Many MindedA text-to-protein model works, though only a sliver of its designs do
Scientists at Westlake University in China have published a Nature paper on Pinal, an AI system that writes protein sequences from a plain-language description of the function wanted. The user does not need to supply structural or sequence details.
Pinal works in two steps. One module turns the text prompt into a structural blueprint, proposing several candidate structures. A second module picks an amino acid sequence using both the blueprint and the original text, and Pinal scores the combinations to choose a final design. Training drew on about 135 million protein sequences, with large language models writing descriptions to produce roughly 1.7 billion protein-text pairs. AlphaFold predicted structures for the training sequences.
The team tested four targets: a green fluorescent protein, a PET-plastic-degrading enzyme, an alcohol-processing enzyme, and a shuttle protein moving chemicals between enzymes in a metabolic pathway. All four succeeded. For the glowing protein, Pinal generated 50,000 sequences, 29 were tested, and one fluoresced. For the shuttle protein, 5 of 20 candidates worked, and the best lifted a multi-enzyme carbon dioxide fixation pathway to 1.7 times the output of the natural version.
The abundance case is that asking in words lowers the skill needed to start designing useful proteins, which could widen who works on medicines, enzymes and plastic breakdown. The limits are real: hit rates are low, success takes many attempts, and the developers say their screening heuristics may confine exploration to a narrow region of design space. Testing the output still needs a well-equipped lab and experienced scientists, so the bottleneck has shifted rather than vanished.
The sheet also mentions Carbon-X, a separate open 14-billion-parameter DNA model that can run on a MacBook. Its benchmark comparisons come from its developers' own post, and it struggles with long sequences and still hallucinates.
Source: Singularity Hub