Rice Mitochondria Optimizer v2

A CodonTransformer-style BigBird model fine-tuned to codon-optimize protein sequences for expression from rice (Oryza sativa) mitochondrial codon usage. Given a protein sequence, it predicts a DNA coding sequence.

This is the second mitochondria model and the "final model" of the project. The first version, Rice_Mitochondria_Optimizer, was trained on rice mitochondrial genomes cut into many small ~6 kb sections. This version was trained on whole genes instead: each training sequence is a complete coding sequence with a start and a stop codon. Fewer, higher-quality sequences were preferred over many small segments.

Training

• Data: Complete rice mitochondrial coding sequences (start codon, single terminal stop codon, no internal stops), 69 genes.
• GC control: GC control was used during fine-tuning but deliberately relaxed in favor of 5′ MFE. GC content and MFE both relate to RNA stability, but MFE measures stability more directly. As a result, the output GC is somewhat below the 43.5% target while the 5′ RNA structure is much better preserved.

Usage

import torch
from transformers import AutoTokenizer, BigBirdForMaskedLM
from CodonTransformer.CodonPrediction import predict_dna_sequence  # from the project repo

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = AutoTokenizer.from_pretrained("gupta-reyansh123/Rice_Mitochondria_Optimizer_v2")
model = BigBirdForMaskedLM.from_pretrained("gupta-reyansh123/Rice_Mitochondria_Optimizer_v2").to(device)

output = predict_dna_sequence(
    protein="MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLA",
    organism=0,
    device=device,
    tokenizer=tokenizer,
    model=model,
)
print(output.predicted_dna)

The model has 2048 positions and a single token type; organism=0 is the only organism it was trained on.

Evaluation

Evaluated on 100 native nifH genes (the test protein set), compared with Geneious, IDT, and two earlier fine-tuned models. The full table is in dataset_comparison_summary.csv.

Metric notes:

• CAI uses relative-adaptiveness weights built from 69 complete rice mitochondrial genes, scored as a geometric mean per sequence.
• Δ5′ MFE is the average change in minimum free energy of the first 50 nucleotides (ViennaRNA, 37 °C), optimized minus original. 0 means the native 5′ structure is preserved.
• Codon similarity is the codon-level similarity to the native sequence.

All 100 predictions translate back to the input protein. The evaluation uses a single gene family (nifH), so these numbers should not be read as a general benchmark.

Limitations

• Trained on rice mitochondrial data only; not intended for other organisms.
• Evaluated on one gene family and a small test set.
• GC and GC3 of the output are below the 43.5% and 40.1% targets. This is the intended trade-off from relaxing GC control in favor of 5′ MFE.

Code and evaluation scripts: see the project repository.