Pinal, a protein-design model from Westlake University and collaborators, took the prompt “glows green like GFP but isn’t GFP” and produced 50,000 candidate sequences. After computational filtering it tested 29 of them in E. coli. One worked. That is a 3.4% hit rate on the shortlist, and a vastly lower one against the original 50,000 — the usual cost of swapping a precise technical specification for a sentence in English.
The model, described in a paper in Nature titled “Programmable design of functional proteins from natural language” and first posted to bioRxiv in 2024, is one of two natural-language biology tools to surface in the past two weeks. The second, Hugging Face’s Carbon-X, does the equivalent for DNA: you paste in a sequence and ask it questions in English rather than running it through specialized bioinformatics tools. Both chase the same bet — that a language-model interface trained on huge amounts of biological data can make expert biological design usable by non-experts. Neither, on its own numbers, gets there reliably yet.
What the old tools require
Existing protein-design pipelines, as Ars Technica explained, work in two stages: a structure-generation model builds a three-dimensional backbone to a set of constraints (bind this molecule, hold this amino acid at this position), then a sequence-design model fills in amino acids to make that shape fold. Specifying the constraints is the hard part. Designing an enzyme means knowing which amino acids carry out the catalytic chemistry and where they need to sit in 3D space — knowledge two protein chemists would compress into a word like “serine protease,” but which current structure-generation models can’t infer from the word alone, because they’re trained on geometry, not biological meaning.
How Pinal closes that gap
Pinal’s bet is that millions of protein database entries already carry function tags in roughly that vocabulary, and a model trained on enough of them can translate a sentence directly into structure. The team built the training set from three protein databases, feeding annotations through an LLM to generate plain-language descriptions, then pairing them with structures predicted by DeepMind’s AlphaFold. The full synthetic training set covers 135 million unique protein sequences and 1.7 billion protein-text pairs, dwarfing Swiss-Prot, the field’s main human-curated database, at about 545,000 proteins and 27 million text pairs — a 247-fold expansion in sequences and 63-fold in annotated pairs. The largest version of Pinal’s structure-generation component, T2struct, runs to 15.5 billion parameters; smaller versions at 200 million and 1.2 billion showed consistent gains from scale.
T2struct doesn’t output coordinates. It outputs a 1D sequence of discrete tokens, borrowed from a method called Foldseek, where each token encodes a residue’s spatial relationship to its nearest neighbors, plus a probability. A second module, SaProt-T, takes those tokens and the text prompt and writes the amino acid sequence, re-ranking the top 50 candidate structures rather than committing to the single most likely one — the authors argue the most probable backbone can still produce a worse protein than a slightly less probable one.
AlphaFold’s predictions aren’t always right, so some of Pinal’s training data is built on guesses rather than measured structures. For well-studied protein families the predictions tend to be close to crystallography; for obscure families, the error compounds.
What the lab actually confirmed
The GFP case was the cleanest test: the team deliberately excluded designs too similar to known fluorescent proteins, to make sure any hit was genuinely novel. The surviving design, pinalGFP, fluoresces at the same wavelengths as natural GFP — 460 nanometres excitation, 510 nanometres emission — while sharing only 30–60% sequence identity with anything in Swiss-Prot.
The other three case studies ran similarly low success rates through similarly large funnels. For PETase, an enzyme that breaks down PET plastic, Pinal generated 200,000 candidates, filters cut that to 2,416, and the team assayed 94; one, pinalPETase36, was active. For an alcohol dehydrogenase enzyme, 2,500 candidates across five prompts were filtered to 1,000, 8 were tested, and 5 expressed successfully. Across all four case studies the paper reports 12 experimentally active Pinal-designed proteins in total. In a separate test of how far Pinal could stray from known structures, 17 of 20 low-similarity designs, or 85%, still folded and expressed correctly in soluble form — a higher hit rate than the function-specific tests, because folding is an easier bar than folding into something that does a specific job. One Pinal-designed protein, built for a carbon-dioxide-fixation pathway without further optimization, produced 1.7 times more of the target product than the natural E. coli version.
Against an existing benchmark, Pinal reportedly recovered 77% of a protein’s functional keywords from its generated sequence, versus about 12% for ESM3, a competing structure model, and an independent evaluation found Pinal ahead of four rival systems (ProteinDT, Chroma, ESM3 and PAAG) on 11 of 14 metrics, according to the paper and the preprint on bioRxiv. None of that changes the core number for anyone thinking about using this: of 29 GFP candidates picked by the model and its filters as most promising, 28 didn’t glow.
DNA gets the same treatment
Carbon-X, released on October 23 by Hugging Face with Zhongguancun Academy, Italy’s Telethon Institute of Genetics and Medicine and the University of Naples Federico II, takes a narrower approach: rather than designing new sequences, it answers questions about existing DNA. It is a 14.8-billion-parameter model built by retraining Alibaba’s Qwen3-14B on a mix of DNA and English, using a hybrid vocabulary that keeps Qwen3’s text tokens and adds 5,515 new tokens for DNA, each covering six base pairs. Mid-training ran 302 billion tokens: 55% raw DNA, 20% English text, 15% biology text and 10% instruction data, followed by supervised fine-tuning and two rounds of reinforcement learning with verifiable rewards.
Across five DNA benchmarks — covering variant-effect prediction, genome understanding and next-base prediction — Carbon-X averages 66.7%, which the team says is 27 percentage points above its Qwen3-14B starting point and ahead of every other open model tested. Against GPT-6.1 Sol, a larger closed model used as the main comparison, Carbon-X trails by 0.8 points on average (66.7% vs 67.5%) but wins outright on two of five benchmarks: coding variant effects (74.9% vs 71.0%) and next-base prediction (39.4% vs 32.9%). It loses by wider margins on complex and single-base variant-effect tasks, where Anthropic’s Opus 5.5 scores 94.8% and 76.2% against Carbon-X’s 85.1% and 70.4%. Splice-site detection is Carbon-X’s weakest spot at 29.8% accuracy.
The economic argument for Carbon-X rests on inference cost rather than accuracy. Hugging Face prices the model’s compute at $5 an hour on an H200 GPU and estimates that scoring 2.4 million variants of uncertain significance from ClinVar would cost about $300 on Carbon-X versus $32,000 to $4 million on frontier closed models — roughly 1,800 times cheaper than GPT-6.1 Sol specifically per the team’s own figure. That gap is the point of releasing open weights that run on a single 80GB GPU, or a 15.8GB quantized version that runs on much less: Carbon-X doesn’t need to match the frontier model’s accuracy to be worth using, it needs to be close enough that the cost difference wins. Lewis Tunstall, a senior research scientist at Hugging Face, described the project to Wired as offering a new way to interact with DNA, aimed at letting people without programming backgrounds query genetic sequences directly.
Hugging Face says the team deliberately excluded viral sequences from training to limit misuse, and the model card lists its caveats plainly: no clinical validation, weaker performance on splice sites and histone marks, degraded output on long generated DNA sequences, and no testing on inputs much longer than the roughly 50,000-base-pair examples it trained on — well short of a full human genome at over 3 billion base pairs. It also, per the card, struggles with exact operations like counting bases or tracking coordinates precisely, the kind of task a conventional script handles trivially and a language model, prone to approximating rather than computing, does not.
Both releases follow the same pattern as other recent attempts to put AI language interfaces on biological design, including Google DeepMind’s work on watermarking AI-designed proteins and the enzyme-design work out of the University of Washington. The pattern is the same: wider access, bigger candidate pools, and a success rate on any individual design that still depends on a lab to sort the handful of hits from the pile of misses. The number worth tracking next isn’t the headline benchmark score; it’s what fraction of candidates survive contact with a lab bench once labs outside Westlake University and Hugging Face’s own partners start running these models on problems nobody has tried yet.