Predicting Protein Cleavage Sites and Designing Peptide Substrates on Neurosnap

Figure 1. Human 20S proteasome (PDB 5LE5, native structure at 1.8 Å), shown intact and with a visual cutaway in heavy-atom space-filling representation. Alpha subunits are blue and beta subunits orange. The structure provides biological context; it is not a Pepsickle input or a simulated cleavage trajectory, and no substrate is inserted.

Where does a protein get cut? The answer depends on which cut you mean. Human serum albumin begins as a 609-residue precursor. Its first 18 residues form a signal peptide, but removing them does not yet produce the annotated mature chain: a six-residue propeptide remains. A model can identify the first boundary correctly without describing the whole maturation pathway.

The distinction matters whenever a prediction becomes an experimental construct. You may want to remove a secretion signal, find potential proteasomal cuts near an epitope, generate a list of peptides for mass spectrometry, or design a short substrate for a particular protease. Each asks a different question of a different input. A score from one should not be read as an answer to all the others.

We use recognizable protein sequences to illustrate these choices on Neurosnap. The examples are exploratory software demonstrations, not validation benchmarks. Their value is in seeing what the outputs mean and where their interpretation stops.

First question: where is the signal peptide removed?

Many secreted proteins are synthesized with an N-terminal signal peptide that directs them into the secretory pathway. Signal peptide predictors look for this kind of sequence and its likely cleavage boundary. The sequence after that boundary is a useful starting point for construct design, but additional processing can follow.

UniProt's human albumin entry annotates three consecutive regions in the precursor:

We submitted the complete precursor, the sequence beginning at residue 19, and the annotated mature chain beginning at residue 25 to DeepSig. For the precursor, DeepSig predicted a signal peptide and a cut after residue 18. It classified both signal-removed constructs as “Other,” rather than calling another signal peptide at their new N termini.

Figure 2. The first 40 residues of the submitted albumin and insulin precursor sequences, with UniProt annotations and DeepSig signal-peptide cuts. Positions refer to the respective precursor sequences.

The output deserves a close look before copying it into an order form. DeepSig's file is named mature_sequences.fasta, but the exported albumin sequence starts RGVFRRDAHK. It still contains the propeptide RGVFRR. We call it the signal-cleaved sequence, not the fully mature chain. The annotated mature chain starts six residues later. If the experiment requires the mature protein, use the relevant annotation and check the intended processing rather than assuming the filename describes every subsequent step.

Insulin provides another example. DeepSig predicted signal removal after residue 24 in the submitted precursor. That prediction concerns the signal boundary; it does not describe the later processing that produces mature insulin. The distinction is especially useful when comparing a model output with a reference entry that describes several products of the same precursor.

Neurosnap also offers USPNet-fast, alignment-based USPNet and TSignal for signal-peptide questions. Their inputs and approaches differ, so choose a service suited to the sequence and available context. Our DeepSig result should not be taken as a comparison of their accuracy. In particular, a plausible boundary remains a computational prediction to check against annotation and, where needed, experiment.

Second question: what can be cut in context?

Proteasomal processing presents a different problem. Instead of identifying a secretion signal at the N terminus, we may want to inspect possible cuts throughout a protein or around an antigenic peptide. Pepsickle scores potential proteasomal cleavage positions from sequence. Its model options include an epitope-trained model and models trained on in-vitro constitutive proteasome or immunoproteasome data. These model families reflect different training questions; their scores should be interpreted with that distinction in mind. The Pepsickle paper describes the method and its training context.

Consider SIINFEKL, a well-known ovalbumin epitope. In the submitted UniProt ovalbumin sequence, which includes the initiator methionine, SIINFEKL occupies residues 258–265. The commonly used name OVA257–264 uses a numbering convention that omits that initial residue. Both names refer to the same peptide here; the coordinate system must travel with any plot or construct.

We ran the epitope model on three inputs: full-length ovalbumin, SIINFEKL with eight natural residues on each side (VSGLEQLESIINFEKLTEWTSSNV), and the isolated octapeptide. The flanked sequence also appears in Pepsickle's own documentation, so this is a demonstration of the software, not an independent benchmark.

Figure 3. Pepsickle scores around SIINFEKL. Panel A compares the epitope model across full ovalbumin, the peptide with eight natural residues on each side, and the isolated peptide; all eight epitope scores agree between the full and flanked inputs. The isolated terminal L is marked “No bond” because there is no downstream bond. Panel B compares the epitope, constitutive proteasome and immunoproteasome model families on the flanked input. SIINFEKL occupies positions 258–265 in the submitted UniProt sequence; OVA257–264 is the common alternative name.

The potential cut after the final L of SIINFEKL has a score of 0.754 in both the full protein and the naturally flanked fragment. In the isolated peptide, the terminal L has a score of zero. That does not mean the natural L–T bond is resistant to cleavage: the isolated input contains no downstream T and therefore no L–T bond. The zero is the service convention for a terminal residue with no downstream bond to score.

Other internal scores change when context is removed, too. A short peptide may be convenient to submit, but it changes the question the model sees. If the biological question concerns processing from a parent protein, retain the natural flanks and record exactly which sequence was used. If you deliberately compare fragments, distinguish a missing bond from a low score for a bond that exists.

The model also assigns scores to potential cuts inside SIINFEKL. A high score at one boundary alone is not a predicted yield of intact epitope, a probability of epitope survival, or proof of antigen presentation. Producing and presenting a peptide involves more than identifying one possible proteasomal cut. Use the scores to select positions or sequence contexts for further investigation, not as a substitute for measuring the products.

A practical inset: a theoretical trypsin digest

Sometimes the useful question is simpler: which fragments follow from a specified digestion rule? Pyteomics can generate theoretical peptides for a chosen enzyme and settings. This is a rule-based list, rather than a learned estimate of how susceptible each bond is in a particular sample.

We submitted the annotated mature albumin chain, allowed one missed trypsin cleavage, retained peptides of length 7–30, and requested precursor charge +2. For example, the output includes DLGEENFK, positions 13–20 of the submitted mature chain, with zero missed cleavages and a calculated monoisotopic mass of 950.434526 Da. It also includes FKDLGEENFK, positions 11–20, with one missed cleavage and a calculated mass of 1225.597903 Da.

Figure 4. Eleven of 103 filtered theoretical trypsin fragments, ending by position 80 of the submitted mature albumin chain. Blue indicates zero missed cleavages and orange one missed cleavage. Positions use mature-chain numbering.

The mature-chain coordinate system begins at precursor residue 25, so add 24 to each reported position to map it back to the precursor. The two examples therefore span precursor residues 37–44 and 35–44, respectively. The reported masses assume unmodified peptides. This list is useful for planning and interpreting a mass-spectrometry search; it is not an experimental spectrum or a prediction of peptide abundance. Actual digestion and detection depend on the sample and protocol.

Third question: can we design a substrate?

The direction of the question can also be reversed. Instead of asking where an existing protein might be cut, you might want a short peptide with a particular activity profile across proteases. CleaveNet addresses this design problem for matrix metalloproteinases (MMPs). The Neurosnap service exposes scores for 18 MMPs, transformer and LSTM predictors, unconditional peptide generation, and generation conditioned on target profiles. Its inputs are short peptides of at most ten residues; it is not a full-protein cleavage-site mapper.

The reported relative Z-scores describe predicted profiles, not cleavage probabilities or reaction rates. Ensemble standard deviation describes disagreement among predictors; it is not an experimental error bar. A candidate with an attractive predicted profile is a hypothesis to synthesize and measure, not a demonstrated selective substrate.

In the CleaveNet paper, Martin-Alonso and colleagues describe models trained on relative cleavage profiles from mRNA-display data and report experimental MMP13 substrate-design examples. Those results motivate the design question, but they do not validate an individual Neurosnap candidate.

A useful design exercise would specify the desired MMP profile before looking at candidates, generate a set, inspect the whole set for diversity and predicted uncertainty, and select peptides for a measured cleavage assay. Scoring generated peptides again with the same predictor can help organize the candidates, but it is not independent validation. Experimental measurements are needed to establish activity and selectivity under the conditions that matter.

Choosing a workflow

Start with the biological decision and retain the sequence and coordinate system that decision requires:

1. Define the product or bond. For a secreted protein, distinguish signal removal from later propeptide or prohormone processing. For an epitope, identify both boundaries and the parent sequence.
2. Choose the matching tool. Use a signal-peptide predictor for an N-terminal processing boundary, Pepsickle for proteasomal cleavage scores in sequence context, a theoretical digest for rule-based fragment enumeration, or CleaveNet for short MMP substrate profiles.
3. Keep context and settings. Save the exact input sequence, model, numbering convention, digestion rule, missed-cleavage allowance and relevant output files. A terminal residue in a truncated input cannot stand in for a natural downstream bond.
4. Read the output in its own units. A boundary prediction, a cleavage score, a calculated peptide mass and a relative substrate Z-score are different quantities. None alone establishes the amount of product in an experiment.
5. Test the decision. Check reference annotations and use appropriate experimental measurements to confirm processing, identify products or compare substrate activity.

A good computational result makes the next experiment more focused. For albumin, it helps separate a signal-cleaved construct from the annotated mature chain. For ovalbumin, it shows why the same peptide needs its natural neighbors when the question concerns a bond in the parent protein. For substrate design, it provides candidates and a profile to test. In each case, the most useful prediction is the one matched to a clearly stated cleavage question.

Sources

• DeepSig paper.
• Pyteomics paper.
• UniProt: human albumin P02768, human insulin P01308, and ovalbumin P01012.
• Pepsickle: paper and official repository.
• CleaveNet: Martin-Alonso et al., Nature Communications (2026) and official repository.
• Human 20S proteasome, PDB 5LE5.