A new deep-learning model can now predict how proteins sort themselves inside the cell. The model has uncovered a hidden layer of molecular code that shapes biological organization, adding new dimensions of complexity to our understanding of life and offering a powerful biotechnology tool for drug design and discovery.
Previous AI systems in biology, such as the Nobel Prize-winning AlphaFold, have focused on predicting protein structure. But this new system, dubbed ProtGPS, allows scientists to predict not just how a protein is built, but where it belongs inside the cell. It also empowers scientists to engineer proteins with defined distributions, directing them to cellular locations with surgical precision.
“Knowledge of where a protein goes is entirely complementary to how it folds,” says Henry Kilgore, a chemical biologist at the Whitehead Institute for Biomedical Research in Cambridge, Mass., who co-led the research. Together, these properties shape its function and interactions within the cell. These insights—and the machine learning tools that make them possible—“will come to have a substantial impact on drug development programs,” he says.
Kilgore and his colleagues described the new tool in a paper published 6 February in the journal Science.
Putting Proteins on the Cellular Map
Over the past few years, AI tools like AlphaFold have revolutionized structural biology by predicting protein shapes—much like the instruction manual that comes with a piece of IKEA furniture, showing how to assemble the chair or bed. But it turns out knowing a protein’s structure isn’t enough to understand its function. ProtGPS fills in this missing piece by determining where each molecular piece of “furniture” belongs within the cell’s open-plan interior.
Some proteins have clear destinations. Researchers have known for decades that proteins headed for places like the nucleus or mitochondria—structures enclosed by membranes and walled off from the rest of the cell—carry short signaling tags that guide them.
But much of the cell is an open environment, where proteins rely on more subtle cues to sort themselves into what are called biomolecular condensates—dynamic, liquid-like clusters that help regulate gene activity, manage cellular stress, and contribute to disease. And just as a cozy armchair might naturally fit into a reading nook
ProtGPS has now begun to decode these rules, uncovering hidden features in the sequence of amino acids that form the backbone of all proteins—intrinsic sorting cues that determine whether and where a protein will localize within different condensates in the cell.
“Our model is learning these localization features,” says co-author Itamar Chinn, a machine-learning scientist at MIT. “And we can use those features to make new proteins that have the localization we want.”
ProtGPS uses a machine-learning framework to predict protein localization within condensate compartments.Henry R. Kilgore et al./Science
Teaching AI the Language of Proteins
ProtGPS is what’s known as a protein language model. It works much like LLMs such as OpenAI’s ChatGPT or Anthropic’s Claude, predicting sequences based on learned patterns. But instead of processing text or speech, ProtGPS analyzes proteins, which are represented as strings of letters, each corresponding to one of 20 amino acid building blocks—L for leucine, S for serine, and so on.
Kilgore, Chinn, and their colleagues built the model using a deep-learning framework called ESM, originally developed by Meta for predicting protein structures, functions, and properties.
Short for Evolutionary Scale Modeling, ESM—like AlphaFold—also extracts meaningful patterns from protein sequences. But instead of using physics to predict precise atomic-level structures, as AlphaFold does, Meta’s model relies on sequence-based learning without complex 3D calculations, making it substantially faster and more scalable for analyzing large datasets. (An upgraded version of ESM with improved capabilities was unveiled last month.)
Kilgore and Chinn’s team used ESM’s architecture to decode cryptic signals embedded in the amino acid sequences. The researchers adapted and refined the tool to both predict where proteins assemble and to enable the design of new kinds of proteins—ones that do not exist in nature, but can be engineered with precise condensate-targeting properties.
Thus, ProtGPS was born. The researchers trained the model on nearly 5,000 human proteins known to localize to one of 12 different condensate compartments. They then tested ProtGPS on an independent dataset, finding that it could accurately place proteins in the correct part of the cell.
An Elusive Code of Compartmentalization
Certain physical and chemical traits, like the charge and water-repelling nature of a protein, seemed to play a role in where things end up in the cell. But, as is often the case with machine-learning models, the exact reasoning behind ProtGPS’s predictions—and, by extension, the biology behind the selective distribution—remain largely a mystery.
That’s not to say the researchers didn’t try to tease it apart. They combed through the model’s predictions, searching for clear sequence patterns or biochemical properties that might explain its sorting rules. “Nothing obvious really falls out,” says co-author Peter Mikhael, a computational biologist at MIT.
That black box opacity is a familiar challenge in AI. Language models, by their very nature, excel at bringing together contributions from many different features and contextual signals, allowing them to detect patterns that aren’t immediately obvious to humans. “So, it’s not all that surprising” that ProtGPS can extract localization cues that even experienced biologists struggle to define, says Ilan Mitnikov, a machine-learning scientist formerly at MIT who helped to develop the model.
“If the rules were simple, people would have already figured them out,” Mitnikov says.
Engineering Proteins, Predicting Diseases
Even without a full understanding of what governs a protein’s cellular destination, the researchers showed that ProtGPS could be used to create proteins with carefully tuned localization properties. The tool also proved capable of predicting how mutations linked to disease might disrupt protein compartmentalization, shedding light on the molecular mechanisms underlying conditions such as cancer and developmental disorders.
Dewpoint Therapeutics—a biotech company co-founded by one of the study’s authors, Whitehead biologist Richard Young—now plans to integrate ProtGPS into its drug discovery efforts, according to chief scientific officer Isaac Klein, who called the tool a “game-changer” for identifying drug targets and designing new therapies. (Young, Kilgore, and MIT computer scientist Regina Barzilay, who also helped lead the study, all hold consulting or advisory roles with Dewpoint.)
Other scientists also see potential for the tool, including Tuomas Knowles, a biophysicist at the University of Cambridge who serves as chief technology officer of Transition Bio, another company focused on drug discovery against condensate targets. “What is particularly exciting is that this paper provides further evidence that there are very specific sequence features that govern localization and partitioning of proteins into condensates in living cells,” says Knowles, who was not involved in the research. “Furthermore, this provides new opportunities to influence and control protein localization—and potentially correct mis-localization, which is at the origin of many diseases,” he adds.
But beyond its applied utility, ProtGPS highlights an emerging paradigm in biology, in which the physical arrangement of the molecules within a cell is as critical to its function as the molecules’ structure, with codes embedded in the amino sequence that impact folding and cellular compartmentalization alike.
Just as a well-designed home is more than a collection of furniture—it relies on intuitive placement to maximize utility—cells, too, require precise molecular organization to function optimally. By uncovering hidden patterns in protein sequences, ProtGPS may serve as the architect of this cellular flow, decoding nature’s blueprint for the cell’s interior design.
