MIT Used AI to Design Cancer Sensors That Turn Up in a Pregnancy-Style Urine Test
MIT researchers used an AI model called CleaveNet to design entirely new peptide sensors that detect cancer-linked enzyme activity through a simple urine test, an early step toward a multiplexed at-home diagnostic that could screen for dozens of cancer types.
Researchers at MIT have used an AI system to invent something no chemist designed by hand: peptide sequences engineered specifically to be cut apart by cancer-linked enzymes, releasing a signal detectable in a simple urine test. The work, led by longtime MIT cancer-diagnostics researcher Sangeeta Bhatia in collaboration with Microsoft Research, published in early January 2026, points toward a future where early cancer detection could look less like a hospital MRI and more like a drugstore test strip.
How the AI Actually Designed the Sensors
The system, called CleaveNet, was built to solve a narrow but important problem in cancer diagnostics: certain enzymes called proteases become overactive in and around tumors, and if you can build a peptide, a short chain of amino acids, that gets efficiently sliced apart specifically by one of those overactive proteases, that cleavage event can be turned into a detectable signal. Historically, finding peptides with the right cleavage specificity has been slow, manual biochemistry work. CleaveNet instead generated entirely novel peptide sequences computationally, then had them validated in the lab against MMP13, a protease strongly implicated in cancer metastasis. Notably, the AI-generated sequences that performed best showed high selectivity and efficiency despite never having appeared anywhere in the model’s training data, meaning the system wasn’t just recombining known biology, it was generating genuinely new molecular designs.
From Enzyme to Urine Strip
The practical pipeline works like this: nanoparticles coated with these AI-designed peptides are introduced into the body, where they circulate until they encounter the target protease near a tumor site. When the protease cleaves the peptide, it releases a small reporter fragment that’s cleared into the urine, where it can be detected using a paper-strip test similar in spirit to a home pregnancy test. Sangeeta Bhatia, the John and Dorothy Wilson Professor of Health Sciences and Technology at MIT, described the underlying strategy plainly, explaining that if you can make a sensor out of these proteases and multiplex them, you find signatures of enzyme activity in diseases. The multiplexing part matters, because different cancers and different disease stages tend to activate different combinations of proteases, so a panel of several AI-designed peptide sensors run together could in principle build a signature specific enough to flag disease type, not just its presence.
Who Did the Work
The research paired MIT’s biological engineering expertise with Microsoft’s machine learning group, with lead authors Carmen Martin-Alonso, now at the biotech firm Amplifyer Bio, and Sarah Alamdari of Microsoft Research handling much of the model design and validation work. Ava Amini, a principal researcher at Microsoft and a former MIT graduate student, helped bridge the computational design work with Bhatia’s diagnostics lab, reflecting a cross-institution structure that’s becoming increasingly common as AI-driven molecular design tools move out of pure computer science departments and into applied biomedical labs.
Why Generative Design Beats Trial-and-Error Here
Traditional protease-sensor development has relied heavily on screening large libraries of candidate peptides experimentally, a slow, expensive process that scales poorly when researchers want sensors tuned to many different proteases across many different cancer types. A generative model that can propose promising candidate sequences computationally, ahead of any lab work, compresses that search dramatically, letting researchers test a smaller, better-targeted set of candidates rather than brute-forcing thousands of them on the bench.
The Skeptic’s View
Diagnostics researchers not involved in the project note that a validated result against a single protease target, MMP13, in controlled lab conditions is an early proof of concept, not a finished diagnostic. Moving from a peptide that gets cleaved efficiently in a test tube or animal model to a reliable, FDA-clearable urine test for a specific cancer in a diverse human population typically takes years of additional validation, and false-positive risk from non-cancer-related protease activity, such as inflammation or infection that also elevate certain proteases, remains a real concern that multiplexed panels are meant to address but haven’t yet been proven to solve at scale in humans.
What’s Next
The MIT team is part of a broader ARPA-H-backed initiative aimed at building an at-home diagnostic kit capable of screening for as many as 30 cancer types simultaneously at early, more treatable stages, a considerably more ambitious goal than the single-protease proof of concept published so far. If that multiplexed vision holds up in further testing, the AI-designed sensor approach could become one of the more concrete near-term payoffs of generative AI in medicine: not a chatbot or an image classifier, but a genuinely new physical molecule, designed by software, doing diagnostic work inside the human body.
Photo: RDNE Stock project / PEXELS via Pexels
