AI Learns to Design Proteins from Text Descriptions
The Pinal system proposes protein sequences with specified functions
Researchers have introduced Pinal, a neural network that designs proteins based on descriptions of desired properties in natural language. The user specifies the required function, and the system suggests amino acid sequences for further study. The work was published in the journal Nature.
Typically, protein design requires detailed specifications: which elements must be present in the molecule and how they should be arranged in space. Pinal allows users to start with a functional description, reducing the volume of such manual adjustments.
The model contains approximately 16 billion parameters and was trained on 1.7 billion pairs linking proteins with textual descriptions. It operates in two stages: first, it forms a representation of the future protein's structure, then selects the corresponding amino acid sequence.
The authors tested the system on four tasks, including the creation of a fluorescent protein and two enzymes. In each category, they succeeded in obtaining variants with the required activity. However, this was preceded by computational screening and laboratory testing.
For example, for the fluorescent protein, the system proposed 50,000 variants. After screening, researchers tested 29 candidates, of which one exhibited fluorescence. This result demonstrates the potential of the approach but does not mean that every text query leads to a successful design.
A similar interaction principle has emerged in tools for working with DNA. The Hugging Face team and partners introduced Carbon-X, a model that processes genetic sequences and questions in English simultaneously. It is designed, in particular, for sequence analysis and assessing the consequences of mutations.
Carbon-X can be run locally, keeping research data on the user's own infrastructure. However, developers warn that the model may confuse positions in long sequences and fabricate explanations when context is insufficient.
Both developments simplify interaction with biological data. However, verifying the proposed solutions still requires specialists, additional computational tools, and, when it comes to new proteins, laboratory experiments.
