Artificial Science

AI

ChatGPT's Training Trick Just Redesigned a Bird-Flu Vaccine Protein

A Nature Methods paper borrows the alignment method used to tune chatbots and applies it to protein design, producing a heat-stable H5N1 hemagglutinin with only 45 tries.

By Editorial Desk 10 min read 1,927 words

The same maths that turns a raw language model into a polite, useful chatbot has now been pointed at a protein. On 14 August 2026, Brian Hie's team at Stanford University and the Arc Institute published a method in Nature Methods that takes the alignment step behind systems like ChatGPT and reuses it to redesign proteins for stability. Their headline demonstration is an H5N1 bird-flu vaccine protein made dramatically more heat-resistant, using only 45 designs. That move, carrying an idea straight from large language models into the wet lab, is why it is worth your attention today.

What happened#

The model is called ProteinDPO, and the paper is open access. The authors, led by Talal Widatalla and Brian Hie, took a protein "language model" that already knew how to read three-dimensional structure and taught it a specific new skill: preferring protein sequences that fold stably. To do this they used direct preference optimization, or DPO, the lightweight alignment algorithm now common in chatbot training. One of the co-authors, Rafael Rafailov, was on the original DPO paper, so this is the technique's own inventors carrying it across into biology.

The results land on two fronts. As a scoring tool, ProteinDPO predicts how mutations change a protein's stability about as well as models built and trained for that one job, while keeping the ability to invent new sequences that the specialist tools lack. As a design tool, it was set loose on the influenza hemagglutinin protein. Starting from a 2004 H5N1 strain and testing just 45 candidates, it found 27 variants that were more stable than the original, with the best raising the melting temperature by up to 17°C. When the team applied the same changes to H5N1 strains that only emerged in 2024, one variant became 32°C more heat-stable while still being recognised by protective antibodies. The code is freely available.

The alignment gap, explained#

Proteins are chains of amino acids that fold into precise shapes. A protein language model treats the amino-acid sequence a bit like text, learning statistical patterns from millions of natural proteins. The specific model here, ESM-IF1, is an "inverse folding" model: give it a target 3D shape and it proposes sequences that should fold into that shape. It learned this by training on huge databases of known structures, without anyone ever telling it what "stable" means.

That is the catch. These models pick up a rough sense of stability by accident, because many proteins in their training data happen to be stable, but stability was never the goal they were trained on. Evolution tunes proteins for function, not for surviving a hot day in a fridge-free clinic. The authors call the mismatch between what the model learned and the property you actually want the "alignment gap". It is the same problem chatbots have before fine-tuning: fluent text, but no sense of how to be helpful.

The usual fix, supervised fine-tuning, just shows the model more examples of the good stuff, and it tends to overfit and forget its general knowledge. DPO works differently. It shows the model pairs and rankings, this sequence is more stable than that one, so the model learns from contrast across the whole landscape rather than memorising a list. The repurposing rests on a neat analogy. In a chatbot, the prompt is text, the response is text, and human ratings are the preference signal. In ProteinDPO, the protein's backbone shape is the "prompt", the generated sequence is the "response", and laboratory stability measurements are the "preference". They trained on the Megascale dataset, a 2023 experiment that measured folding stability for roughly 1.8 million protein variants, curated to about 660,000 measurements across 403 protein domains.

Why this matters#

The most immediate reason is speed in a public-health emergency. Highly pathogenic H5N1 has spread through wild birds, poultry and, since 2024, dairy cattle, with sporadic human cases, and health agencies including the CDC are watching it closely. Hemagglutinin is the protein on the flu virus surface that vaccines are built around, but it is awkward to work with: its useful "prefusion" shape is fragile, which hurts both manufacturing and shelf life where refrigeration is unreliable. A tool that finds heat-stable versions of it in 45 tries, rather than through months of trial and error, shortens the path from a new strain to a testable vaccine ingredient.

The finding that stretches further is generalisation. ProteinDPO was trained only on small single-domain proteins of 40 to 75 amino acids, yet its skill carried over to proteins several times larger, to antibody stability, and to how tightly two proteins bind. It even rediscovered, without ever seeing flu data, the specific "pH-switch" residues that earlier structural biologists had identified by hand as the key to stabilising hemagglutinin. A model that learns a reusable physical principle is worth far more than one that memorises a benchmark.

The broadest point is the one the authors themselves emphasise: this is a general recipe for aligning biological models, not a one-off. The same procedure could in principle steer a protein generator toward solubility, toward binding a chosen target, or toward any property you can measure and rank. The alignment toolkit that reshaped language AI now has a clear route into biology.

Critical analysis#

ProteinDPO did not beat every specialist stability predictor at pure scoring. Its advantage is that it does two jobs at once, scoring and generating, where the supervised tools do only the first. On its home benchmark, it clearly outperformed both the untuned model and the fine-tuned version, but against the best regression models it was competitive rather than dominant.

The flu results, though striking, are early. Melting temperature is a laboratory measure of heat stability. It is not the same as a vaccine that trains the immune system well in a living body. The team showed their variants kept their overall structure and still bound broadly neutralising antibodies, which is encouraging, but they are careful to say that real vaccine efficacy would need animal and eventually human testing. They also stabilised the desired prefusion shape without explicitly discouraging the unwanted post-fusion shape, something they flag as future work. And some of the larger designed sequences were checked only computationally, not expressed in the lab.

There is a quieter caveat too. The method needs a large dataset of ranked experimental examples to align against. Megascale provides that for stability, but for many other properties no such clean dataset exists, and building one is expensive. The recipe is general; its ingredients are not yet everywhere.

On timelines, the scoring and design tools are usable today because the code is public. A ProteinDPO-derived hemagglutinin actually reaching a vaccine would still follow the normal multi-year path of preclinical and clinical testing. The design step at the front of that path has become much faster; the rest has not.

Expert perspective#

The obvious comparison is with the wave of protein AI that preceded this. AlphaFold and the ESM family cracked structure prediction. Generative tools such as ProteinMPNN and RFdiffusion made designing new proteins routine. Supervised predictors such as ThermoMPNN became very good at scoring stability. ProteinDPO does not replace these. It sits on top of one of them and tunes it, the way instruction-tuning sits on top of a base language model.

What makes the work different is the direction of borrowing. Most cross-pollination between AI and biology has run one way, with biologists adopting architectures such as transformers. Here the flow is about method: the idea that a capable but unaligned model can be cheaply steered using preferences, rather than fresh labelled examples, comes straight from the human-feedback playbook that produced ChatGPT. The accompanying Nature Methods briefing frames it the same way, as evidence that experimental data can align a generative model without erasing what it already knows.

It also builds on the group's own work. Hie's lab had shown that a structure-aware language model could guide antibody improvement without task-specific training. ProteinDPO goes further, rewriting the model's preferences rather than just reading them.

Key takeaways#

  1. ProteinDPO applies direct preference optimization, the alignment method behind modern chatbots, to a protein language model, closing the "alignment gap" between what the model learned and the stability property scientists want.
  2. It works as both a scorer and a generator, matching specialist stability predictors while keeping the ability to design new sequences.
  3. In its flagship test it stabilised H5N1 hemagglutinin, finding 27 improved variants from 45 designs and raising melting temperature by up to 17°C, with a 32°C gain on a 2024 strain.
  4. The designs held their shape and stayed recognisable to protective antibodies, and the model rediscovered known stabilising residues it had never seen, suggesting it learned real physical rules.
  5. The approach is a general framework for aligning biological models, limited mainly by the need for large ranked experimental datasets and by the usual gap between lab measurements and clinical proof.

Frequently asked questions#

What is a protein language model? It is an AI trained on large collections of protein sequences or structures, learning statistical patterns much as a text model learns language. Given a target shape, an inverse-folding model like the one used here proposes amino-acid sequences that should fold into it.

What is direct preference optimization, in plain terms? It is a way to nudge a trained model toward preferred outputs by showing it comparisons, "this is better than that", rather than retraining from scratch. It became popular for aligning chatbots to human preferences, and here the "preferences" are experimental stability measurements.

Is this a new bird-flu vaccine? No. It is a computational and laboratory method that produced a more heat-stable version of a vaccine-relevant protein. Whether that translates into an effective vaccine would require animal and human trials.

Why does heat stability matter for vaccines? A more stable protein is easier to manufacture, holds its useful shape for longer, and survives breaks in cold-chain refrigeration, which matters most in resource-limited settings.

Did the mutations break the protein's function? In the tests reported, no. The stabilised variants kept their overall structure and still bound broadly neutralising antibodies, though full immune function was not tested in this study.

Can this be used for proteins other than flu? In principle yes. The authors present it as a general method and released the code. The main practical constraint is having a large, ranked experimental dataset of the property you want to optimise.

Is the model available to use? Yes. The authors made ProteinDPO open-source and publicly available.

References#

Related observations

Adjacent work from the same lines of enquiry.