CleaveNet

CleaveNet webserver overview

CleaveNet is a deep learning model from Microsoft Research for matrix metalloproteinase (MMP) substrates. The ProteinIQ webserver runs CleaveNet 0.1.2 in three modes: Predict cleavage scores peptides against 18 MMPs, Generate peptides samples new substrates, and Generate from z-scores samples substrates for an MMP cleavage profile you supply.

Prediction returns a z-score and an uncertainty for each peptide and MMP, plus the CleaveNet analysis files for the run. Generation returns one peptide per row and the generated sequence file. The method is described in the CleaveNet preprint and is trained on the mRNA display data of Kukreja et al. (2015).

Pricing

Prediction starts at 15 credits for one peptide and grows with the number of peptides. Generation starts at 6 credits and grows with the number of generated sequences. The exact quote is calculated before submission.

Prediction cost depends only on the number of peptides, not on the predictor architecture or evaluation mode. Generation cost depends only on the total number of sequences, which is Sequences per profile multiplied by the number of profiles in the conditioning CSV. Temperature and Repeat penalty do not change the price.

Inputs

Each line is one peptide and goes to CleaveNet exactly as written. Header lines that start with >, spaces and lowercase letters are rejected with the line number. The CleaveNet README says the model accepts peptides up to 10 residues and has not been robustly evaluated on shorter or longer ones, so a longer peptide is submitted as given and its result may be unreliable or fail.

Shorter peptides are padded with - at the end to the length of the longest peptide in the job. Pad on both ends yourself, for example ---AHA---- or --LRVFL---, when the position of the residues matters.

The conditioning CSV must contain MMP1, MMP10, MMP11, MMP12, MMP13, MMP14, MMP15, MMP16, MMP17, MMP19, MMP2, MMP20, MMP24, MMP25, MMP3, MMP7, MMP8 and MMP9. Names are case sensitive, so mmp2 is rejected. Extra columns are ignored and column order does not matter, but a profile with only some MMPs cannot be used. Each value is rounded to the nearest 0.1 before generation.

MMP1,MMP10,MMP11,MMP12,MMP13,MMP14,MMP15,MMP16,MMP17,MMP19,MMP2,MMP20,MMP24,MMP25,MMP3,MMP7,MMP8,MMP9
0.1,0.0,0.2,0.1,0.3,0.0,0.1,0.1,0.2,0.0,0.1,0.1,0.0,0.2,0.3,0.1,0.0,0.2

A truth CSV with a header can hold any subset of MMPs, and the comparison uses the MMPs it shares with the predictor. Without a header, the file must have the 18 columns in the order of the conditioning header above.

Settings

Mode

Predictor settings

Generator settings

The generator is always the CleaveNet transformer trained on rounded z-scores. The lstm option applies to the predictor only.

Outputs

Prediction fills a Results spreadsheet with one row per peptide and a Files list with the CleaveNet files.

Generation fills the spreadsheet with one row per generated peptide.

Generated peptides can be sent to other ProteinIQ tools as sequences.

Understanding results

The Zscore columns are the mean prediction of five models trained on the same data, and the Uncertainty columns are the standard deviation of those five predictions. A large uncertainty means the models disagree about that peptide and MMP. CleaveNet does not define a cutoff for it, so use it to rank which predictions to trust and to flag peptides worth testing first.

Z-scores are relative. They place a peptide against the other substrates in the Kukreja screen, so they rank peptides and do not give a rate. CleaveNet's own tables count a peptide as cleaved when the predicted z-score is above 0, and allmmp_top20_cleaved.csv counts MMPs with a mean z-score of at least 0. Treat 0 as the line between below-average and above-average cleavage, not as a measured cleavage threshold.

Selectivity needs the other MMPs. A peptide with a high score for MMP13 is a good MMP13 substrate only if the other MMPs score lower, and the MMP family shares many substrates. Read the 18 scores of a peptide together, or use weighted_all_scores.csv, which already subtracts the mean of the other 17 MMPs. The per-MMP _weighted_top50 files rank peptides by that selectivity score instead of by raw cleavage.

With Compare against true z-scores, the MAE and RMSE files give the error of each of the five models per MMP against your measurements. The model was trained on 10-residue peptides and has not been robustly evaluated on shorter or longer ones, so expect larger errors for those. The CleaveNet README notes that the lstm predictor can extrapolate better than the transformer to cleavage patterns missing from the training data.

Generated sequences are samples from the model, and every run gives different ones. Unconditional generation draws from the whole training distribution, and conditional generation steers toward the profile you supply. Raising Temperature gives more diverse peptides. Score generated peptides with Predict cleavage before choosing which to synthesize.

Applications

Predicted profiles help design peptide linkers that release a drug payload when a tumor-associated MMP cuts them, and fluorogenic probes for MMP activity that must separate one MMP from its relatives. Comparing the 18 scores across a set of peptides also shows which positions drive selectivity. For rule-based cleavage sites in full-length proteins, use Peptide Cutter instead.

Limitations

• The model is calibrated for 10-residue peptides.
• Scores are relative cleavage efficiencies, not Km or kcat.
• The training data comes from in vitro mRNA display, so folding, localization and inhibitors in tissue can change the result.
• Generated peptides are model samples, not validated substrates.

Sources

• CleaveNet preprint: bioRxiv 2025.
• Kukreja et al. (2015): mRNA display profiling of MMP substrates.
• Source code: microsoft/cleavenet, release 0.1.2.