mRNA optimization model that raises codon adaptation, tRNA adaptation, and folding stability at once while preserving the encoded protein sequence.
No providers recorded yet. Browse all providers
RNop is a deep learning model for mRNA sequence optimization, developed at Tsinghua University's Shenzhen International Graduate School. Given a coding sequence, it rewrites the codons to raise protein expression while leaving the encoded amino acid sequence untouched — the core design problem for mRNA vaccines and protein therapeutics, where the same protein can be encoded by an astronomically large number of synonymous sequences with wildly different expression.
Existing approaches force a three-way trade-off that the authors call the central difficulty of the field. Exact combinatorial optimizers such as dynamic programming methods guarantee fidelity but scale poorly, taking on the order of an hour for a 2.7 kb sequence. Faster tools optimize one or two objectives and ignore the rest. Learned sequence models can generate quickly but offer no hard guarantee that the translated protein is unchanged. RNop attacks all three at once by pushing the biology into the loss function rather than the architecture: four differentiable objectives, each aligned with a distinct mechanism in the mRNA lifecycle, are optimized jointly during training, so a single forward pass at inference produces a sequence that already satisfies them.
Readers looking for the code should note that the repository is named RPLoss, after the loss functions rather than the model — the pretrained weights and the inference script for RNop both live there.
RNop applies a Vision Transformer architecture to the sequence-to-sequence problem, reshaping the tokenized coding sequence into an image-like feature map processed by 16 shifted-window multi-head self-attention layers, with an overlapping convolution (stride 1, kernel 3) for input embedding and a projection head that decodes back to the codon vocabulary. Training used over 3,000,000 mRNA sequences drawn from the NCBI Genome and NCBI Virus databases, spanning eukaryotes, bacteria, and viruses and sampled through 1 May 2024, on a single NVIDIA RTX 3090 for 120,000 iterations. On human sequences the model raises the codon adaptation index from 0.703 to 0.974 and the tRNA adaptation index from 10.945 to 13.16; for E. coli the corresponding figures are 0.655 to 0.992 and 1.679 to 2.109. In vivo assays with eGFP measured fluorescence intensity over four days and showed protein expression up to 4.6 times higher than the unoptimized control.
The immediate use case is designing coding sequences for mRNA therapeutics and vaccines, where expression level directly determines dose and efficacy, and for recombinant protein production in bacterial, yeast, or mammalian hosts. Because the species target is set by the codon usage and tRNA tables rather than by retraining, the same checkpoint serves multiple expression systems, and the throughput makes it practical to optimize whole libraries rather than single constructs.
RNop's contribution is methodological as much as empirical: it shows that mechanistic knowledge about translation — codon bias, tRNA availability, folding energy — can be encoded as differentiable losses and learned once, replacing per-sequence combinatorial search with a single fast inference pass while retaining a hard fidelity guarantee. The in vivo eGFP results are an important step beyond the purely in silico metrics that dominate this area. Practical caveats: the preprint has not been peer reviewed, and the code is MIT-licensed but describes itself as a test-only repository — the four mechanism-aligned loss functions that give RPLoss its name ship as placeholder stubs that return zero, so the paper's central contribution cannot be reproduced from the released code. Weights are distributed through a Google Drive link rather than a model hub, which makes them less discoverable and less durable than a versioned checkpoint, and no model card or data card accompanies the release.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.