p-adic embedding models.

Usage

import torch
from trainable_codon_encoder import TrainableCodonEncoder

# Load model
encoder = TrainableCodonEncoder(latent_dim=16, hidden_dim=64)
checkpoint = torch.load("pytorch_model.bin", map_location="cpu")
encoder.load_state_dict(checkpoint["model_state_dict"])
encoder.eval()

# Get embedding for a codon (e.g., ATG index 14)
codon_idx = torch.tensor([14])
with torch.no_grad():
    z_hyp = encoder(codon_idx)

print(f"Hyperbolic Embedding: {z_hyp}")

Citation

@software{ternary_codon_2026,
  author = {AI Whisperers},
  title = {Ternary Codon Encoder: P-adic Hyperbolic Embeddings},
  year = {2026},
  url = {https://huggingface.co/ai-whisperers/ternary-codon-encoder}
}