OpenAI's GPT-4b micro turns protein design into a prompting problem

What Retro's scientists actually operated

OpenAI and Retro Bio built GPT-4b micro, a miniature version of GPT-4o initialized from a scaled-down base and then trained mostly on protein sequences, plus biological text and tokenized 3D structure data. The point of that mix was not raw prediction accuracy; it was to make the model something a scientist could steer.

A large portion of the training data was enriched with contextual information — textual descriptions of proteins, co-evolutionary homologous sequences, and groups of proteins known to interact. That enrichment is what let the model 'be prompted to generate sequences with specific desired properties.' In effect, the interface Retro's team worked through was a prompt: describe the property you want, get candidate sequences back.

Because most of the data was structure-free, the model handled intrinsically disordered proteins as well as structured ones — which mattered here, since the Yamanaka factors KLF4 and SOX2 are largely unstructured, with flexible arms that attach to other proteins rather than folding into a single stable shape.

The 64,000-token context is the real interface story

Training on proteins wrapped in evolutionary and functional context stretched the effective length of each example well beyond a standalone sequence. OpenAI reports running prompts as large as 64,000 tokens at inference and still seeing gains in controllability and output quality — a context size it calls unprecedented in protein sequence models, though routine in text LLMs.

That detail is where the 'frontend' framing becomes literal. A large context window is what makes a rich prompt possible: the scientist can supply descriptions, homologs, and interaction partners, and the model conditions its output on all of it. The team also observed language-model-style scaling laws — larger models on larger datasets yielded predictable perplexity and benchmark gains — which let them iterate cheaply at small scale before training the final model.

Why the demo had to happen in a wet lab

OpenAI is unusually blunt about the limits of its own metrics: 'in silico evals for protein AI models are often of limited value, as it is unclear if such improvements translate to increased utility in the real world.' So the validation was physical, run on Retro's fibroblast screening platform.

For RetroSOX, over 30% of the model's suggestions outperformed wild-type SOX2 at expressing pluripotency markers, despite differing by more than 100 amino acids on average — against typical traditional-screen hit rates below 10%. For RetroKLF, 14 model-generated variants beat the best RetroSOX cocktails, a hit rate near 50%, where prior expert-guided single-substitution work produced one hit out of 19.

Combining the top variants produced the headline 50-fold increase in reprogramming markers over wild-type controls, with late markers appearing days sooner. The result was later validated across multiple donors, cell types, and delivery methods, including mRNA delivery into mesenchymal stromal cells from donors over 50, where more than 30% of cells began expressing key markers within seven days.

The deep-edit signal that prompting beats mutation screens

The most engineering-relevant number is not the 50x. It is that the winning variants differed from natural human proteins by more than 100 amino acids. Directed-evolution screens mutate a handful of residues at a time; a leading academic effort tested a few thousand SOX2 mutants for a modest triple-mutant gain, and 15 years of chimeric SOX work yielded variants differing by only five residues.

A prompt-driven generative interface explores a different region of a 10^1000-variant space entirely. Instead of perturbing a known sequence, the model proposes distant candidates conditioned on the property description — and does so with hit rates that survived a physical screen. That is the concrete case for treating protein design as a controllability problem rather than a search problem.

What a steerable, non-public model means for domain teams

OpenAI frames this as speed: 'When researchers bring deep domain insight to our language-model tooling, problems that once took years can shift in days,' says Boris Power, who leads research partnerships. The specific enabler was steerability — a prompt interface rich enough that Retro's scientists, not model engineers, drove the design work.

The caveats are stated plainly and matter for anyone tempted to generalize: GPT-4b micro 'was developed for research purposes and is not broadly available,' and Sam Altman is an investor in Retro Biosciences. So the reproducible lesson is not a model you can call today. It is the architecture of the collaboration: a small, domain-tuned model fronted by a long-context prompt, with the real evaluation moved out of the benchmark and into the assay.