Data collection, Cys framework and pattern analysis

The methodology in this study involved collecting mature toxin and protein precursor sequences from ConoServer, a dedicated database for the sequences and structures of conopeptides [29]. A total of 5,985 sequences from 119 Conus species were gathered. Each sequence was analyzed for amino acid variations at species level using a Python script, and they were organized into individual files. For example, sequences for Conus geographus were saved in a file named Conus_geographus58.txt within the sequence folder, containing 58 sequences for this species. The methodology for Cys framework prediction and pattern analysis is illustrated in Fig 1. Multiple analyses were conducted on the dataset to identify Cys frameworks and their patterning. These steps included determining the number of Cys residues and performing count-based categorization. Cys connectivities were predicted using an in-house cone snail PDB library, and sequences with similar Cys frameworks along with their patterns and connectivities were grouped into a separate file (pattern.csv) through comparative analysis.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Flowchart scheme of Cys pattern identification, framework prediction and connectivity analysis.

The arrow directions indicate the flow of work. Cone snail peptides isolated through ConoServer database were subjected to Cys pattern and framework evaluation. The sequences were categorized based on shared patterns and frameworks, and analyzed for connectivity differences among cone snail species.

https://doi.org/10.1371/journal.pone.0327578.g001

Cys pattern identification

After data collection, each sequence underwent analysis to count Cys residues and tag their positions. To simplify this process, intervening residues were replaced by “X.” Python modules Pandas [30] and RegEx [31] facilitated this task: Pandas enabled the data to be stored and managed in a two-dimensional, tabular structure [32], while RegEx matched text based on preset patterns, dividing each pattern into one or more sub-patterns to detect specific sequences. The input was the sequences folder, and the output was a CSV file with detailed sequence information.

Cys pattern classification

To identify unique Cys patterns, sequences were classified as Cys-rich or Cys-poor, and sequences without any Cys residues were excluded. All potential Cys patterns were computed, revealing common patterns among cone snail species. Previously reported patterns from the literature [33] were also compared to annotate any biologically relevant peptides. A Python script containing these known patterns was applied to each sequence, adding the pattern as a dictionary key and identifying both matched and novel patterns. Unique and matching patterns were stored in separate files, which were then compared to identify known and novel Cys patterns. The final patterns were grouped in CSV files based on identical Cys patterns, and redundancy was minimized by removing duplicate patterns within species. This comparative analysis highlighted the diversity and uniqueness of Cys patterns across species.

Disulfide connectivity prediction

Disulfide connectivity was predicted using a PDB library of cone snail peptides, compiling all data available in the PDB database (http://www.rcsb.org). The library included 3D structures of novel peptides with more than 70% sequence similarity to known peptides, totaling 151 PDB files. An in-house Python script was used to predict connectivities, resulting in CSV files with comprehensive connectivity information. Connectivities were then compared across species, and additional analysis contrasted cone snail Cys connectivities with those of FDA-approved or clinical trials peptides.

ML model for therapeutic potential of conopeptides

The positive dataset consisted of known 35 venom-derived peptide drugs, either FDA-approved from DrugBank [34] or currently in clinical trials, with recognized therapeutic potential (Table 1). The negative dataset included 5405 sequences from ConoServer and UniProt, comprising venomous peptide sequences from cone snails, snakes, and other species (5405, 390, and 155 sequences, respectively).

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Overview of 35 venom-derived peptide drugs, including their sequences, pharmacological activities, and regulatory status (FDA-approved or in clinical trials).

https://doi.org/10.1371/journal.pone.0327578.t001

Due to the limited availability of FDA-approved or clinically trials venom-based drugs, 35 unannotated sequences were used as an unregulated or negative dataset, based on the number of Cys residues, frameworks, targets, and connectivity patterns (Fig 2).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Schematic representation of positive and negative datasets used in this study.

https://doi.org/10.1371/journal.pone.0327578.g002

The ML model for predicting the therapeutic potential of cone snail venom peptides was developed through six stepsas illustrated in Fig 3. To create the model, two datasets (positive and negative) were prepared.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Steps involved in the prediction of venom peptide therapeutic potential.

The first step involves the collection of venom peptide data through various databases. Data preprocessing involves cleaning and normalization of data by removing any redundancy followed by preparation of data for analysis. Feature isolation is performed to train the ML model via RF classifier. In model development, generation and optimization of model is carried out to classify individual peptides. Subsequently, predictions are made by applying selected features to identify peptides with high therapeutic potential. Finally, model validation is performed by analyzing and comparing the findings of model-derived peptide sequences with reported venom-derived FDA-approved drugs.

https://doi.org/10.1371/journal.pone.0327578.g003

Data preprocessing

Data preprocessing is a critical step in model development, transforming raw data into a clean and valid format to improve model accuracy. This process involves removing noise, deleting duplicates to prevent redundancy, eliminating any invalid characters, correcting errors, and managing outliers present in data [35]. Prior to applying ML models, data standardization is a crucial step, as it significantly impacts model performance. In this study, data preprocessing was performed using the StandardScaler module from the sklearn package [36], ensuring all features were appropriately scaled. Given the dataset imbalance, where Conus peptides vastly outnumber those from other venomous species, corrective measures were implemented to mitigate model bias toward the majority class. To address this, Stratified Sampling was applied, ensuring that both the training and test sets maintained proportional class distributions, thereby avoiding misleading evaluation results [37]. Furthermore, oversampling was applied using the RandomOverSampler module from the imbalanced-learn package [38]. This method increased the representation of the minority class, enhancing the model’s ability to learn effectively from all classes.

Hyperparameter tuning was conducted with GridSearchCV from sklearn to optimize model parameters [39]. Additionally, data filtration was based on Cys counts and connectivity patterns to maintain data consistency. Cys numbers, positions, and connectivity patterns for both positive and negative datasets were computed and compared using a Python script.

Feature isolation

To use an ML model for interpreting biological data, sequences must undergo featurization, converting variable data types into numerical forms for model training. Defining the feature set, or input space, is essential to the accuracy of any ML model [40]. For estimating the therapeutic potential of venom peptides, three feature representation strategies were applied in ML model development:

Amino acid composition

Amino acid composition (AAC) provides insights into peptide structure by determining the frequency or percentage of each amino acid in a sequence [41]. The descriptor values for the AAC feature set are represented by the following equation:

(1)

Where L is the length of peptide, D(i) is the descriptor value particular to amino acid i in the peptide, and n is the i^th amino acid’s frequency of occurrence.

The amino acid composition (AAC) for each peptide sequence was calculated using the ProteinAnalysis module from the ProtParam tool in the Bio.SeqUtils package [42] and used as a feature set for training the ML model.

Physicochemical property evaluation

Physicochemical properties are crucial for characterizing the therapeutic potential of venom peptides in terms of stability and solubility [43]. Key properties include charge, hydrophobicity, molecular weight, and isoelectric point, all of which significantly influence a peptide’s ability to bind to receptors, enzymes, and ion channels, thereby affecting its therapeutic activity. Amphipathicity is another important factor, as it enables peptides to penetrate cell membranes and target cellular locations effectively [44]. These properties help maintain the peptide’s stability in a biological environment, preventing aggregation and enhancing the predictive accuracy of the ML model for therapeutic peptides. For each peptide in the dataset, various physicochemical properties (listed in Table 2) were calculated using different modules of Bio.SeqUtils, resulting in the creation of a multi-dimensional feature vector.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Physiochemical properties are used as features in developing a ML model.

https://doi.org/10.1371/journal.pone.0327578.t002

Additional descriptor sets

Other descriptors included as input parameters for the ML model were hydrogen bond donors, acceptors, and LogP, calculated using the Descriptors function from the rdkit.chem package [49]. These features influence solubility, molecular interactions, and membrane permeability, which are critical for therapeutic efficacy, thereby enhancing the ML model’s ability to accurately predict the therapeutic potential of venom peptides [50]. These characteristics also govern the pharmacokinetic behavior of peptides in vivo and regulate their ability to bind to various biological targets. The pd.concat function from Pandas was used to combine all these features into a single dataset.

Random Forest classifier

The Random Forest (RF) classifier was employed to predict the therapeutic potential of venom peptides due to its robustness, high accuracy, and stability compared to a single decision tree [51]. RF is an ensemble classifier that integrates a decision tree algorithm, a supervised ML approach used for classification and regression. It minimizes overfitting risks by using limited hyperparameter inputs, balancing bias and variance, making it a valuable tool for prediction, modeling, and data analysis across various domains [52]. RF extends the Classification and Regression Tree (CART) method by applying bagging (bootstrap aggregation) and voting to determine classification results. The RF model consists of ‘N’ decision trees generated from the training dataset, aimed at converting multiple weak classifiers into a robust single classifier. The number of trees in the model is determined by the number of generated bootstrap samples. For each bootstrap sample, a decision tree is constructed following these rules: If there are M input variables, the predictor variables considered at each node are randomly chosen as m, where m is less than M. The variable m is selected randomly through the total M variables. To identify the best predictor variable from the chosen m, a measure of purity, such as Gini or entropy, is calculated. The Gini index, denoted as G_gini (D), is used to determine the optimal binary split point for each feature. G_gini reflects the uncertainty of the set D. In classification tasks with N classes, the G_gini index for a given sample set D is calculated with the following equation:

(2)

In eq (2), C_n represents the subset of samples in D that belong to the nth class [53]. When subsets are defined as follows:

(3)

The optimal split for the predictor variable m is utilized to split the nodes. The value of m remains fixed during the growth of the forest. Each tree is constructed to its full extent without pruning. The final output of the RF model is determined by aggregating the results of all classification trees through majority voting, selecting the most accurate prediction [53]. The most informative predictor variable provides the greatest contribution to the decision-making process. The generation of additional trees and their collective use in the decision-making process ensures a more reliable and robust outcome [54].

In this study, we utilized the sklearn package to classify peptides into two venomous sequence classes. The RFClassifier function was employed to build the model with optimal hyperparameters, and cross_val_score was used for cross-validation. The model was constructed using various threshold values, with the optimal threshold (yielding high accuracy and precision) chosen for the final model. The dataset was divided into training (70%) and testing/validation (30%) subsets. Optimizing the RF algorithm is essential for extracting high-quality features and selecting parameters, significantly reducing the model’s generalization error and improving classification accuracy. Based on ideal hyperparameters, we set 100 optimal estimators (trees). Predictions were made using the rf.predict function and evaluated through various performance metrics.

Support Vector Machine (SVM) classifier

The Support Vector Machine (SVM) classifier, a robust ML model for classification and regression, was chosen for its ability to handle high-dimensional data while minimizing overfitting. SVM identifies an optimal hyperplane that maximizes the margin between classes by utilizing support vectors, which are the data points closest to the hyperplane and define the decision boundaries [55]. Its theoretical foundation is rooted in optimization principles and statistical learning theory [56]. SVM is particularly effective when training data is limited, and traditional statistical methods struggle to guarantee an optimal solution based on large datasets [57].

SVM can be categorized into linear and nonlinear types. Linear SVM utilizes a linear kernel function to separate data with a straight boundary, offering computational efficiency for applications such as drug design and document classification [58]. On the other hand, nonlinear SVM employs kernel functions to map input features into a higher-dimensional space, enabling better classification of complex datasets [59]. Various kernel functions, including radial basis function and polynomial, were selected according to task-specific requirements.

The SVM classifier was implemented using the sklearn package, while feature calculations were performed using the Bio.SeqUtils package. To address class imbalance and mitigate overfitting, the SMOTE (Synthetic Minority Oversampling Technique) algorithm was applied to increase the representation of the minority class. The SVC function was used to build the model with the best hyperparameters, which were identified using GridSearchCV and retrieved via grid_search.best_params. The model was trained and evaluated across different threshold values, selecting the optimal threshold where both accuracy and precision were maximized. Finally, predictions were made using best_svm.predict, and various performance metrics were employed to assess the model’s accuracy.

Performance metrics

The performance evaluation metrics included two types of values: predicted and actual values.

The following metrics were used to assess model performance.

Confusion matrix

A confusion matrix was generated based on the test dataset outcomes. This table provides insights into the model’s accuracy and is also referred to as an error matrix or contingency matrix. It can be utilized for both binary and multi-class classification problems [60]. The confusion matrix comprises four components:

Classification report

A classification report was generated to evaluate the quality of the classification model, calculating recall, precision, F1 score, and support scores. The following equations (4–7) were used to compute these metrics:

(4)

(5)

(6)

(7)

Receiver Operating Characteristic (ROC)

The Area Under the Receiver Operating Characteristic Curve (AUC-ROC) [61] serves as a performance metric for binary classification problems. It is plotted to estimate the True Positive (TP) rate (sensitivity) against the False Positive (FP) rate (1 – specificity) across various threshold settings. The AUC-ROC value ranges from 0 to 1, with higher values indicating better model performance. The ROC curve is used to evaluate how effectively the model distinguishes between positive and negative classes. Ideally, the curve should reside in the upper left corner of the plot, which signifies a high True Positive Rate (TPR) – correctly identifying most positive cases – coupled with a low False Positive Rate (FPR) – indicating few false positives [62].

SHapley Additive exPlanations (SHAP)

SHAP is employed to assess the contribution of each feature to the model’s predictions, identifying the most influential features and their impact on the Random Forest (RF) model’s results. Each feature is assigned a relevance value that reflects its contribution to the output [63], based on principles of game theory. Positive SHAP values indicate features that positively influence predictions, while negative SHAP values denote features that have a detrimental effect; however, in some contexts, negative effects can also be significant. While SHAP analysis enhances the explainability of machine learning models, it has inherent limitations [64]. In SHAP analysis, each feature is considered independent of all others, and it solely clarifies a feature’s contribution to the model’s prediction without implying causation. Therefore, the results from SHAP should be interpreted with caution.

Molecular docking analysis

3D structures of Conus striatus peptide (PDB ID: 1FYG) and Conus magus Ziconotide-bound human N-type voltage-gated calcium channel (Ca_v2.2) (PDB ID: 7VFU) were isolated through RCSB PDB (www.rcsb.org). Interaction of Ca_v2.2 and Conus striatus peptide (referred as peptide-cs) had been evaluated through molecular docking. PatchDock server [65] and embedded refinement tool FireDock [66] was utilized to describe their interaction patterns. PatchDock accomplishes docking analysis through a segmentation algorithm based on the structure geometry. It recapitulates docking transformations that yield good complementary molecular shapes based on a small number of steric clashes and wide interface areas. The PatchDock algorithm classifies the Connolly dot surface representation of the protein molecules as concave, convex, and flat patches [67]. The complementary patches are matched to generate the candidate transformations. A scoring function also evaluates each candidate transformation, which considers both the atomic desolvation energy and geometric fit [68] to measure each candidate transformation. Finally, the most suitable candidate solution is selected among the redundant solutions based on RMSD (Root Mean Square Deviation) clustering. Overall, three major steps are followed in the PatchDock analysis: (i) surface patch matching, (ii) molecular shape representation, and (iii) filtering and scoring [69]. PDBSum [70] and UCSF Chimera version 1.18 were employed to analyze the interactions.

Categorization of sequences based on Cys count

Numerous sequences in our dataset have been classified as rich or poor based on the Cys number. There is no fixed threshold below which a sequence is considered as Cys-poor; instead, sequences with low Cys counts (<= 3) are categorized as poor, while >3 counts are designated as Cys-rich. In our dataset, we isolated 5421 and 563 Cys-rich and Cys-poor sequences, respectively. Fig 4 represents a portion of cone snail sequence classification based on the Cys count. A list of all cone snail sequences having low and high Cys counts is presented in the supplementary data (cys_poor.csv and cys_rich.csv). Evidently, sequences of common species exist in both categories indicating that depending on the amount of Cys count, multiple sequences of similar species may be categorized as either rich or poor. For instance, Lt0.12precursor and Lt5.15precursor [partial] of Conus litteratus exhibit varying Cys content. Although produced by the same species, the sequences differ due to variations in their amino acid composition (with one being partial), leading to differences in peptide structures and functions.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Plot illustrating sequence categorization based on Cys count.

Data along X-axis represents species names along with their unique IDs according to ConoServer database. Cys count is present at the Y-axis. Species having the word “mature” next to its name indicates that its subtype’s sequence is mature; species without this word displays a protein precursor. (A) Species have low Cys count. (B) Cys rich species. Certain species sequences are present multiple times with different Cys counts.

https://doi.org/10.1371/journal.pone.0327578.g004

Prediction of novel Cys patterns and frameworks

In order to differentiate among distinct Cys patterns and frameworks (known and novel), sequence data of 119 cone snail species gathered via ConoServer were compared to the reported ones. Each pattern and framework is labeled as known upon its matching to a previously reported sequence pattern and framework. A significant portion of patterns and frameworks were considered as novel based on this comparison. These patterns and frameworks underwent preprocessing, which involved organizing them based on species name, identifying the corresponding scaffolds (specific Cys frameworks are represented by unique identifiers), and generating a barcode for each pattern and framework to facilitate their classification and comparison. Through this strategy, we extracted 82 cone snail species having novel Cys patterns and frameworks. Due to huge data volume, scaffold-based characterization revealed multiple instances of common Cys patterns and frameworks (Fig 5). The sequence details of these barcodes are listed in S1 Table.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Barcode generation for Cone snail Cys patterns.

For specific species, each single-colored strip represents a unique pattern. Each scaffold represents a unique set of patterns, as we have a large number of residues. We aim to visualize and analyze these patterns by representing them in the form of strips for clearer interpretation. For example, C-CC-C-C pattern is indicated by scaffold S1 in blue color. C-C-C-CC-C-C pattern is represented by S2 in the gray color. Similarly, C-C-C-C-C and C-C patterns are designated by S3 and S4 in yellow and red colors, respectively, while S5 scaffold in green indicates Cys. The intervening region between two scaffolds is indicated by pink color. The separation of two patterns belonging to same specie is indicated by black color comma (,). The number of unique patterns is labeled in small parenthesis along with the names of species.

https://doi.org/10.1371/journal.pone.0327578.g005

Multiple cone snail species exhibited distinct patterns having similar Cys count and arrangement despite having different sequences, indicating the presence of close association and selection among several species. This relationship can be explained through similarities in ancestry, evolution, geographic proximity, and other elements. We selected 9 cone snail species sharing a common Cys pattern (C-C-C-CC-C-C) and performed sequence alignment and phylogenetic analysis, which suggested that these species are related or share a common ancestor (Fig 6). Evidently, Conus ammiralis and Conus dalli shells exhibit similar coloring patterns, which may be due to a common genetic background for pigmentation. As both species are localized in similar ecological niches and are subjected to equivalent selective pressure, their shell designs have been evolved in a convergent manner. Likewise, it has previously been reported that Conus aulicus and Conus omaria share color pattern similarities in their shells as a result of divergence from the same ancestral root [71]. In terms of venom composition, Conus achatinus, Conus aristophanes, Conus aulicus, and Conus aurisacus venoms contain various conopeptides but all of them belong to α-conotoxin class. α-conotoxins are fascinating because they specifically antagonize Nicotinic acetylcholine receptors (nAChRs). These peptides exhibit exquisite subtype selectivity and have been essential in characterizing native nAChR isoforms involved in neurotransmitter release modulation, Parkinson disease pathophysiology, and nociception [72].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Comparative alignment and phylogenetic analysis of a unique Cys pattern in cone snail species.

(A) Depiction of cone snail species (a to i) with a similar novel pattern. (B) Comparative alignment of pattern sequences of mentioned cone snail species. * denotes conserved alignment. The common residues are indicated in distinct colors. Barcode is indicated through the combination of S4 and S1 scaffold. (C) Phylogenetic analysis indicates that these species share a similar ancestral root highlighted by a red circle.

https://doi.org/10.1371/journal.pone.0327578.g006

Cys connectivity analysis

An in-house PDB library of cone snail peptides was utilized to confer the Cys connecting pairs. The connectivities of all available PDB files were computed and compared. Cone snail species having similar Cys count, patterns and connectivity pairs as of approved venom-derived drug were mentioned in S1 Fig. For instance, C-C-CC-C-C pattern of Conus magus extracted from sequence was exactly similar to that of Conus magus in terms of connectivity (Fig 7). We computed connectivity data for individual cone snail species through 151 currently available PDB files of cone snail peptides and presented in S2 Table.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Comparative overview of an FDA-approved drug Ziconotide (PDB ID: 1dw4) and cone snail peptides having similar connectivity pairs.

C1 exhibits connectivity with C16 (red), C8 shows connectivity with C20 (green), and C15 shows connectivity with C25 (blue) and C26 (purple) depending on species.

https://doi.org/10.1371/journal.pone.0327578.g007

Next, we gathered the PDB IDs of available cone snail venom-derived FDA approved drugs for connectivity analysis and compared with our data. The purpose of this process was to isolate novel peptides on the basis of common Cys patterns and their connectivities. For example, Ziconotide, an FDA-approved peptide isolated from Conus magus with a PDB ID: 1dw4, which has been reported for the management of severe chronic pain in patients refractory to other treatments and for whom intrathecal therapy is warranted [16], shared the same connecting pairs with Conus striatus and omega conotoxins_FVI-A (Fig 7). Given that Conus striatus peptide exhibits a high conservation and similar connectivity to ziconotide, our study proposes that it may likewise be utilized to treat pain. Similarly, Conus purpurascens has a comparable similarity in connectivity pattern with ziconotide, except the presence of a single connecting pair, which suggests that Conus purpurascens is a viable option that may result in a positive outcome. There are multiple such cases in our analysis, which are mentioned in S3 Table.

Model performance evaluation

To determine the therapeutic potential of venom peptide sequences, the effectiveness of RF model was assessed through ten distinct threshold values (Table 3). The accuracy, precision, recall, TP, TN, FP, and FN values were increased following the threshold. The model performed efficiently and accurately at the threshold value of 0.5. As the threshold increased further, performance began to decline, indicating an uneven data distribution.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. RF performance evaluation at different threshold values. A row with a bold text indicates the optimal threshold at which the model performed ideally.

https://doi.org/10.1371/journal.pone.0327578.t003

Through ROC, we may assess how well the model discriminates positive and negative instances. ROC curve suggested an AUC of 0.97 for RF model (Fig 8A), indicating its reliability. Similarly, test data set confusion matrix (Fig 8B) revealed higher number of TP and TN, further strengthening the model accuracy.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. ROC and confusion matrix analysis to evaluate the model performance.

(A) ROC curve indicates that with the increase of data, model performance is improved. Rapid fluctuations in the TP and FP of ROC curve have been observed because even slight modifications in the predictions may lead to significant variations in these metrics. It is due to small sample sizes for positive and negative classes. AUC is directly proportional to the model performance. Given that the AUC is 0.97, the model performance is significant, and it favors the randomly selected positive instances more. (B) Confusion matrix reveals a balance between the positive and negative datasets, resulting in a high similarity between TP and TN. Dark blue color indicates TP and TN, while light color shows FP and FN. Only 2 FN instances are predicted by model, indicating that although the sequences belong to a positive class, the model predicts them in the negative class. There is no sequence predicted as FP. 9 sequences are predicted as TN and 10 are identified as TP. Overall, the model generates more FNs than FPs.

https://doi.org/10.1371/journal.pone.0327578.g008

A classification report summarizes the key performance metrics for each class in an ML model. It includes the weighted average of metrics across all classes, as well as precision, recall, F1-score, and support value for each class [73]. This report offers a comprehensive analysis of the model performance per class, highlighting the balance between recall and precision. Additionally, it shows the number of instances (support) for each class, indicating possible dataset or class imbalances. The classification report of our RF model was generated and listed in Table 4.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Classification report representing various performance metrics of RF model.

https://doi.org/10.1371/journal.pone.0327578.t004

Precision, defined as the proportion of TP predictions among all predictions, serves as a critical performance metric. As shown in Table 3, a high precision value of 0.8333 indicates that FN predictions, such as class_1 sequences predicted as 0, were rare. For class_0, the precision value was 1, signifying no FP predictions in this class. The lowest precision value was 0.8333 for class_1, while the highest was 1 for class_0. Notably, class_1 exhibited the highest recall value, indicating very few or no FP predictions. The lowest recall value for class_0 was 0.8182, with the highest for class_1 at 1. This interplay between precision and recall suggests that when one metric is low, the other tends to be high, thereby balancing overall predictions.

The harmonic mean of precision and recall, known as the F1-Score, was 0.9 in most cases, reflecting a balanced evaluation of model performance. Support values, representing the number of actual occurrences for each class in the dataset, confirmed that there was no imbalance. The accuracy of the RF predicted model was 0.9047, with a macro average of 0.9166 and a weighted average of 0.9206, indicating an overall ideal model quality.

SHAP analysis

In this study, multiple features were employed to train the RF model, with their importance contingent upon the specific task. The contribution of each feature to the model’s predictions, quantified by SHAP values, helps identify the most crucial features and their influence on overall results. Ideally, features with positive SHAP values significantly enhance model performance; however, this depends on the nature of the task. Positive values do not always take precedence over negative ones. If numerous features negatively impact the model but are of lesser importance, their overall effect will be less significant compared to more influential positive values. Feature importance and summary plots of the SHAP values are illustrated in Fig 9.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Feature importance plot and summary plot representing SHAP values for RF model features.

(A) Features are presented on the X-axis and their importance has been mentioned on the Y-axis. Among all features Glycine, Cysteine, Lysine, and isoelectric point exhibit more impact on the model performance. In contrast, Valine, Methionine, and Phenylalanine are the least important. (B) The SHAP values are depicted at the X-axis. The feature names are shown at the Y-axis in ascending order with respect to importance. Every point in the plot has a color that corresponds to its related feature value; blue points exhibit low values and red points show high values. A data row (original dataset) is represented by each point. Glycine, isoelectric point, Lysine, and Cysteine are the most important features as described in the feature importance plot. These features affect the model in both positive and negative ways; however, the negative way (blue color) suggests that their affect is negligible. Naturally, for the functioning of certain features, they must have high negative values. For example, higher stability requires a lower instability index [74]. Higher values; however, point to a possible instability. Although the instability index value for our predicted model is negative, indicating a negative impact on the model, in reality, instability index negative value is beneficial for the stability of peptide sequence. More stability is indicated by lower values, and possible instability is indicated by higher values. Overall, this result suggested that the model has the potential of a reliable performance.

https://doi.org/10.1371/journal.pone.0327578.g009

Comparative analysis between RF and SVM

To evaluate the effectiveness of both RF and SVM models, a comprehensive comparative analysis of their performance metrics was performed. Various evaluation metrics, including accuracy, precision, recall, F1-score, TP, TN, FP, and FN, were employed to thoroughly assess their predictive capabilities. A comparative plot was generated to visualize the differences in performance between the two models (Fig 10A).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 10. Comparative model performance analysis.

(A) Detailed comparison of performance metrics between the RF and SVM models, with performance metrics plotted on the X-axis and their corresponding values on the Y-axis. The RF model (blue) outperformed the SVM model (red) across multiple metrics, including accuracy, precision, recall, F1-score, TN, and FN, highlighting its superior reliability in predicting the therapeutic potential of venom peptides. (B) AUC comparison bar plot between the RF and SVM models, with models represented on the X-axis and their corresponding AUC values on the Y-axis. The RF model (blue) demonstrates better performance with a higher AUC value compared to the SVM model (red).

https://doi.org/10.1371/journal.pone.0327578.g010

The ensemble approach of RF, which combines multiple decision trees, enhances its generalization ability by reducing variance and minimizing overfitting. Additionally, RF effectively selects relevant features by randomly choosing subsets at each node [75], improving classification performance by considering multiple peptide characteristics such as shape, length, width, area, and perimeter [76]. Moreover, RF is commonly used for describing non-linear spatial data [77], making it a strong alternative to traditional linear models. Both models were further assessed using the AUC of the ROC curve. The RF model achieved a higher AUC value of 0.97, outperforming the SVM model, which attained an AUC of 0.95 (Fig 10B).

As an ensemble learning method, RF combines multiple decision trees to improve accuracy and generalization, reducing noise and maintaining high performance even with limited data [78]. Unlike SVM, which requires manual feature selection and is prone to overfitting, RF automatically selects and ranks important features, thus enhancing model efficiency [79]. For instance, in audio classification, RF achieved 90% accuracy, while SVM only reached 60% [80]. Additionally, RF has shown superior performance in regional land cover mapping, providing higher accuracy, kappa value, and individual accuracy. It efficiently handles large, coarse-resolution datasets and reduces confusion in classifying mixed classes-tasks where SVM struggles due to memory and computational limitations [81].

While models like tree-based gradient boosting methods, such as XGBoost and Gradient Boosting Machines (GBM), may offer higher accuracy in certain contexts, RF remains more stable and interpretable, making it ideal for understanding the complex interactions in venom peptides [82]. Moreover, RF provides clear feature importance, which helps in understanding the contribution of different peptide features to therapeutic efficacy [83]. These qualities make RF a well-suited model for predicting the therapeutic potential of venom peptides.

Peptide therapeutic potential prediction using RF model

Based on the significant features identified during model training, the RF model was employed to predict the therapeutic potential of specific peptide sequences. The predictions utilized features from the training dataset, including amino acid composition, sequence length, isoelectric properties, physicochemical properties, and molecular weight. Sequences deemed to have therapeutic potential were assigned a value of 1, while those lacking therapeutic significance were given a value of 0. The input dataset, comprising 5,383 sequences with known and unique Cys patterns and frameworks, yielded reliable model predictions (S4 Table). A portion of the model predictions based on the negative dataset is listed in Table 5.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. RF model predictions based on a negative dataset.

https://doi.org/10.1371/journal.pone.0327578.t005

Structural analysis

To validate the model prediction accuracy, we selected a peptide sequence (peptide-cs) of Conus striatus, which was predicted as a positive contributor for therapeutics by our model. Alignment results (Fig 11) showed that peptide-cs sequence was closely matched to a previously FDA approved peptide sequence (Ziconotide, Prialt®) isolated from Conus magus [84].

Download:

• PNG
  larger image
• TIFF
  original image

Fig 11. Sequence alignment between approved drug (1dw4) and peptide-cs (1fyg) sequence.

* indicates the column in which fully conserved residues are present. 18 out of 25 residues are exactly similar.

https://doi.org/10.1371/journal.pone.0327578.g011

PatchDock specific energy value for Ca_v2.2 and peptide-cs complex was −15.28 kcal/mol. Ca_v2.2 binding residues were quite similar for both peptides. Ziconotide exhibited binding with ASP1629, ALA667, ASP664, THE544, PRO642, HIS671, and VAL688, while peptide-cs binding was detected with ASP1629, ALA667, THR664, THR643, PRO642, HIS671, and THR641 residues of Ca_v2.2 (Fig 12). The interaction analysis was conducted using PDBSum [70] to analyze salt bridges, disulfide bonds, hydrogen bonds, and non-bonded contacts. The interaction details are listed in S5 Table. These findings suggest that Conus striatus peptide-cs may exhibit promising therapeutic results like Ziconotide.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 12. Binding analysis of Ziconotide (PDB ID: 1dw4) and Conus striatus peptide-cs (PDB ID: 1fyg) against Ca_v2.2 channel (PDB ID: 7vfu).

(A) Conus magus peptide is shown in magenta color and (B) Ziconotide is indicated in cyan color. Ca_v2.2 ion channel is indicated in light pink colored ribbon representation. (C and D) Surface representation of Ca_v2.2 channel demonstrating binding cavity (pink) for Ziconotide and peptide-cs (green ribbon), respectively.

https://doi.org/10.1371/journal.pone.0327578.g012