FrogPCSP: a propeptide cleavage site predictor for frog antimicrobial peptides

Abstract

Frogs produce and secrete cutaneous antimicrobial peptides (AMPs), which serve as chemical defense against microbial infections and have great biotechnological potential. AMPs are stored in intracellular vesicles in skin glands as propeptides. Before secretion, proprotein convertases (PCs) cleave propeptides into acidic spacer peptides and bioactive peptides. Identifying the correct cleavage site between the acidic spacer and bioactive peptides is a crucial step in AMP prediction. Here, we present Frog Propeptide Cleavage Site Predictor (FrogPCSP), an SVM-based predictor designed to identify propeptide cleavage sites in frog AMPs. The SVM model showed strong performance (global accuracy = 0.981, precision = 0.937, recall = 0.928, F1-score = 0.933, PR-AUC: 0.916) under grouped and stratified 10-fold cross-validation on 424 positive and 2488 negative cleavage sites. Overall, FrogPCSP demonstrated superior performance for propeptide cleavage site prediction of frog AMPs (AUC = 0.990) relative to PSSM (AUC = 0.974) and ProP (AUC = 0.905), a general-purpose reference prohormone cleavage site predictor. As proof of concept, 595 unlabeled frog AMP sequences from UniProtKB/TrEMBL were analyzed. FrogPCSP inferred 926 putative propeptide cleavage sites. Computational physicochemical profiling of the resulting peptides revealed two distinct clusters, one positively charged, with high isoelectric point and strong amphipathicity—consistent with putative bioactive peptides—and another negatively charged, with low isoelectric point and low amphipathicity values—corresponding to putative acidic spacer peptides. Thus, we believe identifying propeptide cleavage sites will assist the discovery and advance the understanding of novel frog AMPs.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-52036-2.

Introduction

Amphibians, especially anurans (frogs), produce and secrete myriad chemical compounds from granular skin glands, including proteins, amines, steroids, alkaloids, and host defense peptides (HDPs; Carrillo et al.¹). Most HDPs are antimicrobial peptides (AMPs) and serve as defensive chemicals against microbial infections^2,3. Multidrug-resistant bacterial⁴ and fungal infections unresponsive to conventional therapies⁵ are considered major public health concerns. Current knowledge indicates that AMPs adopt an amphipathic α-helical structure in aqueous solution and upon interacting with the negatively charged outer leaflet of lipid bilayers (reviewed in⁶. Consequently,, their molecular mechanism of action is receptor-independent, reducing the risk of selecting resistant bacterial strains⁷. Therefore, they have great biotechnological potential to treat these pathologies⁸.

AMPs are synthesized as prepropeptides composed of three regions: (1) a signal peptide, (2) an acidic spacer peptide, and (3) a bioactive peptide⁹. The signal peptide mediates the intracellular transport of the AMP¹⁰. Upon reaching the endoplasmic reticulum (ER), signal peptide are excise from the prepropeptide by signal peptidases, remaining the propeptide¹¹. The acidic spacer peptide maintains the propeptide in an inactive state and is typically negatively charged, with a low isoelectric point, counterbalancing the positively charged (cationic), highly amphipathic bioactive peptide, which generally has a high isoelectric point¹². In frogs, endopeptidases, such as proprotein convertases (PCs), are responsible for cleaving the propeptide into the acid spacer peptide and bioactive peptide prior to its secretion by the granular skin gland¹³. PCs cleavage sites are preferably the C-terminal region of dibasic residues (Arginine [R] and Lysine [K]; Seidah and Chrétien¹⁴). However, the amino acid context surrounding the potential cleavage site influences its recognition by the enzyme¹⁵. Identification of the correct cleavage site between the acidic spacer and the bioactive peptide is a crucial step for AMP prediction¹⁶. However, experimental approaches are highly time-consuming and expensive¹⁷. Therefore, the development of computational tools specifically designed to predict putative PCs cleavage sites is a well-adopted strategy¹⁸. Although standard bioinformatic tools, such as ProP¹⁵, can be employed, they might not be accurate for non-model organism datasets¹⁹. Here, we present Frog Propeptide Cleavage Site Predictor (FrogPCSP), a propeptide cleavage site predictor of frog AMPs based on an SVM-model.

Results

Data pre-processing

Interestingly, an initial manual inspection of UniprotKB/SwissProt AMP cleavage sites showed small inaccuracies in their sequence coordinates (Supplementary Fig. 1A and 1B). Therefore, manual inspection was performed to correct these errors (Supplementary Table 1).

Cleavage sites

As expected, the most frequent amino acid immediately before the correct cleavage sites is Lysine (K), and the most frequent two and three residues upstream to correct cleave sites is glutamic acid (E), as opposed to downstream cleavage site, which does not seem to follow a clear pattern (Fig. 1A).

Fig. 1.

Integrated evaluation of cleavage site prediction. (A) Sequence logo coupled with heatmap showing the relative frequencies of amino acids adjacent to correct cleavage sites. Red indicates acidic amino acids, blue basic amino acids, green polar amino acids, and black hydrophobic amino acids. The heatmap color scale ranges from blue (0) to yellow (1). (B) SVM decision surface using PCA-transformed features. (C) Venn diagram depicting the relationship between cleavage sites across the SwissProt Reference (blue), SVM (magenta) and ProP (green) databases. (D) ROC curve comparing PSSM, ProP and the SVM-model predictions. (E) Calibration curves and brier scores for each cleavage site prediction model. (F) Scatter plot showing the distribution of peptide amphipathicity and isoelectric point, with net charges ranging from − 8 (blue) to + 8 (red) and molecular weights between 2000 Daltons (Da; small circles) and 12,000 Da (large circles). Cross symbols represent outliers (|z-residual| 1.5).

Composition of training and testing datasets

Clustering identity cutoffs did not substantially affect the performance metrics of the model. Although lower cutoff values had higher variance, whereas higher cutoff values had lower variance (Table 1). Therefore, a 90% cutoff was chosen as it provided a suitable trade-off between stability and generalization performance. Clustering outputs are provided in Supplementary Data 1–4. Training and testing datasets generated for each identity cutoff value are provided in Supplementary Tables 2–5.

Table 1.

Model performance metrics under different clustering identity cutoffs.

Identity cutoff	Accuracy	Precision		Recall	F1	σ(K-mers/Fold)
0.70	0.969	0.968	0.968		0.968	± 438
0.80	0.964	0.964	0.964		0.964	± 390
0.90	0.974	0.972	0.973		0.972	± 200
0.95	0.975	0.973	0.974		0.973	± 169

Machine Learning model

The data frame contained 2,488 false cleavage sites and 424 true cleavage sites. After performing a 10-fold grid search, the best estimated parameters were C = 0.5, gamma = ‘scale’, and kernel = ‘rbf’. These parameters yielded the highest AUC of 0.97. Due to the unbalanced data, the model was trained and evaluated using a grouped stratified 10-fold cross-validation approach, achieving global performance metrics above 90% (pooled results: accuracy = 0.981, precision = 0.937, recall = 0.928, F1-score = 0.933, PR-AUC: 0.916) (Table 2). Fold-specific and pooled precision-recall curves, together with the global PR-AUC, are presented in Supplementary Fig. 2. Overall, the model exhibited low rates of false positives and false negatives and displayed a well-defined decision surface (Fig. 1B). Comparison across taxa indicated that the model achieved better performance metrics for the families Ranidae, Hylidae, and Bombinatoridae (Table 3). An example of a successfully cleaved peptide is shown in Supplementary Fig. 3.

Table 2.

Cross-validation performance metrics across folds.

Fold	Accuracy	Precision	Recall	F1-score	PR-AUC	95% CI
1	0.962	0.893	0.926	0.909	0.992	0.970-1.000
2	0.980	0.978	0.898	0.936	0.903	0.787-1.000
3	0.962	0.960	0.774	0.857	0.798	0.634–0.933
4	0.994	0.978	0.978	0.978	0.900	0.782-1.000
5	0.986	0.969	0.939	0.954	0.947	0.867-1.000
6	0.992	0.962	0.981	0.971	0.977	0.941-1.000
7	0.973	0.932	0.891	0.911	0.889	0.772–0.981
8	0.975	0.889	0.941	0.914	0.909	0.760–0.999
9	0.970	0.825	0.940	0.879	0.884	0.768–0.975
10	0.994	1.000	0.961	0.980	0.979	0.948-1.000

Table 3.

Leave-one-out performance metrics across taxa.

Family	Test set	Accuracy	Precision	Recall	F1-Score	PR-AUC	95% CI
Dicroglossidae	35	1.000	1.000	1.000	1.000	1.000	1.000–1.000
Myobatrachidae	14	1.000	1.000	1.000	1.000	1.000	1.000–1.000
Ranidae	936	0.987	0.947	0.962	0.954	0.960	0.926–0.986
Bombinatoridae	796	0.879	0.492	0.957	0.650	0.941	0.889–0.982
Hyperoliidae	35	0.886	0.556	1.000	0.714	0.938	0.644-1.000
Hylidae	742	0.945	0.824	0.844	0.834	0.921	0.883–0.952
Leptodactylidae	75	0.893	0.875	0.500	0.636	0.868	0.659-1.000
Phyllomedusidae	41	0.976	0.857	1.000	0.923	0.829	0.422-1.000
Pelodryadidae	91	0.978	0.926	1.000	0.962	0.757	0.613-1.000
Pipidae	139	0.957	1.000	0.571	0.727	0.658	0.413–0.862
Rhacophoridae	8	0.875	0.000	0.000	0.000	0.100	0.071–0.333

Benchmark

Reference, PSSM and SVM-predicted cleavage sites were non-redundant compared to those inferred by ProP. For example, ProP identified both SICEEEK|RE and ICEEEKR|EE k-mers, whereas in the Reference, PSSM, and SVM datasets, only the k-mer with the strongest cleavage signal (ICEEEKR|EE) was retained. Consequently, ProP inferred a larger number of cleavage sites than the Reference, PSSM and SVM datasets. However, most cleavage sites predicted as positive by ProP were not present in the reference dataset(Fig. 1C). As a result, the SVM model achieved higher true positive rates and lower false positive rates than ProP (Fig. 2D). Additionally, the SVM model presented lower brier score and better calibration curve than ProP and PSSM, demonstrating superior performance for cleavage site prediction of frog AMPs (Fig. 1E).

Fig. 2.

General workflow for cleavage site model development. Manually curated UniProtKB/Swiss-Prot sequences and annotations are retrieved and clustered based on sequence identity. For each sequence, 7-mer windows flanking putative cleavage sites are computationally generated and encoded using one-hot encoder. The cleavage site prediction model is then developed through cross-validation and comprehensive model evaluation. Benchmarking is performed to compare FrogPCSP with the well-established ProP tool. Finally, a proof-of-concept analysis is conducted using UniProtKB/TrEMBL frog AMP precursor sequences.

Prediction of cleavage sites in unlabeled data

As proof-of-concept, 595 AMP automatically curated sequences from Uniprot (TrEMBL) were analyzed by the SVM-model. As a result, 5550 cleave sites were considered incorrect and 623 were considered correct. Signal peptide sequences were also excluded using a Hidden Markov Model (HMM) built based on signal peptide sequences of DADP. To better understand the physicochemical properties of the cleaved peptides, their putative molecular weight, isoelectric point, net charge, and amphipathicity were assessed. As a result, it was found that peptide amphipathicity and isoelectric point values are linearly correlated (r = 0.99, p < 0.001), with positively charged peptides exhibiting higher values of both parameters than negatively charged peptides. Only 38 propeptide fragments were classified as outliers based on standardized residuals (|z-residual| 1.5; Fig. 2F and Supplementary Table 6).

Discussion

Although UniProtKB/SwissProt is a manually curated database, minor inaccuracies in cleavage site annotations were identified (Fig. S1) and corrected. This finding highlights the importance of interpreting manually curated annotations with caution. Analysis of amino acid composition revealed that residues upstream the cleavage site tend to be acidic, particularly glutamic acid (E), with the conserved motif EEKR (Fig. 1A). This result is consistent with the processing quadruple motif (PQM), whereby an arginine residue at position − 1 partners with a glutamic acid residue at positions − 2, − 3, or − 4, suggesting that acidic residues preceding basic amino acids such as lysine or arginine might serve as a stronger recognition signal for PCs in frog AMPs, as toxin observed in spider venom, for example²⁰.

The limited availability of propeptide cleavage site prediction tools represents a constraint for baseline comparison. Nevertheless, the SVM model accurately predicted both correct and incorrect cleavage sites, showing low rates of both false positive and false negatives and a well-defined decision surface (Fig. 1B). Although the AUC values of the SVM and PSSM models were similar (Fig. 1D), the SVM achieved a lower Brier score, indicating better probability calibration and greater predictive stability (Fig. 1E). These results indicate that FrogPCSP is a reliable tool for predicting propeptide cleavage sites in frog AMPs, and underscores the importance of developing tools tailored to frog AMPs rather than relying on general-purpose predictors.

Although this tool produced promising results, the dataset was highly imbalanced across families, with Ranidae, Hylidae, and Bombinatoridae accounting for approximately 85% of all k-mers, while several families are underrepresented with less than 5% of k-mers Pipidae, Pelodryadidae, Leptodactylidae, or even less than 2% (Phyllomedusidae, Dicroglossidae, Hyperoliidae, Myobatrachidaem, Rhacophoridae) (Table 3). This distribution reflects the unbalanced representation of studies across taxonomical groups²¹. Therefore, the model appears to be more realistic for the most represented families, while predictions of propeptide cleavage sites for underrepresented groups should be interpreted with caution.

Bioactive peptides are generally amphipathic and positively charged molecules, whereas acidic spacer peptides are typically negatively charged with low isoelectric points²². Our analysis clearly distinguished two groups of peptides with these characteristic features (Fig. 1F), despite the presence of a limited number of outliers (38 in total; Supplementary Table 6). A more detailed examination of these outliers revealed three distinct clusters (Supplementary Fig. 4). The first cluster was characterized by high isoelectric point, amphipathicity, net charge, and molecular weight, with putative available cleavage sites, suggesting possible miscleavage events, whereas the second and third clusters showed intermediate and lower values, respectively (Supplementary Fig. 5). This pattern indicates that, in most cases, FrogPCSP accurately predicted the correct cleavage site between the acidic spacer peptide and the bioactive peptide of the putative frog AMPs. Therefore, we believe that FrogPCSP will assist the discovery and advance the understanding of novel frog AMPs.

Methods

The general workflow consisted of clustering UniProtKB/Swiss-Prot sequences based on sequence identity to prevent data leakage. Subsequently, sequence fragments of seven residues centered around putative cleavage sites (7-mers) were generated. These 7-mers were encoded using a one-hot encoding scheme and used to train the cleavage site prediction model. Finally, the trained model was applied to UniProtKB/TrEMBL sequences as proof of concept (Fig. 1).

Data acquisition and pre-processing

1089 manually curated AMP sequences and their annotation were obtained from Uniprot/SwissProt database (12/09/2025) using the following search prompt: (taxonomy_id:8342) AND (keyword: KW-0878) AND (reviewed: yes). Only sequences with complete AMP structures (signal peptide, acidic spacer peptide, and bioactive peptide) were used in the subsequent analysis, leaving 348 peptide sequences. Additionally, 595 automatically curated AMP sequences from Uniprot/TrEMBL database (12/09/2025) were used as proof of concept of the tool. To download these sequences, the following search prompt was used: (taxonomy_id:8342) AND (keyword: KW-0878) AND (reviewed: no). Manual curation was performed to correct coordinate errors on the UniprotKB/SwissProt annotated cleavage sites (see Data Pre-processing Results).

Data processing

Putative cleavage sites were searched through a sliding-window along the prepropeptide sequences. Next, short sequences (k-mers) of seven residues (7-mers) were artificially generated around the putative cleavage sites (e.g., the sequence LGKRYGG, where LGK is upstream to the cleavage site and YGG is downstream). Redundant putative cleavage sites were not included (e.g., for LGKR|YGG and ELGK|RYG there are two putative cleavage sites, indicated by the pipe symbol “|”). Annotated cleavage sites (UniProtKB/Swiss-Prot) were considered true and those not annotated were considered false. The k-mer amino acid sequences were coded using one-hot-encoder from the Python library Scikit-learn version 1.7²³. This method converts each amino acid into a numerical vector of length 20, representing the 20 standard amino acids (e.g., the amino acid glutamic acid is represented by the vector [0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]). Signal peptide sequences were removed prior to the development of the model using the UniProtKB/Swiss-Prot annotation as reference.

Composition of training and testing datasets

To avoid sequence redundancy and data leakage, homology filtering was performed at the prepropeptide peptide level. Prepropeptide sequences were clustered with CD-HIT version 4.8.1²⁴ at 70%, 80%, 90% and 95% identity with the command-line: cd-hit -i input.fasta -o output.fasta -c [0.7, 0.8, 0.9, 0.95]. All k-mers derived from a given cluster were assigned to the same fold. Homology control was applied prior to k-mer extraction, ensuring that no sequence segment derived from a precursor in the training folds appeared in the test fold.

Machine learning model construction and evaluation

A support vector machine (SVM) model was developed, as this method is widely applied to the identification of protease cleavage sites²⁵ and has been successfully used to predict propeptide cleavage sites in spider toxins¹⁹. SVMs are also suitable for small training datasets, as in the present study. Hyperparameters were optimized and performance metrics were computed using grid search with 10-fold stratified cross-validation. The search space included (1) C ∈ {0.5, 1, 10, 100}, (2) γ ∈ {‘scale’, 1, 0.1, 0.01, 0.001}, (3) kernel ∈ {linear, rbf, sigmoid, poly}, and (4) degree ∈ {2, 3, 4, 5} for the polynomial kernel.

Model performance (precision, accuracy, recall, and F1-score) was estimated via 10-fold stratified cross-validation, preserving the class balance of true and false cleavage sites. To assess the uncertainty of the performance estimates, a bootstrap resampling strategy was applied to the precision-recall area under the curve (PR-AUC). For each evaluation set, the true labels and predicted scores were resampled with replacement 1,000 times, generating bootstrap samples with the same size as the original dataset. For each resampled dataset, the precision-recall curve was recalculated, and the corresponding PR-AUC was estimated.

The model performance was also evaluated using the leave-one-group-out (LOGO) cross-validation approach to estimate the generalization of the model across different taxa. In this method, the dataset is divided into training set with data from all families except one, and testing set with data of the family excluded from training. This process is iteratively repeated until all taxa are evaluated. Global performance metrics were computed by pooling predictions from all folds, ensuring unbiased estimates. The final model and encoder were saved as .pkl files using the Python library Pickle version 3.14 (https://docs.python.org/3/library/pickle.html).

Benchmarking

To evaluate the predictive performance of the SVM model, a benchmarking analysis was conducted against ProP version 1.0¹⁵, a well-established propeptide cleavage site predictor, and position-specific scoring matrix (PSSM) model.

A PSSM-based approach was evaluated using 10-fold stratified group cross-validation. In each fold, the PSSM was built only from positive training 7-mers containing basic cleavage motifs, using pseudocount-corrected position weights converted to log2 odds scores. K-mers were scored by summing position-specific log-odds values, and these scores were calibrated into probabilities with logistic regression trained only on the corresponding training fold. Test samples were classified using a probability threshold of 0.5, generating out-of-fold predictions for downstream evaluation.

Predictions from both approaches were aligned by sequence identifier and cleavage site position. Performance metrics—including accuracy, precision, recall, F1-score, and AUC—were computed based on experimentally validated cleavage sites. Receiver operating characteristic (ROC) curves were also generated to visualize comparative performance. Probability calibration was assessed using the Brier score, which measures the mean squared difference between predicted probabilities and true binary outcomes. Calibration curves were also generated by grouping predictions into quantile-based bins and comparing the mean predicted probability with the observed fraction of positives in each bin.

Prediction of cleavage sites in unlabeled data

Automatically curated AMP sequences from UniProtKB/TrEMBL database were used as proof-of-concept of FrogPCSP. A Hidden Markov Model (HMM) of AMP signal peptides was built using signal peptide sequences from DADP database as reference²⁶ and the software HMMER version 3.4²⁷. To achieve this, signal peptide sequences from the DADP database were first aligned,, using MUSCLE version 5.3²⁸. The resulting multiple sequence alignments were subsequently used to build the HMM profile. This model was subsequently applied to screen the UniProtKB/TrEMBL database sequences for the identification of putative AMP signal peptides.

The following general command-lines were used:

muscle -align signalp.fa signalp.afa.

hmmbuild signalp.afa signalp.hmm.

hmmsearch -E 1e-5 --domtblout results/database.domtblout --cpu 32 -o results/database.signalp.log database/signalp.hmm database/database.fasta.

After removing signal peptide sequences from the original data, the remaining prepropeptide sequences were used as input to the model. Additionally, putative physicochemical properties of the resulting fragment sequences (isoelectric point and amphipathiciy) were computed using the Python libraries Biopython version 1.87²⁹ and modlamp version 4.3.2³⁰.

The isoelectric point computes the pH at which the net peptide charge is zero based on the pKa values of the N- and C-termini and ionizable side chains. The total charge is calculated across different pH values using the Henderson-Hasselbalch equation, which describes the protonation state of each ionizable group as a function of pH and pKa:

The default pKa values include 3.55 for the C-terminus, 7.50 for the N-terminus, 4.05 for Asp, 4.45 for Glu, 9.00 for Cys, 10.00 for Tyr, 5.98 for His, 10.00 for Lys, and 12.00 for Arg.

For amphipathicity, each residue was assigned with a hydrophobicity value according to the Eisenberg scale. The sequence was projected at a 180º angle (alfa-helix), and the summed vectors were divided by the sequence length., and can be expressed with the equation:

Where:

represents the hydrophobicity of the residue at the position \:i.

represents the residue angle (180º).

\:n represents the sequence length.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: da Silva, E. M. G. Data collection: da Silva, E.M.G. Data analyses: da Silva, E. M. G. Writing (first draft): da Silva, E. M. G. Writing (review and editing): da Silva, E. M. G and Grant, T. Funding acquisition and supervision: Grant, T.

Funding

This research was funded by the Brazilian Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP; grants #2018/15425-0 to Grant, T. and #2024/21125-0 to da Silva, E. M. G.).

Data availability

All Python scripts for FrogPCSP are publicly available without restriction at: https://github.com/Matheusdras/FrogPCSP.

Declarations

Competing interests

The authors declare no competing interests.