Trias

The genetic code is degenerate: most amino acids can be written with several synonymous codons, so any protein of length n corresponds to an astronomically large family of coding sequences. Those sequences are not interchangeable. Codon choice shapes translation elongation speed, mRNA secondary structure, decay rate, and ultimately protein yield — which matters directly now that mRNA vaccines and therapeutics have made coding-sequence design a production problem rather than an academic one. The tools most commonly used to make that choice are simple: pick the most frequent codon per amino acid in the target species, or sample from a codon usage table. Both discard the context in which a codon sits.

Trias, from Liana Lareau's lab at UC Berkeley and the Chan Zuckerberg Biohub, treats reverse translation as a sequence-to-sequence generation problem instead. A protein sequence plus a species tag goes in; a codon sequence written in that species' idiom comes out. Because the decoder is autoregressive and attends to the encoder's representation of the whole protein, the codon it emits at each position depends on local sequence context and on global properties of the transcript, not only on the amino acid being encoded.

The notable result is what Trias learned without being told. It was trained purely on natural coding sequences, with no expression data of any kind in the objective, yet the sequences and likelihood scores it produces correlate strongly with experimental measurements of mRNA stability, ribosome load, and protein output — and it generates sequences that more closely resemble high-expression variants than commercial codon optimization tools do.

#Key Features

• Species-conditioned generation: A species token prefixed to the protein input (>>Homo sapiens<< MTEIT…) selects the codon idiom, letting one checkpoint write sequences for any of the 640 species represented in training without retraining or a per-species usage table.
• Context-aware codon choice: The encoder-decoder attention lets codon selection depend on surrounding sequence, capturing the local and global dependencies that frequency-table optimization cannot represent.
• Zero-shot correlation with expression: Generated sequences and their model scores track experimental mRNA stability, ribosome load, and protein output despite no expression measurements appearing in training.
• Three decoding modes: Greedy decoding for fast deterministic output, beam search for longer or harder proteins, and nucleus sampling with a settable seed for generating diverse candidate libraries.
• Small and openly released: At 46.6M parameters the checkpoint runs on CPU, and code and weights are MIT licensed on GitHub and HuggingFace.

#Technical Details

Trias is a BART-style encoder-decoder transformer with 6 encoder and 6 decoder layers, a model dimension of 512, 8 attention heads per layer, feed-forward dimension 2048, and a 2048-position context — 46.6M parameters in total. Encoder and decoder share a 729-token vocabulary covering amino acids, codons, per-species tags, and control tokens. Training used 10 million eukaryotic coding sequences spanning 640 vertebrate species. The released checkpoint is served through the standard transformers API as a BartForConditionalGeneration model, so inference is a from_pretrained call followed by generate; FlashAttention-2 is used for training but scaled dot-product attention is available for CPU inference.

#Applications

The direct use is designing coding sequences for synthetic mRNA — vaccines, mRNA therapeutics, and gene-replacement constructs — where output is sensitive to codon composition and where the target species may not be human. It also serves recombinant protein production in heterologous eukaryotic hosts, and, because the model assigns likelihoods to natural sequences, it can be used analytically to ask which regions of a native transcript carry unusual codon choices. Related catalog entries approach the same problem from different angles: CodonTranslator conditions on taxonomic lineage, CodonFM learns codon-resolution representations across 20,000 species, and SynCodonLM uses synonymous-codon-constrained masking.

#Impact

Trias is a bioRxiv preprint and has not yet completed peer review, and its training corpus is vertebrate-restricted, so codon idioms for plants, fungi, and invertebrate hosts are outside its demonstrated range. Within that scope it is unusually well-packaged for a research model: a small permissively licensed checkpoint on HuggingFace, a working inference path in a few lines, and a benchmarking notebook that reproduces comparisons against other codon models. The zero-shot agreement with stability and ribosome-load measurements is the substantive claim — evidence that the constraints shaping natural codon usage and the constraints governing expression are close enough that learning one recovers much of the other.