UC Berkeley researchers developed ProteinGuide, a method that uses experimental data to guide existing protein generative models toward desired properties without retraining them.
Perhaps you have tried recreating a pasta dish from your favorite Italian restaurant. You know the basic ingredients, and you may even have found and tested several recipes online. But something is still off: the sauce is too heavy, the pasta is not quite al dente, or the flavors simply do not come together in the right way.
Protein engineers often share the same frustration—except their “pasta” exists on a molecular scale. They 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. These are key properties that can make engineered proteins effective in applications across human health, manufacturing, agriculture, and environmental protection.
To tackle this challenge, UC Berkeley researchers led by Professor Jennifer Listgarten and Professor David Savage developed ProteinGuide, a new strategy for designing proteins with tailored properties. The method uses experimental data to steer existing protein generative models toward improved properties while leaving the generative models unchanged.
In a study recently 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 lab, the computational team joined forces with experimental scientists Maria Lukarska and Luke Oltrogge. 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. This result suggests that ProteinGuide, by reducing the number of experimental iterations, could accelerate protein engineering for biotechnology and medicine.
Teaching AI what scientists want
Recent advances in AI have given protein engineers a growing collection of generative models for designing new proteins. 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. 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.
“These more general protein generation models don’t really know our specific design specification about this particular protein,” explains Junhao Xiong, a recent UC Berkeley PhD graduate and one of the study’s lead authors. “They may also lack the specialized experimental data we have collected for the protein we want to engineer,” he explains.
Researchers can try to bridge this gap in several ways. One traditional approach is post hoc filtering, in which proteins produced by a generative model are scored using predictions based on experimental data. It is like 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.
Another approach is fine-tuning, which involves further training the generative model on proteins known to have desirable properties. This is like honing your cooking skills by studying a small collection of recipes related to the dish you want to recreate.
Fine-tuning can help the model specialize, but it requires additional training and may concentrate the model too narrowly around those examples. ProteinGuide works differently.
“Almost all of these models generate in some sort of iterative fashion,” Xiong points out. ProteinGuide takes advantage of this step-by-step process, using experimental information to steer each successive choice toward the desired protein properties, like you keep tasting and adjusting your pasta sauce as you add different ingredients.
Rather than retraining the generative model, ProteinGuide pairs the model with a separate property predictor trained on experimental measurements. “It’s more flexible,” Xiong says. “You don’t need to change your pre-trained model in any way, and you don’t even need access to its parameters.” This advantage becomes particularly useful for large pre-trained generative models or models whose parameters are not fully accessible.
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. ProteinGuide extended that foundation to a broader range of widely used protein generative models, including ProteinMPNN and ESM3.
In several in silico tests, the team demonstrated ProteinGuide’s power by applying it to improve protein stability and activity and even to balance competing goals, such as increasing a sensor’s binding to lead while reducing unwanted binding to zinc.
A leap in base editor engineering
To test whether ProteinGuide could improve proteins in the lab, the team turned to adenine base editors (ABEs). These CRISPR-based gene-editing tools contain an engineered version of TadA, a bacterial enzyme whose natural function is to modify RNA.
Through previous engineering campaigns, researchers had obtained multiple variants of TadA that can make targeted changes to DNA. These variants provide clear benchmarks for comparison, making ABEs a particularly useful engineering target for testing ProteinGuide.
The team first used two pre-trained protein generative models, ESM3 and FMIF, to design a library of 2,000 variants. Maria Lukarska, a postdoctoral scholar and another lead author of the study, together with her colleagues, measured the DNA-editing activity of these variants using a bacterial high-throughput selection assay. The assay data were used to train an ABE editing-activity predictor. Using this predictor, the team applied ProteinGuide to steer the generative models to design a second library of 2,000 variants predicted to have higher editing activity.
The second library showed a striking shift toward higher editing activity. It contained variants that were more active than any in the first library.
One of the top variants even achieved activity between two previously engineered, highly active editors, ABE7.10 and ABE8. The first had been developed through seven rounds of directed evolution, while the second was further evolved from it using phage-assisted evolution.
“I didn’t think that in one round we would get there,” Lukarska says. Because the most active proteins in the first library still had relatively low activity, she initially doubted that the dataset would contain enough information for the model to learn what made an ABE highly active. “I think my expectation was that we were going to do three rounds from the beginning.”
The successful variants brought another surprise: they were highly diverse rather than converging on the same set of mutations.
“I would have expected that all those winners would have the same mutations,” says Lukarska.
Instead, very different protein sequences reached similarly high activity levels, suggesting that multiple evolutionary paths through sequence space may lead to the same functional destination. Each successful variant could therefore serve as a distinct starting point for another round of engineering.
Learning from every batch
Because ProteinGuide works on the fly, researchers can reuse the same pre-trained generative model and guide it with an iteratively updated property predictor from one design round to the next. But the method can only be as effective as the data used for guidance.
“The experimental limitations are going to be the biggest limitations,” says Lukarska. Without enough measurements from high-throughput assays, researchers may not be able to train a reliable property predictor. And even when the model can propose thousands of diverse sequences, researchers may struggle to find practical and affordable ways to synthesize them.
Xiong reaches a similar conclusion from the computational side: better pre-trained generative models, better property predictors, and better experimental data may now have a greater impact than developing yet another variation of the guidance method.
Overcoming these limitations will require close collaboration between computational and experimental scientists. Lukarska recalls that the two sides initially struggled to understand each other: “We were literally speaking different languages.” Through frequent conversations, the computational researchers learned why certain experiments were impractical or noisy, while the experimental researchers learned how the models worked and where their limitations arose.
Xiong similarly emphasizes the role of timing and collaboration in bringing the project together. He describes feeling fortunate that he could first develop the underlying method with machine learning scientists and later work with Lukarska and other experimentalists to test what it could accomplish in the lab. “It just makes me appreciate having good collaborators,” he says.
Lukarska agrees: “I don’t think any of us could have done any part of this by themselves without the help of each other.”
ProteinGuide does not eliminate the need to cook, taste, and learn from each batch of pasta. But it provides a way to carry those lessons directly into the next recipe—and perhaps discover several very different recipes that work even better than the dish you originally set out to recreate.