CleaveNet

Proteases cut peptide bonds, and a peptide that a chosen protease cuts efficiently — and that its close relatives leave alone — is the working part of an activity-based sensor, an activatable prodrug, or a protease-triggered diagnostic. Finding one is hard for a combinatorial reason: a 10-residue peptide has roughly 20^10 possible sequences, and the screening methods that can survey any meaningful fraction of that space are slow and available to few labs. Selectivity is harder still, because matrix metalloproteinases share overlapping substrate preferences.

CleaveNet turns that search into a conditional generation problem. Developed by Carmen Martin-Alonso, Sarah Alamdari, Tahoura Samad, Kevin Yang, Sangeeta Bhatia, and Ava Amini at Microsoft Research and MIT, and published in Nature Communications in January 2026 after a March 2025 preprint, the pipeline couples two small transformers. A predictor maps a candidate peptide to continuous cleavage scores for 18 matrix metalloproteinases; a generator writes new 10-mers autoregressively, conditioned on a cleavage profile supplied by the user.

That conditioning tag is the design lever. Instead of generating peptides and filtering afterwards, a researcher specifies the profile they want — high cleavage by one protease, low by the other seventeen — and the generator produces sequences aimed at it directly. The paper's hardest test was MMP13 selectivity, a case where sequence preferences overlap heavily with MMP1, MMP8, and other collagenases.

#Key Features

• Profile-conditioned generation: The desired cleavage scores for all 18 proteases, rounded to the nearest tenth, replace the start token, so the target profile steers decoding from the first residue rather than acting as a post-hoc filter.
• Joint conditional and unconditional training: The generator trains on a 50/50 mix, letting the same weights sample freely from the learned substrate distribution or follow a specified profile.
• Ensembled cleavage predictor: Multiple seeds give uncertainty estimates alongside the multi-output regression, so borderline designs can be flagged before synthesis.
• Novel motifs, not memorized ones: Generated peptides recover established cleavage motifs and also produce motifs absent from the training screen, while retaining sound biophysical properties.
• Experimentally validated designs: A 95-substrate fluorogenic screen against 12 recombinant proteases tested the generated sequences directly rather than relying on held-out prediction accuracy.

#Technical Details

Training used the Kukreja mRNA-display dataset: roughly 18,500 synthetic 10-mer peptides profiled against 18 matrix metalloproteinases, with a held-out test set of 2,901 homology-filtered sequences plus 71 sequences characterized by fluorescence assay. Both networks are deliberately small. The predictor is a two-layer encoder-only transformer with model dimension 32 and 6 attention heads, about 56,000 parameters, performing multi-output regression over the 18 protease Z-scores. The generator is a three-layer decoder-only transformer with model dimension 64 and 6 attention heads, about 328,000 parameters; LSTM variants of both were trained as alternatives and are shipped alongside. On held-out data the predicted Z-scores correlated strongly with measured values (Pearson r = 0.80 for MMP13), and thresholded classification reached AUC 0.98 at a Z-score cutoff of 2.5.

Generation produced 20,000 unconditional sequences for distributional analysis, plus conditional sets aimed at MMP13 efficiency and MMP13 selectivity. Ninety-five FRET-paired fluorogenic substrates — 40 selected for efficiency, 40 designed for selectivity, and 15 mRNA-display controls — were assayed against 12 recombinant proteases. Among the selectivity designs, all 24 CleaveNet substrates were cleaved by MMP13, against 1 of 8 from a site-independent baseline, and three were cleaved by MMP13 alone.

#Applications

The immediate beneficiaries are groups building protease-activated tools: activity-based nanosensors, imaging probes, prodrugs, and engineered cell therapies, all of which need a peptide linker that responds to one protease and not its neighbors. Rather than screening a library and hoping a usable substrate falls out, a researcher states the profile they need and synthesizes a short list. The conditioning mechanism carries no assumptions specific to metalloproteinases, so the same recipe applies to any enzyme family for which a comparable activity-profiling dataset exists.

#Impact

CleaveNet is peer reviewed, and its code, model weights for both architectures across five random seeds, generated sequences, computed metrics, and raw in vitro data are released together on GitHub under an MIT license, with a Zenodo archive for citation. The validation is unusually direct for a generative sequence model: designs were synthesized and assayed rather than scored against a held-out split alone. The constraints are equally clear — the model is trained on one mRNA-display dataset covering one protease family, generates only 10-residue peptides, and predicts cleavage on isolated recombinant enzymes rather than in the crowded proteolytic environment of tissue. The authors position it as a template for in silico design across enzyme classes, which will depend on comparable profiling data becoming available for those classes.