Vibe genes? Natural(ish) language comes to protein, genome design software

Judgment

Lab leads deciding whether to hand Pinal to biologists instead of structural biologists must weigh a much lower skill barrier against a pipeline where, by the article's own framing, most designs fail and every hit still has to be sorted and validated at the bench.

Basis

Ars Technica's John Timmer (Oct 27, 2026) reports that a large collaboration of researchers based in China built Pinal, which takes plain-language prompts instead of residue-level constraints. Mechanically it keeps the usual two-step scheme but swaps the input: T2struct, a text-to-structure module, converts the prompt into a linear array of tokens encoding each amino acid's spatial relationship to its nearest neighbors, each with a probability, rather than a 3D shape. SAProt-T then combines the original text with that output to build the sequence, and re-evaluates globally so it does not overlook a slightly less probable structure that could carry a far better functional sequence. Training used function tags fed to an LLM to produce plain-language descriptions, paired with sequences from three databases and AlphaFold-predicted structures; one of those databases alone holds millions of tagged sequences. Separately, Hugging Face said it now hosts a DNA-level tool of similar spirit.