MIT and Microsoft researchers built an AI model that designs molecular sensors to spot cancer-linked enzymes, potentially turning a simple urine test into an early warning system for dozens of cancers. The work could also guide new treatments and a broader map of how these enzymes behave in disease.
Catching cancer when it first appears, long before symptoms show up, could save countless lives. A new study from MIT and Microsoft suggests artificial intelligence may help make that kind of ultra-early detection as simple as taking a home urine test.
The research team has created an AI system that designs tiny molecular sensors able to detect enzymes that are abnormally active in cancer cells. These enzymes, called proteases, help tumors grow and spread by cutting through proteins in the tissue around them.
By engineering nanoparticles coated with custom-designed protein fragments, or peptides, the scientists can turn that hidden enzyme activity into a clear signal that shows up in urine. In the future, that signal could be read on a paper strip similar to an at-home pregnancy test.
The team’s goal is to find cancer when it is still small and more easily treated, according to MIT bioengineer Sangeeta Bhatia.
“We’re focused on ultra-sensitive detection in diseases like the early stages of cancer, when the tumor burden is small, or early on in recurrence after surgery,” she said in a news release.
Bhatia, the John and Dorothy Wilson Professor of Health Sciences and Technology and of Electrical Engineering and Computer Science at MIT, and a member of MIT’s Koch Institute for Integrative Cancer Research and the Institute for Medical Engineering and Science (IMES), co-led the work with Ava Amini, a principal researcher at Microsoft Research and an MIT alum.
The study is published in the journal Nature Communications.
Turning cancer’s tools into beacons
More than a decade ago, Bhatia’s lab proposed using protease activity as an early warning sign for cancer. The human genome encodes hundreds of these enzymes, which can slice through other proteins, including collagen and other structural components that help hold tissues together.
Cancer cells often hijack proteases to break free from their original location and invade surrounding tissue or spread to distant organs. That makes protease activity a promising marker for disease that is otherwise hard to see.
Bhatia’s group previously showed that nanoparticles coated with specially chosen peptides could act as sensors. When these particles are swallowed or inhaled and circulate through the body, cancer-linked proteases cut the peptides. The resulting fragments are filtered into the urine, where they can be detected.
“We have been advancing the idea that if you can make a sensor out of these proteases and multiplex them, then you could find signatures of where these proteases were active in diseases. And since the peptide cleavage is an enzymatic process, it can really amplify a signal,” Bhatia added.
Using arrays of many different peptide sensors, the team has demonstrated diagnostic signatures for lung, ovarian and colon cancers in animal models. But there was a catch: the peptides were chosen largely by trial and error, and many could be cut by more than one protease. That made it difficult to link any one signal to a specific enzyme.
Why AI changes the game
Designing a peptide that is both highly sensitive and highly specific to a single protease is a massive search problem. A short peptide with 10 amino acids can be arranged in roughly 10 trillion different ways. Testing those possibilities one by one in the lab is impossible.
To tackle this, the researchers built an AI system called CleaveNet. It uses a type of protein language model, inspired by the same ideas behind large language models for text, to predict which amino acid sequences are likely to be cut by particular proteases and how efficiently that cutting will happen.
The team trained CleaveNet on publicly available data describing about 20,000 peptides and how they interact with a family of proteases known as matrix metalloproteinases, or MMPs. One model generates candidate peptide sequences that are likely to be cleaved; a second model predicts how efficiently a given protease will cut each candidate.
Users can feed CleaveNet design goals, such as maximizing how strongly a peptide responds to one protease while minimizing its response to others. The system then proposes peptide sequences that fit those criteria, dramatically narrowing down what needs to be tested in the lab.
This computational approach lets the team fine-tune the performance of their sensors.
“If we know that a particular protease is really key to a certain cancer, and we can optimize the sensor to be highly sensitive and specific to that protease, then that gives us a great diagnostic signal,” Amini said in the news release. “We can leverage the power of computation to try to specifically optimize for these efficiency and selectivity metrics.”
Putting CleaveNet to the test
To show what CleaveNet can do, the researchers focused on a protease called MMP13. Cancer cells use MMP13 to cut through collagen, helping them invade nearby tissue and metastasize.
When the team asked CleaveNet to design peptides that would be efficiently and selectively cut by MMP13, the model proposed sequences that had never appeared in its training data. In experiments, those AI-designed peptides turned out to be both highly efficient and highly selective for MMP13.
This was a key moment for the project.
“When we set the model up to generate sequences that would be efficient and selective for MMP13, it actually came up with peptides that had never been observed in training, and yet these novel sequences did turn out to be both efficient and selective,” added co-lead author Carmen Martin-Alonso, a founding scientist at Amplifyer Bio. “That was very exciting to see.”
Having more selective peptides could sharpen the diagnostic power of the sensor system. Instead of relying on large, overlapping panels of peptides, doctors might one day use smaller sets tuned to the most important proteases for a given cancer type. That could help identify new biomarkers, clarify which biological pathways are active in a tumor and guide more targeted therapies.
Toward at-home tests and smarter treatments
Bhatia’s lab is part of a federal ARPA-H project aiming to turn this technology into an at-home diagnostic kit that could potentially detect and distinguish up to 30 types of cancer in their early stages. The idea is to read patterns of protease activity from a simple urine test, using sensors that respond not only to MMPs but also to other enzyme families such as serine and cysteine proteases.
Beyond diagnostics, CleaveNet-designed peptides could be built into cancer treatments. For example, a drug could be attached to an antibody through a peptide linker that is only cut by proteases in the tumor environment. That would keep the drug inactive in healthy tissues and release it where it is needed most, potentially boosting effectiveness and reducing side effects.
The researchers also envision a broader scientific payoff. By combining AI-designed sensors with large-scale experiments across many cancer types, they hope to build a comprehensive “protease activity atlas” that maps how these enzymes behave in different diseases. Such a resource could accelerate basic research in cancer biology and help refine future AI models for peptide and drug design.
For now, the work remains in the research stage, and more testing will be needed before AI-designed sensors reach clinics or homes. But the study points to a future where a small, invisible molecular cut could be turned into an early, lifesaving warning sign.