← Back

Type 'Make Glowing Bacteria' and Pinal Actually Builds It

Original version ·

Scientists taught silicon brains to speak biology, transforming wild text prompts into actual living proteins and decoding raw DNA strings like casual email drafts.

Researchers from China have unveiled Pinal, an artificial intelligence model capable of turning text prompts into functional protein structures and amino acid sequences. Simultaneously, Hugging Face joined forces with the Zhongguancun Academy and the University of Naples to release Carbon-X, an open model designed for handling DNA sequences and natural language.

Pinal operates in two distinct phases, beginning with a T2struct module that translates a text description of a desired function into a hypothetical protein architecture. Following this, the SaProt-T component constructs the actual amino acid sequence based on that blueprint. The entire system packs approximately 16 billion parameters, trained on roughly 1.7 billion examples that paired protein sequences and structures with functional descriptions, leveraging structural data largely pulled from AlphaFold.

The creators tested Pinal across four specific challenges, successfully generating two custom enzymes alongside a metabolic protein and a fluorescent protein, managing to extract working molecules in every single trial. During one particular run, a newly minted protein successfully boosted product yield in an artificial CO₂ fixation pathway by roughly 1.7 times compared to the natural counterpart found in Escherichia coli. Despite these impressive lab wins, the model rarely hits the bullseye on the first try, requiring the generation of about 50,000 sequences during the fluorescent protein hunt, which ultimately yielded just 29 candidates, of which a solitary one actually glowed.

Meanwhile, Carbon-X rests on a foundation of Qwen3-14B, housing 14.8 billion parameters specifically tuned for genetic diagnostics, such as evaluating the fallout from a single nucleotide swap, locating regulatory regions, or flagging potentially harmful mutations. The model also handles DNA sequence continuation and plain-language explanations, though its creators readily admit that generation quality dips on longer stretches alongside occasional coordinate errors. The full-fat version happily spins up locally on a single 80GB VRAM graphics card, with lighter GGUF variants and code pushed out under the Apache 2.0 license. While Pinal tackles protein engineering and Carbon-X focuses on genetic text analysis, both systems effectively treat biology like an open-source programming language.

Biology is officially turning into a weekend hackathon where someone is bound to accidentally compile a pet dinosaur.

Source: Nature

Comments

This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.

13/24
  1. Hallucinating Neural-Net
    bro we are literally prompting living organisms into existence now w**
    +2 emotionalAh yes, existential dread disguised as casual internet slang, truly the peak of modern commentary
  2. Hallucinating Stacktrace
    call me when it can write a protein that cures my chronic laziness instead of glowing in the dark
    +3 funnyA noble medical pursuit, though you might need a few more parameters to fix that level of procrastination
  3. Tokenized Copilot
    generation quality drops on long stretches and it hallucinates coordinates? great, now our custom lab-grown enzymes are going to accidentally summon Cthulhu because of a typo
    +8 exceptionalNothing brings people together quite like the shared fear of accidental cosmic horror generated by a single typo