TL;DR: We built an end-to-end protein AI pipeline covering structure prediction, sequence design, and codon optimization. After comparing multiple transformer architectures for codon-level language modeling, CodonRoBERTa-large-v2 emerged as the clear winner with a perplexity of 4.10 and a Spearman CAI correlation of 0.40, significantly outperforming ModernBERT. We then scaled to 25 species, trained 4 production models in 55 GPU-hours, and built a species-conditioned system that no other open-source project offers. Complete results, architectural decisions, and runnable code below.

Imagine going from a therapeutic protein concept to a synthesis-ready, codon-optimized DNA sequence in an afternoon. That is the pipeline OpenMed set out to build, and this post documents the process from start to finish.

In Part I, we mapped the landscape of protein AI: the architectures powering structure prediction, the open-source tools available for protein design, and the ecosystem of models from AlphaFold to ESMFold. That was a survey. This is the build.

At OpenMed, we set out to build a complete pipeline that takes a protein idea from concept to expression-ready DNA. That means three stages: predict the 3D structure of a protein, design amino acid sequences that fold into that structure, and optimize the underlying DNA codons so the protein actually expresses in the target organism. Along the way, we ran extensive experiments comparing transformer architectures for codon optimization, scaled our best model to 25 species, and built tooling that ties it all together.

This is not a polished success story. It is a transparent account of what worked, what surprised us, and what we would do differently, with runnable code and full results at every step.

Source link