austintw/ecoli_protein2cds_codopt

E. coli Protein → CDS Dataset for training Seq2Seq models

Description

This dataset contains coding sequences from Escherichia coli K-12 MG1655 (RefSeq: GCF_000005845.2). Each example pairs an amino acid sequence with its corresponding coding DNA sequence (CDS), for training Seq2Seq models that map amino acid sequence to CDS that codon usage mirrors native E. coli.

The amino acid sequences:

• Are derived from RefSeq protein annotations.
• Have CDSs that:
  • Are divisible by 3 in length,
  • Begin with ATG (filtered out alternative start codons such as GTG and TTG)
  • End with a single in-frame stop codon (TAA/TAG/TGA),
  • Contain no internal in-frame stop codons.

To prevent data leakage between the training and test sets, protein sequences were clustered with MMseqs2 at 90% sequence identity and 90% coverage, and only cluster representatives that pass the QC filters are included in this dataset.

To ensure that input–output sequence lengths remain ≤ 1024 characters:

• Protein sequences are restricted to those with length ≤ 340 amino acids.
• Note that an underscore _ is appended to each amino acid sequence as a stop token.

Dataset Structure

Each example contains:

• src: amino acid sequence (string) including a trailing _ stop token.
• tgt: CDS nucleotide sequence (string) including the stop codon.

Splits:

• train
• validation
• test

All splits contain only ASCII strings and have sequence lengths ≤ 1024 characters.

Intended Use

This dataset is intended for experiments on Seq2Seq codon optimization models, mapping protein sequences to E. coli coding DNA sequences.