ProtGPS maps where proteins belong inside cells

If you think of a cell as a wide-open studio apartment, AlphaFold is basically the IKEA manual. It tells you exactly how to snap the pieces together to build a chair or a desk.

What it couldn’t tell you was where to put the furniture. Does that armchair go in a cozy reading nook, or does it belong somewhere else entirely?

That missing GPS is what researchers at MIT and the Whitehead Institute set out to build. In a paper published in Science, the team unveiled ProtGPS, a deep-learning model designed to map where proteins actually go once they are made.

Decoding the open-plan cellular layout

For decades, biologists understood the obvious destinations. Proteins heading into membrane-walled rooms like the nucleus carry clear molecular address tags.

However, much of the cell has no walls. Instead, it relies on biomolecular condensates—dynamic, liquid droplets that gather to regulate genes, manage cellular stress, and sometimes drive disease.

Proteins sort themselves into these liquid clusters based on subtle, hidden cues in their amino acid sequences. ProtGPS was built to read that cryptic code.

Instead of crunching heavy 3D atomic coordinates, the team adapted Meta's open-source ESM language model. Much like an LLM reads words, it reads strings of 20 amino acids to spot patterns human eyes miss.

Trained on nearly 5,000 human proteins across 12 distinct condensate compartments, the tool proved it could accurately place proteins where they belong.

There is a familiar catch: ProtGPS remains a black box. The researchers dug into the model's outputs looking for simple rules around charge or water repellency, but found nothing neat.

As the researchers put it, if the rules were simple, human scientists would have cracked them decades ago.

Why it matters

For biotech teams and drug hunters, this is a massive shift in how we think about cellular biology.

ProtGPS does two very practical things right out of the gate. First, it helps researchers engineer entirely synthetic proteins with surgical targeting, sending them to specific liquid compartments inside the cell. Second, it can flag when disease mutations cause proteins to drift into the wrong neighborhood, which is a major driver in cancers and developmental disorders.

Biotech startups are already moving. Dewpoint Therapeutics plans to integrate the model directly into its drug discovery pipeline, and researchers at Transition Bio see it as concrete proof that sequence alone dictates localization.

We think the clever part here is the sidestep: rather than waiting for massive compute to simulate full-cell biophysics, the team treated protein sorting purely as a context-heavy sequence problem.

It proves that in biology, spatial organization matters just as much as 3D shape.

That said, the black-box nature is an obvious hurdle. When designing clinical therapeutics, knowing that a protein lands in the right condensate without understanding why it sticks there means wet-lab scientists will still spend significant time validating safety and unintended interactions.