AI Model Designs Working Proteins Straight from Plain-English Prompts

Most protein design tools need expert input: a natural protein to optimize, or hand-crafted structural and sequence constraints. In a study reported in Nature, researchers describe Pinal, a 16-billion-parameter generative model that takes an ordinary written description of a desired function — say, "an enzyme that breaks down PET plastic" — and returns candidate protein sequences. The work was led by senior author Fajie Yuan at Westlake University in Hangzhou, China, with first author Fengyuan Dai. Crucially, the team did not stop at computer predictions: they synthesized dozens of the designs and showed that several actually work, including two new classes of active enzymes.

Pinal avoids the hardest mapping problem — going directly from language to amino acid sequence — by splitting the task in two. A first module turns the text prompt into a coarse structural blueprint, represented as a string of discrete structural tokens rather than raw 3D coordinates. A second module then designs an amino acid sequence that fits that scaffold while still respecting the text. The model was trained on 1.7 billion synthetically annotated protein–text pairs spanning 135 million protein sequences, far beyond the roughly half-million proteins in the curated Swiss-Prot database. Performance improved with both dataset size and model size, echoing scaling trends in general AI.

On computational benchmarks, Pinal's designs scored close to natural proteins on predicted folding confidence, while matching their text prompts far better than the earlier language-guided models ProteinDT and Chroma, which often produced sequences that scored near random. In a keyword-conditioned test against ESM3 — run on ESM3's own preferred input format — Pinal achieved a mean keyword-recovery rate of 77% versus about 12%. These results are predictions, not proof of function; the experimental work was the real test.

That test covered four proteins: a green fluorescent protein, a PET-degrading hydrolase, an alcohol dehydrogenase, and a metabolic H-protein. A Pinal-designed fluorescent protein, pinalGFP, emitted green light at 510 nm (when excited at 460 nm), and the signal disappeared when its predicted chromophore residues were mutated. A designed PETase, pinalPETase36, released the breakdown products terephthalic acid and MHET from PET film over seven days and left visible pitting on the plastic surface. Five designed alcohol dehydrogenases spanning three distinct subfamilies showed detectable activity. And in a reconstituted multi-enzyme CO2 fixation pathway, a designed H-protein called pinalH02 reached 11.1 mM glycine — 1.7-fold more than the native E. coli H-protein, with no iterative optimization.

Separately, 17 of 20 designs with little structural similarity to anything in the Protein Data Bank (85%) expressed as soluble protein in E. coli. The authors are explicit that soluble expression does not establish function or confirm the predicted structures.

The limitations matter too. The model's precision is capped by the granularity of existing biological annotations — prompting for an exact fluorescence wavelength, for instance, remains difficult. Candidates advanced to the bench were chosen using imperfect in silico proxies, so untested designs should not be assumed inactive; pairing language-guided generation with high-throughput screening will be needed to map the real reach of the approach.

If the results hold up, the practical shift is that protein design becomes something closer to a conversation than a constraint-engineering project. The next question is how well the approach handles functions that current databases describe poorly or not at all.

Write the function you want in a sentence; get a candidate protein back.

Reference: Dai, F., You, S., Zhu, Y. et al. Programmable design of functional proteins from natural language. Nature (2026). https://doi.org/10.1038/s41586-026-11205-z