Loading…
Codon language model for mRNA prediction and coding sequence design, steering synonymous codon choice with a swappable host usage prior at inference.
No providers recorded yet. Browse all providers
Codon optimization tools have historically fixed their preferences at build time. A model trained on human transcripts writes human-like coding sequences; moving the same protein into E. coli means a different model or a hand-tuned objective such as the codon adaptation index. Yet the preference a designer wants is rarely universal — it depends on the expression host, the tissue, the therapeutic format — while the regularities that make a coding sequence viable at all are shared across the tree of life. CodonMamba separates the two: one pretrained model supplies the shared distribution, and the host-specific preference arrives at generation time as a small codon usage table that can be swapped without touching a weight.
CodonMamba was developed at the Hangzhou Institute of Medicine, CAS with the University of Macau and the Hangzhou Institute for Advanced Study of the University of Chinese Academy of Sciences. It is a bidirectional masked language model over non-overlapping codon tokens, built on a state-space rather than an attention backbone, and pretrained on the processed coding-sequence corpus released with CaLM — 8,771,938 sequences from European Nucleotide Archive genomes spanning 1,544 organisms.
Both capabilities run off that one checkpoint: as a frozen feature extractor it feeds a lightweight prediction head across a 12-task mRNA benchmark, and as a generator it performs constrained synonymous decoding, where a hard mask enforces the target protein while a soft prior steers codon choice toward a chosen host.
CodonMamba tokenizes coding sequences into non-overlapping triplets over a vocabulary of the 64 standard codons plus five special tokens, mapped to learned 768-dimensional embeddings. The backbone is eight bidirectional Mamba blocks at hidden dimension 768, roughly 71 million trainable parameters, supporting up to 1,024 codon tokens. Pretraining used masked language modelling with 15% of positions corrupted, optimized with AdamW at a peak learning rate of 4 × 10⁻⁴, and ran 47 epochs — about seven days on four NVIDIA A100 GPUs.
Static codon embeddings organize by encoded amino acid and physicochemical class, with mean cosine similarity of 0.37 between synonymous codon pairs against 0.02 for non-synonymous pairs, despite no amino-acid labels during training. On the CaLM held-out species panel, nearest-centroid classification from sequence-level embeddings reached 91.4% accuracy, ahead of CaLM at 90.0% and mRNA-FM at 86.1%. Generation runs as a 12-step mask-and-predict schedule with greedy selection, and model-derived naturalness — a length-normalized pseudo-log-likelihood — correlated with measured expression and stability across three independent experimental datasets at a mean Pearson r of 0.68.
The design mode targets the practical middle of mRNA and protein production: a therapeutic antigen that must express well in human cells, a research construct destined for yeast, an enzyme produced in bacteria. Because the host preference is a table rather than a training decision, one deployment covers all of them, and a designer can dial the prior's strength to trade host adaptation against fidelity to the learned distribution. Separately, frozen embeddings serve reporter-expression, half-life, degradation and localization models without codon-specific feature engineering.
CodonMamba's contribution is architectural in the design sense rather than the neural-network sense: a generative coding-sequence model can hold its learned biology fixed while exposing application preferences as a runtime input, a pattern that generalizes to any preference expressible as a per-token weight. It also strengthens the case for codon-resolution pretraining, since protein language models — which discard synonymous variation by construction — fell furthest behind on the codon-sensitive tasks. The principal caveat is stated plainly by the authors: the work reports no direct experimental validation, so the gains in codon adaptation, nucleotide composition and predicted structure are computational properties rather than demonstrated improvements in yield or stability. The preprint has not yet been peer reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.