Researchers from China unveil AI for protein design from text descriptions — Ars Technica
Researchers from China have unveiled the Pinal system, which makes it possible to describe a required protein function in ordinary language and obtain candidate protein sequences for further laboratory testing. As Ars Technica reports, the tool is intended to simplify protein design, which previously required detailed knowledge of proteins' spatial structure, amino acid locations, and enzyme mechanisms.
From text to sequence
Conventional protein design systems work in two stages: first, they create a three-dimensional scaffold with specified parameters, and then select an amino acid sequence capable of folding into that structure. Pinal also uses two stages, but begins with a natural-language request for the required function.
The T2struct module converts a text description into a structural plan containing data on spatial relationships between amino acids and probabilities of different variants. The SAProt-T system then uses both the text request and this plan to form an amino acid sequence. According to the developers, global re-evaluation of variants should reduce the risk of discarding a structure that initially appears less probable but may provide better protein function.
To train the model, the team used protein sequences from three databases combined with structural data predicted by AlphaFold. One of the databases contained 800 million “keyword-protein” pairs. Protein functional tags were also converted into text descriptions using a large language model.
More current news is available on the UA.News Telegram channel Telegram.
Testing candidates in the laboratory
During testing, Pinal was tasked with creating four different types of proteins: a fluorescent protein, two enzymes, and a protein for a large multiprotein complex. To create an analogue of the green fluorescent protein GFP, the system generated 50,000 variants. After computer-based selection according to structural features and dissimilarity to known fluorescent proteins, 29 candidates produced in Escherichia coli were selected. One of them showed fluorescence.
The authors of the approach selected only dozens of candidates from large sets of generated sequences for experiments. For practical use of such tools, specialized selection tools, an equipped biological laboratory, and experts able to verify whether the created protein actually works will probably still be needed.
A separate tool for DNA
Separately, Hugging Face, together with Zhongguancun Academy and the University of Naples Federico II, unveiled Carbon X, a model with open weights for working with DNA sequences and natural language. It is based on Qwen3-14B and can be run locally, including on relatively affordable hardware. The developers warned that Carbon X is worse at identifying specific positions in long sequences and may generate false answers when the request is insufficiently precise.
