ppLM-CO

Any protein can be written as an enormous number of synonymous coding sequences (CDSs), and the choice among them is a design lever rather than a detail. Codon usage shapes translation speed, tRNA demand, and the secondary structure of the transcript, so the same antigen encoded two different ways can differ substantially in how much protein a host cell makes and how long the message survives — which for mRNA vaccines and recombinant biologics is the difference between a potent construct and a weak one.

Deep learning recast codon optimization as a sequence problem — tag each amino acid with its best codon, or translate protein into nucleotides — but most such models train a full encoder from scratch, spending millions of parameters relearning the protein context that pretrained protein language models already capture. ppLM-CO, from Shashank Pathak and Guohui Lin at the University of Alberta, inverts that. It freezes the ProtBert encoder from the ProtTrans suite outright, trains a single linear classifier on its per-residue embeddings, and masks the output layer so only codons synonymous with the input residue can be chosen. The entire trainable component is 62,000 parameters.

The work was first posted in December 2024 and revised through October 2025, and the peer-reviewed version appeared at the ACM Conference on Bioinformatics, Computational Biology and Health Informatics in 2026. Earlier preprint versions described adapter-based fine-tuning under the names Adaptive-ProtBert and Adasel-ProtBert — with host- and objective-specific variants labelled -short, -E.coli, -CHO, -long-mfe and -random-mfe — before the method was simplified to a frozen backbone with a masked classification head. Those are alternative names for this model, not separate releases.

#Key Features

• Frozen language model encoder: ProtBert supplies 1,024-dimensional contextual embeddings for each residue and is never updated, so amino-acid representations are inherited rather than learned.
• Biological masking: Logits for codons that do not encode the residue at a given position are set to negative infinity before the softmax, making every designed CDS reverse-translate to the input protein by construction.
• A 62K-parameter trainable head: One linear layer maps each embedding onto the 61 sense codons — a 92% reduction in trainable parameters against Codon-Box and 99% against CodonBERT.
• Per-host instances: Separate classifiers are trained for human, Escherichia coli, and Chinese hamster coding sequences, so host-specific codon preferences live in the head while the encoder stays shared.
• Runnable release: The implementation ships with trained checkpoints for all three hosts and a Gradio interface that optimizes a pasted protein sequence and reports CAI, MFE, and GC content.

#Technical Details

Codon optimization is posed as per-residue multi-class classification over the 61 sense codons, the three stop codons excluded because they admit no synonymous variation. The human training set was assembled from UCSC and NCBI GenBank, keeping CDSs between 250 and 2,500 nucleotides and retaining only transcripts with a minimum free energy (MFE) at or below −76 kcal/mol computed with ViennaRNA 2.6.4, so the model learns codon usage from structurally stable messages. The E. coli and Chinese hamster datasets were adopted unchanged from earlier codon optimization work.

On an 820-sequence human test set, ppLM-CO reached a mean codon adaptation index (CAI) of 0.98 ± 0.02, against 0.97 for CodonBERT, 0.77 for LinearDesign, and 0.75 for the wild-type sequences, with a mean MFE of −282 ± 212 kcal/mol. The mask yielded 100% synonymous accuracy, compared with 99.89% for CodonBERT and 56.51% for LinearDesign. Training on the other two hosts lifted mean CAI from 0.67 to 0.97 for E. coli and from 0.78 to 0.96 for Chinese hamster, with GC content staying inside the 30–70% band throughout.

#Applications

The immediate use is designing the coding region of an mRNA vaccine or therapeutic. For the SARS-CoV-2 spike antigen, ppLM-CO produced a CDS with CAI 0.99 and MFE −1,467 kcal/mol, ahead of Pfizer's BNT162b2 (0.94) and Moderna's mRNA-1273 (0.97) on CAI; for the varicella-zoster virus gE antigen behind shingles vaccines it reached CAI 0.99 and MFE −744 kcal/mol. The same per-host classifiers serve recombinant protein production in E. coli and Chinese hamster ovary cells, the standard bacterial and mammalian expression systems. Because only the head is trained, adapting to a new host needs a modest set of paired protein and coding sequences rather than a full pretraining run.

#Impact

ppLM-CO joins a growing family of learned codon optimizers, including High-Codon, which also fine-tunes ProtBert but for E. coli expression, and codon-native models such as CodonTranslator, CodonFM, and TransCodon. Its distinctive claim is architectural: a frozen protein language model's embeddings already separate codon contexts well enough that a linear readout suffices, and restricting the label space to synonymous codons converts biological validity from something a model must learn into something the architecture enforces. Two limits bound the results. All evaluation is in silico, resting on CAI, MFE, and GC content rather than measured expression or immunogenicity, and stability trails LinearDesign, which optimizes free energy directly. Code is released under Apache 2.0 alongside a checkpoint for each host, though the three training datasets the paper points to have not been posted.