Berkeley's ProteinGuide Steers AI Toward Better Protein Design

TL;DR: UC Berkeley researchers created ProteinGuide, a method that steers existing protein-generating AI models toward desired properties using experimental data without retraining the models.

UC Berkeley researchers developed a new protein-design method called ProteinGuide that uses experimental data to guide existing protein generative models toward desired properties without retraining them, according to Guiding AI toward better protein design – QB3 Berkeley.

The report, written by Hanlun Jiang and published on September 17, 2026, describes work led by Professor Jennifer Listgarten and Professor David Savage to design proteins with tailored properties while leaving the generative models unchanged.

In a study published in Nature Biotechnology, the machine learning team including Junhao Xiong, Ishan Gaur and Hunter Nisonoff showed that ProteinGuide can be applied across several widely used classes of protein generative models and can optimize one or multiple properties at once.

To test the method in the laboratory, the computational team joined forces with experimental scientists Maria Lukarska and Luke Oltrogge, according to the same QB3 Berkeley report.

Together, the researchers used ProteinGuide to improve an adenine base editor, a CRISPR-based molecular tool that can convert adenine to guanine at targeted sites in DNA.

After a single round of guided design, they produced a highly active editor that rivaled variants developed through more labor-intensive, multi-round engineering campaigns.

That result suggests that ProteinGuide, by reducing the number of experimental iterations, could accelerate protein engineering for biotechnology and medicine, the QB3 Berkeley account stated.

A listing on the EECS at Berkeley publications page dated September 18, 2026 also describes the project as UC Berkeley research led by Professor Jennifer Listgarten and Professor David Savage that developed ProteinGuide to guide existing protein generative models toward desired properties without retraining them, according to publications Archives - EECS at Berkeley.

Why protein engineers need better steering

Protein engineers may know what a functional protein generally looks like yet still struggle to find the precise changes that will make it more stable, more active or more selective, according to QB3 Berkeley.

These are key properties that can make engineered proteins effective in applications across human health, manufacturing, agriculture and environmental protection, the report stated.

The publication compares the challenge to recreating a pasta dish from a favorite Italian restaurant, where the basic ingredients are known but the sauce is too heavy, the pasta is not quite al dente, or the flavors do not come together in the right way.

For protein engineers, that frustration exists on a molecular scale, because they seek specific changes that improve stability, activity or selectivity.

Recent advances in AI have given protein engineers a growing collection of generative models for designing new proteins, according to the QB3 Berkeley report.

Trained on enormous datasets of natural protein sequences and structures, these models can generate diverse, plausible protein sequences beyond those previously found in nature.

But researchers do not want just any new protein, the report explained.

They want functional proteins for specific purposes: therapeutics, biosensors, and industrial enzymes such as engineered proteases used in laundry detergents to break down protein-based stains.

Compared with sequence and structure data, however, there is far less information available about protein functions and how changes to their sequences affect those functions, according to QB3 Berkeley.

Junhao Xiong, a recent UC Berkeley PhD graduate and one of the study's lead authors, said the more general protein generation models lack knowledge of a particular design specification.

“These more general protein generation models don’t really know our specific design specification about this particular protein,” Xiong explained in the QB3 Berkeley report.

Xiong added that those models may also lack the specialized experimental data collected for the protein that researchers want to engineer.

Xiong is a recently graduated PhD student in Professor Jennifer Listgarten's research group in the Department of Electrical Engineering and Computer Sciences at UC Berkeley, according to QB3 Berkeley.

Filtering and fine-tuning leave a gap

Researchers can try to bridge the gap between general protein generation and specific functional goals in several ways, according to QB3 Berkeley.

One traditional approach is post hoc filtering, in which proteins produced by a generative model are scored using predictions based on experimental data.

The report compares post hoc filtering to preparing many pasta dishes, tasting them and keeping the best ones.

But this becomes inefficient when proteins with the desired properties rarely appear among the model's usual outputs, the report stated.

Another approach is fine-tuning, which involves further training the generative model on proteins known to have desirable properties.

The QB3 Berkeley report compares fine-tuning to honing cooking skills by studying a small collection of recipes related to the dish a person wants to recreate.

Fine-tuning can help the model specialize, but it requires additional training and may concentrate the model too narrowly around those examples, according to the report.

ProteinGuide works differently from both post hoc filtering and fine-tuning, the report stated.

Rather than scoring a large batch after generation or further training the generative model on desirable examples, ProteinGuide intervenes during generation itself.

That distinction matters because post hoc filtering depends on desired proteins already appearing in the usual outputs, while fine-tuning depends on additional training that can narrow the model.

The QB3 Berkeley report presents ProteinGuide as a new strategy for designing proteins with tailored properties that steers existing models toward improved properties.

The method uses experimental data to steer existing protein generative models toward improved properties while leaving the generative models unchanged, according to the report.

How ProteinGuide steers step by step

Almost all of these models generate in some sort of iterative fashion, Xiong pointed out in the QB3 Berkeley report.

ProteinGuide takes advantage of this step-by-step process, using experimental information to steer each successive choice toward the desired protein properties.

The report compares that process to tasting and adjusting a pasta sauce while adding different ingredients, rather than waiting until many finished dishes are prepared.

Rather than retraining the generative model, ProteinGuide pairs the model with a separate property predictor trained on experimental measurements, according to QB3 Berkeley.

That separate predictor is trained on experimental measurements and supplies property information during the iterative generation process.

“It’s more flexible,” Xiong said in the report, describing the pairing of an unchanged generative model with a separate predictor.

“You don’t need to change your pre-trained model in any way, and you don’t even need access to its parameters,” Xiong said.

This advantage becomes particularly useful for large pre-trained generative models or models whose parameters are not fully accessible, according to QB3 Berkeley.

The EECS at Berkeley publications listing similarly emphasizes that the method guides existing protein generative models toward desired properties without retraining them.

The approach grew out of Discrete Guidance, an earlier framework developed in the Listgarten group by Hunter Nisonoff, Stephan Allenspach and Xiong himself for steering models that generate discrete objects, including protein and DNA sequences, according to QB3 Berkeley.

ProteinGuide extended that foundation to a broader range of widely used protein generative models, including ProteinMPNN and ESM3, the report stated.

In several in silico tests, the team demonstrated ProteinGuide's power by applying it to improve protein stability and activity, according to QB3 Berkeley.

Those computer-based tests also showed balancing of competing goals, such as increasing a sensor's binding to lead while reducing unwanted binding to zinc.

The Nature Biotechnology study showed that ProteinGuide can be applied across several widely used classes of protein generative models and can optimize one or multiple properties at once, according to the report.

That ability to handle one or multiple properties at once is central to the reported advance, because protein engineers often need stability, activity or selectivity together rather than in isolation.

Test case: a more active adenine base editor

To test whether ProteinGuide could improve proteins in the laboratory, the Berkeley team turned to an adenine base editor, according to QB3 Berkeley.

An adenine base editor is a CRISPR-based molecular tool that can convert adenine to guanine at targeted sites in DNA, the report stated.

The computational team of Xiong, Gaur and Nisonoff joined forces with experimental scientists Maria Lukarska and Luke Oltrogge for that laboratory test.

Together, the researchers used ProteinGuide to improve the adenine base editor in a single round of guided design.

After that single round, they produced a highly active editor that rivaled variants developed through more labor-intensive, multi-round engineering campaigns, according to QB3 Berkeley.

The comparison to multi-round campaigns is important because traditional protein engineering often requires repeated cycles of design and testing.

By matching variants from more labor-intensive, multi-round engineering campaigns after only one guided round, the ProteinGuide-designed editor demonstrated reduced experimental iteration.

That result suggests that ProteinGuide, by reducing the number of experimental iterations, could accelerate protein engineering for biotechnology and medicine, the report stated.

The project brought together machine learning researchers and experimental scientists, with Listgarten leading the computational direction and Savage contributing experimental leadership.

The EECS at Berkeley listing identifies Listgarten and Savage as the leaders of the UC Berkeley team behind ProteinGuide.

Social-media posts by participating scientists describe the Nature Biotechnology paper as “Property guidance for protein sequence generative models with ProteinGuide” and summarize it as on-the-fly conditioning of pretrained protein generative models toward specific properties.

What changes for United States biotechnology

For United States science and technology readers, the practical meaning is straightforward: existing protein AI models can be reused more efficiently when paired with focused experimental data, according to QB3 Berkeley.

Instead of retraining a large pre-trained generative model, a laboratory can train a separate property predictor on its own experimental measurements and use it to guide generation.

That structure preserves the generative model unchanged while adding project-specific direction from laboratory measurements.

It also avoids the additional training required by fine-tuning and avoids concentrating the generative model too narrowly around a small set of examples, according to the report.

It further avoids reliance on post hoc filtering, which becomes inefficient when desired proteins rarely appear among the model's usual outputs.

Because ProteinGuide does not require access to the pre-trained model's parameters, it can be useful for large pre-trained generative models or models whose parameters are not fully accessible, the report stated.

The reported compatibility with ProteinMPNN and ESM3 indicates application across widely used classes of protein generative models, rather than a single custom system.

The reported in silico results on protein stability and activity, plus the lead-versus-zinc sensor example, indicate both single-property improvement and balancing of competing goals.

The adenine base editor experiment provides the laboratory demonstration, showing a CRISPR-based tool that converts adenine to guanine improved in one guided round to a highly active form.

For biotechnology and medicine, fewer experimental iterations can mean faster movement from computer design to laboratory-tested protein candidates, according to the interpretation offered in the QB3 Berkeley report.

The work was reported by QB3 Berkeley on September 17, 2026, and noted by EECS at Berkeley on September 18, 2026, placing it within that week's Berkeley research announcements.