A cleavage-site predictor takes a sequence and returns a ranked list of scissile bonds. It does not answer the question you actually have, which is whether a given bond will be hydrolysed in your tube, on your timescale, in your matrix. Keeping those two things apart — what the models encode, what their training sets can support, and what only an experiment settles — is the difference between using a prediction to design an assay and mistaking it for the assay's result.
The subsite grid the models encode
Every sequence-based predictor is a model of the same object: an extended substrate-binding cleft divided into subsites. In the nomenclature in use since 1967, subsites of the active site are numbered S1, S2 and onwards, and the substrate residues occupying them P1, P2 and onwards. The scheme followed from the finding that protease active sites are larger than had been expected, with important interactions occurring in regions remote from the catalytic site — predictions later confirmed by crystallography, and a model still used to evaluate enzyme activity [1]. Residues on the carboxyl side of the scissile bond and the pockets holding them carry primes, so specificity is properly reported over a window straddling the bond rather than at one position: proteome-derived library screens read sequence preference on the non-prime (P) and prime (P') sides simultaneously, out to P6 and P6' [2].
Predictors are built by turning collections of observed cleavages into position-specific frequencies over that window. MEROPS assembles cleavages from the literature and renders each peptidase's specificity as a logo, a distribution matrix and tables summarising which residues are acceptable or unacceptable in each binding pocket, for those peptidases with ten or more known substrate cleavages [3]. Two experimental technologies are prominent sources of that density. Soluble positional fluorogenic substrate libraries deliver subsite preferences directly: libraries of 137, 180 and 6,859 members, with amino acid diversity at P4-P3-P2-P1 and P4-P3-P2, were used to profile thrombin, plasmin, factor Xa, urokinase-type and tissue-type plasminogen activator, granzyme B, trypsin, chymotrypsin, human neutrophil elastase, papain and cruzain [4]. Proteome-derived peptide libraries read both sides at once: prime-side cleavage products are biotin-tagged, isolated and identified by tandem MS, and the corresponding non-prime sequences are recovered from proteome databases — which also makes subsite cooperativity visible rather than forcing pockets to be treated as independent [2].
Specificity is a preference distribution, not a rule
What comes out of that machinery is a distribution over each pocket, and reading it as a rule is the commonest way to misuse a prediction. An information-theoretic treatment makes the point quantitative: computing an entropy value for each subpocket from MEROPS cleavage data and summing the subpocket contributions yields a measure of overall substrate specificity that ranks proteases on a continuum, separating unspecific digestive enzymes showing high total cleavage entropy from the specific proteases involved in signalling cascades [5]. Nothing on that scale is binary.
Even rules that look safe are frequency statements. The Keil rule for trypsin — cleaves next to arginine or lysine, but not before proline — was derived from a small number of experimentally confirmed cleavages and is built into peptide identification algorithms as a filtering criterion. Re-examined against 14.5 million tandem spectra, it did not hold: the analysis revealed a surprisingly large number of cleavages before proline, and the authors argued that trypsin specificity rules used in identification algorithms should be modified to legitimatise those cleavages, and that specificity rules for other enzymes should be re-evaluated the same way against large MS/MS datasets [6]. The generalisation is worth internalising. A residue that a matrix scores as forbidden is usually a residue that is rare, and rare is a statement about rate, not about possibility.
What the predictors actually do
The first generation scored candidate bonds against a frequency matrix and ranked them. PoPS let a user assemble a specificity model from any available experimental data or expert knowledge and use it to predict and rank likely cleavages within a single substrate or across entire proteomes; its authors noted that computational attempts of the period were limited by representing cleavage sequences as patterns or frequency matrices, and that other factors such as the secondary or tertiary structure of the substrate could be used separately to screen out unlikely sites [7].
The second generation folded predicted structural context into the feature set itself. PROSPER covers twenty-four proteases using specificity information derived from MEROPS and encodes, alongside the local amino acid sequence profile, predicted secondary structure, predicted solvent accessibility and predicted native disorder [8]. Where the size of that gain has been reported, it is real but bounded. For caspase substrates, Cascleave built on local sequence-derived profiles reached a Matthews correlation coefficient of 0.667; incorporating predicted solvent accessibility and whether the cleavage sequence lies in a region most likely to be natively unstructured, together with bi-profile Bayesian signatures, produced a best model at 0.747 [9]. That is a worthwhile gain, not a change of category.
The current generation reflects data outgrowing tools. Substrate cleavage data now exist for more than a hundred protease types and keep accumulating, which makes publishing a separate predictor for each new protease impractical; ProsperousPlus instead supplies a platform on which a user trains, empirically assesses and deploys a model for the substrate type they care about, with benchmarking that reports predictors on average exceeding current state-of-the-art performance [10]. A parallel line abandons sequence patterns altogether: a protein graph convolutional network represents the protease–substrate complex as a structure-based molecular interaction graph describing topology and interaction energetics, predicts the specificity landscapes of several variants of two model proteases, and can accurately assess designs featuring diversity at positions not present in the training data [11]. The stated motivation for that work is precisely the limitation under discussion — that current methods are limited to sequence pattern recognition in cleavage data generated separately for each protease variant [11].
The training set is the ceiling
Before weighting a score, ask what the model for your protease is standing on. The MEROPS substrate collection had accumulated 66,615 cleavages, with at least one cleavage known for 1,700 of 2,457 peptidases. Specificity profiles are shown for the 556 peptidases with ten or more known cleavages — but it has been proposed that at least forty cleavages in disparate proteins are required for specificity analysis to be meaningful, and only 163 peptidases, 6.6% of the total, fulfil that criterion [3]. For a protease outside that fraction, a clean-looking matrix is an extrapolation from a handful of events, and those events are the substrates investigators happened to choose, test and publish.
MEROPS ships a partial correction for the resulting noise: a tool that assesses whether a cleavage site is conserved across orthologous sequences, on the hypothesis that a site conserved in orthologues is more likely to be physiologically relevant, counting a substitution as conserved only where the substituting residue is known to occupy the same binding pocket in another substrate cleaved by the same peptidase [3]. That is a filter on physiological relevance rather than on chemical possibility; it does not tell you a bond can be cut, only that if it is cut, the site has been kept.
A deeper limitation is that most collections record cleavage as a binary event and discard the rate. Quantitative mass spectrometry applied to hundreds of natural protein substrates in cell lysate found cleavage rates for caspase-2 and caspase-6 varying up to 500-fold, and little correlation in catalytic efficiency among five caspases across shared substrates, suggesting each has a unique set of preferred substrates rather than the shared repertoire that pooled substrate lists imply [12]. A classifier trained on presence and absence cannot reproduce a spread it never saw, which is why two sites carrying similar scores can differ enormously in how fast they are actually cut.
The variable sequence cannot see: accessibility
A predicted site is a sequence match. Whether the backbone at that position can reach the catalytic machinery is a separate question. Limited-proteolysis work on folded proteins established the principle: the sites of limited proteolysis along a polypeptide chain are characterised by enhanced backbone flexibility, so proteolytic probes pinpoint the positions of local unfolding in that chain [13].
On a peptide, the cleanest demonstration is the cyclotide kalata B1, which combines a head-to-tail cyclised backbone with a cystine knot core. Following incubation with a range of proteases, native kalata B1 showed no evidence of enzymatic digestion when monitored by LC-MS; the reduced form — same sequence, disulfides broken — was susceptible to proteolysis by trypsin, endoproteinase Glu-C and thermolysin [14]. A sequence-only predictor scores both states identically.
Some tools attempt the correction. Procleave represents structural features of known cleavage sites as discrete values using a LOWESS data-smoothing optimisation the authors report as critical to performance, encodes them in a conditional random field alongside sequence and chemical group-based features, and was applied across a human structural proteome of 17,628 protein structures [15]. The correction is only as good as the structure you can supply or predict, which for a synthetic peptide in solution is frequently no structure at all.
The accessibility argument should not be over-extended. In the same quantitative caspase study, peptide substrates synthesised on the basis of the identified protein cleavage sites showed catalytic efficiencies similar to those of the protein substrates, which the authors read as the rates of proteolysis being dominated more by local primary sequence and less by the tertiary protein fold [12]. For a short linear peptide with no persistent structure, sequence carries most of the signal; for a disulfide-constrained or cyclised one, it does not.
A cleavage site is not a half-life
These are different measurements, and the second does not follow from the first. A prediction names a bond; a stability assay measures disappearance of the parent in one specific matrix under the combined action of every active protease in it. How far the matrix governs that answer was shown by drawing blood from CD-1 mice by cardiac puncture directly into a syringe already containing peptide solution, so that coagulation activators such as glass surfaces and damaged cells were absent and degradation could be followed in unclotted blood for one hour, then comparing against serum derived freshly from the same animal and against commercial serum and plasma. Peptides of three different families were generally degraded faster in serum than in plasma, all were more stable in fresh blood than in either, and the order of degradation rates among the peptides varied across the six different incubation experiments [16].
Behind that spread is the fact that the spectrum of active or activated proteases appears to vary with how the matrix was prepared: anticoagulants inhibit particular classes — EDTA metallo- and Ca2+-dependent proteases, heparin enzymes such as thrombin and factor Xa — while coagulation activates others [16]. The practical warning the authors draw is that degradation measured this way may be misleading, stimulating effort to stabilise a peptide at sites relevant only to the serum or plasma assay and of lower importance in the animal [16]. A predictor queried for one purified protease has access to none of this, which places cleavage-site prediction upstream of a plasma or serum stability assay rather than in place of one.
Validating a prediction by MS of the digest
The test of a predicted site is the mass of the products. Run the digest with the protease of interest, quench at several time points, and analyse by LC-MS: loss of parent gives the rate and the product masses give the position. Because hydrolysis of one amide bond splits the peptide into two fragments whose masses sum to the parent plus water, a predicted site is confirmed when the complementary mass pair appears, and a predicted site with no such pair is simply not being used under those conditions. This is how peptide degradation products were identified in the blood, serum and plasma comparison above [16], and how the absence of digestion of the native cyclotide was established [14].
Two complications arrive as soon as the sample is more than one peptide and one enzyme. Prime- and non-prime-side products behave differently in a proteomic workflow, which is why the proteome-derived library design tags and isolates the prime-side products for tandem MS and derives the non-prime sequences bioinformatically [2]. That design also shows how sharp a genuine profile can be: for the highly specific GluC protease, more than 95% of the 558 cleavage sites identified displayed the canonical selectivity, whereas for the broad-specificity matrix metalloproteinase 2 the same approach identified more than 1,200 peptidic cleavage sites [2].
The second complication is background proteolysis. N-terminomic methods handle it by labelling terminal amines and depleting internal peptides — dendritic polyglycerol aldehyde polymers remove tryptic and C-terminal peptides, and peptide isotope quantification discriminates substrates of the protease of interest from products of background proteolysis, an approach that identified 288 MMP-2 cleavage sites in mouse fibroblast secretomes [17]. Incubating a single peptide in plasma rather than in buffer raises the same problem at smaller scale: a fragment you observe is not automatically the work of the enzyme you had in mind, and matrix-only and enzyme-free controls are what separate the two.
Using a prediction properly
Treat the ranked output as a list of hypotheses and a guide to where to look in the chromatogram, not as a verdict. Check the evidence behind the specificity model for your protease before weighting its score, since the number of known cleavages underpinning a profile varies by orders of magnitude between the well-studied enzymes and the rest [3]. If the peptide is conformationally constrained, expect the sequence-only score to be wrong in a predictable direction, and run the reduced form alongside the native one to see it [14]. Then keep the questions apart: what a single purified enzyme does to a defined sequence is the subject of protease specificity profiling, and what a biological matrix does to a peptide over time is the subject of a plasma or serum stability assay. A cleavage-site prediction is an input to designing both experiments and a substitute for neither.