MitoSeqGen — Mitochondrial Codon-Aware Sequence Generator

MitoSeqGen is a sequence-to-sequence Transformer that generates mitochondrial coding sequences (CDS) from an input amino-acid (protein) sequence, respecting the vertebrate mitochondrial genetic code and codon usage patterns learned from mitochondrial genomes.

This checkpoint is the cross-entropy baseline model (referred to as "Model 1" in the project's internal comparisons) — of the variants evaluated so far it has the best BLEU / mt-CAI tradeoff and the lowest GC-content drift from natural sequences.

• Repo / training code: https://github.com/Maheshbonthada/MItoDNA

Model details

• Architecture: encoder-decoder Transformer (MitoSeqTransformer), amino-acid sequence in, codon sequence out.
• Parameters: ~25.2M
• d_model: 384 · heads: 6 · encoder/decoder layers: 6 each · feedforward dim: 1536 · dropout: 0.1 · max position embeddings: 768
• Training objective: token-level cross-entropy over the codon vocabulary
• Training data: vertebrate mitochondrial genome CDS records, QC-filtered, deduplicated, and split by phylogeny into train/val/test (see the training repo's data/ and src/data/ for the pipeline)
• Epochs trained: 30 (this checkpoint is the best-validation-loss snapshot, epoch 20)
• Final validation loss: 0.901 (train loss 0.940 at epoch 30)
• Hardware used for training: single NVIDIA RTX 3050 (8.6GB VRAM), bf16 mixed precision

Evaluation (n=5,155 held-out test proteins, full test set)

Metric	MitoSeqGen (this model)
mean mt-CAI	0.866
genetic-code compliance rate	1.000
mean BLEU-4 vs. natural CDS	0.313
mean GC-content deviation from natural	0.040
mean MFE deviation from natural (kcal/mol)	36.07
novel-sequence rate	1.000

mt-CAI = mitochondrial codon adaptation index; MFE = minimum free energy (RNA secondary structure, via ViennaRNA). Compared against lookup-table, random-synonymous, most-frequent-codon, and CodonTransformer-remap baselines in the source repo's evaluation reports.

Files

• pytorch_model.pt — inference-only checkpoint: {"model_state_dict", "config", "epoch", "val_loss"}. Optimizer/scheduler state was stripped (not needed for inference); this is not a drop-in replacement for resuming training.
• config.json — the full training config (data paths, model hyperparameters, training hyperparameters, hardware settings) for this run.

Usage

Requires the MitoSeqTransformer class and vocabularies from the training repo (src/models/transformer.py, src/genetic_codes.py). This checkpoint does not include a HF transformers-compatible wrapper — load it directly with PyTorch:

import torch
from src.models.transformer import MitoSeqTransformer
from src.genetic_codes import AA_VOCAB, VOCAB  # from the training repo

ckpt = torch.load("pytorch_model.pt", map_location="cpu")
cfg = ckpt["config"]["model"]

model = MitoSeqTransformer(
    src_vocab_size=len(AA_VOCAB),
    tgt_vocab_size=len(VOCAB),
    d_model=cfg["d_model"],
    nhead=cfg["nhead"],
    num_encoder_layers=cfg["num_encoder_layers"],
    num_decoder_layers=cfg["num_decoder_layers"],
    dim_feedforward=cfg["dim_feedforward"],
    dropout=cfg["dropout"],
    max_position_embeddings=cfg["max_position_embeddings"],
)
model.load_state_dict(ckpt["model_state_dict"])
model.eval()

# then use src.models.generate.generate_cds(model, protein_sequence, device="cpu")

Limitations

• Research checkpoint, not benchmarked against a large external test set.
• Two multi-objective variants (GC-content / mt-CAI regularized) were trained alongside this baseline; one regressed on held-out evaluation and a fourth (synonym-class-restricted) has not yet been fully evaluated. This baseline was selected as the best-performing checkpoint among those evaluated so far, not necessarily the final model for the project.