Researchers led by Ludwig MIT’s Sangeeta Bhatia reported in a January issue of Nature Communications a deep learning model for the design of peptides that are clipped with high specificity by defined proteases, a useful tool for cancer therapeutics and detection. Capitalizing on the tendency of cancers to overexpress particular proteases—which help enable metastasis, among other things—the Bhatia lab has over the past 10-plus years developed nanoparticles coated with peptides as diagnostic reporters and established proof of concept for their use in detecting lung, ovarian and colon cancers. The peptide substrates of proteases are roughly 10 amino acids long, which works out to 200 billion possible sequences for each. To work as diagnostic reporters, the peptides must be cleaved both efficiently and selectively by specific proteases. Previously, Sangeeta and her colleagues employed a trial-and-error approach to identify candidate peptides that was informed by published sequences. The end-to-end AI pipeline presented in their paper, CleaveNet, dispenses with that approach. The researchers showed that when applied to matrix metalloproteinases (MMPs), CleaveNet dramatically improved the scale, tunability, and efficiency of substrate design and, in the case of MMP13, captured not only known cleavage motifs but also many that hadn’t yet been identified. Aside from its utility in cancer diagnostics, CleaveNet could also be used to design peptide linkers that release antibody-guided drugs exclusively into tumors.
Deep learning guided design of protease substrates
Nature Communications, 2026 January 6