How We Built a Multi-Species mRNA AI Pipeline for Under $200

From Protein Concept to DNA Sequence: An AI Pipeline Revolution

Imagine having a brilliant idea for a therapeutic protein on Monday morning and, by the afternoon, possessing a fully optimized DNA sequence ready for synthesis and expression. This isn’t a distant dream—it’s the reality OpenMed’s open-source team is building toward. In a groundbreaking project documented on the Hugging Face blog, they constructed an end-to-end protein AI pipeline covering structure prediction, sequence design, and mRNA codon optimization. The most staggering achievement? Training specialized language models for 25 different species cost a mere $165 in compute.

This work represents a significant leap in democratizing advanced bioengineering tools. While large corporations and well-funded labs have access to sophisticated protein design software, OpenMed’s mission is to make these capabilities available to researchers, startups, and open-source developers everywhere. Their transparent build log, complete with runnable code, offers a rare look at the practical realities of building AI for healthcare and life sciences.

!A conceptual diagram showing the three stages of the protein AI pipeline: Structure Prediction, Sequence Design, and mRNA Optimization.

The Three Pillars of the Protein Pipeline

The team’s system addresses the three critical, sequential challenges in protein engineering:

1. Protein Folding (Structure Prediction): Determining the 3D shape a protein will adopt. They used ESMFold v1, Meta’s powerful open-source tool, to predict structures for 30 protein chains, achieving an average predicted TM-score (pTM) of 0.79—a strong indicator of accurate modeling.
2. Sequence Design: Figuring out which specific chain of amino acids will fold into the desired structure. For this, they employed ProteinMPNN, a neural network from the Baker Lab, which successfully recovered 42% of the native sequence for a given scaffold.
3. mRNA Optimization (Codon Optimization): This is where OpenMed made its most substantial original contribution. The genetic code is “degenerate,” meaning most amino acids can be encoded by multiple different three-letter DNA combinations called codons. The choice of codons dramatically impacts how efficiently a cell can produce the protein. Their goal was to build an AI that learns the optimal codon patterns directly from nature.

Why Codon Optimization Matters: The Pfizer-BioNTech COVID-19 vaccine’s mRNA sequence was meticulously codon-optimized for high expression in human cells. This step can mean the difference between a therapeutic that works at microgram doses versus one that requires milligram doses, impacting safety, cost, and manufacturability.

The Architecture Showdown: Finding the Best Model for Codons

Most biological language models, like the famed ESM-2 that powers ESMFold, are adaptations of NLP architectures. OpenMed’s first major research question was: which transformer architecture is best suited for the unique language of codons?

Codon sequences have distinct properties. They are composed of triplets from a 64-token “alphabet,” with strong positional rules and biases that vary wildly between species (what works well in E. coli bacteria may express poorly in human cells). The team tested a range of models, starting with a minimal 6-million parameter CodonBERT as a baseline.

Their contenders included:
ModernBERT-base: Incorporating the latest NLP efficiency innovations like Rotary Position Embedding (RoPE).
CodonRoBERTa variants: Based on the robust RoBERTa architecture, the same family as Meta’s ESM models.

The hypothesis was that the proven RoBERTa framework, already successful for amino acids, might also master the language of DNA codons.

And the Winner Is: CodonRoBERTa

The results were decisive. After training on 250,000 protein-coding sequences, the CodonRoBERTa-large-v2 model emerged as the clear champion. It achieved a perplexity of 4.10 (a measure of prediction confidence, where lower is better) and a Spearman correlation of 0.40 with the Codon Adaptation Index (CAI), a standard metric for expression efficiency.

This performance significantly outperformed the ModernBERT variant. The success validated their architectural choice, demonstrating that the RoBERTa framework could indeed capture the complex, species-specific statistical patterns of codon usage.

Scaling to 25 Species on a Budget

With a winning architecture in hand, the next challenge was scaling. A truly useful tool couldn’t be limited to one organism. The team expanded their dataset to 381,000 coding sequences across 25 diverse species, from humans and mice to yeast, plants, and bacteria.

Here’s where the project’s ethos of accessibility shines. They trained a suite of four production-ready models covering all 25 organisms in just 55 GPU-hours on cloud infrastructure. The total cost? Approximately $165. This cost-effectiveness opens the door for widespread use and further experimentation.

They built a species-conditioned system, allowing a user to specify a target organism (e.g., “Homo sapiens” or “Escherichia coli”) and receive codon optimization tailored specifically for it. This multi-species capability is something they note no other open-source project currently offers in this integrated form.

The Integrated Workflow and Practical Impact

So, what does the complete pipeline look like in practice? A researcher starts with a protein of interest. ESMFold generates a predicted 3D structure. That structure is fed to ProteinMPNN, which designs a novel amino acid sequence likely to fold into that shape. Finally, that amino acid sequence is passed to OpenMed’s CodonRoBERTa model, which, conditioned on the target species, outputs an optimized DNA/mRNA sequence with codons chosen for high expression.

Potential use cases are vast:
Therapeutic Protein & Antibody Development: Accelerating the design of biologics for diseases.
mRNA Vaccine Design: Rapidly optimizing sequences for new viral targets.
Synthetic Biology & Metabolic Engineering: Designing enzymes for bio-manufacturing pathways.
Research Tooling: Providing a fast, affordable benchmark for academic labs.

Looking Ahead: The Future of Open-Source Bio-AI

OpenMed is candid that this is a work in progress, not a polished final product. They share lessons learned, surprises encountered, and what they would do differently—a transparency that accelerates community learning.

The next steps are exciting. Future work could involve:
Training even larger models on more extensive genomic datasets.
Incorporating ribosomal profiling data to model translation dynamics directly.
Extending the pipeline to consider mRNA secondary structure and stability.
Integrating feedback from wet-lab experiments to close the design-build-test loop.

By open-sourcing their models, code, and detailed methodology, OpenMed isn’t just building a tool; they’re building a foundation. They are proving that sophisticated, multi-species AI for molecular biology can be developed efficiently and shared openly, potentially leveling the playing field in drug discovery and bio-engineering. For $165 and a lot of ingenuity, they’ve taken a significant stride toward that future.

Interested in exploring the models or code? The complete project, including trained model weights and notebooks, is available on the Hugging Face Hub under the OpenMed organization.